advertisement
▼
Scroll to page 2
of
647
42300 Rev 3.12 - July 14, 2015 BKDG for AMD Family 15h Models 10h-1Fh Processors Cover page BIOS and Kernel Developer’s Guide (BKDG) for AMD Family 15h Models 10h-1Fh Processors Advanced Micro Devices 1 42300 Rev 3.12 - July 14, 2015 BKDG for AMD Family 15h Models 10h-1Fh Processors © 2011–2015 Advanced Micro Devices, Inc. All rights reserved. The information contained herein is for informational purposes only, and is subject to change without notice. While every precaution has been taken in the preparation of this document, it may contain technical inaccuracies, omissions and typographical errors, and AMD is under no obligation to update or otherwise correct this information. Advanced Micro Devices, Inc. makes no representations or warranties with respect to the accuracy or completeness of the contents of this document, and assumes no liability of any kind, including the implied warranties of noninfringement, merchantability or fitness for particular purposes, with respect to the operation or use of AMD hardware, software or other products described herein. No license, including implied or arising by estoppel, to any intellectual property rights is granted by this document. Terms and limitations applicable to the purchase or use of AMD's products are as set forth in a signed agreement between the parties or in AMD's Standard Terms and Conditions of Sale. Trademarks AMD, the AMD Arrow logo, and combinations thereof are trademarks of Advanced Micro Devices, Inc. Other product names used in this publication are for identification purposes only and may be trademarks of their respective companies. ARM, AMBA, Cortex and TrustZone are registered trademarks of ARM Limited. HDMI is a trademark of HDMI Licensing, LLC. HyperTransport is a licensed trademark of the HyperTransport Technology Consortium. Microsoft and DirectX are registered trademarks of Microsoft Corporation. MMX is a trademark of Intel Corporation. PCI Express and PCIe are registered trademarks of PCI-Special Interest Group (PCI-SIG). Dolby Laboratories, Inc. Manufactured under license from Dolby Laboratories. Rovi Corporation This device is protected by U.S. patents and other intellectual property rights. The use of Rovi Corporation's copy protection technology in the device must be authorized by Rovi Corporation and is intended for home and other limited pay-per-view uses only, unless otherwise authorized in writing by Rovi Corporation. Reverse engineering or disassembly is prohibited. USE OF THIS PRODUCT IN ANY MANNER THAT COMPLIES WITH THE MPEG ACTUAL OR DE FACTO VIDEO AND/OR AUDIO STANDARDS IS EXPRESSLY PROHIBITED WITHOUT ALL NECESSARY LICENSES UNDER APPLICABLE PATENTS. SUCH LICENSES MAY BE ACQUIRED FROM VARIOUS THIRD PARTIES INCLUDING, BUT NOT LIMITED TO, IN THE MPEG PATENT PORTFOLIO, WHICH LICENSE IS AVAILABLE FROM MPEG LA, L.L.C., 6312 S. FIDDLERS GREEN CIRCLE, SUITE 400E, GREENWOOD VILLAGE, COLORADO 80111. 2 42300 Rev 3.12 - July 14, 2015 BKDG for AMD Family 15h Models 10h-1Fh Processors Table of Contents 1 Overview . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 18 1.1 Intended Audience . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 18 1.2 Reference Documents . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 18 1.3 Conventions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 18 1.3.1 Numbering . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 18 1.3.2 Arithmetic And Logical Operators . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 19 1.4 Definitions. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 20 1.5 Changes Between Revisions and Product Variations . . . . . . . . . . . . . . . . . . . . . . . . 25 1.5.1 Revision Conventions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 25 1.5.2 Major Changes Relative to Family 15h Models 00h-0Fh Processors . . . . . 25 1.5.2.1 Major Changes to Core/NB Performance Counters . . . . . . . . . . . . . . . . . 26 1.5.3 Changes For Revision RL-A1 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 26 2 Functional Description . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 27 2.1 Processor Overview . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 27 2.2 System Overview . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 27 2.3 Processor Initialization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 28 2.3.1 BSC initialization. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 28 2.3.2 AP initialization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 28 2.3.3 Using L2 Cache as General Storage During Boot . . . . . . . . . . . . . . . . . . . . 29 2.4 Core. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 31 2.4.1 Compute Unit. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 31 2.4.1.1 Registers Shared by Cores in a Compute Unit . . . . . . . . . . . . . . . . . . . . . 31 2.4.2 Virtual Address Space . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 31 2.4.3 Processor Cores and Downcoring . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 31 2.4.3.1 Software Downcoring using D18F3x190[DisCore] . . . . . . . . . . . . . . . . . 32 2.4.4 Physical Address Space . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 32 2.4.5 System Address Map . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 32 2.4.5.1 Memory Access to the Physical Address Space . . . . . . . . . . . . . . . . . . . . 32 2.4.5.1.1 Determining Memory Type . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 32 2.4.5.1.2 Determining The Access Destination for Core Accesses . . . . . . . . . . . . 33 2.4.6 Timers . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 33 2.4.7 Implicit Conditions for TLB Invalidation . . . . . . . . . . . . . . . . . . . . . . . . . . 33 2.4.8 Interrupts . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 34 2.4.8.1 Local APIC . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 34 2.4.8.1.1 Detecting and Enabling . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 34 2.4.8.1.2 APIC Register Space . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 34 2.4.8.1.3 ApicId Enumeration Requirements . . . . . . . . . . . . . . . . . . . . . . . . . . . . 34 2.4.8.1.4 Physical Destination Mode . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 35 2.4.8.1.5 Logical Destination Mode . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 35 2.4.8.1.6 Interrupt Delivery . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 35 2.4.8.1.7 Vectored Interrupt Handling . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 36 2.4.8.1.8 Interrupt Masking . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 36 2.4.8.1.9 Spurious Interrupts . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 36 2.4.8.1.10 Spurious Interrupts Caused by Timer Tick Interrupt . . . . . . . . . . . . . . . 36 2.4.8.1.11 Lowest-Priority Interrupt Arbitration . . . . . . . . . . . . . . . . . . . . . . . . . . . 37 2.4.8.1.12 Inter-Processor Interrupts . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 37 2.4.8.1.13 APIC Timer Operation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 37 3 42300 Rev 3.12 - July 14, 2015 2.5 BKDG for AMD Family 15h Models 10h-1Fh Processors 2.4.8.1.14 Generalized Local Vector Table . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 37 2.4.8.1.15 State at Reset . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 38 2.4.8.2 System Management Mode (SMM) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 38 2.4.8.2.1 SMM Overview . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 38 2.4.8.2.2 Operating Mode and Default Register Values . . . . . . . . . . . . . . . . . . . . 38 2.4.8.2.3 SMI Sources And Delivery . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 39 2.4.8.2.4 SMM Initial State . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 39 2.4.8.2.5 SMM Save State . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 40 2.4.8.2.6 Exceptions and Interrupts in SMM . . . . . . . . . . . . . . . . . . . . . . . . . . . . 45 2.4.8.2.7 The Protected ASeg and TSeg Areas . . . . . . . . . . . . . . . . . . . . . . . . . . . 45 2.4.8.2.8 SMM Special Cycles . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 45 2.4.8.2.9 Locking SMM . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 46 2.4.8.2.10 Synchronizing SMM Entry (Spring-Boarding) . . . . . . . . . . . . . . . . . . . 46 2.4.9 Secure Virtual Machine Mode (SVM). . . . . . . . . . . . . . . . . . . . . . . . . . . . . 47 2.4.9.1 BIOS support for SVM Disable . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 47 2.4.10 CPUID Instruction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 48 2.4.10.1 Multi-Core Support . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 48 Power Management. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 49 2.5.1 Processor Power Planes And Voltage Control. . . . . . . . . . . . . . . . . . . . . . . 49 2.5.1.1 Serial VID Interface . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 49 2.5.1.1.1 SVI2 Features . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 49 2.5.1.2 Internal VID Registers and Encodings . . . . . . . . . . . . . . . . . . . . . . . . . . . 50 2.5.1.2.1 MinVid and MaxVid Check . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 50 2.5.1.3 Low Power Features . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 50 2.5.1.3.1 PSIx_L Bit . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 50 2.5.1.3.1.1 BIOS Requirements for PSI0_L . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 50 2.5.1.3.1.2 BIOS Requirements for PSI1_L . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 52 2.5.1.3.2 Low Power Voltages . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 52 2.5.1.4 Voltage Transitions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 52 2.5.1.4.1 Hardware-Initiated Voltage Transitions . . . . . . . . . . . . . . . . . . . . . . . . . 52 2.5.1.4.2 Software-Initiated Voltage Transitions . . . . . . . . . . . . . . . . . . . . . . . . . 52 2.5.1.4.2.1 Software-Initiated NB Voltage Transitions . . . . . . . . . . . . . . . . . . . . 52 2.5.1.4.2.2 Software-Initiated Core Voltage Transitions . . . . . . . . . . . . . . . . . . . 53 2.5.2 Frequency and Voltage Domain Dependencies. . . . . . . . . . . . . . . . . . . . . . 53 2.5.2.1 Dependencies Between Cores . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 53 2.5.2.2 Dependencies Between Subcomponents on VDDNB . . . . . . . . . . . . . . . . 54 2.5.2.3 BIOS Requirements for Power Plane Initialization . . . . . . . . . . . . . . . . . . 54 2.5.3 CPU Power Management. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 54 2.5.3.1 Core P-states . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 54 2.5.3.1.1 Application Power Management (APM) . . . . . . . . . . . . . . . . . . . . . . . . 54 2.5.3.1.2 Core P-state Naming and Numbering . . . . . . . . . . . . . . . . . . . . . . . . . . 55 2.5.3.1.2.1 Software P-state Numbering . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 55 2.5.3.1.2.2 Hardware P-state Numbering . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 56 2.5.3.1.3 Core P-state Control . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 56 2.5.3.1.4 Core P-state Visibility . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 57 2.5.3.1.5 Core P-state Limits . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 57 2.5.3.1.6 Core P-state Transition Behavior . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 57 2.5.3.1.7 BIOS Requirements for Core P-state Initialization and Transitions . . . 58 2.5.3.1.8 Processor-Systemboard Power Delivery Compatibility Check . . . . . . . 59 2.5.3.1.9 BIOS COF and VID Requirements After Warm Reset . . . . . . . . . . . . . 60 2.5.3.1.9.1 Core Maximum P-state Transition Sequence After Warm Reset . . . . 61 4 42300 Rev 3.12 - July 14, 2015 2.6 BKDG for AMD Family 15h Models 10h-1Fh Processors 2.5.3.1.9.2 Core Minimum P-state Transition Sequence After Warm Reset . . . . 61 2.5.3.1.9.3 ACPI Processor P-state Objects . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 61 2.5.3.1.9.4 Fixed ACPI Description Table (FADT) Entries . . . . . . . . . . . . . . . . . 63 2.5.3.1.9.5 XPSS (Microsoft® Extended PSS) Object . . . . . . . . . . . . . . . . . . . . . 63 2.5.3.2 Core C-states . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 63 2.5.3.2.1 C-state Names and Numbers . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 63 2.5.3.2.2 C-state Request Interface . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 63 2.5.3.2.3 C-state Actions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 64 2.5.3.2.3.1 C-state Probes and Cache Flushing . . . . . . . . . . . . . . . . . . . . . . . . . . . 64 2.5.3.2.3.2 Core C1 (CC1) State . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 64 2.5.3.2.3.3 Core C6 (CC6) State . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 64 2.5.3.2.3.4 Package C6 (PC6) State . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 65 2.5.3.2.4 C-state Request Monitors . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 65 2.5.3.2.4.1 FCH Messaging . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 65 2.5.3.2.4.2 Timer Tick Monitor . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 66 2.5.3.2.4.3 Cache Flush On Halt Saturation Counter . . . . . . . . . . . . . . . . . . . . . . 66 2.5.3.2.5 Exiting C-states . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 66 2.5.3.2.6 ACPI Processor C-state Objects . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 66 2.5.3.2.6.1 _CST . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 66 2.5.3.2.6.2 _CSD . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 67 2.5.3.2.6.3 _CRS . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 67 2.5.3.2.6.4 Fixed ACPI Description Table (FADT) Entries . . . . . . . . . . . . . . . . . 67 2.5.3.2.7 BIOS Requirements for Initialization . . . . . . . . . . . . . . . . . . . . . . . . . . . 67 2.5.3.3 Effective Frequency . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 67 2.5.4 NB Power Management. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 68 2.5.4.1 NB P-states . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 68 2.5.4.1.1 NB P-state Transitions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 68 2.5.4.1.2 BIOS NB P-state Configuration . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 69 2.5.4.1.2.1 NB P-state COF and VID Synchronization After Warm Reset . . . . . 69 2.5.4.1.2.2 NB P-state Voltage Adjustment for MEMCLK . . . . . . . . . . . . . . . . . 70 2.5.4.1.2.3 NB P-state Configuration for Runtime . . . . . . . . . . . . . . . . . . . . . . . . 70 2.5.4.2 NB C-states . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 70 2.5.5 P-state Bandwidth Requirements. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 71 2.5.6 GPU and Root Complex Power Management . . . . . . . . . . . . . . . . . . . . . . . 71 2.5.6.1 Dynamic Power Management (DPM) . . . . . . . . . . . . . . . . . . . . . . . . . . . . 71 2.5.6.1.1 Activity Monitors . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 72 2.5.6.1.2 SCLK DPM . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 72 2.5.6.1.3 LCLK DPM . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 72 2.5.6.2 GPU and Root Complex Power Gating . . . . . . . . . . . . . . . . . . . . . . . . . . . 72 2.5.7 DRAM Power Management . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 73 2.5.7.1 Memory P-states . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 73 2.5.7.2 DRAM Self-Refresh . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 73 2.5.7.3 Stutter Mode . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 74 2.5.7.3.1 System BIOS Requirements for Stutter Mode Operation During POST 74 2.5.7.4 EVENT_L . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 74 2.5.8 System Power Management. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 74 2.5.8.1 S-states . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 75 2.5.8.1.1 ACPI Suspend to RAM State (S3) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 75 2.5.9 Bidirectional Application Power Management (BAPM). . . . . . . . . . . . . . . 75 2.5.9.1 Hybrid Boost . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 75 Performance Monitoring . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 77 5 42300 Rev 3.12 - July 14, 2015 2.7 2.8 2.9 BKDG for AMD Family 15h Models 10h-1Fh Processors 2.6.1 Core Performance Monitor Counters. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 77 2.6.2 NB Performance Monitor Counters . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 78 2.6.3 Instruction Based Sampling (IBS) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 79 Configuration Space . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 80 2.7.1 MMIO Configuration Coding Requirements. . . . . . . . . . . . . . . . . . . . . . . . 80 2.7.2 MMIO Configuration Ordering . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 80 2.7.3 Processor Configuration Space . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 81 Northbridge (NB) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 82 2.8.1 NB Architecture . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 82 2.8.2 NB Routing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 82 2.8.2.1 Address Space Routing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 82 2.8.2.1.1 DRAM and MMIO Memory Space . . . . . . . . . . . . . . . . . . . . . . . . . . . . 82 2.8.2.1.2 IO Space . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 83 2.8.2.1.3 Configuration Space . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 83 DRAM Controllers (DCTs) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 84 2.9.1 DCT Configuration Registers . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 87 2.9.2 DDR Pad to Processor Pin Mapping . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 87 2.9.2.1 DDR Chip to Pad Mapping . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 88 2.9.3 DRAM Controller Direct Response Mode . . . . . . . . . . . . . . . . . . . . . . . . . 89 2.9.4 DRAM Data Burst Mapping . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 89 2.9.5 DCT/DRAM Initialization and Resume . . . . . . . . . . . . . . . . . . . . . . . . . . . 90 2.9.5.1 Low Voltage DDR3 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 90 2.9.5.2 NB P-state Specific Configuration . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 91 2.9.5.3 Memory P-state Specific Configuration . . . . . . . . . . . . . . . . . . . . . . . . . . 92 2.9.5.4 DDR Phy Initialization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 92 2.9.5.4.1 Phy Voltage Level Programming . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 93 2.9.5.4.2 DRAM Channel Frequency Change . . . . . . . . . . . . . . . . . . . . . . . . . . . . 93 2.9.5.4.2.1 Requirements for DRAM Frequency Change During Training . . . . . 94 2.9.5.4.3 Phy Fence Programming . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 94 2.9.5.4.3.1 Phy Fence Training . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 95 2.9.5.4.4 Phy Compensation Initialization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 96 2.9.5.5 SPD ROM-Based Configuration . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 101 2.9.5.6 Non-SPD ROM-Based Configuration . . . . . . . . . . . . . . . . . . . . . . . . . . . 102 2.9.5.6.1 TrdrdSdSc, TrdrdSdDd, and TrdrdDd (Read to Read Timing) . . . . . . 103 2.9.5.6.2 TwrwrSdSc, TwrwrSdDc, TwrwrDd (Write to Write Timing) . . . . . . 103 2.9.5.6.3 Twrrd (Write to Read DIMM Termination Turn-around) . . . . . . . . . . 104 2.9.5.6.4 TrwtTO (Read-to-Write Turnaround for Data, DQS Contention) . . . . 104 2.9.5.6.5 DRAM ODT Control . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 104 2.9.5.6.6 DRAM Address Timing and Output Driver Compensation Control . . 105 2.9.5.7 DCT Training Specific Configuration . . . . . . . . . . . . . . . . . . . . . . . . . . . 114 2.9.5.8 DRAM Device and Controller Initialization . . . . . . . . . . . . . . . . . . . . . . 115 2.9.5.8.1 Software DDR3 Device Initialization . . . . . . . . . . . . . . . . . . . . . . . . . 115 2.9.5.8.1.1 DDR3 MR Initialization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 116 2.9.5.9 DRAM Training . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 118 2.9.5.9.1 Write Levelization Training . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 118 2.9.5.9.1.1 Write Leveling Seed Value . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 120 2.9.5.9.2 DQS Receiver Enable Training . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 120 2.9.5.9.2.1 DQS Receiver Enable Training Seed Value . . . . . . . . . . . . . . . . . . . 121 2.9.5.9.3 DQS Receiver Enable Cycle Training . . . . . . . . . . . . . . . . . . . . . . . . . 122 2.9.5.9.4 DQS Position Training . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 122 2.9.5.9.5 Calculating MaxRdLatency . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 125 6 42300 Rev 3.12 - July 14, 2015 BKDG for AMD Family 15h Models 10h-1Fh Processors 2.9.5.9.5.1 MaxRdLatency Training . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 126 2.9.5.9.6 Continuous Pattern Generation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 126 2.9.5.9.6.1 DRAM Training Pattern Generation . . . . . . . . . . . . . . . . . . . . . . . . . 126 2.9.5.10 DRAM Channel Disable . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 129 2.9.5.11 DRAM Phy Power Savings . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 129 2.9.6 Memory Interleaving Modes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 130 2.9.6.1 Chip Select Interleaving . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 130 2.9.6.2 Channel Interleaving . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 132 2.9.7 Memory Hoisting . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 132 2.9.7.1 DramHoleOffset Programming . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 132 2.9.7.2 DctSelBaseOffset Programming . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 133 2.9.8 DRAM CC6/PC6 Storage . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 134 2.9.9 DRAM On DIMM Thermal Management and Power Capping . . . . . . . . 134 2.10 Thermal Functions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 136 2.10.1 The Tctl Temperature Scale. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 136 2.10.2 Temperature Slew Rate Control. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 137 2.10.3 Sideband Temperature Sensor Interface (SB-TSI) . . . . . . . . . . . . . . . . . . 137 2.10.4 Temperature-Driven Logic . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 137 2.10.4.1 PROCHOT_L and Hardware Thermal Control (HTC) . . . . . . . . . . . . . . 137 2.10.4.2 Local Hardware Thermal Control (LHTC) . . . . . . . . . . . . . . . . . . . . . . . 138 2.10.4.3 Software P-state Limit Control . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 138 2.10.4.4 THERMTRIP . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 138 2.11 Root Complex . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 139 2.11.1 Overview . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 139 2.11.2 Interrupt Routing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 139 2.11.3 Links . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 140 2.11.3.1 Overview . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 140 2.11.3.2 Link Configurations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 140 2.11.3.3 Clocking . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 143 2.11.4 Root Complex Configuration. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 143 2.11.4.1 LPC MMIO Requirements . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 143 2.11.4.2 Link Configuration and Initialization . . . . . . . . . . . . . . . . . . . . . . . . . . . 144 2.11.4.2.1 Clock Configuration . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 144 2.11.4.2.2 Link Configuration and Core Initialization . . . . . . . . . . . . . . . . . . . . . 144 2.11.4.2.3 Link Training . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 145 2.11.4.3 Miscellaneous Features . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 145 2.11.4.3.1 Straps . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 145 2.11.4.3.2 Lane Reversal . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 145 2.11.4.3.3 Link Speed Changes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 146 2.11.4.3.3.1 Software Initiated Link Speed Changes . . . . . . . . . . . . . . . . . . . . . . 146 2.11.4.3.3.2 Autonomous Link Speed Changes . . . . . . . . . . . . . . . . . . . . . . . . . . 146 2.11.4.3.4 Deemphasis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 146 2.11.4.4 Power Management . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 146 2.11.4.4.1 Link States . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 146 2.11.4.4.2 Dynamic Link-width Control . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 147 2.11.4.5 Link Test and Debug Features . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 147 2.11.4.5.1 Compliance Mode . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 147 2.11.5 BIOS Timer . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 147 2.12 IOMMU . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 147 2.12.1 IOMMU Configuration Space . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 147 2.12.2 IOMMU Initialization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 148 7 42300 Rev 3.12 - July 14, 2015 BKDG for AMD Family 15h Models 10h-1Fh Processors 2.13 System Management Unit (SMU). . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 149 2.13.1 Software Interrupts. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 149 2.14 Graphics Processor (GPU) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 150 2.14.1 GPU PCI Interface . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 150 2.14.2 Graphics Memory Controller (GMC) . . . . . . . . . . . . . . . . . . . . . . . . . . . . 150 2.14.3 Frame Buffer (FB) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 150 2.15 RAS Features . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 151 2.15.1 Machine Check Architecture . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 151 2.15.1.1 Machine Check Registers . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 151 2.15.1.2 Machine Check Errors Classes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 152 2.15.1.3 Error Detection, Action, Logging, and Reporting . . . . . . . . . . . . . . . . . . 153 2.15.1.3.1 MCA conditions that cause Shutdown . . . . . . . . . . . . . . . . . . . . . . . . . 154 2.15.1.3.2 Error Logging During Overflow . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 154 2.15.1.4 MCA Initialization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 155 2.15.1.5 Error Code . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 155 2.15.1.6 Handling Machine Check Exceptions . . . . . . . . . . . . . . . . . . . . . . . . . . . 157 2.15.1.6.1 MCA Differentiation Between System-Fatal and Process-Fatal Errors 159 2.15.1.7 Error Thresholding . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 159 2.15.1.8 Error Diagnosis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 160 2.15.1.8.1 Common Diagnosis Information . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 161 2.15.2 Error Injection and Simulation. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 162 3 Registers. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 163 3.1 Register Descriptions and Mnemonics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 163 3.1.1 Northbridge MSRs In Multi-Core Products. . . . . . . . . . . . . . . . . . . . . . . . 165 3.1.2 Software Recommendation (BIOS, SBIOS, CBIOS, etc.) . . . . . . . . . . . . 166 3.1.3 Mapping Tables . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 166 3.1.3.1 Register Mapping . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 166 3.1.3.2 Index Mapping . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 166 3.1.3.3 Field Mapping . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 166 3.1.3.4 Broadcast Mapping . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 166 3.1.3.5 Valid Values . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 167 3.2 IO Space Registers . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 167 3.3 Device 0 Function 0 (Root Complex) Configuration Registers . . . . . . . . . . . . . . . 168 3.4 Device 0 Function 2 (IOMMU) Configuration Registers. . . . . . . . . . . . . . . . . . . . 236 3.5 Device 1 Function 0 Configuration Registers. . . . . . . . . . . . . . . . . . . . . . . . . . . . . 268 3.6 Device 1 Function 1 (Audio Controller) Configuration Registers . . . . . . . . . . . . . 280 3.7 Device [8:2] Function 0 (Root Port) Configuration Registers . . . . . . . . . . . . . . . . 291 3.8 Device 18h Function 0 Configuration Registers. . . . . . . . . . . . . . . . . . . . . . . . . . . 320 3.9 Device 18h Function 1 Configuration Registers. . . . . . . . . . . . . . . . . . . . . . . . . . . 328 3.10 Device 18h Function 2 Configuration Registers. . . . . . . . . . . . . . . . . . . . . . . . . . . 338 3.11 Device 18h Function 3 Configuration Registers. . . . . . . . . . . . . . . . . . . . . . . . . . . 419 3.12 Device 18h Function 4 Configuration Registers. . . . . . . . . . . . . . . . . . . . . . . . . . . 446 3.13 Device 18h Function 5 Configuration Registers. . . . . . . . . . . . . . . . . . . . . . . . . . . 453 3.14 GPU Memory Mapped Registers . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 464 3.15 IOMMU Memory Mapped Registers . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 467 3.16 APIC Registers . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 482 3.17 CPUID Instruction Registers . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 493 3.18 MSRs - MSR0000_xxxx. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 519 3.19 MSRs - MSRC000_0xxx . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 562 3.20 MSRs - MSRC001_0xxx . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 567 8 42300 Rev 3.12 - July 14, 2015 BKDG for AMD Family 15h Models 10h-1Fh Processors 3.21 MSRs - MSRC001_1xxx . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3.22 Core Performance Counter Events . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3.22.1 PMCx0[1F:00] Events (FP) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3.22.2 PMCx0[3F:20] Events (LS). . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3.22.3 PMCx0[5F:40] Events (DC) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3.22.4 PMCx[1,0][7F:60] Events (CU) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3.22.5 PMCx[1,0][9F:80] Events (IC) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3.22.6 PMCx0[BF:A0] Events . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3.22.7 PMCx[1,0][DF:C0] Events (EX, DE) . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3.23 NB Performance Counter Events. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3.23.1 0E[7:0] Events (Memory Controller). . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3.23.2 0E[F:8] Events (Crossbar) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3.23.3 0F[F:0] Events (Crossbar) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3.23.4 1E[F:0] Events (Crossbar) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3.23.5 1F[F:0] Events (Memory Controller, Crossbar) . . . . . . . . . . . . . . . . . . . . 4 592 606 606 607 609 612 615 617 617 622 622 624 626 626 630 Register List. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 631 9 42300 Rev 3.12 - July 14, 2015 BKDG for AMD Family 15h Models 10h-1Fh Processors List of Figures Figure 1: Figure 2: Figure 3: Figure 4: Figure 5: Figure 6: Figure 7: Figure 8: Figure 9: Figure 10: Figure 11: A processor ................................................................................................................................. 27 System Diagram.......................................................................................................................... 28 DQS Position Training Example Results.................................................................................. 124 DQS Position Training Insertion Delay Recovery Example Results........................................ 125 Example Cases for Programming DramHoleOffset.................................................................. 133 Example Cases for Programming DctSelBaseOffset................................................................ 134 Tctl scale ................................................................................................................................... 136 Root complex topology ............................................................................................................ 139 Phy clock configuration ............................................................................................................ 143 Phy recovered clock and sample clock ..................................................................................... 213 Address/Command Timing at the Processor Pins..................................................................... 353 10 42300 Rev 3.12 - July 14, 2015 BKDG for AMD Family 15h Models 10h-1Fh Processors List of Tables Table 1: Table 2: Table 3: Table 4: Table 5: Table 6: Table 7: Table 8: Table 9: Table 10: Table 11: Table 12: Table 13: Table 14: A1) Table 15: Table 16: Table 17: Table 18: Table 19: Table 20: Table 21: Table 22: Table 23: Table 24: Table 25: Table 26: Table 27: Table 28: Table 29: Table 30: Table 31: Table 32: Table 33: Table 34: Table 35: Table 36: Arithmetic and Logical Operators............................................................................................... 19 Functions..................................................................................................................................... 19 Definitions................................................................................................................................... 20 Processor revision conventions................................................................................................... 25 SMM Initial State........................................................................................................................ 39 SMM Save State.......................................................................................................................... 40 Power Management Support....................................................................................................... 49 Software P-state Naming ............................................................................................................ 55 Software P-state Control ............................................................................................................. 56 Core PMC mapping to PERF_CTL[5:0] .................................................................................... 77 DCT Definitions.......................................................................................................................... 84 DDR3 unbuffered DIMM maximum frequency support for FM2 (per channel)........................ 85 DDR3 SO-DIMM maximum frequency support for FS1r2 (per channel).................................. 85 DDR3 unbuffered DIMM maximum frequency support for FP2 (per channel) for (TN-A0 || TN85 DDR3 SO-DIMM maximum frequency support for FP2 (per channel) for (TN-A0 || TN-A1). 86 DDR3 unbuffered DIMM maximum frequency support for FP2 (per channel) for RL-A1....... 86 DDR3 SO-DIMM maximum frequency support for FP2 (per channel) for RL-A1................... 86 DDR3 Soldered-down DRAM maximum frequency support for FP2 (per channel) ................. 86 Package pin mapping ................................................................................................................. 87 Pad (from chiplet) pin mapping ................................................................................................. 89 DDR PLL Lock Time................................................................................................................. 94 Phy predriver calibration codes for Data/DQS at 1.5V .............................................................. 96 Phy predriver calibration codes for Data/DQS at 1.35V ............................................................ 97 Phy Predriver Calibration Codes for Data/DQS at 1.25V .......................................................... 97 Phy predriver calibration codes for Cmd/Addr at 1.5V .............................................................. 98 Phy predriver calibration codes for Cmd/Addr at 1.35V ............................................................ 99 Phy Predriver Calibration Codes for Cmd/Addr at 1.25V .......................................................... 99 Phy predriver calibration codes for Clock at 1.5V.................................................................... 100 Phy predriver calibration codes for Clock at 1.35V.................................................................. 100 Phy Predriver Calibration Codes for Clock at 1.25V................................................................ 101 DDR3 DIMM ODT Pattern ...................................................................................................... 105 BIOS recommendations for UDIMM address timings and output driver control for FM2 or FS1r2 Package...................................................................................................................... 105 BIOS recommendations for UDIMM address timings and output driver control for FP2 Package...............................................................................................................................108 BIOS recommendations for SO-DIMM address timings and output driver control for FM2 or FS1r2 Package............................................................................................................. 109 BIOS recommendations for SO-DIMM address timings and output driver control for FP2 Package................................................................................................................................112 BIOS recommendations for DRAM soldered down address timings and output driver control for FP2 Package..........................................................................................................................113 11 42300 Rev 3.12 - July 14, 2015 Table 37: Table 38. Table 39. Table 40. Table 41. Table 42. Table 43. Table 44. Table 45. Table 46. Table 47: Table 48: Table 49: Table 50: Table 51: Table 52: Table 53: Table 54: Table 55: Table 56: Table 57: Table 58: Table 59: Table 60: Table 61: Table 62: Table 63: Table 64: Table 65: Table 66: Table 67: Table 68: Table 69: Table 70: Table 71: Table 72: Table 73: Table 74: Table 75: Table 76: Table 77: Table 78: Table 79: Table 80: BKDG for AMD Family 15h Models 10h-1Fh Processors DCT Training Specific Register Values.....................................................................................114 DDR3 MR0 ................................................................................................................................116 DDR3 MR1 ................................................................................................................................117 DDR3 MR2 ................................................................................................................................117 DDR3 MR3 ................................................................................................................................118 DDR3 Write Leveling Seed Values........................................................................................... 120 DDR3 DQS Receiver Enable Training Seed Values................................................................. 122 Command Generation and Data Comparison ........................................................................... 129 DDR3 Swapped Normalized Address Lines for CS Interleaving............................................. 130 Example storage region configuration ...................................................................................... 134 INTx Mapping........................................................................................................................... 139 Lane Id Mapping....................................................................................................................... 140 Supported Gfx Port Configurations .......................................................................................... 140 Supported Gfx Port Configurations for D0F0xE4_x013[3:1]_804[3:0]................................... 141 Supported DP0/DP1 DDI Link Configurations ........................................................................ 142 Supported DP2 DDI Link Configurations ................................................................................ 142 Supported General Purpose (GPP) Link Configurations .......................................................... 142 BIOS Services ........................................................................................................................... 149 Recommended Frame Buffer Configurations ........................................................................... 150 MCA Register Cross-Reference Table...................................................................................... 152 Overwrite Priorities for All Banks ............................................................................................ 155 Error Code Types ...................................................................................................................... 156 Error Codes: Transaction Type ................................................................................................. 156 Error codes: cache level ............................................................................................................ 156 Error Codes: Memory Transaction Type................................................................................... 156 Error Codes: Participation Processor ........................................................................................ 157 Error Codes: Memory or IO...................................................................................................... 157 Error Scope Hierarchy .............................................................................................................. 159 Registers Commonly Used for Diagnosis................................................................................. 161 Terminology in Register Descriptions ...................................................................................... 164 BIOS recommendations for D0F0x64_x5[B,9,7,5,3,1] ............................................................ 176 Register Mapping for D0F0xBC_x1F2[E0:00:step20]............................................................. 182 Register Mapping for D0F0xBC_x1F2[E8:08:step20]............................................................. 182 Register Mapping for D0F0xBC_x1F2[F0:10:step20] ............................................................. 183 Index addresses for D0F0xE4_x0[2:1]1[3:0]_0010 ................................................................. 201 Index addresses for D0F0xE4_x0[2:1]1[3:0]_0011 ................................................................. 202 Index addresses for D0F0xE4_x0[2:1]1[3:0]_001[3:2]............................................................ 203 Index addresses for D0F0xE4_x0[2:1]1[3:0]_0015 ................................................................. 204 Per phy register addresses to pin mappings .............................................................................. 205 Per nibble register addresses to pin mappings .......................................................................... 205 Index Mapping for D0F0xE4_x0[2:1]2[3:0]_0000 .................................................................. 205 Index Mapping for D0F0xE4_x0[2:1]2[3:0]_000[2:1]............................................................. 206 Index Mapping for D0F0xE4_x0[2:1]2[3:0]_000[A:9]............................................................ 207 Index Mapping for D0F0xE4_x0[2:1]2[3:0]_000[C:B] ........................................................... 208 12 42300 Rev 3.12 - July 14, 2015 Table 81: Table 82: Table 83: Table 84: Table 85: Table 86: Table 87: Table 88: Table 89: Table 90: Table 91: Table 92: Table 93: Table 94: Table 95: Table 96: Table 97: Table 98: Table 99: Table 100: Table 101: Table 102: Table 103: Table 104: Table 105: Table 106: Table 107: Table 108: Table 109: Table 110: Table 111: Table 112: Table 113: Table 114: Table 115: Table 116: Table 117: Table 118: Table 119: Table 120: Table 121: Table 122: Table 123: Table 124: BKDG for AMD Family 15h Models 10h-1Fh Processors Index Mapping for D0F0xE4_x0[2:1]2[3:0]_000D ................................................................. 209 Index Mapping for D0F0xE4_x0[2:1]2[3:0]_2000 .................................................................. 209 Index Mapping for D0F0xE4_x0[2:1]2[3:0]_2002 .................................................................. 210 Index Mapping for D0F0xE4_x0[2:1]2[3:0]_2005 .................................................................. 210 Index Mapping for D0F0xE4_x0[2:1]2[3:0]_2008 ...................................................................211 Phy per receiver lane register addresses ................................................................................... 212 Phy receiver broadcast register addresses................................................................................. 212 Index Mapping for D0F0xE4_x0[2:1]2[1:0]_[5:4][7:6,3:0][8,0]1 ........................................... 213 Broadcast Mapping for D0F0xE4_x0[2:1]2[1:0]_[5:4][7:6,3:0][8,0]1 .................................... 213 Index Mapping for D0F0xE4_x0[2:1]2[1:0]_[5:4][7:6,3:0][8,0]2 ........................................... 214 Broadcast Mapping for D0F0xE4_x0[2:1]2[1:0]_[5:4][7:6,3:0][8,0]2 .................................... 214 Index Mapping for D0F0xE4_x0[2:1]2[1:0]_[5:4][7:6,3:0][8,0]5 ........................................... 216 Broadcast Mapping for D0F0xE4_x0[2:1]2[1:0]_[5:4][7:6,3:0][8,0]5 .................................... 216 Recommended DCV settings .................................................................................................... 216 Index Mapping for D0F0xE4_x0[2:1]2[1:0]_[5:4][7:6,3:0][8,0]6 ........................................... 217 Broadcast Mapping for D0F0xE4_x0[2:1]2[1:0]_[5:4][7:6,3:0][8,0]6 .................................... 217 Index Mapping for D0F0xE4_x0[2:1]2[1:0]_[5:4][7:6,3:0][8,0]A .......................................... 217 Broadcast Mapping for D0F0xE4_x0[2:1]2[1:0]_[5:4][7:6,3:0][8,0]A ................................... 218 Index Mapping for D0F0xE4_x0[2:1]2[1:0]_[5:4][7:6,3:0][8,0]F........................................... 219 Broadcast Mapping for D0F0xE4_x0[2:1]2[1:0]_[5:4][7:6,3:0][8,0]F.................................... 219 Phy per transmitter lane register addresses ............................................................................... 220 Phy transmitter broadcast register addresses ............................................................................ 220 Index Mapping for D0F0xE4_x0[2:1]2[3:0]_[7:6][7:6,3:0][8,0]0 ........................................... 220 Broadcast Mapping for D0F0xE4_x0[2:1]2[3:0]_[7:6][7:6,3:0][8,0]0 .................................... 221 Recommended link configuration............................................................................................. 221 Index Mapping for D0F0xE4_x0[2:1]2[3:0]_[7:6][7:6,3:0][8,0]5 ........................................... 221 Broadcast Mapping for D0F0xE4_x0[2:1]2[3:0]_[7:6][7:6,3:0][8,0]5 .................................... 222 Recommended preemphasis settings ........................................................................................ 222 Index Mapping for D0F0xE4_x0[2:1]2[3:0]_[7:6][7:6,3:0][8,0]6 ........................................... 223 Broadcast Mapping for D0F0xE4_x0[2:1]2[3:0]_[7:6][7:6,3:0][8,0]6 .................................... 223 Index Mapping for D0F0xE4_x0[2:1]2[1:0]_[D:C][7:0][8,0]5................................................ 223 Broadcast Mapping for D0F0xE4_x0[2:1]2[1:0]_[D:C][7:0][8,0]5......................................... 223 Mapping for wrapper registers.................................................................................................. 224 Index address mapping for D0F0xE4_x013[1:0]_0[C:8]00 ..................................................... 225 Index address mapping for D0F0xE4_x013[1:0]_0[C:8]03 ..................................................... 225 Reserved field mappings for D0F0xE4_x013[3:0]_8013......................................................... 226 Reserved field mappings for D0F0xE4_x013[3:0]_8014......................................................... 227 Reserved field mappings for D0F0xE4_x013[3:0]_8016......................................................... 229 Index address mapping for D0F0xE4_x013[3:1]_804[3:0]...................................................... 234 Register Stream mappings for D0F0xE4_x013[3:1]_804[E:8] ................................................ 234 Valid values for D0F2xFC_x20_L1sel[3:0].............................................................................. 264 Reset mapping for D[8:2]F0x00. .............................................................................................. 291 Link controller state encodings ..................................................................................................311 Register Mapping for D18F1x[1B8,1B0,1A8,1A0,B8,B0,A8,A0,98,90,88,80] ...................... 330 13 42300 Rev 3.12 - July 14, 2015 Table 125: Table 126: Table 127: Table 128: Table 129: Table 130: Table 131: Table 132: Table 133: Table 134: Table 135: Table 136: Table 137: Table 138: Table 139: Table 140: Table 141: Table 142: Table 143: Table 144: Table 145: Table 146: Table 147: Table 148: Table 149: Table 150: Table 151: Table 152: Table 153: Table 154: Table 155: Table 156: Table 157: Table 158: Table 159: Table 160: Table 161: Table 162: Table 163: Table 164: Table 165: Table 166: Table 167: Table 168: BKDG for AMD Family 15h Models 10h-1Fh Processors Register Mapping for D18F1x[1BC,1B4,1AC,1A4,BC,B4,AC,A4,9C,94,8C,84] .................. 331 Register Mapping for D18F1x[D18F1x[1CC:1C0,19C:180] ................................................... 332 Register Mapping for D18F1x[D8,D0,C8,C0] ......................................................................... 333 Register Mapping for D18F1x[DC,D4,CC,C4] ........................................................................ 334 Register Mapping for D18F1x[EC:E0]..................................................................................... 335 DIMM, Chip Select, and Register Mapping ............................................................................. 338 DDR3 DRAM Address Mapping.............................................................................................. 342 Memory Clock Frequency Value Definition............................................................................. 345 Memory P-state Specific Indirect CSR Access by non-BIOS Software................................... 349 Memory P-state Specific DCT CSR In Indirect Space ............................................................. 350 Index Mapping for D18F2x9C_x0000_0[3:0]0[2:1]_dct[1:0]_mp[1:0]................................... 351 Byte Lane Mapping for D18F2x9C_x0000_0[3:0]0[2:1]_dct[1:0]_mp[1:0] ........................... 352 Index Mapping for D18F2x9C_x0000_0[3:0]0[6:5]_dct[1:0]_mp[1:0]................................... 354 Byte Lane Mapping for D18F2x9C_x0000_0[3:0]0[6:5]_dct[1:0]_mp[1:0] ........................... 354 Index Mapping for D18F2x9C_x0000_00[2A:10]_dct[1:0]_mp[1:0] ..................................... 358 Byte Lane Mapping for D18F2x9C_x0000_00[2A:10]_dct[1:0]_mp[1:0] .............................. 359 Index Mapping for D18F2x9C_x0000_00[4A:30]_dct[1:0]_mp[1:0] ..................................... 359 Byte Lane Mapping for D18F2x9C_x0000_00[4A:30]_dct[1:0]_mp[1:0] .............................. 360 Register Mapping for D18F2x9C_x0000_00[51:50]_dct[1:0] ................................................. 361 Byte Lane Mapping for D18F2x9C_x0000_00[51:50]_dct[1:0].............................................. 361 Index Mapping for D18F2x9C_x0D0F_0[F,7:0]0[8,0]_dct[1:0]_mp[1:0] .............................. 362 Index Mapping for D18F2x9C_x0D0F_0[F,7:0]02_dct[1:0]................................................... 363 Broadcast Mapping for D18F2x9C_x0D0F_0[F,7:0]02_dct[1:0] ............................................ 363 Valid Values for D18F2x9C_x0D0F_0[F,7:0]02_dct[1:0][TxPreP]........................................ 363 Index Mapping for D18F2x9C_x0D0F_0[F,7:0]04_dct[1:0]_mp[1:0] .................................... 364 Index Mapping for D18F2x9C_x0D0F_0[F,7:0]06_dct[1:0] ................................................... 365 Index Mapping for D18F2x9C_x0D0F_0[F,7:0]0A_dct[1:0] .................................................. 365 Broadcast Mapping for D18F2x9C_x0D0F_0[F,7:0]0[A,6]_dct[1:0] ..................................... 365 Index addresses for D18F2x9C_x0D0F_0[F,7:0]0F_dct[1:0].................................................. 366 Broadcast write index address for D18F2x9C_x0D0F_0[F,7:0]0F_dct[1:0] ........................... 366 Index Addresses for D18F2x9C_x0D0F_0[F,7:0]10_dct[1:0]_mp[1:0] .................................. 366 Index Addresses for D18F2x9C_x0D0F_0[F,7:0]11_dct[1:0]_mp[1:0] .................................. 367 Index Addresses for D18F2x9C_x0D0F_0[F,7:0]13_dct[1:0]_mp[1:0] .................................. 368 Index Addresses for D18F2x9C_x0D0F_0[F,7:0]1C_dct[1:0]_mp[1:0] ................................. 369 Index Addresses for D18F2x9C_x0D0F_0[F,7:0]1F_dct[1:0]_mp[1:0].................................. 369 Index Addresses for D18F2x9C_x0D0F_0[F,7:0]30_dct[1:0]................................................. 370 Broadcast Write Index Address for D18F2x9C_x0D0F_0[F,7:0]30_dct[1:0] ......................... 370 Index addresses for D18F2x9C_x0D0F_0[F,7:0]31_dct[1:0] .................................................. 371 Broadcast write index address for D18F2x9C_x0D0F_0[F,7:0]31_dct[1:0] ........................... 371 Index Addresses for D18F2x9C_x0D0F_2[F,2:0]00_dct[1:0]_mp[1:0] .................................. 372 Index Mapping for D18F2x9C_x0D0F_2[2:0]02_dct[1:0] ...................................................... 372 Index address mapping for D18F2x9C_x0D0F_[C,8,2][2:0]1F............................................... 373 Index Addresses for D18F2x9C_x0D0F_2[F,2:0]30_dct[1:0]................................................. 373 Index Mapping for D18F2x9C_x0D0F_8[1:0]0[8,4,0]_dct[1:0]_mp[1:0]............................... 376 14 42300 Rev 3.12 - July 14, 2015 Table 169: Table 170: Table 171: Table 172: Table 173: Table 174: Table 175: Table 176: Table 177: Table 178: Table 179: Table 180: Table 181: Table 182: Table 183: Table 184: Table 185: Table 186: Table 187: Table 188: Table 189: Table 190: Table 191: Table 192: Table 193: Table 194: Table 195: Table 196: Table 197: Table 199: Table 200: Table 201: Table 202: Table 203: Table 204: Table 205: Table 206: Table 207: Table 208: Table 209: Table 210: Table 211: Table 212: Table 213: BKDG for AMD Family 15h Models 10h-1Fh Processors Index Mapping for D18F2x9C_x0D0F_[C,8][1:0]02_dct[1:0]................................................ 376 Index Mapping for D18F2x9C_x0D0F_[C,8][1:0][12,0E,0A,06]_dct[1:0]............................. 377 Index Mapping for D18F2x9C_x0D0F_8[1:0]20_dct[1:0]_mp[1:0] ....................................... 377 Index Mapping for D18F2x9C_x0D0F_[C,8][F:0]30_dct[1:0]................................................ 379 Index Mapping for D18F2x9C_x0D0F_C0[10,0C,08,04]_dct[1:0]_mp[1:0] .......................... 380 Index Mapping for D18F2x9C_x0D0F_C020_dct[1:0]_mp[1:0] ............................................ 380 Index Mapping for D18F2x9C_x0D0F_C021_dct[1:0]_mp[1:0] ............................................ 381 D18F2x1B[4:0] Recommended Settings .................................................................................. 395 BIOS Recommendations for RdPtrInit ..................................................................................... 401 XBAR Definitions..................................................................................................................... 426 SMAF Action Definition .......................................................................................................... 429 Buffer Count Definitions .......................................................................................................... 443 D18F5x80[Enabled, DualCore] Definition............................................................................... 454 Register Mapping for D18F5x1[6C:60].................................................................................... 458 NB P-state Definitions .............................................................................................................. 458 Block to register mapping for IOMMUx4[1,0][3:0]00............................................................. 477 Block to register mapping for IOMMUx4[1,0][3:0]04............................................................. 478 Block to register mapping for IOMMUx4[1,0][3:0]08............................................................. 478 Block to register mapping for IOMMUx4[1,0][3:0]10............................................................. 478 Block to register mapping for IOMMUx4[1,0][3:0]14............................................................. 479 Block to register mapping for IOMMUx4[1,0][3:0]18............................................................. 479 Block to register mapping for IOMMUx4[1,0][3:0]1C ............................................................ 480 Block to register mapping for IOMMUx4[1,0][3:0]20............................................................. 480 Block to register mapping for IOMMUx4[1,0][3:0]24............................................................. 480 Block to register mapping for IOMMUx4[1,0][3:0]28............................................................. 481 Block to register mapping for IOMMUx4[1,0][3:0]2C ............................................................ 481 Register Mapping for APIC[170:100] ...................................................................................... 484 Register Mapping for APIC[1F0:180] ...................................................................................... 485 Register Mapping for APIC[270:200] ...................................................................................... 485 Register Mapping for APIC3[60:50] ........................................................................................ 488 Div[3,1:0] Value Table .............................................................................................................. 490 Register Mapping for APIC[4F0:480] ...................................................................................... 491 Register Mapping for APIC[530:500] ...................................................................................... 491 CPUID Fn0000_0000_E[B,C,D]X Value ................................................................................. 493 CPUID Fn8000_0000_E[B,C,D]X Value ................................................................................. 500 Value mapping for CPUID Fn8000_000[4:2]_E[D,C,B,A]X Value......................................... 503 ECX mapping to Cache Type for CPUID Fn8000_001D_E[D,C,B,A]X................................. 512 Register Mapping for MSR0000_020[E,C,A,8,6,4,2,0] ........................................................... 523 Memory Type Definition .......................................................................................................... 523 Register Mapping for MSR0000_020[F,D,B,9,7,5,3,1]............................................................ 524 Register Mapping for MSR0000_02[6F:68,59:58,50].............................................................. 525 Fixed-size MTRR size and Range Mapping ............................................................................. 525 LS Error Descriptions ............................................................................................................... 529 LS Error Signatures................................................................................................................... 530 15 42300 Rev 3.12 - July 14, 2015 Table 214: Table 215: Table 216: Table 217: Table 218: Table 219: Table 220: Table 221: Table 222: Table 223: Table 224: Table 225: Table 226: Table 227: Table 228: Table 229: Table 230: Table 231: Table 232: Table 233: Table 234: Table 235: Table 236: Table 237: Table 238: Table 239: Table 240: Table 241: Table 242: Table 243: Table 244: Table 245: Table 246: BKDG for AMD Family 15h Models 10h-1Fh Processors LS Address Register.................................................................................................................. 530 IF Error Descriptions ................................................................................................................ 533 IF Error Signatures.................................................................................................................... 536 IF Address Register................................................................................................................... 537 MB, SBU, and SBC Definitions ............................................................................................... 541 CU Error Descriptions .............................................................................................................. 541 CU Error Signatures.................................................................................................................. 543 CU Address Register................................................................................................................. 545 NB Error Descriptions .............................................................................................................. 549 NB Error Signatures, Part 1 ...................................................................................................... 550 NB Error Signatures, Part 2 ...................................................................................................... 551 NB Address Register Default Encoding ................................................................................... 552 NB Address Register for Protocol Errors ................................................................................. 552 Protocol Error Type................................................................................................................... 552 NB Address Register for NB Array Errors ............................................................................... 553 NB Address Register for Watchdog Timer Errors .................................................................... 553 EX Error Descriptions............................................................................................................... 557 EX Error Signatures .................................................................................................................. 558 EX Address Register................................................................................................................. 558 FP Error Descriptions................................................................................................................ 560 FP Error Signatures................................................................................................................... 561 Register Mapping for MSRC001_00[03:00] ............................................................................ 567 Register Mapping for MSRC001_00[07:04] ............................................................................ 567 Register Mapping for MSRC001_00[35:30] ............................................................................ 572 BIOS recommendation for MSRC001_00[35:30] .................................................................... 573 Register Mapping for MSRC001_00[53:50] ............................................................................ 577 Register Mapping for MSRC001_00[6B:64]............................................................................ 580 P-state Definitions..................................................................................................................... 580 Register Mapping for MSRC001_020[A,8,6,4,2,0].................................................................. 587 Register Mapping for MSRC001_020[B,9,7,5,3,1].................................................................. 589 Register Mapping for MSRC001_024[6,4,2,0] ........................................................................ 589 Register Mapping for MSRC001_024[7,5,3,1] ........................................................................ 591 Register Mapping for PMCx0D[F:C] ....................................................................................... 620 16 42300 Rev 3.12 - July 14, 2015 BKDG for AMD Family 15h Models 10h-1Fh Processors Revision History Revision 3.12 Changes • 2.10.4.2 [Local Hardware Thermal Control (LHTC)]: Clarified. Revision 3.10 Changes • • • • • • • • • • • • • • • • • • • • 2.5.3.1.1 [Application Power Management (APM)]: Clarified. Table 50: [Supported Gfx Port Configurations for D0F0xE4_x013[3:1]_804[3:0]]: Updated. 2.9 [DRAM Controllers (DCTs)]: Updated. 2.10.1 [The Tctl Temperature Scale]: Updated. 2.14.3 [Frame Buffer (FB)]: Updated. Table 131: Updated. D18F2x404_dct[1:0][UrMctTokenLimit]: Updated. D18F3x64 [Hardware Thermal Control (HTC)], D18F3x68 [Software P-state Limit]: Updated. MSRC001_1021[DisWayFilter]: Added . MSRC001_1030[IbsFetchVal]: Updated. D18F2xA8_dct[1:0][BankSwap]: Corrected. 1.5.1 [Revision Conventions]: Updated. 1.5.3 [Changes For Revision RL-A1]Added. 2.5.4.1.1 [NB P-state Transitions]: Updated. 2.9.5.6.6 [DRAM Address Timing and Output Driver Compensation Control]: Updated. 2.12 [IOMMU]: Updated. D18F2x110[DctSelIntLvAddr[1:0]]: Updated. D18F2x114[DctSelIntLvAddr[2]]: Updated. MSRC001_0071: Updated. MSRC001_024[6,4,2,0][IntCoreSel, IntCoreEn]: Added. Revision 3.00 Changes • Initial public release. 17 42300 Rev 3.12 - July 14, 2015 1 BKDG for AMD Family 15h Models 10h-1Fh Processors Overview The AMD Family 15h Models 10h-1Fh processor (in this document referred to as the processor) is a processing unit that supports x86-based instruction sets. The processor includes (a) up to two independent processing units referred to as compute units (each compute unit containing two cores), (b) one PCIe® root complex (in this document referred to as the root complex or RC) with generation 2 link support, and (c) up to 2 system memory DRAM interfaces. AMD Family 15h Models 10h-1Fh processors are distinguished by the combined ExtFamily and BaseFamily fields of the CPUID instruction (see CPUID Fn8000_0001_EAX in 3.17 [CPUID Instruction Registers]). 1.1 Intended Audience This document provides the processor behavioral definition and associated design notes. It is intended for platform designers and for programmers involved in the development of low-level BIOS (basic input/output system) functions, drivers, and operating system kernel modules. It assumes prior experience in personal computer platform design, microprocessor programming, and legacy x86 and AMD64 microprocessor architecture. The reader should also have familiarity with various platform technologies, such as DDR DRAM. 1.2 • • • • • • • • • • • • • • • • Reference Documents Advanced Configuration and Power Interface (ACPI) Specification. www.acpi.info. AMD64 Architecture Programmer's Manual Volume 1: Application Programming, #24592. AMD64 Architecture Programmer's Manual Volume 2: System Programming, #24593. AMD64 Architecture Programmer's Manual Volume 3: Instruction-Set Reference, #24594. AMD64 Architecture Programmer's Manual Volume 4: 128-Bit and 256-Bit Media Instructions, #26568. AMD64 Architecture Programmer's Manual Volume 5: 64-Bit Media and x87 Floating-Point Instructions, #26569. AMD I/O Virtualization Technology (IOMMU) Specification, #34434. Software Optimization Guide for AMD Family 15h Processors, #47414. CPUID Specification, #25481. Electrical Data Sheet for AMD Family 15h Models 10h-1Fh Processors, #47080. Revision Guide for AMD Family 15h Models 10h-1Fh Processors, #48931. JEDEC standards. www.jedec.org. PCI local bus specification. (www.pcisig.org). PCI Express specification. (www.pcisig.org). System Management Bus (SMBus) specification. (www.smbus.org). AMD64 Technology Lightweight Profiling Specification, #43724. 1.3 1.3.1 Conventions Numbering • Binary numbers. Binary numbers are indicated by appending a “b” at the end, e.g., 0110b. • Decimal numbers. Unless specified otherwise, all numbers are decimal. This rule does not apply to the register mnemonics described in 3.1 [Register Descriptions and Mnemonics]; register mnemonics all utilize hexadecimal numbering. • Hexadecimal numbers. hexadecimal numbers are indicated by appending an “h” to the end, e.g., 45f8h. • Underscores in numbers. Underscores are used to break up numbers to make them more readable. They do not imply any operation. E.g., 0110_1100b. 18 42300 Rev 3.12 - July 14, 2015 1.3.2 BKDG for AMD Family 15h Models 10h-1Fh Processors Arithmetic And Logical Operators In this document, formulas follow some Verilog conventions for logic equations. Table 1: Arithmetic and Logical Operators Operator {} | || & && ^ ~ ! == != <= >= * / << >> ?: Definition Curly brackets are used to indicate a group of bits that are concatenated together. Each set of bits is separated by a comma. E.g., {Addr[3:2], Xlate[3:0]} represents a 6-bit value; the two MSBs are Addr[3:2] and the four LSB’s are Xlate[3:0]. Bitwise OR operator. E.g. (01b | 10b == 11b). Logical OR operator. E.g. (01b || 10b == 01b). Bitwise AND operator. E.g. (01b & 10b == 00b). Logical AND operator. E.g. (01b && 10b == 01b). Bitwise exclusive-OR operator; sometimes used as “raised to the power of” as well, as indicated by the context in which it is used. E.g. (01b ^ 10b == 11b). E.g. (2^2 == 4). Bitwise NOT operator (also known as one’s complement). E.g. (~10b == 01b). Logical NOT operator. E.g. (!10b == 00b). Logical “is equal to” operator. Logical “is not equal to” operator. Less than or equal operator. Greater than or equal operator. Arithmetic multiplication operator. Arithmetic division operator. Shift left first operand by the number of bits specified by the 2nd operand. E.g. (01b << 01b == 10b). Shift right first operand by the number of bits specified by the 2nd operand. E.g. (10b >> 01b == 01b). Conditional operator. E.g. condition ? value if true : value if false. Equivalent to IF condition THEN value if true ELSE value if false. Table 2: Functions Function ABS FLOOR CEIL MIN MAX Definition ABS(integer-expression): Remove sign from signed value. FLOOR(integer-expression): Rounds real number down to nearest integer. CEIL(real-expression): Rounds real number up to nearest integer. MIN(integer-expression-list): Picks minimum integer or real value of comma separated list. MAX(integer-expression-list): Picks maximum integer or real value of comma separated list. 19 42300 Rev 3.12 - July 14, 2015 BKDG for AMD Family 15h Models 10h-1Fh Processors Table 2: Functions Function COUNT ROUND Definition COUNT(integer-expression): Returns the number of binary 1’s in the integer. ROUND(real-expression): Rounds to the nearest integer; halfway rounds away from zero. UNIT(fieldName UnitOfMeasure): Input operand is a register field name that defines all values with the same unit of measure. Returns the value expressed in the unit of measure for the current value of the register field. E.g. If (D18F2x210_dct[1:0]_nbp[3:0][RdPtrInit]==0101b) then (UNIT(D18F2x210_dct[1:0]_nbp[3:0][RdPtrInit] MEMCLK)==1.5). UNIT The order in which logical operators are applied is: ~ first, & second, and | last. For example, the equation: Output[3:0] = {A[1:0], B[3:2]} & C[3:0] | ~D[3:0] & E[9:6], is translated as: Output[3] = (A[1] & C[3]) | (~D[3] & E[9]); Output[2] = (A[0] & C[2]) | (~D[2] & E[8]); Output[1] = (B[3] & C[1]) | (~D[1] & E[7]); Output[0] = (B[2] & C[0]) | (~D[0] & E[6]); 1.4 Definitions Table 3: Definitions Term AP BatteryPower BCS BERT BIST Boot VID BSC BSP CAR C-states Canonical address Channel Definition Application processor. See 2.3 [Processor Initialization]. The system is running from a battery power source or otherwise undocked from a continuous power supply. Setting using this definition may be required to change during runtime. Base configuration space. See 2.7 [Configuration Space]. Bit error rate tester. A piece of test equipment that generates arbitrary test patterns and checks that a device under test returns them without errors. Built-in self-test. Hardware within the processor that generates test patterns and verifies that they are stored correctly (in the case of memories) or received without error (in the case of links). Boot voltage ID. This is the VDD and VDDNB voltage level that the processor requests from the external voltage regulator during the initial phase of the cold boot sequence. See 2.5.1.2 [Internal VID Registers and Encodings]. Boot strap core. Core 0 of the BSP. Specified by MSR0000_001B[BSC]. Boot strap processor. See 2.3 [Processor Initialization]. Use of the L2 cache as RAM during boot. See 2.3.3 [Using L2 Cache as General Storage During Boot]. These are ACPI-defined core power states. C0 is operational. All other C-states are low-power states in which the processor is not executing code. See 2.5.3.2 [Core C-states]. An address in which the state of the most-significant implemented bit is duplicated in all the remaining higher-order bits, up to bit 63. See DRAM channel. 20 42300 Rev 3.12 - July 14, 2015 BKDG for AMD Family 15h Models 10h-1Fh Processors Table 3: Definitions Term Channel interleaved mode CMP COF Cold reset Compute Unit Core CPB CpuCoreNum CPUID function X CS DCT DCQ DDR3 DID Definition Mode in which DRAM address space is interleaved between DRAM channels. See 2.9.6 [Memory Interleaving Modes]. Chip multi-processing. Refers to processors that include multiple cores. See 2.1 [Processor Overview]. Current operating frequency of a given clock domain. See 2.5.3 [CPU Power Management]. PWROK is deasserted and RESET_L is asserted. See 2.3 [Processor Initialization]. Two Cores that share IC, DE, FP and L2 resources. See 2.1 [Processor Overview]. The instruction execution unit of the processor. See 2.1 [Processor Overview]. Core performance boost. See 2.5.3.1.1 [Application Power Management (APM)]. Specifies the core number. See 2.4.3 [Processor Cores and Downcoring]. Refers to the CPUID instruction when EAX is preloaded with X. See 3.17 [CPUID Instruction Registers]. Chip select. See D18F2x[5C:40]_dct[1:0] [DRAM CS Base Address]. DRAM controller. See 2.9 [DRAM Controllers (DCTs)]. DRAM controller queue. DDR3 memory technology. See 2.9 [DRAM Controllers (DCTs)]. Divisor identifier. Specifies the post-PLL divisor used to reduce the COF. See 2.5.3 [CPU Power Management]. Doubleword A 32-bit value. Downcoring Removal of cores. See 2.4.3 [Processor Cores and Downcoring]. DRAM The part of the DRAM interface that connects to a 64-bit DIMM. For example, a processor with channel a 128-bit DRAM interface is said to support two DRAM channels. See 2.9 [DRAM Controllers (DCTs)]. Dual-Plane Refers to a processor or systemboard where VDD and VDDNB are separate and may operate at independent voltage levels. Refer to 2.5.1 [Processor Power Planes And Voltage Control]. DW Doubleword. A 32-bit value. ECS Extended configuration space. See 2.7 [Configuration Space]. EDS Electrical data sheet. See 1.2 [Reference Documents]. FCH Fusion Controller Hub. The platform device that contains the bridge to the system BIOS. FDS Functional data sheet; there is one FDS for each package type. FID Frequency identifier. Specifies the PLL frequency multiplier for a given clock domain. See 2.5.3 [CPU Power Management]. FreeRAn internal free running timer used by many power management features. The timer increunSample- ments at the rate specified by D18F4x110[CSampleTimer]. Timer GB Gbyte or Gigabyte; 1,073,741,824 bytes. #GP A general-protection exception. Always produces an error code of 0. #GP(0) GpuEnabled GpuEnabled = ( D1F0x00 != FFFF_FFFFh ). GT/s Giga-transfers per second. 21 42300 Rev 3.12 - July 14, 2015 BKDG for AMD Family 15h Models 10h-1Fh Processors Table 3: Definitions Term HTC Definition Hardware thermal control. See 2.10.4.1 [PROCHOT_L and Hardware Thermal Control (HTC)]. HTC-active Hardware-controlled lower-power, lower-performance state used to reduce temperature. See state 2.10.4.1 [PROCHOT_L and Hardware Thermal Control (HTC)]. IBS Instruction based sampling. See 2.6 [Performance Monitoring]. IFCM Isochronous flow-control mode, as defined in the link specification. ILM Internal loopback mode. Mode in which the link receive lanes are connected directly to the transmit lanes of the same link for testing and characterization. See D18F0x170 [Link Extended Control]. IO configu- Access to configuration space through IO ports CF8h and CFCh. See 2.7 [Configuration ration Space]. IORR IO range register. See MSRC001_00[18,16] [IO Range Base (IORR_BASE[1:0])]. IOMMU I/O Memory Management Unit. Also known as AMD I/O Virtualization Technology. See 2.12 [IOMMU]. KB Kbyte or Kilobyte; 1024 bytes. L1 cache The level 1 caches (instruction cache and the data cache) and the level 2 caches. See 2.1 [Processor Overview]. L2 cache Linear (vir- The address generated by a core after the segment is applied. tual) address Link Generic term that refers to a refer to PCIe® link. LINT Logical address LVT Master abort MB MCT MCQ Micro-op MEMCLK MMIO Local interrupt. The address generated by a core before the segment is applied. Local vector table. A collection of APIC registers that define interrupts for local events. E.g., APIC[530:500] [Extended Interrupt [3:0] Local Vector Table]. This is a PCI-defined term that is applied to transactions on other than PCI buses. It indicates that the transaction is terminated without affecting the intended target; reads return all 1’s; writes are discarded; the master abort error code is returned in the response, if applicable; master abort error bits are set if applicable. Megabyte; 1024 KB. Memory controller. See 2.8 [Northbridge (NB)]. Memory controller queue. See 2.8 [Northbridge (NB)]. Micro-op. Instructions have variable-length encoding and many perform multiple primitive operations. The processor does not execute these complex instructions directly, but, instead, decodes them internally into simpler fixed-length instructions called macro-ops. Processor schedulers subsequently break down macro-ops into sequences of even simpler instructions called micro-ops, each of which specifies a single primitive operation. See Software Optimization Guide for AMD Family 15h Processors. Refers to the clock signals, M[B, A][3:0]_CLK, that are driven from the processor to DDR DIMMs. Memory-mapped input-output range. This is physical address space that is mapped to the IO functions such as the IO links or MMIO configuration. The IO link MMIO ranges are specified by D18F1x[1CC:180,BC:80] [MMIO Base/Limit]. 22 42300 Rev 3.12 - July 14, 2015 BKDG for AMD Family 15h Models 10h-1Fh Processors Table 3: Definitions Term Definition MMIO con- Access to configuration space through memory space. See 2.7 [Configuration Space]. figuration MSR Model-specific register. The core includes several MSRs for general configuration and control. See 3.18 [MSRs - MSR0000_xxxx] for the beginning of the MSR register definitions. MTRR Memory-type range register. The MTRRs specify the type of memory associated with various memory ranges. See MSR0000_00FE, MSR0000_020[F:0], MSR0000_02[6F:68,59:58,50], and MSR0000_02FF. NB Northbridge. The transaction routing block of the node. See 2.1 [Processor Overview]. NBC Node Base Core. The lowest numbered core in the node. NCLK The main northbridge clock. The NCLK frequency is the NB COF. Node ID The identifier assigned to each node, D18F0x60[NodeId]. Node See 2.1 [Processor Overview]. Normalized Addresses used by DCTs. See 2.8 [Northbridge (NB)]. address OW Octword. An 128-bit value. ODM On-DIMM mirroring. See D18F2x[5C:40]_dct[1:0][OnDimmMirror]. ODT On-die termination, which is applied DRAM interface signals. ODTS DRAM On-die thermal sensor. Operational The frequency at which the processor operates. See 2.5 [Power Management]. frequency PCI Express. PCIe® PDS Physical address PRBS Processor PSI P-state PTE QW RAS RDQ RX Shutdown Product data sheet. Addresses used by cores in transactions sent to the NB. Pseudo-random bit sequence. See 2.1 [Processor Overview]. Power Status Indicator. See 2.5.1.3.1 [PSIx_L Bit]. Performance state. See 2.5 [Power Management]. Page table entry. Quadword. A 64-bit value. Reliability, availability and serviceability (industry term). See 2.15 [RAS Features]. Read data queue. Receiver. A state in which the affected core waits for either INIT, RESET, or NMI. When shutdown state is entered, a shutdown special cycle is sent on the IO links. Single-Plane Refers to a processor or systemboard where VDD and VDDNB are tied together and operate at the same voltage level. Refer to 2.5.1 [Processor Power Planes And Voltage Control]. Slam Refers to change the voltage to a new value in one step (as opposed to stepping). See 2.5.1.4.1 [Hardware-Initiated Voltage Transitions]. 23 42300 Rev 3.12 - July 14, 2015 BKDG for AMD Family 15h Models 10h-1Fh Processors Table 3: Definitions Term SMAF Definition System management action field. This is the code passed from the SMC to the processors in STPCLK assertion messages. The action taken by the processors in response to this message is specified by D18F3x[84:80] [ACPI Power State Control]. SMBus System management bus. Refers to the protocol on which the serial VID interface (SVI) commands and SBI are based. See 2.5.1 [Processor Power Planes And Voltage Control] and 1.2 [Reference Documents]. SMC System management controller. This is the platform device that communicates system management state information to the processor through an IO link, typically the system IO hub. SMI System management interrupt. See 2.4.8.2.1 [SMM Overview]. SMM System management mode. See 2.4.8.2 [System Management Mode (SMM)]. Speculative A performance monitor event counter that counts all occurrences of the event even if the event event occurs during speculative code execution. SVI2 Serial VID 2.0 interface. See 2.5.1.1 [Serial VID Interface] SVM Secure virtual machine. See 2.4.9 [Secure Virtual Machine Mode (SVM)]. Sync flood The propagation of continuous sync packets to all links. This is used to quickly stop the transmission of potentially bad data when there are no other means to do so. See the link specification for additional information. TCC Temperature calculation circuit. See 2.10 [Thermal Functions]. Tctl Processor temperature control value. See 2.10.4 [Temperature-Driven Logic]. TDP Thermal design power. A power consumption parameter that is used in conjunction with thermal specifications to design appropriate cooling solutions for the processor. Token A scheduler entry used in various northbridge queues to track outstanding requests. See D18F3x140 [SRI to XCS Token Count] on Page 461. TX Transmitter. UI Unit interval. This is the amount of time equal to one half of a clock cycle. UMA Unified memory architecture. This is a type of display device that uses a frame buffer located in main memory. UMI Unified Media Interface. The link between the processor and the FCH. VDD Main power supply to the processor core logic. VDDNB Main power supply to the processor NB logic. VID Voltage level identifier. See 2.5.1 [Processor Power Planes And Voltage Control]. Virtual CAS The clock in which CAS is asserted for the burst, N, plus the burst length (in MEMCLKs), minus 1; so the last clock of virtual CAS = N + (BL/2) - 1. VRM Voltage regulator module. W Word. A 16-bit value. Warm reset RESET_L is asserted only (while PWROK stays high). See 2.3 [Processor Initialization]. WDT Watchdog timer. A timer that detects activity and triggers an error if a specified period of time expires without the activity. For example, see the NB watchdog timer in D18F3x40 [MCA NB Control]. WDQ Write data queue. XBAR Cross bar; command packet switch. See 2.8 [Northbridge (NB)]. 24 42300 Rev 3.12 - July 14, 2015 1.5 BKDG for AMD Family 15h Models 10h-1Fh Processors Changes Between Revisions and Product Variations 1.5.1 Revision Conventions The processor revision is specified by CPUID Fn0000_0001_EAX [Family, Model, Stepping Identifiers] or CPUID Fn8000_0001_EAX [Family, Model, Stepping Identifiers]. This document uses a revision letter instead of specific model numbers. The following table contains the definitions based on model and stepping used in this document. Where applicable, the processor stepping is indicated after the revision letter. All behavior marked with a revision letter apply to future revisions unless they are superseded by a change in a later revision. See the Revision Guide for AMD Family 15h Models 10h-1Fh Processors for additional information about revision determination. Table 4: Processor revision conventions Term Revision TN-A0 TN-A1 RL-A1 1.5.2 Definition Revision = {CPUID Fn0000_0001_EAX[BaseModel], CPUID Fn0000_0001_EAX[Stepping]}. TN-A0 = {0, 0}. TN-A1 = {0, 1}. RL-A1 = {3, 1}. Major Changes Relative to Family 15h Models 00h-0Fh Processors • CPU core changes: • Instruction Set Architecture changes: • Added FMA instruction support. See CPUID Fn0000_0001_ECX[FMA]. • Added F16C support. See CPUID Fn0000_0001_ECX[F16C]. • Added BMI1 and TBM instruction support. See CPUID Fn0000_0007_EBX_x0[BMI1] and CPUID Fn8000_0001_ECX[TBM]. • Added TCE support. See CPUID Fn8000_0001_ECX[TCE]. • No L3 cache. • Other core changes: • Increased L1 DTLB size to 64. See CPUID Fn8000_0005_EAX, CPUID Fn8000_0005_EAX, and CPUID Fn8000_0019_EAX. • Added read-only effective frequency registers. See CPUID Fn8000_0007_EDX[EffFreqRO]. • Memory controller (MCT) and DRAM controllers (DCTs) additions: • Low-voltage DDR3 support; Added 1.25V in addition to 1.35V. • Support for 1066 MHz (2133 MT/s) MEMCLK frequency. • Support for NCLK:MEMCLK frequency ratio down to 1.25:1 instead of 2:1. • Links and IO additions: • PCIe® with generation 2 link support. • Per-lane power gating. • RAS-related additions: • General Northbridge additions: • NB power gating. • Integrated root complex. • Integrated graphics processor. • Integrated IOMMUv2. • Power management: • Package C6 support. 25 42300 Rev 3.12 - July 14, 2015 BKDG for AMD Family 15h Models 10h-1Fh Processors • Bidirectional APM. • SVI 2.0 infrastructure. • DRAM Power Management: • Memory P-states. 1.5.2.1 Major Changes to Core/NB Performance Counters • Core performance counters: • Added PMCx032 [Misaligned Stores]. • Added PMCx034 [FP Load Buffer Stall]. • Updated PMCx052 [Ineffective Software Prefetches]. • Updated PMCx0D[F:C]. • Added PMCx1C0 [Retired x87 Floating Point Operations]. 1.5.3 Changes For Revision RL-A1 • Changes that may result in BIOS modifications. • Added 2.5.9.1 [Hybrid Boost] • Added D0F0xBC_x1F428[HybridBoostEn] • Other changes • Added D0F0xBC_x1F8EC 26 42300 Rev 3.12 - July 14, 2015 2 Functional Description 2.1 Processor Overview BKDG for AMD Family 15h Models 10h-1Fh Processors The processor is a package that contains a node consisting of (1) one to four cores (one or two compute units), (2) one PCIe® root complex with generation 2 link support, (3) two 64-bit DDR3 interfaces for communication to system memory, and (4) one communication packet routing block referred to as the northbridge (NB). 1 to 2 CPU compute units - Each compute unit shares: FP, IC, L2 Core 0: Core 1: - x86 execution unit - x86 execution unit - MSR registers - MSR registers - APIC and registers - APIC and registers - DC - DC Northbridge (NB) - Transaction routing - Configuration- and IO-space registers - Root Complex - Graphics Controller DDR A DDR B Figure 1: A processor Each compute unit includes 2 cores each having an x86 instruction execution logic and first-level (L1) data cache. The FP unit, second level (L2) general-purpose cache, and first-level instruction cache are shared between the cores of a compute unit. There is a set of MSRs and APIC registers associated with each core. Processors that include multiple cores are said to incorporate chip multi-processing or CMP. The links are input-output links, as defined by the PCI Express Base Specification. Each DRAM interface supports a 64-bit DDR3 memory channel. The NB routes transactions between the cores, the link, and the DRAM interfaces. It includes the configuration register space for the device. 2.2 System Overview The following diagram illustrates the expected system architecture. 27 42300 Rev 3.12 - July 14, 2015 BKDG for AMD Family 15h Models 10h-1Fh Processors DDI Processor Syste m Memory UMI IO Subsystem FCH System BIO S Figure 2: System Diagram 2.3 Processor Initialization This section describes the initialization sequence after a cold reset. Core 0 of the processor, the bootstrap core (BSC), begins executing code from the reset vector. The remaining cores do not fetch code until their enable bits are set (D18F0x1DC[CpuEn]). 2.3.1 BSC initialization The BSC must perform the following tasks as part of boot. • Store BIST information from the EAX register into an unused processor register. • D18F0x6C[InitDet] may be used by BIOS to differentiate between INIT and cold/warm reset. • Determine type of startup using D18F0x6C[ColdRstDet]. • If this is a warm reset then BIOS may check for valid MCA errors and if present save the status for later use. See 2.15.1.6 [Handling Machine Check Exceptions]. • Enable the cache, program the MTRRs for CAR and initialize CAR. See 2.3.3 [Using L2 Cache as General Storage During Boot]. • Setup the SMU. • Setup of APIC (2.4.8.1.3 [ApicId Enumeration Requirements]). • Setup the link configuration 2.11.3.2 [Link Configurations]. • Setup the root complex and initialize the I/O links 2.11.4.2 [Link Configuration and Initialization]. • If required, reallocate data and flow control buffers of the links (see D18F0x90 [Upstream Base Channel Buffer Count] and D18F0x94 [Link Isochronous Channel Buffer Count]) and issue system warm reset. • Configure the DRAM controllers. • Configure processor power management. See 2.5 [Power Management]. • If supported, allow other cores to begin fetching instructions by setting D18F0x1DC[CpuEn] in the PCI configuration space of the nodes. See 2.4.3 [Processor Cores and Downcoring]. 2.3.2 AP initialization All other processor cores other than core 0 begin executing code from the reset vector. They must perform the following tasks as part of boot. • Store BIST information from the eax register into an unused processor register. 28 42300 Rev 3.12 - July 14, 2015 BKDG for AMD Family 15h Models 10h-1Fh Processors • D18F0x6C[InitDet] may be used by BIOS to differentiate between INIT and cold/warm reset. • Determine the history of this reset using the D18F0x6C [Link Initialization Control] [ColdRstDet] bit: • If this is a warm reset then BIOS may check for valid MCA errors and if present save the status for use later. See 2.15.1.6 [Handling Machine Check Exceptions]. • Set up the local APIC. See 2.4.8.1.3 [ApicId Enumeration Requirements]. • Configure processor power management. See 2.5 [Power Management]. 2.3.3 Using L2 Cache as General Storage During Boot Prior to initializing the DRAM controller for system memory, BIOS may use the L2 cache of each core as general storage. The L2 cache as general storage is described as follows: • Each compute unit has its own L2 cache. • BIOS manages the mapping of the L2 storage such that cacheable accesses do not cause L2 victims. • The L2 size, L2 associativity, and L2 line size is determined by reading CPUID Fn8000_0006_ECX[L2Size, L2Assoc, L2LineSize]. L2WayNum is defined to be the number of ways indicated by the L2Assoc code. • The L2 cache is viewed as (L2Size/L2LineSize) cache lines of storage, organized as L2WayNum ways, each way being (L2Size/L2WayNum) in size. • E.g. L2Assoc=8 so L2WayNum=16 (there are 16 ways). If L2Size=512KB then there are 16 blocks of cache, each 512KB/16 in size, or 32KB each. • For each of the following values of L2Size, the following values are defined: • L2Size=1 MB: L2Tag=PhysAddr[39:16], L2WayIndex=PhysAddr[15:6]. • L2Size=2 MB: L2Tag=PhysAddr[39:17], L2WayIndex=PhysAddr[16:6]. • PhysAddr[5:0] addresses the L2LineSize number of bytes of storage associated with the cache line. • The L2 cache, when allocating a line at L2WayIndex: • Picks an invalid way before picking a valid way. • Prioritizes the picking of invalid ways such that way L2WayNum-1 is the highest priority and 0 is the lowest priority. • It is recommended that BIOS assume a simpler allocation of L2 cache memory, being L2WayNum sizealigned blocks of memory, each being L2Size/L2WayNum bytes. • BIOS can rely on a minimum L2Size of 512 KB and can rely on being able to use a minimum of 14 ways for general storage. See CPUID Fn8000_0006_ECX[L2Size]. See initialization requirements below for MSRC001_1023[L2WayLock, L2FirstLockedWay]. The following memory types are supported: • WP-IO: BIOS ROM may be assigned the write-protect IO memory type and may be accessed read-only as data and fetched as instructions. • WP-IO accesses, both read and write, do not get evicted to the L2 and therefore do not need to be considered for allocation into the L2. • WB-DRAM: General storage may be assigned the write-back DRAM memory type and may be accessed as read-write data, but not accessed by instruction fetch. • BIOS initializes an L2LineSize sized and aligned location in the L2 cache, mapped as write-back DRAM, with 1 read to at least 1 byte of the L2LineSize sized and aligned WB-DRAM address. BIOS may store to a line only after it has been allocated by a load. • Fills, sent to the disabled memory controller, return undefined data. • All of memory space that is not accessed as WP-IO or WB-DRAM space must be marked as UC memory type. • In order to prevent victimizing L2 data, no more than L2WayNum cache lines accessed as WB-DRAM may have the same L2WayIndex. 29 42300 Rev 3.12 - July 14, 2015 BKDG for AMD Family 15h Models 10h-1Fh Processors • Software does not need to know which ways the L2WayNum lines are allocated to for any given value of L2WayIndex, only that invalid ways will be selected for allocation before valid ways will be selected for allocation. • Software is not allowed to deallocate a line in the L2 by using CLFLUSH. See 2.9.5.9.6.1 [DRAM Training Pattern Generation]. Performance monitor event PMCx07F[1], titled “L2 Writebacks to system“, can be used to indicate whether L2 dirty data was lost by being victimized and sent to the disabled memory controller. The following requirements must be satisfied prior to using the cache as general storage: • Paging must be disabled. • MSRC001_0015[INVDWBINVD]=0. • MSRC001_1020[DisSS]=1. • MSRC001_1021[DisSpecTlbRld]=1. Disable speculative ITLB reloads. • MSRC001_1022[DisSpecTlbRld]=1. Disable speculative DTLB reloads. • MSRC001_1022[DisHwPf]=1. • MSRC001_102B[CombineCr0Cd]=0. • CLFLUSH, INVD, and WBINVD must not be used during CAR but may be used when tearing down CAR for all compute units on a node. • The BIOS must not use SSE, or MMX™ instructions, with the exception of the following list: MOVD, MOVQ, MOVDQA, MOVQ2DQ, MOVDQ2Q. • The BIOS must not enable exceptions, page-faults, and other interrupts. • BIOS must not use software prefetches. • UC-DRAM: All of DRAM that is not accessed as WB-DRAM space must be marked as UC memory type. • If (MSRC001_1023[L2WayLock]==1) then: • Only the ways 0 through (MSRC001_1023[L2FirstLockedWay]-1) may be used for general storage. • BIOS can rely on MSRC001_1023[L2FirstLockedWay] to have a minimum value of Eh. • If (MSRC001_1023[L2WayLock]==0) then: • Set MSRC001_1023[L2WayLock]=1. • Set MSRC001_1023[L2FirstLockedWay]=Fh. When BIOS has completed using the cache for general storage the following steps are followed: 1.An INVD instruction should be executed on each core that used cache as general storage; an INVD should only be issued when all cores on all nodes have completed using the cache for general storage. 2.If DRAM is initialized and there is data in the cache that needs to get moved to main memory, CLFLUSH or WBINVD may be used instead of INVD, but software must ensure that needed data in main memory is not overwritten. 3.Program the following configuration state: • MSRC001_0015[INVDWBINVD]=1. • MSRC001_1020[DisSS]=0. • MSRC001_1021[DisSpecTlbRld]=0. • MSRC001_1022[DisSpecTlbRld]=0. • MSRC001_1022[DisHwPf]=0. • If ((MSRC001_1023[L2WayLock]==1) & (MSRC001_1023[L2FirstLockedWay]==Fh)), program MSRC001_1023[L2WayLock]=0. 30 42300 Rev 3.12 - July 14, 2015 2.4 BKDG for AMD Family 15h Models 10h-1Fh Processors Core The majority of the behavioral definition of the core is specified in the AMD64 Architecture Programmer’s Manual. See 1.2 [Reference Documents]. 2.4.1 Compute Unit Unless otherwise specified the processor configuration interface hides the compute unit implementation and presents software with homogenous cores, each independent of the other. Software may use D18F5x80[Enabled, DualCore] in order to associate a core with a compute unit. This information can be useful because some configuration settings are determined based on active compute units and core performance may vary based on resource sharing within a compute unit. 2.4.1.1 Registers Shared by Cores in a Compute Unit Some MSRs are implemented one copy per compute unit instead of per core; these MSRs are designated as SharedC (shared coherent) or SharedNC (shared non-coherent). The absence of SharedC/SharedNC implies not-shared, which is the normal per-core instance programming model for RDMSR/WRMSR. Programing rules for SharedC and SharedNC registers: • Software must ensure that a shared MSR written by one core on a compute unit will not cause a problem for software that is running on the other core of the compute unit. • SharedC: A write to a SharedC MSR does not have to be written to the other core of the compute unit in order for the other core to see the updated value. • SharedNC: A write to a SharedNC MSR has to be written to both cores of the compute unit in order for both cores to see the updated value. • If software can know that the other core has not read the SharedNC MSR since the last warm reset, then a write is not needed to the SharedNC MSR on the other core. • Software may not rely on the other core maintaining the previous value of the SharedNC MSR. • The SharedNC MSRs are: MSRC001_00[35:30], MSRC001_00[53:50], MSRC001_0054, MSRC001_0055. • A read-modify-write of a shared MSR register is not atomic. Software must ensure atomicity between the cores that could simultaneously read-modify-write the shared register. 2.4.2 Virtual Address Space The processor supports 48 address bits of virtual memory space (256 TB) as indicated by CPUID Fn8000_0008_EAX. 2.4.3 Processor Cores and Downcoring The processor supports downcoring as follows: • The cores of a compute unit may be software downcored by D18F3x190[DisCore]. See 2.4.3.1 [Software Downcoring using D18F3x190[DisCore]]. • Both cores of a compute unit must be downcored if either core needs to be downcored. • The cores of a compute unit are even core x and odd core x+1.. • Clocks are turned off and power is gated to downcored compute units. The power savings is the same as CC6. • There must be at least 1 compute unit enabled. • Downcoring affects: D18F5x84[CmpCap], CPUID Fn8000_0008_ECX[NC]. • An implemented (physical) core that is downcored is not visible to software. Cores that are not downcored 31 42300 Rev 3.12 - July 14, 2015 BKDG for AMD Family 15h Models 10h-1Fh Processors are numbered logically in a contiguous manner. • D18F5x80 [Compute Unit Status] reports core topology information to software. • The number of cores specified in CPUID Fn8000_0008_ECX[NC] must be the same as the number of cores enabled in D18F0x1DC[CpuEn]. • The core number, CpuCoreNum, is provided to SW running on each core through CPUID Fn0000_0001_EBX[LocalApicId] and APIC20[ApicId]; CpuCoreNum also affects D18F0x1DC[CpuEn]. CpuCoreNum, varies as the lowest integers from 0 to D18F5x84[CmpCap], based on the number of enabled cores; e.g., a 4-core node with 1 core disabled results in cores reporting CpuCoreNum values of 0, 1, and 2 regardless of which core is disabled. The boot core is always the core reporting CpuCoreNum=0. 2.4.3.1 Software Downcoring using D18F3x190[DisCore] Cores may be downcored by D18F3x190[DisCore]. Software is required to use D18F3x190[DisCore] as follows: • Once a core has been removed by D18F3x190[DisCore]=1, it cannot be added back without a cold reset. E.g. Software may only set DisCore bits, never clear them. • If the number of cores in the system is changed, then D18F0x60[CpuCnt] must be updated to reflect the new value after the warm reset. • BIOS should configure MSRC001_102A[ThrottleNbInterface] to reflect the number of enabled compute units. 2.4.4 Physical Address Space The core supports 48 address bits of coherent memory space (256 terabytes) as indicated by CPUID Fn8000_0008_EAX [Long Mode Address Size Identifiers]. However the NB only supports 40 address bits. The processor master aborts the following upper-address transactions (to address PhysAddr): • Link or core requests with non-zero PhysAddr[63:40]. 2.4.5 System Address Map The processor defines a reserved memory address region starting at 0000_00FD_0000_0000h and extending up to 0000_0100_0000_0000h. System software must not map memory into this region. Downstream host accesses to the reserved address region results in a page fault. Upstream system device accesses to the reserved address region results in an undefined operation. 2.4.5.1 Memory Access to the Physical Address Space All memory accesses to the physical address space from a core are sent to its associated northbridge (NB). All memory accesses from a link are routed through the NB. A core access to physical address space has two important attributes that must be determined before issuing the access to the NB: the memory type (e.g., WB, WC, UC; as described in the MTRRs) and the access destination (DRAM or MMIO). This mechanism is managed by the BIOS and does not require any setup or changes by system software. 2.4.5.1.1 Determining Memory Type The memory type for a core access is determined by the highest priority of the following ranges that the access falls in: 1==Lowest priority. 1. The memory type as determined by architectural mechanisms. 32 42300 Rev 3.12 - July 14, 2015 BKDG for AMD Family 15h Models 10h-1Fh Processors • See the APM2 chapter titled “Memory System”, sections “Memory-Type Range Registers”and “Page-Attribute Table Mechanism”. • See the APM2 chapter titled “Nested Paging”, section “Combining Memory Types, MTRRs”. • See MSR0000_02FF [MTRR Default Memory Type (MTRRdefType)], MSR0000_020[F:0] [Variable-Size MTRRs Base/Mask], MSR0000_02[6F:68,59:58,50] [Fixed-Size MTRRs]. 2. TSeg & ASeg SMM mechanism. (see MSRC001_0112 and MSRC001_0113) 3. CR0[CD]: If (CR0[CD]==1) then MemType=CD. 4. MMIO config space, APIC space. • MMIO APIC space and MMIO config space must not overlap. • MemType=UC. • See 2.4.8.1.2 [APIC Register Space] and 2.7 [Configuration Space]. 5. If (“In SMM Mode”) && ((MSRC001_0113[AValid] && “The address does not fall within the ASeg region”) || (MSRC001_0113[TValid] && “The address does not fall within the TSeg region”))) then MemType=CD. 2.4.5.1.2 Determining The Access Destination for Core Accesses The access destination, DRAM or MMIO, is based on the highest priority of the following ranges that the access falls in: 1==Lowest priority. 1. RdDram/WrDram as determined by MSRC001_001A [Top Of Memory (TOP_MEM)] and MSRC001_001D [Top Of Memory 2 (TOM2)]. 2. The IORRs. (see MSRC001_00[18,16] and MSRC001_00[19,17]). 3. The fixed MTRR’s. (see MSR0000_02[6F:68,59:58,50] [Fixed-Size MTRRs]) 4. TSeg & ASeg SMM mechanism. (see MSRC001_0112 and MSRC001_0113) 5. MMIO config space, APIC space. • MMIO APIC space and MMIO config space must not overlap. • RdDram=IO, WrDram=IO. • See 2.4.8.1.2 [APIC Register Space] and 2.7 [Configuration Space]. 6. NB address space routing. See 2.8.2.1.1 [DRAM and MMIO Memory Space]. 2.4.6 Timers Each core includes the following timers. These timers do not vary in frequency regardless of the current P-state or C-state. • MSR0000_0010 [Time Stamp Counter (TSC)]; the TSC increments at the rate specified by the P0 P-state; see 2.5.3.1.2.1 [Software P-state Numbering] and MSRC001_00[6B:64] [P-state [7:0]]. • The APIC timer (APIC380 and APIC390), which increments at the rate of 2xCLKIN; the APIC timer may increment in units of between 1 and 8. 2.4.7 Implicit Conditions for TLB Invalidation The following family specific conditions will cause all TLB’s for both cores of the compute unit to be invalidated; except MSR0000_0277 which will only clear the TLB’s for the core that did the MSR write. The architectural conditions that cause TLB invalidation are documented by the APM2 section titled “TranslationLookaside Buffer (TLB)”; see “Implicit Invalidations”. • MSR0000_020[F:0] [Variable-Size MTRRs Base/Mask] • MSR0000_02[6F:68,59:58,50] [Fixed-Size MTRRs] • MSR0000_0277 [Page Attribute Table (PAT)] (TLB’s not cleared for the other core) • MSR0000_02FF [MTRR Default Memory Type (MTRRdefType)] • MSRC001_0010 [System Configuration (SYS_CFG)] write. 33 42300 Rev 3.12 - July 14, 2015 • • • • • • BKDG for AMD Family 15h Models 10h-1Fh Processors MSRC001_00[18,16] [IO Range Base (IORR_BASE[1:0])] write. MSRC001_00[19,17] [IO Range Mask (IORR_MASK[1:0])] write. MSRC001_001A [Top Of Memory (TOP_MEM)] write. MSRC001_001D [Top Of Memory 2 (TOM2)] write. MSRC001_1023 [Combined Unit Configuration (CU_CFG)] write. MSRC001_102A [Combined Unit Configuration 2 (CU_CFG2)] write. 2.4.8 2.4.8.1 Interrupts Local APIC The local APIC contains logic to receive interrupts from a variety of sources and to send interrupts to other local APICs, as well as registers to control its behavior and report status. Interrupts can be received from: • • • • • • • IO devices including the IO hub (IO APICs) Other local APICs (inter-processor interrupts) APIC timer Thermal events Performance counters Legacy local interrupts from the IO hub (INTR and NMI) APIC internal errors The APIC timer, thermal events, performance counters, local interrupts, and internal errors are all considered local interrupt sources, and their routing is controlled by local vector table entries. These entries assign a message type and vector to each interrupt, allow them to be masked, and track the status of the interrupt. IO and inter-processor interrupts have their message type and vector assigned at the source and are unaltered by the local APIC. They carry a destination field and a mode bit that together determine which local APIC(s) accepts them. The destination mode (DM) bit specifies if the interrupt request packet should be handled in physical or logical destination mode. If the destination field is FFh, the interrupt is broadcast and is accepted by all local APICs regardless of destination mode. If the destination field matches the broadcast value of FFh, then the interrupt is a broadcast interrupt and is accepted by all local APICs regardless of destination mode. 2.4.8.1.1 Detecting and Enabling APIC is detected and enabled via CPUID Fn0000_0001_EDX[APIC]. The local APIC is enabled via MSR0000_001B[ApicEn]. Reset forces APIC disabled. 2.4.8.1.2 APIC Register Space MMIO APIC space: • Memory mapped to a 4 KB range. The memory type of this space is forced to the UC memory type. The base address of this range is specified by {MSR0000_001B[ApicBar[47:12]], 000h}. • The mnemonic is defined to be APICXX; XX is the byte address offset from the base address. • MMIO APIC registers in xAPIC mode is defined by the register from APIC20 to APIC[530:500]. • Treated as normal memory space when APIC is disabled, as specified by MSR0000_001B[ApicEn]. 2.4.8.1.3 ApicId Enumeration Requirements System hardware and BIOS must ensure that the number of cores per processor (NC) exposed to the operating system by all tables, registers, and instructions across all cores in the processor is identical. See2.4.10.1 [Multi- 34 42300 Rev 3.12 - July 14, 2015 BKDG for AMD Family 15h Models 10h-1Fh Processors Core Support] to derive NC. Operating systems are expected to use CPUID Fn8000_0008_ECX[ApicIdCoreIdSize], the number of least significant bits in the Initial APIC ID that indicate core ID within a processor, in constructing per-core CPUID masks. (ApicIdCoreIdSize[3:0] determines the maximum number of cores (MNC) that the processor could theoretically support, not the actual number of cores that are actually implemented or enabled on the processor, as indicated by CPUID Fn8000_0008_ECX[NC].) BIOS must use the ApicId MNC rule when assigning APIC20 [APIC ID][ApicId] values as described below. ApicId MNC rule: The ApicId of core j must be enumerated/assigned as: ApicId[core=j] = (OFFSET_IDX) * MNC + j Where OFFSET_IDX is an integer offset (0 to N) used to shift up the core ApicId values to allow room for IOAPIC devices. It is recommended that BIOS use the following APIC ID assignments for the broadest operating system support. Given N = MNC and M = Number_Of_IOAPICs: • Assign the core ApicId’s first from 0 to N-1, and the IOAPIC IDs from N to N+(M-1). 2.4.8.1.4 Physical Destination Mode The interrupt is only accepted by the local APIC whose APIC20[ApicId] matches the destination field of the interrupt. Physical mode allows up to 255 APICs to be addressed individually. 2.4.8.1.5 Logical Destination Mode A local APIC accepts interrupts selected by APICD0 [Logical Destination (LDR)] and the destination field of the interrupt using either cluster or flat format as configured by APICE0[Format]. If flat destinations are in use, bits 7-0 of APICD0[Destination] are checked against bits 7-0 of the arriving interrupt’s destination field. If any bit position is set in both fields, the local APIC is a valid destination. Flat format allows up to 8 APICs to be addressed individually. If cluster destinations are in use, bits 7-4 of APICD0[Destination] are checked against bits 7-4 of the arriving interrupt’s destination field to identify the cluster. If all of bits 7-4 match, then bits 3-0 of APICD0[Destination] and the interrupt destination are checked for any bit positions that are set in both fields to identify processors within the cluster. If both conditions are met, the local APIC is a valid destination. Cluster format allows 15 clusters of 4 APICs each to be addressed. 2.4.8.1.6 Interrupt Delivery SMI, NMI, INIT, Startup, and External interrupts are classified as non-vectored interrupts. When an APIC accepts a non-vectored interrupt, it is handled directly by the processor instead of being queued in the APIC. When an APIC accepts a fixed or lowest-priority interrupt, it sets the bit in APIC[270:200] [Interrupt Request (IRR)] corresponding to the vector in the interrupt. For local interrupt sources, this comes from the vector field in that interrupt’s local vector table entry. The corresponding bit in APIC[1F0:180] [Trigger Mode (TMR)] is set if the interrupt is level-triggered and cleared if edge-triggered. If a subsequent interrupt with the same vector arrives when the corresponding bit in APIC[270:200][RequestBits] is already set, the two 35 42300 Rev 3.12 - July 14, 2015 BKDG for AMD Family 15h Models 10h-1Fh Processors interrupts are collapsed into one. Vectors 15-0 are reserved. 2.4.8.1.7 Vectored Interrupt Handling APIC80 [Task Priority (TPR)] and APICA0 [Processor Priority (PPR)] each contain an 8-bit priority divided into a main priority (bits 7-4) and a priority sub-class (bits 3-0). The task priority is assigned by software to set a threshold priority at which the processor is interrupted. The processor priority is calculated by comparing the main priority (bits 7-4) of APIC80[Priority] to bits 7-4 of the 8-bit encoded value of the highest bit set in APIC[170:100] [In-Service (ISR)]. The processor priority is the higher of the two main priorities. The processor priority is used to determine if any accepted interrupts (indicated by APIC[270:200][RequestBits]) are high enough priority to be serviced by the processor. When the processor is ready to service an interrupt, the highest bit in APIC[270:200][RequestBits] is cleared, and the corresponding bit is set in APIC[170:100][InServiceBits]. When the processor has completed service for an interrupt, it performs a write to APICB0 [End of Interrupt], clearing the highest bit in APIC[170:100][InServiceBits] and causing the next-highest interrupt to be serviced. If the corresponding bit in APIC[1F0:180][TriggerModeBits] is set, a write to APICB0 is performed on all APICs to complete service of the interrupt at the source. 2.4.8.1.8 Interrupt Masking Interrupt masking is controlled by the APIC410 [Extended APIC Control]. If APIC410[IerCap] is set, APIC[4F0:480] [Interrupt Enable] are used to mask interrupts. Any bit in APIC[4F0:480][InterruptEnableBits] that is clear indicates the corresponding interrupt is masked. A masked interrupt is not serviced and the corresponding bit in APIC[270:200][RequestBits] remains set. 2.4.8.1.9 Spurious Interrupts In the event that the task priority is set to or above the level of the interrupt to be serviced, the local APIC delivers a spurious interrupt vector to the processor, as specified by APICF0 [Spurious-Interrupt Vector (SVR)]. APIC[170:100] is not changed and no write to APICB0 occurs. 2.4.8.1.10 Spurious Interrupts Caused by Timer Tick Interrupt A typical interrupt is asserted until it is serviced. An interrupt is deasserted when software clears the interrupt status bit within the interrupt service routine. Timer tick interrupt is an exception, since it is deasserted regardless of whether it is serviced or not. The processor is not always able to service interrupts immediately (i.e. when interrupts are masked by clearing EFLAGS.IM). If the processor is not able to service the timer tick interrupt for an extended period of time, the INTR caused by the first timer tick interrupt asserted during that time is delivered to the local APIC in ExtInt mode and latched, and the subsequent timer tick interrupts are lost. The following cases are possible when the processor is ready to service interrupts: • An ExtInt interrupt is pending, and INTR is asserted. This results in timer tick interrupt servicing. This occurs 50 percent of the time. • An ExtInt interrupt is pending, and INTR is deasserted. The processor sends the interrupt acknowledge 36 42300 Rev 3.12 - July 14, 2015 BKDG for AMD Family 15h Models 10h-1Fh Processors cycle, but when the PIC receives it, INTR is deasserted, and the PIC sends a spurious interrupt vector. This occurs 50 percent of the time. There is a 50 percent probability of spurious interrupts to the processor. 2.4.8.1.11 Lowest-Priority Interrupt Arbitration Fixed, remote read, and non-vectored interrupts are accepted by their destination APICs without arbitration. Delivery of lowest-priority interrupts requires all APICs to arbitrate to determine which one accepts the interrupt. If APICF0[FocusDisable] is clear, then the focus processor for an interrupt always accepts the interrupt. A processor is the focus of an interrupt if it is already servicing that interrupt (corresponding bit in APIC[170:100][InServiceBits] is set) or if it already has a pending request for that interrupt (corresponding bit in APIC[270:200][RequestBits] is set). If APIC410[IerCap] is set the interrupt must also be enabled in APIC[4F0:480][InterruptEnableBits] for a processor to be the focus processor. If there is no focus processor for an interrupt, or focus processor checking is disabled, then each APIC calculates an arbitration priority value, stored in APIC90 [Arbitration Priority (APR)], and the one with the lowest result accepts the interrupt. The arbitration priority value is calculated by comparing APIC80[Priority] with the 8-bit encoded value of the highest bit set in APIC[270:200][RequestBits] (IRRVec) and the 8-bit encoded value of the highest bit set APIC[170:100][InServiceBits] (ISRVec). If APIC410[IerCap] is set the IRRVec and ISRVec are based off the highest enabled interrupt. The main priority bits 7-4 are compared as follows: If (APIC80[Priority[7:4]] >= IRRVec[7:4] and APIC80[Priority[7:4]] > ISRVec[7:4]) Then APIC90[Priority] = APIC80[Priority] Else if (IRRVec[7:4] > ISRVec[7:4]) APIC90[Priority] = {IRRVec[7:4],0h} Else APIC90[Priority] = {ISRVect[7:4],0h} 2.4.8.1.12 Inter-Processor Interrupts APIC300 [Interrupt Command Low (ICR Low)] and APIC310 [Interrupt Command High (ICR High)] provide a mechanism for generating interrupts in order to redirect an interrupt to another processor, originate an interrupt to another processor, or allow a processor to interrupt itself. A write to register APIC300 causes an interrupt to be generated with the properties specified by the APIC300 and APIC310 fields. 2.4.8.1.13 APIC Timer Operation The local APIC contains a 32-bit timer, controlled by APIC320 [LVT Timer], APIC380 [Timer Initial Count], and APIC3E0 [Timer Divide Configuration]. The processor bus clock is divided by the value in APIC3E0[Div] to obtain a time base for the timer. When APIC380[Count] is written, the value is copied into APIC390 [Timer Current Count]. APIC390[Count] is decremented at the rate of the divided clock. When the count reaches 0, a timer interrupt is generated with the vector specified in APIC320[Vector]. If APIC320[Mode] specifies periodic operation, APIC390[Count] is reloaded with the APIC380[Count] value, and it continues to decrement at the rate of the divided clock. If APIC320[Mask] is set, timer interrupts are not generated. 2.4.8.1.14 Generalized Local Vector Table All LVTs (APIC330 to APIC3[60:50], and APIC[530:500]) support a generalized message type as follows: • 000b=Fixed • 010b=SMI • 100b=NMI • 111b=ExtINT 37 42300 Rev 3.12 - July 14, 2015 BKDG for AMD Family 15h Models 10h-1Fh Processors • All other messages types are reserved. 2.4.8.1.15 State at Reset At power-up or reset, the APIC is hardware disabled (MSR0000_001B[ApicEn]=0) so only SMI, NMI, INIT, and ExtInt interrupts may be accepted. The APIC can be software disabled through APICF0[APICSWEn]. The software disable has no effect when the APIC is hardware disabled. When a processor accepts an INIT interrupt, the APIC is reset as at power-up, with the exception that: • APIC20[ApicId] is unaffected. • Pending APIC register writes complete. 2.4.8.2 System Management Mode (SMM) System management mode (SMM) is typically used for system control activities such as power management. These activities are typically transparent to the operating system. 2.4.8.2.1 SMM Overview SMM is entered by a core on the next instruction boundary after a system management interrupt (SMI) is received and recognized. A core may be programmed to broadcast a special cycle to the system, indicating that it is entering SMM mode. The core then saves its state into the SMM memory state save area and jumps to the SMI service routine (or SMI handler). The pointer to the SMI handler is specified by MSRs. The code and data for the SMI handler are stored in the SMM memory area, which may be isolated from the main memory accesses. The core returns from SMM by executing the RSM instruction from the SMI handler. The core restores its state from the SMM state save area and resumes execution of the instruction following the point where it entered SMM. The core may be programmed to broadcast a special bus cycle to the system, indicating that it is exiting SMM mode. 2.4.8.2.2 Operating Mode and Default Register Values The software environment after entering SMM has the following characteristics: • Addressing and operation is in Real mode. • A far jump, call or return in the SMI handler can only address the lower 1M of memory, unless the SMI handler first switches to protected mode. • If (MSRC001_0111[SmmBase]>=0010_0000h) then: • The value of the CS selector is undefined upon SMM entry. • The undefined CS selector value should not be used as the target of a far jump, call, or return. • 4-Gbyte segment limits. • Default 16-bit operand, address, and stack sizes (instruction prefixes can override these defaults). • Control transfers that do not override the default operand size truncate the EIP to 16 bits. • Far jumps or calls cannot transfer control to a segment with a base address requiring more than 20 bits, as in Real mode segment-base addressing, unless a change is made into protected mode. • A20M# is disabled. A20M# assertion or deassertion have no affect during SMM. • Interrupt vectors use the Real mode interrupt vector table. • The IF flag in EFLAGS is cleared (INTR is not recognized). • The TF flag in EFLAGS is cleared. 38 42300 Rev 3.12 - July 14, 2015 BKDG for AMD Family 15h Models 10h-1Fh Processors • The NMI and INIT interrupts are masked. • Debug register DR7 is cleared (debug traps are disabled). The SMM base address is specified by MSRC001_0111 [SMM Base Address (SMM_BASE)][SmmBase]. Important offsets to the base address pointer are: • MSRC001_0111[SmmBase] + 8000h: SMI handler entry point. • MSRC001_0111[SmmBase] + FE00h - FFFFh: SMM state save area. 2.4.8.2.3 SMI Sources And Delivery The processor accepts SMIs as link-defined interrupt messages only. The core/node destination of these SMIs is a function of the destination field of these messages. However, the expectation is that all such SMI messages are specified to be delivered globally (to all cores of all nodes). There are also several local events that can trigger SMIs. However, these local events do not generate SMIs directly. Each of them triggers a programmable IO cycle that is expected to target the SMI command port in the IO hub and trigger a global SMI interrupt message back to the coherent fabric. Local sources of SMI events that generate the IO cycle specified in MSRC001_0056 [SMI Trigger IO Cycle] are: • In the core, as specified by: • MSRC001_0022 [Machine Check Exception Redirection]. • MSRC001_00[53:50] [IO Trap (SMI_ON_IO_TRAP_[3:0])]. • All local APIC LVT registers programmed to generate SMIs. The status for these is stored in SMMFEC4. 2.4.8.2.4 SMM Initial State After storing the save state, execution starts at MSRC001_0111[SmmBase] + 08000h. The SMM initial state is specified in the following table. Table 5: SMM Initial State Register CS DS ES FS GS SS General-Purpose Registers EFLAGS RIP CR0 CR4 GDTR LDTR SMM Initial State SmmBase[19:4] 0000h 0000h 0000h 0000h 0000h Unmodified 0000_0002h 0000_0000_0000_8000h Bits 0, 2, 3, and 31 cleared (PE, EM, TS, and PG); remainder is unmodified 0000_0000_0000_0000h Unmodified Unmodified 39 42300 Rev 3.12 - July 14, 2015 BKDG for AMD Family 15h Models 10h-1Fh Processors Table 5: SMM Initial State Register IDTR TR DR6 DR7 EFER 2.4.8.2.5 SMM Initial State Unmodified Unmodified Unmodified 0000_0000_0000_0400h All bits are cleared except bit 12 (SVME) which is unmodified. SMM Save State In the following table, the offset field provides the offset from the SMM base address specified by MSRC001_0111 [SMM Base Address (SMM_BASE)]. Table 6: SMM Save State Offset FE00h FE02h FE08h FE10h FE12h FE18h FE20h FE22h FE28h FE30h FE32h FE38h FE40h FE42h FE44h Size Word 6 Bytes Quadword Word 6 Bytes Quadword Word 6 Bytes Quadword Word 6 Bytes Quadword Word 2 Bytes Doubleword FE48h FE50h FE52h FE54h Quadword Word 2 Bytes Doubleword FE58h FE60h FE64h FE66h FE68h FE70h FE72h FE74h FE78h Quadword 4 Bytes Word 2 Bytes Quadword Word Word Doubleword Quadword Contents ES Selector Reserved Descriptor in memory format CS Selector Reserved Descriptor in memory format SS Selector Reserved Descriptor in memory format DS Selector Reserved Descriptor in memory format FS Selector Reserved GS FS Base {16'b[47], 47:32}1 Descriptor in memory format Selector Reserved GS Base {16'b[47], 47:32}1 Descriptor in memory format GDTR Reserved Limit Reserved Descriptor in memory format LDTR Selector Attributes Limit Base Access Read-only Read-only Read-only Read-only Read-only Read-only Read-only Read-only 40 42300 Rev 3.12 - July 14, 2015 BKDG for AMD Family 15h Models 10h-1Fh Processors Table 6: SMM Save State Offset FE80h FE84h FEB6h FE88h FE90h FE92h FE94h FE98h FEA0h FEA8h FEB0h FEB8h FEC0h FEC4 FEC8h FEC9h FECAh FECBh FED0h FED8h FEE0h FEE8h FEF0h FEFCh FF00h FF04h FF20h FF28h Size 4 Bytes Word 2 Bytes Quadword Word Word Doubleword Quadword Quadword Quadword Quadword Quadword Doubleword Doubleword Byte Byte Byte 5 Bytes Quadword Quadword Quadword Quadword 16 Bytes Doubleword Doubleword 28 Bytes Quadword Quadword Contents IDTR Reserved Limit Reserved Base TR Selector Attributes Limit Base IO_RESTART_RIP IO_RESTART_RCX IO_RESTART_RSI IO_RESTART_RDI SMMFEC0 [SMM IO Trap Offset] SMMFEC4 [Local SMI Status] SMMFEC8 [SMM IO Restart Byte] SMMFEC9 [Auto Halt Restart Offset] SMMFECA [NMI Mask] Reserved EFER SMMFED8 [SMM SVM State] Guest VMCB physical address SVM Virtual Interrupt Control Reserved SMMFEFC [SMM-Revision Identifier] SMMFF00 [SMM Base Address (SMM_BASE)] Reserved Guest PAT FF30h Quadword Host CR42 FF38h Quadword Nested CR32 FF40h Quadword FF48h FF50h FF58h FF60h FF68h FF70h Quadword Quadword Quadword Quadword Quadword Quadword Host Cr02 CR4 CR3 CR0 DR7 DR6 RFLAGS Host EFER Access Read-only Read-only Read-only Read-only Read-only Read-write Read-write Read-write Read-only Read-only Read-only Read-only Read-only Read-write Read-only 2 Read-only Read-write 41 42300 Rev 3.12 - July 14, 2015 BKDG for AMD Family 15h Models 10h-1Fh Processors Table 6: SMM Save State Offset FF78h FF80h FF88h FF90h FF98h FFA0h FFA8h FFB0h FFB8h FFC0h FFC8h FFD0h FFD8h FFE0h FFE8h FFF0h FFF8h Size Quadword Quadword Quadword Quadword Quadword Quadword Quadword Quadword Quadword Quadword Quadword Quadword Quadword Quadword Quadword Quadword Quadword Contents RIP R15 R14 R13 R12 R11 R10 R9 R8 RDI RSI RBP RSP RBX RDX RCX RAX Access Read-write Read-write Notes: 1. This notation specifies that bit[47] is replicated in each of the 16 MSB’s of the DW (sometimes called sign extended). The 16 LSB’s contain bits[47:32]. 2. Only used for an SMI in guest mode with nested paging enabled. The SMI save state includes most of the integer execution unit. Not included in the save state are: the floating point state, MSRs, and CR2. In order to be used by the SMI handler, these must be saved and restored. The save state is the same, regardless of the operating mode (32-bit or 64-bit). The following are some offsets in the SMM save state area. The mnemonic for each offset is in the form SMMxxxx, where xxxx is the offset in the save state. SMMFEC0 SMM IO Trap Offset If the assertion of SMI is recognized on the boundary of an IO instruction, SMMFEC0 [SMM IO Trap Offset] contains information about that IO instruction. For example, if an IO access targets an unavailable device, the system can assert SMI and trap the IO instruction. SMMFEC0 then provides the SMI handler with information about the IO instruction that caused the trap. After the SMI handler takes the appropriate action, it can reconstruct and then re-execute the IO instruction from SMM. Or, more likely, it can use SMMFEC8 [SMM IO Restart Byte], to cause the core to re-execute the IO instruction immediately after resuming from SMM. Bits Description 31:16 Port: trapped IO port address. Read-only. This provides the address of the IO instruction. 15:12 BPR: IO breakpoint match. Read-only. 11 TF: EFLAGS TF value. Read-only. 42 42300 Rev 3.12 - July 14, 2015 BKDG for AMD Family 15h Models 10h-1Fh Processors 10:7 Reserved 6 SZ32: size 32 bits. Read-only. 1=Port access was 32 bits. 5 SZ16: size 16 bits. Read-only. 1= Port access was 16 bits. 4 SZ8: size 8 bits. Read-only. 1=Port access was 8 bits. 3 REP: repeated port access. Read-only. 2 STR: string-based port access. Read-only. 1 V: IO trap word valid. Read-only. 1=The core entered SMM on an IO instruction boundary; all information in this offset is valid. 0=The other fields of this offset are not valid. 0 RW: port access type. Read-only. 0=IO write (OUT instruction). 1=IO read (IN instruction). SMMFEC4 Local SMI Status This offset stores status bits associated with SMI sources local to the core. For each of these bits, 1=The associated mechanism generated an SMI. Bits Description 31:20 Reserved. 19 SmiSrcThrCntHt: SMI source link thresholding. Read-only. This bit is associated with the SMI source specified in the link thresholding register (see MSR0000_0403 [LS Machine Check Miscellaneous (MC0_MISC)]). 18 Reserved. 17 SmiSrcLvtExt: SMI source LVT extended entry. Read-only. This bit is associated with the SMI sources specified in APIC[530:500] [Extended Interrupt [3:0] Local Vector Table]. 16 SmiSrcLvtLcy: SMI source LVT legacy entry. Read-only. This bit is associated with the SMI sources specified by the non-extended LVT entries of the APIC. 15:11 Reserved. 10 IntPendSmiSts: interrupt pending SMI status. Read-only. This bit is associated with the SMI source specified in MSRC001_0055 [Interrupt Pending][IntrPndMsg] (when that bit is high). 9 Reserved. 8 MceRedirSts: machine check exception redirection status. Read-only. This bit is associated with the SMI source specified in MSRC001_0022 [Machine Check Exception Redirection][RedirSmiEn]. 7:4 Reserved. 3:0 IoTrapSts: IO trap status. Read-only. Each of these bits is associated with each of the respective SMI sources specified in MSRC001_00[53:50] [IO Trap (SMI_ON_IO_TRAP_[3:0])]. SMMFEC8 SMM IO Restart Byte 00h on entry into SMM. If the core entered SMM on an IO instruction boundary, the SMI handler may write this to FFh. This causes the core to re-execute the trapped IO instruction immediately after resuming from SMM. The SMI handler should only write to this byte if SMMFEC0[V]=1; otherwise, the behavior is undefined. If a second SMI is asserted while a valid IO instruction is trapped by the first SMI handler, the core services the second SMI prior to re-executing the trapped IO instruction. SMMFEC0[V]=0 during the second entry into 43 42300 Rev 3.12 - July 14, 2015 BKDG for AMD Family 15h Models 10h-1Fh Processors SMM, and the second SMI handler must not rewrite this byte. If there is a simultaneous SMI IO instruction trap and debug breakpoint trap, the processor first responds to the SMI and postpones recognizing the debug exception until after resuming from SMM. If debug registers other than DR6 and DR7 are used while in SMM, they must be saved and restored by the SMI handler. If SMMFEC8 [SMM IO Restart Byte], is set to FFh when the RSM instruction is executed, the debug trap does not occur until after the IO instruction is re-executed. Bits Description 7:0 RST: SMM IO Restart Byte. Read-write. SMMFEC9 Auto Halt Restart Offset Bits Description 7:1 0 Reserved. HLT: halt restart. Read-write. Upon SMM entry, this bit indicates whether SMM was entered from the Halt state. 0=Entered SMM on a normal x86 instruction boundary. 1=Entered SMM from the Halt state. Before returning from SMM, this bit can be written by the SMI handler to specify whether the return from SMM should take the processor back to the Halt state or to the instruction-execution state specified by the SMM state save area (normally, the instruction after the halt). 0=Return to the instruction specified in the SMM save state. 1=Return to the halt state. If the return from SMM takes the processor back to the Halt state, the HLT instruction is not refetched and re-executed. However, the Halt special bus cycle is broadcast and the processor enters the Halt state. SMMFECA NMI Mask Bits Description 7:1 0 Reserved. NmiMask. Read-write. Specifies whether NMI was masked upon entry to SMM. 0=NMI not masked. 1=NMI masked. SMMFED8 SMM SVM State Read-only. This offset stores the SVM state of the processor upon entry into SMM. Bits Description 63:4 Reserved. 44 42300 Rev 3.12 - July 14, 2015 3 2:0 BKDG for AMD Family 15h Models 10h-1Fh Processors HostEflagsIf: host Eflags IF. SvmState. Bits 000b 001b 010b 101b-011b 110b 111b Definition SMM entered from a non-guest state. Reserved. SMM entered from a guest state. Reserved. SMM entered from a guest state with nested paging enabled. Reserved. SMMFEFC SMM-Revision Identifier SMM entry state: 0003_0064h Bits Description 31:18 Reserved. 17 BRL. Read-only. Base relocation supported. 16 IOTrap. Read-only. IO trap supported. 15:0 Revision. Read-only. SMMFF00 SMM Base Address (SMM_BASE) Bits Description 31:0 See: MSRC001_0111[SmmBase]. 2.4.8.2.6 Exceptions and Interrupts in SMM When SMM is entered, the core masks INTR, NMI, SMI, INIT, and A20M interrupts. The core clears the IF flag to disable INTR interrupts. To enable INTR interrupts within SMM, the SMM handler must set the IF flag to 1. A20M is disabled so that address bit 20 is never masked when in SMM. Generating an INTR interrupt can be used for unmasking NMI interrupts in SMM. The core recognizes the assertion of NMI within SMM immediately after the completion of an IRET instruction. Once NMI is recognized within SMM, NMI recognition remains enabled until SMM is exited, at which point NMI masking is restored to the state it was in before entering SMM. While in SMM, the core responds to the DBREQ and STPCLK interrupts, as well as to all exceptions that may be caused by the SMI handler. 2.4.8.2.7 The Protected ASeg and TSeg Areas These ranges are controlled by MSRC001_0112 and MSRC001_0113; see those registers for details. 2.4.8.2.8 SMM Special Cycles Special cycles can be initiated on entry and exit from SMM to acknowledge to the system that these transitions are occurring. These are controlled by MSRC001_0015[SMISPCYCDIS, RSMSPCYCDIS]. 45 42300 Rev 3.12 - July 14, 2015 2.4.8.2.9 BKDG for AMD Family 15h Models 10h-1Fh Processors Locking SMM The SMM registers (MSRC001_0112 and MSRC001_0113) can be locked from being altered by setting MSRC001_0015[SmmLock]. The BIOS can lock the SMM registers after initialization to prevent unexpected changes to these registers. 2.4.8.2.10 Synchronizing SMM Entry (Spring-Boarding) The BIOS must take special care to ensure that all cores have entered SMM prior to accessing shared IO resources and all core SMI interrupt status bits are synchronized. This generally requires that BIOS waits for all cores to enter SMM. The following conditions can cause one or more cores to enter SMM without all cores entering SMM: • More than one IO device in the system is enabled to signal an SMI without hardware synchronization (e.g. using an end of SMI gate). • A single device may signal multiple SMI messages without hardware synchronization (e.g. using an end of SMI gate). • An SMI is received while one or more AP cores are in the INIT state. This may occur either during BIOS or secure boot. • A hardware error prevents a core from entering SMM. The act of synchronizing cores into SMM is called spring-boarding. Because not all of the above conditions can be avoided, it is recommended that all systems support spring-boarding. An ACPI-compliant IO hub is required for spring-boarding. Depending on the IO hub design, BIOS may have to set additional end-of-SMI bits to trigger an SMI from within SMM. The software requirements for the suggested spring-boarding implementation are listed as follows. • A binary semaphore located in SMRAM, accessible by all cores. For the purpose of this discussion, the semaphore is called CheckSpringBoard. CheckSpringBoard is initialized to zero. • Two semaphores located in SMRAM, accessible by all cores. For the purpose of this discussion, the semaphores are called NotInSMM and WaitInSMM. NotInSMM and WaitInSMM are initialized to a value equal to the number of cores in the system (NumCPUs). The following BIOS algorithm describes spring-boarding and is optimized to reduce unnecessary SMI activity. This algorithm must be made part of the SMM instruction sequence for each core in the system. 1. Attempt to obtain ownership of the CheckSpringBoard semaphore with a read-modify-write instruction. If ownership was obtained then do the following, else proceed to step 2: • Check all enabled SMI status bits in the IO hub. Let Status=enable1&status1 | enable2&status2 | enable3&status3 … enable n & status n. • If (Status==0) then perform the following sub-actions. • Trigger an SMI broadcast assertion from the IO hub by writing to the software SMI command port. • Resume from SMM with the RSM instruction. //Example: InLineASM{ BTS CheckSpringBoard,0; Try to obtain ownership of semaphore JC Step_2: CALL CheckIOHUB_SMIEVT; proc returns ZF=1 for no events JNZ Step_2: CALL Do_SpringBoard;Trigger SMI and then RSM 46 42300 Rev 3.12 - July 14, 2015 2. 3. 4. 5. 6. 7. 8. BKDG for AMD Family 15h Models 10h-1Fh Processors Step_2: } Decrement the NotInSMM variable. Wait for (NotInSMM==0). See Note 1. Execute the core-local event SMI handler. Using a third semaphore (not described here), synchronize core execution at the end of the task. After all cores have executed, proceed to step 4. The following is a brief description of the task for each core: • Check all enabled core-local SMI status bits in the core’s private or MSR address space. Handle the event if possible, or pass information necessary to handle the event to a mailbox for the BSC to handle. • An exclusive mailbox must exist for each core for each core local event. • On-line spare events should be handled in this task by the individual core for optimal performance. Assign one core of a dual core processor to handle On-line spare. These events may be optionally handled by the BSC just as other global events. • Wait for all cores to complete this task at least once. If the current core executing instructions is not the BSC then jump to step 5. If the core executing instructions is the BSC then jump to the modified main SMI handler task, described below. • Check all enabled SMI status bits in the IO hub. Check mailboxes for event status. • For each event, handle the event and clear the corresponding status bit. • Repeat until all enabled SMI status bits are clear and no mailbox events remain. • Set NotInSMM=NumCPUs. (Jump to step 5.) Decrement the WaitInSMM variable. Wait for WaitInSMM=0. See Note 2. Increment the WaitInSMM variable. Wait for WaitInSMM=NumCPUs. If the current processor core executing instructions is the BSC then reset CheckSpringBoard to zero. Resume from SMM with the RSM instruction. Notes: 1. To support a secure startup by the secure loader the BIOS must provide a timeout escape from the otherwise endless loop. The timeout value should be large enough to account for the latency of all cores entering SMM. The maximum SMM entrance latency is defined by the platform’s IO sub-system, not the processor. A value of twice the watchdog timer count is recommended. See D18F3x44 [MCA NB Configuration] for more information on the watchdog time-out value. If a time-out occurs in the wait loop, the BIOS (the last core to decrement NotInSMM) should record the number of cores that have not entered SMM and all cores must fall out of the loop. 2. If a time-out occurs in the wait loop in step 2, the BIOS must not wait for WaitInSMM=0. Instead it must wait for WaitInSMM=”the number of cores recorded in step 2”. 2.4.9 Secure Virtual Machine Mode (SVM) Support for SVM mode is indicated by CPUID Fn8000_0001_ECX[SVM]. 2.4.9.1 BIOS support for SVM Disable The BIOS should include the following user setup options to enable and disable AMD Virtualization™ technology. • Enable AMD Virtualization™. • MSRC001_0114[Svm_Disable] = 0. • MSRC001_0114[Lock] = 1. • MSRC001_0118[SvmLockKey] = 0000_0000_0000_0000h. • Disable AMD Virtualization™. • MSRC001_0114[Svm_Disable]=1. • MSRC001_0114[Lock]=1. 47 42300 Rev 3.12 - July 14, 2015 BKDG for AMD Family 15h Models 10h-1Fh Processors • MSRC001_0118[SvmLockKey] = 0000_0000_0000_0000h. The BIOS may also include the following user setup options to disable AMD Virtualization™. • Disable AMD Virtualization™, with a user supplied key. • MSRC001_0114[Svm_Disable]=1. • MSRC001_0114[Lock]=1. • MSRC001_0118[SvmLockKey] programmed with value supplied by user. This value should be stored in NVRAM. 2.4.10 CPUID Instruction The CPUID instruction provides data about the features supported by the processor. See 3.17 [CPUID Instruction Registers]. 2.4.10.1 Multi-Core Support There are two methods for determining multi-core support. A recommended mechanism is provided and a legacy method is also available for existing operating systems. System software should use the correct architectural mechanism to detect the number of physical cores by observing CPUID Fn8000_0008_ECX[NC]. The legacy method utilizes the CPUID Fn0000_0001_EBX[LogicalProcessorCount]. 48 42300 Rev 3.12 - July 14, 2015 2.5 BKDG for AMD Family 15h Models 10h-1Fh Processors Power Management The processor supports many power management features in a variety of systems. Table 7 provides a summary of ACPI states and power management features and indicates whether they are supported. Table 7: Power Management Support ACPI/Power Management State Supported Description G0/S0/C0: Working Yes G0/S0/C0: Core P-state transitions Yes G0/S0/C0: NB P-state transitions Yes 2.5.4.1 [NB P-states] G0/S0/C0: Hardware thermal control (HTC) Yes 2.10.4.1 [PROCHOT_L and Hardware Thermal Control (HTC)] G0/S0/C0: P-state limit control Yes 2.10.4.3 [Software P-state Limit Control] G0/S0/C0: Thermal clock throttling (SMC controlled) No G0/S0/Per-core IO-based C-states Yes G0/S0/C1: Halt Yes G0/S0/PC4: Altvid (VDD power plane) No G0/S0/C5: Deeper altvid support (VDD power plane) No G0/S0/CC6: Power gating Yes G0/S0/Cx: Cache flushing support Yes G1/S1: Stand By (Powered On Suspend) No G1/S3: Stand By (Suspend to RAM) Yes G1/S4, S5: Hibernate (Suspend to Disk), Shut Down (Soft Off) Yes G3 Mechanical Off Yes Parallel VID Interface No Serial VID Interface 1 No Serial VID Interface 2 Yes Single-plane systems No Dual-plane systems Yes Triple-plane systems No 2.5.1 2.5.3.1 [Core P-states] 2.5.3.2 [Core C-states] 2.5.8.1 [S-states] 2.5.1 [Processor Power Planes And Voltage Control] Processor Power Planes And Voltage Control Refer to the Electrical Data Sheet for AMD Family 15h Models 10h-1Fh Processors for power plane definitions. 2.5.1.1 Serial VID Interface The processor includes an interface to control external voltage regulators, called the serial VID interface (SVI). Only SVI2 is supported. The frequency of SVC for SVI2 is controlled by D18F3xA0[Svi2HighFreqSel]. 2.5.1.1.1 SVI2 Features The processor supports the following SVI2 features: • Voltage offsets: • VDD: D18F5x12C[CoreOffsetTrim]. • VDDNB: D18F5x188[NbOffsetTrim]. 49 42300 Rev 3.12 - July 14, 2015 BKDG for AMD Family 15h Models 10h-1Fh Processors • Load line trim: • VDD: D18F5x12C[CoreLoadLineTrim]. • VDDNB: D18F5x188[NbLoadLineTrim]. 2.5.1.2 Internal VID Registers and Encodings All VID register fields within the processor are 8-bits wide. The SVI2 VID encoding to voltage level is calculated by the following formula: IF (Vid[7:0] < 1111_1000b) THEN Voltage (V) = 1.5500 - 0.00625 * Vid[7:0] ELSE Voltage (V) = OFF END The boot VID is 1.0 volts. 2.5.1.2.1 MinVid and MaxVid Check Hardware limits the minimum and maximum VID code that is sent to the voltage regulator. The allowed limits are specified in D18F5x17C[MinVid, MaxVid]. Prior to generating VID-change commands to SVI, the processor filters the InputVid value to the OutputVid as follows (higher VID codes correspond to lower voltages and lower VID codes correspond to higher voltages): • If InputVid < MaxVid, OutputVid=MaxVid. • Else if (InputVid > MinVid) & (MinVid!=00h), OutputVid=MinVid. • Else OutputVid=InputVid. This filtering is applied regardless of the source of the VID-change command. 2.5.1.3 2.5.1.3.1 Low Power Features PSIx_L Bit The processor supports indication of whether the processor is in a low-voltage state or not, which may be used by the regulator to place itself into a more power efficient mode. The PSIx_L bit can be controlled for the VDD and VDDNB power planes independently. • The processor supports the PSI0_L and the PSI1_L bits in the data fields of the SVI2 command. • PSI0_L: PSI0_L is enabled using D18F3xA0[PsiVidEn] and D18F5x17C[NbPsi0VidEn]. Once enabled, the state of PSI0_L is controlled by D18F3xA0[PsiVid[7:0]] and D18F5x17C[NbPsi0Vid[7:0]]. • PSI1_L: The PSI1_L bit for VDD functions as specified by D18F5x12C[CorePsi1En]. The PSI1_L bit for VDDNB functions as specified by D0F0xBC_x1F5F8[EnableNbPsi1]. The processor sends PSIx_L changes to the voltage regulator whenever the criteria associated with PSIx_L changes, no voltage change is necessary. 2.5.1.3.1.1 BIOS Requirements for PSI0_L Enabling PSI0_L for the VDD and VDDNB planes depends on support from the voltage regulator and is therefore system specific. The voltage regulator must be able to supply the current required for the processor to operate at the VID code specified in D18F3xA0[PsiVid[7:0]] and D18F5x17C[NbPsi0Vid[7:0]]. Depending on the regulator used, AMD recommends one of the following methods: 50 42300 Rev 3.12 - July 14, 2015 BKDG for AMD Family 15h Models 10h-1Fh Processors • PSI0_L is disabled: • VDD: To set PSI0_L for the VDD plane, program D18F3xA0[PsiVidEn]=0. • VDDNB: To set PSI0_L for the VDDNB plane, program D18F5x17C[NbPsi0VidEn]=0. • PSI0_L set/clear based on current requirements: • VDD: The following algorithm describes how to program PSI0_L on VDD: PSI_vrm_current = current at which the regulator allows PSI0_L. previous_voltage = FFh for (each P-state from P0 to D18F3xDC[HwPstateMaxVal]) { pstate_current = ProcIddMax for the current P-state, see 2.5.3.1.8 [Processor-Systemboard Power Delivery Compatibility Check]; pstate_voltage = MSRC001_00[6B:64][CpuVid] of the current P-state; if (current P-state == D18F3xDC[HwPstateMaxVal]) { next_pstate_current = 0; } else { next_pstate_current = ProcIddMax for the next P-state, see 2.5.3.1.8 [Processor-Systemboard Power Delivery Compatibility Check]; } if ((pstate_current <= PSI_vrm_current) && (next_pstate_current <= PSI_vrm_current) && (pstate_voltage != previous_voltage)) { Program D18F3xA0[PsiVid] = pstate_voltage; Program D18F3xA0[PsiVidEn] = 1; break; } previous_voltage = pstate_voltage; } • VDDNB: The following algorithm describes how to program PSI0_L on VDDNB: NbIddMax = D18F5x1[6C:60][NbIddDiv] current. PSI_vrm_current = current at which the VDDNB regulator allows PSI0_L. previous_voltage = FFh for (each valid NB P-state starting with NBP0) { pstate_current = NbIddMax of the current NB P-state; pstate_voltage = D18F5x1[6C:60][NbVid] of the current NB P-state; if (current NB P-state is the last valid NB P-state) { next_pstate_current = 0; } else { next_pstate_current = NbIddMax for the next NB P-state; } if ((pstate_current <= PSI_vrm_current) && (next_pstate_current <= PSI_vrm_current) && (pstate_voltage != previous_voltage)) { Program D18F5x17C[NbPsi0Vid] = pstate_voltage; Program D18F5x17C[NbPsi0VidEn] = 1; break; } previous_voltage = pstate_voltage; } 51 42300 Rev 3.12 - July 14, 2015 2.5.1.3.1.2 BKDG for AMD Family 15h Models 10h-1Fh Processors BIOS Requirements for PSI1_L • VDD: Program D18F5x12C[CorePsi1En] to its recommended setting. • VDDNB: Please see your AMD representative for details. 2.5.1.3.2 Low Power Voltages In order to save power, voltages lower than those normally needed for operation may be applied to the VDD power plane while the processor is in a C-state or S-state. The lower voltage is defined as follows: • PC6Vid: D18F5x128[PC6Vid] specifies a voltage that does not retain the CPU caches or the core microarchitectural state. PC6Vid does not allow execution and is only applied to the cores. See 2.5.3.2.3.4 [Package C6 (PC6) State]. 2.5.1.4 Voltage Transitions The processor supports dynamic voltage transitions on the VDD and VDDNB planes. These transitions are requested by either hardware or software during state changes such as reset, P-state changes, and C-state changes. In all cases the VID code passed to the voltage regulator changes from the old value to the new value without stepping through intermediate values. The voltage regulator ramps the voltage directly from the starting voltage to the final voltage, no stepping occurs. See the AMD Serial VID Interface 2.0 (SVI2) Specification for additional details. • If a voltage increase is requested, the processor waits as specified by D18F5x12C[WaitVidCompDis] before sending any additional voltage change requests to the voltage regulator or before beginning a frequency transition. • If a voltage decrease is requested, the processor waits the amount of time specified by D18F5x128[FastSlamTimeDown] before sending any additional voltage change requests to the voltage regulator. For voltage decreases, the processor does not wait any time before beginning frequency changes. The processor continues code execution during voltage changes when in the C0 state. 2.5.1.4.1 Hardware-Initiated Voltage Transitions When software requests any of the following state changes, or hardware determines that any of the following state changes are necessary, hardware coordinates the necessary voltage changes: • VDD: • Core P-state transition. See 2.5.3.1 [Core P-states]. • Package C-state transition. See 2.5.3.2 [Core C-states]. • S-state transition. See 2.5.8.1 [S-states]. • VDDNB: • NB P-state transition. See 2.5.4.1 [NB P-states]. • S-state transition. See 2.5.8.1 [S-states]. 2.5.1.4.2 2.5.1.4.2.1 Software-Initiated Voltage Transitions Software-Initiated NB Voltage Transitions Software can request voltage changes on the VDDNB power plane using the following control/status register 52 42300 Rev 3.12 - July 14, 2015 BKDG for AMD Family 15h Models 10h-1Fh Processors pairs: • GMMx63C and GMMx640 • GMMx770 and GMMx774 The voltage requests from each register pair are considered independently by hardware when taking voltage plane dependencies into account (see 2.5.2.2 [Dependencies Between Subcomponents on VDDNB]). To make a voltage change request, software uses the following sequence: 1. Ensure VoltageChangeEn==1 in the control register. If software needs to program VoltageChangeEn=1, software must perform this register write independently of the writes in the following steps. 2. Program VoltageLevel to the desired voltage and toggle VoltageChangeReq in the control register. 3. The voltage change is complete when VoltageChangeReq in the control register is equal to VoltageChangeAck in the status register. Software can force a VDDNB voltage change using GMMx63C and GMMx770. To do so, software programs VoltageForceEn=1 in the respective register before toggling VoltageChangeReq when making a voltage change request. If this is done, the voltage requested overrides any other VDDNB voltage requests made by software or DPM state transitions when determining voltage plane dependencies (see 2.5.2.2 [Dependencies Between Subcomponents on VDDNB]). NB P-state transitions still request voltage transitions as normal when software forces a voltage change using this mechanism. If software forces a voltage change using both GMMx63C and GMMx770, the voltage requested in GMMx63C takes precedence. The following registers also cause VDDNB voltage transitions: • MSRC001_0070[NbVid]: See 2.5.4.1 [NB P-states]. 2.5.1.4.2.2 Software-Initiated Core Voltage Transitions To force VDD voltage changes, software must take the following steps: 1. Write the destination CpuVid to MSRC001_0070[CpuVid]. 2. Wait the specified D18F3xD8[VSRampSlamTime] . 2.5.2 2.5.2.1 Frequency and Voltage Domain Dependencies Dependencies Between Cores Whenever a P-state or C-state is requested on a core (see 2.5.3.1 [Core P-states] and 2.5.3.2 [Core C-states]), hardware must take the following frequency and voltage domain dependencies into account when deciding whether to make the requested change: • Cores within a compute unit share a common frequency and voltage domain. • Compute units within a processor share a common voltage domain, but have independent frequency domains. The voltage is determined by the highest-performance P-state requested on any core. As a result, the P-state and C-state change requests have the following results: • If different compute units request different voltages, the VDD voltage is determined by the highest voltage (lowest VID) requested. • If the cores within a compute unit request different P-states while in C0, frequency and voltage are deter- 53 42300 Rev 3.12 - July 14, 2015 BKDG for AMD Family 15h Models 10h-1Fh Processors mined by the highest-performance P-state requested. • If one core within a compute unit requests a C-state while the other core is in C0, the frequency and voltage of the compute unit is determined by the core in C0. • If both cores request non-C0 states, the behavior is specified by D18F4x128[CoreCstatePolicy]. 2.5.2.2 Dependencies Between Subcomponents on VDDNB Many subcomponents of the processor including the NB, the GPU, and the root complex reside on the VDDNB power plane. Hardware must take voltage domain dependencies into account when determining whether to make a voltage change requested by one of the subcomponents. Whenever a state transition occurs that causes a voltage change request (see 2.5.1.4.1 [Hardware-Initiated Voltage Transitions]), or software makes a voltage change request (see 2.5.1.4.2 [Software-Initiated Voltage Transitions]), the VDDNB voltage requested by the processor is determined by the highest voltage (lowest VID) request made by any of the subcomponents or by software. In addition, software can force VDDNB voltage changes. See 2.5.1.4.2 [SoftwareInitiated Voltage Transitions]. 2.5.2.3 BIOS Requirements for Power Plane Initialization • Ensure the following fields are configured to their BIOS recommendations: • D18F3xA0[Svi2HighFreqSel]. • D18F3xD8[VSRampSlamTime]. 2.5.3 2.5.3.1 CPU Power Management Core P-states Core P-states are operational performance states characterized by a unique combination of core frequency and voltage. The processor supports up to 8 core P-states (P0 through P7), specified in MSRC001_00[6B:64]. Out of cold reset, the voltage and frequency of the compute units is specified by MSRC001_0071[StartupPstate]. Support for dynamic core P-state changes is indicated by more than one enabled selection in MSRC001_00[6B:64][PstateEn]. At least one enabled P-state (P0) is specified for all processors. Software requests core P-state changes for each core independently using the hardware P-state control mechanism (a.k.a. fire and forget). Support for hardware P-state control is indicated by CPUID Fn8000_0007_EDX[HwPstate]=1b. Software may not request any P-state transitions using the hardware Pstate control mechanism until the P-state initialization requirements defined in 2.5.3.1.7 [BIOS Requirements for Core P-state Initialization and Transitions] are complete. 2.5.3.1.1 Application Power Management (APM) Application Power Management (APM) allows the processor to deterministically provide maximum performance while remaining within the specified power delivery and removal envelope. APM dynamically monitors processor activity and generates an approximation of power consumption. If power consumption exceeds a defined power limit, a P-state limit is applied by APM hardware to reduce power consumption. APM ensures that average power consumption over a thermally significant time period remains at or below the defined power limit. This allows P-states to be defined with higher frequencies and voltages than could be used without APM. These P-states are referred to as boosted P-states. • Support for APMis specified by CPUID Fn8000_0007_EDX[CPB]. • APM is enabled if all of the following conditions are true: • MSRC001_0015[CpbDis] = 0 for all cores. 54 42300 Rev 3.12 - July 14, 2015 • • • • • • • BKDG for AMD Family 15h Models 10h-1Fh Processors • D18F4x15C[ApmMasterEn] = 1. • D18F4x15C[BoostSrc] = 01b. • D18F4x15C[NumBoostStates] != 0. APM can be dynamically enabled and disabled through MSRC001_0015[CpbDis]. If core performance boost (CPB) is disabled , a P-state limit is applied . The P-state limit restricts cores to the highest performance non-boosted P-state. All P-states, both boosted and non-boosted, are specified in MSRC001_00[6B:64]. The number of boosted P-states is specified by D18F4x15C[NumBoostStates]. • The number of boosted P-states may vary from product to product. There are two stages of boost supported. Compute units can be placed in the first stage of boosted P-states if the processor power consumption remains within the TDP limit. The second stage of boosted P-states can only be achieved if a subset of compute units are in CC6 and the processor power consumption remains within the TDP limit. See D18F4x16C[CstateCnt, CstateBoost]. All boosted P-states are always higher performance than non-boosted P-states. To ensure proper operation, boosted P-states should be hidden from the operating system. BIOS should not provide ACPI _PSS entries for boosted P-states. See 2.5.3.1.9.3.2 [_PSS (Performance Supported States)]. The lowest-performance P-state CPB limits the processor to is the highest-performance non-boosted P-state. 2.5.3.1.2 Core P-state Naming and Numbering Since the number of boosted P-states may vary from product to product, the mapping between MSRC001_00[6B:64] and the indices used to request P-state changes or status also varies. In order to clarify this, two different numbering schemes are used. 2.5.3.1.2.1 Software P-state Numbering When referring to software P-state numbering, the following naming convention is used: • Non-boosted P-states are referred to as P0, P1, etc. • P0 is the highest power, highest performance, non-boosted P-state. • Each ascending P-state number represents a lower-power, lower performance non-boosted P-state than the prior P-state number. • Boosted P-states are referred to as Pb0, Pb1, etc. • Pb0 is the highest-performance, highest-power boosted P-state. • Each higher numbered boosted P-state represents a lower-power, lower-performance boosted P-state. For example, if D18F4x15C[NumBoostStates] contains the values shown below, then the P-states would be named as follows: Table 8: Software P-state Naming D18F4x15C[NumBoostD18F4x15C[NumBoostStates]=1 States]=3 P-state Name Corresponding P-state Name Corresponding MSR Address MSR Address Pb0 MSRC001_0064 Pb0 MSRC001_0064 P0 MSRC001_0065 Pb1 MSRC001_0065 P1 MSRC001_0066 Pb2 MSRC001_0066 P2 MSRC001_0067 P0 MSRC001_0067 P3 MSRC001_0068 P1 MSRC001_0068 55 42300 Rev 3.12 - July 14, 2015 BKDG for AMD Family 15h Models 10h-1Fh Processors Table 8: Software P-state Naming D18F4x15C[NumBoostD18F4x15C[NumBoostStates]=1 States]=3 P-state Name Corresponding P-state Name Corresponding MSR Address MSR Address P4 MSRC001_0069 P2 MSRC001_0069 P5 MSRC001_006A P3 MSRC001_006A P6 MSRC001_006B P4 MSRC001_006B All sections and register definitions use software P-state numbering unless otherwise specified. 2.5.3.1.2.2 Hardware P-state Numbering When referring to hardware P-state numbering, the following naming convention is used: • All P-states are referred to as P0, P1, etc. • P0 is the highest power, highest-performance P-state, regardless of whether it is a boosted P-state or a non-boosted P-state. • Each ascending P-state number represents a lower-power, lower-performance P-state, regardless of whether it is a boosted P-state or not. 2.5.3.1.3 Core P-state Control Core P-states are dynamically controlled by software and are exposed through ACPI objects (refer to 2.5.3.1.9.3 [ACPI Processor P-state Objects]). Software requests a core P-state change by writing a 3 bit index corresponding to the desired P-state number to MSRC001_0062[PstateCmd] of the appropriate core. For example, to request P3 for core 0 software would write 011b to core 0’s MSRC001_0062[PstateCmd]. Boosted P-states may not be directly requested by software. Whenever software requests the P0 state on a processor that supports CPB (i.e. writes 000b to MSRC001_0062[PstateCmd]), hardware dynamically places the core into the highest-performance P-state possible as determined by CPB. See 2.5.3.1.1 [Application Power Management (APM)]. Hardware sequences the frequency and voltage changes necessary to complete a P-state transition as specified by 2.5.3.1.6 [Core P-state Transition Behavior] with no additional software interaction required. Hardware also coordinates frequency and voltage changes when differing P-state requests are made on cores that share a frequency or voltage plane. See 2.5.2 [Frequency and Voltage Domain Dependencies] for details about hardware coordination. Table 9: Software P-state Control D18F4x15C[NumBoostStates]=1 D18F4x15C[NumBoostStates]=3 P-state Name Index Used for Corresponding P-state Name Index Used for Corresponding Requests/Status MSR Address Requests/Status MSR Address Pb0 n/a MSRC001_0064 Pb0 n/a MSRC001_0064 P0 0 MSRC001_0065 Pb1 n/a MSRC001_0065 P1 1 MSRC001_0066 Pb2 n/a MSRC001_0066 P2 2 MSRC001_0067 P0 0 MSRC001_0067 P3 3 MSRC001_0068 P1 1 MSRC001_0068 56 42300 Rev 3.12 - July 14, 2015 BKDG for AMD Family 15h Models 10h-1Fh Processors Table 9: Software P-state Control D18F4x15C[NumBoostStates]=1 D18F4x15C[NumBoostStates]=3 P-state Name Index Used for Corresponding P-state Name Index Used for Corresponding Requests/Status MSR Address Requests/Status MSR Address P4 4 MSRC001_0069 P2 2 MSRC001_0069 P5 5 MSRC001_006A P3 3 MSRC001_006A P6 6 MSRC001_006B P4 4 MSRC001_006B Hardware controls the VID for each voltage domain according to the highest requirement of the frequency domain(s) on each plane. For example, the VID for a 4 compute unit dual-plane system must be maintained at the highest level required for all 4 frequency domains. The number of frequency domains in a voltage domain is package/platform specific. Refer to 2.5.3.1.6 [Core P-state Transition Behavior] for details on hardware Pstate voltage control. 2.5.2.3 [BIOS Requirements for Power Plane Initialization] specifies the processor initialization requirements for voltage plane control. 2.5.3.1.4 Core P-state Visibility MSRC001_0063 [P-state Status][CurPstate] reflects the current non-boosted P-state number for each compute unit. For example, if MSRC001_0063[CurPstate]=010b on compute unit 1, then compute unit 1 is in the P2 state. If a compute unit is in a boosted P-state, MSRC001_0063[CurPstate] reads back as 0. The voltage on a compute unit may not correspond to the VID code specified by the current P-state of the compute unit due to voltage plane dependencies. See 2.5.2 [Frequency and Voltage Domain Dependencies]. If a compute unit is in the P0 state (i.e. if MSRC001_0063[CurPstate]=0), the frequency of the compute unit could be the frequency specified by P0 or any boosted P-state. To determine the frequency of a compute unit, see 2.5.3.3 [Effective Frequency]. 2.5.3.1.5 Core P-state Limits Core P-states may be limited to lower-performance values under certain conditions, including: • HTC. See D18F3x64 [Hardware Thermal Control (HTC)][HtcPstateLimit]. • Software. See D18F3x68[SwPstateLimit]. • Core Performance Boost. See 2.5.3.1.1 [Application Power Management (APM)]. • PROCHOT_L assertion. See 2.10.4.1 [PROCHOT_L and Hardware Thermal Control (HTC)]. • SMU. See D18F4x13C[SmuPstateLimit]. P-state limits are applied to all cores on the processor. The current P-state limit is provided in MSRC001_0061 [P-state Current Limit][CurPstateLimit]. Changes to the MSRC001_0061[CurPstateLimit] can be programmed to trigger interrupts through D18F3x64[PslApicLoEn and PslApicHiEn] In addition, the maximum P-state value, regardless of the source, is limited as specified in MSRC001_0061[PstateMaxVal]. 2.5.3.1.6 Core P-state Transition Behavior The following rules specify how P-states changes function and interact with other system or processor states: • If the P-state number is increasing (the compute unit is moving to a lower-performance state), then the COF is changed first, followed by the VID change. If the P-state number is decreasing, then the VID is changed first followed by the COF. • When the processor initiates a VID change that increases voltage for a voltage domain, no new voltage or frequency changes occur until D18F3xD8[VSRampSlamTime] has expired, regardless of whether any new 57 42300 Rev 3.12 - July 14, 2015 • • • • BKDG for AMD Family 15h Models 10h-1Fh Processors requests are received. When the processor initiates a VID change that decreases voltage for a voltage domain, new voltage or frequency changes are allowed to occur immediately. • This is true regardless of whether the frequency or voltages changes occur as a result of P-state or C-state changes. If multiple commands are issued that affect the P-state of a domain prior to when the processor initiates the change of the P-state of that domain, then the processor operates on the last one issued. Once a P-state change starts, the P-state state machine (PSSM) continues through completion unless interrupted by a PWROK deassertion. If multiple P-state changes are requested concurrently, the PSSM may group the associated VID changes separately from the associated COF changes. Behavior during RESET_L assertions: • All compute units are transitioned to C0. • If there is no P-state transition activity, then the compute units and NB remain in the current P-state. If a RESET_L assertion interrupts a P-state transition, then the COF remains in it’s current state at the time RESET_L is asserted (either the value of the old or the new P-state) and the VID remains in it’s current state (perhaps at a VID between the old and the new P-states, if the VID was being stepped). BIOS is required to transition to valid COF and VID settings after a warm reset according to the sequence defined in 2.5.3.1.9 [BIOS COF and VID Requirements After Warm Reset]. • After a warm reset MSRC001_0063 [P-state Status] is consistent with MSRC001_0071[CurPstate]. MSRC001_0062 [P-state Control] may not be consistent with MSRC001_0071[CurPstate]. • If D18F5x1[6C:60][NbFid] has changed, then the new value is applied to the NB PLL on the assertion of RESET_L. It is assumed that BIOS adjusts the NB VID to the appropriate value prior to the warm reset. See 2.5.1.4.2 [Software-Initiated Voltage Transitions]. The OS controls the P-state through MSRC001_0062 [P-state Control], independent of P-state limits described in D18F3x64 [Hardware Thermal Control (HTC)][HtcPstateLimit], D18F3x68 [Software P-state Limit][StcPstateLimit]. P-state limits interact with OS-directed P-state transitions as follows: • Of all the active P-state limits, the one that represents the lowest-performance P-state number, at any given time, is treated as an upper limit on performance. • As the limit becomes active or inactive, or if it changes, the P-state for each compute unit is placed in either the last OS-requested P-state or the new limit P-state, whichever is a lower performance P-state number. • If the resulting P-state number exceeds MSRC001_0061 [P-state Current Limit][PstateMaxVal], regardless of whether it is a limit or OS-requested, then the PstateMaxVal is used instead. 2.5.3.1.7 BIOS Requirements for Core P-state Initialization and Transitions 1. Check that CPUID Fn8000_0007_EDX[HwPstate]=1. If not, P-states are not supported and BIOS skips the rest of these steps. 2. Complete the 2.5.2.3 [BIOS Requirements for Power Plane Initialization]. 3. Ensure the following fields are configured to their BIOS recommendations: • D18F3xA0[PllLockTime]. • D18F3xD4[PowerStepUp, PowerStepDown]. 4. Transition all cores to the minimum performance P-state using the algorithm detailed in 2.5.3.1.9.2 [Core Minimum P-state Transition Sequence After Warm Reset]. 5. Complete the 2.5.4.1.2.1 [NB P-state COF and VID Synchronization After Warm Reset]. All cores on a processor must be in the minimum performance P-state prior to executing this sequence. 6. Complete the 2.5.3.1.8 [Processor-Systemboard Power Delivery Compatibility Check]. 7. Perform the following steps in any order: A. Enable 2.5.3.1.1 [Application Power Management (APM)] as follows: • Ensure the following fields are configured to their BIOS recommendations: • D18F4x110[CSampleTimer]. • D18F4x15C[ApmMasterEn]. 58 42300 Rev 3.12 - July 14, 2015 BKDG for AMD Family 15h Models 10h-1Fh Processors • D18F5xE0[RunAvgRange]. • See your AMD representative for details on how to enable the GPU aspects of 2.5.3.1.1 [Application Power Management (APM)]. • If D18F4x15C[NumBoostStates]!=0, program D18F4x15C[BoostSrc]=1. B. Transition all cores to the maximum performance P-state by writing 0 to MSRC001_0062[PstateCmd]. C. Create ACPI objects if neccessary: • Determine the valid set of P-states as indicated byMSRC001_00[6B:64][PstateEn]. • If P-states are not supported, as indicated by only one enabled selection in MSRC001_00[6B:64][PstateEn], then BIOS must not generate ACPI-defined P-state objects described in 2.5.3.1.9.3 [ACPI Processor P-state Objects]. Otherwise, the ACPI objects should be generated to enable P-state support. D. Configure the COF and VID for each processor appropriately based on the sequence described in 2.5.4.1.2 [BIOS NB P-state Configuration]. 8. Configure PSIx_L. Refer to 2.5.1.3.1 [PSIx_L Bit] for additional details. 2.5.3.1.8 Processor-Systemboard Power Delivery Compatibility Check BIOS must disable processor P-states that require higher power delivery than the systemboard can support. This power delivery compatibility check is designed to prevent system failures caused by exceeding the power delivery capability of the systemboard for the power plane(s) that contain the core(s). Refer to 2.5.1 [Processor Power Planes And Voltage Control] for power plane definitions and configuration information. BIOS can optionally notify the user if P-states are detected that exceed the systemboard power delivery capability. Modifications to MSRC001_00[6B:64] [P-state [7:0]] must be applied equally to all cores on the same node. This check does not ensure functionality for all package/socket compatible processor/systemboard combinations. MSRC001_00[6B:64][PstateEn] must be set to 0 for any P-state MSR where PstateEn=1 and the processor current requirement (ProcIddMax), defined by the following equation, is greater than the systemboard current delivery capability.ProcIddMax = MSRC001_00[6B:64][IddValue] current * 1/10^MSRC001_00[6B:64][IddDiv] * (D18F5x84[CmpCap]+1) ; The power delivery check should be applied starting with hardware P0 and continue with increasing P-state indexes (1, 2, 3, and 4) for all enabled P-states. Once a compatible P-state is found using the ProcIddMax equation the check is complete. All processor P-states with higher indexes are defined to be lower power and performance, and are therefore compatible with the systemboard. Example: • MSRC001_0065[IddValue] = 32d • MSRC001_0065[IddDiv] = 0d • D18F5x84[CmpCap] = 1d • ProcIddMax = 32 * 1 * 2 = 64A per plane The systemboard must be able to supply >= 64A for the unified core power plane in order to support P1 for this processor. If the systemboard current delivery capability is < 64A per plane then BIOS must set MSRC001_0065[PstateEn]=0 for all cores on this node, and continue by checking P2 in the same fashion. If no P-states are disabled while performing the power delivery compatibility check then BIOS does not need to take any action. If at least one P-state is disabled by performing the power delivery compatibility check and at least one P-state remains enabled, then BIOS must perform the following steps: 1. If the P-state pointed to by MSRC001_0063[CurPstate] is disabled by the power delivery compatibility 59 42300 Rev 3.12 - July 14, 2015 2. 3. 4. 5. 6. BKDG for AMD Family 15h Models 10h-1Fh Processors check, then BIOS must request a transition to an enabled P-state using MSRC001_0062[PstateCmd] and wait for MSRC001_0063[CurPstate] to reflect the new value. Copy the contents of the enabled P-state MSRs (MSRC001_00[6B:64]) to the highest performance P-state locations. E.g. if P0 and P1 are disabled by the power delivery compatibility check and P2 - P4 remain enabled, then the contents of P2 - P4 should be copied to P0 - P2 and P3 and P4 should be disabled (PstateEn=0). This step uses software P-state numbering. See 2.5.3.1.2.1 [Software P-state Numbering]. Request a P-state transition to the P-state MSR containing the COF/VID values currently applied. E.g. If MSRC001_0063[CurPstate]=100b and P4 P-state MSR information is copied to P2 in step 2, then BIOS should write 010b to MSRC001_0062[PstateCmd] and wait for MSRC001_0063[CurPstate] to reflect the new value. If a subset of boosted P-states are disabled, then copy the contents of the P-state MSR pointed to by the highest performance boosted P-state that is enabled to the P-state MSRs pointed to by the boosted P-states that are disabled. If all boosted P-states are disabled, then program D18F4x15C[BoostSrc]=0. Adjust the following P-state parameters affected by the P-state MSR copy by subtracting the number of software P-states that are disabled by the power delivery compatibility check. This calculation should not wrap, but saturate at 0. E.g. if P0 and P1 are disabled, then each of the following register fields should have 2 subtracted from them: • D18F3x64[HtcPstateLimit] • D18F3x68[SwPstateLimit] • D18F3xDC[HwPstateMaxVal] If any node has all P-states disabled after performing the power delivery compatibility check, then BIOS must perform the following steps. This does not ensure operation and BIOS should notify the user of the incompatibility between the processor and systemboard if possible. 1. If MSRC001_0063[CurPstate]!=MSRC001_0061[PstateMaxVal], then write MSRC001_0061[PstateMaxVal] to MSRC001_0062[PstateCmd] and wait for MSRC001_0063[CurPstate] to reflect the new value. 2. If MSRC001_0061[PstateMaxVal]!=000b copy the contents of the P-state MSR pointed to by MSRC001_0061[PstateMaxVal] to MSRC001_0064 and set MSRC001_0064[PstateEn]; Write 000b to MSRC001_0062[PstateCmd] and wait for MSRC001_0063[CurPstate] to reflect the new value. This step uses software P-state numbering. See 2.5.3.1.2.1 [Software P-state Numbering]. 3. Adjust the following fields to 000b. • D18F3x64[HtcPstateLimit] • D18F3x68[SwPstateLimit] • D18F3xDC[HwPstateMaxVal] 4. Program D18F4x15C[BoostSrc]=0. 2.5.3.1.9 BIOS COF and VID Requirements After Warm Reset Warm reset is asynchronous and can interrupt P-state transitions leaving the processor in a VID state that does not correspond to MSRC001_0063[CurPstate] on any core. The processor frequency after warm reset corresponds to MSRC001_0063[CurPstate]. See 2.5.3.1.6 [Core P-state Transition Behavior] for P-state transition behavior when RESET_L is asserted. BIOS is required to transition the processor to valid COF and VID settings corresponding to an enabled P-state following warm reset. The cores may be transitioned to either the maximum or minimum P-state COF and VID settings using the sequences defined in 2.5.3.1.9.1 [Core Maximum P-state Transition Sequence After Warm Reset] and 2.5.3.1.9.2 [Core Minimum P-state Transition Sequence After Warm Reset]. Transitioning to the minimum P-state after warm reset is recommended to prevent undesired system behavior if a warm reset occurs before the 2.5.3.1.8 [Processor-Systemboard Power Delivery Compatibility Check] is complete. BIOS is not required to manipulate NB COF and VID settings following warm reset if the warm reset was issued by BIOS to update D18F5x1[6C:60][NbFid]. 60 42300 Rev 3.12 - July 14, 2015 2.5.3.1.9.1 BKDG for AMD Family 15h Models 10h-1Fh Processors Core Maximum P-state Transition Sequence After Warm Reset 1. Write MSRC001_0061[PstateMaxVal] to MSRC001_0062[PstateCmd] on all cores in the processor. 2. Wait for MSRC001_0071[CurCpuFid, CurCpuDid] = [CpuFid, CpuDid] from MSRC001_00[6B:64] indexed by D18F3xDC[HwPstateMaxVal]. 3. Step 2 must be completed on all cores prior to executing step 4 since a compute unit transitions to the highest performance P-state requested on either core. 4. Write 0 to MSRC001_0062[PstateCmd] on all cores in the processor. 5. Wait for MSRC001_0071[CurCpuFid, CurCpuDid] = [CpuFid, CpuDid] from MSRC001_00[6B:64] indexed by MSRC001_0071[CurPstateLimit]. 6. If MSRC001_0071[CurPstateLimit] != D18F3xDC[HwPstateMaxVal], wait for MSRC001_0071[CurCpuVid] = [CpuVid] from MSRC001_00[6B:64] indexed by MSRC001_0071[CurPstateLimit]. 7. Wait for MSRC001_0063[CurPstate] = MSRC001_0061[CurPstateLimit]. 2.5.3.1.9.2 Core Minimum P-state Transition Sequence After Warm Reset 1. Write 0 to MSRC001_0062[PstateCmd] on all cores in the processor. 2. Wait for MSRC001_0071[CurCpuFid, CurCpuDid] = [CpuFid, CpuDid] from MSRC001_00[6B:64] indexed by MSRC001_0071[CurPstateLimit]. 3. Write MSRC001_0061[PstateMaxVal] to MSRC001_0062[PstateCmd] on all cores in the processor. 4. Wait for MSRC001_0071[CurCpuFid, CurCpuDid] = [CpuFid, CpuDid] from MSRC001_00[6B:64] indexed by D18F3xDC[HwPstateMaxVal]. 5. If MSRC001_0071[CurPstateLimit] != MSRC001_0071[CurPstate], wait for MSRC001_0071[CurCpuVid] = [CpuVid] from MSRC001_00[6B:64] indexed by D18F3xDC[HwPstateMaxVal]. 6. Wait for MSRC001_0063[CurPstate] = MSRC001_0062[PstateCmd]. 2.5.3.1.9.3 ACPI Processor P-state Objects Processor performance control is implemented through the _PCT, _PSS and _PSD objects in ACPI 2.0 and later revisions. The presence of these objects indicates to the OS that the platform and processor are capable of supporting multiple performance states. Processor performance states are not supported with ACPI 1.0b. BIOS must provide the _PCT, _PSS, and _PSD objects, and define other ACPI parameters to support operating systems that provide native support for processor P-state transitions. The following rules apply to BIOS generated ACPI objects in multi-core systems. Refer to the appropriate ACPI specification (http://www.acpi.info) for additional details: • All cores must expose the same number of performance states to the OS. • The respective performance states displayed to the OS for each core must have identical performance and power-consumption parameters (e.g. P0 on core 0 must have the same performance and power-consumptions parameters as P0 on core 1, P1 on core 0 must have the same parameters as P1 on core 1, however P0 can be different than P1). • Performance state objects must be present under each processor object in the system. 2.5.3.1.9.3.1 _PCT (Performance Control) BIOS must declare the performance control object parameters as functional fixed hardware. This definition indicates the processor driver understands the architectural definition of the P-state interface associated with CPUID Fn8000_0007_EDX[HwPstate]=1. • Perf_Ctrl_Register = Functional Fixed Hardware • Perf_Status_Register = Functional Fixed Hardware 61 42300 Rev 3.12 - July 14, 2015 2.5.3.1.9.3.2 BKDG for AMD Family 15h Models 10h-1Fh Processors _PSS (Performance Supported States) A unique _PSS entry is created for each non-boosted P-state. The value contained in the _PSS Control field is written to MSRC001_0062 [P-state Control] to request a P-state change to the CoreFreq of the associated _PSS object. The value contained in MSRC001_0063 [P-state Status] can be used to identify the _PSS object of the current P-state request by equating MSRC001_0063[CurPstate] to the value of the Status field. See 2.5.3.1 [Core P-states]. BIOS loops through each of MSRC001_00[6B:64] applying the following formulas to create the fields for the _PSS object for for each valid P-state (see MSRC001_00[6B:64][PstateEn]). BIOS skips over any P-state MSRs that specify boost P-states (see D18F4x15C[NumBoostStates]). • CoreFreq (MHz) = Calculated using the formula for CoreCOF. • Power (mW) = MSRC001_00[6B:64][CpuVid] voltage * MSRC001_00[6B:64][IddValue] current * 1000. • TransitionLatency (us) and BusMasterLatency (us): • If MSRC001_00[6B:64][CpuFid] is the same for all enabled P-states (see MSRC001_00[6B:64][PstateEn]) and all boosted P-states: • TransitionLatency = BusMasterLatency = (15 steps * D18F3xD4[PowerStepDown] time * 1000 us/ns) + (15 steps * D18F3xD4[PowerStepUp] time * 1000 us/ns) • Else if MSRC001_00[6B:64][CpuFid] is different for any enabled (see MSRC001_00[6B:64][PstateEn]) or boost P-states: • TransitionLatency = BusMasterLatency = (15 steps * D18F3xD4[PowerStepDown] time * 1000 us/ns) + D18F3xA0[PllLockTime] time + (15 steps * D18F3xD4[PowerStepUp] time * 1000 us/ns) • Example: • MSRC001_00[6B:64][CpuFid] is not the same for all P-states • D18F3xD4[PowerStepDown] = D18F3xD4[PowerStepUp] = 8h (50 ns/step) • D18F3xA0[PllLockTime] = 001b (2 us) • TransitionLatency = BusMasterLatency = (15 steps * 50 ns/step / 1000 us/ns) + 2us + (15 steps * 50 ns/step / 1000 us/ns) = 3.5 us (round up to 4 us) • Control/Status: • The highest performance non-boosted P-state must have the _PSS control and status fields programmed to 0. • Any lower performance non-boosted P-states must have the _PSS control and status fields programmed in ascending order. 2.5.3.1.9.3.3 _PPC (Performance Present Capabilities) The _PPC object is optional. Refer to the ACPI specification for details on use and content. 2.5.3.1.9.3.4 _PSD (P-state Dependency) AMD recommends the ACPI 3.0 _PSD object be generated for each core as follows to cause the cores to transition between P-states together: • • • • • NumberOfEntries = 5. Revision = 0. Domain = 0. CoordType = FCh. (SW_ALL) NumProcessors = CPUID Fn8000_0008_ECX[NC] + 1. 62 42300 Rev 3.12 - July 14, 2015 BKDG for AMD Family 15h Models 10h-1Fh Processors A vendor may optionally choose to generate _PSD object to allow cores to transition between P-states independently as follows: • • • • • NumberOfEntries = 5. Revision = 0. Domain = CPUID Fn0000_0001_EBX[LocalApicId[7:1]]. CoordType = FEh. (HW_ALL) NumProcessors = 2. BIOS provides an option to choose between either _PSD definition. 2.5.3.1.9.4 Fixed ACPI Description Table (FADT) Entries Declare the following FADT entries: • PSTATE_CNT = 00h. • DUTY_WIDTH = 00h. 2.5.3.1.9.5 XPSS (Microsoft® Extended PSS) Object Some Microsoft® operating systems require an XPSS object to make P-state changes function properly. A BIOS that implements an XPSS object has special requirements for the _PCT object. See the Microsoft Extended PSS ACPI Method Specification for the detailed requirements to implement these objects. 2.5.3.2 Core C-states C-states are processor power states. C0 is the operational state in which instructions are executed. All other Cstates are low-power states in which instructions are not executed. When coming out of warm and cold reset, the processor is transitioned to the C0 state. 2.5.3.2.1 C-state Names and Numbers C-states are often referred to by an alphanumeric naming convention, C1, C2, C3, etc. The mapping between ACPI defined C-states and AMD specified C-states is not direct. AMD specified C-states are referred to as IObased C-states. Up to three IO-based C-states are supported, IO-based C-state 0, 1, and 2. The IO-based C-state index corresponds to the offset added to MSRC001_0073[CstateAddr] to initiate a C-state request. See 2.5.3.2.2 [C-state Request Interface]. The actions taken by the processor when entering a low-power C-state are configured by software. See 2.5.3.2.3 [C-state Actions] for information about AMD specific actions. 2.5.3.2.2 C-state Request Interface C-states are dynamically requested by software and are exposed through ACPI objects (see 2.5.3.2.6 [ACPI Processor C-state Objects]). C-states can be requested on a per-core basis. Software requests a C-state change in one of two ways: • Reading from an IO address: The IO address must be the address specified by MSRC001_0073[CstateAddr] plus an offset of 0 through 7. The processor always returns 0 for this IO read. Offsets 2 through 7 result in an offset of 2. • Executing the HLT instruction. This is equivalent to reading from the IO address specified by D18F4x128[HaltCstateIndex]. When software requests a C-state transition, hardware evaluates any frequency and voltage domain dependencies and determines which C-state actions to execute. See 2.5.2 [Frequency and Voltage Domain Dependen- 63 42300 Rev 3.12 - July 14, 2015 BKDG for AMD Family 15h Models 10h-1Fh Processors cies] and 2.5.3.2.3 [C-state Actions]. 2.5.3.2.3 C-state Actions A core takes one of several different possible actions based upon a C-state change request from software. The C-state action fields are defined in D18F4x11[C:8] and D18F4x11C. 2.5.3.2.3.1 C-state Probes and Cache Flushing If probes occur after a core enters a non-C0 state, and the caches are not flushed by hardware, the core clock may be ramped back up to the C0 frequency to service the probes, as specified by D18F3x[84:80][CpuPrbEn] or D18F4x11[C:8]/D18F4x11C[*CpuPrbEn]. If a core enters a non-C0 state and cache flush is enabled (see D18F3xDC[CacheFlushOnHaltCtl] and D18F4x11[C:8][CacheFlushEn]), a timer counts down for a programmable period of time as specified by D18F3xDC[CacheFlushOnHaltTmr] or D18F4x11[C:8][CacheFlushTmrSel]. When the timer expires, the core flushes its L1 and L2 caches to DRAM and the core clocks are ramped down to a divisor specified by D18F3xDC[CacheFlushOnHaltCtl]. The timer is reset if the core exits the C-state for any reason. See 2.5.3.2.4.3 [Cache Flush On Halt Saturation Counter]. Once a core flushes its caches, probes are no longer sent to that core. This improves probing performance for cores that are in C0. 2.5.3.2.3.2 Core C1 (CC1) State When a core enters the CC1 state, its clock ramps down to the frequency specified by D18F4x11[C:8][ClkDivisorCstAct]. 2.5.3.2.3.3 Core C6 (CC6) State A core can gate off power to its internal logic when it enters any non-C0 state. This power gated state is known as CC6. In order to enter CC6, hardware first enters CC1 and then flushes the caches (see 2.5.3.2.3.1 [C-state Probes and Cache Flushing]) before checking D18F4x11[C:8][PwrGateEnCstAct]. Power gating reduces the amount of power consumed by the core. VDD voltage is not reduced when a core is in CC6. The following sequence occurs when a core enters the CC6 state: 1. If MSRC001_0071[CurPstate] < D18F3xA8[PopDownPstate], transition the core P-state to D18F3xA8[PopDownPstate]. 2. L1 and L2 caches are flushed to DRAM. See 2.5.3.2.3.1 [C-state Probes and Cache Flushing]. 3. Internal core state is saved to DRAM. 4. Power is removed from the core and the core PLL/voltage regulator is powered down as specified by D18F5x128[CC6PwrDwnRegEn]. All of the following must be true in order for a core to be placed into CC6: • D18F4x11[C:8][CacheFlushEn]=1 for the corresponding C-state action field. • D18F4x11[C:8][CacheFlushTmrSel] != 11b for the corresponding C-state action field. • D18F4x11[C:8][PwrGateEnCstAct]=1 for the corresponding C-state action field. • D18F2x118[CC6SaveEn]=1. • D18F2x118[LockDramCfg]=1 . • The CC6 storage area in DRAM is configured. See 2.9.8 [DRAM CC6/PC6 Storage]. 64 42300 Rev 3.12 - July 14, 2015 BKDG for AMD Family 15h Models 10h-1Fh Processors The events which cause a core to exit the CC6 state are specified in 2.5.3.2.5 [Exiting C-states]. If a warm reset occurs while a core is in CC6, all MCA registers in the core shown in Table 56 are cleared to 0. See 2.15.1[Machine Check Architecture]. The time required to enter and exit CC6 is directly proportional to the core P-state frequency. Slower core frequencies require longer entry and exit times. Latency issues may occur with core P-state frequencies less than 800MHz. 2.5.3.2.3.4 Package C6 (PC6) State When all cores enter a non-C0 state, VDD can be reduced to a non-operational voltage that does not retain core state. This state is known as PC6 and reduces the amount of static and dynamic power consumed by all cores. The following actions are taken by hardware prior to PC6 entry: 1. If MSRC001_0071[CurPstate] < D18F3xA8[PopDownPstate], transition the core P-state to D18F3xA8[PopDownPstate]. 2. For all cores not in CC6, L1 and L2 caches are flushed to DRAM. See 2.5.3.2.3.1 [C-state Probes and Cache Flushing]. 3. For all cores not in CC6, internal core state is saved to DRAM. 4. VDD is transitioned to the VID specified by D18F5x128[PC6Vid]. 5. If the core PLLs are not powered down during CC6 entry (see 2.5.3.2.3.3 [Core C6 (CC6) State]), then they are powered down as specified by D18F5x128[PC6PwrDwnRegEn]. All of the following must be true on all cores in order for a package to be placed into PC6: • D18F4x11[C:8][CacheFlushEn]=1 for the corresponding C-state action field • D18F4x11[C:8][CacheFlushTmrSel] != 11b for the corresponding C-state action field. • D18F4x11[C:8][PwrOffEnCstAct]=1 for the corresponding C-state action field. • D18F2x118[CC6SaveEn]=1. • D18F2x118[LockDramCfg]=1. • MSRC001_0015[HltXSpCycEn]=1. • Timer Tick Monitor does not disallow PC6 entry. See 2.5.3.2.4.2 [Timer Tick Monitor]. 2.5.3.2.4 C-state Request Monitors Deeper C-states have higher entry and exit latencies but provide greater power savings than shallower C-states. To help balance the performance and power needs of the system, the processor can limit access to specific Cstates in certain scenarios. 2.5.3.2.4.1 FCH Messaging The FCH can be notified when the processor transitions package C-states. See the following: • D18F4x128[CstateMsgDis]. • D18F5x178[CstateFusionDis]. • MSRC001_0015[HltXSpCycEn]. If the processor sends a message when entering PC6, the FCH sends a return message notifying the processor whether PC6 is allowed. The processor can ignore the FCH response as specified by D18F5x178[CstateFusionHsDis]. A C-state message indicating the package C-state entered is sent to the FCH as specified by D18F5x178[CstateThreeWayHsEn]. 65 42300 Rev 3.12 - July 14, 2015 2.5.3.2.4.2 BKDG for AMD Family 15h Models 10h-1Fh Processors Timer Tick Monitor The timer tick monitor (enabled using D18F4x124[IntMonPC6En]) tracks interrupts that are delivered from the FCH on a regular interval, generally referred to as timer tick interrupts. When software changes the period of any timers in the FCH, the FCH reports the new period to the processor. The timer tick monitor cannot be used to track periodic local APIC timer interrupts. The monitor compares the timer tick period to the threshold defined by D18F4x124[IntMonPC6Limit]. If the period is less than or equal to the threshold, access to PC6 is disallowed. 2.5.3.2.4.3 Cache Flush On Halt Saturation Counter A cache flush success monitor tracks the success rate of cache flush timer expirations relative to the core exiting a C-state. Based on the success rate, caches may be flushed immediately without waiting for the cache flush timer to expire. See D18F4x128[CacheFlushSucMonThreshold]. When the core resumes normal execution, the caches refill as normal. 2.5.3.2.5 Exiting C-states The following events may cause the processor to exit a non-C0 C-state and return to C0: • INTR • NMI • SMI • INIT • RESET_L assertion If an INTR is received while a core is in a non-C0 C-state, the state of EFLAGS[IF] and the mechanism used to enter the non-C0 C-state determine the actions taken by the processor. • Entry via HLT, EFLAGS[IF]==0: The interrupt does not wake up the core. • Entry via HLT, EFLAGS[IF]==1: The interrupt wakes the core and the core begins execution at the interrupt service routine. • Entry via IO read, EFLAGS[IF]==0: The interrupt wakes the core and the core begins execution at the instruction after the IN instruction that was used to enter the non-C0 C-state. • Entry via IO read, EFLAGS[IF]==1: The interrupt wakes the core and the core begins execution at the interrupt service routine. 2.5.3.2.6 ACPI Processor C-state Objects Processor power control is implemented through the _CST object in ACPI 2.0 and later revisions. The presence of the _CST object indicates to the OS that the platform and processor are capable of supporting multiple power states. BIOS must provide the _CST object and define other ACPI parameters to support operating systems that provide native support for processor C-state transitions. See 2.5.3.2.6.1 [_CST] and 2.5.3.2.6.2 [_CSD]. The _CST object is not supported with ACPI 1.0b. BIOS should provide FADT entries to support operating systems that are not compatible with ACPI 2.0 and later revisions. See 2.5.3.2.6.4 [Fixed ACPI Description Table (FADT) Entries]. 2.5.3.2.6.1 _CST The _CST object should be generated for each core as follows: • Count = 1. 66 42300 Rev 3.12 - July 14, 2015 • • • • BKDG for AMD Family 15h Models 10h-1Fh Processors Register = MSRC001_0073[CstateAddr] + 1. Type = 2. Latency = 100. Power = 0. 2.5.3.2.6.2 _CSD The _CSD object should be generated for each core as follows: • NumberOfEntries = 6. • Revision = 0. • Domain = CPUID Fn0000_0001_EBX[LocalApicId[7:1]]. • CoordType = FEh. (HW_ALL) • NumProcessors = 2. • Index = 0. 2.5.3.2.6.3 _CRS BIOS must declare in the root host bridge _CRS object that the IO address range from MSRC001_0073[CstateAddr] to MSRC001_0073[CstateAddr]+7 is consumed by the host bridge. 2.5.3.2.6.4 Fixed ACPI Description Table (FADT) Entries Declare the following FADT entries: • P_LVL2_LAT = 100. • P_LVL3_LAT = 1001. • FLAGS.PROC_C1 = 1. • FLAGS.P_LVL2_UP = 1. Declare the following P_BLK entries: • P_LVL2 = MSRC001_0073[CstateAddr] + 1. • P_LVL3 = 0. BIOS must declare the PSTATE_CNT entry as 00h. 2.5.3.2.7 BIOS Requirements for Initialization 1. Initialize MSRC001_0073[CstateAddr] with an available IO address. See 2.5.3.2.6.3 [_CRS]. 2. Initialize D18F4x11[C:8]. 3. Generate ACPI objects as described in 2.5.3.2.6 [ACPI Processor C-state Objects]. 2.5.3.3 Effective Frequency The effective frequency interface allows software to discern the average, or effective, frequency of a given core over a configurable window of time. This provides software a measure of actual performance rather than forcing software to assume the current frequency of the core is the frequency of the last P-state requested. This can be useful when the P-state is limited by HTC, D18F3x68, SBI, or CPB. The following procedure calculates effective frequency using MSR0000_00E7 [Max Performance Frequency Clock Count (MPERF)] and MSR0000_00E8 [Actual Performance Frequency Clock Count (APERF)]: 1. At some point in time, write 0 to both MSRs. 2. At some later point in time, read both MSRs. 67 42300 Rev 3.12 - July 14, 2015 BKDG for AMD Family 15h Models 10h-1Fh Processors 3. Effective frequency = (value read from MSR0000_00E8 / value read from MSR0000_00E7) * P0 frequency using software P-state numbering. Additional notes: • The amount of time that elapses between steps 1 and 2 is determined by software. • It is software’s responsibility to disable interrupts or any other events that may occur in between the write of MSR0000_00E7 and the write of MSR0000_00E8 in step 1 or between the read of MSR0000_00E7 and the read of MSR0000_00E8 in step 2. • The behavior of MSR0000_00E7 and MSR0000_00E8 may be modified by MSRC001_0015[EffFreqCntMwait]. • The effective frequency interface provides +/- 50MHz accuracy if the following constraints are met: • Effective frequency is read at most one time per millisecond. • When reading or writing MSR0000_00E7 and MSR0000_00E8 software executes only MOV instructions, and no more than 3 MOV instructions, between the two RDMSR or WRMSR instructions. • MSR0000_00E7 and MSR0000_00E8 are invalid if an overflow occurs. 2.5.4 2.5.4.1 NB Power Management NB P-states The processor supports up to four NB P-states (NBP0 through NBP3), specified in D18F5x1[6C:60]. Each NB P-state consists of the following: • • • • Enable: D18F5x1[6C:60][NbPstateEn]. NCLK frequency: D18F5x1[6C:60][NbFid, NbDid]. VDDNB voltage: D18F5x1[6C:60][NbVid]. Memory P-state: D18F5x1[6C:60][MemPstate]. See 2.5.7.1 [Memory P-states]. Although four NB P-states are defined, only two NB P-states are used at any given time, specified by D18F5x170[NbPstateHi, NbPstateLo]. Hardware dynamically changes which two of the four NB P-states are in use based on GPU activity as follows: • When the GPU is active: • The high NB P-state is specified by D0F0xBC_x1F5F8[DpmXNbPsHi]. • The low NB P-state is specified by D0F0xBC_x1F5F8[DpmXNbPsLo]. • When the GPU is idle , and the timer specified by D0F0xBC_x1F5F8[Hysteresis] has expired: • The high NB P-state is specified by D0F0xBC_x1F5F8[Dpm0PgNbPsHi]. • The low NB P-state is specified by D0F0xBC_x1F5F8[Dpm0PgNbPsLo]. In addition, hardware forces the NB P-state to the active high or low NB P-state based on the GPU activity level. These decisions are all made at the rate specified by D0F0xBC_x1F638[NbDpmPeriod]. Out of cold reset, the NB P-state is specified by D18F5x174[StartupNbPstate] and D18F3xA0[CofVidProg]. The current NB P-state is specified by D18F5x174[CurNbFid, CurNbDid, CurNbVid]. 2.5.4.1.1 NB P-state Transitions Hardware selects whether to use the high or low NB P-state based on several criteria as follows: • Core P-state: 68 42300 Rev 3.12 - July 14, 2015 BKDG for AMD Family 15h Models 10h-1Fh Processors • MSRC001_00[6B:64][NbPstate]. • D18F5x170[NbPstateThreshold]. • GPU activity: • The GPU driver selects levels of GPU activity that force the NB P-state to either the high or low state or allow either NB P-state. • Hysteresis timer: • D18F5x170[NbPstateHiRes, NbPstateLoRes]. • The following configuration registers: • D18F5x170[SwNbPstateLoDis, NbPstateDisOnP0]. • MSRC001_0071[NbPstateDis]. Note: BIOS should set D18F5x170[SwNbPstateLoDis] to 1b, to be cleared by the GPU driver when it loads. Once hardware determines that an NB P-state transition is necessary, hardware executes the following sequence: 1. 2. 3. 4. 5. 6. 7. 8. 9. 10. If transitioning from the low NB P-state to the high NB P-state, transition VDDNB voltage. If the GPU is enabled as specified by D18F5x178[SwGfxDis], wait for display buffer to fill. Quiesce all active cores. If the internal GPU is enabled as specified by D18F5x178[SwGfxDis] and D18F5x178[Dis2ndGnbAllowPsWait]==0, wait for display buffer to fill. Stop memory traffic and place DRAM into self-refresh. Transition NCLK frequency. Update NB P-state specific DRAM settings within hardware, see D18F2x210_dct[1:0]_nbp[3:0]. Take DRAM out of self-refresh and allow memory traffic. Wake up cores. If transitioning from the high NB P-state to the low NB P-state, transition VDDNB voltage. 2.5.4.1.2 2.5.4.1.2.1 BIOS NB P-state Configuration NB P-state COF and VID Synchronization After Warm Reset BIOS performs the following sequence on one core. All APs must be initialized prior to running the steps in this section. See 2.3.2 [AP initialization]. This is done after any warm reset and before 2.9.5 [DCT/DRAM Initialization and Resume]: 1. 2. 3. 4. 5. 6. 7. 8. 9. 10. 11. 12. 13. Temp1=D18F5x170[SwNbPstateLoDis]. Temp2=D18F5x170[NbPstateDisOnP0]. Temp3=D18F5x170[NbPstateThreshold]. Temp4=D18F5x170[NbPstateGnbSlowDis]. If MSRC001_0070[NbPstate]=0, go to step 6. If MSRC001_0070[NbPstate]=1, go to step 11. Write 1 to D18F5x170[NbPstateGnbSlowDis]. Write 0 to D18F5x170[SwNbPstateLoDis, NbPstateDisOnP0, NbPstateThreshold]. Wait for D18F5x174[CurNbPstate] = D18F5x170[NbPstateLo] and D18F5x174[CurNbFid, CurNbDid]=[NbFid, NbDid] from D18F5x1[6C:60] indexed by D18F5x170[NbPstateLo]. Set D18F5x170[SwNbPstateLoDis]=1. Wait for D18F5x174[CurNbPstate] = D18F5x170[NbPstateHi] and D18F5x174[CurNbFid, CurNbDid]=[NbFid, NbDid] from D18F5x1[6C:60] indexed by D18F5x170[NbPstateHi]. Go to step 15. Write 1 to D18F5x170[SwNbPstateLoDis]. Wait for D18F5x174[CurNbPstate] = D18F5x170[NbPstateHi] and D18F5x174[CurNbFid, CurNbDid]=[NbFid, NbDid] from D18F5x1[6C:60] indexed by D18F5x170[NbPstateHi]. Write 0 to D18F5x170[SwNbPstateLoDis, NbPstateDisOnP0, NbPstateThreshold]. 69 42300 Rev 3.12 - July 14, 2015 BKDG for AMD Family 15h Models 10h-1Fh Processors 14. Wait for D18F5x174[CurNbPstate] = D18F5x170[NbPstateLo] and D18F5x174[CurNbFid, CurNbDid]=[NbFid, NbDid] from D18F5x1[6C:60] indexed by D18F5x170[NbPstateLo]. 15. Set D18F5x170[SwNbPstateLoDis]=Temp1, D18F5x170[NbPstateDisOnP0]=Temp2, D18F5x170[NbPstateThreshold]=Temp3, and D18F5x170[NbPstateGnbSlowDis]=Temp4. 2.5.4.1.2.2 NB P-state Voltage Adjustment for MEMCLK During 2.9.5 [DCT/DRAM Initialization and Resume], BIOS updates the NB P-state voltages based on the target MEMCLK for M0 and M1 using the following steps. See 2.9.5.2 [NB P-state Specific Configuration]. 1. Determine the target MEMCLK for M0 and M1 (see 2.9.5 [DCT/DRAM Initialization and Resume] and 2.5.7.1 [Memory P-states]). 2. MemClkVidHi = read D0F0xBC_xE0104168 through D0F0xBC_xE0104170 to find the VID code corresponding to the M0 MEMCLK. If the M0 MEMCLK is not found, use the next higher defined MEMCLK as the target. If the M0 MEMCLK is greater than frequency associated wtih D0F0xBC_xE0104170[MemClkVid8], use D0F0xBC_xE0104170[MemClkVid8] as the target. 3. MemClkVidLo = read D0F0xBC_xE0104168 through D0F0xBC_xE0104170 to find the VID code corresponding to the M1 MEMCLK. If the M1 MEMCLK is not found, use the next higher defined MEMCLK as the target. If the M1 MEMCLK is greater than frequency associated wtih D0F0xBC_xE0104170[MemClkVid8], use D0F0xBC_xE0104170[MemClkVid8] as the target. 4. For each NB P-state from NBP0 through D18F5x170[NbPstateMaxVal]: • If ((D18F5x1[6C:60][MemPstate] == 0) && (MemClkVidHi voltage > D18F5x1[6C:60][NbVid] voltage)): • Program D18F5x1[6C:60][NbVid] == MemClkVidHi. • If ((D18F5x1[6C:60][MemPstate] == 1) && (MemClkVidLo voltage > D18F5x1[6C:60][NbVid] voltage)): • Program D18F5x1[6C:60][NbVid] == MemClkVidLo. 5. Program = MemClkVidHi. 2.5.4.1.2.3 NB P-state Configuration for Runtime Please see your AMD representative for details. 2.5.4.2 NB C-states NB C-states are package-level actions that occur only when all cores enter a non-C0 state (see 2.5.3.2 [Core Cstates]). The NB C-state actions are: • DRAM self-refresh (see 2.5.7.2 [DRAM Self-Refresh]): • Enable bit: D18F4x11[C:8][SelfRefr]. • Entry requirements: • No outstanding GPU traffic or traffic from a link. • Exit conditions (any of the following must be true): • The local APIC timer expires. See 2.4.8.1 [Local APIC]. • New GPU traffic or traffic from a link. • A P-state limit update (see 2.5.3.1.5 [Core P-state Limits]) causes the most restrictive P-state limit to become a higher number than the current P-state for any core in CC1. • NB power gating: • Enable bit: D18F4x11[C:8][NbPwrGate]. • Entry requirements (all of the following must be true): 70 42300 Rev 3.12 - July 14, 2015 BKDG for AMD Family 15h Models 10h-1Fh Processors • No outstanding GPU traffic or traffic from a link. • All cores are in CC6. • DRAM is either in or entering self-refresh. • Exit conditions (any of the following must be true): • The local APIC timer expires. See 2.4.8.1 [Local APIC]. • New GPU traffic or traffic from a link. • A P-state limit update (see 2.5.3.1.5 [Core P-state Limits]) causes the most restrictive P-state limit to become a higher number than the current P-state for any core in CC1. • NB clock gating: • Enable bit: D18F4x11[C:8][NbClkGate]. • Entry requirements: • No outstanding GPU traffic or traffic from a link. • Exit conditions (any of the following must be true): • The local APIC timer expires. See 2.4.8.1 [Local APIC]. • New GPU traffic or traffic from a link. • A P-state limit update (see 2.5.3.1.5 [Core P-state Limits]) causes the most restrictive P-state limit to become a higher number than the current P-state for any core in CC1. When entering NB C-states, the actions are taken in the following order: 1. DRAM self-refresh. 2. NB clock gating. 3. NB power gating. When exiting NB C-states, the actions are taken in the following order: 1. NB power gating. 2. NB clock gating. 3. DRAM self-refresh. 2.5.5 P-state Bandwidth Requirements • The frequency relationship of (core COF / NB COF) <= 6 must be maintained for all supported P-state combinations. E.g., a core COF of 4.0 GHz could not be combined with a NB COF of 0.6 GHz; the NB COF would have to be 0.7 GHz or greater. • All core P-states are required to be defined such that (NB COF/core COF) <= 32, for all NB/core P-state combinations. E.g., if the NB COF is 4.8 GHz then the core COF must be no less than 150 MHz. • All core P-states must be defined such that CoreCOF >= 900 MHz. • All core P-states must be defined such that MSRC001_00[6B:64][CpuFid] <= 22h. • All NB P-states must be defined such that D18F5x1[6C:60][NbFid] <= 2Eh. • NBCOF > 1.25 * MEMCLK frequency. • NBCOF >= 700Mhz. • If NBCOF==700Mhz then MEMCLK must be 333Mhz. 2.5.6 2.5.6.1 GPU and Root Complex Power Management Dynamic Power Management (DPM) The processor supports dynamic SCLK and LCLK frequency changes along with VDDNB voltage change requests, known as Dynamic Power Management (DPM). Once initialized, hardware dynamically monitors processor utilization and adjusts the frequencies and voltage based on that utilization. For both LCLK and SCLK DPM, higher numbered states represent higher performance and lower numbered states represent lower 71 42300 Rev 3.12 - July 14, 2015 BKDG for AMD Family 15h Models 10h-1Fh Processors performance. All family 15h model 10h-1Fh processors support DPM. 2.5.6.1.1 Activity Monitors The processor contains two activity monitors, one for 2.5.6.1.2 [SCLK DPM] and one for 2.5.6.1.3 [LCLK DPM]. Each activity monitor tracks the usage level of different processor subcomponents. A binary signal from each subcomponent is used to determine whether that subcomponent is busy. On each SCLK or LCLK cycle, the respective activity monitors sample the signal from each unmasked subcomponent (see D0F0xBC_xE0000120). The output of the activity monitor is then used to determine whether the DPM state should be changed. See 2.5.6.1.2 [SCLK DPM] and 2.5.6.1.3 [LCLK DPM]. 2.5.6.1.2 SCLK DPM SCLK DPM consists of up to 8 states. Any number of states up through 8 may be used and there is no requirement that the states be contiguous. When SCLK DPM is enabled, the DPM state is transitioned to D0F0xBC_x1F100[SclkDpmBootState] and the activity monitor begins calculating activity levels (see 2.5.6.1.1 [Activity Monitors]). Each time a new activity value is calculated, it is placed into a loop filter and delta-sigma modulator. At that point, the SCLK DPM monitor determines whether to transition to next higher numbered (higher performance) valid state or the next lower numbered (lower performance) valid state. SCLK DPM is enabled by setting D0F0xBC_x1F100[SclkDpmEn]. During runtime, SCLK DPM parameters are programmed by the graphics driver. To support the driver, BIOS creates a data structure in memory containing information about the processor. Please see your AMD representative for more information. 2.5.6.1.3 LCLK DPM LCLK DPM is functionally identical to 2.5.6.1.2 [SCLK DPM] except for the locations of the following parameters: • • • • Valid bit: D0F0xBC_x1F2[E0:00:step20][StateValid]. Voltage change hysteresis threshold: D0F0xBC_x1F2[E0:00:step20][LowVoltageReqThreshold]. Divisor: D0F0xBC_x1F2[E0:00:step20][SclkDivider]. VID: D0F0xBC_x1F2[E0:00:step20][VID]. • See 2.5.2.2 [Dependencies Between Subcomponents on VDDNB]. • State change hysteresis thresholds: D0F0xBC_x1F2[E8:08:step20][HysteresisUp, HysteresisDown]. • Activity thresholds: D0F0xBC_x1F2[F0:10:step20][ActivityThreshold]. • Residency counter: See D0F0xBC_x1F2[E8:08:step20][ResidencyCounter]. LCLK DPM is enabled by setting D0F0xBC_x1F300[LclkDpmEn]. LCLK DPM voltage changes are enabled using D0F0xBC_x1F300[VoltageChgEn]. When LCLK DPM is first enabled, the DPM state is transitioned to D0F0xBC_x1F300[LclkDpmBootState]. 2.5.6.2 GPU and Root Complex Power Gating Several subcomponents of the GPU and root complex can be power gated when not in use. 72 42300 Rev 3.12 - July 14, 2015 BKDG for AMD Family 15h Models 10h-1Fh Processors • GPU: GPU power gating is initialized and enabled by software (see GpuEnabled and D0F0x7C[ForceIntGfxDisable]). Once initialized and enabled, the GPU is power gated by hardware when inactive and is ungated by hardware when needed. When internal GPU is disabled by BIOS, BIOS is reponsible for power gating the GPU. • GMC: GMC power gating is initialized and enabled by software. Once initialized and enabled, GMC is power gated by hardware when inactive and is ungated by hardware when needed. GMC’s state is saved internally. If the internal GPU is disabled, either by hardware (fusing) or by software, software is repsonsible for power gating GMC. • UVD: UVD is power gated by software when not in use and is ungated by software when needed. UVD’s internal state is not saved and UVD goes through an internal reset when power is restored. • VCE: VCE is power gated by software when not in use and is ungated by software when needed. VCE’s internal state is not saved and VCE goes through an internal reset when power is restored. • DCE: DCE is power gated by software when there is no display connected. DCE’s internal state is not saved and DCE goes through an internal reset when power is restored. • PCIe: The Gfx link core can be power gated when it is not in use. In addition, the individual phys on the TX and RX sides of each link can be power gated when the links are not connected. During POST and runtime, several software components inform the SMU whether the Gfx link core or the link phys are in use. Hardware dynamically power gates or ungates the Gfx link core and the link phys as needed. 2.5.7 2.5.7.1 DRAM Power Management Memory P-states The processor supports up to 2 memory P-states, M0 and M1. Each memory P-state consists of the following: • MEMCLK frequency: D18F2x94_dct[1:0][MemClkFreq] and D18F2x2E0_dct[1:0][M1MemClkFreq]. • A set of frequency dependent timing and configuration registers: See 2.9.1 [DCT Configuration Registers]. All valid memory P-states are associated with a specific NB P-state, as specified by D18F5x1[6C:60][MemPstate]. When hardware transitionsto a new NB P-state, the memory P-state is transitioned to that specified by the new NB P-state. Out of cold reset the current memory P-state is M0. The P-state value specified by D18F5x1[6C:60][MemPstate] of the NB P-state indexed by D18F5x174[StartupNbPstate] is invalid. Support for dynamic memory Pstate changes is indicated by D18F3xE8[MemPstateCap]=1 and one or more D18F5x1[6C:60][MemPstate]=1; otherwise M0 is used by hardware for configuration purposes. 2.5.7.2 DRAM Self-Refresh DRAM is placed into self-refresh on S3 entry (see 2.5.8.1.1 [ACPI Suspend to RAM State (S3)]). In addition to S3, DRAM is placed into self-refresh in S0 in the following two scenarios: • NB P-state transitions (see 2.5.4.1 [NB P-states]). • NB C-states (see 2.5.4.2 [NB C-states]). 73 42300 Rev 3.12 - July 14, 2015 BKDG for AMD Family 15h Models 10h-1Fh Processors The following requirements must be met before hardware places DRAM into self-refresh: • No pending traffic. • One of the following is true: • The GPU is idle and the internal display buffer is full. • The internal GPU is disabled. Once the above requirements are met, hardware places DRAM into self-refresh. Early self-refresh occurs when DRAMs are placed in self-refresh before expiration of the cache flush timer. See D18F4x11[C:8][SelfRefrEarly] and D18F5x128[SelfRefrEarlyDis]. If early self-refresh is enabled, the DRAMs are taken out of self-refresh to perform the flush operation when the cache flush timer expires and then placed back into self-refresh. The following are events that cause DRAM to transition out of self-refresh: • Core transitioning to C0. • Incoming request from any link or the GMC • P-state limit update, only in the case when all cores are not in the power gated (CC6) state. To save additional power, hardware always tristates MEMCLK when entering self-refresh. 2.5.7.3 Stutter Mode DRAM is most commonly placed in self-refresh due to stutter mode when the internal GPU is in use. The display buffer in the GPU is a combination of a large buffer known as the DMIF (Display Memory Interface FIFO) and a smaller line buffer. The DMIF takes data originating from DRAM, and sends it to the line buffer to draw to the screen. When the data level in the DMIF is full, DRAM is placed in self-refresh, and incoming DRAM requests are queued. As the DMIF drains, it eventually falls below a predefined watermark level, at which point hardware pulls DRAM out of self-refresh and services all the requests in the queue. Once all the requests are complete and the DMIF is full again, a transition back into self-refresh occurs if the stutter mode conditions are still met. 2.5.7.3.1 System BIOS Requirements for Stutter Mode Operation During POST BIOS creates a data structure in memory containing information about the processor for use by the driver. Please see your AMD representative for more information. 2.5.7.4 EVENT_L EVENT_L is a level sensitive input to the processor. When asserted, the actions specified by D18F2xA4 are taken. EVENT_L is generally asserted to indicate that a DRAM high temperature condition exists. The minimum assertion time for EVENT_L is 15 ns. The minimum deassertion time for EVENT_L is 15 ns. • EVENT_L is pulled to VDDIO on the motherboard. • EVENT_L is ignored while: • PWROK is de-asserted. • RESET_L is asserted. • BIOS must ensure that throttling is disabled (see D18F2xA4[CmdThrottleMode]) until DRAM training is complete. See 2.9.9 [DRAM On DIMM Thermal Management and Power Capping]. 2.5.8 System Power Management 74 42300 Rev 3.12 - July 14, 2015 2.5.8.1 BKDG for AMD Family 15h Models 10h-1Fh Processors S-states S-states are ACPI defined sleep states. S0 is the operational state. All other S-states are low-power states in which various voltage rails in the system may or may not be powered. See the ACPI specification for descriptions of each S-state. 2.5.8.1.1 ACPI Suspend to RAM State (S3) The processor supports the ACPI-defined S3 state. Software is responsible for restoring the state of the processor’s registers when resuming from S3. All registers in the processor that BIOS initialized during the initial boot must be restored. The method used to restore the registers is system specific. During S3 entry, software is responsible for transitioning the processor to Memory Pstate0. See 2.5.7.1 [Memory P-states]. During S3 entry, system memory enters self-refresh mode (see 2.5.7.2 [DRAM Self-Refresh]). Software is responsible for bringing memory out of self-refresh mode when resuming from S3. To bring memory out of self-refresh mode, see 2.9.5 [DCT/DRAM Initialization and Resume]. Many of the systemboard power planes for the processor are powered down during S3. Refer to the Electrical Data Sheet for AMD Family 15h Models 10h-1Fh Processors for the following: • Power plane electrical requirements during S3. • Power plane sequencing requirements on S3 entry and exit. • System signal states for both inputs (e.g. PWROK and RESET_L) and outputs (e.g. VID[*], PSI_L bit, THERMTRIP_L, etc.) during S3. • System signal sequencing requirements on S3 entry and exit. • System management message sequencing on S3 entry and exit. 2.5.9 Bidirectional Application Power Management (BAPM) Bidirectional Application Power Management (BAPM) is the algorithm used to contain temperature within a defined limit. BAPM controls several thermal entities. A thermal entity (TE) is any subcomponent of the processor that reports power consumption to BAPM and controls itself to a power limit specified by BAPM. On Trinity, each compute unit and the GPU core are the thermal entities controlled. On a regular time interval, hardware does the following: 1. Calculates a digital estimate of power consumption for each TE. 2. Converts the power estimates into temperature estimates for each TE. 3. Uses the temperature estimates to assign new power limits to each TE. Once BAPM has assigned power limits, each TE controls its own frequency and voltage to fit within that limit. IF (Revision == RL-A1) THEN 2.5.9.1 Hybrid Boost Hybrid boost is a performance enhancement feature that improves BAPM performance when favorable thermal conditions (like cooler than max-spec ambient, over-the-spec heat sink, etc.) enable the die sensor temperature to be well below the max design temperature limits. When hardware detects such favorable conditions, it 75 42300 Rev 3.12 - July 14, 2015 BKDG for AMD Family 15h Models 10h-1Fh Processors boosts the thermal entities further, beyond what is defined by the baseline BAPM algorithm. Hybrid Boost is enabled if D0F0xBC_x1F8EC[HybridBoostNotSupported] is 0 and D0F0xBC_x1F428[HybridBoostEn] is 1. ENDIF 76 42300 Rev 3.12 - July 14, 2015 2.6 BKDG for AMD Family 15h Models 10h-1Fh Processors Performance Monitoring The processor includes support for two methods of monitoring processor performance: performance monitor counters and instruction based sampling (IBS). There are two types of performance counters: 2.6.1 [Core Performance Monitor Counters], consisting of one set located in each core of each compute unit and 2.6.2 [NB Performance Monitor Counters]. 2.6.1 Core Performance Monitor Counters The core performance monitor counters are used by software to count specific events that occur in a core of the compute unit. Unless otherwise specified, core performance events count only the activity of the core, not activity caused by the other core of the compute unit. Each core of each compute unit provides six 48-bit performance counters. MSRC001_020[A,8,6,4,2,0] [Performance Event Select (PERF_CTL[5:0])] specify the events to be monitored and how they are monitored. MSRC001_020[B,9,7,5,3,1] [Performance Event Counter (PERF_CTR[5:0])] are the counters. MSRC001_00[03:00] is the legacy alias for MSRC001_020[6,4,2,0]. MSRC001_00[07:04] is the legacy alias for MSRC001_020[7,5,3,1]. All of the events are specified in 3.22 [Core Performance Counter Events]. Some performance monitor events have a maximum count per clock that exceeds one event per clock. These performance events are called multi-events. Some counters support a greater multi-event count per clock than others. Events that are multi-events will specify the maximum multi-event count per clock. E.g. The number of events logged per cycle can vary from 0 to X. An event that doesn’t specify multi-event is implied to be a maximum of 1 event per clock. Undefined results will be produced if an multi-event is selected that exceeds that counters capabilities. The following list specifies the maximum number of multi-events supported by each counter: • • • • • • PERF_CTL[0]: 31 multi-event per clock maximum. PERF_CTL[1]: 7 multi-event per clock maximum. PERF_CTL[2]: 7 multi-event per clock maximum. PERF_CTL[3]: 63 multi-event per clock maximum. PERF_CTL[4]: 7 multi-event per clock maximum. PERF_CTL[5]: 7 multi-event per clock maximum. Not all performance monitor events can be counted on all counters. The performance counter registers are generally assigned to specific blocks of the core according to Table 10; however, there are exceptions when an events is implemented by another block of the core and therefore has the counter restrictions of that block. Each core event description starts with one of the following terms to indicate which counters support that event. Selecting an event for a counter that does not support that counter will produce undefined results. Table 10: Core PMC mapping to PERF_CTL[5:0] Term PERF_CTL[5:0] PERF_CTL[3,0] PERF_CTL[0] Definition PERF_CTL[5:0] are used to count events in the LS/DC and EX where the number of events logged per cycle can vary up to 7. PERF_CTL[3,0] are used to count events in the LS/DC and EX where the number of events logged per cycle can vary up to 31. PERF_CTL[0] are used to count events in the LS/DC, EX, IF/DE and CU where the number of events logged per cycle can vary up to 31. 77 42300 Rev 3.12 - July 14, 2015 BKDG for AMD Family 15h Models 10h-1Fh Processors Table 10: Core PMC mapping to PERF_CTL[5:0] Term PERF_CTL[3] PERF_CTL[2:0] PERF_CTL[5:3] Definition PERF_CTL[3] are used to count events in the LS/DC, EX and FP where the number of events logged per cycle can vary up to 63. PERF_CTL[2:0] are used to count events in the IF/DE and CU; The number of events logged per cycle can vary up to 7. PERF_CTL[5:3] are used to count events in the FP; The number of events logged per cycle can vary up to 7. Writing the performance counters can be useful if there is an intention for software to count a specific number of events, and then trigger an interrupt when that count is reached. An interrupt can be triggered when a performance counter overflows. Software should use the WRMSR instruction to load the count as a two’s-complement negative number into the performance counter. This causes the counter to overflow after counting the appropriate number of times. In addition to the RDMSR instruction, the PERF_CTR[5:0] registers can be read using a special read performance-monitoring counter instruction, RDPMC. The accuracy of the performance counters is not ensured. The performance counters are not assured of producing identical measurements each time they are used to measure a particular instruction sequence, and they should not be used to take measurements of very small instruction sequences. The RDPMC instruction is not serializing, and it can be executed out-of-order with respect to other instructions around it. Even when bound by serializing instructions, the system environment at the time the instruction is executed can cause events to be counted before the counter value is loaded into EDX:EAX. 2.6.2 NB Performance Monitor Counters The NB performance monitor counters are used by software to count specific events that occur in the NB. Each node provides four 48-bit performance counters. MSRC001_024[6,4,2,0] [Northbridge Performance Event Select (NB_PERF_CTL[3:0])] and MSRC001_024[7,5,3,1] [Northbridge Performance Event Counter (NB_PERF_CTR[3:0])] specify the events to be monitored and how they are monitored. All of the events are specified in 3.23 [NB Performance Counter Events]. All NB performance monitor events can be counted on all counters. E.g. Unlike core performance events, there are no NB events that can not be counted on some NB performance counters. All NB performance events are one event per clock. E.g. Unlike core performance events, there are no NB multi-event performance events. NB performance counters do not support APIC interrupt capability. In addition to the RDMSR instruction, the NB_PERF_CTR[3:0] registers can be read using a special read performance-monitoring counter instruction, RDPMC. The accuracy of the performance counters is not ensured. The performance counters are not assured of producing identical measurements each time they are used to measure a particular instruction sequence, and they should not be used to take measurements of very small instruction sequences. The RDPMC instruction is not 78 42300 Rev 3.12 - July 14, 2015 BKDG for AMD Family 15h Models 10h-1Fh Processors serializing, and it can be executed out-of-order with respect to other instructions around it. Even when bound by serializing instructions, the system environment at the time the instruction is executed can cause events to be counted before the counter value is loaded into EDX:EAX. 2.6.3 Instruction Based Sampling (IBS) IBS is a code profiling mechanism that enables the processor to select a random instruction fetch or micro-op after a programmed time interval has expired and record specific performance information about the operation. An interrupt is generated when the operation is complete as specified by MSRC001_103A [IBS Control]. An interrupt handler can then read the performance information that was logged for the operation. The IBS mechanism is split into two parts: instruction fetch performance controlled by MSRC001_1030 [IBS Fetch Control (IC_IBS_CTL)]; and instruction execution performance controlled by MSRC001_1033 [IBS Execution Control (SC_IBS_CTL)]. Instruction fetch sampling provides information about instruction TLB and instruction cache behavior for fetched instructions. Instruction execution sampling provides information about micro-op execution behavior. The data collected for instruction fetch performance is independent from the data collected for instruction execution performance. Support for the IBS feature is indicated by the CPUID Fn8000_0001_ECX[IBS]. Instruction fetch performance is profiled by recording the following performance information for the tagged instruction fetch: • If the instruction fetch completed or was aborted. See MSRC001_1030. • The number of clock cycles spent on the instruction fetch. See MSRC001_1030. • If the instruction fetch hit or missed the IC, hit/missed in the L1 and L2 TLBs, and page size. See MSRC001_1030. • The linear address, physical address associated with the fetch. See MSRC001_1031, MSRC001_1032. Instruction execution performance is profiled by tagging one micro-op associated with an instruction. Instructions that decode to more than one micro-op return different performance data depending upon which micro-op associated with the instruction is tagged. These micro-ops are associated with the RIP of the next instruction to retire. The following performance information is returned for the tagged micro-op: • Branch and execution status for micro-ops. See MSRC001_1035. • Branch target address for branch micro-ops. See MSRC001_103B. • The logical address associated with the micro-op. See MSRC001_1034. • The linear and physical address associated with a load or store micro-op. See MSRC001_1038, MSRC001_1039. • The data cache access status associated with the micro-op: DC hit/miss, DC miss latency, TLB hit/miss, TLB page size. See MSRC001_1037. • The number clocks from when the micro-op was tagged until the micro-op retires. See MSRC001_1035. • The number clocks from when the micro-op completes execution until the micro-op retires. See MSRC001_1035. • Source information for DRAM, MMIO, L3 hit and state. See MSRC001_1036. 79 42300 Rev 3.12 - July 14, 2015 2.7 BKDG for AMD Family 15h Models 10h-1Fh Processors Configuration Space PCI-defined configuration space was originally defined to allow up to 256 bytes of register space for each function of each device; these first 256 bytes are called base configuration space (BCS). It was expanded to support up to 4096 bytes per function; bytes 256 through 4095 are called extended configuration space (ECS). The processor includes configuration space registers located in both BCS and ECS. Processor configuration space is accessed through bus 0, devices 18h to 1Fh, where device 18h corresponds to node 0 and device 1Fh corresponds to node 7. See 2.7.3 [Processor Configuration Space]. Configuration space is accessed by the processor through two methods: • IO-space configuration: IO instructions to addresses CF8h and CFCh. • Enabled through IOCF8 [IO-Space Configuration Address][ConfigEn], which allows access to BCS. • Access to ECS enabled through MSRC001_001F [Northbridge Configuration 1 (NB_CFG1)][EnableCf8ExtCfg]. • Use of IO-space configuration can be programmed to generate GP faults through MSRC001_0015 [Hardware Configuration (HWCR)][IoCfgGpFault]. • SMI trapping for these accesses is specified by MSRC001_0054 [IO Trap Control (SMI_ON_IO_TRAP_CTL_STS)] and MSRC001_00[53:50] [IO Trap (SMI_ON_IO_TRAP_[3:0])]. • MMIO configuration: configuration space is a region of memory space. • The base address and size of this range is specified by MSRC001_0058 [MMIO Configuration Base Address]. The size is controlled by the number of configuration-space bus numbers supported by the system. Accesses to this range are converted configuration space as follows: • Address[31:0] = {0h, bus[7:0], device[4:0], function[2:0], offset[11:0]}. BIOS may use either configuration space access mechanism during boot. Before booting the OS, BIOS must disable IO access to ECS, enable MMIO configuration and build an ACPI defined MCFG table. BIOS ACPI code must use MMIO to access configuration space. Per the link specification, BCS accesses utilize link addresses starting at FD_FE00_0000h and ECS accesses utilize link addresses starting at FE_0000_0000h. 2.7.1 MMIO Configuration Coding Requirements MMIO configuration space accesses must use the uncacheable (UC) memory type. Instructions used to read MMIO configuration space are required to take the following form: mov eax/ax/al, <any_address_mode>; Instructions used to write MMIO configuration space are required to take the following form: mov <any_address_mode>, eax/ax/al; No other source/target registers may be used other than eax/ax/al. In addition, all such accesses are required not to cross any naturally aligned DW boundary. Access to MMIO configuration space registers that do not meet these requirements result in undefined behavior. 2.7.2 MMIO Configuration Ordering Since MMIO configuration cycles are not serializing in the way that IO configuration cycles are, their ordering rules relative to posted may result in unexpected behavior. 80 42300 Rev 3.12 - July 14, 2015 BKDG for AMD Family 15h Models 10h-1Fh Processors Therefore, processor MMIO configuration space is designed to match the following ordering relationship that exists naturally with IO-space configuration: if a core generates a configuration cycle followed by a postedwrite cycle, then the posted write is held in the processor until the configuration cycle completes. As a result, any unexpected behavior that might have resulted if the posted-write cycle were to pass MMIO configuration cycle is avoided. 2.7.3 Processor Configuration Space The processor includes configuration space as described in 3 [Registers]. Accesses to unimplemented registers of implemented functions are ignored: writes dropped; reads return 0’s. Accesses to unimplemented functions also ignored: writes are dropped; however, reads return all F’s. The processor does not log any master abort events for accesses to unimplemented registers or functions. Accesses to device numbers of devices not implemented in the processor are routed based on the configuration map registers. If such requests are master aborted, then the processor can log the event. 81 42300 Rev 3.12 - July 14, 2015 2.8 BKDG for AMD Family 15h Models 10h-1Fh Processors Northbridge (NB) Each node includes a single northbridge that provides the interface to the local core(s), the interface to system memory, and the interface to system IO devices. The NB includes all power planes except VDD; see 2.5.1 [Processor Power Planes And Voltage Control]. The NB is responsible for routing transactions sourced from cores and link to the appropriate core, cache, DRAM, or link. See 2.4.5 [System Address Map]. 2.8.1 NB Architecture Major NB blocks are: System Request Interface (SRI), Memory Controller (MCT), DRAM Controllers (DCTs), and crossbar (XBAR). SRI interfaces with the core(s). MCT maintains cache coherency and interfaces with the DCTs; MCT maintains a queue of incoming requests called MCQ. XBAR is a switch that routes packets between SRI, MCT, and the link. The MCT operates on physical addresses. Before passing transactions to the DCTs, the MCT converts physical addresses into normalized addresses that correspond to the values programmed into D18F2x[5C:40]_dct[1:0] [DRAM CS Base Address]. Normalized addresses include only address bits within the DCTs’ range. The normalized address varies based on DCT interleave and hoisting settings in D18F2x110 [DRAM Controller Select Low] and D18F2x114 [DRAM Controller Select High] as well as node interleaving based on D18F1x[144:140,44:40] [DRAM Base/Limit]. 2.8.2 2.8.2.1 NB Routing Address Space Routing There are four main types of address space routed by the NB: 1. Memory space targeting system DRAM 2. Memory space targeting IO (MMIO) 3. IO space 4. Configuration space. 2.8.2.1.1 DRAM and MMIO Memory Space Memory space transactions provide the NB with the physical address, cacheability type, access type, and DRAM/MMIO destination type as specified in section 2.4.5.1.2 [Determining The Access Destination for Core Accesses]. Memory-space transactions are handled by the NB as follows: • IO-device accesses are compared against: • If the access matches D18F1x[1CC:180,BC:80] [MMIO Base/Limit], then the transaction is routed to the root complex; • Else, if the access matches D18F1x[144:140,44:40] [DRAM Base/Limit], then the access is routed to the DCT; • Else, the access is routed to the UMI. • For core accesses the routing is determined based on the DRAM/MMIO destination: • If the destination is DRAM: • If the access matches D18F1x[144:140,44:40] [DRAM Base/Limit], then the transaction is routed to the DCT; • Else, the access is routed to the UMI. • If the destination is MMIO: • If the access matches D18F1x[1CC:180,BC:80] [MMIO Base/Limit], then the transaction is routed 82 42300 Rev 3.12 - July 14, 2015 BKDG for AMD Family 15h Models 10h-1Fh Processors to the root complex; • Else, If the access matches the VGA-compatible MMIO address space and D18F1xF4[VE]=1 then D18F1xF4 describes how the access is routed and controlled. • Else, the access is routed to the UMI. 2.8.2.1.2 IO Space IO-space transactions from IO links or cores are routed as follows: • If the access matches D18F1x[DC:C0] [IO-Space Base/Limit], then the transaction is routed to the root complex; • Else, If the access matches the VGA-compatible IO address space and D18F1xF4[VE]=1 then D18F1xF4 describes how the access is routed and controlled. • Else, the access is routed to the UMI. 2.8.2.1.3 Configuration Space Configuration-space transactions from IO links are master aborted. Configuration-space transactions from cores are routed as follows: • If the access targets the configuration space of an existing node (based on the configuration-space address and D18F0x60[NodeCnt]), then it is routed to that node. • Else, if the access matches D18F1x[EC:E0] [Configuration Map], then the transaction is routed to the specified link; • Else, the access is routed to the node or link that contains compatibility (subtractive) address space, specified by D18F0x60[SbNode] and D18F0x64[SbLink]. 83 42300 Rev 3.12 - July 14, 2015 2.9 BKDG for AMD Family 15h Models 10h-1Fh Processors DRAM Controllers (DCTs) The processor includes two DRAM controllers (DCTs). Each DCT controls one 64-bit DDR3 DRAM channel. A DRAM channel consists of the group of DRAM interface pins connecting to one series of DIMMs. The DCTs operate on normalized addresses corresponding to the values programmed into the D18F2x[5C:40]_dct[1:0] [DRAM CS Base Address]. Normalized addresses only include address bits within a DCT’s range. The physical to normalized address translation varies based on DCT interleave and hoisting settings. See 2.9.6 [Memory Interleaving Modes] and 2.9.7 [Memory Hoisting]. The following restrictions limit the DIMM types and configurations supported by the DCTs: • All DIMMs connected to the processor are required to operate at the same MEMCLK frequency, regardless of the channel. All DCTs must be programmed to the same frequency. Table 11: DCT Definitions Term Definition SR Single Rank DR Dual Rank any Any: SR or DR. MRS JEDEC defined DRAM Mode Register Set. NP No DIMM populated DIMM0 DIMM slots 0-n. The DIMMs on each channel are numbered from 0 to n where DIMM0 is the DIMM closest to the processor on that channel and DIMMn is the DIMM1 DIMM farthest from the processor on that channel. DimmsPopulated The number of DIMMs populated per channel NumDimmSlots The number of motherboard DIMM slots per channel DIMM The DIMM being configured Rank The rank being configured NumRegisters The number of registers on the DIMM being configured NumDimmSlots The number of motherboard DIMM slots per channel RowAddrBits SPDByte[5][5:3] of the DIMM being configured. DramCapacity SPDByte[4][3:0] of the DIMM being configured. RankMap SPDByte[63][0] of the DIMM being configured. NumRanks SPDByte[7][5:3] of the DIMM being configured, or the number of ranks soldered down. DeviceWidth SPDByte[7][2:0] of the DIMM being configured. AutoSelfRefresh SPDByte[31][2] of the DIMM being configured. ExtendedTemperature- SPDByte[31][0] of the DIMM being configured. Range DdrRate The DDR data rate (MT/s). VDDIO DDR VDDIO in V. UDIMM DCT is configured for UDIMM if (D18F2x90_dct[1:0][UnbuffDimm]==1). SODIMM DCT is configured for UDIMM if (D18F2x90_dct[1:0][UnbuffDimm]==1) and the processor is connected to SO-DIMM(s) . This definition is used to distinguish between settings optimized for SO-DIMMs, as in a laptop, or standard sized UDIMMs as in a desktop. 84 42300 Rev 3.12 - July 14, 2015 BKDG for AMD Family 15h Models 10h-1Fh Processors The following restrictions limit the DIMM types and configurations supported by the DCTs: • All DIMMs connected to a processor are required to operate at the same MEMCLK frequency, regardless of the channel. All DCTs must be programmed to the same frequency. The tables below list the maximum DIMM speeds supported by the processor for different configurations. The motherboard should comply with the relevant AMD socket motherboard design guideline (MBDG) to achieve the rated speeds. In cases where MBDG design options exist, lower-quality options may compromise the maximum achievable speed; motherboard designers should assess the tradeoffs. Table 12: DDR3 unbuffered DIMM maximum frequency support for FM2 (per channel) Frequency (MT/s)5 1.5V 1.35V 1.25V 2 1 2133 1600 1333 2 SR 1866 1600 1333 6 1600 1066 1 DR + (TN-A0 || TN-A1) : 1 SR/DR 1600 RL-A16: 1866 5. Population restrictions (including the order for partially populated channels) may apply. See 2.9.5.6.6. 6. See 1.5 for revision information. DIMM Slots Installed DIMMs Table 13: DDR3 SO-DIMM maximum frequency support for FS1r2 (per channel) Frequency (MT/s)5 1.5V 1.35V 1.25V 1 1 1866 1600 1333 7. Population restrictions (including the order for partially populated channels) may apply. See 2.9.5.6.6. DIMM Slots Installed DIMMs Table 14: DDR3 unbuffered DIMM maximum frequency support for FP2 (per channel) for (TN-A0 || TN-A1) DIMM Slots 1 2 Installed DIMMs 1.5V 1600 1333 1333 1333 Frequency (MT/s)8 1.35V 1600 1333 1333 1333 1.25V 1333 1333 1333 1066 1 1 2 SR 1 DR + 1 SR/DR 8. For (TN-A0 || TN-A1). See 1.5 for revision information. Population restrictions (including the order for partially populated channels) may apply. See 2.9.5.6.6. 85 42300 Rev 3.12 - July 14, 2015 BKDG for AMD Family 15h Models 10h-1Fh Processors Table 15: DDR3 SO-DIMM maximum frequency support for FP2 (per channel) for (TN-A0 || TN-A1) DIMM Slots Installed DIMMs 1 2 1.5V 1600 1333 1333 1333 Frequency (MT/s)9 1.35V 1600 1333 1333 1333 1.25V 1333 1066 1066 1066 1 1 2 SR 1 DR + 1 SR/DR 9. For (TN-A0 || TN-A1). See 1.5 for revision information. Population restrictions (including the order for partially populated channels) may apply. See 2.9.5.6.6. Table 16: DDR3 unbuffered DIMM maximum frequency support for FP2 (per channel) for RL-A1 DIMM Slots Installed DIMMs 1 2 2 1.5V 1600 1600 1600 1600 Frequency (MT/s)5 1.35V 1600 1333 1333 1333 1.25V 1333 1333 1333 1066 1 1 2 SR 1 DR + 1 SR/DR 1. For RL-A1. See 1.5 for revision information. Population restrictions (including the order for partially populated channels) may apply. See 2.9.5.6.6. Table 17: DDR3 SO-DIMM maximum frequency support for FP2 (per channel) for RL-A1 DIMM Slots 1 2 2 Installed DIMMs 1.5V 1600 1333 1333 1333 Frequency (MT/s)5 1.35V 1600 1333 1333 1333 1.25V 1333 1333 1333 1066 1 1 2 SR 1 DR + 1 SR/DR 1. For RL-A1. See 1.5 for revision information. Population restrictions (including the order for partially populated channels) may apply. See 2.9.5.6.6. Table 18: DDR3 Soldered-down DRAM maximum frequency support for FP2 (per channel) Frequency (MT/s)5 1.5V 1.35V 1.25V Up to 4 1333 1333 1333 1. Population restrictions (including the order for partially populated channels) may apply. See 2.9.5.6.6. Ranks 86 42300 Rev 3.12 - July 14, 2015 2.9.1 BKDG for AMD Family 15h Models 10h-1Fh Processors DCT Configuration Registers There are two types of DCT configuration registers: • Registers for which there is one instance for all DCT’s. E.g. D18F2xAC [DRAM Controller Temperature Status]. • Registers for which there is one instance per DCT. E.g. D18F2x78_dct[1:0] [DRAM Control]. • For D18F2x78_dct[x], x=D18F1x10C[DctCfgSel]; see D18F1x10C[DctCfgSel]. • The syntax for this register type is described by example as follows: • D18F2x78_dct[1:0] refers to all instances of the D18F2x78 register. • D18F2x78_dct[1] refers to the D18F2x78 register instance for DCT1. Registers for which there is one instance per memory P-state use D18F1x10C[MemPsSel] for software accesses. The syntax for this register type is described by example as follows: • D18F2x2E8_dct[1:0]_mp[1:0] refers to all instances of the D18F2x2E8 register. • D18F2x2E8_dct[1:0]_mp[1] refers to the register for memory P-state 1 of either or both DCTs. • Software must ensure consistency of context when accessing registers with both controller fields and DDR phy fields. See D18F2x98_dct[1:0] for more information. 2.9.2 DDR Pad to Processor Pin Mapping The relationship of pad drivers to processor pins varies by package as shown in the following table. Table 19: Package pin mapping Pad MEMCLK0_H[0] MEMCLK0_H[1] MEMCLK0_H[2] MEMCLK0_H[3] MEMCLK0_H[4] MEMCLK0_H[5] MEMCLK0_H[6] MEMCLK0_H[7] MEMCLK1_H[0] MEMCLK1_H[1] MEMCLK1_H[2] MEMCLK1_H[3] MEMCLK1_H[4] MEMCLK1_H[5] MEMCLK1_H[6] MEMCLK1_H[7] MEMCS0_L[0] MEMCS0_L[1] MEMCS0_L[2] MEMCS0_L[3] MEMCS0_L[4] FS1r2 NC NC NC NC NC NC Pin1 FM2 FP2 MA_CLK_H[0] MA_CLK_H[1] MA_CLK_H[2] MA_CLK_H[3] NC NC NC NC MB_CLK_H[0] MB_CLK_H[1] MB_CLK_H[2] MB_CLK_H[3] NC NC NC NC MA0_CS_L[0] MA0_CS_L[1] MA1_CS_L[0] MA1_CS_L[1] NC 87 42300 Rev 3.12 - July 14, 2015 BKDG for AMD Family 15h Models 10h-1Fh Processors Table 19: Package pin mapping Pin1 FS1r2 FM2 FP2 MEMCS0_L[5] NC MEMCS0_L[6] NC MEMCS0_L[7] NC MEMCS1_L[0] MB0_CS_L[0] MEMCS1_L[1] MB0_CS_L[1] MEMCS1_L[2] NC MB1_CS_L[0] MEMCS1_L[3] NC MB1_CS_L[1] MEMCS1_L[4] NC MEMCS1_L[5] NC MEMCS1_L[6] NC MEMCS1_L[7] NC MEMODT0[0] MA0_ODT[0] MEMODT0[1] MA0_ODT[1] MEMODT0[2] NC MA1_ODT[0] MEMODT0[3] NC MA1_ODT[1] MEMODT1[0] MB0_ODT[0] MEMODT1[1] MB0_ODT[1] MEMODT1[2] NC MB1_ODT[0] MEMODT1[3] NC MB1_ODT[1] MEMCKE0[0] MA_CKE[0] MEMCKE0[1] MA_CKE[1] MEMCKE0[2] NC MEMCKE0[3] NC MEMCKE1[0] MB_CKE[0] MEMCKE1[1] MB_CKE[1] MEMCKE1[2] NC MEMCKE1[3] NC 1. For differential pins, only positive polarity pins are shown; negative polarity pins have corresponding mapping and are controlled by the same CSR field. NC = Not connected. BIOS should tri-state or disable the pad for maximum power savings. Pad 2.9.2.1 DDR Chip to Pad Mapping The relationship of chip to pad drivers is shown in the following table. BIOS should disable or power down unused chips for maximum power savings. 88 42300 Rev 3.12 - July 14, 2015 BKDG for AMD Family 15h Models 10h-1Fh Processors Table 20: Pad (from chiplet) pin mapping Pad1 FS1r2 FM2 FP2 Clk chip 0 0 MEMCLK0_H[1:0] Clk chip 1 0 MEMCLK0_H[3:2] Cmd/Addr 0 0 MEMCS0_L[7,5,3,1] 1 MEMODT0[3:0] 2 MEMCS0_L[6,4,2,0] Cmd/Addr 1 0 MEMRAS0_L MEMCAS0_L MEMWE0_L MEMADD0[13] 1 MEMADD0[10,0] MEMBANK0[1:0] 2 MEMPAR0 Address 2 0 MEMCKE0[3:0] 1 MEMADD0[0,1,8,9] 2 MEMADD0[2,3,10,11] 3 MEMADD0[4,5,12,13] 4 MEMADD0[6,7,14,15] 2 0 MEMDATA[5,4,1,0] Data 1 MEMDQS_H[0], MEMDQSDM[0] 2 MEMDATA[7,6,3,2] 1. For differential pads, only positive polarity pads are shown; negative polarity pads have corresponding mapping and are controlled by the same chip. Only channel A map is shown. Channel B is similar. 2. Only Data chip 0 is shown. Data chips [7:1] are repeated with sequential DQ/DQS/DM pin numbers. Chip 2.9.3 Group DRAM Controller Direct Response Mode The DCT supports direct response mode for responding to a cache line fill request before the DCT is initialized. In direct response mode, the target DCT responds to a cache line fill request by returning 64 bytes of all ones without issuing a read transaction on the DRAM bus. The BIOS uses this featureto allocate cache lines for temporary data storage. The controller exits direct response mode when either D18F2x7C_dct[1:0][EnDramInit] or D18F2x90_dct[1:0][InitDram] is set to 1. See 2.9.5.8 [DRAM Device and Controller Initialization] and 2.3.3 [Using L2 Cache as General Storage During Boot]. 2.9.4 DRAM Data Burst Mapping DRAM requests are mapped to data bursts on the DDR bus in the following order: • When D18F2x110[DctDatIntLv] = 0, a 64 B request is mapped to each of the eight sequential data beats as QW0, QW1...QW7. 89 42300 Rev 3.12 - July 14, 2015 BKDG for AMD Family 15h Models 10h-1Fh Processors • When D18F2x110[DctDatIntLv] = 1, the order of cache data to QW on the bus is the same except that even and odd bits are interleaved on the DRAM bus as follows: • For every 8 bytes in the cache line, even bits map to QW0, QW2, QW4, and QW6 on the DRAM bus. • For every 8 bytes in the cache line, odd bits map to QW1, QW3, QW5, and QW7 on the DRAM bus. 2.9.5 DCT/DRAM Initialization and Resume DRAM initialization involves several steps in order to configure the DRAM controllers and the DRAM, and to tune the DRAM channel for optimal performance. DRAM resume requires several steps to configure the DCTs to properly resume from the S3 state. The following sequence describes the steps needed after a reset for initialization or resume: • To disable an unused DRAM channel see DRAM Channel Disable [2.9.5.10]. 1. 2. 3. 4. 5. 6. 7. 8. 9. 10. Configure the DDR supply voltage regulator. See 2.9.5.1. Force NB P-state to NBP0. See 2.9.5.2. Force register accesses to M0. See 2.9.5.3 for further requirements used in steps below. DDR phy initialization. See 2.9.5.4. DRAM device and controller initialization. • If BIOS is booting from an unpowered state (ACPI S4, S5 or G3), then it performs the following: a. Program SPD configuration. See 2.9.5.5. b. Program Non-SPD configuration. See 2.9.5.6. c. Program DCT training specific configuration. See 2.9.5.7. d. Program the remaining DCT registers not covered by an explicit sequence dependency. e. DRAM device initialization. See 2.9.5.8. • If BIOS is resuming the platform from S3 state, then it performs the following: a. Restore all DCT and phy registers that were programmed during the first boot from non-volatile storage. See 2.9.5.5 and 2.9.5.6 for a review of registers. b. Program D18F2x90_dct[1:0][ExitSelfRef] = 1. c. Restore the trained delayed values (found during the initial boot in step 6 below) from nonvolatile storage. d. Continue at step 9. DRAM data training. A. Write levelization training. See 2.9.5.9.1. B. DQS receiver enable training. See 2.9.5.9.2. C. DQS position training. See 2.9.5.9.4. D. MaxRdLatency training. See 2.9.5.9.5.1. NB P-state specific training. For each NB P-state from NBP1 to D18F5x170[NbPstateMaxVal]: A. Force the NB P-state. See 2.9.5.2. B. MaxRdLatency training. See 2.9.5.9.5.1. Release NB P-state force. Program DCT for normal operation. See 2.9.5.7. Program DRAM phy for power savings. See 2.9.5.11. The DRAM subsystem is ready for use. 2.9.5.1 Low Voltage DDR3 The processor supports JEDEC defined DDR3L and DDR3U devices which operate at voltages lower than 1.5V. 90 42300 Rev 3.12 - July 14, 2015 BKDG for AMD Family 15h Models 10h-1Fh Processors Platforms supporting low voltage devices should power up VDDIO at 1.35V. BIOS should not operate DIMMs at voltages higher than supported as indicated by SPD Byte 6: Module Nominal Voltage, VDD. BIOS should consult vendor data sheets for the supply voltage regulator programming requirements. On supported platforms, BIOS must take steps to configure the supply voltage regulators as follows: 1. Read the SPD of all DIMMs within the programmable VDDIO domain and check all of the defined bits within the SPD byte to determine the common operating voltages. 2. Configure VDDIO to match the lowest common supported voltage based on the SPD values. • If the DIMMs do not specify a common operating voltage then BIOS must take platform vendor defined action to notify the end user of the mismatch and to protect DIMMs from damage. 3. Additional derating of the DDR speed may be necessary for reliable operation at lower voltage. See 2.9.5.6.6 [DRAM Address Timing and Output Driver Compensation Control] for information on voltage and electrical load specific maximum speed. 2.9.5.2 NB P-state Specific Configuration A subset of DCT configuration and training must be repeated for each enabled NB P-state. To accomplish this, BIOS forces the processor to the desired NB P-state and releases the force once DRAM initialization and training is complete. When D18F5x174[NbPstateDis]=0, BIOS performs the following to configure NB P-states for MEMCLK frequency specific operation: 1. Adjust the NB P-state northbridge voltage. See 2.5.4.1.2.2. • BIOS may perform this step once for all defined NB P-states, or for each time an NB P-state is forced during DRAM initialization to the target state only. When D18F5x174[NbPstateDis]=0, BIOS performs the following to configure the DCT for NB P-states: Before DRAM training, BIOS performs the following to transition to the highest performance NB P-state, NBP0: 1. Save the initial value of D18F5x170. 2. If D18F5x174[CurNbPstate]!= NBP0 and D18F5x174[CurNbPstate]==D18F5x170[NbPstateHi], transition to target NB P-state, NBP0, via Steps 3-7 below. Else if D18F5x174[CurNbPstate]!= NBP0 and D18F5x174[CurNbPstate]==D18F5x170[NbPstateLo], transition to target NB P-state, NBP0, via Steps 3,6-7 below. 3. Program the configuration registers which contain multiple internal copies for each NB P-state. See D18F1x10C[NbPsSel]. 4. Program D18F5x170 to transition the NB P-state: • SwNbPstateLoDis = NbPstateDisOnP0 = NbPstateThreshold = 0. • NbPstateLo = NbPstateMaxVal. 5. Wait for D18F5x174[CurNbPstate] to equal NbPstateLo. 6. Program D18F5x170 to force the NB P-state: • NbPstateHi = target NB P-state. • SwNbPstateLoDis = 1 7. Wait for D18F5x174[CurNbPstate] to equal the target NB P-state. BIOS performs the following to release the NB P-state force: 8. Restore the initial D18F5x170[SwNbPstateLoDis, NbPstateDisOnP0, NbPstateLo] values. 9. Restore the initial D18F5x170[NbPstateThreshold, NbPstateHi] values. 91 42300 Rev 3.12 - July 14, 2015 BKDG for AMD Family 15h Models 10h-1Fh Processors See also 2.5.4.1 [NB P-states]. 2.9.5.3 Memory P-state Specific Configuration A subset of DCT configuration and training must be repeated for each enabled memory P-state. See also 2.5.7.1 [Memory P-states]. To accomplish this, BIOS forces the processor to the NBP0 state and uses the M0 context to train for values that are used by all memory P-states. For example, BIOS uses D18F2x94_dct[1:0][MemClkFreq] to specify the dram frequency during each pass of training. BIOS saves the trained configuration values at each step and uses them appropriately as outlined below: 1. 2. 3. 4. Force NB P-state to NBP0. See 2.9.5.2. Program D18F1x10C[MemPsSel]=0. See 2.9.1. Program D18F2x9C_x0D04_E008_dct[1:0][PStateToAccess]=0. Using the M0 context, BIOS optimally configures the controller and trains at DDR667 according to 2.9.5 steps 4 thru 7. A. BIOS saves the trained values for use in M1 and for subsequent higher frequencies. B. BIOS saves the controller configuration for use in M1. C. BIOS optimally writes all the values to registers in step 6. It may optionally write the values as each is computed or trained. If this is done then BIOS must ensure that D18F1x10C[MemPsSel] and D18F2x9C_x0D04_E008_dct[1:0][PStateToAccess] are configured correctly before and after each register access. 5. BIOS increases the frequency and repeats step 4 until the target frequency is trained. The target frequency is always associated with M0. 6. BIOS programs the M1 controller context with the DDR667 configuration and trained results. This must be done before the next subsequent NB P-state change. a. Program D18F1x10C[MemPsSel]=1. b. Program D18F2x9C_x0D04_E008_dct[1:0][PStateToAccess] = 1. c. Program D18F2x2E0_dct[1:0][M1MemClkFreq] and D18F2x9C_x0D0F_E000_dct[1:0]_mp[1][Rate, FreqValid] = {04h, 1}. d. Program DCT configuration and trained delay values saved earlier. Three groups of registers must be written in the same order as the initialization sequence: • Fence values: D18F2x9C_x0000_000C_dct[1:0], D18F2x9C_x0D0F_0[F,7:0]31_dct[1:0], D18F2x9C_x0D0F_E019_dct[1:0]. • Address/Command timing delays: D18F2x9C_x0000_0004_dct[1:0]_mp[1:0]. • All other Data/Dqs delays: See step 6 in section 2.9.5. e. Program D18F2x9C_x0D04_E008_dct[1:0][PStateToAccess] = 0. f. Program D18F1x10C[MemPsSel] = 0. 2.9.5.4 DDR Phy Initialization The BIOS initializes the phy and the internal interface from the DCT to the phy, including the PLLs and the fence value, after each reset and for each time a MEMCLK frequency change is made. BIOS obtains size, loading, and frequency information about the DIMMs and channels using SPDs prior to phy initialization. BIOS then performs the following actions: 1. Program D18F2x9C_x0D0F_E013_dct[1:0] = 0118h. 2. For each byte lane and each memory P-state: Program D18F2x9C_x0D0F_0[F,7:0]13_dct[1:0]_mp[1:0][RxSsbMntClkEn] = 0. 3. Program D18F2xA8_dct[1:0][MemPhyPllPdMode] = 00b. 4. Force the phy to M0 with the following sequence: 92 42300 Rev 3.12 - July 14, 2015 5. 6. 7. 8. BKDG for AMD Family 15h Models 10h-1Fh Processors A. Program D18F2x9C_x0D0F_E006_dct[1:0][PllLockTime] = 190h. B. For each DCT: Program D18F2x9C_x0000_000B_dct[1:0] = 80800000h. C. Program D18F2x9C_x0D0F_E018_dct[0][PhyPSMasterChannel] = 0. D. Program D18F2x9C_x0000_000B_dct[0] = 40000000h. E. For each DCT: Program D18F2x9C_x0000_000B_dct[1:0] = 80000000h. Phy Voltage Level Programming. See 2.9.5.4.1. DRAM channel frequency change. See 2.9.5.4.2. Phy fence programming. See 2.9.5.4.3. Phy compensation initialization. See 2.9.5.4.4. 2.9.5.4.1 Phy Voltage Level Programming BIOS programs the following according to the desired phy VDDIO voltage level: • Program D18F2x9C_x0D0F_0[F,7:0]1F_dct[1:0]_mp[1:0][RxVioLvl]. • Program D18F2x9C_x0D0F_[C,8,2][2:0]1F[RxVioLvl]. • Program D18F2x9C_x0D0F_4009_dct[1:0][CmpVioLvl,ComparatorAdjust]. See 2.9.5.1 [Low Voltage DDR3]. 2.9.5.4.2 DRAM Channel Frequency Change The following sequence is used to change the DRAM frequency under all boot conditions, including restoring the DCT state when resuming from the S3 state. BIOS performs the sequence only on DCTs with an enabled interface (D18F2x94_dct[1:0][DisDramInterface]==0): For each DCT: 1. Program D18F2x9C_x0D0F_E006_dct[1:0][PllLockTime] = 190h. 2. Program D18F2x94_dct[1:0][MemClkFreqVal] = 0. 3. Program D18F2x94_dct[1:0][MemClkFreq] to the desired DRAM frequency. 4. Program the following parameters which must be configured prior to setting MemClkFreqVal: • D18F2x90_dct[1:0][X4Dimm] • D18F2x94_dct[1:0][ProcOdtDis] • D18F2x9C_x0000_0004_dct[1:0]_mp[1:0] • D18F2x9C_x0D0F_0[F,7:0]13_dct[1:0]_mp[1:0][ProcOdtAdv] • D18F2x9C_x0D0F_E00A_dct[1:0][SkewMemClk] • D18F2x210_dct[1:0]_nbp[3:0][RdPtrInit, DataTxFifoWrDly] of the current NB P-state for the target MEMCLK frequency. See also 2.9.5.2. For each DCT: 5. Program D18F2x94_dct[1:0][MemClkFreqVal] = 1. • Wait for D18F2x94_dct[1:0][FreqChangeInProg] = 0. For each DCT: 6. Program D18F2x9C_x0D0F_E006_dct[1:0][PllLockTime] according to Table 21. BIOS must observe the following requirements: • BIOS must not change the PLL frequency after DRAM has exited from self-refresh. • BIOS must not change the PLL frequency after the DRAM training for DDR3 DIMMs is complete. 93 42300 Rev 3.12 - July 14, 2015 BKDG for AMD Family 15h Models 10h-1Fh Processors Table 21: DDR PLL Lock Time D18F2xA8_dct[1:0][MemPhyPllPdMode] 00b 10b 2.9.5.4.2.1 PllLockTime 0Fh 190h Requirements for DRAM Frequency Change During Training During DRAM training, BIOS may be required to change the DRAM(MEMCLK) frequency. The steps below describe what is required to prepare the processor and memory subsystem for the new MEMCLK frequency. It is assumed that the memory subsystem has previously been initialized at the current MEMCLK frequency, and this procedure describes only the steps that must be repeated at the new MEMCLK frequency. See 2.9.5.9.1 [Write Levelization Training] and 2.9.5.9.2 [DQS Receiver Enable Training]. 1. Force the NB P-state. See 2.9.5.2. 2. Enter self-refresh: A. Program D18F2x90_dct[1:0][DisDllShutDownSR] = 1. B. Program D18F2x90_dct[1:0][EnterSelfRef] = 1. C. Wait for D18F2x90_dct[1:0][EnterSelfRef] = 0. 3. DRAM channel frequency change. See 2.9.5.4.2. 4. Change NCLK frequency to meet NCLK-MEMCLK ratio requirements. See 2.5.3.1.6. 5. Phy fence programming. See 2.9.5.4.3. 6. Phy compensation initialization. See 2.9.5.4.4. 7. Program SPD configuration. See 2.9.5.5. 8. Program Non-SPD configuration. See 2.9.5.6. 9. Exit self-refresh: A. Program D18F2x90_dct[1:0][ExitSelfRef] = 1. B. Wait for D18F2x90_dct[1:0][ExitSelfRef] = 0. C. Program D18F2x90_dct[1:0][DisDllShutDownSR] = 0. 10. Re-program devices with frequency dependent mode register field values. See 2.9.5.8. 2.9.5.4.3 Phy Fence Programming The DDR phy fence logic is used to adjust the phase relationship between the data fifo and the data going to the pad. After any MEMCLK frequency change and before any memory training, BIOS must perform phy fence training for each channel using the following steps: 1. 2. 3. 4. 5. 6. 7. 8. 9. 10. 11. 12. 13. 14. Program D18F2x9C_x0D0F_0[F,7:0]31_dct[1:0] = 0000_0000h. Program D18F2x9C_x0D0F_E019_dct[1:0] = 0000_0000h. Program D18F2x9C_x0000_0008_dct[1:0]_mp[1:0][FenceTrSel]=10b. Program D18F2x9C_x0000_00[51:50]_dct[1:0]=1313_1313h. Perform phy fence training. See 2.9.5.4.3.1 [Phy Fence Training]. Write the calculated fence value to D18F2x9C_x0000_000C_dct[1:0][FenceThresholdTxDll]. Program D18F2x9C_x0000_0008_dct[1:0]_mp[1:0][FenceTrSel]=01b. Program D18F2x9C_x0000_00[51:50]_dct[1:0]=1313_1313h. Perform phy fence training. See 2.9.5.4.3.1 [Phy Fence Training]. Write the calculated fence value to D18F2x9C_x0000_000C_dct[1:0][FenceThresholdRxDll]. Program D18F2x9C_x0000_0008_dct[1:0]_mp[1:0][FenceTrSel]=11b. Program D18F2x9C_x0000_00[51:50]_dct[1:0]=1313_1313h. Perform phy fence training. See 2.9.5.4.3.1 [Phy Fence Training]. Write the calculated fence value to D18F2x9C_x0000_000C_dct[1:0][FenceThresholdTxPad]. 94 42300 Rev 3.12 - July 14, 2015 BKDG for AMD Family 15h Models 10h-1Fh Processors 15. Program Fence2 threshold for data as follows: A. IF (D18F2x9C_x0000_000C_dct[1:0][FenceThresholdTxPad] < 16) THEN Fence2_TxPad[4:0] = {1b, D18F2x9C_x0000_000C_dct[1:0][19:16]} ELSE Fence2_TxPad[4:0] = 00000b ENDIF. B. IF (D18F2x9C_x0000_000C_dct[1:0][FenceThresholdRxDll] < 16) THEN Fence2_RxDll[4:0] = {1b, D18F2x9C_x0000_000C_dct[1:0][24:21]} ELSE Fence2_RxDll[4:0] = 00000b ENDIF. C. IF (D18F2x9C_x0000_000C_dct[1:0][FenceThresholdTxDll] < 16) THEN Fence2_TxDll[4:0] = {1b, D18F2x9C_x0000_000C_dct[1:0][29:26]} ELSE Fence2_TxDll[4:0] = 00000b ENDIF. D. Program D18F2x9C_x0D0F_0[F,7:0]31_dct[1:0] = {000000b, Fence2_TxDll[4:0], Fence2_TxPad[4:0]}. E. Program D18F2x9C_x0D0F_E019_dct[1:0] = {0b, Fence2_RxDll[4:0], 00000b, Fence2_TxPad[4:0]}. 16. Reprogram D18F2x9C_x0000_0004_dct[1:0]_mp[1:0]. 17. If motherboard routing requires CS[7:6] to adopt address timings, BIOS performs the following: A. Program D18F2xA8_dct[1:0][CSTimingMux67] = 1. B. Program D18F2x9C_x0D0F_8021_dct[1:0]_mp[1:0]: • DiffTimingEn = 1. • IF (D18F2x9C_x0000_0004_dct[1:0]_mp[1:0][AddrCmdFineDelay] >= D18F2x9C_x0D04_E008_dct[1:0][FenceValue]) THEN Fence = 1 ELSE Fence = 0. • Delay = D18F2x9C_x0000_0004_dct[1:0]_mp[1:0][AddrCmdFineDelay]. 18. Reprogram D18F2x9C_x0000_0004_dct[1:0]_mp[1:0]. When resuming from S3, BIOS reprograms D18F2x9C_x0000_000C_dct[1:0][FenceThresholdTxDll, FenceThresholdRxDll, FenceThresholdTxPad], D18F2x9C_x0D0F_E019_dct[1:0], and D18F2x9C_x0D0F_8021_dct[1:0]_mp[1:0] from values stored in non-volatile storage instead of training. BIOS reprograms D18F2x9C_x0D0F_0[F,7:0]31_dct[1:0] and D18F2x9C_x0D0F_E019_dct[1:0] as indicated in the steps above. BIOS may use D18F2x9C_x0000_000C_dct[1:0], D18F2x9C_x0D0F_4003_dct[1], and D18F2x9C_x0D0F_4003_dct[1] for local storage of fence values during the intermediate time between training and writing the values into non-volatile storage in preparation for S3. 2.9.5.4.3.1 1. 2. 3. 4. 5. Phy Fence Training Program D18F2x9C_x0000_0008_dct[1:0]_mp[1:0][PhyFenceTrEn]=1. Wait 2000 MEMCLKs. Program D18F2x9C_x0000_0008_dct[1:0]_mp[1:0][PhyFenceTrEn]=0. Read the phase recovery engine registers D18F2x9C_x0000_00[51:50]_dct[1:0]. Calculate the average of the fine delay values of all byte lanes and subtract 8. If the result is negative then the fence value is zero. 95 42300 Rev 3.12 - July 14, 2015 2.9.5.4.4 BKDG for AMD Family 15h Models 10h-1Fh Processors Phy Compensation Initialization Each DDR IO driver has a programmable slew rate controlled by the pre-driver calibration code. The recommended slew rate is a function of the DC drive strength. BIOS initializes the recommended nominal slew rate values as follows: 1. Program D18F2x9C_x0000_0008_dct[1:0]_mp[1:0][DisAutoComp] = 1. 2. Program D18F2x9C_x0000_0008_dct[1:0]_mp[1:0][DisablePredriverCal]=1. 3. Program TxPreP/TxPreN for Data and DQS according toTable 22, Table 23, or Table 24. A. Program D18F2x9C_x0D0F_0[F,7:0]0[A,6]_dct[1:0]={0000b, TxPreP, TxPreN}. B. Program D18F2x9C_x0D0F_0[F,7:0]02_dct[1:0]={1000b, TxPreP, TxPreN}. 4. Program TxPreP/TxPreN for Cmd/Addr according to Table 25, Table 26, or Table 27. A. Program D18F2x9C_x0D0F_[C,8][1:0][12,0E,0A,06]_dct[1:0]={0000b, TxPreP, TxPreN}. B. Program D18F2x9C_x0D0F_[C,8][1:0]02_dct[1:0]={1000b, TxPreP, TxPreN}. 5. Program TxPreP/TxPreN for Clock according to Table 28, Table 29, or Table 30. A. Program D18F2x9C_x0D0F_2[2:0]02_dct[1:0]={1000b, TxPreP, TxPreN}. 6. Program D18F2x9C_x0000_0008_dct[1:0]_mp[1:0][DisAutoComp] = 0. Table 22: Phy predriver calibration codes for Data/DQS at 1.5V 667 - 800 Drive Strength1 000b 100_100b 100_100b 100_100b 100_100b 100_100b 100_100b 1066 - 1333 001b 010b 011b 000b 111_111b 110_110b 111_111b 111_111b 111_111b 110_110b 110_110b 110_110b 1600 - 2133 001b 010b 011b 000b 111_111b 110_110b DDR Rate TxPreP2 TxPreN2 100_100b 100_100b 001b 111_111b 110_110b 010b 111_111b 110_110b 011b 111_111b 110_110b 1. IF (D18F2x9C_x0D0F_0[F,8:0]06) THEN See D18F2x9C_x0000_0000_dct[1:0]_mp[1:0][DqsDrvStren] ELSE See D18F2x9C_x0000_0000_dct[1:0]_mp[1:0][DataDrvStren] ENDIF. 2. See D18F2x9C_x0D0F_0[F,7:0]0[A,6]_dct[1:0] and D18F2x9C_x0D0F_0[F,7:0]02_dct[1:0]. 96 42300 Rev 3.12 - July 14, 2015 BKDG for AMD Family 15h Models 10h-1Fh Processors Table 23: Phy predriver calibration codes for Data/DQS at 1.35V 667 - 800 Drive Strength1 000b 101_101b 101_101b 100_100b 101_101b 101_101b 100_100b 1066 - 1333 001b 010b 011b 000b 111_111b 110_110b 111_111b 111_111b 111_111b 110_110b 110_110b 110_110b 1600 - 1866 001b 010b 011b 000b 111_111b 110_110b DDR Rate TxPreP2 TxPreN2 111_111b 110_110b 001b 111_111b 110_110b 010b 111_111b 110_110b 011b 111_111b 110_110b 1. IF (D18F2x9C_x0D0F_0[F,8:0]06) THEN See D18F2x9C_x0000_0000_dct[1:0]_mp[1:0][DqsDrvStren] ELSE See D18F2x9C_x0000_0000_dct[1:0]_mp[1:0][DataDrvStren] ENDIF. 2. See D18F2x9C_x0D0F_0[F,7:0]0[A,6]_dct[1:0] and D18F2x9C_x0D0F_0[F,7:0]02_dct[1:0]. Table 24: Phy Predriver Calibration Codes for Data/DQS at 1.25V 667 - 800 Drive Strength1 000b 110_110b 110_110b 100_100b 101_101b 101_101b 100_100b 1066 - 1333 001b 010b 011b 000b 111_111b 110_110b 001b 111_111b 110_110b 010b 111_111b 110_110b 011b 111_111b 110_110b DDR Rate TxPreP2 TxPreN2 111_111b 110_110b 97 42300 Rev 3.12 - July 14, 2015 BKDG for AMD Family 15h Models 10h-1Fh Processors Table 24: Phy Predriver Calibration Codes for Data/DQS at 1.25V DDR Rate 1600 - 1866 Drive Strength1 000b TxPreP2 TxPreN2 111_111b 110_110b 001b 111_111b 110_110b 010b 111_111b 110_110b 011b 111_111b 110_110b 1. IF (D18F2x9C_x0D0F_0[F,8:0]06) THEN See D18F2x9C_x0000_0000_dct[1:0]_mp[1:0][DqsDrvStren] ELSE See D18F2x9C_x0000_0000_dct[1:0]_mp[1:0][DataDrvStren] ENDIF. 2. See D18F2x9C_x0D0F_0[F,7:0]0[A,6]_dct[1:0] and D18F2x9C_x0D0F_0[F,7:0]02_dct[1:0]. Table 25: Phy predriver calibration codes for Cmd/Addr at 1.5V 667 - 800 Drive Strength1 000b 010_010b 010_010b 010_010b 010_010b 010_010b 010_010b 1066 - 1333 001b 010b 011b 000b 011_011b 011_011b 011_011b 011_011b 011_011b 011_011b 011_011b 011_011b 1600 - 1866 001b 010b 011b 000b 101_101b 101_101b DDR Rate TxPreP2 TxPreN2 010_010b 010_010b 001b 101_101b 101_101b 010b 101_101b 101_101b 011b 101_101b 101_101b 1. IF (D18F2x9C_x0D0F_C002)THEN See D18F2x9C_x0000_0000_dct[1:0]_mp[1:0][CkeDrvStren] ELSEIF (D18F2x9C_x0D0F_800[A,6,2])THEN See D18F2x9C_x0000_0000_dct[1:0]_mp[1:0][CsOdtDrvStren] ELSE See D18F2x9C_x0000_0000_dct[1:0]_mp[1:0][AddrCmdDrvStren] ENDIF. 2. See D18F2x9C_x0D0F_[C,8][1:0][12,0E,0A,06]_dct[1:0] and D18F2x9C_x0D0F_[C,8][1:0]02_dct[1:0]. 98 42300 Rev 3.12 - July 14, 2015 BKDG for AMD Family 15h Models 10h-1Fh Processors Table 26: Phy predriver calibration codes for Cmd/Addr at 1.35V 667 - 800 Drive Strength1 000b 1066 - 1333 001b 010b 011b 000b 1600 - 1866 001b 010b 011b 000b DDR Rate TxPreP2 TxPreN2 010_010b 010_010b 010_010b 010_010b 010_010b 010_010b 010_010b 010_010b 100_100b 100_100b 011_011b 011_011b 011_011b 011_011b 011_011b 011_011b 101_101b 101_101b 001b 101_101b 101_101b 010b 100_100b 100_100b 011b 100_100b 100_100b 1. IF (D18F2x9C_x0D0F_C002)THEN See D18F2x9C_x0000_0000_dct[1:0]_mp[1:0][CkeDrvStren] ELSEIF (D18F2x9C_x0D0F_800[A,6,2])THEN See D18F2x9C_x0000_0000_dct[1:0]_mp[1:0][CsOdtDrvStren] ELSE See D18F2x9C_x0000_0000_dct[1:0]_mp[1:0][AddrCmdDrvStren] ENDIF. 2. See D18F2x9C_x0D0F_[C,8][1:0][12,0E,0A,06]_dct[1:0] and D18F2x9C_x0D0F_[C,8][1:0]02_dct[1:0]. Table 27: Phy Predriver Calibration Codes for Cmd/Addr at 1.25V DDR Rate 667- 800 1066 - 1333 Drive Strength1 000b TxPreP2 TxPreN2 010_010b 010_010b 001b 010_010b 010_010b 010b 010_010b 010_010b 011b 010_010b 010_010b 000b 110_110b 101_101b 001b 010b 011b 100_100b 011_011b 010_010b 100_100b 011_011b 010_010b 99 42300 Rev 3.12 - July 14, 2015 BKDG for AMD Family 15h Models 10h-1Fh Processors Table 27: Phy Predriver Calibration Codes for Cmd/Addr at 1.25V DDR Rate 1600 - 1866 Drive Strength1 000b TxPreP2 TxPreN2 111_111b 110_110b 001b 110_110b 101_101b 010b 101_101b 100_100b 011b 101_101b 100_100b 1. IF (D18F2x9C_x0D0F_C002)THEN See D18F2x9C_x0000_0000_dct[1:0]_mp[1:0][CkeDrvStren] ELSEIF (D18F2x9C_x0D0F_800[A,6,2])THEN See D18F2x9C_x0000_0000_dct[1:0]_mp[1:0][CsOdtDrvStren] ELSE See D18F2x9C_x0000_0000_dct[1:0]_mp[1:0][AddrCmdDrvStren] ENDIF. 2. See D18F2x9C_x0D0F_[C,8][1:0][12,0E,0A,06]_dct[1:0] and D18F2x9C_x0D0F_[C,8][1:0]02_dct[1:0]. Table 28: Phy predriver calibration codes for Clock at 1.5V 667 - 800 Drive Strength1 000b 100_100b 100_100b 100_100b 100_100b 100_100b 100_100b 1066 - 1333 001b 010b 011b 000b 111_111b 110_110b 111_111b 111_111b 101_101b 110_110b 110_110b 101_101b 1600 - 1866 001b 010b 011b 000b 111_111b 110_110b 001b 111_111b 010b 111_111b 110_110b 110_110b DDR Rate TxPreP2 TxPreN2 100_100b 100_100b 011b 111_111b 110_110b 1. See D18F2x9C_x0000_0000_dct[1:0]_mp[1:0][ClkDrvStren]. 2. See D18F2x9C_x0D0F_2[2:0]02_dct[1:0]. Table 29: Phy predriver calibration codes for Clock at 1.35V DDR Rate 667 - 800 Drive Strength1 000b TxPreP2 TxPreN2 110_110b 101_101b 001b 010b 011b 110_110b 100_100b 100_100b 101_101b 100_100b 100_100b 100 42300 Rev 3.12 - July 14, 2015 BKDG for AMD Family 15h Models 10h-1Fh Processors Table 29: Phy predriver calibration codes for Clock at 1.35V 1066 - 1333 Drive Strength1 000b 111_111b 111_111b 110_110b 110_110b 110_110b 101_101b 1600 - 1866 001b 010b 011b 000b 111_111b 110_110b 001b 111_111b 010b 111_111b 110_110b 110_110b DDR Rate TxPreP2 TxPreN2 111_111b 110_110b 011b 110_110b 101_101b 1. See D18F2x9C_x0000_0000_dct[1:0]_mp[1:0][ClkDrvStren]. 2. See D18F2x9C_x0D0F_2[2:0]02_dct[1:0]. Table 30: Phy Predriver Calibration Codes for Clock at 1.25V 667 - 800 Drive Strength1 000b 110_110b 100_100b 100_100b 101_101b 100_100b 100_100b 1066 - 1333 001b 010b 011b 000b 111_111b 110_110b 111_111b 111_111b 111_111b 110_110b 110_110b 110_110b 1600 - 1866 001b 010b 011b 000b 111_111b 110_110b 001b 111_111b 110_110b 010b 111_111b 110_110b 011b 111_111b 110_110b DDR Rate TxPreP2 TxPreN2 110_110b 101_101b 1. See D18F2x9C_x0000_0000_dct[1:0]_mp[1:0][ClkDrvStren]. 2. See D18F2x9C_x0D0F_2[2:0]02_dct[1:0]. 2.9.5.5 SPD ROM-Based Configuration The Serial Presence Detect (SPD) ROM is a non-volatile memory device on the DIMM encoded by the DIMM manufacturer. The description of the SPD is usually provided on a data sheet for the DIMM itself along with data describing the memory devices used. The data describes configuration and speed characteristics of the DIMM and the SDRAM components mounted on the DIMM. The associated data sheet also contains the DIMM byte values that are encoded in the SPD on the DIMM. BIOS reads the values encoded in the SPD ROM through a system-specific interface. BIOS acquires DIMM configuration information, such as the amount of memory on each DIMM, from the SPD ROM on each DIMM and uses this information to program the DRAM controller registers. 101 42300 Rev 3.12 - July 14, 2015 BKDG for AMD Family 15h Models 10h-1Fh Processors For solder-down DRAM, in the absence of an SPD ROM, BIOS provides the information necessary for DRAM configuration. For convenience, this document may refer to solder-down DRAM as a DIMM, and to DRAM data sheet or Jedec DRAM specifications as SPD ROM based, unless otherwise noted. The SPD ROM provides values for several DRAM timing parameters that are required by the DCT. In general, BIOS should use the optimal value specified by the SPD ROM. These parameters are: • • • • • • • • • • • • D18F2x22C_dct[1:0]_mp[1:0][Twr]: Write recovery time D18F2x8C_dct[1:0][Tref] D18F2x200_dct[1:0]_mp[1:0][Tras]: Active to precharge time D18F2x200_dct[1:0]_mp[1:0][Trp]: Precharge time D18F2x200_dct[1:0]_mp[1:0][Trcd]: RAS to CAS delay D18F2x200_dct[1:0]_mp[1:0][Tcl]: CAS latency D18F2x204_dct[1:0]_mp[1:0][Trtp]: Internal read to precharge command delay time D18F2x204_dct[1:0]_mp[1:0][FourActWindow]: Four activate window delay time D18F2x204_dct[1:0]_mp[1:0][Trrd]: Row active to row active delay D18F2x204_dct[1:0]_mp[1:0][Trc]: Active to active/refresh time D18F2x208_dct[1:0][Trfc3, Trfc2, Trfc1, Trfc0]: Refresh recovery delay time D18F2x20C_dct[1:0]_mp[1:0][Twtr]: Internal write to read command delay time Optimal cycle time is specified for each DIMM and is used to limit or determine bus frequency. See 2.9.5.8 [DRAM Device and Controller Initialization]. The DRAM data sheet or the Jedec DRAM specification provides values for some DRAM timing parameters that are required by the DCT, regardless of whether the DRAMs are solder-down. Most of these parameters have a fixed time component as part of their specification. If memory P-states are supported (See 2.5.7.1 [Memory P-states]) and a timing register does not have per memory P-state contexts, BIOS must evaluate the minimum value for each frequency and choose the most pessimistic value for all frequencies to program into the register. These parameters are: • • • • • • • • • • • D18F2x20C_dct[1:0]_mp[1:0][Tcwl]: CAS write latency D18F2x220_dct[1:0][Tmrd]: MRS command cycle time D18F2x220_dct[1:0][Tmod]: MRS command recovery time D18F2x224_dct[1:0][Tzqcs]: Short calibration command recovery time D18F2x224_dct[1:0][Tzqoper]: Long calibration command recovery time D18F2x248_dct[1:0]_mp[1:0][Txpdll]: Exit precharge power down (with DLL frozen) to command delay. D18F2x248_dct[1:0]_mp[1:0][Txp]: Exit precharge power down to command delay. D18F2x24C_dct[1:0][Tcksrx]: Clock stable to self refresh exit delay D18F2x24C_dct[1:0][Tcksre]: Self refresh entry to clock removal delay D18F2x24C_dct[1:0][Tckesr]: Self refresh entry to command delay D18F2x24C_dct[1:0][Tpd]: Power down entry to exit delay. BIOS may increase this value above the Jedec minimum for certain power down modes. 2.9.5.6 Non-SPD ROM-Based Configuration There are several DRAM timing parameters and DCT configurations that need to be programmed for optimal memory performance. These values are not derived from the SPD ROM. Several of these timing parameters are functions of other configuration values. These interdependencies must be considered when programming values into several DCT register timing fields. The factors to consider when specifying a value for a specific non-SPD timing parameter are: • Mixed or non-mixed DIMMs (x4 with x8). 102 42300 Rev 3.12 - July 14, 2015 • • • • • • BKDG for AMD Family 15h Models 10h-1Fh Processors Training delay values. See 2.9.5.9 [DRAM Training]. Read and write latency differences. The phy's idle clock requirements on the data bus. DDR3 ODT timing requirements. NCLK frequency MEMCLK frequency The following sub-sections describe how BIOS programs each non-SPD related timing field to a recommended minimum timing value with respect to the above factors. The following terms are defined to simplify calculations and are calculated in MEMCLKs: • Latency Difference (LD) = D18F2x200_dct[1:0]_mp[1:0][Tcl] - D18F2x20C_dct[1:0]_mp[1:0][Tcwl]. • Read ODT Delay (ROD) = MAX(0, D18F2x240_dct[1:0]_mp[1:0][RdOdtOnDuration] - 6) + MAX(0, D18F2x240_dct[1:0]_mp[1:0][RdOdtTrnOnDly] - LD). • Write ODT Delay (WOD) = MAX(0, D18F2x240_dct[1:0]_mp[1:0][WrOdtOnDuration] - 6). • WrEarly = ABS(D18F2x20C_dct[1:0]_mp[1:0][WrDqDqsEarly]) /2 2.9.5.6.1 TrdrdSdSc, TrdrdSdDd, and TrdrdDd (Read to Read Timing) The optimal values for D18F2x218_dct[1:0]_mp[1:0][TrdrdSdSc, TrdrdSdDc, TrdrdDd] are platform and configuration specific and should be characterized for best performance. Prior to DRAM training, BIOS should program these parameters to the largest defined value. After DRAM training, BIOS should use the guidelines below to configure the recommended platform generic timing values: • TrdrdSdSc (in MEMCLKs) = 1. • TrdrdSdDc (in MEMCLKs) = MAX(TrdrdSdSc, 3 + (IF (D18F2xA8_dct[1:0][PerRankTimingEn]) THEN CEIL(CDDTrdrdSdDc / 2 ) ELSE 0 ENDIF)). • TrdrdDd (in MEMCLKs) = MAX(TrdrdSdDc, CEIL(MAX(ROD + 3, CDDTrdrdDd/2 + 3.5))). The Critical Delay Difference (CDD) is the largest delay difference of the channel. • Each delay difference is D18F2x9C_x0000_00[2A:10]_dct[1:0]_mp[1:0][DqsRcvEnGrossDelay] minus D18F2x9C_x0000_00[2A:10]_dct[1:0]_mp[1:0][DqsRcvEnGrossDelay]. • For CDDTrdrdSdDc, the subtraction terms are the delays of different chip selects within the same DIMM within the same byte lane. • For CDDTrdrdDd, the subtraction terms are the delays of different DIMMs within the same byte lane. 2.9.5.6.2 TwrwrSdSc, TwrwrSdDc, TwrwrDd (Write to Write Timing) The optimal values for D18F2x214_dct[1:0]_mp[1:0][TwrwrSdSc, TwrwrSdDc, TwrwrDd] are platform and configuration specific and should be characterized for best performance. Prior to DRAM training, BIOS should program these parameters to the largest defined value. After DRAM training, BIOS should use the guidelines below to configure the recommended platform generic timing values: • TwrwrSdSc (in MEMCLKs) = 1. • TwrwrSdDc (in MEMCLKs) = CEIL(MAX(WOD + 3, CDDTwrwrSdDc / 2 + (IF (D18F2xA8_dct[1:0][PerRankTimingEn]) THEN 3.5 ELSE 3 ENDIF))). • TwrwrDd (in MEMCLKs) = CEIL(MAX(WOD + 3, CDDTwrwrDd / 2 + 3.5)). The Critical Delay Difference (CDD) is the largest delay difference of the channel. • Each delay difference is D18F2x9C_x0000_00[4A:30]_dct[1:0]_mp[1:0][WrDqsGrossDly] minus D18F2x9C_x0000_00[4A:30]_dct[1:0]_mp[1:0][WrDqsGrossDly]. • For CDDTwrwrSdDc, the subtraction terms are the delays of different chip selects within the same DIMM 103 42300 Rev 3.12 - July 14, 2015 BKDG for AMD Family 15h Models 10h-1Fh Processors within the same byte lane. • For CDDTwrwrDd, the subtraction terms are the delays of different DIMMs within the same byte lane. BIOS must program these parameters as follows: TwrwrSdSc <= TwrwrSdDc <= TwrwrDd. 2.9.5.6.3 Twrrd (Write to Read DIMM Termination Turn-around) The optimal value for D18F2x218_dct[1:0]_mp[1:0][Twrrd] is platform and configuration specific and should be characterized for best performance. Prior to DRAM training, BIOS should program these parameters to the largest defined value. After DRAM training, BIOS should use the guidelines below to configure the recommended platform generic timing values: • Twrrd (in MEMCLKs) = MAX(1, CEIL(MAX(WOD, CDDTwrrd / 2 + 0.5 - WrEarly) - LD + 3)). The Critical Delay Difference (CDD) is the largest delay difference of the channel. • Each delay difference is D18F2x9C_x0000_00[4A:30]_dct[1:0]_mp[1:0][WrDqsGrossDly] minus D18F2x9C_x0000_00[2A:10]_dct[1:0]_mp[1:0][DqsRcvEnGrossDelay]. • For CDDTwrrd, the subtraction terms are the delays of different chip selects within the same byte lane. 2.9.5.6.4 TrwtTO (Read-to-Write Turnaround for Data, DQS Contention) The optimal value for D18F2x21C_dct[1:0]_mp[1:0][TrwtTO] is platform and configuration specific and should be characterized for best performance. Prior to DRAM training, BIOS should program this parameter to the largest defined value. After DRAM training, BIOS should use the guidelines below to configure the recommended platform generic timing values after DDR training is complete: • TrwtTO (in MEMCLKs) = CEIL(MAX(ROD, CDDTrwtTO / 2 - 0.5 + WrEarly) + LD + 3). • If 1 DIMM/ch, substitute ROD = 0 in the above equation. The Critical Delay Difference (CDD) is the largest delay difference of the channel. • Each delay difference is D18F2x9C_x0000_00[2A:10]_dct[1:0]_mp[1:0][DqsRcvEnGrossDelay] minus D18F2x9C_x0000_00[4A:30]_dct[1:0]_mp[1:0][WrDqsGrossDly]. • For CDDTrwtTO, the subtraction terms are the delays of all chip selects within the same byte lane. 2.9.5.6.5 DRAM ODT Control The following describes the general ODT behavior for various system configurations. In all cases, the processor ODT is off for writes and is on for reads: • For one single or dual rank DIMM on a channel: • For writes, the ODT is on for the target rank. • For reads, the ODT is off for all ranks. • For two single or dual rank DIMMs on a channel: • For writes, the ODT is on for the target rank of the target DIMM and also on for the first rank of the nontarget DIMM. • For reads, the ODT is on for the first rank of the non-target DIMM. BIOS configures the DIMM ODT behavior per chip select according to the DIMM population. In all cases, the processor ODT is off for writes and is on for reads. The ODT pin patterns for reads and writes are programmed using D18F2x[234:230]_dct[1:0] and D18F2x[23C:238]_dct[1:0], respectively. BIOS also configures the ODT pin pattern used during write leveling by programming 104 42300 Rev 3.12 - July 14, 2015 BKDG for AMD Family 15h Models 10h-1Fh Processors D18F2x9C_x0000_0008_dct[1:0]_mp[1:0][WrLvOdt]. BIOS programs WrLvOdt with the D18F2x[23C:238]_dct[1:0] value provided for writes to the rank targeted by training. See 2.9.5.9.1 [Write Levelization Training]. Table 31: DDR3 DIMM ODT Pattern D18F2x[234:230]_dct[1:0] D18F2x[23C:238]_dct[1:0] D18F2x230 D18F2x234 D18F2x238 D18F2x23C SR 0000_0000h 0000_0000h 0004_0000h 0000_0000h DR 0000_0000h 0000_0000h 0804_0000h 0000_0000h SR 0000_0000h 0000_0000h 0000_0001h 0000_0000h DR 0000_0000h 0000_0000h 0000_0201h 0000_0000h SR/DR SR/DR 0101_0404h 0000_0000h 0905_0605h 0000_0000h 1. DIMM0: MEMCSx[1:0], MEMODTx[1,0]. DIMM1: MEMCSx[3:2], MEMODTx[3,2]. Population restrictions may apply. See 2.9.5.6.6 for details. DIMM01 DIMM11 2.9.5.6.6 DRAM Address Timing and Output Driver Compensation Control This section describes the settings required for programming the timing on the address pins, the CS/ODT pins, and the CKE pins. The following tables document the address timing and output driver settings on a per channel basis for DDR3 DIMM types. DIMM0 is the DIMM closest to the processor on that channel and DIMM1 is the DIMM farthest from the processor on that channel. DIMMs must be populated from farthest slot to closest slot to the processor on a per channel basis . Populations that are not shown in these tables are not supported. These tables document the optimal settings for motherboards which meet the relevant motherboard design guidelines. • The presence of settings for higher speeds does not constitute support for those speeds. See 2.9 [DRAM Controllers (DCTs)] for an overview of the DIMM and memory bus speed support. Table 32: BIOS recommendations for UDIMM address timings and output driver control for FM2 or FS1r2 Package D18F2x9C_x0000_0000_dct[1:0]_mp[1:0] DIMM1 D18F2x9C_x0000_0004_dct[1:0]_mp[1:0] DIMM0 RTT_Wr VDDIO SlowAccessMode NumDimmSlots DdrRate RTT_Nom D18F2x94_dct[1:0] Condition 1 667 1.25, 1.35, 1.5 SR - 0 120 Off 00000000h 00112222h 1 667 1.25, 1.35, 1.5 DR - 0 120 Off 003B0000h 00112222h 1 800 1.25, 1.35, 1.5 SR - 0 120 Off 00000000h 00112222h 105 42300 Rev 3.12 - July 14, 2015 BKDG for AMD Family 15h Models 10h-1Fh Processors Table 32: BIOS recommendations for UDIMM address timings and output driver control for FM2 or FS1r2 Package D18F2x9C_x0000_0000_dct[1:0]_mp[1:0] DIMM1 D18F2x9C_x0000_0004_dct[1:0]_mp[1:0] DIMM0 RTT_Wr VDDIO SlowAccessMode NumDimmSlots DdrRate RTT_Nom D18F2x94_dct[1:0] Condition 1 800 1.25, 1.35, 1.5 DR - 0 120 Off 003B0000h 00112222h 1 1066 1.25, 1.35, 1.5 SR - 0 120 Off 00000000h 10112222h 1 1066 1.25, 1.35, 1.5 DR - 0 120 Off 00380000h 10112222h 1 1333 1.25, 1.35, 1.5 SR - 0 60 Off 00000000h 20112222h 1 1333 1.25, 1.35, 1.5 DR - 0 60 Off 00360000h 20112222h 1 1600 1.35, 1.5 SR - 0 60 Off 00000000h 30112222h 1 1600 1.35, 1.5 DR - 1 40 Off 00000000h 30112222h 1 1866 1.5 SR - 0 40 Off 00000000h 30112222h 1 1866 1.5 DR - 1 40 Off 00000000h 30112222h 1 2133 1.5 SR - 0 40 Off 00000000h 30112222h 1 2133 1.5 DR - 1 40 Off 00000000h 30112222h 2 667 1.25, 1.35, 1.5 NP SR 0 120 Off 00000000h 00112222h 2 667 1.25, 1.35, 1.5 NP DR 0 120 Off 003B0000h 00112222h 2 667 1.25, 1.35, 1.5 SR SR 0 40 120 00390039h 10222322h 2 667 1.25, 1.35, 1.5 DR DR 0 40 120 00390039h 10222322h 2 667 1.25, 1.35, 1.5 DR SR 0 40 120 00390039h 10222322h 2 667 1.25, 1.35, 1.5 SR DR 0 40 120 00390039h 10222322h 2 800 1.25, 1.35, 1.5 NP SR 0 120 Off 00000000h 00112222h 2 800 1.25, 1.35, 1.5 NP DR 0 120 Off 003B0000h 00112222h 2 800 1.25, 1.35, 1.5 SR SR 0 40 120 00390039h 20222322h 2 800 1.25, 1.35, 1.5 DR DR 0 40 120 00390039h 20222322h 2 800 1.25, 1.35, 1.5 DR SR 0 40 120 00390039h 20222322h 2 800 1.25, 1.35, 1.5 SR DR 0 40 120 00390039h 20222322h 2 1066 1.25, 1.35, 1.5 NP SR 0 120 Off 00000000h 10112222h 2 1066 1.25, 1.35, 1.5 NP DR 0 120 Off 00380000h 10112222h 2 1066 1.25, 1.35, 1.5 SR SR 0 40 120 00350037h 30222322h 106 42300 Rev 3.12 - July 14, 2015 BKDG for AMD Family 15h Models 10h-1Fh Processors Table 32: BIOS recommendations for UDIMM address timings and output driver control for FM2 or FS1r2 Package D18F2x9C_x0000_0000_dct[1:0]_mp[1:0] DIMM1 D18F2x9C_x0000_0004_dct[1:0]_mp[1:0] DIMM0 RTT_Wr VDDIO SlowAccessMode NumDimmSlots DdrRate RTT_Nom D18F2x94_dct[1:0] Condition 2 1066 1.25, 1.35, 1.5 DR DR 0 40 120 00350037h 30222322h 2 1066 1.25, 1.35, 1.5 DR SR 0 40 120 00350037h 30222322h 2 1066 1.25, 1.35, 1.5 SR DR 0 40 120 00350037h 30222322h 2 1333 1.25, 1.35, 1.5 NP SR 0 60 Off 00000000h 20112222h 2 1333 1.25, 1.35, 1.5 NP DR 0 60 Off 00360000h 20112222h 2 1333 1.25, 1.35, 1.5 SR SR 1 30 120 00000035h 30222322h 2 1333 1.25, 1.35, 1.5 DR DR 1 30 120 00000035h 30222322h 2 1333 1.25, 1.35, 1.5 DR SR 1 30 120 00000035h 30222322h 2 1333 1.25, 1.35, 1.5 SR DR 1 30 120 00000035h 30222322h 2 1600 1.35, 1.5 NP SR 0 40 Off 00000000h 30112222h 2 1600 1.35, 1.5 NP DR 1 40 Off 00000000h 30112222h 2 1600 1.35, 1.5 SR SR 1 20 60 0000002Bh 30222322h 2 1600 1.35 DR DR 1 20 60 0000002Bh 30222322h 2 1600 1.35 DR SR 1 20 60 0000002Bh 30222322h 2 1600 1.35 SR DR 1 20 60 0000002Bh 30222322h 2 1600 1.5 DR DR 1 20 60 0000002Bh 30222322h 2 1600 1.5 DR SR 1 20 60 0000002Bh 30222322h 2 1600 1.5 SR DR 1 20 60 0000002Bh 30222322h 2 1866 1.5 NP SR 0 40 Off 00000000h 30112222h 2 1866 1.5 NP DR 1 40 Off 00000000h 30112222h 2 1866 1.5 SR SR 1 20 60 00000031h 30222322h 2 1866 1.5 DR DR 1 20 60 00000031h 30222322h 2 1866 1.5 DR SR 1 20 60 00000031h 30222322h 2 1866 1.5 SR DR 1 20 60 00000031h 30222322h 2 2133 1.5 NP SR 0 40 Off 00000000h 30112222h 2 2133 1.5 NP DR 1 40 Off 00000000h 30112222h 107 42300 Rev 3.12 - July 14, 2015 BKDG for AMD Family 15h Models 10h-1Fh Processors Table 33: BIOS recommendations for UDIMM address timings and output driver control for FP2 Package D18F2x9C_x0000_0000_dct[1:0]_mp[1:0] DIMM1 D18F2x9C_x0000_0004_dct[1:0]_mp[1:0] DIMM0 RTT_Wr VDDIO SlowAccessMode NumDimmSlots DdrRate RTT_Nom D18F2x94_dct[1:0] Condition 1 667 1.25, 1.35, 1.5 SR - 0 120 Off 00000000h 00112222h 1 667 1.25, 1.35, 1.5 DR - 0 120 Off 003B0000h 00112222h 1 800 1.25, 1.35, 1.5 SR - 0 120 Off 00000000h 00112222h 1 800 1.25, 1.35, 1.5 DR - 0 120 Off 003B0000h 00112222h 1 1066 1.25, 1.35, 1.5 SR - 0 120 Off 00000000h 10112222h 1 1066 1.25, 1.35, 1.5 DR - 0 120 Off 00380000h 10112222h 1 1333 1.25, 1.35, 1.5 SR - 0 60 Off 00000000h 20112222h 1 1333 1.25, 1.35, 1.5 DR - 0 60 Off 00360000h 20112222h 2 667 1.25, 1.35, 1.5 NP SR 0 120 Off 00000000h 00112222h 2 667 1.25, 1.35, 1.5 NP DR 0 120 Off 003B0000h 00112222h 2 667 1.25, 1.35, 1.5 SR SR 0 40 120 00390039h 10222322h 2 667 1.25, 1.35, 1.5 DR DR 0 40 120 00390039h 10222322h 2 667 1.25, 1.35, 1.5 DR SR 0 40 120 00390039h 10222322h 2 667 1.25, 1.35, 1.5 SR DR 0 40 120 00390039h 10222322h 2 800 1.25, 1.35, 1.5 NP SR 0 120 Off 00000000h 00112222h 2 800 1.25, 1.35, 1.5 NP DR 0 120 Off 003B0000h 00112222h 2 800 1.25, 1.35, 1.5 SR SR 0 40 120 00390039h 20222322h 2 800 1.25, 1.35, 1.5 DR DR 0 40 120 00390039h 20222322h 2 800 1.25, 1.35, 1.5 DR SR 0 40 120 00390039h 20222322h 2 800 1.25, 1.35, 1.5 SR DR 0 40 120 00390039h 20222322h 2 1066 1.25, 1.35, 1.5 NP SR 0 120 Off 00000000h 10112222h 2 1066 1.25, 1.35, 1.5 NP DR 0 120 Off 00380000h 10112222h 2 1066 1.25, 1.35, 1.5 SR SR 0 40 120 00350037h 30222322h 2 1066 1.25, 1.35, 1.5 DR DR 0 40 120 00350037h 30222322h 2 1066 1.25, 1.35, 1.5 DR SR 0 40 120 00350037h 30222322h 108 42300 Rev 3.12 - July 14, 2015 BKDG for AMD Family 15h Models 10h-1Fh Processors Table 33: BIOS recommendations for UDIMM address timings and output driver control for FP2 Package D18F2x9C_x0000_0000_dct[1:0]_mp[1:0] DIMM1 D18F2x9C_x0000_0004_dct[1:0]_mp[1:0] DIMM0 RTT_Wr VDDIO SlowAccessMode NumDimmSlots DdrRate RTT_Nom D18F2x94_dct[1:0] Condition 2 1066 1.25, 1.35, 1.5 SR DR 0 40 120 00350037h 30222322h 2 1333 1.25, 1.35, 1.5 NP SR 0 60 Off 00000000h 20112222h 2 1333 1.25, 1.35, 1.5 NP DR 0 60 Off 00360000h 20112222h 2 1333 1.25, 1.35, 1.5 SR SR 1 30 120 00000035h 30222322h 2 1333 1.35, 1.5 DR DR 1 30 120 00000035h 30222322h 2 1333 1.35, 1.5 DR SR 1 30 120 00000035h 30222322h 2 1333 1.35, 1.5 SR DR 1 30 120 00000035h 30222322h Table 34: BIOS recommendations for SO-DIMM address timings and output driver control for FM2 or FS1r2 Package D18F2x9C_x0000_0000_dct[1:0]_mp[1:0] DIMM1 D18F2x9C_x0000_0004_dct[1:0]_mp[1:0] DIMM0 RTT_Wr VDDIO SlowAccessMode NumDimmSlots DdrRate RTT_Nom D18F2x94_dct[1:0] Condition 1 667 1.25, 1.35, 1.5 SR - 0 120 Off 00000000h 00002222h 1 667 1.25, 1.35, 1.5 DR - 0 120 Off 00000000h 00002222h 1 800 1.25, 1.35, 1.5 SR - 0 120 Off 00000000h 00002222h 1 800 1.25, 1.35, 1.5 DR - 0 120 Off 00000000h 00002222h 1 1066 1.25, 1.35, 1.5 SR - 0 120 Off 003D3D3Dh 10002222h 109 42300 Rev 3.12 - July 14, 2015 BKDG for AMD Family 15h Models 10h-1Fh Processors Table 34: BIOS recommendations for SO-DIMM address timings and output driver control for FM2 or FS1r2 Package D18F2x9C_x0000_0000_dct[1:0]_mp[1:0] DIMM1 D18F2x9C_x0000_0004_dct[1:0]_mp[1:0] DIMM0 RTT_Wr VDDIO SlowAccessMode NumDimmSlots DdrRate RTT_Nom D18F2x94_dct[1:0] Condition 1 1066 1.25, 1.35, 1.5 DR - 0 120 Off 00000000h 10002222h 1 1333 1.25, 1.35, 1.5 SR - 0 60 Off 003D3D3Dh 20002222h 1 1333 1.25, 1.35, 1.5 DR - 0 60 Off 00003D3Dh 20002222h 1 1600 1.35, 1.5 SR - 0 40 Off 003C3C3Ch 30112222h 1 1600 1.35, 1.5 DR - 1 40 Off 00003C3Ch 30112222h 1 1866 1.5 SR - 0 40 Off 003C3C3Ch 30112222h 1 1866 1.5 DR - 1 40 Off 00003C3Ch 30112222h 2 667 1.25, 1.35, 1.5 NP SR 0 120 Off 00000000h 00002222h 2 667 1.25, 1.35, 1.5 NP DR 0 120 Off 00000000h 00002222h 2 667 1.25, 1.35, 1.5 SR SR 1 40 120 00000039h 10222323h 2 667 1.25, 1.35, 1.5 DR DR 1 40 120 00000039h 10222323h 2 667 1.25, 1.35, 1.5 DR SR 1 40 120 00000039h 10222323h 2 667 1.25, 1.35, 1.5 SR DR 1 40 120 00000039h 10222323h 2 800 1.25, 1.35, 1.5 NP SR 0 120 Off 00000000h 00002222h 2 800 1.25, 1.35, 1.5 NP DR 0 120 Off 00000000h 00002222h 2 800 1.25, 1.35, 1.5 SR SR 1 40 120 00000039h 20222323h 2 800 1.25, 1.35, 1.5 DR DR 1 40 120 00000039h 20222323h 2 800 1.25, 1.35, 1.5 DR SR 1 40 120 00000039h 20222323h 2 800 1.25, 1.35, 1.5 SR DR 1 40 120 00000039h 20222323h 2 1066 1.25, 1.35, 1.5 NP SR 0 120 Off 003D3D3Dh 10002222h 2 1066 1.25, 1.35, 1.5 NP DR 0 120 Off 00000000h 10002222h 2 1066 1.25, 1.35, 1.5 SR SR 1 30 120 00000037h 30222323h 2 1066 1.25, 1.35, 1.5 DR DR 1 30 120 00000037h 30222323h 2 1066 1.25, 1.35, 1.5 DR SR 1 30 120 00000037h 30222323h 2 1066 1.25, 1.35, 1.5 SR DR 1 30 120 00000037h 30222323h 2 1333 1.25, 1.35, 1.5 NP SR 0 60 Off 003D3D3Dh 20002222h 110 42300 Rev 3.12 - July 14, 2015 BKDG for AMD Family 15h Models 10h-1Fh Processors Table 34: BIOS recommendations for SO-DIMM address timings and output driver control for FM2 or FS1r2 Package D18F2x9C_x0000_0000_dct[1:0]_mp[1:0] DIMM1 D18F2x9C_x0000_0004_dct[1:0]_mp[1:0] DIMM0 RTT_Wr VDDIO SlowAccessMode NumDimmSlots DdrRate RTT_Nom D18F2x94_dct[1:0] Condition 2 1333 1.25, 1.35, 1.5 NP DR 0 60 Off 00003D3Dh 20002222h 2 1333 1.25, 1.35, 1.5 SR SR 1 30 120 00000035h 30222323h 2 1333 1.35, 1.5 DR DR 1 30 120 00000035h 30222323h 2 1333 1.35, 1.5 DR SR 1 30 120 00000035h 30222323h 2 1333 1.35, 1.5 SR DR 1 30 120 00000035h 30222323h 2 1600 1.35, 1.5 NP SR 0 40 Off 003C3C3Ch 30112222h 2 1600 1.35, 1.5 NP DR 1 40 Off 00003C3Ch 30112222h 2 1600 1.5 SR SR 1 20 60 00000033h 30222323h 2 1600 1.5 DR DR 1 20 60 00000033h 30222323h 2 1600 1.5 DR SR 1 20 60 00000033h 30222323h 2 1600 1.5 SR DR 1 20 60 00000033h 30222323h 2 1866 1.5 NP SR 0 40 Off 003C3C3Ch 30112222h 2 1866 1.5 NP DR 1 40 Off 00003C3Ch 30112222h 111 42300 Rev 3.12 - July 14, 2015 BKDG for AMD Family 15h Models 10h-1Fh Processors Table 35: BIOS recommendations for SO-DIMM address timings and output driver control for FP2 Package D18F2x9C_x0000_0000_dct[1:0]_mp[1:0] DIMM1 D18F2x9C_x0000_0004_dct[1:0]_mp[1:0] DIMM0 RTT_Wr VDDIO SlowAccessMode NumDimmSlots DdrRate RTT_Nom D18F2x94_dct[1:0] Condition 1 667 1.25, 1.35, 1.5 SR - 0 120 Off 00000000h 00002222h 1 667 1.25, 1.35, 1.5 DR - 0 120 Off 00000000h 00002222h 1 800 1.25, 1.35, 1.5 SR - 0 120 Off 00000000h 00002222h 1 800 1.25, 1.35, 1.5 DR - 0 120 Off 00000000h 00002222h 1 1066 1.25, 1.35, 1.5 SR - 0 120 Off 003D3D3Dh 10002222h 1 1066 1.25, 1.35, 1.5 DR - 0 120 Off 00000000h 10002222h 1 1333 1.25, 1.35, 1.5 SR - 0 60 Off 003D3D3Dh 20002222h 1 1333 1.25, 1.35, 1.5 DR - 0 60 Off 00003D3Dh 20002222h 2 667 1.25, 1.35, 1.5 NP SR 0 120 Off 00000000h 00002222h 2 667 1.25, 1.35, 1.5 NP DR 0 120 Off 00000000h 00002222h 2 667 1.25, 1.35, 1.5 SR SR 1 40 120 00000039h 10222323h 2 667 1.25, 1.35, 1.5 DR DR 1 40 120 00000039h 10222323h 2 667 1.25, 1.35, 1.5 DR SR 1 40 120 00000039h 10222323h 2 667 1.25, 1.35, 1.5 SR DR 1 40 120 00000039h 10222323h 2 800 1.25, 1.35, 1.5 NP SR 0 120 Off 00000000h 00002222h 2 800 1.25, 1.35, 1.5 NP DR 0 120 Off 00000000h 00002222h 2 800 1.25, 1.35, 1.5 SR SR 1 40 120 00000039h 20222323h 2 800 1.25, 1.35, 1.5 DR DR 1 40 120 00000039h 20222323h 2 800 1.25, 1.35, 1.5 DR SR 1 40 120 00000039h 20222323h 2 800 1.25, 1.35, 1.5 SR DR 1 40 120 00000039h 20222323h 2 1066 1.25, 1.35, 1.5 NP SR 0 120 Off 003D3D3Dh 10002222h 2 1066 1.25, 1.35, 1.5 NP DR 0 120 Off 00000000h 10002222h 2 1066 1.25, 1.35, 1.5 SR SR 1 30 120 00000037h 30222323h 2 1066 1.25, 1.35, 1.5 DR DR 1 30 120 00000037h 30222323h 2 1066 1.25, 1.35, 1.5 DR SR 1 30 120 00000037h 30222323h 112 42300 Rev 3.12 - July 14, 2015 BKDG for AMD Family 15h Models 10h-1Fh Processors Table 35: BIOS recommendations for SO-DIMM address timings and output driver control for FP2 Package D18F2x9C_x0000_0000_dct[1:0]_mp[1:0] DIMM1 D18F2x9C_x0000_0004_dct[1:0]_mp[1:0] DIMM0 RTT_Wr VDDIO SlowAccessMode NumDimmSlots DdrRate RTT_Nom D18F2x94_dct[1:0] Condition 2 1066 1.25, 1.35, 1.5 SR DR 1 30 120 00000037h 30222323h 2 1333 1.25, 1.35, 1.5 NP SR 0 60 Off 003D3D3Dh 20002222h 2 1333 1.25, 1.35, 1.5 NP DR 0 60 Off 00003D3Dh 20002222h 2 1333 1.25, 1.35, 1.5 SR SR 1 30 120 00000035h 30222323h 2 1333 1.35, 1.5 DR DR 1 30 120 00000035h 30222323h 2 1333 1.35, 1.5 DR SR 1 30 120 00000035h 30222323h 2 1333 1.35, 1.5 SR DR 1 30 120 00000035h 30222323h Table 36: BIOS recommendations for DRAM soldered down address timings and output driver control for FP2 Package D18F2x9C_x0000_0000_dct[1:0]_mp[1:0] D18F2x9C_x0000_0004_dct[1:0]_mp[1:0] NumRanks RTT_Wr VDDIO SlowAccessMode DdrRate RTT_Nom D18F2x94_dct[1:0] Condition 667 1.25, 1.35, 1.5 1 0 120 Off 00000000h 00002222h 667 1.25, 1.35, 1.5 2 0 120 Off 00000000h 00002222h 667 1.25, 1.35, 1.5 3 1 40 120 00000035h 30222323h 667 1.25, 1.35, 1.5 4 1 40 120 00000035h 30222323h 800 1.25, 1.35, 1.5 1 0 120 Off 00000000h 00002222h 113 42300 Rev 3.12 - July 14, 2015 BKDG for AMD Family 15h Models 10h-1Fh Processors Table 36: BIOS recommendations for DRAM soldered down address timings and output driver control for FP2 Package D18F2x9C_x0000_0000_dct[1:0]_mp[1:0] D18F2x9C_x0000_0004_dct[1:0]_mp[1:0] NumRanks RTT_Wr VDDIO SlowAccessMode DdrRate RTT_Nom D18F2x94_dct[1:0] Condition 800 1.25, 1.35, 1.5 2 0 120 Off 00000000h 00002222h 800 1.25, 1.35, 1.5 3 1 40 120 00000035h 30222323h 800 1.25, 1.35, 1.5 4 1 40 120 00000035h 30222323h 1066 1.25, 1.35, 1.5 1 0 120 Off 003D3D3Dh 10002222h 1066 1.25, 1.35, 1.5 2 0 120 Off 00000000h 10002222h 1066 1.25, 1.35, 1.5 3 1 30 120 00000035h 30222323h 1066 1.25, 1.35, 1.5 4 1 30 120 00000035h 30222323h 1333 1.25, 1.35, 1.5 1 0 60 Off 003D3D3Dh 20002222h 1333 1.25, 1.35, 1.5 2 0 60 Off 00003D3Dh 20002222h 1333 1.25, 1.35, 1.5 3 1 30 120 00000035h 30222323h 1333 1.25, 1.35, 1.5 4 1 30 120 00000035h 30222323h 2.9.5.7 DCT Training Specific Configuration The DCT requires certain features be disabled during DRAM device initialization and training. BIOS should program the registers in Table 37 before DRAM device initialization and training. For normal operation, BIOS programs the recommended values if provided in Table 37. BIOS must quiesce all other forms of DRAM traffic on the channel being trained. See 2.9.5 [DCT/DRAM Initialization and Resume]. Table 37: DCT Training Specific Register Values Register D18F2x78_dct[1:0][AddrCmdTriEn] D18F2x8C_dct[1:0][DisAutoRefresh] D18F2x90_dct[1:0][ForceAutoPchg] D18F2x90_dct[1:0][DynPageCloseEn] D18F2x94_dct[1:0][BankSwizzleMode] D18F2x94_dct[1:0][DcqBypassMax] D18F2x94_dct[1:0][PowerDownEn] D18F2x94_dct[1:0][ZqcsInterval] Training 0 1 0 0 0 0 0 00b Normal Operation 1 0 0 0 1 1Fh 1 10b 114 42300 Rev 3.12 - July 14, 2015 BKDG for AMD Family 15h Models 10h-1Fh Processors Table 37: DCT Training Specific Register Values Register D18F2x9C_x0000_000D_dct[1:0]_mp[1:0][RxMax DurDllNoLock] D18F2x9C_x0000_000D_dct[1:0]_mp[1:0][TxMax DurDllNoLock] D18F2x9C_x0D0F_0[F,7:0]10_dct[1:0]_mp[1:0][E nRxPadStandby] D18F2xA4[CmdThrottleMode] Training 000b Normal Operation See 2.9.5.11 000b See 2.9.5.11 0 See 2.9.5.11 000b xxxb1 D18F2xA4[ODTSEn] 000b x1 D18F2xA4[BwCapEn] 0 D18F2xA8_dct[1:0][BankSwap] D18F2x110[DctSelIntLvEn] 0 0 x1 1 x1 1. Programmed specific to the current platform or memory configuration. 2.9.5.8 DRAM Device and Controller Initialization BIOS initializes the DRAM devices and the controller using a software controlled sequence. See 2.9.5.8.1 [Software DDR3 Device Initialization]. BIOS must observe additional requirements for changing the PLL frequency when setting D18F2x7C_dct[1:0][EnDramInit]. See 2.9.5.4.2 [DRAM Channel Frequency Change]. DRAM initialization is complete after the value of D18F2x7C_dct[1:0][EnDramInit] is written by BIOS from 1 to 0 in the software-controlled sequence. 2.9.5.8.1 Software DDR3 Device Initialization BIOS should apply the following procedure to each DCT to initialize the DDR3 DIMMs on the channel. This procedure should be run only when booting from an unpowered state (ACPI S4, S5 or G3; not S3, suspend to RAM). 1.Configure the DCT registers, including MemClkFreq and MemClkFreqVal. 2.Program D18F2x7C_dct[1:0][EnDramInit] = 1. 3.Wait 200 us. 4.Program D18F2x7C_dct[1:0][DeassertMemRstX] = 1. 5.Wait 500 us. 6.Program D18F2x7C_dct[1:0][AssertCke] = 1. 7.Wait 360 ns. 8.Send MRS(2). 9.Send MRS(3). Ordinarily at this time, MrsAddress[2:0] = 000b. 10.Send MRS(1) with MrsAddress[7] = 0. 11.Send MRS(0) with MrsAddress[8] = 1. 12.Send two ZQCL commands. • BIOS instructs the DCT to send a ZQCL command by programming D18F2x7C_dct[1:0] as follows: • Program MrsAddress[10] = 1. • Program SendZQCmd = 1. • Wait for SendZQCmd = 0. • Wait 512 MEMCLKs. 115 42300 Rev 3.12 - July 14, 2015 BKDG for AMD Family 15h Models 10h-1Fh Processors 13.Program D18F2x7C_dct[1:0][EnDramInit] = 0. 14.Program D18F2x2E8_dct[1:0]_mp[1:0], D18F2x2EC_dct[1:0]_mp[1:0] with a copy of the mode register data sent in the steps above, except with D18F2x2E8_dct[1:0]_mp[1:0][8] = 0 (do not reset the DLL with a memory P-state change). 2.9.5.8.1.1 DDR3 MR Initialization BIOS instructs the DCT to send MRS commands by programming D18F2x7C_dct[1:0] as follows: 1.Program MrsBank and MrsAddress as specified by Table 38, Table 39, Table 40, and Table 41: • MrsBank[2:0] = BA2:BA0. • MrsAddress[15:0] = A15:A0. • See also D18F2x[5C:40]_dct[1:0][OnDimmMirror]. 2.Program MrsChipSel as appropriate. 3.Program SendMrsCmd = 1. 4.Wait for SendMrsCmd = 0. Table 38. DDR3 MR0 Address Field Field BA2:BA0 A15:A13 A12 A11:A9 MR Select Reserved PPD WR A8 A7 DLL TM Value 000b 000b See D18F2x84_dct[1:0][PchgPDModeSel]. Twr (See Bits D18F2x22C_dct[1:0]_mp[1:0][Twr]) 000b 16 MEMCLKs 001b 5 MEMCLKs 010b 6 MEMCLKs 011b 7 MEMCLKs 100b 8 MEMCLKs 101b 10 MEMCLKs 110b 12 MEMCLKs 111b 14 MEMCLKs Controlled as required by the initialization sequence 0 116 42300 Rev 3.12 - July 14, 2015 Table 38. DDR3 MR0 Address Field Field A6:A4,A2 CAS Latency A3 A1:A0 RBT BL Table 39. BA2:BA0 A15:A13 A12 A11 A10 A8 A7 A4:A3 A9, A6, A2 A5, A1 A0 Bits 0000b 0001b 0010b 0011b 0100b 0101b 0110b 0111b 1000b 1001b 1010b 1011b 1100b 1101b 1110b 1111b D18F2x200[Tcl] Reserved 12 MEMCLKs 5 MEMCLKs 13 MEMCLKs 6 MEMCLKs 14 MEMCLKs 7 MEMCLKs 15 MEMCLKs 8 MEMCLKs 16 MEMCLKs 9 MEMCLKs Reserved 10 MEMCLKs Reserved 11 MEMCLKs Reserved 1 D18F2x84_dct[1:0][BurstCtrl] Field MR Select Reserved Qoff TDQS Reserved Reserved Level AL Rtt_Nom DIC DLL Value 001b 000b 0b D18F2x94_dct[1:0][RDqsEn] 0b 0b Controlled as required by the initialization sequence 00b See 2.9.5.6.5 [DRAM ODT Control] 01b 0 DDR3 MR2 Address Field BA2:BA0 A15:A11 A10:A9 A8 Value DDR3 MR1 Address Field Table 40. BKDG for AMD Family 15h Models 10h-1Fh Processors Field MR Select Reserved Rtt_Wr Reserved Value 010b 0_0000b See 2.9.5.6.5 [DRAM ODT Control] 0b 117 42300 Rev 3.12 - July 14, 2015 Table 40. DDR3 MR2 Address Field A7 A6 A5:A3 A2:A0 Table 41. 2.9.5.9 Field SRT ASR CWL PASR Value See ASR See DIMM SPD Byte 31: SDRAM Thermal and Refresh Options D18F2x20C_dct[1:0]_mp[1:0][Tcwl]-5 000b DDR3 MR3 Address Field BA2:BA0 A15:A3 A2 A1:A0 BKDG for AMD Family 15h Models 10h-1Fh Processors Field MR Select Reserved MPR MPR Loc Value 011b 000_0000_0000b 0b 00b DRAM Training This section describes detailed methods used to train the processor DDR interface to DRAM for optimal functionality and performance. DRAM training is performed by BIOS after initializing the DRAM controller. See 2.9.5.8 [DRAM Device and Controller Initialization]. Some of the DRAM training steps described in this section require two passes if the target MEMCLK frequency is greater than 333 MHz.For optimal software performance, software may defer the second pass (at target MEMCLK frequency) for each training step until after the first pass (at lowest supported frequency) of all other training steps are complete. See D18F2x94_dct[1:0][MemClkFreq]. Product specific training requirements are as follows: • Program D18F2x9C_x0000_0008_dct[1:0]_mp[1:0][TrNibbleSel]=0. • Program D18F2xA8_dct[1:0][PerRankTimingEn]=1. See 2.9.5.7 [DCT Training Specific Configuration] for additional training requirements. In the following subsections, lane is used to describe an 8-bit wide data group, each with its own timing control. 2.9.5.9.1 Write Levelization Training Write levelization involves using the phy to detect the edge of DQS with respect to the memory clock on the DIMM for write accesses to each lane. Training is accomplished on a per channel, per rank basis. If the target frequency is greater than 333 MHz then BIOS performs two passes; otherwise, only one pass is required. See 2.9.5.4.2 [DRAM Channel Frequency Change]. • Pass 1: Configure the memory subsystem for 333 MHz MEMCLK frequency. • Pass 2: Configure the memory subsystem for the target MEMCLK frequency. 118 42300 Rev 3.12 - July 14, 2015 BKDG for AMD Family 15h Models 10h-1Fh Processors The following describes the steps used for each pass of write levelization training for each channel: For each rank: 1. Prepare the DIMMs for write levelization using DDR3-defined MR commands. A. Configure the output driver and on-die termination of the target DIMM as follows: • For the target rank of the target DIMM, enable write leveling mode and enable the output driver. • For all other ranks of the target DIMM, enable write leveling mode and disable the output driver. • For two or more DIMMs per channel, program Rtt_Nom of the target rank to the corresponding specified Rtt_Wr termination. Otherwise, configure Rtt_Nom of the target DIMM as normal. See 2.9.5.6.6 [DRAM Address Timing and Output Driver Compensation Control]. B. Configure Rtt_Nom on the non-target DIMMs as normal. See 2.9.5.6.6 [DRAM Address Timing and Output Driver Compensation Control]. 2. Wait 40 MEMCLKs. 3. Configure the phy for write levelization training: A. Program D18F2x9C_x0000_0008_dct[1:0]_mp[1:0][WrtLvTrEn]=0. B. Program D18F2x9C_x0000_0008_dct[1:0]_mp[1:0][TrChipSel] to specify the target rank to be trained. C. Program D18F2x9C_x0000_0008_dct[1:0]_mp[1:0][WrLvOdt] to the proper ODT settings for the current memory subsystem configuration. See 2.9.5.6.5 [DRAM ODT Control]. D. Program D18F2x9C_x0000_0008_dct[1:0]_mp[1:0][WrLvOdtEn]=1. E. MFENCE. F. Wait 10 MEMCLKs to allow for ODT signal settling. G. For each lane program an initial value to registers D18F2x9C_x0000_00[51:50]_dct[1:0] to set the gross and fine delay. See 2.9.5.9.1.1 [Write Leveling Seed Value]. 4. Perform write leveling of the devices on the DIMM: A. Program D18F2x9C_x0000_0008_dct[1:0]_mp[1:0][WrtLvTrEn]=1. B. MFENCE. C. Wait 200 MEMCLKs. D. Program D18F2x9C_x0000_0008_dct[1:0]_mp[1:0][WrtLvTrEn]=0. E. Read from registers D18F2x9C_x0000_00[51:50]_dct[1:0] to get the gross and fine delay settings for the target DIMM and save these values. 5. Disable write levelization training so that the phy stops driving write levelization ODT. A. Program D18F2x9C_x0000_0008_dct[1:0]_mp[1:0][WrLvOdtEn]=0. B. MFENCE. C. Wait 10 MEMCLKs to allow for ODT signal settling. 6. Program the target DIMM back to normal operation by configuring the following: A. Configure all ranks of the target DIMM for normal operation. B. Enable the output drivers of all ranks of the target DIMM. C. For a two or more DIMM system, program the Rtt_Nom value for the target DIMM to the normal operating termination. For each rank: • BIOS calculates and programs the final saved gross and fine delay values for each lane into D18F2x9C_x0000_00[4A:30]_dct[1:0]_mp[1:0] [DRAM DQS Write Timing]. • WrDqsFineDly = PhRecFineDlyByte. • GrossDly = SeedGross + PhRecGrossDlyByte - SeedPreGross. • The Critical Gross Delay (CGD) is the minimum GrossDly of all byte lanes and all DIMMs. • If (CGD < 0) Then • D18F2x20C_dct[1:0]_mp[1:0][WrDqDqsEarly] = ABS(CGD) • WrDqsGrossDly = GrossDly + WrDqDqsEarly • Else 119 42300 Rev 3.12 - July 14, 2015 BKDG for AMD Family 15h Models 10h-1Fh Processors • D18F2x20C_dct[1:0]_mp[1:0][WrDqDqsEarly] = 0. • WrDqsGrossDly = GrossDly. 2.9.5.9.1.1 Write Leveling Seed Value The seed value for pass 1 of write leveling is design and platform specific. The platform vendor may need to characterize and adjust this value for proper write levelization training. The seed delay value must fall within +/- 1/2 MEMCLK, including silicon PVT margin and jitter of the measured clock delay. 1. Calculate the total seed based on the following: • Pass 1: SeedTotal = configuration specific seed value found in Table 42. • Pass N: • SeedTotalPreScaling = the total delay values obtained from the previous (N-1) pass of write levelization training. • Write Leveling Total Delay = D18F2x9C_x0000_00[4A:30]_dct[1:0]_mp[1:0][WrDqsGrossDly, WrDqs-FineDly] - (0x20 * D18F2x20C_dct[1:0]_mp[1:0][WrDqDqsEarly]) • SeedTotal = SeedTotalPreScaling*(target frequency)/(frequency from previous pass). 2. SeedGross = SeedTotal DIV 32. 3. SeedFine = SeedTotal MOD 32. 4. If (SeedGross is odd) then SeedPreGross = 1 else SeedPreGross = 2. 5. Program D18F2x9C_x0000_00[51:50]_dct[1:0][PhRecFineDlyByte] = SeedFine. 6. Program D18F2x9C_x0000_00[51:50]_dct[1:0][PhRecGrossDlyByte] = SeedPreGross. Table 42. DDR3 Write Leveling Seed Values DIMM Type Seed Value1 SO-DIMM Unbuffered 1. DDR3-667 (333 MHz). 2.9.5.9.2 0Eh 15h DQS Receiver Enable Training Receiver enable delay training is used to dynamically determine the optimal delay value for D18F2x9C_x0000_00[2A:10]_dct[1:0]_mp[1:0] [DRAM DQS Receiver Enable Timing]. The optimal DQS receiver enable delay value is platform and load specific, and occurs in the middle of a received read preamble. The timing of the preamble includes the inbound DQS propagation delay, which is unknown by BIOS. The training for delay values involves: 1. 2. 3. 4. Configuring the phy for an initial expected phase value (seed). Generating a stream of read DQS edges from the DRAM by issuing multiple read commands. The phy determining the phase between the received DQS edges and a reference clock. Calculating a final delay value for enabling receivers during normal read operations using the phase determined by the phy. BIOS should program D18F2x210_dct[1:0]_nbp[3:0][MaxRdLatency] to 55h. Training is accomplished on a per channel, per rank basis. If the target frequency is greater than 333 MHz then BIOS performs two passes; otherwise only one pass is required. See 2.9.5.4.2 [DRAM Channel Frequency 120 42300 Rev 3.12 - July 14, 2015 BKDG for AMD Family 15h Models 10h-1Fh Processors Change]. • Pass 1: Configure the memory subsystem for 333 MHz MEMCLK frequency. • Pass 2: Configure the memory subsystem for the target MEMCLK frequency. The following describes the steps used for each pass of receiver enable training for each channel: For each rank: 1. Program D18F2x9C_x0000_0008_dct[1:0]_mp[1:0][TrChipSel] to specify the target rank to be trained. 2. For each lane program an initial value to registers D18F2x9C_x0000_00[51:50]_dct[1:0] to set the gross and fine delay as specified in 2.9.5.9.2.1 [DQS Receiver Enable Training Seed Value]. 3. Program D18F2x9C_x0000_0008_dct[1:0]_mp[1:0][DqsRcvTrEn]=1. 4. Issue 192 read requests to the target rank. See 2.9.5.9.6 [Continuous Pattern Generation]. 5. Program D18F2x9C_x0000_0008_dct[1:0]_mp[1:0][DqsRcvTrEn]=0. 6. Read D18F2x9C_x0000_00[51:50]_dct[1:0][PhRecGrossDlyByte, PhRecFineDlyByte] to get the gross and fine delay values for each lane. 7. For each lane, calculate and program the corresponding receiver enable delay values for D18F2x9C_x0000_00[2A:10]_dct[1:0]_mp[1:0][DqsRcvEnGrossDelay, DqsRcvEnFineDelay]. Save the result for use later. • DqsRcvEnFineDelay = PhRecFineDlyByte • DqsRcvEnGrossDelay = SeedGross + PhRecGrossDlyByte - SeedPreGross + 1. 2.9.5.9.2.1 DQS Receiver Enable Training Seed Value The seed value for pass 1 of receiver enable delay training is design and platform specific and should be determined by characterization for best performance. The seed value represents the total delay from a reference point to the first expected rise edge of DQS on a read CAS measured at the processor pins, in 1 UI/32 increments. The reference point is defined as the clock in which CAS is asserted + CL - 1. This value is expected to be larger than 2 UI in the steps below. The following steps are taken to determine the seed values needed to program the DRAM Phase Recovery Control Registers: For each pass and each lane: 1. Calculate the total seed based on the following: • Pass 1: SeedTotal = The seed value found in Table 43 + the total delay values obtained from the first pass of write levelization training. See 2.9.5.9.1 [Write Levelization Training]. • Write Leveling Total Delay = D18F2x9C_x0000_00[4A:30]_dct[1:0]_mp[1:0][WrDqsGrossDly, WrDqs-FineDly] - (0x20 * D18F2x20C_dct[1:0]_mp[1:0][WrDqDqsEarly]) • Pass N: • RegisterDelay = 0 • SeedTotalPreScaling = (the total delay values in D18F2x9C_x0000_00[2A:10]_dct[1:0]_mp[1:0] from the previous (N-1) pass of training) - RegisterDelay - 20h. • SeedTotal = RegisterDelay + FLOOR(SeedTotalPreScaling*(target frequency)/(frequency from previous pass)). 2. SeedGross = SeedTotal DIV 32. 3. SeedFine = SeedTotal MOD 32. 4. If (SeedGross is odd) then SeedPreGross =1 else SeedPreGross = 2. 5. Program D18F2x9C_x0000_00[51:50]_dct[1:0][PhRecFineDlyByte] = SeedFine. 121 42300 Rev 3.12 - July 14, 2015 BKDG for AMD Family 15h Models 10h-1Fh Processors 6. Program D18F2x9C_x0000_00[51:50]_dct[1:0][PhRecGrossDlyByte] = SeedPreGross. 7. Program D18F2x9C_x0000_00[2A:10]_dct[1:0]_mp[1:0][DqsRcvEnGrossDelay] = SeedGross. Table 43. DDR3 DQS Receiver Enable Training Seed Values DIMM Type Seed Value1 SO-DIMM 32h Unbuffered 32h 1. DDR3-667 (333 MHz) 2.9.5.9.3 DQS Receiver Enable Cycle Training Receiver enable delay cycle training is used to train the gross delay settings of D18F2x9C_x0000_00[2A:10]_dct[1:0]_mp[1:0] to the middle of the received read preamble using the phy phase results. For each rank and lane: 1. Program D18F2x9C_x0D0F_0[F,7:0]30_dct[1:0][BlockRxDqsLock] = 1. 2. RxEnOrig = D18F2x9C_x0000_00[2A:10]_dct[1:0]_mp[1:0][DqsRcvEnGrossDelay, DqsRcvEnFineDelay] result from 2.9.5.9.2 [DQS Receiver Enable Training]. 3. RxEnOffset = MOD(RxEnOrig + 10h, 40h) 4. For each DqsRcvEn value beginning from RxEnOffset incrementing by 1 MEMCLK: A. Program D18F2x9C_x0000_00[2A:10]_dct[1:0]_mp[1:0][DqsRcvEnGrossDelay, DqsRcvEnFineDelay] with the current value. B. Perform 2.9.5.9.4 [DQS Position Training]. • Record the result for the current DqsRcvEn setting as a pass or fail depending if a data eye is found. 5. Process the array of results and determine a pass-to-fail transition. A. DqsRcvEnCycle = the total delay value of the pass result. B. Program D18F2x9C_x0000_00[2A:10]_dct[1:0]_mp[1:0][DqsRcvEnGrossDelay, DqsRcvEnFineDelay] = DqsRcvEnCycle - 10h. 6. Program D18F2x9C_x0D0F_0[F,7:0]30_dct[1:0][BlockRxDqsLock] = 0. 2.9.5.9.4 DQS Position Training DQS position training is used to place the DQS strobe in the center of the read DQ data eye and to center the write DQ data eye across the write DQS strobe. Determining the correct DRAM DQS and DQ delay settings for both reads and writes is conducted by performing a two dimensional search of the delay settings found in D18F2x9C_x0000_0[3:0]0[6:5]_dct[1:0]_mp[1:0] [DRAM Read DQS Timing] and D18F2x9C_x0000_0[3:0]0[2:1]_dct[1:0]_mp[1:0] [DRAM Write Data Timing]. Training is accomplished on a per channel, per rank, and per lane basis. For DQS position training, BIOS generates a training pattern using continuous read or write data streams. A 2k-bit-time training pattern is recommended for optimal results. To achieve this, BIOS programs D18F2x260_dct[1:0][CmdCount] = 256, D18F2x250_dct[1:0][CmdTgt]=01b, and D18F2x25[8,4]_dct[1:0] to access two different banks of the same CS. See 2.9.5.9.6.1 [DRAM Training Pattern Generation]. Prior to DQS position training, BIOS must program D18F2x210_dct[1:0]_nbp[3:0][MaxRdLatency] based on 122 42300 Rev 3.12 - July 14, 2015 BKDG for AMD Family 15h Models 10h-1Fh Processors the current greatest value of D18F2x9C_x0000_00[2A:10]_dct[1:0]_mp[1:0]. See 2.9.5.9.5 [Calculating MaxRdLatency]. The following describes the steps used for DQS position training for each channel: • Program D18F2x9C_x0D0F_0[F,7:0]1F_dct[1:0]_mp[1:0] for all lanes. • RxBypass3rd4thStg = 1. • Rx4thStgEn = 0. For each rank and lane: 1. Select a 64 byte aligned test address. 2. For each write data delay value in D18F2x9C_x0000_0[3:0]0[2:1]_dct[1:0]_mp[1:0] from Wr-DQS to WrDQS plus 1 UI, using the Wr-DQS delay value found in 2.9.5.9.1 [Write Levelization Training]: • Program the write data delay value for the current lane. • Write the DRAM training pattern to the test address. 3. For each read DQS delay value in D18F2x9C_x0000_0[3:0]0[6:5]_dct[1:0]_mp[1:0] from 0 to 1 UI: • Program the read DQS delay value for the current lane. • Read the DRAM training pattern from the test address. • Record the result for the current settings as a pass or fail depending if the pattern is read correctly. • Process the array of results and determine the longest string of consecutive passing read DQS delay values. • If the read DQS delay results for the current lane contain three or more consecutive passing delay values, then program D18F2x9C_x0000_0[3:0]0[6:5]_dct[1:0]_mp[1:0] with the average value of the smallest and largest delay values in the string of consecutive passing results. • If the average value of passing read DQS delay for the lane is negative, then adjust the input receiver DQ delay in D18F2x9C_x0D0F_0[F,7:0]1F_dct[1:0]_mp[1:0] for the lane as follows: • IF (RxBypass3rd4thStg == 1) program RxBypass3rd4thStg=0 and repeat step 3 above for all ranks and lanes. • ELSEIF (Rx4thStgEn == 0) program Rx4thStgEn=1 and repeat step 3 above for all ranks and lanes. • ELSE program the read DQS delay for the lane with a value of zero. 4. Process the array of results and determine the longest string of consecutive passing write data delay values for the read DQS delay value found in the step above. • If the write data delay results for the current lane contain three or more consecutive passing delay values, then program D18F2x9C_x0000_0[3:0]0[2:1]_dct[1:0]_mp[1:0] with the average value of the smallest and largest delay values in the string of consecutive passing results. See Figure 3. 123 WrDatFineDly (Wr-DQS to Wr-DQS + 1UI) 42300 Rev 3.12 - July 14, 2015 31 30 29 28 27 26 25 24 23 22 21 20 19 18 17 16 15 14 13 12 11 10 9 8 7 6 5 4 3 2 1 0 /32 F F F F F F F F F F F F F F F F F F F F F F F F F F F F F F F F 0 F F F F F F F F F F F F F F F F F F F F F F F F F F F F F F F F 1 F F F F F F F F F F F F F F F F F F F F F F F F F F F F F F F F 2 F F F F F F F F F F F F F F F F F F F F F F F F F F F F F F F F 3 F F F F F F F F F F F F F F F F F F F F F F F F F F F F F F F F 4 F F F F F F F F F F F F F F F F F F F F F F F F F F F F F F F F 5 F F F F F F F F F F F F F F F F F F F F F F F F F F F F F F F F 6 F F F F F F F F F F F F F F F F F F F F F F F F F F F F F F F F 7 BKDG for AMD Family 15h Models 10h-1Fh Processors F F F F F F F F F F F F F F F F F F F F F F F F F F F F F F F F 8 F F F F F F F F F F F F F F F F F F F F F F F F F F F F F F F F F F F F F F F F F F F F F F F F F F F F F F F F F F F F F F F F 9 10 F F F F F F F F F F F F F F F F F F F F F F F F F F F F F F F F 11 F F F F F F F F F F F F F F F F F F F F F F F F F F F F F F F F 12 F F F F F F F F F F F F F F F F F F F F F F F F F F F F F F F F 13 F F F F F F F F F F F F F F F F F F F F F F F F F F F F F F F F 14 F F F F F F F F F F F F F F F F F F F F F F F F F F F F F F F F 15 F F F F F F P P P P P P P P P F F F F F F F F F F F F F F F F F 16 F F F F F F P P P P P P P P P F F F F F F F F F F F F F F F F F 17 F F F F F F P P P P P P P P P F F F F F F F F F F F F F F F F F 18 F F F F F F P P P P P P P P P F F F F F F F F F F F F F F F F F 19 F F F F F F P P P P P P P P P F F F F F F F F F F F F F F F F F 20 F F F F F F P P P P P P P P P F F F F F F F F F F F F F F F F F 21 F F F F F F P P P P P P P P P F F F F F F F F F F F F F F F F F 22 F F F F F F P P P P P P P P P F F F F F F F F F F F F F F F F F 23 F F F F F F P P P P P P P P P F F F F F F F F F F F F F F F F F 24 F F F F F F F F F F F F F F F F F F F F F F F F F F F F F F F F 25 F F F F F F F F F F F F F F F F F F F F F F F F F F F F F F F F 26 F F F F F F F F F F F F F F F F F F F F F F F F F F F F F F F F 27 F F F F F F F F F F F F F F F F F F F F F F F F F F F F F F F F 28 F F F F F F F F F F F F F F F F F F F F F F F F F F F F F F F F 29 F F F F F F F F F F F F F F F F F F F F F F F F F F F F F F F F 30 F F F F F F F F F F F F F F F F F F F F F F F F F F F F F F F F 31 RdDqsTime Figure 3: DQS Position Training Example Results In some cases, a non-zero process, voltage, and temperature dependent insertion delay is added to the DLL programmed read DQS delay. This has the effect of sampling data later than intended and can result in missing the left edge of the passing region when sweeping from 0 to 1 UI because a read DQS delay value of 0 is already in the passing region. Since DQS is periodic, BIOS can recover the missing information by adjusting the algorithm described above to analyze both the in phase data and the data shifted by one bit time at each step of the read DQS delay sweep. See D18F2x268_dct[1:0][NibbleErrSts] and D18F2x26C_dct[1:0][NibbleErr180Sts]. As shown in Figure 4, for each delay setting BIOS records a passing result of PΦ for the data comparison shifted by one bit time if the data at bit times N=0, 1, …, 6, is read correctly when compared against the data written at bit times N=1, 2, …, 7. In the array of results, these passing values make up the left piece of information that had been lost due to insertion delay. In order to process the array of results, BIOS calculates the read DQS delay value for a PΦ result as RdDqsTimeByte minus 1 UI. 124 WrDatFineDly (Wr-DQS to Wr-DQS + 1UI) 42300 Rev 3.12 - July 14, 2015 BKDG for AMD Family 15h Models 10h-1Fh Processors 31 F F F F F F F F F F F F F F F F F F F F F F F 30 F F F F F F F F F F F F F F F F F F F F F F F 29 F F F F F F F F F F F F F F F F F F F F F F F 28 F F F F F F F F F F F F F F F F F F F F F F F 27 F F F F F F F F F F F F F F F F F F F F F F F 26 F F F F F F F F F F F F F F F F F F F F F F F 25 P P P P P P P F F F F F F F F F F F F F F F F 24 P P P P P P P F F F F F F F F F F F F F F F F 23 P P P P P P P F F F F F F F F F F F F F F F F 22 P P P P P P P F F F F F F F F F F F F F F F F 21 P P P P P P P F F F F F F F F F F F F F F F F 20 P P P P P P P F F F F F F F F F F F F F F F F 19 P P P P P P P F F F F F F F F F F F F F F F F 18 P P P P P P P F F F F F F F F F F F F F F F F 17 P P P P P P P F F F F F F F F F F F F F F F F 16 F F F F F F F F F F F F F F F F F F F F F F F 15 F F F F F F F F F F F F F F F F F F F F F F F 14 F F F F F F F F F F F F F F F F F F F F F F F 13 F F F F F F F F F F F F F F F F F F F F F F F 12 F F F F F F F F F F F F F F F F F F F F F F F 11 F F F F F F F F F F F F F F F F F F F F F F F 10 F F F F F F F F F F F F F F F F F F F F F F F 9 F F F F F F F F F F F F F F F F F F F F F F F 8 F F F F F F F F F F F F F F F F F F F F F F F 7 F F F F F F F F F F F F F F F F F F F F F F F 6 F F F F F F F F F F F F F F F F F F F F F F F 5 F F F F F F F F F F F F F F F F F F F F F F F 4 F F F F F F F F F F F F F F F F F F F F F F F 3 F F F F F F F F F F F F F F F F F F F F F F F 2 F F F F F F F F F F F F F F F F F F F F F F F 1 F F F F F F F F F F F F F F F F F F F F F F F 0 F F F F F F F F F F F F F F F F F F F F F F F /32 0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 Φ -32 -31 -30 -29 -28 -27 -26 -25 -24 -23 -22 -21 -20 -19 -18 -17 -16 -15 -14 -13 -12 -11 -10 F F F F F F F F F F F F F F F F F F F F F F F F F F F F F F F F 23 -9 F F F F F F F F F F F F F F F F F F F F F F F F F F F F F F F F 24 -8 F F F F F F F F F F F F F F F F F F F F F F F F F F F F F F F F 25 -7 F F F F F F F F F F F F F F F F F F F F F F F F F F F F F F F F 26 -6 F F F F F F F F F F F F F F F F F F F F F F F F F F F F F F F F 27 -5 F F F F F F F F F F F F F F F F F F F F F F F F F F F F F F F F 28 -4 F F F F F F F F F F F F F F F F F F F F F F F F F F F F F F F F 29 -3 F F F F F F F F F F F F PΦ PΦ PΦ PΦ PΦ PΦ PΦ PΦ PΦ PΦ PΦ PΦ PΦ PΦ PΦ PΦ PΦ PΦ F F F F F F F F F F F F F F F F F F F F F F F F F F F F F F F F F F 30 31 -2 -1 RdDqsTime Figure 4: DQS Position Training Insertion Delay Recovery Example Results 2.9.5.9.5 Calculating MaxRdLatency The MaxRdLatency value determines when the memory controller can receive incoming data from the DCTs. Calculating MaxRdLatency consists of summing all the synchronous and asynchronous delays in the path from the processor to the DRAM and back at a given MEMCLK frequency. BIOS incrementally calculates the MaxRdLatency and then finally programs the value into D18F2x210_dct[1:0]_nbp[3:0][MaxRdLatency]. The following steps describe the algorithm used to compute D18F2x210_dct[1:0]_nbp[3:0][MaxRdLatency] used for DRAM training. P, N, and T are used as a temporary placeholders for the incrementally summed value. 1. P = N = T = 0 2. If (D18F2x9C_x0000_0004_dct[1:0]_mp[1:0][AddrCmdSetup] = 0 & D18F2x9C_x0000_0004_dct[1:0]_mp[1:0][CsOdtSetup] = 0 & D18F2x9C_x0000_0004_dct[1:0]_mp[1:0][CkeSetup] = 0) then P = P + 1 else P = P + 2 125 42300 Rev 3.12 - July 14, 2015 3. 4. 5. 6. 7. 8. 9. 10. BKDG for AMD Family 15h Models 10h-1Fh Processors P = P + (8 - D18F2x210_dct[1:0]_nbp[3:0][RdPtrInit]) P=P+5 P = P + (2 * (D18F2x200_dct[1:0]_mp[1:0][Tcl] - 1 clocks)) P = P + CEIL(MAX(D18F2x9C_x0000_00[2A:10]_dct[1:0]_mp[1:0][DqsRcvEnGrossDelay, DqsRcvEnFineDelay] + D18F2x9C_x0000_0[3:0]0[6:5]_dct[1:0]_mp[1:0][RdDqsTime] PCLKs)) + 1 If (NclkFreq/MemClkFreq < 2) then P = P + 4.5 Else P = P + 2.5 T = T + 1050 ps N = (P/(MemClkFreq * 2) + T) * NclkFreq • See D18F5x1[6C:60][NbDid, NbFid] and D18F2x88_dct[1:0][MemClkFreq]. D18F2x210_dct[1:0]_nbp[3:0][MaxRdLatency] = CEIL(N) - 1 2.9.5.9.5.1 MaxRdLatency Training After DRAM DQS receiver enable training, BIOS optimizes D18F2x210_dct[1:0]_nbp[3:0][MaxRdLatency] using the following algorithm. For MaxRdLatency training, BIOS generates a training pattern using continuous read or write data streams. See 2.9.5.9.6.1 [DRAM Training Pattern Generation]. For each channel: 1. Calculate a starting MaxRdLatency delay value by executing the steps in 2.9.5.9.5, excluding steps 4, 7, 8, and 10. 2. Select 32 64-byte aligned test addresses associated with the rank that has the worst case (D18F2x9C_x0000_00[2A:10]_dct[1:0]_mp[1:0][DqsRcvEnGrossDelay, DqsRcvEnFineDelay] + D18F2x9C_x0000_0[3:0]0[6:5]_dct[1:0]_mp[1:0][RdDqsTime]) register setting. 3. Write the DIMM test addresses with the training pattern. 4. For each MaxRdLatency value incrementing from the value calculated in step 1: A. Program D18F2x210_dct[1:0]_nbp[3:0][MaxRdLatency] with the current value. B. Read the DIMM test addresses. C. Compare the values read against the pattern written. • If the pattern is read correctly, go to step 5. D. Program D18F2x9C_x0000_0050_dct[1:0]=00000000h 5. Program D18F2x210_dct[1:0]_nbp[3:0][MaxRdLatency] = CEIL(current value + 1 NCLK + 1.5 MEMCLK). 2.9.5.9.6 Continuous Pattern Generation DRAM training relies on the ability to generate a string of continuous reads or writes between the processor and DRAM, such that worst case electrical interactions can be created. This section describes how these continuous strings of accesses may be generated. 2.9.5.9.6.1 DRAM Training Pattern Generation DRAM training pattern generation uses PRBS generators in the DCT to generate controlled read and write traffic streams. During write pattern generation, data values based off of the seeded PRBS are burst to the DRAM interface. Conversely for reads, data bursts from the DRAM interface are compared against expected data values based off of the seeded PRBS on a per nibble basis. Two address modes are available for DRAM training pattern generation, as configured by D18F2x250_dct[1:0][CmdTgt]. For generating a stream of reads or writes to the same rank, address target A mode is used. To generate a stream of accesses to up to two different ranks, address target A and B mode is used. 126 42300 Rev 3.12 - July 14, 2015 BKDG for AMD Family 15h Models 10h-1Fh Processors An overview of the BIOS sequence to generate training patterns is as follows: • Configure the DCT for pattern generation. See 2.9.5.7 [DCT Training Specific Configuration]. • Ensure DIMMs are configured to support 8-beat bursts (BL8 or dynamic burst length on the fly). See 2.9.5.8.1.1 [DDR3 MR Initialization]. • Wait for D18F2x250_dct[1:0][CmdSendInProg] = 0. • Program D18F2x250_dct[1:0][CmdTestEnable] = 1. • Send activate commands as appropriate. See 2.9.5.9.6.1.1 [Activate and Precharge Command Generation]. • Send read or write commands as desired. See 2.9.5.9.6.1.2 [Read and Write Command Generation]. • Send precharge commands as appropriate. See 2.9.5.9.6.1.1 [Activate and Precharge Command Generation]. • Program D18F2x250_dct[1:0][CmdTestEnable] = 0. 2.9.5.9.6.1.1 Activate and Precharge Command Generation Prior to sending read or write commands, BIOS must send an activate command to a row in a particular bank of the DRAM devices for access. To send an activate command, BIOS performs the following steps: • Program D18F2x28C_dct[1:0] to the desired address as follows: • CmdChipSelect = CS[7:0]. • CmdBank = BA[2:0]. • CmdAddress = A[17:0]. • Program D18F2x28C_dct[1:0][SendActCmd] = 1. • Wait until D18F2x28C_dct[1:0][SendActCmd] = 0. • Wait 75 MEMCLKs. After completing its accesses, BIOS must deactivate open rows in the DRAM devices. To send a precharge or precharge all command to deactivate open rows in a bank or in all banks, BIOS performs the following steps: • Wait 25 MEMCLKs. • Program D18F2x28C_dct[1:0] to the desired address as follows: • CmdChipSelect = CS[7:0]. • Precharge all command: • CmdAddress[10] = 1. • Precharge command: • CmdAddress[10] = 0. • CmdBank = BA[2:0]. • Program D18F2x28C_dct[1:0][SendPchgCmd] = 1. • Wait until D18F2x28C_dct[1:0][SendPchgCmd] = 0. • Wait 25 MEMCLKs. On an activate command, the LR-DIMM stores the address of the physical rank so it can direct subsequent CAS commands. Therefore, only one row per bank can be active for a logical rank of a LR-DIMM. For LR-DIMMs, the behavior of precharge and precharge all commands issued to physical ranks associated with the logical rank selected by D18F2x28C_dct[1:0][CmdChipSelect] depends on the setting of F0RC14. In broadcast mode, the command is issued to each physical rank. In rank addressable mode, the command is issued to a specific physical rank BIOS specifies by programming D18F2x28C_dct[1:0][CmdAddress[17:14]]. 2.9.5.9.6.1.2 Read and Write Command Generation BIOS performs the following steps for read pattern generation: • Program D18F2x278_dct[1:0], and D18F2x274_dct[1:0] with the data comparison masks for bit lanes of 127 42300 Rev 3.12 - July 14, 2015 • • • • • • • BKDG for AMD Family 15h Models 10h-1Fh Processors interest. Program D18F2x270_dct[1:0][DataPrbsSeed] the seed for the desired PRBS. Program D18F2x260_dct[1:0][CmdCount] equal to the number of cache line commands. Program D18F2x25[8,4]_dct[1:0] to the initial address. Program D18F2x250_dct[1:0] as follows: • ResetAllErr and StopOnErr as desired. See 2.9.5.9.6.1.3 [PRBS Data Comparison]. • CmdTgt corresponding to D18F2x25[8,4]_dct[1:0]. • CmdType = 000b. • SendCmd = 1. If D18F2x260_dct[1:0][CmdCount] != 0 then Wait for D18F2x250_dct[1:0][TestStatus] = 1 and D18F2x250_dct[1:0][CmdSendInProg] = 0. Program D18F2x250_dct[1:0][SendCmd] = 0. Read D18F2x264_dct[1:0], D18F2x268_dct[1:0], and D18F2x26C_dct[1:0] if applicable. BIOS performs the following steps for write pattern generation: • Program D18F2x270_dct[1:0][DataPrbsSeed] the seed for the desired PRBS. • Program D18F2x260_dct[1:0][CmdCount] equal to the number of cache line commands desired. • Program D18F2x25[8,4]_dct[1:0] to the initial address. • Program D18F2x250_dct[1:0] as follows: • CmdTgt corresponding to D18F2x25[8,4]_dct[1:0]. • CmdType = 001b. • SendCmd = 1. • If D18F2x260_dct[1:0][CmdCount] != 0 then Wait for D18F2x250_dct[1:0][TestStatus] = 1 and D18F2x250_dct[1:0][CmdSendInProg] = 0. • Program D18F2x250_dct[1:0][SendCmd] = 0. BIOS combines the two sets of steps listed above for alternating write and read pattern generation. 2.9.5.9.6.1.3 PRBS Data Comparison The DCT compares the incoming read data against the expected PRBS sequence specified by D18F2x270_dct[1:0][DataPrbsSeed] during pattern generation. BIOS may choose to continue command generation and accumulate errors or stop command generation on the first error occurrence by programming D18F2x250_dct[1:0][StopOnErr]. Error information is reported via D18F2x264_dct[1:0], D18F2x268_dct[1:0], D18F2x26C_dct[1:0], D18F2x294_dct[1:0], and D18F2x298_dct[1:0]can be masked on per-bit basis by programming D18F2x274_dct[1:0] and D18F2x278_dct[1:0]. BIOS resets the error information by programming D18F2x250_dct[1:0][ResetAllErr]=1. Error information is only valid in certain modes of D18F2x250_dct[1:0][CmdType, CmdTgt] and D18F2x260_dct[1:0][CmdCount] and when using 64 byte aligned addresses in D18F2x25[8,4]_dct[1:0][TgtAddress]. Some modes require a series of writes to setup a DRAM data pattern. See Table 44. 128 42300 Rev 3.12 - July 14, 2015 Table 44. BKDG for AMD Family 15h Models 10h-1Fh Processors Command Generation and Data Comparison Commands Read CmdType CmdTgt 000b 00b1 1 Write-Read 010b 01b 00b Maximum CmdCount 128 2562 Infinite 01b3 2562 1. Requires setup writes to store a data pattern in DRAM. The write commands are generated using the same CmdTgt, CmdCount, and DataPrbsSeed settings. 2. D18F2x254[TgtAddress] != D18F2x258[TgtAddress]. 3. Requires setup writes to store a data pattern in DRAM. The write commands are generated programming D18F2x254[TgtAddress] to the intended Target B, CmdTgt=00b, CmdCount to 1/2 of the intended command count, and the same DataPrbsSeed setting. 2.9.5.10 DRAM Channel Disable The following steps are performed to disable an unused DRAM channel: For the channel to be disabled: 1. 2. 3. 4. Program D18F2x9C_x0000_000C_dct[1:0][CKETri] = 1111b. Wait 24 MEMCLKs. Program D18F2x94_dct[1:0][DisDramInterface] = 1. If channel 0 is disabled, then program D18F2x9C_x0D0F_E018_dct[0][PhyPSMasterChannel] = 1. 2.9.5.11 DRAM Phy Power Savings For power savings, BIOS should perform the following actions for each channel: 1. 2. 3. 4. 5. 6. 7. 8. 9. Program D18F2x88_dct[1:0][MemClkDis] to disable unused MEMCLK pins. Program D18F2x9C_x0D0F_2[F,2:0]30_dct[1:0][PwrDn] for unused MEMCLK pairs. Program D18F2x9C_x0000_000C_dct[1:0][CKETri, ODTTri, ChipSelTri] to disable unused pins. Program D18F2x9C_x0D0F_0[F,7:0]13_dct[1:0]_mp[1:0][DllDisEarlyU] = 1. then Program D18F2x9C_x0D0F_0[F,7:0]13_dct[1:0]_mp[1:0][DllDisEarlyL] = 1. D18F2x9C_x0D0F_0[F,7:0]13_dct[1:0]_mp[1:0][RxDqsUDllPowerDown] = 1. D18F2x9C_x0D0F_812F_dct[1:0][PARTri] = 1. D18F2x9C_x0D0F_812F_dct[1:0][Add17Tri, Add16Tri] = {1b, 1b}. IF (DimmsPopulated == 1) && ((D18F2x9C_x0000_0000_dct[1:0]_mp[1:0][CkeDrvStren]==010b) || (D18F2x9C_x0000_0000_dct[1:0]_mp[1:0][CkeDrvStren]==011b)) THEN program D18F2x9C_x0D0F_C000_dct[1:0]_mp[1:0][LowPowerDrvStrengthEn] = 1 ELSE program D18F2x9C_x0D0F_C000_dct[1:0]_mp[1:0][LowPowerDrvStrengthEn] = 0 ENDIF. 10. Program D18F2x9C_x0D0F_0[F,7:0]10_dct[1:0]_mp[1:0][EnRxPadStandby] = IF (D18F2x94_dct[1:0][MemClkFreq] <= 800 MHz) THEN 1 ELSE 0 ENDIF. 11. Program D18F2x9C_x0000_000D_dct[1:0]_mp[1:0] as follows: • If (DDR rate < = 1600) TxMaxDurDllNoLock = RxMaxDurDllNoLock = 8h else TxMaxDurDllNoLock = RxMaxDurDllNoLock = 7h. • TxCPUpdPeriod = RxCPUpdPeriod = 011b. 129 42300 Rev 3.12 - July 14, 2015 BKDG for AMD Family 15h Models 10h-1Fh Processors • TxDLLWakeupTime = RxDLLWakeupTime = 11b. 12. Program D18F2x248_dct[1:0]_mp[1:0] and then D18F2x9C_x0D0F_0[F,7:0]13_dct[1:0]_mp[1:0] as follows: • For M1 context program RxChMntClkEn=RxSsbMntClkEn=0. • For M0 context program RxChMntClkEn=RxSsbMntClkEn=1. 2.9.6 Memory Interleaving Modes The processor supports two different types of memory interleaving modes: • Chip select: interleaves the physical address space over multiple DIMM ranks on a channel, as opposed to each DIMM owning single consecutive address spaces. See 2.9.6.1 [Chip Select Interleaving]. • Channel: interleaves the physical address space over multiple channels, as opposed to each channel owning single consecutive address spaces. See 2.9.6.2 [Channel Interleaving]. Any combination of these interleaving modes may be enabled concurrently. 2.9.6.1 Chip Select Interleaving The chip select memory interleaving mode has the following requirements: • The number of chip selects interleaved is a power of two. • The chip selects are the same size and type. A BIOS algorithm for programming D18F2x[5C:40]_dct[1:0] [DRAM CS Base Address] and D18F2x[6C:60]_dct[1:0] [DRAM CS Mask] in memory interleaving mode is as follows: 1.Program all DRAM CS Base Address and DRAM CS Mask registers using contiguous normalized address mapping. 2.For each enabled chip select, swap the corresponding BaseAddr[38:27] bits with the BaseAddr[21:11] bits as defined in Table 45. 3.For each enabled chip select, swap the corresponding AddrMask[38:27] bits with the AddrMask[21:11] bits as defined in Table 45. Table 45. DDR3 Swapped Normalized Address Lines for CS Interleaving DIMM Chip Select Address Size Map1 0001b 256-MB 0010b 512-MB 0101b 1-GB Swapped Base Address and Address (BankSwap, Mask bits DctSelIntLvAddr)2 0b,N/A 1b,100b 1b,101b 0b,N/A 1b,100b 1b,101b 0b,N/A 1b,100b 1b,101b 4 way CS interleaving 2 way CS interleaving [29:28] and [17:16] [29:28] and [12:11] [29:28] and [13:12] [30:29] and [17:16] [30:29] and [12:11] [30:29] and [13:12] [31:30] and [17:16] [31:30] and [12:11] [31:30] and [13:12] [28] and [16] [28] and [11] [28] and [12] [29] and [16] [29] and [11] [29] and [12] [30] and [16] [30] and [11] [30] and [12] 130 42300 Rev 3.12 - July 14, 2015 Table 45. DDR3 Swapped Normalized Address Lines for CS Interleaving DIMM Chip Select Address Size Map1 0111b BKDG for AMD Family 15h Models 10h-1Fh Processors Swapped Base Address and Address (BankSwap, Mask bits 2 DctSelIntLvAddr) 4 way CS interleaving 2 way CS interleaving 2-GB 0b,N/A [32:31] and [17:16] [31] and [16] 1b,100b [32:31] and [12:11] [31] and [11] 1b,101b [32:31] and [13:12] [31] and [12] 1010b 4-GB 0b,N/A [33:32] and [17:16] [32] and [16] 1b,100b [33:32] and [12:11] [32] and [11] 1b,101b [33:32] and [13:12] [32] and [12] 1011b 8-GB 0b,N/A [34:33] and [18:17] [33] and [17] 1b,100b [34:33] and [12:11] [33] and [11] 1b,101b [34:33] and [13:12] [33] and [12] 1. See D18F2x80_dct[1:0] [DRAM Bank Address Mapping]. 2. See D18F2xA8_dct[1:0][BankSwap] and D18F2x110[DctSelIntLvAddr]. If BankSwap==1 && (D18F2x110[DctSelIntLvEn]==0 || DctSelIntLvAddr!=100b) then software swaps the bits shown in the rows for address bit 9 channel interleaving. The following is an example of interleaving a 64-bit interface to DDR3 DRAM. The DRAM memory consists of two 512 MB dual rank DDR3 DIMMs. 1. The register settings for contiguous memory mapping are: • D18F2x80_dct[1:0] = 0000_0011h. // CS0/1 = 256 MB; CS2/3 = 256 MB • D18F2x40 = 0000_0001h. // 0 MB base • D18F2x44 = 0010_0001h. // 256 MB base = 0 MB + 256 MB • D18F2x48 = 0020_0001h. // 512 MB base = 256 MB + 256 MB • D18F2x4C = 0030_0001h. // 768 MB base = 512 MB + 256 MB • D18F2x60 = 0008_FFE0h. // CS0/CS1 = 256 MB • D18F2x64 = 0008_FFE0h. // CS2/CS3 = 256 MB 2. The base address bits to be swapped are defined in Table 45, 256MB chip select size, 4 way CS interleaving column. The BaseAddr[29:28] bits are specified by D18F2x[5C:40]_dct[1:0][21:20]. The BaseAddr[17:16] bits are specified by D18F2x[5C:40]_dct[1:0][11:10]. • D18F2x40 = 0000_0001h. • D18F2x44 = 0000_0401h. • D18F2x48 = 0000_0801h. • D18F2x4C = 0000_0C01h. 3. The AddrMask bits to be swapped are the same as the BaseAddr bits defined in the previous step. The AddrMask[29:28] bits are specified by D18F2x[6C:60]_dct[1:0][21:20]. The AddrMask[17:16] bits are specified by D18F2x[6C:60]_dct[1:0][11:10]. • D18F2x60 = 0038_F3E0h. • D18F2x64 = 0038_F3E0h. 4. If BankSwap is enabled and DCT channel interleaving is enabled on system address bit 8, then the Base and AddrMask bits to be swapped are as follows: • D18F2x40 = 0000_0001h. • D18F2x44 = 0000_0021h. 131 42300 Rev 3.12 - July 14, 2015 BKDG for AMD Family 15h Models 10h-1Fh Processors • D18F2x48 = 0000_0041h. • D18F2x4C = 0000_0061h. • D18F2x60 = 0038_FF90h. • D18F2x64 = 0038_FF90h. 5. If BankSwap is enabled and DCT channel interleaving is enabled on system address bit 9, then the Base and AddrMask bits to be swapped are as follows: • D18F2x40 = 0000_0001h. • D18F2x44 = 0000_0041h. • D18F2x48 = 0000_0081h. • D18F2x4C = 0000_00C1h. • D18F2x60 = 0038_FF20h. • D18F2x64 = 0038_FF20h. 2.9.6.2 Channel Interleaving The channel memory interleaving mode requires that DIMMs are present on both channels. Channel interleaving is enabled by programming D18F2x110[DctSelIntLvEn] and D18F2x110[DctSelIntLvAddr] to specify how interleaving is performed between the DCTs. If the channels do not have the same amount of DRAM, D18F2x110[DctSelBaseAddr, DctSelHi, DctSelHiRngEn] are used to configure the interleaved region. See also 2.9.7 [Memory Hoisting]. 2.9.7 Memory Hoisting Memory hoisting reclaims the otherwise inaccessible DRAM that would naturally reside in memory regions used by MMIO. When memory hoisting is configured by BIOS, DRAM physical addresses are repositioned above the 4 GB address level in the address map. In operation, the physical addresses are remapped in hardware to the normalized addresses used by a DCT. The region of DRAM that is hoisted is defined to be from D18F1xF0[DramHoleBase] to the 4 GB level or from FD_0000_0000h to the 1 TB level. Hoisting is enabled by programming D18F1xF0 [DRAM Hole Address] and configuring the DCTs per the equations in this section. DramHoleSize is defined in order to simplify the following equations in this section and is calculated as follows: • Define the DRAM hole region as DramHoleSize[31:24] = 100h - D18F1xF0[DramHoleBase[31:24]]. 2.9.7.1 DramHoleOffset Programming D18F1xF0[DramHoleOffset] is programmed based on the scenarios shown in Figure 5: • Case 1: If D18F2x110[DctSelHiRngEn] = 0 OR D18F2x110[DctSelHiRngEn] = 1 AND DctSelBaseAddr > DramHoleBase then: DramHoleOffset[31:23] = {DramHoleSize[31:24], 0b} + {DramBaseAddr[31:27], 0000b}; • Case 2: If D18F2x110[DctSelIntLvEn] = 0 AND D18F2x110[DctSelHiRngEn] = 1 AND DctSelBaseAddr < DramHoleBase then: DramHoleOffset[31:23] = {DramHoleSize[31:24], 0b} + {DctSelBaseAddr[31:27], 0000b}; • Case 3: If D18F2x110[DctSelIntLvEn] = 1 AND D18F2x110[DctSelHiRngEn] = 1 AND DctSelBaseAddress < DramHoleBase then DramHoleOffset[31:23] = {DramHoleSize[31:24], 0b} + {DramBaseAddr[31:27], 0000b} + {0b, (DctSelBaseAddr[31:27] - DramBaseAddr[31:27]), 000b}; 132 42300 Rev 3.12 - July 14, 2015 BKDG for AMD Family 15h Models 10h-1Fh Processors 7G - DramLimitAddr DctHi DctLo 6G - DctSelBaseAddr 5G MMIO hole DCTx or DCT0/1 interleaved 7G - DramLimitAddr DctHi MMIO hole 6G DctHi 5G 4G 4G 3G - DramHoleBase 3G - DramHoleBase 2G 2G - DctSelBaseAddr DctLo Case 1 7G - DramLimitAddr 1G - DramBaseAddr 1G - DramBaseAddr 0G 0G Case 2 6G 5G MMIO hole 4G 3G - DramHoleBase 2G - DctS elBase Addr DCT0/1 interleaved 1G - DramBase Addr 0G Case 3 Figure 5: Example Cases for Programming DramHoleOffset 2.9.7.2 DctSelBaseOffset Programming When D18F2x110[DctSelHiRngEn] = 1, D18F2x114[DctSelBaseOffset] is programmed based on the scenarios shown in Figure 6: • Case 1: If D18F2x110[DctSelIntLvEn] = 0 then: DctSelBaseOffset[47:26] = {DctSelBaseAddr[47:27], 0b}; • Case 2: If D18F2x110[DctSelIntLvEn] = 1 AND (D18F1xF0[DramHoleValid] = 1 AND DctSelBaseAddr < DramHoleBase OR D18F1xF0[DramHoleValid] = 0) then: DctSelBaseOffset[47:26] = {DramBaseAddr[47:27], 0b} + {0b, (DctSelBaseAddr[47:27] – DramBaseAddr[47:27])}; • Case 3: • If D18F2x110[DctSelIntLvEn] = 1 AND D18F1xF0[DramHoleValid] = 1 AND DctSelBaseAddr > DramHoleBase then: DctSelBaseOffset[47:26] = {DramBaseAddr[47:27], 0b} + {0000h, DramHoleSize[31:26]} + {0b, (DctSelBaseAddr[47:27] – {0000h, (DramBaseAddr[31:27] + DramHoleSize[31:27])}}; 133 42300 Rev 3.12 - July 14, 2015 BKDG for AMD Family 15h Models 10h-1Fh Processors 7G - DramLimitAddr DctHi DctLo 6G - DctSelBaseAddr 7G - DramLimitAddr DctHi DctHi 6G 5G 6G - DctS elBase Addr 5G 4G MMIO Hole 7G - DramLimitAddr 3G - DramHoleBase 5G MMIO Hole 4G MMIO Hole 3G - DramHoleBase 2G 2G - DctSelBaseAddr 1G - DramBaseAddr DCT0/1 interleaved 0G Case 1 4G 3G - DramHoleBase DCT0/1 interleaved 2G 1G - DramBaseAddr 1G - DramBase Addr 0G 0G Case 3 Case 2 Figure 6: Example Cases for Programming DctSelBaseOffset 2.9.8 DRAM CC6/PC6 Storage DRAM is used to hold the state information of cores entering the CC6 power management state. As part of the system setup if CC6 or PC6 is enabled, BIOS configures a special region of DRAM to hold the state information. In operation, hardware protects this region from general system accesses while allowing the cores access during C-state transitions. The size of each special DRAM storage region is defined to be a fixed 16MB. BIOS configures the storage region at the top of the DRAM range and adjusts D18F1x[144:140,44:40][DramLimit] downward accordingly. See Table 46. After finalizing the system DRAM configuration, BIOS must set D18F2x118[LockDramCfg] = 1 to enable the hardware protection. Table 46. Example storage region configuration DRAM Node Populated 0 2.9.9 256 MB CC6 D18F1x[144:140,44:40] DRAM [DramBase, DramLimit] Range 0 MB, 240 MB - 1 240 MB, 256 MB - 1 D18F4x128 D18F1x120[DramBaseAddr], [CoreStateSa D18F1x124[DramLimitAddr] veDestNode] 0 0 MB, 256 MB - 1 DRAM On DIMM Thermal Management and Power Capping Each DCT can throttle commands based on the state of the channel EVENT_L pin or when D18F2xA4[BwCapEn]=1. The EVENT_L pin is used for thermal management while D18F2xA4[BwCapEn] limits memory power independent of the thermal management solution. 134 42300 Rev 3.12 - July 14, 2015 BKDG for AMD Family 15h Models 10h-1Fh Processors The EVENT_L pin for each channel must be wire OR’ed. If two DCTs are enabled then both throttle commands in lockstep using the amount specified in D18F2xA4[CmdThrottleMode] and D18F2xA4[BwCapCmdThrottleMode]. The recommended BIOS configuration for the EVENT_L pin is as follows: • BIOS may enable command throttling on a DRAM controller if the platform supports the EVENT_L pin by programming D18F2xA4[ODTSEn] = 1. • The recommended usage is for this pin to be connected to one or more JEDEC defined on DIMM temperature sensors. The DIMM SPD ROM indicates on DIMM temperature sensor support. • BIOS configures the temperature sensor(s) to assert EVENT_L pin active low when the trip point is exceeded and deassert EVENT_L when the temperature drops below the trip point minus the sensor defined hysteresis. • BIOS programs D18F2xA4[CmdThrottleMode] with the throttling mode to employ when the trip point has been exceeded. • The hardware enforces a refresh rate of 3.9 us while EVENT_L is asserted. • BIOS configures D18F2x8C_dct[1:0][Tref] based on JEDEC defined temperature range options, as indicated by the DIMM SPD ROM. The two defined temperature ranges are normal (with a case temperature of 85 °C) and extended (with a case temperature of 95 °C). • If all DIMMs support the normal temperature range, or if normal and extended temperature range DIMMs are mixed, BIOS programs D18F2x8C_dct[1:0][Tref] to 7.8 us and D18F2xA4[ODTSEn] = 1. BIOS configures the temperature sensor trip point for all DIMMs according to the 85 °C case temperature specification. • If all DIMMs support the extended temperature range, BIOS has two options: a. Follow the recommendation for normal temperature range DIMMs. b. Program D18F2x8C_dct[1:0][Tref] = 3.9 us and configure the temperature sensor trip point for all DIMMs according to the 95 °C case temperature specification. • At startup, the BIOS determines if the DRAMs are hot before enabling a DCT and delays for an amount of time to allow the devices to cool under the influence of the thermal solution. This is accomplished by checking the temperature status in the temperature sensor of each DIMM. • The latched status of the EVENT_L pin for both DCTs can be read by system software in D18F2xAC[MemTempHot1, MemTempHot0]. The relationship between the DRAM case temperature, trip point, and EVENT_L pin sampling interval is outlined as follows: • The trip point for each DIMM is ordinarily configured to the case temperature specification minus a guardband temperature for the DIMM. • The temperature guardband is vendor defined and is used to account for sensor inaccuracy, EVENT_L pin sample interval, and platform thermal design. • The sampling interval is vendor defined. It is expected to be approximately 1 second. BIOS may enable bandwidth capping on a DRAM controller by setting D18F2xA4[BwCapEn] = 1 and programming D18F2xA4[BwCapCmdThrottleMode] with the throttling mode to employ. The DCT will employ the larger of the two throttling percentages as specified by D18F2xA4[BwCapCmdThrottleMode] and D18F2xA4[CmdThrottleMode] if the EVENT_L pin is asserted when both D18F2xA4[BwCapEn] = 1 and D18F2xA4[ODTSEn] = 1. 135 42300 Rev 3.12 - July 14, 2015 BKDG for AMD Family 15h Models 10h-1Fh Processors 2.10 Thermal Functions Thermal functions SB-TSI, HTC, PROCHOT_L and THERMTRIP are intended to maintain processor temperature in a valid range by: • Providing a signal to external circuitry for system thermal management like fan control. • Lowering power consumption by switching to lower-performance P-state. • Sending processor to the THERMTRIP state to prevent it from damage. The processor thermal-related circuitry includes (1) the temperature calculation circuit (TCC) for determining the temperature of the processor and (2) logic that uses the temperature from the TCC. 2.10.1 The Tctl Temperature Scale Tctl is a processor temperature control value used for processor thermal management. Tctl is accessible through SB-TSI and D18F3xA4[CurTmp]. Tctl is a temperature on its own scale aligned to the processors cooling requirements. Therefore Tctl does not represent a temperature which could be measured on the die or the case of the processor. Instead, it specifies the processor temperature relative to the maximum operating temperature, Tctl,max. Tctl is defined as follows for all parts: A: For Tctl = Tctl_max to 255.875: the temperature of the part is [Tctl - Tctl_max] over the maximum operating temperature. The processor may take corrective actions that affects performance, such as HTC, to support the return to Tctl range A. B: For Tctl = 0 to Tctl_max - 0.125: the temperature of the part is [Tctl_max - Tctl] under the maximum operating temperature. Tctl 255.875 A Maximum operating temperature Tctl_max B 0.000 Figure 7: Tctl scale 136 42300 Rev 3.12 - July 14, 2015 2.10.2 BKDG for AMD Family 15h Models 10h-1Fh Processors Temperature Slew Rate Control The temperature slew rate controls in D18F3xA4 are used to filter processor the processor temperature provided in D18F3xA4[CurTmp] and through SB-TSI. Separate controls are provided for increasing and decreasing temperatures. The latest measured temperature is referred to as Tctlm below. If downward slew control is enabled (D18F3xA4[TmpSlewDnEn]), Tctl is not updated down unless Tctlm remains below Tctl for a time specified by D18F3xA4[PerStepTimeDn]. If at any point before the timer expires Tctlm equals or exceeds Tctl, then the timer resets and Tctl is not updated. If the timer expires, then Tctl is reduced by 0.125. If downard slew control is disabled, then if Tctlm is less than Tctl, Tctl is immediately updated to Tctlm. The upward slew control works similar to downward slew control except that if Tctlm exceeds Tctl by a value defined by D18F3xA4[TmpMaxDiffUp] then Tctl is immediately updated to Tctlm. Otherwise, Tctlm must remain above Tctl for time specified by D18F3xA4[PerStepTimeUp] before Tctl is incremented by 0.125. 2.10.3 Sideband Temperature Sensor Interface (SB-TSI) SB-TSI is used by an external SMBus master to access the internal temperature sensor and to specify temperature thresholds.100 kHz standard-mode and 400 kHz fast-mode are supported. 3.4 MHz high-speed mode is not supported. 2.10.4 Temperature-Driven Logic The temperature calculated by the TCC is used by HTC,THERMTRIP, PROCHOT_L, and the serial interface, SB-TSI. 2.10.4.1 PROCHOT_L and Hardware Thermal Control (HTC) The processor HTC-active state is characterized by (1) the assertion of PROCHOT_L, (2) reduced power consumption, and (3) reduced performance. While in the HTC-active state, the processor reduces power consumption by limiting all cores to a P-state (specified by D18F3x64[HtcPstateLimit]). See 2.5.3 [CPU Power Management]. While in the HTC-active state, software should not change the following: All D18F3x64 fields (except for HtcActSts and HtcEn). Any change to the previous list of fields when in the HTC-active state can result in undefined behavior. HTC status and control is provided through D18F3x64. The PROCHOT_L pin acts as both an input and as an open-drain output. As an output, PROCHOT_L is driven low to indicate that the HTC-active state has been entered due to an internal condition, as described by the following text. The minimum assertion and deassertion time for PROCHOT_L is 200 us. The processor enters the HTC-active state if all of the following conditions are true: • D18F3xE8[HtcCapable]=1 • D18F3x64[HtcEn]=1 • PWROK=1 • THERMTRIP_L=1 • The processor is not in the C3 ACPI state. and any of the following conditions are true: • Tctl is greater than or equal to the HTC temperature limit (D18F3x64[HtcTmpLmt]). • PROCHOT_L=0 The processor exits the HTC-active state when all of the following are true: 137 42300 Rev 3.12 - July 14, 2015 BKDG for AMD Family 15h Models 10h-1Fh Processors • Tctl is less than the HTC temperature limit (D18F3x64[HtcTmpLmt]). • Tctl has become less than the HTC temperature limit (D18F3x64[HtcTmpLmt]) minus the HTC hysteresis limit (D18F3x64[HtcHystLmt]) since being greater than or equal to the HTC temperature limit (D18F3x64[HtcTmpLmt]). • PROCHOT_L=1. 2.10.4.2 Local Hardware Thermal Control (LHTC) The LHTC-active state is characterized by (1) reduced power consumption and (2) reduced performance. While in the LHTC-active state, the processor reduces power consumption by limiting the maximum P-state specified by D0F0xBC_x1F628[LhtcActivePstateLimit]. See 2.5.3.1 [Core P-states]. LHTC is supported only under Battery Power with BAPM disabled. While in the LHTC-active state, software should not change D0F0xBC_x1F628[LhtcActivePstateLimit]. Any change to the previous list of fields when in the LHTC-active state can result in undefined behavior. The LHTC trip point is specified by D0F0xBC_x1F89C[LhtcGtempLimitHi]. The LHTC-active state is independent from the HTC-active state. LHTC does not affect PROCHOT_L output and is not affected by PROCHOT_L input. The processor enters the LHTC-active state if all of the following conditions are true: • PWROK is asserted. • The processor is not in the package C6 (PC6) state. • Tctl is greater than or equal to LHTC temperature limit D0F0xBC_x1F89C[LhtcGtempLimitHi]. • BAPM is disabled. See 2.5.9 [Bidirectional Application Power Management (BAPM)]. The processor exits the LHTC-active state when the following is true: • Tctl is less than LHTC temperature low limit (D0F0xBC_x1F89C[LhtcGtempLimitLo]) since being greater than or equal to the LHTC temperature high limit (D0F0xBC_x1F89C[LhtcGtempLimitHi]). 2.10.4.3 Software P-state Limit Control D18F3x68 [Software P-state Limit] provides a software mechanism to limit the P-state MSRC001_0061[CurPstateLimit]. See 2.5.3 [CPU Power Management]. 2.10.4.4 THERMTRIP If the processor supports the THERMTRIP state (as specified by D18F3xE4 [Thermtrip Status][ThermtpEn] or CPUID Fn8000_0007_EDX[TTP], which are the same) and the temperature approaches the point at which the processor may be damaged, the processor enters the THERMTRIP state. The THERMTRIP function is enabled after cold reset (after PWROK asserts and RESET_L deasserts). It remains enabled in all other processor states, except during warm reset (while RESET_L is asserted). The THERMTRIP state is characterized as follows: • • • • The THERMTRIP_L signal is asserted. Nearly all clocks are gated off to reduce dynamic power. A low-value VID is generated. In addition, the external chipset is expected to place the system into the S5 ACPI state (power off) if THERMTRIP_L is detected to be asserted. A cold reset is required to exit the THERMTRIP state. 138 42300 Rev 3.12 - July 14, 2015 BKDG for AMD Family 15h Models 10h-1Fh Processors 2.11 Root Complex 2.11.1 Overview NB Bus 0 D evice 0 Function 0 Root com plex D evice 0 Function 2 IO M M U D evice 1 Function 0 Internal G raphics D evice 1 Function 1 H D A udio D evice 2 G fx0 Root Port B ridge D evice 6 G PP2 Root Port B ridge D evice 3 G fx1 Root Port B ridge D evice 7 G PP3 Root Port B ridge D evice 4 G PP0 R oot Port B ridge D evice 8* B ridge to FC H D evice 5 G PP1 R oot Port B ridge * D evice 8 is hidden by default so that the FC H appears to be on bus 0. Figure 8: Root complex topology 2.11.2 Interrupt Routing The RC remaps PCI defined INTx interrupts based on the device number of the virtual bridge or internal device that the interrupt is received from. Table 47: INTx Mapping Device Number 1 2 3 4 INTA INTB INTC INTD INTA INTB INTC INTD INTA INTB INTC INTA INTB INTC INTD INTB INTC INTD 139 42300 Rev 3.12 - July 14, 2015 BKDG for AMD Family 15h Models 10h-1Fh Processors Table 47: INTx Mapping Device Number 5 6 7 2.11.3 INTA INTB INTC INTD INTB INTC INTD INTA INTC INTD INTA INTB INTD INTA INTB INTC Links 2.11.3.1 Overview There are 7 configurable ports, which can be divided into 2 groups: • Gfx: Contains 2 x8 ports. Each port can be limited to lower link widths for applications that require fewer lanes. Additionally, the two ports can be combined to create a single x16 link. • GPP: Contains 1 x4 UMI and 4 General Purpose Ports (GPP). Gfx and GPP ports each have a Type 1 Virtual PCI-to-PCI bridge header in the PCI configuration space mapped to devices according to Figure 8. FS1r2 and FM2 processors can be configured for either Gen1 or Gen2 mode. For FP2 processors see D18F3x1FC[Fp2PcieGen2Sup]. Each PCIe and DDI lane is assigned a unique lane ID that software uses to communicate configuration information to the SMU. Table 48 shows the mappings between lane ID’s and lanes. Table 48: Lane Id Mapping Lane Id Lane Lane Id Lane Lane Id Lane Lane Id Lane 0 P_UMI_[T,R]X[P,N]0 10 P_GFX_[T,R]X[P,N]2 20 P_GFX_[T,R]X[P,N]12 30 DP1_TX[P,N]2 1 P_UMI_[T,R]X[P,N]1 11 P_GFX_[T,R]X[P,N]3 21 P_GFX_[T,R]X[P,N]13 31 DP1_TX[P,N]3 2 P_UMI_[T,R]X[P,N]2 12 P_GFX_[T,R]X[P,N]4 22 P_GFX_[T,R]X[P,N]14 32 DP2_TX[P,N]0 3 P_UMI_[T,R]X[P,N]3 13 P_GFX_[T,R]X[P,N]5 23 P_GFX_[T,R]X[P,N]15 33 DP2_TX[P,N]1 4 P_GPP_[T,R]X[P,N]0 14 P_GFX_[T,R]X[P,N]6 24 DP0_TX[P,N]0 34 DP2_TX[P,N]2 5 P_GPP_[T,R]X[P,N]1 15 P_GFX_[T,R]X[P,N]7 25 DP0_TX[P,N]1 35 DP2_TX[P,N]3 6 P_GPP_[T,R]X[P,N]2 16 P_GFX_[T,R]X[P,N]8 26 DP0_TX[P,N]2 36 DP2_TX[P,N]4 7 P_GPP_[T,R]X[P,N]3 17 P_GFX_[T,R]X[P,N]9 27 DP0_TX[P,N]3 37 DP2_TX[P,N]5 8 P_GFX_[T,R]X[P,N]0 18 P_GFX_[T,R]X[P,N]10 28 DP1_TX[P,N]0 38 DP2_TX[P,N]6 9 P_GFX_[T,R]X[P,N]1 19 P_GFX_[T,R]X[P,N]11 29 DP1_TX[P,N]1 2.11.3.2 Link Configurations Lanes of the Gfx ports can be assigned to IO links or DDI links. The following link configurations are supported for the Gfx links: Table 49: Supported Gfx Port Configurations D0F0xE4 Gfx Port Lanes1 x0131_0080 x0111_0011 x0211_0011 x0131_8021 x0131_8022 x0131_8013 0000_0000h 0200_0000h 0200_0000h 7654_3210h 7654_3210h 0000_0001h 0000_0005h 0001_0000h 0001_0000h 7654_3210h 7654_3210h 0000_0001h 3:0 7:4 11:8 15:12 x16 Link x8 Link x8 Link 140 42300 Rev 3.12 - July 14, 2015 BKDG for AMD Family 15h Models 10h-1Fh Processors Table 49: Supported Gfx Port Configurations D0F0xE4 Gfx Port Lanes1 x0131_0080 x0111_0011 x0211_0011 x0131_8021 x0131_8022 x0131_8013 3:0 7:4 11:8 15:12 0000_0005h 0001_0000h 0000_0300h 7654_3210h 7654_3210h 0000_0001h x8 Link x4 Link DDI 0000_0005h 0001_0000h 0000_0300h 5476_3210h 5476_3210h 0000_0001h x8 Link DDI x4 Link 0000_0000h 0001_0000h 0000_0300h 7654_3210h 7654_3210h 0000_0001h x8 Link DDI DDI 0000_0000h 0001_0000h 0001_0000h 7654_3210h 7654_3210h 0000_0001h x8 Link 0000_0005h 0000_0300h 0000_0300h 7632_5410h 7632_5410h 0000_0001h x4 Link x4 Link 0000_0005h 0000_0300h 0001_0000h 7632_5410h 7632_5410h 0000_0001h x4 Link x4 Link 0000_0005h 0000_0300h 0001_0000h 7654_3210h 7654_3210h 0000_0004h x4 Link DDI 0000_0005h 0000_0300h 0000_0300h 7654_3210h 7654_3210h 0000_0004h x4 Link DDI Dual-DVI DDI DDI Dual-DVI x8 Link x4 Link DDI 0000_0005h 0000_0300h 0000_0300h 5476_3210h 5476_3210h 0000_0008h x4 Link DDI DDI x4 Link 0000_0000h 0000_0300h 0000_0300h 7654_3210h 7654_3210h 0000_0001h x4 Link DDI DDI DDI 0000_0000h 0000_0300h 0001_0000h 7654_3210h 7654_3210h 0000_0001h x4 Link DDI Dual-DVI 0000_0005h 0000_0300h 0000_0300h 5410_7632h 5410_7632h 0000_0004h unused DDI x4 Link x4 Link 0000_0000h 0001_0000h 0001_0000h 3210_7654h 3210_7654h 0000_0004h Dual-DVI x8 Link 0000_0005h 0001_0000h 0000_0300h 5410_7632h 5410_7632h 0000_0004h Dual-DVI x4 Link x4 Link 0000_0000h 0001_0000h 0000_0300h 7610_3254h 7610_3254h 0000_0004h Dual-DVI x4 Link DDI 0000_0000h 0001_0000h 0000_0300h 1054_3276h 1054_3276h 0000_0008h Dual-DVI DDI x4 Link 0000_0000h 0001_0000h 0000_0300h 7654_3210h 7654_3210h 0000_0001h Dual-DVI DDI DDI 0000_0000h 0001_0000h 0001_0000h 7654_3210h 7654_3210h 0000_0001h Dual-DVI Dual-DVI Table 50: Supported Gfx Port Configurations for D0F0xE4_x013[3:1]_804[3:0] D0F0xE4 Gfx Port Lanes1 x0131_8040 x0131_8041 x0131_8042 x0131_8043 3:0 7:4 11:8 15:12 0000_0000h 0000_0000h 0000_0000h 0000_0000h 0000_0000h 0000_0000h 0000_0000h 0000_0000h x8 Link 0000_0000h 0000_0000h 0000_0000h 0000_0001h x8 Link x4 Link DDI 0000_0000h 0000_0000h 0000_0001h 0000_0000h x8 Link DDI x4 Link 0000_0000h 0000_0000h 0000_0001h 0000_0001h x8 Link DDI DDI 0000_0000h 0000_0000h 0000_0001h 0000_0001h x8 Link 0000_0000h 0000_0000h 0000_0001h 0000_0001h x4 Link x4 Link 0000_0000h 0000_0000h 0000_0001h 0000_0001h x4 Link x4 Link 0000_0000h 0000_0001h 0000_0000h 0000_0000h x4 Link DDI 0000_0000h 0000_0001h 0000_0000h 0000_0001h x4 Link DDI x16 Link x8 Link Dual-DVI DDI DDI Dual-DVI x8 Link x4 Link DDI 0000_0000h 0000_0001h 0000_0001h 0000_0000h x4 Link DDI DDI x4 Link 0000_0000h 0000_0001h 0000_0001h 0000_0001h x4 Link DDI DDI DDI 0000_0000h 0000_0001h 0000_0001h 0000_0001h x4 Link DDI Dual-DVI 0000_0000h 0000_0001h 0000_0000h 0000_0000h unused DDI x4 Link x4 Link 0000_0001h 0000_0001h 0000_0000h 0000_0000h Dual-DVI x8 Link 0000_0001h 0000_0001h 0000_0000h 0000_0000h Dual-DVI x4 Link x4 Link 0000_0001h 0000_0001h 0000_0000h 0000_0001h Dual-DVI x4 Link DDI 0000_0001h 0000_0001h 0000_0001h 0000_0000h Dual-DVI DDI x4 Link 0000_0001h 0000_0001h 0000_0001h 0000_0001h Dual-DVI DDI DDI 141 42300 Rev 3.12 - July 14, 2015 BKDG for AMD Family 15h Models 10h-1Fh Processors Table 50: Supported Gfx Port Configurations for D0F0xE4_x013[3:1]_804[3:0] D0F0xE4 Gfx Port Lanes1 x0131_8040 x0131_8041 x0131_8042 x0131_8043 0000_0001h 0000_0001h 0000_0001h 0000_0001h 3:0 7:4 11:8 Dual-DVI 15:12 Dual-DVI 1. FP2 package processors with TDP <= 25W support 8 lanes only. These can be either lanes [7:0] or [15:8]. To achieve the above configurations, program the following registers: • Program Gfx function in D0F0xE4_x013[1:0]_0080[StrapBifLinkConfig]. • Program Gfx PIF 0 and Gfx PIF 1 in D0F0xE4_x0[2:1]1[3:0]_0011. • Program Gfx TX Lane Mux in D0F0xE4_x013[3:0]_8021. • Program Gfx RX Lane Mux in D0F0xE4_x013[3:0]_8022. • Program Gfx MasterPciePll in D0F0xE4_x013[3:0]_8013. • Program OwnSlice in the Gfx registers in D0F0xE4_x013[3:1]_804[3:0][OwnSlice]. The following DP0/DP1 DDI configurations are supported: Table 51: Supported DP0/DP1 DDI Link Configurations Lanes[3:0] Lanes[7:4] DDI DDI DDI (Dual-link DVI) The following DP2 DDI configurations are supported: Table 52: Supported DP2 DDI Link Configurations Lanes[3:0] Lanes[6:4] DDI Unused DDI (Dual-link DVI) The following link configurations are supported for the GPP links: Table 53: Supported General Purpose (GPP) Link Configurations D0F0xE4 GPP Port Lane x0130_0080 x0110_0011 Lanes[0:3] 0000_0001h 0000_0300h 4 5 x4 UMI 6 7 x4 Link 0000_0002h 0000_010Ch x4 UMI x2 Link x2 Link 0000_0003h 0000_0104h x4 UMI x2 Link x1 Link x1 Link 0000_0004h 0000_0100h x4 UMI x1 Link x1 Link x1 Link x1 Link 142 42300 Rev 3.12 - July 14, 2015 2.11.3.3 BKDG for AMD Family 15h Models 10h-1Fh Processors Clocking DISP_CLKIN CLKIN DPLL 0 DisplayStream DDI Phy PCIe PLL CascadedPllSel GppTxClkA 8 Instances (1 Per Nibble) PllMode[1:0]=10b Gfx Phy 0 GfxTxClkC PCIe PLL GfxTxClkD ÷2 PllMode[1:0]=11b GPP PCIe Phy PLL GfxTxClkA GfxTxClkB BypassClk DPLL 1 GppTxClkB PCIePllSel TxClk DdiTxClkA DdiTxClkB TxClk to Phy PllMode[1:0]=00b Gfx PCIe Phy 1 PLL PllMode[1:0]=01b Figure 9: Phy clock configuration Each phy contains a PLL and 2 clock picker circuits. Each PLL can be controlled independently of the phy and has its clock signal routed to the phys. The PLL in Gfx[8:15] is not connected and should be turned off to conserve power. The GPP PLL is always placed in PCI Express mode and is used as the transmit clock (TxClk) for all PCI Express links. The remaining two PLLs can be configured as clocks for DDI modes. Setting up the link clocking scheme includes: • Configuring PLL behavior. See 2.11.4.2.1 [Clock Configuration]. • Selecting the clock source for phys, DDI streams, cores, and phy interface. 2.11.4 2.11.4.1 Root Complex Configuration LPC MMIO Requirements To ensure proper operation of LPC generated DMA requests, the UMI must be configured to send processor generated MMIO writes that target the LPC bus to the FCH as non-posted writes. To ensure this requirement the MMIO address space of the LPC bus must not be included in the ranges specified by D18F1x[1CC:180,BC:80] [MMIO Base/Limit] and non-posted protocol for memory writes must be enabled using the following sequence before LPC DMA transactions are initiated. 1. Configure the FCH to use the non-posted write protocol. See the FCH register specification for configuration details. 2. Program D0F0xE4_x0101_0010[UmiNpMemWrite] = 1. 3. Program D0F0x98_x06[UmiNpMemWrEn] = 1. 143 42300 Rev 3.12 - July 14, 2015 2.11.4.2 BKDG for AMD Family 15h Models 10h-1Fh Processors Link Configuration and Initialization Link configuration and initialization is performed by the fowllowing sequence: 1. IF(UMI) THEN program D0F0x64_x00[NbFchCfgEn] = 1. 2. Clock Configuration (see 2.11.3.3 [Clocking] and 2.11.4.2.1 [Clock Configuration]). 3. 2.11.4.2.2 [Link Configuration and Core Initialization] 4. 2.11.4.2.3 [Link Training] 5. 2.11.4.4 [Power Management] 6. Lock link configuration registers. • Program D0F0xE4_x0[2:1]01_0010[HwInitWrLock] = 1. • Program D0F0x64_x00[HwInitWrLock] = 1. 7. IF(UMI) THEN program D0F0x64_x00[NbFchCfgEn] = 0. 2.11.4.2.1 Clock Configuration Clock configuration is required after each cold or warm reset performed. The configuration applies to all phys, regardless of being configured as an IO link or DDI link. 1. Program the PLL to be powered off: A. Program D0F0xE4_x0[2:1]1[3:0]_001[3:2][PllPowerStateInOff]=111b. B. Program D0F0xE4_x0[2:1]1[3:0]_001[3:2][PllRampUpTime]=000b. C. Program D0F0xE4_x0[2:1]1[3:0]_0010[Ls2ExitTime]=000b. 2. Power down sub-link PLLs: A. Program D0F0xE4_x0130_8023[LaneEnable]=0Fh. See D0F0xE4_x013[3:0]_8023 B. Program D0F0xE4_x0131_8023[LaneEnable]=00h. C. Program D0F0xE4_x0132_8023[LaneEnable]=00h. 3. Wait for D0F0xE4_x0[2:1]1[3:0]_0015[7:0]==FFh. 4. Program the PLL mode: A. Program D0F0xE4_x0121_2005[PllMode]=01b. See D0F0xE4_x0[2:1]2[3:0]_2005 B. Program D0F0xE4_x0221_2005[PllMode] = 00b. C. Program D0F0xE4_x0122_2005[PllMode] = 10b. 5. Enable the individual lanes: A. Program D0F0xE4_x0130_8023[LaneEnable]=FFh. See D0F0xE4_x013[3:0]_8023 B. Program D0F0xE4_x0131_8023[LaneEnable]=FFFFh. C. Program D0F0xE4_x0132_8023[LaneEnable]=FFh. D. Wait for D0F0xE4_x0[2:1]1[3:0]_0015[7:0]==FFh. 6. Program D0F0xE4_x0[2:1]1[3:0]_001[3:2][PllPowerStateInOff]=000b. 2.11.4.2.2 Link Configuration and Core Initialization Link configuration is done on a per link basis. Lane reversal, IO link/DDI link selection, and lane enablement is configured through this sequence. 1. Place software-reset module into blocking mode: A. Program D0F0xE4_x013[3:0]_8062[ConfigXferMode]=0. B. Program D0F0xE4_x013[3:0]_8062[BlockOnIdle]=0. 2. If the link is an IO link, Program D0F0xE4_x0[2:1]01_0011[DynClkLatency]=Fh. 3. Program D0F0xE4_x013[1:0]_0080 per Table 49 and Table 53. 4. Program D0F0xE4_x013[3:0]_8021 per Table 49 and Table 53. 5. Program D0F0xE4_x0[2:1]1[3:0]_0010[RxDetectTxPwrMode]=1. 6. Program D0F0xE4_x0[2:1]1[3:0]_0010[Ls2ExitTime]=000b. 144 42300 Rev 3.12 - July 14, 2015 BKDG for AMD Family 15h Models 10h-1Fh Processors 7. Program D0F0xE4_x013[3:0]_8013[MasterPciePllA, MasterPciePllB, MasterPciePllC, MasterPciePllD] per Table 49. 8. Initiate core reconfiguration sequence: A. Program D0F0xE4_x013[3:0]_8062[ReconfigureEn]=1. B. Program D0F0xE4_x013[3:0]_8060[Reconfigure]=1. C. Wait for D0F0xE4_x013[3:0]_8060[Reconfigure]==0. D. Program D0F0xE4_x013[3:0]_8062[ReconfigureEn]=0. 9. Return software-reset module to non-blocking mode: A. Program D0F0xE4_x013[3:0]_8062[ConfigXferMode]=1. 10. Program D[8:2]F0xE4_xC1[StrapReverseLanes] if necessary. 11. Program D0F0xE4_x0[2:1]1[3:0]_0011 per Table 49 and Table 53. 12. Program D0F0xE4_x0[2:1]2[3:0]_[7:6][7:6,3:0][8,0]5 per Table 105. 13. For each link mapped to DDI: A. Program D0F0xE4_x0[2:1]1[3:0]_001[3:2][PllPowerStateInTxs2]=111b. B. Program D0F0xE4_x0[2:1]1[3:0]_001[3:2][PllPowerStateInOff]=111b. C. Program D0F0xE4_x0[2:1]1[3:0]_001[3:2][PllRampUpTime]=010b. 14. For each nibble that has no PCIe lanes in use: A. Program D0F0xE4_x0[2:1]1[3:0]_001[3:2][PllPowerStateInOff]=111b. B. Program D0F0xE4_x0[2:1]1[3:0]_001[3:2][PllPowerStateInTxs2]=111b. C. Program D0F0xE4_x0[2:1]1[3:0]_001[3:2][TxPowerStateInTxs2]=111b. D. Program D0F0xE4_x0[2:1]1[3:0]_001[3:2][RxPowerStateInRxs2]=111b. 15. For each lane that is not in use, program the corresponding D0F0xE4_x013[3:0]_8023[LaneEnable]=0. 16. If the link is a DDI link: A. Program D0F0xE4_x013[3:1]_804[3:0][OwnSlice] per Table 49. 17. Configure PIF parings and disable ganged mode for UMI: A. Program D0F0xE4_x0110_0011=0000_0300h. B. Program D0F0xE4_x0120_6[3:2][8,0]5[GangedModeEn]=0. 2.11.4.2.3 Link Training Link training is performed on a per link basis. BIOS may train the links in parallel. 2.11.4.3 2.11.4.3.1 Miscellaneous Features Straps 1. Program D0F0xE4_x013[3:0]_8011[StrapBifValid]=1. 2. Program strap values. 3. Program D0F0xE4_x013[3:0]_8011[StrapBifValid]=0. 2.11.4.3.2 Lane Reversal Normally, the lanes of each port are physically numbered from n-1 to 0 where n is the number of lanes assigned to the port. Physical lane numbering can be reversed according to the following methods: • To reverse the physical lane numbering for a specific port, program D[8:2]F0xE4_xC1[StrapReverseLanes]=1 according to the sequence in 2.11.4.3.1 [Straps]. • To reverse the physical lane numbering for all ports in the GPP or GFX interfaces, program D0F0xE4_x0[2:1]01_00C0[StrapReverseAll]=1 according to the sequence in 2.11.4.3.1 [Straps]. Note that logical port numbering is established during link training regardless of the physical lane numbering. 145 42300 Rev 3.12 - July 14, 2015 2.11.4.3.3 BKDG for AMD Family 15h Models 10h-1Fh Processors Link Speed Changes Link speed changes can only occur on Gen2 capable links. To verify that Gen2 speeds are supported verify D[8:2]F0x64[LinkSpeed]==02h. 2.11.4.3.3.1 Software Initiated Link Speed Changes The following programming sequence describes a software initiated speed change from Gen1 to Gen2: 1. Verify D[8:2]F0xE4_xA4[LcGen2EnStrap]==1. If LcGen2EnStrap is not 1, program it to 1. 2. Program D[8:2]F0x88[TargetLinkSpeed]=2. 3. Program D[8:2]F0xE4_xA4[LcGoToRecovery]=1. 4. Verify D[8:2]F0xE4_xA4[LcOtherSideSupportsGen2]==1, otherwise stop. 5. If D[8:2]F0x68[LinkSpeed]==2h, stop. The link is already at Gen2 speed. 6. If D[8:2]F0xE4_xA4[LcSpeedChangeAttemptFailed]==1, stop. The maximum number of speed negotiation failures have been reached. 7. Ensure D[8:2]F0xE4_xA4[LcForceDisSwSpeedChange]=0. 8. Ensure D[8:2]F0x88[HwAutonomousSpeedDisable]=0. 9. Program D[8:2]F0xE4_xA4[LcInitiateLinkSpeedChange]=1. The following programming sequence describes a software initiated speed change from Gen2 to Gen1: 1. Program D[8:2]F0x88[TargetLinkSpeed]=1h. 2. If D[8:2]F0x68[LinkSpeed]==1h, stop. The link is already at Gen1 speed. 3. If D[8:2]F0xE4_xA4[LcSpeedChangeAttemptFailed]==1, stop. The maximum number of speed negotiation failures have been reached. 4. Ensure D[8:2]F0xE4_xA4[LcForceDisSwSpeedChange]=0. 5. Ensure D[8:2]F0x88[HwAutonomousSpeedDisable]=0. 6. Program D[8:2]F0xE4_xA4[LcInitiateLinkSpeedChange]=1. 2.11.4.3.3.2 Autonomous Link Speed Changes To enable autonomous speed changes on a per port basis: 1. Program D[8:2]F0x88[TargetLinkSpeed]=2h. 2. Program D0F0xE4_x013[1:0]_0[C:8]03[StrapBifDeemphasisSel]=1. 3. Program D[8:2]F0xE4_xA4[LcGen2EnStrap]=1. 4. Program D[8:2]F0xE4_xC0[StrapAutoRcSpeedNegotiationDis]=0. 5. Program D[8:2]F0xE4_xA4[LcMultUpstreamAutoSpdChangEn]=1. 6. Program D[8:2]F0xE4_xA2[LcUpconfigureDis]=0. 2.11.4.3.4 Deemphasis Deemphasis strength can be changed on a per-port basis by programming D[8:2]F0xE4_xB5[LcSelectDeemphasis]. 2.11.4.4 2.11.4.4.1 Power Management Link States To enable support for L1 program D[8:2]F0xE4_xA0[LcL1Inactivity]=6h. To enable support for L0s: • Program D[8:2]F0xE4_xA1[LcDontGotoL0sifL1Armed]=1. • Program D[8:2]F0xE4_xA0[LcL0sInactivity]=9h. 146 42300 Rev 3.12 - July 14, 2015 2.11.4.4.2 BKDG for AMD Family 15h Models 10h-1Fh Processors Dynamic Link-width Control Dynamic link-width control is a power saving feature that reconfigures the link to run with fewer lanes. The inactive lanes are turned off to conserve power. The GFX links can switch among widths of: x1, x2, x4, x8, and x16. The GPP links can switch among widths of: x1, x2, and x4. There are 3 link-width control mechanisms: • Long Reconfiguration: The link is brought down and retrained to a different width. This mechanism is only supported on links between AMD products. • Short Reconfiguration: The link is retrained to a different width by going through the recovery state of the LTSSM. This mechanism is only supported on links between AMD products. • Up/Down Reconfiguration: The link is retrained according to the PCI Express specification. This mechanism is only available between Gen2 devices. The core has the capability to turn off the inactive lanes of trained links. To enable this feature program D[8:2]F0xE4_xA2[LcDynLanesPwrState]=11b. 2.11.4.5 2.11.4.5.1 Link Test and Debug Features Compliance Mode To enable Gen1 software compliance mode program D[8:2]F0xE4_xC0[StrapForceCompliance]=1 for each port to be placed in compliance mode. To enable Gen2 software compliance mode: 1. BIOS enables Gen2 capability by programming D0F0xE4_x0[2:1]01_00C1[StrapGen2Compliance]=1. 2. Program D0F0xE4_x013[1:0]_0[C:8]03[StrapBifDeemphasisSel]=1 for each port to be placed in compliance mode. 3. Program D[8:2]F0x88[TargetLinkSpeed]=2h for each port to be placed in compliance mode. 4. Program D[8:2]F0x88[EnterCompliance]=1 for each port to be placed in compliance mode. 2.11.5 BIOS Timer The root complex implements a 32-bit microsecond timer (see D0F0xE4_x0130_80F0 and D0F0xE4_x0130_80F1) that the BIOS can use to accurately time wait operations between initialization steps. To ensure that BIOS waits a minimum number of microseconds between steps BIOS should always wait for one microsecond more than the required minimum wait time. 2.12 IOMMU The processor includes an IOMMU revision 2 for FM2 packages only. See the AMD I/O Virtualization Technology (IOMMU) Specification. 2.12.1 IOMMU Configuration Space The IOMMU configuration space consists of the following four groups: • PCI Configuration space. See 3.4 [Device 0 Function 2 (IOMMU) Configuration Registers]. • IOMMU Memory Mapped Register space. See 3.15 [IOMMU Memory Mapped Registers]. 147 42300 Rev 3.12 - July 14, 2015 BKDG for AMD Family 15h Models 10h-1Fh Processors • IOMMU L1 Indexed space accessed through D0F2xF8 [IOMMU L1 Config Index]. • IOMMU L2 Indexed space accessed through D0F2xF0 [IOMMU L2 Config Index]. 2.12.2 IOMMU Initialization BIOS should perform the following steps to initialize the IOMMU: 1. Program D0F0x64_x0D[Dev0Fn2RegEn]=1. 2. Program D0F2x44 [IOMMU Base Address Low] and D0F2x48 [IOMMU Base Address High] to allocate a 512K region of MMIO space for IOMMU memory mapped registers. This region of MMIO space is reserved for IOMMU and BIOS must not allocate it for use by system software. 3. Program D0F2x44[IommuEnable]=1. 4. Program D[8:2]F0xE4_xC1[StrapExtendedFmtSupported]=1 and D[8:2]F0xE4_xC1[StrapE2EPrefixEn]=1. 5. Program the registers with BIOS recommendations in L1 (D0F2xF8) and L2 (D0F2xF0) indexed space. 6. Check if any PCIe devices in the system support the Phantom Function. For each PCIe core that has a connected device advertising support for the Phantom Function, program D0F2xFC_x07_L1sel[3:0][PhantomSupEn]=1 for the L1 corresponding to that PCIe core. 7. If a PCIe port is hot-plug capable, then program D0F2xFC_x07_L1sel[3:0][PhantomSupEn]=1 for the L1 corresponding to the PCIe core. 8. If at least one PCIe to PCI-x bridge exists on a PCIe port or a HotPlug capable PCIe slot is present on a PCIe port then program D0F2xFC_x0D_L1sel[3:0][VOQPortBits]=111b for the L1 corresponding to the particular PCIE core. 148 42300 Rev 3.12 - July 14, 2015 BKDG for AMD Family 15h Models 10h-1Fh Processors 2.13 System Management Unit (SMU) The system management unit (SMU) is a subcomponent of the northbridge that is responsible for a variety of system and power management tasks during boot and runtime. The SMU contains a microcontroller to assist with many of these tasks. 2.13.1 Software Interrupts The microcontroller can be interrupted to cause it to perform several initialization and runtime tasks. BIOS and ACPI methods can interrupt the SMU to request a specific action using the following sequence: 1. Wait for D0F0xBC_xE0003004[IntDone]==1. 2. Program D0F0xBC_xE0003000[ServiceIndex] to the desired service index and toggle D0F0xBC_xE0003000[IntReq]. This may be done in a single write. 3. Wait for D0F0xBC_xE0003004[IntAck]==1. After performing the steps above, software may continue execution before the interrupt has been serviced. However, software should not rely on the results of the interrupt until the service is complete (see D0F0xBC_xE0003004[IntDone]). Interrupting the SMU with a service index that does not exist results in undefined behavior. Table 54: BIOS Services Service Index 16h Notes Description: SMC_MSG_CONFIG_NBDPM. Enables NB P-state Adjustments. See 2.5.4.1 [NB P-states]. Input: D0F0xBC_x1F5F8 Output: D0F0xBC_x1F5FC Firmware revision: 149 42300 Rev 3.12 - July 14, 2015 BKDG for AMD Family 15h Models 10h-1Fh Processors 2.14 Graphics Processor (GPU) The processor contains an integrated DX11 compliant graphics processor. 2.14.1 GPU PCI Interface BIOS must configure the PCI interface block of the GPU before the GPU can be accessed. The PCI interface block is configured by programming the D0F0x64_x1C [Internal Graphics PCI Control 1] register. Two writes to this register are required. The first write sets the configuration and must have D0F0x64_x1C[WriteDis]=0. The second write uses the same write data with D0F0x64_x1C[WriteDis]=1. 2.14.2 Graphics Memory Controller (GMC) The graphics memory controller is responsible for servicing memory requests from the different blocks within the GPU and forwarding routing them to the appropriate interface. The GMC is also responsible for translating GPU virtual address to GPU physical addresses and for translating GPU physical addresses to system addresses. 2.14.3 Frame Buffer (FB) The frame buffer is defined as the portion of system memory dedicated for GPU use. Table 55: Recommended Frame Buffer Configurations System Memory Size < 1GB >= 1GB & < 2GB >= 2GB & < 4GB >= 4GB Frame Buffer Size 64 MB 256 MB 512 MB 768 MB 150 42300 Rev 3.12 - July 14, 2015 BKDG for AMD Family 15h Models 10h-1Fh Processors 2.15 RAS Features This section applies reliability, availability, and serviceability, or RAS, and related considerations. 2.15.1 Machine Check Architecture The processor contains logic and registers to control detection, corrective action, logging, and reporting of errors in the data or control paths in each core and the northbridge. This section assumes familiarity with the AMD64 Architecture Programmer's Manual Volume 2: System Programming chapter titled “Machine Check Mechanism”. See 1.2 [Reference Documents]. The ability of hardware to generate a machine check exception upon an error is indicated by CPUID Fn0000_0001_EDX[MCE]. 2.15.1.1 Machine Check Registers CPUID Fn0000_0001_EDX[MCA] indicates the presence of the following machine check registers: • MSR0000_0179 [Global Machine Check Capabilities (MCG_CAP)] • Reports how many machine check register banks are supported. • MSR0000_017A [Global Machine Check Status (MCG_STAT)] • MSR0000_017B [Global Machine Check Exception Reporting Control (MCG_CTL)] The error reporting register banks supported are: • MC0: load-store unit (LS), including data cache. • MC1: instruction fetch unit (IF), including instruction cache. • MC2: combined unit (CU), including L2 cache. • MC3: Reserved. • MC4: northbridge (NB). These MSRs are also accessible from configuration space. There is only one NB error reporting bank, independent of the number of cores. • MC5: execution unit (EX), including mapper/scheduler/retire/execute functions and fixed-issue reorder buffer. • MC6: floating point unit (FP). The register types within each bank are: • MCi_CTL: The Machine Check Control Register: Enables error reporting. • MCi_STATUS: The Machine Check Status Register: Logs information associated with errors. • MCi_ADDR: The Machine Check Address Register: Logs address information associated with errors. • MCi_MISC: The Machine Check Miscellaneous Register: Log miscellaneous information associated with errors, as defined by each error type. • MCi_CTL_MASK: The Machine Check Control Mask Register: Inhibit detection of an error source unless otherwise specified. Table 56 identifies the registers associated with each register bank: 151 42300 Rev 3.12 - July 14, 2015 BKDG for AMD Family 15h Models 10h-1Fh Processors Table 56: MCA Register Cross-Reference Table Register Bank MC0 MC1 MC2 MC3 MC4 CTL MSR0000_0400 MSR0000_0404 MSR0000_0408 MSR0000_040C MSR0000_0410 STATUS MSR0000_0401 MSR0000_0405 MSR0000_0409 MSR0000_040D MSR0000_0411 MC5 MC6 MSR0000_0414 MSR0000_0415 MSR0000_0418 MSR0000_0419 MCA Register ADDR MISC MSR0000_0402 MSR0000_0403 MSR0000_0406 MSR0000_0407 MSR0000_040A MSR0000_040B MSR0000_040E MSR0000_040F MSR0000_0412 MSR0000_0413 MSRC000_0408 MSR0000_0416 MSR0000_0417 MSR0000_041A MSR0000_041B CTL_MASK MSRC001_0044 MSRC001_0045 MSRC001_0046 MSRC001_0047 MSRC001_0048 MSRC001_0049 MSRC001_004A Uncorrectable errors that are enabled in MCi_CTL result in reporting to software via machine check exceptions. Some errors increment a counter in MCi_MISC, which may trigger an interrupt (see 2.15.1.7 [Error Thresholding]). Each register bank i implements a number of machine check miscellaneous registers, denoted as MCi_MISCj, where j goes from 0 to a maximum of 8. The presence of valid information in the first MISC register in the bank (MCi_MISC0) is indicated by MCi_STATUS[MiscV], and in subsequent MISC registers by MCi_MISCj[Val] in the target register. If there is more than one MISC register in a given bank, a non-zero value in MCi_MISC0[BlkPtr] points to the contiguous block of additional registers. 2.15.1.2 Machine Check Errors Classes The classes of machine check errors are: • Uncorrectable • Deferred • Correctable Uncorrectable errors cannot be corrected by hardware or microcode and may have caused the loss of data or corruption of processor state. Uncorrectable errors, if not masked from logging, update the status and address registers, and if enabled for reporting, cause a machine check exception. If there is information in the status and address registers from a previously logged deferred or correctable error, it is overwritten. If an uncorrectable error is masked from logging, the error is ignored by hardware (exceptions are noted in the register definitions). If an uncorrectable error is disabled from reporting, containment of the error and logging/reporting of subsequent errors may be affected. Therefore, unmasked uncorrectable errors should be enabled for reporting for normal operation. Uncorrectable errors should only be disabled from reporting for debug purposes. Deferred errors are errors that cannot be corrected by hardware, but that do not cause an immediate interruption in program flow, loss of data integrity, or corruption of processor state. They are errors that indicate that data has been corrupted but contained; no exception is generated because the data has not been referenced by a core or an IO link. If deferred errors are enabled for logging, the status and address registers in the corresponding register bank are written with information that identifies the source of the error. If there is information in the status and address registers from a previously logged correctable error, it is overwritten. Correctable errors are corrected by hardware or microcode and cause no loss of data or corruption of processor state (unless disabled by implementation-specific bits in the control registers for test or debug reasons). If they are enabled for logging, the status and address registers in the corresponding register bank are written with 152 42300 Rev 3.12 - July 14, 2015 BKDG for AMD Family 15h Models 10h-1Fh Processors information that identifies the source of the error. Correctable errors are not reported via machine check exceptions; some correctable errors may be reported via error thresholding (see 2.15.1.7 [Error Thresholding]). The implications of these categories of errors are: 1. Uncorrected error; the problem was not dealt with by hardware. • Operationally (error handling), action does need to be taken, because program flow is affected. • Diagnostically (fault management), software may collect information to determine if any components should be de-configured or serviced. • Examples include: • Uncorrectable ECC, no way to avoid passing it to process. • Poison data consumed (as opposed to created), no way to avoid passing it to process or link. 2. Deferred error; the problem was partly dealt with by containment. • Operationally, no immediate action needs to be taken, because program flow has not been affected. However, steps may be taken by software to prevent access to the data in error. • Diagnostically, software may collect information to determine if any components should be de-configured or serviced. • Examples include: • Uncorrectable ECC, converted to poison data. 3. Corrected error; the problem was dealt with by hardware. • Operationally, no action needs to be taken, because program flow is unaffected. • Diagnostically, software may collect information to determine if any components should be de-configured or serviced. • Examples include: • Correctable ECC, corrected in-line. For debug observability, D18F3x180[ChgUcToCeEn] can be used to convert NB uncorrectable errors to correctable errors. Machine check conditions can be simulated by using MSRC001_0015[McStatusWrEn]. This is useful for debugging machine check handlers. See 2.15.2 [Error Injection and Simulation] for more detail. 2.15.1.3 Error Detection, Action, Logging, and Reporting Error detection: • Core error detection is enabled if not masked by MCi_CTL_MASK (MSRC001_0044-MSRC001_0046, MSRC001_0049-MSRC001_004A). • NB error detection (MC4) is not affected by MC4_CTL_MASK (MSRC001_0048). • Error masking is performed regardless of MCA bank enablement in MCG_CTL (MSR0000_017B). Error actions are enabled if all of the following are true: • Error detection is enabled. • The condition that enables the action to be taken for a detected error varies. The EAC (Error Action Condition) column of the error description tables defines whether MCG_CTL[i] affects whether the error action is taken. • D: Detected. The error action is taken if the error is detected. MCG_CTL[i] does not affect if the error action is taken. • E: Enabled. The error action is taken only if the bank is enabled by MCG_CTL[i]==1. Error logging is enabled if all of the following are true: • Error detection is enabled. 153 42300 Rev 3.12 - July 14, 2015 BKDG for AMD Family 15h Models 10h-1Fh Processors • The MCA bank is enabled in MCG_CTL (MSR0000_017B). • MC4/NB only: Error logging is not masked. See MC4_CTL_MASK (MSRC001_0048). Error reporting is enabled if all of the following are true: • Error logging is enabled. • The corresponding enable bit for the error in MCi_CTL is set to 1. A machine check exception will be generated if all the following are true: • The error is enabled for reporting. • The error is uncorrectable. • CR4.MCE is enabled. Notes: 1. If error reporting is enabled but CR4.MCE is disabled or the machine check handler cannot be invoked for other reasons, then a reportable error will cause the system to enter shutdown. 2. If error reporting is disabled, the setting of CR4.MCE has no effect. 3. If an uncorrectable error is disabled from reporting, containment of the error and logging/reporting of subsequent errors may be affected. Therefore, unmasked uncorrectable errors should be enabled for reporting for normal operation. Uncorrectable errors should only be disabled from reporting for debug purposes. 4. Errors not associated with a specific core are reflected to core 0 of the compute unit. The error description tables identify which errors are associated or not associated with a specific core of the compute unit. Throughout the MCA register descriptions, the terms “enabled” and “disabled” generally refer to reporting, and the terms “masked” and “unmasked” generally refer to logging, unless otherwise noted. Some logged errors increment a counter in MCi_MISC, which may trigger an interrupt (see 2.15.1.7 [Error Thresholding]). 2.15.1.3.1 MCA conditions that cause Shutdown The following architectural conditions cause the processor to enter the Shutdown state; see “Machine-Check Errors”and subsections in APM volume 2 for more detail; see 1.2 [Reference Documents]: • Attempting to generate an MCE when CR4.MCE=0. • Attempting to generate an MCE when MSR0000_017A[MCIP]=1. The following non-architectural conditions cause the processor to enter the Shutdown state: • EX “Retire dispatch queue parity” error. See Table 230 [EX Error Descriptions]. • EX “Mapper checkpoint array parity” error if UC=1. See Table 230 [EX Error Descriptions]. 2.15.1.3.2 Error Logging During Overflow An error to be logged when the status register contains valid data can result in an overflow condition. During error overflow conditions, the new error may not be logged or an error which has already been logged in the status register may be overwritten. For the rules on error overflow, priority, and overwriting, see MSR0000_0401[Overflow] and MSR0000_0411[Overflow]. Uncorrectable errors require software intervention. Therefore, when an uncorrectable error cannot be logged, critical error information may have been lost, and MCi_STATUS[PCC] may be set. If PCC is indicated, software should terminate system processing to prevent data corruption (see 2.15.1.6 [Handling Machine Check Exceptions]). If PCC is not indicated, any MCA data lost due to overflow was informational only and not critical to system hardware operation. 154 42300 Rev 3.12 - July 14, 2015 BKDG for AMD Family 15h Models 10h-1Fh Processors Table 57 indicates which errors are overwritten in the error status registers. Table 57: Overwrite Priorities for All Banks Newer Error 2.15.1.4 Uncorrectable Correctable Enabled Disabled Enabled Disabled Older Error Uncorrectable Correctable Enabled Disabled Enabled Disabled Overwrite Overwrite Overwrite Overwrite - MCA Initialization The following initialization sequence must be followed: • MCi_CTL_MASK registers (MSRC001_0044 to MSRC001_004A): • BIOS must initialize the mask registers to inhibit error detection prior to the initialization of MCi_CTL and MSR0000_017B. • BIOS must not clear MASK bits that are reset to 1. • The MCi_CTL registers must be initialized prior to enabling the error reporting banks in MCG_CTL. If initializing after a cold reset (see D18F0x6C[ColdRstDet]), then BIOS must clear the MCi_STATUS MSRs (see Table 56). If initializing after a warm reset, then BIOS should check for valid MCA errors and if present save the status for later use (see 2.15.1.6 [Handling Machine Check Exceptions]). BIOS that wishes to ensure continued operation in the event that a machine check occurs during boot may write MCG_CTL with all ones and write zeros into each MCi_CTL. With these settings, a machine check error will result in MCi_STATUS being written without generating a machine check exception or a system shutdown. BIOS may then poll MCi_STATUS during critical sections of boot to ensure system integrity. Before passing control to the operating system, BIOS should restore the values of those registers to what the operating system is expecting. (Note that using MCi_CTL to disable error reporting on uncorrectable errors will affect error containment; see 2.15.1.3 [Error Detection, Action, Logging, and Reporting].) Alternatively, the BIOS initialize MCA without setting CR4.MCE; This will result in a system shutdown on any machine check which would have caused a machine check exception (followed by a reboot if configured in the chipset). 2.15.1.5 Error Code The MCi_STATUS[ErrorCode] field contains information on the logged error. Table 58 [Error Code Types] identifies how to decode ErrorCode. The MCi_STATUS[ErrorCodeExt] field contains detailed, model-specific information that is used for error diagnosis but not error handling; see 2.15.1.6 [Handling Machine Check Exceptions]. For a given error reporting bank, Error Code is used in conjunction with other MCi_STATUS fields to identify the Error Type. Details for each Error Type are described in the error signatures tables accompanying the MCi_STATUS register for each bank: • MC0; Table 213. • MC1; Table 216. • MC2; Table 220. • MC3; Reserved. • MC4; Table 223 and Table 224. 155 42300 Rev 3.12 - July 14, 2015 BKDG for AMD Family 15h Models 10h-1Fh Processors • MC5; Table 231. • MC6; Table 234. Table 58: Error Code Types Error Code Error Code Type Description 0000_0000_0001_TTLL TLB Errors in the GART TLB cache. TT = Transaction Type LL = Cache Level 0000_0001_RRRR_TTLL MEM Errors in the cache hierarchy (not in NB) RRRR = Memory Transaction Type TT = Transaction Type LL = Cache Level 0000_1PPT_RRRR_IILL BUS General bus errors including link and DRAM PP = Participation Processor T = Timeout RRRR = Memory Transaction Type II = Memory or IO LL = Cache Level Table 59: Error Codes: Transaction Type TT 00 01 10 11 Transaction Type I: Instruction D: Data G: Generic Reserved Table 60: Error codes: cache level LL 00 01 10 11 Cache Level Reserved L1: Level 1 L2: Level 2 LG: Generic Table 61: Error Codes: Memory Transaction Type RRRR 0000 0001 0010 0011 0100 0101 Memory Transaction Type GEN: Generic. Includes scrub errors. RD: Generic Read WR: Generic Write DRD: Data Read DWR: Data Write IRD: Instruction Fetch 156 42300 Rev 3.12 - July 14, 2015 BKDG for AMD Family 15h Models 10h-1Fh Processors Table 61: Error Codes: Memory Transaction Type RRRR 0110 0111 1000 Memory Transaction Type Prefetch Evict Probe (Snoop) Table 62: Error Codes: Participation Processor PP 00 01 10 11 Participation Processor SRC: Local node originated the request RES: Local node responded to the request OBS: Local node observed the error as a third party GEN: Generic Table 63: Error Codes: Memory or IO II 00 01 10 11 2.15.1.6 Memory or IO MEM: Memory Access Reserved IO: IO Access GEN: Generic Handling Machine Check Exceptions A machine check handler is invoked to handle an exception for a particular core. Because MCA registers are generally not shared among cores, the handler does not need to coordinate register usage with handler instances on other cores. (Those few MCA registers which are shared are noted in the register description. See also 2.4.1.1 [Registers Shared by Cores in a Compute Unit].) For access to the NB MCA registers, D18F3x44[NbMcaToMstCpuEn] allows a single core (the NBC) to access the registers through MSR space without contention from other cores. This organization of registers on a per core basis allows independent execution, simplifies exception handling, and reduces the number of conditions which are globally fatal. At a minimum, the machine check handler must be capable of logging error information for later examination. The handler should log as much information as is needed to diagnose the error. More thorough exception handler implementations can analyze errors to determine if each error is recoverable. If a recoverable error is identified, the exception handler can attempt to correct the error and restart the interrupted program. An error may not be recoverable for the process or virtual machine it directly affects, but may be containable, so that other processes or virtual machines in the system are unaffected and system operation is recovered; see 2.15.1.6.1 [MCA Differentiation Between System-Fatal and Process-Fatal Errors]. Machine check exception handlers that attempt to recover must be thorough in their analysis and the corrective actions they take. The following guidelines should be used when writing such a handler: • Data collection: • All status registers in the error reporting banks must be examined to identify the cause of the machine check exception. 157 42300 Rev 3.12 - July 14, 2015 BKDG for AMD Family 15h Models 10h-1Fh Processors • Read MSR0000_0179[Count] to determine the number of status registers visible to the core. The status registers are numbered from 0 to one less than the value found in MSR0000_0179[Count]. For example, if the Count field indicates five status registers are supported, they are numbered MC0_STATUS to MC4_STATUS. These are generically referred to as MCi_STATUS. • Check the valid bit in each status register (MCi_STATUS[Val]). The remainder of the status register does not need to be examined when its valid bit is clear. • When identifying the error condition and determining how to handle the error, portable exception handlers should examine the following MCi_STATUS fields: ErrorCode, UC, PCC, CECC, UECC, Deferred, Poison. The expected settings of these and other fields in MCi_STATUS are identified in the error signatures tables which accompany the descriptions of each MCA status register. See 2.15.1.5 [Error Code] for a discussion of error codes and pointers to the error signatures tables. • MCi_STATUS[ErrorCodeExt] should generally not be used by portable code to identify the error condition because it is model specific. ErrorCodeExt is useful in determining the error subtype for root cause analysis. • Error handlers should collect all available MCA information (status register, address register, miscellaneous register, etc.), but should only interrogate details to the level which affects their actions. Lower level details may be useful for diagnosis and root cause analysis, but not for error handling. • Recovery: • Check the valid MCi_STATUS registers to see if error recovery is possible. Error recovery is not possible when the processor context corrupt indicator (MCi_STATUS[PCC]) is set to 1. • The error overflow status indicator (MCi_STATUS[Overflow]) does not indicate whether error recovery is possible. If error recovery is not possible, the handler should log the error information and return to the operating system for system termination. • Check MCi_STATUS[UC] to see if the processor corrected the error. If UC is set, the processor did not correct the error, and the exception handler must correct the error prior to attempting to restart the interrupted program. If the handler cannot correct the error, it should log the error information and return to the operating system. If the error affects only process data, it may be possible to terminate only the affected process or virtual machine. If the error affects processor state, continued use of that processor should not occur. See individual error descriptions for further guidance. • If MSR0000_017A[RIPV] is set, the interrupted program can be restarted reliably at the instruction pointer address pushed onto the exception handler stack if any uncorrectable error has been corrected by software. If RIPV is clear, the interrupted program cannot be restarted reliably, although it may be possible to restart it for debugging purposes. As long as PCC is clear, it may be possible to terminate only the affected process or virtual machine. • When logging errors, particularly those that are not recoverable, check MSR0000_017A[EIPV] to see if the instruction pointer address pushed onto the exception handler stack is related to the machine check. If EIPV is clear, the address is not ensured to be related to the error. • See 2.15.1.6.1 [MCA Differentiation Between System-Fatal and Process-Fatal Errors] for more explanation on the relationship between PCC, RIPV, and EIPV. • Exit: • When an exception handler is able to successfully log an error condition, clear the MCi_STATUS registers prior to exiting the machine check handler. • Prior to exiting the machine check handler, be sure to clear MSR0000_017A[MCIP]. MCIP indicates that a machine check exception is in progress. If this bit is set when another machine check exception occurs in the same core, the processor enters the shutdown state. Additional machine check handler portability can be added by having the handler use the CPUID instruction to identify the processor and its capabilities. Implementation specific software can be added to the machine check exception handler based on the processor information reported by CPUID. 158 42300 Rev 3.12 - July 14, 2015 BKDG for AMD Family 15h Models 10h-1Fh Processors In cases where sync flood is the recommended response to a particular error, a machine check exception cannot be used in lieu of the sync flood to stop the propagation of potentially bad data. 2.15.1.6.1 MCA Differentiation Between System-Fatal and Process-Fatal Errors The bits MCi_STATUS[PCC], MSR0000_017A[RIPV], and MSR0000_017A[EIPV] form a hierarchy, used by software to determine the degree of corruption and recoverability in the system. Table 64 shows how these bits are interpreted. Table 64: Error Scope Hierarchy PCC RIPV EIPV 1 0 0 0 2.15.1.7 1 0/1 0 0/1 Comments Error has corrupted system state (PCC=1), and the process program cannot be restarted (RIPV=0). The error is fatal to the system. Error is confined to the process (PCC=0), and the process program can be restarted (RIPV=1) if any uncorrectable error is corrected by software. Corruption is confined to the process (PCC=0), but the process program cannot be restarted (RIPV=0). Continued operation of this process may not be possible without intervention, however system processing or other processes can continue with appropriate software clean up. Error Thresholding For some types of errors, the hardware maintains counts of the number of errors. When the counter reaches a programmable threshold, an event may optionally be triggered to inform software. This is known as error thresholding. The primary purpose of error thresholding is to help software recognize an excessive rate of errors, which may indicate marginal or failing hardware. This information can be used to make decisions about deconfiguring hardware or scheduling service actions. Counts are incremented for correctable, and uncorrectable errors. The error thresholding hardware counts only the number of errors; it is up to software to track the errors reported over time in order to determine the rate of errors. Thresholding gives error counts on groups of resources. In order to make decisions on individual resources, a finer granularity of error information, such as MCA information for specific errors, must be utilized in order to obtain more accurate counts and to limit the scope of actions to affected hardware. Thresholding is performed for “Error Threshold Groups” identified in the list below. For all error threshold groups, some number of correctable errors is expected and normal. There are numerous factors influencing error rates, including temperature, voltage, operating speed, and geographic location. In order to accommodate the various factors, including software latency to respond and track the error thresholding, additional guardband above the normal rates is recommended before error rates are considered abnormal for purposes of hardware action. The {MC0, MC1, MC2, MC5} error thresholding banks maintains counters, but do not provide interrupts when the threshold is reached; these counters must be polled. Error thresholding groups: • LS (MC0) • LS errors are counted and polled via MSR0000_0403. • LS errors are listed in Table 213 [LS Error Signatures]. • IF (MC1) 159 42300 Rev 3.12 - July 14, 2015 BKDG for AMD Family 15h Models 10h-1Fh Processors • IF errors are counted and polled via MSR0000_0407. • IF errors are listed in Table 216 [IF Error Signatures]. • CU (MC2) • CU errors are counted and polled via MSR0000_040B. • CU errors are listed in Table 220 [CU Error Signatures]. • EX (MC5) • EX errors are counted and polled via MSR0000_0417. • EX errors are listed in Table 231 [EX Error Signatures]. • Links (MC4) • Link errors are counted and reported via MSRC000_0408. • Link errors are the errors listed in Table 222 [NB Error Descriptions] as “L” (Cache) in the ETG (Error Threshold Group) column. • For a link exhibiting excessive errors, it may be possible to reduce errors by lowering the link frequency or reducing the link width (if a bad lane can be avoided). See 2.11 [Root Complex] for details and restrictions on configuring links. In rare circumstances, such as two simultaneous errors in the same error thresholding group, it is possible for one error not to increment the counter. In these conditions, MCi_STATUS[Overflow] may indicate that an overflow occurred, but the error counter may only indicate one error. 2.15.1.8 Error Diagnosis This section describes generalized information and algorithms for diagnosing errors. The primary goal of diagnosis is to identify the failing component for repair purposes. The secondary goal is to identify the smallest possible sub-component for de-allocation, de-configuration, or design/manufacturing root cause analysis. Indictment means identifying the part in error. The simplest form of indictment is self-indictment, where the bank reporting the error is also the unit in error. The next simplest form of indictment is eyewitness indictment, where the part in error is not the bank reporting the error, but is identified unambiguously. Both of these forms can be considered direct indictment; the information for indictment is contained in the MCA error information. If an error is not directly indicted, then identifying the part in error is more difficult and may not be an explicit part of the error log. In general, an address logged in the MCA is useful for direct indictment only if the address identifies a physical location in error, such as a cache index. Logical addresses, while identifying the data, do not identify the location of the data. If possible, physical storage locations in caches should be checked to determine whether the error is a soft error (a temporary upset of the stored value) or a hard fault (malfunctioning hardware). A location which has had a soft error can be corrected by writing a new value to the location; a reread of the location should see the new value. Hard faults cannot be corrected by writing a new value; the hardware persistently returns the previous value. If such checking is not possible, a grossly simplifying assumption can be made that uncorrected errors are hard and corrected errors are soft. Repeated corrected errors from the same location are an indication that the fault is actually hard. Determining whether corrected errors represent a hard fault or a soft error requires understanding the access patterns and any attempts to correct the faulty data in place. An attempt to correct the data in place creates two epochs, one before the correction event and one after. If an error is seen at the same location in two different epochs (especially back-to-back epochs), it is more likely that the cause is a hard fault, since the error has persisted or repeated through an in place correction. The more epochs in which a error is seen, the higher the likelihood of it being caused by a hard fault. 160 42300 Rev 3.12 - July 14, 2015 BKDG for AMD Family 15h Models 10h-1Fh Processors As an example, consider a correctable error found during a read from DRAM. If the DRAM redirect scrubber is enabled (D18F3x5C[ScrubReDirEn]), the data in error is corrected in place, and this event conceptually creates a new epoch. If the original fault was due to a soft error, a read of the same data in the new epoch should not encounter a data error. If the original fault was due to a hard fault (e.g., a stuck bit), a read of the data in the new epoch will likely result in another corrected or uncorrected error. There are numerous correction events that can be used to separate time periods into epochs. These include DRAM redirect scrubs, DRAM sequential scrubs, cache scrubs, cache writes, cache flushes, resets, and others. 2.15.1.8.1 Common Diagnosis Information A common set of diagnosis information is useful for many problems. Table 65 indicates the minimum set of generally useful diagnostic information that should be collected by software, unless the specifics of the problem are known to be narrower, based on the error code or other information. It is useful to collect configuration information to ensure that the behavior is not caused by misconfiguration. Table 65: Registers Commonly Used for Diagnosis MCA Status Bank MC0 MSR0000_0401 MSR0000_0402 MSR0000_0403 MC1 MSR0000_0405 MSR0000_0406 MSR0000_0407 MC2 MSR0000_0409 MSR0000_040A MSR0000_040B MC3 Reserved MC4 MSR0000_0411 MSR0000_0412 MSR0000_0413 MSRC000_0408 D18F3x54 D18F2xAC MC5 MC6 MSR0000_0415 MSR0000_0416 MSR0000_0417 MSR0000_0419 MSR0000_041A MSR0000_041B Configuration MSR0000_0400 MSRC001_1022 MSRC001_0044 MSR0000_0404 MSRC001_0045 MSR0000_0408 MSRC001_0046 MSRC001_1023 Reserved MSR0000_0410 MSRC001_0048 D18F3x40 D18F3x44 D18F3xE4 D18F3xE8 MSRC001_001F D18F3x180 MSR0000_0414 MSRC001_0049 MSR0000_0418 MSRC001_004A If examining MCA registers after startup, determine the cause of the startup: • INIT; D18F0x6C[InitDet]. • Cold reset; D18F0x6C[ColdRstDet]. 161 42300 Rev 3.12 - July 14, 2015 BKDG for AMD Family 15h Models 10h-1Fh Processors • Warm reset; if not INIT or cold reset. To see if a link failure occurred, examine D18F0x84[LinkFail]. If set, look for additional information: • Receipt of a sync, such as during a sync flood, saves a status of Sync Error in MC4_STATUS. • CRC error saves a status of CRC Error in MC4_STATUS. See D18F0x84[CrcErr and CrcFloodEn]. • Link not present does not save status in MC4_STATUS. See D18F0x84[InitComplete]. Other registers may be needed depending on the specific error symptoms. 2.15.2 Error Injection and Simulation Error injection allows the introduction of errors into the system for test and debug purposes. See the following sections for error injection details: • Link: • D18F3x44[GenLinkSel, GenSubLinkSel, GenCrcErrByte1, GenCrcErrByte0]. Error simulation involves creating the appearance to software that an error occurred. This is performed by manually setting the MCA registers with desired values (see MSRC001_0015[McStatusWrEn]), and then driving the software via INT18. McStatusWrEn can be used to debug machine check interrupt handlers. When McStatusWrEn is set, privileged software can write non-zero values to the specified registers without generating exceptions, and then simulate a machine check using the INT18 instruction (INTn instruction with an operand of 18). Setting a reserved bit in these registers does not generate an exception when this mode is enabled. However, setting a reserved bit may result in undefined behavior. 162 42300 Rev 3.12 - July 14, 2015 3 BKDG for AMD Family 15h Models 10h-1Fh Processors Registers This section provides detailed field definitions for the core register sets in the processor. 3.1 Register Descriptions and Mnemonics Each register in this document is referenced with a mnemonic. Each mnemonic is a concatenation of the register-space indicator and the offset of the register. Here are the mnemonics for the various register spaces: • IOXXX: x86-defined input and output address space registers; XXX specifies the hexidecimal byte address of the IO instruction. This space includes IO-space configuration access registers IOCF8 [IO-Space Configuration Address] and IOCFC [IO-Space Configuration Data Port]. • APICXX0: APIC memory-mapped registers; XX0 is the hexidecimal byte address offset from the base address. See 2.4.8.1.2 [APIC Register Space]. • CPUID FnXXXX_XXXX_EiX[_xYYY]: processor capabilities information returned by the CPUID instruction. See 3.17 [CPUID Instruction Registers]. Each core may only access this information for itself. • MSRXXXX_XXXX: MSRs; XXXX_XXXX is the hexidecimal MSR number. This space is accessed through x86-defined RDMSR and WRMSR instructions. Unless otherwise specified there is one set of these registers Per-core. See 2.4.1 [Compute Unit]. • DXFYxZZZ: PCI-defined configuration space; X specifies the hexadecimal device number (this may be 1or 2 digits), Y specifies the function number, and ZZZ specifies the hexidecimal byte address (this may be 2 or 3 digits); e.g., D18F3x40 specifies the register at device 18h, function 3, and address 40h. See 2.7 [Configuration Space], for details about configuration space. • Some register in D18F2xXXX have the _dct[1:0] and/or _mp[1:0] mnemonic suffix. See 2.9.1 [DCT Configuration Registers]. • IOMMUxX_XXXX: IOMMU memory mapped registers; X_XXXX specifies the hexadecimal byte address offset (this may be 2 to 5 digits) from the base address register; The base address for this space is specified by D0F2x44 [IOMMU Base Address Low] and D0F2x48 [IOMMU Base Address High]. See 3.15 [IOMMU Memory Mapped Registers]. • PMCxXXX: performance monitor events; XXX is the hexidecimal event counter number programmed into MSRC001_020[A,8,6,4,2,0] [Performance Event Select (PERF_CTL[5:0])][EventSelect]; See 2.6.1 [Core Performance Monitor Counters]. NBPMCxXXX: NB performance monitor events; XXX is the hexadecimal event counter number programmed into MSRC001_024[6,4,2,0] [Northbridge Performance Event Select (NB_PERF_CTL[3:0])][EventSelect]; See 2.6.2 [NB Performance Monitor Counters]. • When PMCxXXX or NBPMCxXXX is followed by [z:y] then UnitMask[z:y] is being specified. Each mnemonic may specify the location of one or more registers that share the same base definition. A mnemonic that specifies more than one register will contain one or more ranges within braces. The ranges are specified as follows: • Comma separated lists [A,B]: Define specific instances of a register, e.g., D0F3x[1,0]40 defines two registers D0F3x40 and D0F3x140. • Colon separated ranges [A:B]: Defines all registers that contain the range between A and B. Examples: • D0F3x[50:40] defines five registers D0F3x40, D0F3x44, D0F3x48, D0F3x4C, and D0F3x50. • D[8:2]F0x40 defines seven registers D2F0x40, D3F0x40, D4F0x40, D5F0x40, D6F0x40, D7F0x40, and D8F0x40. • D0F0xE4_x013[2:0]_0000 defines three registers D0F0xE4_x0130_0000, D0F0xE4_x0131_0000, and D0F0xE4_x0132_0000. • Colon separated ranges with a explicit step [A:BstepC]: Defines the registers from A to B, C defines the offset between registers., e.g., D0F3x[50:40:step8] defines three registers D0F3x40, D0F3x48, and D0F3x50. 163 42300 Rev 3.12 - July 14, 2015 BKDG for AMD Family 15h Models 10h-1Fh Processors The processor includes a single set of IO-space and configuration-space registers. However, APIC, CPUID, and MSR register spaces are implemented once per processor core. Access to IO-space and configuration space registers may require software-level techniques to ensure that no more than one core attempts to access a register at a time. The following is terminology found in the register descriptions. Table 66: Terminology in Register Descriptions Term BIOS SBIOS CBIOS IBIOS VBIOS OS Driver See Alias IF THEN ELSEIF ELSE ENDIF Access Types Read Read-only Write Write-only Read-write Set-by-hardware Cleared-by-hardware Updated-by-hardware Write-1-to-clear Write-1-only SharedC SharedNC Reset-applied GP-read Definition Software recommendation syntax. See 3.1.2 [Software Recommendation (BIOS, SBIOS, CBIOS, etc.)]. Reference to remote definition. The alias keyword allows the definition of a soft link between two registers. • X is an alias of Y: X is a soft link to the register Y. • X1, X2 are an alias of Y: Both X1 and X2 are soft links to Y. Allows conditional definition as a function of register fields. The syntax is: • IF (conditional-expression) THEN definition ENDIF. • IF (conditional-expression) THEN definition ELSE definition ENDIF. • IF (conditional-expression) THEN definition ELSEIF (conditional-expression) definition ELSE definition ENDIF. Capable of being read by software. Capable of being read but not written by software. Capable of being written by software. Write-only. Capable of being written by software. Reads are undefined. Capable of being written by software and read by software. Register field is set high by hardware, set low by hardware, or updated by hardware. Software must write a 1 to the bit in order to clear it. Writing a 0 to these bits has no affect. Software can set the bit high by writing a 1 to it. Writes of 0 have no effect. Shared by both cores of a compute unit: SharedC (shared coherent) or SharedNC (shared non-coherent). See 2.4.1.1 [Registers Shared by Cores in a Compute Unit] for a definition of coherent. Takes affect on warm reset. GP exception occurs on read. 164 42300 Rev 3.12 - July 14, 2015 BKDG for AMD Family 15h Models 10h-1Fh Processors Table 66: Terminology in Register Descriptions Term GP-write GP-read-write Per-core Per-compute-unit Per-node Not-same-for-all Same-for-all Field Definitions Reserved Unused MBZ RAZ Reset Definitions Reset Cold reset Value 3.1.1 Definition GP exception occurs on write. GP exception occurs on a read or a write. One instance per core. Only valid for MMIO config space. Writes of these bits from one core only affect that core’s register. Reads return the values appropriate to that core. One instance per compute unit. Writes of these bits from one core only affect that compute unit’s register. Reads return the values appropriate to that compute unit. One instance per node. See 3.1.1 [Northbridge MSRs In Multi-Core Products]. Provide indication as to whether all instances of a given register should be the same across all cores/nodes according to the following equation: SameOnAllCheckEnabled = (Writable && (same-for-all | MSR) && ~(not-samefor-all || UpdatedByHw)). UpdatedByHw = (Updated-by-hardware || set-byhardware || cleared-by-hardware || set-when-done || cleared-when-done). Field is reserved for future use. Software is required to preserve the state read from these bits when writing to the register. Software may not depend on the state of reserved fields nor on the ability of such fields to return the state previously written. Field is reserved for future use. Software is not required to preserve the state read from these bits when writing to the register. Software may not depend on the state of unused fields nor on the ability of such fields to return the state previously written. Must be zero. If software attempts to set an MBZ bit to 1, a general-protection exception (#GP) occurs. Read as zero. Writes are ignored, unless RAZ is combined with write, write-1-only or write-once. The reset value of each register is provided below the mnemonic or in the field description. Unless otherwise noted, the register state matches the reset value when RESET_L is asserted (either a cold or a warm reset). Reset values may include: • X: an X in the reset value indicates that the field resets (warm or cold) to an unspecified state. The field state is not affected by a warm reset (even if the field is labeled “cold reset: X”); it is placed into the reset state when PWROK is deasserted. See "Reset" above for the definition of characters that may be found in the cold reset value. The current value of a read-only field or register. A value statement explicitly defines the field or register as read-only and the value returned under all conditions including after reset events. A field labeled “Value:” will not have a separate reset definition. Northbridge MSRs In Multi-Core Products MSRs that control Northbridge functions are shared between all cores on the node in a multi-core processor (e.g. MSR0000_0410). If control of Northbridge functions is shared between software on all cores, software must ensure that only one core at a time is allowed to access the shared MSR. Some MSR’s are conditionally shared; see D18F3x44[NbMcaToMstCpuEn]. 165 42300 Rev 3.12 - July 14, 2015 3.1.2 BKDG for AMD Family 15h Models 10h-1Fh Processors Software Recommendation (BIOS, SBIOS, CBIOS, etc.) The following keywords specify the recommended value to be set by software. • BIOS: AMD BIOS. • SBIOS: Platform BIOS. • CBIOS: Chip-set BIOS. • OS: Operating system. • Driver: Device driver software settings invoked by the OS. Syntax: BIOS:<integer-expression>. Any of the supported tags can be substituted for BIOS. If “BIOS:” occurs in a register field then the recommended value is applied to the field. If “BIOS:” occurs after a register name but outside of a register field table row then the recommended value is applied to the width of the register. 3.1.3 Mapping Tables The following mapping table types are defined. 3.1.3.1 Register Mapping The register mapping table specifies the specific function for each register in a range of registers. Table 182, for example, specifies that the D18F5x160 function is for NB P-state 0. 3.1.3.2 Index Mapping The index mapping table is similar to the register mapping table, but specifies the register by index instead of by full register mnemonic. Table 146, for example, specifies that the D18F2x98_dct[1:0][31:0]==0D0F_0002h, or D18F2x9C_x0D0F_0002_dct[1:0], function is for Byte 0. 3.1.3.3 Field Mapping The field mapping table maps the fields of a range of registers. The rows are the registers that are mapped. Each column specifies a field bit range that is mapped by that column for all registers. The cell at the intersection of the register and the field bit range specifies the suffix that is appended to the register field. “Reserved” specifies that the field is reserved for the register of that row. Table 136, for example, specifies that the fields at D18F2x9C_x0000_0[3:0]01[31:24] should have the suffix Byte3, resulting in WrDatGrossDlyByte3 and WrDatFineDlyByte3. 3.1.3.4 Broadcast Mapping The broadcast mapping table maps a register address to a range of register addresses. The register address is formed by the concatenation of the row address with the column address. The cell at the intersection of the row and column address is a range of register addresses that will be read or written as a group when the row and column address is read or written. Table 163, for example, specifies that a read or write to D18F2x98_dct[1:0][31:0]==0D0F_0F31h will result 166 42300 Rev 3.12 - July 14, 2015 BKDG for AMD Family 15h Models 10h-1Fh Processors in a broadcast read or write to the D18F2x9C_x0D0F_0[8:0]31 range of registers. 3.1.3.5 Valid Values The valid values table defines the valid values for one or more register fields. The valid values table is equivalent in function to the Bits/Description tables in register fields (E.g. D18F0x16C[ForceFullT0]) and is most often used when the table becomes too large and unwieldy to be included into the register field. (E.g. Table 132 [Memory Clock Frequency Value Definition]) 3.2 IO Space Registers See 3.1 [Register Descriptions and Mnemonics] for a description of the register naming convention. IOCF8 IO-Space Configuration Address Reset: 0. IOCF8 [IO-Space Configuration Address], and IOCFC [IO-Space Configuration Data Port], are used to access system configuration space, as defined by the PCI specification. IOCF8 provides the address register and IOCFC provides the data port. Software sets up the configuration address by writing to IOCF8. Then, when an access is made to IOCFC, the processor generates the corresponding configuration access to the address specified in IOCF8. See 2.7 [Configuration Space]. IOCF8 may only be accessed through aligned, DW IO reads and writes; otherwise, the accesses are passed to the appropriate IO link. Accesses to IOCF8 and IOCFC received from an IO link are treated as all other IO transactions received from an IO link and are forwarded based on the settings in D18F1x[DC:C0] [IO-Space Base/Limit]. IOCF8 and IOCFC in the processor are not accessible from an IO link. Bits 31 Description ConfigEn: configuration space enable. Read-write. 1=IO read and write accesses to IOCFC are translated into configuration cycles at the configuration address specified by this register. 0=IO read and write accesses are passed to the appropriate IO link and no configuration access is generated. 30:28 Reserved. 27:24 ExtRegNo: extended register number. Read-write. ExtRegNo provides bits[11:8] and RegNo provides bits[7:2] of the byte address of the configuration register. ExtRegNo is reserved unless it is enabled by MSRC001_001F[EnableCf8ExtCfg]. 23:16 BusNo: bus number. Read-write. Specifies the bus number of the configuration cycle. 15:11 Device: device number. Read-write. Specifies the device number of the configuration cycle. 10:8 Function. Read-write. Specifies the function number of the configuration cycle. 7:2 RegNo: register address. Read-write. See IOCF8[ExtRegNo]. 1:0 Reserved. IOCFC IO-Space Configuration Data Port Bits Description 31:0 Data. Read-write. Reset: 0. See IOCF8. 167 42300 Rev 3.12 - July 14, 2015 3.3 BKDG for AMD Family 15h Models 10h-1Fh Processors Device 0 Function 0 (Root Complex) Configuration Registers See 3.1 [Register Descriptions and Mnemonics] for a description of the register naming convention. See 2.7 [Configuration Space] for details about how to access this space. D0F0x00 Device/Vendor ID Reset: 1410_1022h. Bits Description 31:16 DeviceID: device ID. Read-only. 15:0 VendorID: vendor ID. Read-only. D0F0x04 Status/Command Reset: 0220_0004h. Bits Description 31 ParityErrorDetected: parity error detected. Read-only. 30 SignaledSystemError: signaled system error. Read; write-1-to-clear. 1=FCH generated a system error. 29 ReceivedMasterAbort: received master abort. Read; write-1-to-clear. 28 ReceivedTargetAbort: received target abort. Read; write-1-to-clear. 27 SignalTargetAbort: signaled target abort. Read-only. 26:25 DevselTiming: DEVSEL# Timing. Read-only. 24 Reserved. 23 FastBackCapable: fast back-to-back capable. Read-only. 22 Reserved. 21 PCI66En: 66 MHz capable. Read-only. 20 CapList: capability list. Read-only. 1=Capability list supported. 19:10 Reserved. 9 FastB2BEn: fast back-to-back enable. Read-write. 8 SerrEn: system error enable. Read-write. 7 Reserved. 6 ParityErrorEn: parity error response enable. Read-only. 5 PalSnoopEn: VGA palette snoop enable. Read-only. 4 MemWriteInvalidateEn: memory write and invalidate enable. Read-only. 3 SpecialCycleEn: special cycle enable. Read-only. 2 BusMasterEn: bus master enable. Read-only. 1 MemAccessEn: memory access enable. Read-write. This bit controls if memory accesses by this device are accepted or not. 1=Enabled. 0=Disabled. 0 IoAccessEn: IO access enable. Read-only. 168 42300 Rev 3.12 - July 14, 2015 BKDG for AMD Family 15h Models 10h-1Fh Processors D0F0x08 Class Code/Revision ID Reset: 0600_0000h. Bits Description 31:8 ClassCode: class code. Read-only. Provides the host bridge class code as defined in the PCI specification. 7:0 RevID: revision ID. Read-only. D0F0x0C Header Type Reset: 0080_0000h. Bits Description 31:24 BIST. Read-only. 23:16 HeaderTypeReg. Read-only. 00h=Single function device. 15:8 LatencyTimer. Read-write. 7:0 CacheLineSize. Read-only. D0F0x2C Subsystem and Subvendor ID Reset: 1410_1022h. Bits Description 31:16 SubsystemID. Read-only. 15:0 SubsystemVendorID. Read-only. D0F0x34 Capabilities Pointer Reset: 0000_0000h. Bits Description 31:8 Reserved. 7:0 CapPtr: capabilities pointer. Read-only. There is no capability list. D0F0x4C PCI Control Reset: 0000_1002h. Bits Description 31:27 Reserved. 26 HPDis: hot plug message disable. Read-write. 1=Hot plug message generation is disabled. 25:24 Reserved. 23 MMIOEnable: memory mapped IO enable. Read-write. 1=Decoding of MMIO cycles is enabled. 22:15 Reserved. 169 42300 Rev 3.12 - July 14, 2015 BKDG for AMD Family 15h Models 10h-1Fh Processors 14:12 CfgRdTime. Read-write. Specifies the propagation delay for read data on the configuration bus. Bits Definition 111b-000b <7-CfgRdTime> Clocks. 11 CRS: configuration request retry detected. Read; Write-1-to-clear. 1=Configuration request retry was detected. 10:6 Reserved. 5 SerrDis: system error message disable. Read-write. 1=The generation of SERR messages is disabled. 4 PMEDis: PME disable. Read-write. 1=The generation of PME messages is disabled. 3 Cf8Dis: Cf8 disable. Read-write. 1=Configuration accesses through IO address Cf8h to this device are disabled. 2 Reserved. 1 ApicEnable: APIC enable. Read-write. 1=APIC is enabled. 0 Function1Enable: device 0 function 1 enable. Read-write. 1=Configuration accesses to device 0 function 1 are enabled. D0F0x60 Miscellaneous Index Reset: 0000_0000h. The index/data pair registers D0F0x60 and D0F0x64 is used to access the registers D0F0x64_x[FF:00]. To read or write to one of these register, the address is written first into the address register D0F0x60 and then the data are read or written by read or write the data register D0F0x64. Bits Description 31:8 Reserved. 7 6:0 MiscIndWrEn: miscellaneous index write enable. Read-write. If set writes to D0F0x64 are enabled. MiscIndAddr: miscellaneous index register address. Read-write. D0F0x64 Miscellaneous Index Data See D0F0x60. D0F0x64_x00 Northbridge Control Reset: 0000_0002h. Bits Description 31:8 Reserved. 7 HwInitWrLock. Read-write. 1=Lock HWInit registers. 0=Unlock HWInit registers. 6 NbFchCfgEn: device 8 enable. Read-write. 1=Bridge device 8 is enabled. 5:0 Reserved. 170 42300 Rev 3.12 - July 14, 2015 BKDG for AMD Family 15h Models 10h-1Fh Processors D0F0x64_x0B IOC Link Control Reset: 0000_0000h. Bits Description 31:24 Reserved. 23 IocFchSetPmeTurnOffEn. Read-write. 1=Enables the PME_Turn_Off/PME_To_Ack mechanism between northbridge and FCH. 22 Reserved. 21 IocFchSetPowEn: set slot power enable. Read-write. 1=Enables sending set_slot_power_limit/scale messages to the FCH. 20 SetPowEn: set slot power enable. Read-write. 1=Enables sending set_slot_power messages to the FCH. 19:0 Reserved. D0F0x64_x0C IOC Bridge Control Reset: 0000_0000h. Bits Description 31:8 Reserved. 7 Dev7BridgeDis. Read-write. 1=Bus 0, device 7 bridge functionality is hidden. 6 Dev6BridgeDis. Read-write. 1=Bus 0, device 6 bridge functionality is hidden. 5 Dev5BridgeDis. Read-write. 1=Bus 0, device 5 bridge functionality is hidden. 4 Dev4BridgeDis. Read-write.1=Bus 0, device 4 bridge functionality is hidden. 3 Dev3BridgeDis. Read-write.1=Bus 0, device 3 bridge functionality is hidden. 2 Dev2BridgeDis. Read-write.1=Bus 0, device 2 bridge functionality is hidden. 1:0 Reserved. D0F0x64_x0D IOC PCI Configuration Reset: 0000_0001h. Bits Description 31:1 Reserved. 0 PciDev0Fn2RegEn. Read-write. BIOS: See 2.12.2 [IOMMU Initialization]. 1=Enable configuration accesses to device 0 function 2. 171 42300 Rev 3.12 - July 14, 2015 BKDG for AMD Family 15h Models 10h-1Fh Processors D0F0x64_x16 IOC Advanced Error Reporting Control Reset: 0000_0001h. Bits Description 31:1 Reserved. 0 AerUrMsgEn: AER unsupported request message enable. Read-write. 1=AER unsupported request messages are enabled. D0F0x64_x19 Top of Memory 2 Low Reset: 0000_0000h. Bits Description 31:23 Tom2[31:23]: top of memory 2. Read-write. BIOS: MSRC001_001D[Tom2[31:23]]. This field specifies the maximum system address for upstream read and write transactions that are forwarded to the host bridge. All addresses less than or equal to this system address are forwarded to DRAM and are not checked to determine if the transaction is a peer-to-peer transaction. All upstream reads with addresses greater than this system address are master aborted. 22:1 Reserved. 0 TomEn: top of memory enable. Read-write. BIOS: MSRC001_0010[MtrrTom2En]. 1=Top of memory check enabled. D0F0x64_x1A Top of Memory 2 High Reset: 0000_0000h. Bits Description 31:8 Reserved. 7:0 Tom2[39:32]: top of memory 2. Read-write. BIOS: MSRC001_001D[Tom2[39:32]]. See D0F0x64_x19[Tom2]. D0F0x64_x1C Internal Graphics PCI Control 1 Reset: 0080_0000h. Bits Description 31:24 Reserved. 23 RcieEn. IF (D0F0x64_x1C[WriteDis]==1) THEN Read-only. ELSE Read-write. ENDIF. BIOS: 1. 1=Root complex integrated endpoint mode. 0=Legacy PCI device mode. 22:18 Reserved. 17 F0En. IF (D0F0x64_x1C[WriteDis]==1) THEN Read-only. ELSE Read-write. ENDIF. BIOS:1. 1=Internal graphics enabled. 0=Internal Graphics disabled. 16 IoBarDis. IF (D0F0x64_x1C[WriteDis]==1) THEN Read-only. ELSE Read-write. ENDIF. 1=Graphics IO base address register disabled. 0= Graphics IO base address register enabled. 15:12 Reserved. 172 42300 Rev 3.12 - July 14, 2015 BKDG for AMD Family 15h Models 10h-1Fh Processors 11 Audio64BarEn. IF (D0F0x64_x1C[WriteDis]==1) THEN Read-only. ELSE Read-write. ENDIF. Controls the size of the audio BAR. 1=64-bit BAR. 0=32-bit BAR. 10 AudioNonlegacyDeviceTypeEn. IF (D0F0x64_x1C[WriteDis]==1) THEN Read-only. ELSE Readwrite. ENDIF. BIOS: 0. 1=PCIe device. 0=Legacy PCI device. 9 MsiDis. IF (D0F0x64_x1C[WriteDis]==1) THEN Read-only. ELSE Read-write. ENDIF. BIOS: 0. 1=MSI interrupts disabled. 0=MSI interrupts enabled. 8 AudioEn. IF (D0F0x64_x1C[WriteDis]==1) THEN Read-only. ELSE Read-write. ENDIF. BIOS: 1. 1=HDMITM audio enabled. 0=HDMI audio disabled. 7 Reserved. 6 RegApSize. IF (D0F0x64_x1C[WriteDis]==1) THEN Read-only. ELSE Read-write. ENDIF. BIOS: 1. Specifies the size of the graphics register aperture. 0=64KB. 1=256KB. 5:3 MemApSize. IF (D0F0x64_x1C[WriteDis]==1) THEN Read-only. ELSE Read-write. ENDIF. Specifies the size of the frame buffer aperture. Bits Definition Bits Definition 000b 128MB 100b 512MB 001b 256MB 101b 1GB 010b 64MB 110b 2GB 011b 32MB 111b 4GB 2 F064BarEn. IF (D0F0x64_x1C[WriteDis]==1) THEN Read-only. ELSE Read-write. ENDIF. 1=64bit base address registers. 0=32-bit base address registers. 1 F0NonlegacyDeviceTypeEn. IF (D0F0x64_x1C[WriteDis]==1) THEN Read-only. ELSE Readwrite. ENDIF. BIOS: 0. 1=PCIe device. 0=Legacy PCI device. 0 WriteDis. IF (D0F0x64_x1C[WriteDis]==1) THEN Read-only. ELSE Read-write. ENDIF. 1=Register is read-only and the GPU PCI interface is configured. 0=Register is read-write. D0F0x64_x1D Internal Graphics PCI Control 2 Reset: 0000_0001h. Bits Description 31:4 Reserved. 3 Vga16En: VGA IO 16 bit decoding enable. Read-write. 1=Address bits 15:10 for VGA IO cycles are decoded. 0=Address bits 15:10 for VGA IO cycles are ignored. 2 Reserved. 1 VgaEn: VGA enable. Read-write. Affects the response by the internal graphics to compatible VGA addresses when IntGfxAsPcieEn=1. 1=The internal graphics decodes the following accesses: • Memory accesses in the range of A0000h to BFFFFh. • IO address where address bits 9:0 are in the ranges of 3B0h to 3BBh or 3C0h to 3DFh. For IO cycles the decoding of address bits 15:10 depends on Vga16En. 0 IntGfxAsPcieEn: internal graphics is RC integrated device. Read-write. BIOS: 1. 1=Integrated graphics is device 1 on bus 0 and operates as a RC integrated device. 0=Integrated graphics is located behind a PCI-to-PCI bridge and is device 5 on the bus behind the bridge. The bridge device is device 1 on bus 0. 173 42300 Rev 3.12 - July 14, 2015 BKDG for AMD Family 15h Models 10h-1Fh Processors D0F0x64_x20 Programmable Device Remap 0 Reset: BA97_6542h. Bits Description 31:20 Reserved. 19:16 GppPortEDevmap. Read-write. Maps any downstream PCIE bridge, except SB, to port Eof PCIE GPPSB core. 15:12 GppPortDDevmap. Read-write. Maps any downstream PCIE bridge, except SB, to port D of PCIE GPPSB core. 11:8 GppPortCDevmap. Read-write. Maps any downstream PCIE bridge, except SB, to port C of PCIE GPPSB core. 7:4 GppPortBDevmap. Read-write. Maps any downstream PCIE bridge, except SB, to port B of PCIE GPPSB core. 3:2 Reserved. 1 IocPcieDevRemapDis. Read-write. 1=Disables remapping. 0 ProgDevMapEn. Read-write. 1=Enable software device remapping. D0F0x64_x21 Programmable Device Remap 1 Reset: 0003_2EDCh. Bits Description 31:20 Reserved. 19:16 GfxPortBDevmap. Read-write. Maps any downstream PCIE bridge, except SB, to port B of PCIE GFX core. 15:12 GfxPortADevmap. Read-write. Maps any downstream PCIE bridge, except SB, to port A of PCIE GFX core. 11:0 Reserved. D0F0x64_x22 LCLK Control 0 Reset: 7F3F_8100h. Bits Description 31 Reserved. 30 SoftOverrideClk0. Read-write. BIOS: 0. 1=Dynamic clock gating disabled for the host request path to the PCIe cores. 29 SoftOverrideClk1. Read-write. BIOS: 0. 1=Dynamic clock gating disabled for the host request path to the internal graphics and the host response path. 28 SoftOverrideClk2. Read-write. BIOS: 0. 1=Dynamic clock gating disabled for the host configuration requests. 27 SoftOverrideClk3. Read-write. BIOS: 0. 1=Dynamic clock gating disabled for the debug bus path. 174 42300 Rev 3.12 - July 14, 2015 26 BKDG for AMD Family 15h Models 10h-1Fh Processors SoftOverrideClk4. Read-write. BIOS: 0. 1=Dynamic clock gating disabled for the host request path to the configuration block. 25:0 Reserved. D0F0x64_x23 LCLK Control 1 Reset: 7F3F_8100h. Bits Description 31 Reserved. 30 SoftOverrideClk0. Read-write. BIOS: IF (D0F2x44[IommuEnable]==1) THEN 1 ELSE 0 ENDIF. 1=Dynamic clock gating disabled for upstream DMA requests from all sources. 29 SoftOverrideClk1. Read-write. BIOS: IF (D0F2x44[IommuEnable]==1) THEN 1 ELSE 0 ENDIF. 1=Dynamic clock gating disabled for upstream DMA requests from the GPPFCH link core. 28 SoftOverrideClk2. Read-write. BIOS: 0. 1=Dynamic clock gating disabled for upstream DMA requests from internal graphics and its DMA response reordering path. 27 SoftOverrideClk3. Read-write. BIOS: IF (D0F2x44[IommuEnable]==1) THEN 1 ELSE 0 ENDIF. 1=Dynamic clock gating disabled for upstream DMA requests from internal graphics. 26 SoftOverrideClk4. Read-write. BIOS: IF (D0F2x44[IommuEnable]==1) THEN 1 ELSE 0 ENDIF. 1=Dynamic clock gating disabled for upstream DMA requests from the Gfx link core. 25:0 Reserved. D0F0x64_x46 IOC Features Control Reset: 0001_3063h. Bits Description 31:17 Reserved. 16 Msi64bitEn: 64 bit MSI enable. Read-write. 1=64-bit MSI support enabled. 0=64-bit MSI support disabled. 15:3 Reserved. 2:1 0 P2PMode: peer-to-peer mode. Read-write. BIOS: 00b. Specifies how upstream write transactions above D0F0x64_x19[Tom2] are completed. Bits Definition 00b Master abort writes that do not hit one of the internal PCI bridges. Forward writes that hit one of the internal PCI bridges to the bridge. 11b-01b Reserved. Reserved. 175 42300 Rev 3.12 - July 14, 2015 BKDG for AMD Family 15h Models 10h-1Fh Processors D0F0x64_x5[B,9,7,5,3,1] IOC PCIe Device Control Reset: 0000_0000h. Table 67: BIOS recommendations for D0F0x64_x5[B,9,7,5,3,1] D0F0x60 [6:0] 51h 53h 55h 57h 59h 5Bh Bits Function BIOS Device 2 Device 3 Device 4 Device 5 Device 6 Device 7 IF (D0F0x64_x0C[Dev2BridgeDis]==0) THEN 0010_0000h ELSE 0000_0000h ENDIF. IF (D0F0x64_x0C[Dev3BridgeDis]==0) THEN 0010_0000h ELSE 0000_0000h ENDIF. IF (D0F0x64_x0C[Dev4BridgeDis]==0) THEN 0010_0000h ELSE 0000_0000h ENDIF. IF (D0F0x64_x0C[Dev5BridgeDis]==0) THEN 0010_0000h ELSE 0000_0000h ENDIF. IF (D0F0x64_x0C[Dev6BridgeDis]==0) THEN 0010_0000h ELSE 0000_0000h ENDIF. IF (D0F0x64_x0C[Dev7BridgeDis]==0) THEN 0010_0000h ELSE 0000_0000h ENDIF. Description 31:21 Reserved. 20 SetPowEn: set slot power enable. Read-write. 1=Enables the set_slot_power message to the FCH. 19:0 Reserved. D0F0x7C IOC Configuration Control Reset: 0000_0000h. Bits Description 31:1 Reserved. 0 ForceIntGFXDisable: internal graphics disable. Read-write. Setting this bit disables bridge device 1 on bus 0 and all devices behind this bridge. D0F0x84 Link Arbitration Reset: 0300_0018h. Bits Description 31:10 Reserved. 9 PmeTurnOff: PME_Turn_Off message trigger. Read-write. 1=Trigger a PME_Turn_Off message to all downstream devices if PmeMode=1. 8 PmeMode: PME message mode. Read-write. 1=PME_Turn_Off message is triggered by writing PmeTurnOff. 0=PME_Turn_Off message is triggered by a message from the FCH. 7:5 Reserved. 4 Ev6Mode: EV6 mode. Read-write. BIOS: 1. 1=The links decode the memory range from 640K to 1M. 3 VgaHole: vga memory hole. Read-write. This bit creates a hole in memory for the VGA memory range. 1=Requests hitting the VGA range are checked against PCI bridge memory ranges instead of being forwarded to system memory. 2:0 Reserved. 176 42300 Rev 3.12 - July 14, 2015 BKDG for AMD Family 15h Models 10h-1Fh Processors D0F0x90 Northbridge Top of Memory Reset: 0000_0000h. Bits Description 31:23 TopOfDram. Read-write. BIOS: MSRC001_001A[TOM[31:23]]. Specifies the address that divides between MMIO and DRAM. From TopOfDram to 4G is MMIO; below TopOfDram is DRAM. See 2.4.3 [Access Type Determination]. 22:0 Reserved. D0F0x94 Northbridge ORB Configuration Offset Reset: 0000_0000h. The index/data pair D0F0x94 and D0F0x98 are used to access D0F0x98_x[FF:00]. To read or write to one of these register, the address is written first into the address register D0F0x94 and then the data are read or written by read or write the data register D0F0x98. Bits Description 31:9 Reserved. 8 OrbIndWrEn: ORB index write enable. Read-write. 1=Writes to D0F0x98 are enabled. 7 Reserved. 6:0 OrbIndAddr: ORB index register address. Read-write. D0F0x98 Northbridge ORB Configuration Data Port See D0F0x94. D0F0x98_x06 ORB Downstream Control 0 Reset: 0000_0000h. Bits Description 31:27 Reserved. 26 UmiNpMemWrEn. Read-write. BIOS: See 2.11.4.1. 1=NP protocol over UMI for memory-mapped writes targeting LPC enabled. This bit may be set to avoid a deadlock condition. 25:0 Reserved. D0F0x98_x07 ORB Upstream Arbitration Control 0 Reset: 0000_0000h. Bits Description 31 SMUCsrIsocEn. Read-write. BIOS: 1. 1=CSR accesses go through ISOC channel. If this bit is set, D0F0x98_x1E[HiPriEn] must also be set. 30 UnadjustThrottlingStpclk. Read-write. BIOS: 1. 1=Enable normal UnitID for STPCLK message. 29:16 Reserved. 177 42300 Rev 3.12 - July 14, 2015 BKDG for AMD Family 15h Models 10h-1Fh Processors 15 DropZeroMaskWrEn. Read-write. BIOS: 1. 1=Drop byte write request that have all bytes masked. 0=Forward byte write request that have all bytes masked. 14 MSIHTIntConversionEn. Read-write. BIOS: 0. 1=MSI to HT interrupt conversion enabled. 13:8 Reserved. 7 IommuIsocPassPWMode. Read-write. BIOS: 1. 1=Always set PassPW for IOMMU upstream isochronous requests. 6 DmaReqRespPassPWMode. Read-write. BIOS: 0. Specifies the RespPassPW bit for non-posted upstream DMA requests. Bit Description 0 Always 1. 1 Value passed from IOC. 5 Reserved. 4 IommuBwOptEn. Read-write. BIOS: 1. 1=Optimize IOMMU L2 byte write by detecting consecutive DW mask and translate the request to DW write. 3:1 0 Reserved. IocBwOptEn. Read-write. BIOS: 1. 1=Enable optimization of byte writes by detecting consecutive DW masks and translating the request to DW writes. D0F0x98_x08 ORB Upstream Arbitration Control 1 Reset: 0008_0808h. BIOS: 0001_0808h. This register specifies the weights of the weighted round-robin arbiter in stage 1 of the upstream arbitration for non-posted reads. Bits Description 31:24 Reserved. 23:16 NpWrrLenC. Read-write. This field defines the maximum number of non-posted read requests from the GPU that are serviced before the arbiter switches to the next client. 15:8 NpWrrLenB. Read-write. This field defines the maximum number of non-posted read requests from IOMMU that are serviced before the arbiter switches to the next client. 7:0 NpWrrLenA. Read-write. This field defines the maximum number of non-posted read requests from IOC that are serviced before the arbiter switches to the next client. D0F0x98_x09 ORB Upstream Arbitration Control 2 Reset: 0000_0808h. This register specifies the weights of the weighted round-robin arbiter in stage 1 of the upstream arbitration for posted writes. Bits Description 31:16 Reserved. 178 42300 Rev 3.12 - July 14, 2015 BKDG for AMD Family 15h Models 10h-1Fh Processors 15:8 PWrrLenB. Read-write. This field defines the maximum number of posted write requests from 7:0 the IOMMUthat are serviced before the arbiter switches to the next client. PWrrLenA. Read-write. This field defines the maximum number of posted write requests from the IOC that are serviced before the arbiter switches to the next client. D0F0x98_x0C ORB Upstream Arbitration Control 5 Reset: 0000_0808h. This register specifies the weights of the weighted round-robin arbiter in stage 2 of the upstream arbitration. Bits Description 31 Reserved. 30 StrictSelWinnerEn. Read-write. BIOS: 1. 1=Select arbitration winner when TX is idle and the FIFO is not full. 0=Select arbitration winner when TX is idle. 29:16 Reserved. 15:8 GcmWrrLenB. Read-write. BIOS: 08h. This field defines the maximum number of posted write requests from stage 1 that are getting serviced in the round-robin before the stage 2 arbiter switches to the next client. 7:0 GcmWrrLenA. Read-write. BIOS: 08h. This field defines the maximum number of non-posted read requests from stage 1 that are getting serviced in the round-robin before the stage 2 arbiter switches to the next client. D0F0x98_x1E ORB Receive Control 0 Reset: 0000_0000h. Bits Description 31:2 Reserved. 1 HiPriEn. Read-write. BIOS: 1. 1=High priority channel enabled. See D0F0x98_x27[IOMMUUrAddr[31:6]]. IF (D0F0x98_x1E[HiPriEn]==0) THEN (D0F0x98_x07[SMUCsrIsocEn]==0) && (IOMMUx18[Isoc]==0). 0 Reserved. D0F0x98_x26 ORB IOMMU Control 0 Reset: 0000_0000h. Bits Description 31:8 Reserved. 7:0 IOMMUUrAddr[39:32]. Read-write. See: D0F0x98_x27[IOMMUUrAddr[31:6]]. 179 42300 Rev 3.12 - July 14, 2015 BKDG for AMD Family 15h Models 10h-1Fh Processors D0F0x98_x27 ORB IOMMU Control 1 Reset: 0000_0000h. Bits Description 31:6 IOMMUUrAddr[31:6]. Read-write. BIOS: IOMMUUrAddr[39:6] must be programmed to a safe system memory address when D0F0x98_x1E[HiPriEn]=1. IOMMUUrAddr[39:6] = {D0F0x98_x26[IOMMUUrAddr[39:32]], IOMMUUrAddr[31:6]}. IOMMU requests that are not directed to system memory are redirected to IOMMUUrAddr. 5:0 Reserved. D0F0x98_x28 ORB Transmit Control 0 Reset: 0000_0000h. Bits Description 31:2 Reserved. 1 ForceCoherentIntr. Read-write. BIOS: 1. 1=Interrupt request are forced to have coherent bit set. 0 Reserved. D0F0x98_x2C ORB Clock Control Reset: 000F_0000h. Bits Description 31:16 WakeHysteresis.Read-write. BIOS: 64h. Specifies the amount of time hardware waits after ORB becomes idle before deasserting the wake signal to the NB. Wait time = WakeHysteresis * 50ns. Values less than 64h may result in undefined behavior. Changes to this field should be done prior to setting DynWakeEn. 15:11 Reserved. 10 NBOutbWakeMask. Read-write. BIOS: 1. 0=OnOutbWake state affects OnInbWake state. 1=OnOutbWake state is masked out. 9 SBDmaActiveMask. Read-write. BIOS: 0. 0=SB_DMA_ACTIVE_L state affects OnInbWake state. 1= SB_DMA_ACTIVE_L state is masked out. 8 OrbRxIdlesMask. Read-write. BIOS: 1. 0=ORB RX idle states affect OnInbWake state. 1=ORB RX idle states are masked out. 7:2 Reserved. 1 DynWakeEn.Read-write. BIOS: 1. 1=Enable dynamic toggling of the wake signal between ORB and NB. 0=Disable dynamic toggling of the wake signal. See WakeHysteresis. 0 Reserved. D0F0x98_x3A ORB Source Tag Translation Control 2 Reset: 0000_0000h. 180 42300 Rev 3.12 - July 14, 2015 BKDG for AMD Family 15h Models 10h-1Fh Processors Clumping allows device 2 port to have up to 64 outstanding non-posted requests when device 3 is not enabled. Bits Description 31:0 ClumpingEn. Read-write. Valid for clumping internal unit IDs 2 and 3 and/or unit IDs 14, 15, 16, and 17. D0F0x98_x4[A,9] ORB LCLK Clock Control 1-0 Reset: 7F3F_8100h. Bits Description 31 Reserved. 30 SoftOverrideClk0. Read-write. BIOS: IF (REG==D0F0x98_x4A) THEN 1 ELSE 0 ENDIF. See: SoftOverrideClk6. 29 SoftOverrideClk1. Read-write. BIOS: 0. See: SoftOverrideClk6. 28 SoftOverrideClk2. Read-write. BIOS: 0. See: SoftOverrideClk6. 27 SoftOverrideClk3. Read-write. BIOS: 0. See: SoftOverrideClk6. 26 SoftOverrideClk4. Read-write. BIOS: 0. See: SoftOverrideClk6. 25 SoftOverrideClk5. Read-write. BIOS: 0. See: SoftOverrideClk6. 24 SoftOverrideClk6. Read-write. BIOS: 0. 1=Clock gating disabled. 0=Clock gating enabled. 23:0 Reserved. D0F0xB8 SMU Index Address The index/data pair D0F0xB8 and D0F0xBC are used to access D0F0xBC_x[FFFFFFFF:00000000]. To read or write to one of these register, the address is written first into the address register D0F0xB8 and then the data are read or written by read or write to the data register D0F0xBC. Bits Description 31:0 NbSmuIndAddr: smu index address. Read-write. Reset: 0. D0F0xBC SMU Index Data See D0F0xB8. D0F0xBC_x1F100 SCLK DPM Control Reset: xxxx_xxxxh. See 2.5.6.1.2 [SCLK DPM]. Bits Description 31:25 Reserved. 24 VoltageChgEn. Read-write 23:16 SclkDpmBootState. Read-write. 15:1 0 Reserved. Read-write. SclkDpmEn. Read-write. 181 42300 Rev 3.12 - July 14, 2015 BKDG for AMD Family 15h Models 10h-1Fh Processors D0F0xBC_x1F2[E0:00:step20] LCLK DPM Control 0 Reset: xxxx_xxxxh. Each register in D0F0xBC_x1F2[E0:00:step20] corresponds to one LCLK DPM state as follows. Table 68: Register Mapping for D0F0xBC_x1F2[E0:00:step20] Register D0F0xBC_x1F200 D0F0xBC_x1F220 D0F0xBC_x1F240 D0F0xBC_x1F260 Bits Function State 0 State 1 State 2 State 3 Register D0F0xBC_x1F280 D0F0xBC_x1F2A0 D0F0xBC_x1F2C0 D0F0xBC_x1F2E0 Function State 4 State 5 State 6 State 7 Description 31:24 LowVoltageReqThreshold. Read-write. 23:16 VID. Read-write. Specifies the VDDNB VID for this DPM state. 15:8 LclkDivider. Read-write. Specifies the LCLK divisor for this DPM state. 7:1 Reserved. 0 StateValid. Read-write. 1=DPM state is valid. 0=DPM state is invalid. D0F0xBC_x1F2[E8:08:step20] LCLK DPM Control 2 Reset: xxxx_xxxxh. Each register in D0F0xBC_x1F2[E8:08:step20] corresponds to one LCLK DPM state as follows. Table 69: Register Mapping for D0F0xBC_x1F2[E8:08:step20] Register D0F0xBC_x1F208 D0F0xBC_x1F228 D0F0xBC_x1F248 D0F0xBC_x1F268 Bits Function State 0 State 1 State 2 State 3 Register D0F0xBC_x1F288 D0F0xBC_x1F2A8 D0F0xBC_x1F2C8 D0F0xBC_x1F2E8 Function State 4 State 5 State 6 State 7 Description 31:16 ResidencyCounter. Read-write. 15:8 HysteresisDown. Read-write. 7:0 HysteresisUp. Read-write. D0F0xBC_x1F2[F0:10:step20] LCLK DPM Activity Thresholds Reset: xxxx_xxxxh. Each register in D0F0xBC_x1F2[F0:10:step20] corresponds to one LCLK DPM state as follows. 182 42300 Rev 3.12 - July 14, 2015 BKDG for AMD Family 15h Models 10h-1Fh Processors Table 70: Register Mapping for D0F0xBC_x1F2[F0:10:step20] Register D0F0xBC_x1F210 D0F0xBC_x1F230 D0F0xBC_x1F250 D0F0xBC_x1F270 Function State 0 State 1 State 2 State 3 Bits Description 31:8 Reserved. 7:0 ActivityThreshold. Read-write. Register D0F0xBC_x1F290 D0F0xBC_x1F2B0 D0F0xBC_x1F2D0 D0F0xBC_x1F2F0 Function State 4 State 5 State 6 State 7 D0F0xBC_x1F300 SMU_LCLK_DPM_CNTL Reset: xxxx_xxxxh. Bits Description 31:25 Reserved. 24 VoltageChgEn. Read-write. 1=Enable voltage change during LCLK DPM state transition. 23:16 LclkDpmBootState. Read-write. 15:9 8 7:1 0 Reserved. LclkDpmType. Read-write. Reserved. LclkDpmEn. Read-write. 1=Enable LCLK DPM. D0F0xBC_x1F308 SMU_LCLK_DPM_THERMAL_THROTTLING_CNTL Reset: xxxx_xxxxh. Bits Description 31:25 Reserved. 24 TtHtcActive. Read-write. 23:19 Reserved. 18:16 15:9 8 7:1 0 LclkTtMode. Read-write. Reserved. TemperatureSel. Read-write. Reserved. LclkThermalThrottlingEn. Read-write. D0F0xBC_x1F30C SMU_LCLK_DPM_THERMAL_THROTTLING_THRESHOLDS Reset: xxxx_xxxxh. 183 42300 Rev 3.12 - July 14, 2015 Bits BKDG for AMD Family 15h Models 10h-1Fh Processors Description 31:16 HighThreshold. Read-write. Specifies the high thermal threshold for LCLK thermal throttling. 15:0 LowThreshold. Read-write. Specifies the low thermal threshold for LCLK thermal throttling. D0F0xBC_x1F380 FIRMWARE_FLAGS Reset: xxxx_xxxxh. Bits Description 31:24 TestCount. Read-write. Test count. 23:1 0 Reserved. Read-write. InterruptsEnabled. Read-write. Bits Definition 0 Firmware has not yet enabled interrupts. BIOS/Driver cannot yet send message interrupts to SMC. 1 Firmware has enabled interrupts. BIOS/Driver can send message interrupts to SMC. D0F0xBC_x1F384 FIRMWARE_VID Reset: xxxx_xxxxh. Bits Description 31:8 Reserved. 7:0 FirmwareVid. Read-write. Current voltage set by firmware voltage controller. D0F0xBC_x1F388 TEMPERATURE_READ_ADDR Reset: xxxx_xxxxh. Bits Description 31:10 Reserved. 9:6 TcenId. Read-write. 5:0 CsrAddr. Read-write. D0F0xBC_x1F39C PCIE_PG_ARGS Read-write. Reset: xxxx_xxxxh. Bits Description 31:24 UpperLaneID. 0=Upper Lane ID. PHY Lanes UpperLaneID:LowerLaneID (inclusive) are affected. 23:16 LowerLaneID. 0=Lower Lane ID. PHY Lanes UpperLaneID:LowerLaneID (inclusive) are affected. 15:5 Reserved. 184 42300 Rev 3.12 - July 14, 2015 4 3 2 SkipCore. Bits 0 1 Definition Do not skip Core Power Gating actions. Skip Core Power Gating actions. SkipPhy. Bits 0 1 Definition Do not skip PHY Lane Power Gating actions. Skip PHY Lane Power Gating actions. Core. Bits 0 1 1 0 BKDG for AMD Family 15h Models 10h-1Fh Processors Definition Do not record in Housekeeping Structure. Will NOT result in PCIe x16 Core Power Up/Down Record in Housekeeping Structure. May result in PCIe x16 Core Power Up/Down Tx. Bits 0 1 Definition Do not take action for PHY Tx Lanes Take action for PHY Tx Lanes Rx. Read-write. Bits 0 1 Definition Do not take action for PHY Rx Lanes Take action for PHY Rx Lanes D0F0xBC_x1F3D8 LOAD_LINE_TRIM_TABLE1 Reset: xxxx_xxxxh. Bits Description 31:24 LoadLineTrim0. Read-write. 23:16 LoadLineTrim1. Read-write. 15:8 LoadLineTrim2. Read-write. 7:0 LoadLineTrim3. Read-write. D0F0xBC_x1F3DC LOAD_LINE_TRIM_TABLE2 Reset: xxxx_xxxxh. Bits Description 31:24 LoadLineTrim4. Read-write. 23:16 LoadLineTrim5. Read-write. 15:8 LoadLineTrim6. Read-write. 7:0 LoadLineTrim7. Read-write. D0F0xBC_x1F3F8 CSR_GNB_1 Reset: xxxx_xxxxh. 185 42300 Rev 3.12 - July 14, 2015 Bits BKDG for AMD Family 15h Models 10h-1Fh Processors Description 31:24 SviTrimValueVddNB. Read-write. BIOS: D0F0xBC_xE0104184[SviLoadLineTrimVddNb]. Loadline trim to use for VDDNB. Used to index into D0F0xBC_x1F3D8 and D0F0xBC_x1F3DC. See D18F5x188[NbLoadLineTrim]. 23:16 SviTrimValueVdd. Read-write. BIOS: D0F0xBC_xE0104184[SviLoadLineTrimVdd]. Loadline trim to use for VDD. Used to index into D0F0xBC_x1F3D8 and D0F0xBC_x1F3DC. See D18F5x12C[CoreLoadLineTrim]. 15:8 SviInitLoadLineVddNB. Read-write. 7:0 SviInitLoadLineVdd. Read-write. D0F0xBC_x1F3FC CSR_GNB_2 Reset: xxxx_xxxxh. Bits Description 31:16 SviVidStep. Read-write. 15:0 SviVidStepBase. Read-write. D0F0xBC_x1F400 CSR_GNB_3 Reset: xxxx_xxxxh. Bits Description 31:24 Reserved. Read-write. 23:16 PstateMax. Read-write. Highest enabled Pstate. This is the lowest power enabled pstate. 15:8 SviLoadLineOffsetVddNB. Read-write. Loadline offset to use for VddNB. Used to index into D0F0xBC_x1F404. See D18F5x188[NbOffsetTrim]. 7:0 SviLoadLineOffsetVdd. Read-write. BIOS: D0F0xBC_xE0104184[SviLoadLineOffsetVdd] Loadline offset to use for Vdd. Used to index into D0F0xBC_x1F404. See D18F5x12C[CoreOffsetTrim]. D0F0xBC_x1F404 LOAD_LINE_OFFSET_TABLE Reset: xxxx_xxxxh. Bits Description 31:24 LoadLineOffset0. Read-write. 23:16 LoadLineOffset1. Read-write. 15:8 LoadLineOffset2. Read-write. 7:0 LoadLineOffset3. Read-write. D0F0xBC_x1F428 PM_CONFIG Reset: xxxx_xxxxh. 186 42300 Rev 3.12 - July 14, 2015 Bits BKDG for AMD Family 15h Models 10h-1Fh Processors Description 31:30 Reserved. Read-write. 29 SviMode. Read-write. Bits Definition 0 SVI1. 1 SVI2. 28 BapmCoeffOverride. Read-write. Bits Definition 0 Calculate filter coefficients. 1 Use SW programmed filter coefficients; ignored if D18F4x15C[BoostLock]==1. 27 NbPstateAllCpusIdle. Read-write. Bits Definition 0 Use low NB P-state voltage when AllCpusIdle. 1 Use high NB P-state voltage when AllCpusIdle. 26:24 PstateAllCpusIdle. Read-write. This field specifies the core P-state to use for IDD calculation when AllCpusIdle. 23:21 Reserved. 20 19:6 IF (Revision == RL-A1) THEN HybridBoostEn. Read-write. This bit controls the Hybrid Boost feature. 2.5.9.1 [Hybrid Boost]. SMU initializes this bit to 1. Also see D0F0xBC_x1F8EC [HybridBoostNotSupported]. Bits Definition 0 Disable Hybrid Boost. 1 Enable Hybrid Boost. ELSE Reserved. ENDIF Reserved. 5 EnableNbDpm. Read-write. Bits Definition 0 Disable Dynamic NB Pstate Management. Should be cleared before sending SMC_MSG_CONFIG_NBDPM message to SMU. 1 Enable Dynamic NB Pstate Management. Should be set before sending SMC_MSG_CONFIG_NBDPM message to SMU. 4 Reserved. 3 EnableLpmx. Read-write. Bits Definition 0 Disable LPMx. Should be cleared before sending SMC_MSG_CONFIG_LPMx message to SMU. 1 Enable LPMx. Should be set before sending SMC_MSG_CONFIG_LPMx message to SMU. 2 EnableTdcLimit. Read-write. 1 EnableBapm. Read-write. 0 EnableVpcAccumulators. Read-write. 187 42300 Rev 3.12 - July 14, 2015 BKDG for AMD Family 15h Models 10h-1Fh Processors D0F0xBC_x1F460 PM_INTERVAL_CNTL_0 Reset: xxxx_xxxxh. Bits Description 31:24 Loadline. Read-write. 23:16 VoltageCntl. Read-write. 15:8 ThermalCntl. Read-write. 7:0 LclkDpm. Read-write. D0F0xBC_x1F468 PM_TIMER_PERIOD Bits Description 31:0 TimerPeriod. Read-write. Reset: X. Specifies the period at which various power management related algorithms are run. Period = TimerPeriod / REFCLK. D0F0xBC_x1F46C PM_TIMERS_1 Bits Description 31:24 LpmxPeriod. Read-write. Reset: 0. 23:16 BapmPeriod. Read-write. Reset: 0. 15:0 VpcPeriod. Read-write. Reset: 0. D0F0xBC_x1F5F8 NB_PSTATE_CONFIG Reset: xxxx_xxxxh. Bits Description 31:24 Reserved. 23 EnableDpmPstatePoll. Read-write 22 EnableNbPsi1. Read-write. Specifies how PSI1_L functions for VDDNB. 0=PSI1_L is deasserted. 1=PSI1_L is asserted whenever the GPU is idle. 21:18 Reserved. 17 SkipDPM0. Read-write. 16 SkipPG. Read-write. 15:8 Hysteresis. Read-write. Specifies the time the GPU must be idle before transitioning to the NB Pstates indexed by Dpm0PgNbPsHi and Dpm0PgNbPsLo. Time = Hysteresis * D0F0xBC_x1F638[NbDpmPeriod] time. 7:6 DpmXNbPsHi. Read-write. BIOS: D0F0xBC_xE010703C[NbPstateHi]. See: Dpm0PgNbPsLo. 5:4 DpmXNbPsLo. Read-write. BIOS: D0F0xBC_xE010703C[NbPstateLo]. See: Dpm0PgNbPsLo. 188 42300 Rev 3.12 - July 14, 2015 BKDG for AMD Family 15h Models 10h-1Fh Processors 3:2 Dpm0PgNbPsHi. Read-write. BIOS: D0F0xBC_xE010703C[NbPstateHi]. See: Dpm0PgNbPsLo. 1:0 Dpm0PgNbPsLo. Read-write. BIOS: D0F0xBC_xE010703C[NbPstateLo]. Indexes the NB Pstate used during specific levels of GPU activity. See 2.5.4.1 [NB P-states]. Bits NB P-state Indexed 00b D18F3x160 (see D18F5x1[6C:60]). 01b D18F3x164 (see D18F5x1[6C:60]). 10b D18F3x168 (see D18F5x1[6C:60]). 11b D18F3x16C (see D18F5x1[6C:60]). D0F0xBC_x1F5FC NB_PSTATE_STATUS Reset: xxxx_xxxxh. Bits Description 31:9 Reserved. 8 7:2 PSI1Sts. Read-write. 1=PSI1_L is enabled for VDDNB. 0=PSI1_L is disabled for VDDNB. Reserved. 1 CurrentPstatePair. Read-write, updated-by-hardware. Specifies the NB P-state pair in use. 1=D0F0xBC_x1F5F8[Dpm0PgNbPsHi, Dpm0PgNbPsLo]. 0=D0F0xBC_x1F5F8[DpmXNbPsHi, DpmXNbPsLo]. 0 ChangeInProgress. Read-write, updated-by-hardware. 1=Hardware is updating D18F5x170[NbPstateHi, NbPstateLo]. D0F0xBC_x1F628 LHTC CONFIG 1 See 2.10.4.2 [Local Hardware Thermal Control (LHTC)]. Bits Description 31:24 Reserved. 23:16 LhtcActivePstateLimit. Read-write. Reset: 0. BIOS: D0F0xBC_xE0104188[LhtcPstateLimit]. Specifies the P-state limit of all cores when in the LHTC-active state. LhtcPstateLimit must be greater than or equal to its reset value (same or lower performing P-state) but not exceed MSRC001_0061[PstateMaxVal]+D18F4x15C[NumBoostStates]. This field uses hardware Pstate numbering, see 2.5.3.1.2.2 [Hardware P-state Numbering]. 15:0 Reserved. D0F0xBC_x1F62C TDC_VRM_LIMIT Bits Description 31:16 Iddnb. Read-write. Reset: 0. 15:0 Idd. Read-write. Reset: 0. D0F0xBC_x1F638 PM_TIMERS_2 Reset: xxxx_xxxxh. 189 42300 Rev 3.12 - July 14, 2015 Bits BKDG for AMD Family 15h Models 10h-1Fh Processors Description 31:24 PgInterlockPeriod. Read-write. 23:16 NbdpmPeriod. Read-write. Specifies the rate at which hardware determines which two NB Pstates should be in use. Time = NbdpmPeriod * D0F0xBC_x1F468[TimerPeriod] time. 15:8 HtcPeriod. Read-write. 7:0 TdcPeriod. Read-write. D0F0xBC_x1F86C LHTC CONFIG 2 Reset: xxxx_xxxxh. Bits Description 31:24 Reserved. 23 22:0 LhtcCap. IF D18F4x15C[BoostLock] THEN Read-only, updated-by-SMU. ELSE Read-write, updated-by-SMU. ENDIF. 1=D0F0xBC_x1F8555[LhtcPstateLimit] is valid. 0=D0F0xBC_x1F8555[LhtcPstateLimit] is not valid. See 2.10.4.2 [Local Hardware Thermal Control (LHTC)]. Reserved. D0F0xBC_x1F89C LHTC CONFIG 3 Reset: xxxx_xxxxh. Bits Description 31:24 LhtcGtempLimitLo. IF D18F4x15C[BoostLock] THEN Read-only, updated-by-SMU. ELSE Read-write, updated-by-SMU. ENDIF. See 2.10.4.2 [Local Hardware Thermal Control (LHTC)]. See HtcGtempLimitHi for encoding. 23:16 LhtcGtempLimitHi. IF D18F4x15C[BoostLock] THEN Read-only, updated-by-SMU. ELSE Read-write, updated-by-SMU. ENDIF. See 2.10.4.2 [Local Hardware Thermal Control (LHTC)]. Range is -49°C to 206°C and is encoded as <Limit-49>°C. Bits Temperature Limit 00h -49°C. FEh-01h <Limit - 49> °C. FFh 206°C. 15:0 Reserved. D0F0xBC_x1F8EC HYBRID_BOOST_CONFIG Reset: xxxx_xxxxh. IF (Revision==RL-A1) THEN Bits Description 31:25 Reserved. 190 42300 Rev 3.12 - July 14, 2015 24 23:0 BKDG for AMD Family 15h Models 10h-1Fh Processors HybridBoostNotSupported. IF D18F4x15C[BoostLock] THEN Read-only, updated-by-SMU. ELSE Read-write, updated-by-SMU. ENDIF. Hybrid Boost cannot be enabled if this bit is set. Reserved. ELSE Bits Description 31:0 Reserved. ENDIF. D0F0xBC_xE0000120 Activity Monitor Control Bits Description 31:11 Reserved. 10 EnOrbDsCnt. Read-write. Reset: 0. 1=Enable downstream counter. 9 EnOrbUsCnt. Read-write. Reset: 0. 1=Enable upstream counter. 8 EnBifCnt. Read-write. Reset: 0. 1=Enable BIF counter. 7:5 Reserved. 4:3 BusyCntSel. Read-write. Reset: 0. Specifies subcomponents or activity monitored by the LCLK activity monitor. Bits Definition Bits Definition 00b GFX DMA (BIF) 10b Downstream activity 01b Upstream activity 11b Up/downstream activity max 2 Reserved. 1 PeriodCntRst. Read-write. Reset: 1. 0 ActivityCntRst. Read-write. Reset: 1. D0F0xBC_xE0003000 CPU Interrupt Request See 2.13.1 [Software Interrupts]. Bits Description 31:17 Reserved. 16:1 0 ServiceIndex. Read-write. Reset: 0. IntToggle. Read-write. Reset: 0. D0F0xBC_xE0003004 CPU Interrupt Status See 2.13.1 [Software Interrupts]. Bits Description 31:2 Reserved. 1 IntDone. Read-only; updated-by-hardware. Reset: 0. 0 IntAck. Read-only; updated-by-hardware. Reset: 0. 191 42300 Rev 3.12 - July 14, 2015 BKDG for AMD Family 15h Models 10h-1Fh Processors D0F0xBC_xE0003048 SCLK_MIN_DIV Bits Description 31:19 Reserved. 18:12 Intv. Read-write. Reset: 02h. Divider integer value 11:0 Fracv. Read-write. Reset: 0. Divider fraction value D0F0xBC_xE0003088 SMU_AUTH_STATUS Bits Description 31:2 Reserved. 1 SmuAuthPass. Read-write. Reset: 0. 0 SmuAuthDone. Read-write. Reset: 0. D0F0xBC_xE00030A4 SMU_FIRMWARE_AUTH Bits Description 31:17 Reserved. 16 15:0 SmuProtectedMode. Read-only. Reset: 0. Reserved. D0F0xBC_xE0104168 MEMCLK VID Configuration Bits Description 31:30 MemClkVid3[1:0]. Read-only. Reset: value varies by product. See MemClkVid0 29:22 MemClkVid2. Read-only. Reset: value varies by product. See MemClkVid0. 21:14 MemClkVid1. Read-only. Reset: value varies by product. See MemClkVid0. 13:6 MemClkVid0. Read-only. Reset: value varies by product. Specifies the VDDNB voltage required for a specific MEMCLK. See the AMD Serial VID Interface 2.0 (SVI2) Specification. Each MemClkVid field corresponds to a MEMCLK frequency as follows: • MemClkVid8: 1200MHz. • MemClkVid7: 1066MHz. • MemClkVid6: 1050MHz. • MemClkVid5: 933MHz. • MemClkVid4: 800MHz. • MemClkVid3: 667MHz. • MemClkVid2: 533MHz. • MemClkVid1: 400MHz. • MemClkVid0: 333 MHz. 5:0 Reserved. 192 42300 Rev 3.12 - July 14, 2015 BKDG for AMD Family 15h Models 10h-1Fh Processors D0F0xBC_xE010416C MEMCLK VID Configuration Bits Description 31:30 MemClkVid7[1:0]. Read-only. Reset: value varies by product. See D0F0xBC_xE0104168[MemClkVid0]. 29:22 MemClkVid6. Read-only. Reset: value varies by product. See D0F0xBC_xE0104168[MemClkVid0]. 21:14 MemClkVid5. Read-only. Reset: value varies by product. See D0F0xBC_xE0104168[MemClkVid0]. 13:6 MemClkVid4. Read-only. Reset: value varies by product. See D0F0xBC_xE0104168[MemClkVid0]. 5:0 MemClkVid3[7:2]. Read-only. Reset: value varies by product. See D0F0xBC_xE0104168[MemClkVid0]. D0F0xBC_xE0104170 MEMCLK VID Configuration Bits Description 31:14 Reserved. 13:6 MemClkVid8. Read-only. Reset: value varies by product. See D0F0xBC_xE0104168[MemClkVid0]. 5:0 MemClkVid7[7:2]. Read-only. Reset: value varies by product. See D0F0xBC_xE0104168[MemClkVid0]. D0F0xBC_xE0104184 SVI Loadline Configuration Bits Description 31:10 Reserved. 9:8 SviLoadLineOffsetVddNb. Read-only. Reset: value varies by product. See D18F5x188[NbOffsetTrim]. 7:6 SviLoadLineOffsetVdd. Read-only. Reset: value varies by product. See D18F5x12C[CoreOffsetTrim]. 5:3 SviLoadLineTrimVddNb. Read-only. Reset: value varies by product. See D18F5x188[NbLoadLineTrim]. 2:0 SviLoadLineTrimVdd. Read-only. Reset: value varies by product. See D18F5x12C[CoreLoadLineTrim]. D0F0xBC_xE0104188 SMU Power Management Config 1 Bits Description 31 Reserved. 30 BapmDisable. Read-only. 29 BapmMeasuredTemp. Read-only. 193 42300 Rev 3.12 - July 14, 2015 BKDG for AMD Family 15h Models 10h-1Fh Processors 28:26 LhtcPstateLimit. Read-only. Reset: value varies by product. See D0F0xBC_x1F628[LhtcPstateLimit] and 2.10.4.2 [Local Hardware Thermal Control (LHTC)]. 25:0 Reserved. D0F0xBC_xE010418C SMU Power Management Config 2 Bits Description 31:0 Reserved. D0F0xBC_xE010703C SMU Power Management Config 3 Bits Description 31:7 Reserved. 6:5 NbPstateLo. Read-only. Reset: value varies by product. See D0F0xBC_x1F5F8[Dpm0PgNbPsLo]. 4:3 NbPstateHi. Read-only. Reset: value varies by product. See D0F0xBC_x1F5F8[Dpm0PgNbPsHi]. 2:0 Reserved. D0F0xBC_xE0300000 SBLK_MC_PGFSM_CONFIG Read-write. Reset: 0. Bits Description 31:28 RegAddr. 27:14 Reserved. 13 ReadOp. 12 WriteOp. 11 P2Select. 10 P1Select. 9 PowerUp. 8 PowerDown. 7:0 FsmAddr. D0F0xBC_xE030000C SBLK_PGD_PGFSM_CONFIG Bits Description 31:28 RegAddr. Read-write. Reset: 0. 27:14 Reserved. 13 ReadOp. Read-write. Reset: 0. 12 WriteOp. Read-write. Reset: 0. 194 42300 Rev 3.12 - July 14, 2015 11 P2Select. Read-write. Reset: 0. 10 P1Select. Read-write. Reset: 0. 9 PowerUp. Read-write. Reset: 0. 8 PowerDown. Read-write. Reset: 0. 7:0 BKDG for AMD Family 15h Models 10h-1Fh Processors FsmAddr. Read-write. Reset: 0. D0F0xBC_xE0300018 SBLK_IOMMU_PGFSM_CONFIG Bits Description 31:28 RegAddr. Read-write. Reset: 0. 27:14 Reserved. 13 ReadOp. Read-write. Reset: 0. 12 WriteOp. Read-write. Reset: 0. 11 P2Select. Read-write. Reset: 0. 10 P1Select. Read-write. Reset: 0. 9 PowerUp. Read-write. Reset: 0. 8 PowerDown. Read-write. Reset: 0. 7:0 FsmAddr. Read-write. Reset: 0. D0F0xBC_xE0300024 SBLK_VCE_PGFSM_CONFIG Bits Description 31:28 RegAddr. Read-write. Reset: 0. 27:14 Reserved. 13 ReadOp. Read-write. Reset: 0. 12 WriteOp. Read-write. Reset: 0. 11 P2Select. Read-write. Reset: 0. 10 P1Select. Read-write. Reset: 0. 9 PowerUp. Read-write. Reset: 0. 8 PowerDown. Read-write. Reset: 0. 7:0 FsmAddr. Read-write. Reset: 0. D0F0xBC_xE0300030 SBLK_DC2_PGFSM_CONFIG Bits Description 31:28 RegAddr. Read-write. Reset: 0. 27:14 Reserved. 13 ReadOp. Read-write. Reset: 0. 12 WriteOp. Read-write. Reset: 0. 195 42300 Rev 3.12 - July 14, 2015 11 P2Select. Read-write. Reset: 0. 10 P1Select. Read-write. Reset: 0. 9 PowerUp. Read-write. Reset: 0. 8 PowerDown. Read-write. Reset: 0. 7:0 BKDG for AMD Family 15h Models 10h-1Fh Processors FsmAddr. Read-write. Reset: 0. D0F0xBC_xE030003C UVD_CHAIN_PGFSM_CONFIG Bits Description 31:28 RegAddr. Read-write. Reset: 0. 27:14 Reserved. 13 ReadOp. Read-write. Reset: 0. 12 WriteOp. Read-write. Reset: 0. 11 P2Select. Read-write. Reset: 0. 10 P1Select. Read-write. Reset: 0. 9 PowerUp. Read-write. Reset: 0. 8 PowerDown. Read-write. Reset: 0. 7:0 FsmAddr. Read-write. Reset: 0. D0F0xBC_xE0300200 SBLK_MC_PGFSM_DEBUG Bits Description 31:11 Reserved. 10 P1IsoN. Read-only. Reset: 1. P1 isolation 9:0 Reserved. D0F0xBC_xE0300208 SBLK_IOMMU_PGFSM_DEBUG Bits Description 31:11 Reserved. 10 P1IsoN. Read-only. Reset: 1. P1 isolation 9:0 Reserved. D0F0xBC_xE030020C SBLK_VCE_PGFSM_DEBUG Bits Description 31:11 Reserved. 10 P1IsoN. Read-only. Reset: 1. P1 isolation 9:0 Reserved. 196 42300 Rev 3.12 - July 14, 2015 BKDG for AMD Family 15h Models 10h-1Fh Processors D0F0xBC_xE0300210 SBLK_DC2_PGFSM_DEBUG Bits Description 31:11 Reserved. 10 P1IsoN. Read-only. Reset: 1. P1 isolation 9:0 Reserved. D0F0xBC_xE0300218 SBLK_UVDU_PGFSM_DEBUG Bits Description 31:11 Reserved. 10 P1IsoN. Read-only. Reset: 1. P1 isolation 9:0 Reserved. D0F0xBC_xE0300320 PGFSM_CLK_CNTL_0 Bits Description 31:2 Reserved. 1 IommuPgfsmClockEn. Read-write. Reset: 1. 0 PgdPgfsmClockEn. Read-write. Reset: 1. D0F0xBC_xE0300324 PGFSM_CLK_CNTL_1 Bits Description 31:3 Reserved. 2 Dc2PgfsmClockEn. Read-write. Reset: 1. 1 UvdPgfsmClockEn. Read-write. Reset: 1. 0 VcePgfsmClockEn. Read-write. Reset: 1. D0F0xE0 Link Index Address Reset: 0130_8001h. D0F0xE0 and D0F0xE4 are used to access D0F0xE4_x[FFFF_FFFF:0000_0000]. To read or write to one of these register, the address is written first into the address register D0F0xE0 and then the data is read from or written to the data register D0F0xE4. The phy index registers (D0F0xE4_x[2xxx_xxxx]) mapping to a specific phy, pin or pin group is shown in a table in the register definition. For example, to perform a read or write operation to configure Gfx phy 0 (P_GFX_[T,R]X[P,N][7:0] pin group) compensation, software should program D0F0xE0[31:0]=0121_0000h. Accessing any register number that is not listed in the mapping table may result in undefined behavior. Some phy registers support broadcast write operations to groups of 4 or 8 lanes. For example, to perform 197 42300 Rev 3.12 - July 14, 2015 BKDG for AMD Family 15h Models 10h-1Fh Processors broadcast write operation to configure Gfx Link[3:0] (P_GFX_RX[P,N][3:0] lanes) receiver phase loop filter, software should program D0F0xE0[31:0]=0221_5602h. Bits Description 31:24 BlockSelect: block select. Read-write. This field is used to select the specific register block to access. The encodings supported depends on the FrameType selected. FrameType Encoding 01h 1=GPP link core, 2=Gfx link core 1xh 1=Phy interface 0, 2=Phy interface 1 (FrameType 11h only) 2xh 1=Phy 0, 2=Phy 1 (FrameType 21h only) 3xh 1=Wrapper 23:16 FrameType: frame type. Read-write. This field is used to select the type of register block to access. Bits Destination 01h Link core registers 1Nh Phy interface block registers. 2Nh Phy registers. 3Nh Wrapper registers. N 0h 1h 2h 3h Register Block GPP PCIe Links Gfx PCIe links DDI DDI2 15:0 PcieIndxAddr: index address. Read-write. D0F0xE4 Link Index Data See D0F0xE0. Bits Description 31:0 PcieIndxData: index data. Read-write. 3.3.1 IO Link Registers D0F0xE4_x0[2:1]01_0002 IO Link Hardware Debug Reset: 0000_0000h. Bits Description 31:1 Reserved. 0 HwDebug[0]: ignore DLLPs in L1. Read-write. BIOS: 1. 1=DLLPs are ignored in L1 so the TXCLK can be turned off. D0F0xE4_x0[2:1]01_0010 IO Link Control 1 Reset: 80E3_0800h. Bits Description 31:13 Reserved. 198 42300 Rev 3.12 - July 14, 2015 BKDG for AMD Family 15h Models 10h-1Fh Processors 12:10 RxSbAdjPayloadSize. Read-write. BIOS: 100b. Payload size for DMA requests from the PCIe controller buffer. Bits Definition Bits Definition 00xb Reserved. 100b 64 bytes. 010b 16 bytes. 101b Reserved. 011b 32 bytes. 11xb Reserved. 9 8:1 0 UmiNpMemWrite: memory write mapping enable. Read-write. 1=Internal non-posted memory writes are transferred to UMI. Reserved. HwInitWrLock: hardware init write lock. Read-write. 1=Lock HWInit registers. 0=Unlock HWInit registers. D0F0xE4_x0[2:1]01_0011 IO Link Config Control Reset: 0000_0007h. Bits Description 31:30 Reserved. 29:27 CiPrivMaxCplPayloadSize. Read-write. Private value for MAX_PAYLOAD_SIZE, used when CiMaxCplPayloadSizeMode=1. 26 CiMaxCplPayloadSizeMode. Read-write. Used for Slave completions. Private override for cfg register DEVICE_CNTL:MAX_PAYLOAD_SIZE. 1=Use CiPrivMaxCplPayloadSize. 25 CiExtendedTagEnOverride. Read-write. 1=Private override for DEVICE_CNTL: EXTENDED_TAG_EN. 24 CiMaxReadSafeMode. Read-write. 1=Only 32 or 64 byte address aligned reads are generated. 23:21 CiPrivMaxReadRequestSize. Read-write. Specifies the maximum read request size when CiMaxReadRequestSizeMode=1. Bits Definition Bits Definition 000b 128 bytes. 100b 2 Kbytes. 001b 256 bytes. 101b 4 Kbytes. 010b 512 bytes. 110b Reserved. 011b 1 Kbyte. 111b Reserved. 20 CiMaxReadRequestSizeMode. Read-write. 1=CiPrivMaxReadRequestSize defines the maximum read request size. 0=D[8:2]F0x60[MaxRequestSize] defines the maximum read request size. 19:17 CiPrivMaxPayloadSize. Read-write. Specifies the maximum payload size when CiMaxPayloadSizeMode=1. Bits Definition Bits Definition 000b 128 bytes. 100b 2 Kbytes. 001b 256 bytes. 101b 4 Kbytes. 010b 512 bytes. 110b Reserved. 011b 1 Kbyte. 111b Reserved. 16 CiMaxPayloadSizeMode. Read-write. 1=CiPrivMaxPayloadSize defines the maximum payload size. 0=D[8:2]F0x60[MaxPayloadSize] defines the maximum payload size. 15:4 Reserved. 3:0 DynClkLatency: dynamic clock latency. Read-write. BIOS: See 2.11.4.2.2. Specifies the number of clock cycles after logic goes idle before clocks are gated off. 199 42300 Rev 3.12 - July 14, 2015 BKDG for AMD Family 15h Models 10h-1Fh Processors D0F0xE4_x0[2:1]01_001C IO Link Control 2 Reset: 0E00_0000h. Bits Description 31:11 Reserved. 10:6 TxArbMstLimit: transmitter arbitration master limit. Read-write. BIOS: 4h. Defines together with TxArbSlvLimit a round robin arbitration pattern for downstream accesses. TxArbMstLimit defines the weight for downstream CPU requests and TxArbSlvLimit for the downstream read responses. 5:1 0 TxArbSlvLimit: transmitter arbitration slave limit. Read-write. BIOS: 4h. See TxArbMstLimit for details TxArbRoundRobinEn: transmitter round robin arbitration enabled. Read-write. BIOS: 1. 1=Enable transmitter round robin arbitration. 0=Disable transmitter round robin arbitration. D0F0xE4_x0[2:1]01_0020 IO Link Chip Interface Control Reset: 0000_0050h. Bits Description 31:10 Reserved. 9 CiRcOrderingDis: chip interface RC ordering disable. Read-write. 0=RC ordering logic is enabled. 1=RC ordering logic is disabled. 8 CiSlvOrderingDis: slave interface ordering disable. Read-write. 0=Slave ordering logic is enabled. 1=Slave ordering logic is disabled. 7:0 Reserved. D0F0xE4_x0[2:1]01_0040 IO Link Phy Control Reset: 0000_0000h. Bits Description 31:16 Reserved. 15:14 PElecIdleMode: electrical idle mode for physical layer. Read-write. BIOS: 10b. Defines which electrical idle signal is used, either inferred by link controller of from phy. Bits Definition 00b Gen1 - entry:phy, exit:phy; Gen2 - entry:infer, exit:phy 01b Gen1 - entry:infer, exit:phy; Gen2 - entry:infer, exit:phy 10b Gen1 - entry:phy, exit:phy; Gen2 - entry:PHY, exit:phy 11b Reserved 13:0 Reserved. 200 42300 Rev 3.12 - July 14, 2015 BKDG for AMD Family 15h Models 10h-1Fh Processors D0F0xE4_x0[2:1]01_00B0 IO Link Strap Link Strap Control Reset: 0000_0001h. Bits Description 31:6 Reserved. 5 4:3 2 1:0 StrapF0AerEn. Read-write. BIOS: 0. 1=AER support enabled. 0=AER support disabled. Reserved. StrapF0MsiEn. Read-write. BIOS: 1. 1=MSI enabled. 0=MSI disabled. Reserved. D0F0xE4_x0[2:1]01_00C0 IO Link Strap Miscellaneous Bits Description 31 Reserved. 30 StrapFlrEn. Read-write. Reset:0. 29 StrapMstAdr64En. Read-write. Reset: 0. 28 StrapReverseAll. Read-write. Reset: 0. Reverse All strap override. 27:0 Reserved. D0F0xE4_x0[2:1]01_00C1 IO Link Strap Miscellaneous Bits Description 31:2 Reserved. 1 StrapGen2Compliance. Read-write. Reset: 1. 0 StrapLinkBwNotificationCapEn. Read-write. Reset: 0. 3.3.2 PIF Registers D0F0xE4_x0[2:1]1[3:0]_0010 PIF Control Reset: 018A_0059h. Table 71: Index addresses for D0F0xE4_x0[2:1]1[3:0]_0010 D0F0xE0[31:16] D0F0xE0[15:0] 0110h GPPFCH PIF 0111h Gfx PIF 0 0211h Gfx PIF 1 0112h DDI PIF 0113h DDI2 PIF 0010h 201 42300 Rev 3.12 - July 14, 2015 Bits BKDG for AMD Family 15h Models 10h-1Fh Processors Description 31:20 Reserved. 19:17 Ls2ExitTime: LS2 exit time. Read-write. Definition Bits 000b 10us 001b 5us 010b 2.5us 011b 1.25us Bits 100b 101b 110b 111b Definition 625ns 0s Reserved Reserved 16:8 Reserved. 7 RxDetectTxPwrMode: receiver detection transmitter power mode. Read-write. 1=Transmitter is powered on. 6 RxDetectFifoResetMode: receiver detect FIFO reset mode. Read-write. BIOS: 1. 1=The transmit FIFO is reset after receiver detection. 0=The transmit FIFO is not reset after receiver detection. 5 Reserved. 4 EiDetCycleMode: electrical idle detect mode. Read-write. 1=Electrical idle cycle detection mode is enabled in L1. 0=Electrical idle detection is always enabled in L1. 3:0 Reserved. D0F0xE4_x0[2:1]1[3:0]_0011 PIF Pairing Reset: 0200_0000h. Table 72: Index addresses for D0F0xE4_x0[2:1]1[3:0]_0011 D0F0xE0[31:16] D0F0xE0[15:0] 0011h Bits 0110h GPPFCH PIF 0111h Gfx PIF 0 0211h Gfx PIF 1 0112h DDI PIF 0113h DDI2 PIF Description 31:26 Reserved. 25 MultiPif: x16 link. Read-write. 1=Lanes 7:0 are paired with a second PIF to create a x16 link. 24:17 Reserved. 16 X8Lane70: x8 link lanes 7:0. Read-write. 1=Lanes 7:0 are paired to create a x8 link. 15:13 Reserved. 12 X4Lane52: x4 link lanes 5:2. Read-write. 1=Lanes 5:2 are paired to create a x4 link. 11:10 Reserved. 9 X4Lane74: x4 link lanes 7:4. Read-write. 1=Lanes 7:4 are paired to create a x4 link. 8 X4Lane30: x4 link lanes 3:0. Read-write. 1=Lanes 3:0 are paired to create a x4 link. 7:4 Reserved. 202 42300 Rev 3.12 - July 14, 2015 BKDG for AMD Family 15h Models 10h-1Fh Processors 3 X2Lane76: x2 link lanes 7:6. Read-write. 1=Lanes 7:6 are paired to create a x2 link. 2 X2Lane54: x2 link lanes 5:4. Read-write. 1=Lanes 5:4 are paired to create a x2 link. 1 X2Lane32: x2 link lanes 3:2. Read-write. 1=Lanes 3:2 are paired to create a x2 link. 0 X2Lane10: x2 link lanes 1:0. Read-write. 1=Lanes 1:0 are paired to create a x2 link. D0F0xE4_x0[2:1]1[3:0]_001[3:2] PIF Power Down Control [1:0] Reset: 0001_0022h. Table 73: Index addresses for D0F0xE4_x0[2:1]1[3:0]_001[3:2] D0F0xE0[31:16] Bits D0F0xE0[15:0] 0013h 0012h 0110h GPPFCH PIF Lanes 7-4 GPPFCH PIF Lanes 3-0 0111h Gfx PIF 0 Lanes 7-4 Gfx PIF 0 Lanes 3-0 0211h Gfx PIF 1 Lanes 7-4 Gfx PIF 1 Lanes 3-0 0112h DDI PIF Lanes 7-4 DDI PIF Lanes 3-0 0113h DDI2 PIF Lanes 6-4 DDI2 PIF Lanes 3-0 Description 31:29 PllPwrOverrideVal: PLL power state override value. Read-write. See TxPowerStateInTxs2. 28 PllPwrOverrideEn: PLL power state override enable. Read-write. 1=PLL forced to the power state specified by PllPwrOverrideVal. 27 Reserved. 26:24 PllRampUpTime: PLL ramp time. Read-write. Bits Definition 000b 5us 001b 10us 010b 300us 011b 500us Bits 100b 101b 110b 111b Definition 0s 0s Reserved Reserved 23:17 Reserved. 16 Tx2p5clkClockGatingEn. Read-write. 1=The 2.5x TxClk is gated if the lane is idle 0=The 2.5x TxClk is never gated. 15:13 Reserved. 12:10 PllPowerStateInOff: PLL off power state. Read-write. See: TxPowerStateInTxs2. All links associated with the PLL must be in the off state to transition the PLL to this state. 9:7 PllPowerStateInTxs2: PLL L1 power state. Read-write. See: TxPowerStateInTxs2. All links associated with the PLL must be in L1 to transition the PLL to this state. 6:4 RxPowerStateInRxs2: receiver L1 power state. Read-write. See: TxPowerStateInTxs2. 203 42300 Rev 3.12 - July 14, 2015 3 2:0 BKDG for AMD Family 15h Models 10h-1Fh Processors ForceRxEnInL0s: force receiver enable in L0s. Read-write. 1=The phy CDR is always enabled in L0s. TxPowerStateInTxs2: transmitter L1 power state. Read-write. Definition Bits Bits 000b L0 100b 001b LS1 101b 010b LS2 110b 011b Reserved 111b Definition Reserved Reserved Reserved Off D0F0xE4_x0[2:1]1[3:0]_0015 PIF Transmitter Status Reset: 0000_0000h. Table 74: Index addresses for D0F0xE4_x0[2:1]1[3:0]_0015 D0F0xE0[31:16] D0F0xE0[15:0] 0015h Bits 0110h GPPFCH PIF 0111h Gfx PIF 0 0211h Gfx PIF 1 0112h DDI PIF 0113h DDI2 PIF Description 31:8 Reserved. 7 TxPhyStatus07: lane 7 TxPhyStatus. Read-only. See: TxPhyStatus00. 6 TxPhyStatus06: lane 7 TxPhyStatus. Read-only. See: TxPhyStatus00. 5 TxPhyStatus05: lane 5 TxPhyStatus. Read-only. See: TxPhyStatus00. 4 TxPhyStatus04: lane 4 TxPhyStatus. Read-only. See: TxPhyStatus00. 3 TxPhyStatus03: lane 3 TxPhyStatus. Read-only. See: TxPhyStatus00. 2 TxPhyStatus02: lane 2 TxPhyStatus. Read-only. See: TxPhyStatus00. 1 TxPhyStatus01: lane 1 TxPhyStatus. Read-only. See: TxPhyStatus00. 0 TxPhyStatus00: lane 0 TxPhyStatus. Read-only. Returns the state of the TxPhyStatus signal From the PIF to the BIF. 3.3.3 Phy Registers There are three categories of phy registers: 3.3.3.1 [Global Phy Control Registers], receiver lane control registers and transmitter lane control registers. 3.3.3.1 Global Phy Control Registers Each global phy control register may have one instance per phy or two instances per phy (one per nibble). When a global register is implemented per phy the mapping to signal pins is shown in Table 75. When a global register is implemented per nibble the mapping to pins is shown in Table 75. 204 42300 Rev 3.12 - July 14, 2015 BKDG for AMD Family 15h Models 10h-1Fh Processors Table 75: Per phy register addresses to pin mappings D0F0xE0[31:0] Pin Names 0120h_[2:0]xxxh GPP Links: P_GPP_[T,R]X[P,N][3:0] & FCH ports: P_UMI_[T,R]X[P,N][3:0] 0121h_[2:0]xxxh Gfx Links[7:0]: P_GFX_[T,R]X[P,N][7:0] 0221h_[2:0]xxxh Gfx Links[15:8]: P_GFX_[T,R]X[P,N][15:8] 0122h_[2:0]xxxh DDI 1: DP1_TX[P,N][3:0] & DDI 0: DP0_TX[P,N][3:0] 0123h_[2:0]xxxh DDI 2: DP2_TX[P,N][6:0] Table 76: Per nibble register addresses to pin mappings D0F0xE0[31:12] Pin Names Address N+1 Address N 0120h_[2:0]xxxh GPP ports: P_GPP_[T,R]X[P,N][3:0] FCH ports: P_UMI_[T,R]X[P,N][3:0] 0121h_[2:0]xxxh Graphics port lower: P_GFX_[T,R]X[P,N][7:4] Graphics port lower: P_GFX_[T,R]X[P,N][3:0] 0221h_[2:0]xxxh Graphics port upper: P_GFX_[T,R]X[P,N][15:12] Graphics port upper: P_GFX_[T,R]X[P,N][11:8] 0122h_[2:0]xxxh DDI 0: DP0_TX[P,N][3:0] DDI 1: DP1_TX[P,N][3:0] 0123h_[2:0]xxxh DDI 2:DP2_TX[P,N][6:4] DDI 2:DP2_TX[P,N][3:0] D0F0xE4_x0[2:1]2[3:0]_0000 Phy Compensation Control and Calibration Control I This register provides general control of various circuits that perform auto-calibration. Table 77: Index Mapping for D0F0xE4_x0[2:1]2[3:0]_0000 D0F0xE0[31:16] D0F0xE0[15:0] 0000h Bits 0120h GPP Links: P_GPP_[T,R]X[P,N][3:0] & FCH ports: P_UMI_[T,R]X[P,N][3:0] 0121h Gfx Links[7:0]: P_GFX_[T,R]X[P,N][7:0] 0221h Gfx Links[15:8]: P_GFX_[T,R]X[P,N][15:8] 0122h DDI 1: DP1_TX[P,N][3:0] & DDI 0: DP0_TX[P,N][3:0] 0123h DDI 2: DP2_TX[P,N][6:0] Description 31:28 Reserved. 27:23 RttRawCal: receiver termination resistance (Rtt) raw calibration value. Read-only; Updated-byhardware. Reset: 0. This field provides the raw Rtt calibration value as determined by the compensation circuit. 22:18 RonRawCal: transmitter resistance (Ron) raw calibration value. Read-only; Updated-by-hardware. Reset: 0. This field provides the raw Ron calibration value as determined by the compensation circuit. 17:0 Reserved. D0F0xE4_x0[2:1]2[3:0]_000[2:1] Phy Impedance Control Updates to these registers that result in a change to impedance may not take effect in the phy for up to 2 micro- 205 42300 Rev 3.12 - July 14, 2015 BKDG for AMD Family 15h Models 10h-1Fh Processors seconds after the update to this register completes. Table 78: Index Mapping for D0F0xE4_x0[2:1]2[3:0]_000[2:1] D0F0xE0[31:16] Bits D0F0xE0[15:0] 0002h 0001h 0120h GPP ports: P_GPP_[T,R]X[P,N][3:0] FCH ports: P_UMI_[T,R]X[P,N][3:0] 0121h Graphics port lower: P_GFX_[T,R]X[P,N][7:4] Graphics port lower: P_GFX_[T,R]X[P,N][3:0] 0221h Graphics port upper: P_GFX_[T,R]X[P,N][15:12] Graphics port upper: P_GFX_[T,R]X[P,N][11:8] 0122h DDI 0: DP0_TX[P,N][3:0] DDI 1: DP1_TX[P,N][3:0] 0123h DDI 2: DP2_TX[P,N][6:4] DDI 2: DP2_TX[P,N][3:0] Description 31:29 RttCtl: receiver termination resistance (Rtt) control. Read-write. Reset: 0. This field specifies how the receiver termination resistance value is calculated. All values between 00h and 1Fh are valid. Bits Definition 000b Rtt is as determined by the compensation circuit, D0F0xE4_x0[2:1]2[3:0]_0000[RttRawCal]. 001b Rtt is as specified by (RttIndex - 3). 010b Rtt is as specified by the difference: RttRawCal - RttIndex. If this results in a value that is less than 00h, then 00h is used. 011b Rtt is as specified by the sum: RttRawCal + RttIndex. If this results in a value that is greater than 1Fh, then 1Fh is used. 100b Enable only one tap of the Rtt resistor, as specified by RttIndex, and disable the base resistor that is normally always enabled. This is intended for testing purposes only. 111b-101b Reserved. For all modes (except 100b), higher values reduce the resistance of Rtt and lower values increase the resistance of Rtt. If RttCtl is programmed to either 011b or 100b, the value of RttRawCal + RttIndex must be less than or equal to 24. 28:21 Reserved 20:16 RttIndex: receiver termination resistance (Rtt) index. Read-write. Reset: 0. See RttCtl. 206 42300 Rev 3.12 - July 14, 2015 BKDG for AMD Family 15h Models 10h-1Fh Processors 15:13 RonCtl: transmitter resistance (Ron) control. Read-write. Reset: 0. This field specifies how the transmitter resistance value is calculated. Definition Bits 000b Ron is as determined by the compensation circuit, D0F0xE4_x0[2:1]2[3:0]_0000[RonRawCal]. 001b Ron is as specified by the RonIndex field. 010b Ron is as specified by the difference: RonRawCal - RonIndex. If this results in a value that is less than 00h, then 00h is used. 011b Ron is as specified by the sum: RonRawCal + RonIndex. If this results in a value that is greater than 1Fh, then 1Fh is used. 100b Enable only one tap of the Ron resistor, as specified by RonIndex, and disable the base resistor that is normally always enabled. This is intended for testing purposes only. 111b-101b Reserved. For all modes (except 100b), higher values reduce the resistance of Ron and lower values increase the resistance of Ron. If RonCtl is programmed to either 011b or 100b, the value of RonRawCal + RonIndex must be less than or equal to 23. 12:5 Reserved. 4:0 RonIndex: transmitter resistance (Ron) index. Read-write.Reset: 0. See RonCtl. D0F0xE4_x0[2:1]2[3:0]_000[A:9] Phy Clock Tree Control BIOS: See 2.11.4.2.2 [Link Configuration and Core Initialization]. Table 79: Index Mapping for D0F0xE4_x0[2:1]2[3:0]_000[A:9] D0F0xE0[31:16] Bits 31 D0F0xE0[15:0] 000Ah 0009h 0120h GPP ports: P_GPP_[T,R]X[P,N][3:0] FCH ports: P_UMI_[T,R]X[P,N][3:0] 0121h Graphics port upper: P_GFX_[T,R]X[P,N][15:12] Graphics port upper: P_GFX_[T,R]X[P,N][11:8] 0221h Graphics port lower: P_GFX_[T,R]X[P,N][7:4] Graphics port lower: P_GFX_[T,R]X[P,N][3:0] 0122h DDI 0: DP0_TX[P,N][3:0] DDI 1: DP1_TX[P,N][3:0] 0123h DDI 2: DP2_TX[P,N][6:4] DDI 2: DP2_TX[P,N][3:0] Description PCIePllSel. Read-write. Reset: 1. 1=Selects PCIe clock. 0=Selects the display clock. 30:29 Reserved. 28 CascadedPllSel. Read-write. Reset: 0. 1=Selects cascaded PLL clocks. 0=Selects the external display PLL clocks. 27:26 Reserved. 25 DisplayStream. Read-write. Reset: 0. 1=Selects display PLL1 and cascaded PLL clock 1. 0=Selects display PLL0 and cascaded PLL clock 0. 207 42300 Rev 3.12 - July 14, 2015 24 BKDG for AMD Family 15h Models 10h-1Fh Processors ClkOff. Read-write. Reset: 0. 1=Turns off clock to 4 lane wide sub-link within phy. 0=Enables clock to 4 lane wide sublink within phy. 23:0 Reserved. D0F0xE4_x0[2:1]2[3:0]_000[C:B] Phy Serial Bus Packet Control This register provides control to enable or disable various fields contained in the phy serial bus primary control packet, the margining packet and the miscellaneous control packet. Table 80: Index Mapping for D0F0xE4_x0[2:1]2[3:0]_000[C:B] D0F0xE0[31:16] Bits D0F0xE0[15:0] 000Ch 000Bh 0120h GPP ports: P_GPP_[T,R]X[P,N][3:0] FCH ports: P_UMI_[T,R]X[P,N][3:0] 0121h Graphics port lower: P_GFX_[T,R]X[P,N][7:4] Graphics port lower: P_GFX_[T,R]X[P,N][3:0] 0221h Graphics port upper: P_GFX_[T,R]X[P,N][15:12] Graphics port upper: P_GFX_[T,R]X[P,N][11:8] 0122h DDI 0: DP0_TX[P,N][3:0] DDI 1: DP1_TX[P,N][3:0] 0123h DDI 2: DP2_TX[P,N][6:4] DDI 2: DP2_TX[P,N][3:0] Description 31:16 Reserved. 15 PllCmpPktSbiEn. Read-write. Reset: 1. 1=Enables the serial bus PLL component packet used for controlling PLL features such as mode of operation and power states. 14 MargPktSbiEn. Read-write. Reset: 1. IF (REG==D0F0xE4_x012[3:2]_000[C:B]) THEN BIOS: 0. ENDIF. 1=Enables the serial bus margining update packet used for controlling PCIe transmit margining test. 13:9 Reserved. 8 EiDetSbiEn. Read-write. Reset: 1. 1=Enables the electrical idle detector control field in the primary control packet. 7 IncoherentClkSbiEn. Read-write. Reset: 1. 1=Enables the incoherent clock control field in the primary control packet. 6 SkipBitSbiEn. Read-write. Reset: 1. 1=Enables the skip bit control field in the primary control packet. 5 OffsetCancelSbiEn. Read-write. Reset: 1. 1=Enables the offset cancellation control field in the primary control packet. 4 DllLockSbiEn. Read-write. Reset: 1. 1=Enables the DLL lock control field in the primary control packet. 3 FreqDivSbiEn. Read-write. Reset: 1. 1=Enables the frequency divider control field in the primary control packet. 2 PcieModeSbiEn. Read-write. Reset: 1. IF(REG==D0F0xE4_x012[3:2]_000[C:B]) THEN BIOS: 0. ENDIF 1=Enables the phy mode control field in the primary control packet. 1 RxPwrSbiEn. Read-write. Reset: 1. 1=Enables the Rx power state control field in the primary control packet. 0 TxPwrSbiEn. Read-write. Reset: 1. 1=Enables the Tx power state control field in the primary control packet. 208 42300 Rev 3.12 - July 14, 2015 BKDG for AMD Family 15h Models 10h-1Fh Processors D0F0xE4_x0[2:1]2[3:0]_000D Phy Serial Bus Compensation Component Packet Enable This register provides control to enable or disable the phy serial bus compensation component packet. Table 81: Index Mapping for D0F0xE4_x0[2:1]2[3:0]_000D D0F0xE0[31:16] D0F0xE0[15:0] 000Dh Bits 0120h GPP ports: P_GPP_[T,R]X[P,N][3:0] & FCH ports: P_UMI_[T,R]X[P,N][3:0] 0121h Gfx Links[7:0]: P_GFX_[T,R]X[P,N][7:0] 0221h Gfx Links[15:8]: P_GFX_[T,R]X[P,N][15:8] 0122h DDI 1: DP1_TX[P,N][3:0] & DDI 0: DP0_TX[P,N][3:0] 0123h DDI 2: DP2_TX[P,N][6:0] Description 31:1 Reserved. 0 CmpCmpPktSbiEn. Read-write. Reset: 1. 1=Enables the serial bus compensation component packet used for controlling compensation circuit features and power states. D0F0xE4_x0[2:1]2[3:0]_2000 Phy PLL Power State Control This register provides control of the phy PLL component power state. Table 82: Index Mapping for D0F0xE4_x0[2:1]2[3:0]_2000 D0F0xE0[31:16] D0F0xE0[15:0] 2000h Bits 0120h GPP ports: P_GPP_[T,R]X[P,N][3:0] & FCH port: P_UMI_[T,R]X[P,N][3:0] 0121h Gfx Links[7:0]: P_GFX_[T,R]X[P,N][7:0] 0221h Gfx Links[15:8]: P_GFX_[T,R]X[P,N][15:8] 0122h DDI 1: DP1_TX[P,N][3:0] & DDI 0: DP0_TX[P,N][3:0] 0123h DDI 2: DP2_TX[P,N][6:0] Description 31:4 Reserved. 3 PllAutoPwrDownDis. Read-write. Reset: 0. 1=Disables the automatic power down feature. 0=Enables the automatic power down feature; PLL powers down when it determines that it is unused. 2:0 PllPowerDownEn. Read-write. Reset: 0. This filed specifies the power state of the phy PLL circuit component. Bits Definition 000b L0 power state; all circuits are enabled. 001b LS1 power state; all circuits are enabled. 010b LS2 power state; PLL is powered down and clock tree is gated off. 110b-011b Reserved. 111b PHYOFF power state; all circuits are disabled to achieve the lowest power consumption. 209 42300 Rev 3.12 - July 14, 2015 BKDG for AMD Family 15h Models 10h-1Fh Processors D0F0xE4_x0[2:1]2[3:0]_2002 Phy PLL Control The newly written values to this register do not take effect until an explicit shadow update event takes place via setting D0F0xE4_x0[2:1]2[3:0]_2008[PllControlUpdate], or until the PLL component exits the LS2 or PHYOFF power state. Any read from this register always returns the current register value, not the value pending until the next update event. See D0F0xE4_x0[2:1]2[3:0]_2008[PllControlUpdate]. Table 83: Index Mapping for D0F0xE4_x0[2:1]2[3:0]_2002 D0F0xE0[31:16] D0F0xE0[15:0] 2002h Bits 31 0120h GPP ports: P_GPP_[T,R]X[P,N][3:0] & FCH port: P_UMI_[T,R]X[P,N][3:0] 0121h Graphics port upper: P_GFX_[T,R]X[P,N][15:8] 0221h Graphics port lower: P_GFX_[T,R]X[P,N][7:0] 0122h DDI 0: DP0_TX[P,N][3:0] & DDI 1: DP1_TX[P,N][3:0] 0123h DDI 2: DP2_TX[P,N][6:0] Description IsLc: PLL select. Read-write. Reset: 0. 1=Selects the LC tank raw PLL. 0=Selects the ring oscillator raw PLL. 30:28 Reserved. 27 RoCalEn: Ring oscillator calibration enable. Read-write. Reset: 0. BIOS: IF (REG==D0F0xE4_x0120_2002) THEN 1 ELSE 0 ENDIF. A 0 to 1 edge transition of this bit triggers a calibration cycle for the ring oscillator VCO. Reset and set this bit again to trigger another calibration cycle if needed. 26:0 Reserved. D0F0xE4_x0[2:1]2[3:0]_2005 Phy PLL Frequency and Mode Control This register provides PLL operation mode control and frequency selection. The newly written values to this register do not take effect until an explicit shadow update event takes place via setting D0F0xE4_x0[2:1]2[3:0]_2008[PllControlUpdate], or until the PLL component exits the LS2 or PHYOFF power state. Any read from this register always returns the current register value, not the value pending until the next update event. See D0F0xE4_x0[2:1]2[3:0]_2008[PllControlUpdate]. Table 84: Index Mapping for D0F0xE4_x0[2:1]2[3:0]_2005 D0F0xE0[31:16] D0F0xE0[15:0] 2005h 0120h GPP ports: P_GPP_[T,R]X[P,N][3:0] & FCH port: P_UMI_[T,R]X[P,N][3:0] 0121h Gfx Links[7:0]: P_GFX_[T,R]X[P,N][7:0] 0221h Gfx Links[15:8]: P_GFX_[T,R]X[P,N][15:8] 0122h DDI 1: DP1_TX[P,N][3:0] & DDI 0: DP0_TX[P,N][3:0] 0123h DDI 2: DP2_TX[P,N][6:0] 210 42300 Rev 3.12 - July 14, 2015 Bits BKDG for AMD Family 15h Models 10h-1Fh Processors Description 31:15 Reserved. 14:13 PllMode: PLL operation mode. Read-write. Reset: 0. BIOS: See 2.11.4.2.1 [Clock Configuration]. Bits Definition 00b PCIe mode 01b Disabled 10b IsCascadedClk0 mode 11b IsCascadedClk1 mode 12:11 Reserved. 10:9 PllClkFreqExt. Read-write. Reset: 10b. See PllClkFreq. 8:4 Reserved. 3:0 PllClkFreq: PLL frequency select. Read-write. Reset: 0011b. BIOS: See 2.11.4.2.1 [Clock Configuration]. This field is used together with PllClkFreqExt to specify the PLL frequency. {PllClkFreqExt, PllClkFreq} Definition 100000b 810MHz (display port) 100001b 2.7GHz (display port) 100010b 1.35GHz (display port) 100011b 2.5GHz (PCIe Gen 2) 110000b 125MHz - 156.25MHz (DVI/HDMI) 110001b 156.25MHz - 206.25MHz (DVI/HDMI) 110010b 206.25MHz - 250MHz (DVI/HDMI) 110011b 250MHz - 312.5MHz (DVI/HDMI) 110100b 312.5MHz - 412.5MHz (DVI/HDMI) 110101b 412.5MHz - 500MHz (DVI/HDMI) 110110b 500MHz - 625MHz (DVI/HDMI) 110111b 625MHz - 825MHz (DVI/HDMI) all others Reserved D0F0xE4_x0[2:1]2[3:0]_2008 Phy PLL Update Control Table 85: Index Mapping for D0F0xE4_x0[2:1]2[3:0]_2008 D0F0xE0[31:16] D0F0xE0[15:0] 2008h Bits 0120h GPP ports: P_GPP_[T,R]X[P,N][3:0] & FCH port: P_UMI_[T,R]X[P,N][3:0] 0121h Gfx Links[7:0]: P_GFX_[T,R]X[P,N][7:0] 0221h Gfx Links[15:8]: P_GFX_[T,R]X[P,N][15:8] 0122h DDI 1: DP1_TX[P,N][3:0] & DDI 0: DP0_TX[P,N][3:0] 0123h DDI 2: DP2_TX[P,N][6:0] Description 31:30 Reserved. 211 42300 Rev 3.12 - July 14, 2015 BKDG for AMD Family 15h Models 10h-1Fh Processors VdDetectEn: PLL output clock gating enable. Read-write. Reset: 0. BIOS: 1. 1=Enables clock gating during LS2 exit. The internally regulated power supply to the phy is turned off and then back on during LS2 entry and exit. A power sniffer circuit detects the internally regulated power supply level. PLL output clock distribution can be gated off before the regulated power supply reaches the desired level during power up, this effectively prevents possible reliability issue arising from undesirable drivers contention and crow bar current stress during power up upon LS2 exit. Clock distribution is turned on only when the regulated power reaches the right level. 0=PLL output clock distribution is not gated during power up upon LS2 exit. 29 28:26 Reserved. 25:23 MeasCycCntVal[2:0]. Read-write. Reset: 0. This field specifies the time interval over which the LC VCO feedback clock cycles is counted. Bits Definition 000b 3167 count for 31.5kHz modulation 001b 3023 count for 33.0kHz modulation (max spec) 010b 3325 count for 33.0 kHz modulation (min spec) 011b 6334 count, 2 times default (long count) 100b 12667 count, 4 times default (very long count) 101b 2048, 11bits (shorter count for debug) 110b 1024, 10bits 111b 512, 9bits 22:1 Reserved. PllControlUpdate: PLL control register update. Read-write. Reset: 0. 0 to 1 transition of this bit triggers the explicit shadow update event for the phy PLL registers that are shadowed. 0 3.3.3.2 Phy Receiver Lane Control Registers Each receiver lane has a group of registers for controlling the operation of the lane. The mapping from address register to receiver lane is shown in Table 86. Multiple receiver lanes may be written at the same time using per nibble and per byte broadcast write addresses. The mapping from broadcast address to receiver lanes is shown in Table 87. Table 86: Phy per receiver lane register addresses Pin Group D0F0xE0[31:16] D0F0xE0[15:0] 438xh 430xh 428xh 420xh 418xh 410xh 408xh 400xh P_GPP_ 0120h RX[P,N]3 RX[P,N]2 RX[P,N]1 RX[P,N]0 - - - - P_UMI_ 0120h - - - - RX[P,N]3 RX[P,N]2 RX[P,N]1 RX[P,N]0 P_GFX_ 0121h RX[P,N]7 RX[P,N]6 RX[P,N]5 RX[P,N]4 RX[P,N]3 RX[P,N]2 RX[P,N]1 RX[P,N]0 P_GFX_ 0221h RX[P,N]15 RX[P,N]14 RX[P,N]13 RX[P,N]12 RX[P,N]11 RX[P,N]10 RX[P,N]9 RX[P,N]8 Table 87: Phy receiver broadcast register addresses D0F0xE0[31:16] D0F0xE0[15:0] 57[1,0]xh 56[1,0]xh 50[1,0]xh 0120h P_GPP_RX[P,N][3:0] P_UMI_RX[P,N][3:0] P_GPP_RX[P,N][3:0], P_UMI_RX[P,N][3:0] 0121h P_GFX_RX[P,N][7:4] P_GFX_RX[P,N][3:0] P_GFX_RX[P,N][7:0] P_GFX_RX[P,N][15:12] P_GFX_RX[P,N][11:8] P_GFX_RX[P,N][15:8] 0221h 212 42300 Rev 3.12 - July 14, 2015 BKDG for AMD Family 15h Models 10h-1Fh Processors D0F0xE4_x0[2:1]2[1:0]_[5:4][7:6,3:0][8,0]1 Phy Receiver DLL Control and Test 1 These registers provide control of the DLL and duty cycle correction circuit associated with the receive lanes on the phys. Table 88: Index Mapping for D0F0xE4_x0[2:1]2[1:0]_[5:4][7:6,3:0][8,0]1 Pin Group D0F0xE0[31:16] D0F0xE0[15:0] 4381h 4301h 4281h 4201h 4181h 4101h 4081h 4001h P_GPP_ 0120h RX[P,N]3 RX[P,N]2 RX[P,N]1 RX[P,N]0 - - - - P_UMI_ 0120h - - - - RX[P,N]3 RX[P,N]2 RX[P,N]1 RX[P,N]0 P_GFX_ 0121h RX[P,N]7 RX[P,N]6 RX[P,N]5 RX[P,N]4 RX[P,N]3 RX[P,N]2 RX[P,N]1 RX[P,N]0 P_GFX_ 0221h RX[P,N]15 RX[P,N]14 RX[P,N]13 RX[P,N]12 RX[P,N]11 RX[P,N]10 RX[P,N]9 RX[P,N]8 Table 89: Broadcast Mapping for D0F0xE4_x0[2:1]2[1:0]_[5:4][7:6,3:0][8,0]1 D0F0xE0[31:16] D0F0xE0[15:0] 5701h 5601h 5001h 0120h D0F0xE4_x0120_4[3:2][8,0]1 D0F0xE4_x0120_4[1:0][8,0]1 D0F0xE4_x0120_4[3:0][8,0]1 0121h D0F0xE4_x0121_4[3:2][8,0]1 D0F0xE4_x0121_4[1:0][8,0]1 D0F0xE4_x0121_4[3:0][8,0]1 0221h D0F0xE4_x0221_4[3:2][8,0]1 D0F0xE4_x0221_4[1:0][8,0]1 D0F0xE4_x0221_4[3:0][8,0]1 Bits Description 31:16 Reserved. 15 ForceDccRecalc: Force DCC code recalculation. Read-write. Reset: 0. BIOS: IF (REG==D0F0xE4_x0120_4[3:0][8,0]1) THEN 1 ELSE 0 ENDIF. A 0 to 1 edge transition of this bit forces the DCC code to be recalculated when the DLL core loop is locked the next time. This could be useful for debug. When transitioning between the RO PLL to the LC PLL this bit should be set. 14:0 Reserved. D0F0xE4_x0[2:1]2[1:0]_[5:4][7:6,3:0][8,0]2 Phy Receiver Phase Loop Filter Control Data eye from lane Bit time 0 Bit time 1 Bit time 2 Recovered clock Nominal sample clock Offset sample clock SamClkPiOffset Figure 10: Phy recovered clock and sample clock When the link is in a mode that relies on dynamic phase alignment (automatic sample-clock correction), then the processor generates a recovered clock for each lane based on transitions in the lane. The ideal recovered clock transitions at exactly the same time as the transitions in the lane. Phase detection logic detects if the recovered clock transitions before or after the lane transition. The digital loop filter (DLF) is logic that adjusts 213 42300 Rev 3.12 - July 14, 2015 BKDG for AMD Family 15h Models 10h-1Fh Processors the phase of the recovered clock such that its transitions match the transition time of the lane as much as possible. The DLF counts the number of times the lane transitions before the recovered clock versus after to determine whether the recovered clock phase requires adjustment. The DLF uses an 8-bit counter, called the loop filter counter (LFC) for this purpose. The LFC controls are included in this register. They specify DLF behavior as follows: • LfcMax is programmed to be greater than LfcMin. • The LFC is initialized to LfcMin. • The LFC is updated periodically. The logic keeps a tally of the number of lane transitions occurring before and after the recovered clock transition within each update period. • To start, if there is a net lane transition occurs after the recovered clock transition within the update period, the LFC is incremented by the net value; on the other hand, if there is a net lane transition occurs before the recovered clock transition, the LFC is decremented. However, if the LFC is ever decremented while it is zero, these rules are reversed (and the LFC is incremented instead). Thus, if there is a phase correction needed, the LFC trends either upward or downward; if it trends downward, it hits zero and then trends upward again. • If the LFC reaches LfcMax value, then (1) the phase of the recovered clock is adjusted in the appropriate direction, (2) the LFC is set to the LfcMin value. The LfcMin and LfcMax fields are designed to improve the stability of the recovered clock phase while improving the response time for multiple phase updates in the same direction. For example, if the recovered clock phase needs several adjustments in the same direction, then the LFC increments until it hits LfcMax value and then be set to LfcMin (and trigger a phase adjustment); then it would increment to LfcMax value again to trigger the next phase adjustment. If, however, the next phase adjustment needs to be in the opposite direction, the LFC would decrement to zero, change direction, and then increment up to LfcMax again. In this way, phase adjustments in the same direction occur more quickly than phase adjustments in the opposite direction of the prior phase adjustment. The nominal sample clock is offset by 90 degrees from the recovered clock. An offset can be inserted to move the sample clock from the nominal position, based on SamClkPiOffset and SamClkPiOffsetSign. Table 90: Index Mapping for D0F0xE4_x0[2:1]2[1:0]_[5:4][7:6,3:0][8,0]2 Pin Group D0F0xE0[31:16] D0F0xE0[15:0] 4382h 4302h 4282h 4202h 4182h 4102h 4082h 4002h P_GPP_ 0120h RX[P,N]3 RX[P,N]2 RX[P,N]1 RX[P,N]0 - - - - P_UMI_ 0120h - - - - RX[P,N]3 RX[P,N]2 RX[P,N]1 RX[P,N]0 P_GFX_ 0121h RX[P,N]7 RX[P,N]6 RX[P,N]5 RX[P,N]4 RX[P,N]3 RX[P,N]2 RX[P,N]1 RX[P,N]0 P_GFX_ 0221h RX[P,N]15 RX[P,N]14 RX[P,N]13 RX[P,N]12 RX[P,N]11 RX[P,N]10 RX[P,N]9 RX[P,N]8 Table 91: Broadcast Mapping for D0F0xE4_x0[2:1]2[1:0]_[5:4][7:6,3:0][8,0]2 D0F0xE0[31:16] D0F0xE0[15:0] 5702h 5602h 5002h 0120h D0F0xE4_x0120_4[3:2][8,0]2 D0F0xE4_x0120_4[1:0][8,0]2 D0F0xE4_x0120_4[3:0][8,0]2 0121h D0F0xE4_x0121_4[3:2][8,0]2 D0F0xE4_x0121_4[1:0][8,0]2 D0F0xE4_x0121_4[3:0][8,0]2 0221h D0F0xE4_x0221_4[3:2][8,0]2 D0F0xE4_x0221_4[1:0][8,0]2 D0F0xE4_x0221_4[3:0][8,0]2 214 42300 Rev 3.12 - July 14, 2015 Bits BKDG for AMD Family 15h Models 10h-1Fh Processors Description 31:30 Reserved. 29:22 LfcMax: loop filter counter maximum value. Read-write. Reset: 08h. BIOS: 08h. 21:14 LfcMin: loop filter counter minimum value. Read-write. Reset: 00h. BIOS: 00h. 13:8 Reserved. 7 SamClkPiOffsetEn: sample clock phase interpolator offset enable. Read-write. Reset: 0. 1=Enable offset insertion around the nominal sample clock position. 6:4 SamClkPiOffset: sample clock phase interpolator offset setting. Read-write. Reset: X. This field specifies the magnitude of the offset of the sample clock from the nominal position. See Figure 10. This field is encoded as follows. • Sample clock phase interpolator offset = (SamClkPiOffset + 1) * step size. • If link speed is >3.6GT/s, the expected typical step size is 2ps with a +/-1ps error. • If link speed is <=3.6GT/s, the expected typical step size is 3ps with a +/-1ps error. 3 SamClkPiOffsetSign: sample clock phase interpolator offset setting sign bit. Read-write. Reset: X. 0=Sample clock is moved to before the nominal position. 1=Sample clock is moved to after the nominal position. See SamClkPiOffset and Figure 10. 2:0 Reserved. D0F0xE4_x0[2:1]2[1:0]_[5:4][7:6,3:0][8,0]5 Phy Receiver Timing Margin Test The built in jitter injection test mode is useful for checking the clock data recovery tracking bandwidth of the receiver. By forcing the sample clock to move from the lock position by a controlled amount and then observing the time it takes to recover, the tracking rate and bandwidth can be estimated. This register provides the control of the test mode. The jitter injection test mode works as follows. • The circuit is clocked by a jitter injection clock derived from dividing the link forwarded clock by 2.5; for example, if the link speed is 5.2GT/s and the link forwarded clock frequency is 2.6GHz, the jitter injection clock frequency becomes 1.04GHz. • There are 2 phases, the on phase and the off phase. It starts with the on phase once the test mode is enabled. • During the on phase, at every tick of jitter injection clock, the sample clock is moved away from the nominal lock position by 1/96*UI. • The direction of adjustment is specified by JitterInjDir. • The on phase adjustment continues for a number of times as specified by JitterInjOnCnt. • Then the adjustment turns off for a duration specified by {JitterInjOffCnt, JitterInjOnCnt} * jitter injection clock period, this is known as the off phase. During this time, clock data recovery resumes to try to adjust the position of the sample clock back to the center of the data eye. • The off phase is followed by the on phase again. The process continues to alternate between the on phase and the off phase until the jitter injection test mode is disabled. In addition, the JitterInjHold bit may be set to inject a hold state at the end of the on phase. This stops clock data recovery from resuming after the on phase, hence holding the sample clock at its last adjusted position until the JitterInjHold bit is cleared. This test mode may be useful for margining the width of the input data eye. This margining mechanism is not characterized for precision jitter adjustments or measurements. 215 42300 Rev 3.12 - July 14, 2015 BKDG for AMD Family 15h Models 10h-1Fh Processors Table 92: Index Mapping for D0F0xE4_x0[2:1]2[1:0]_[5:4][7:6,3:0][8,0]5 Pin Group D0F0xE0[31:16] D0F0xE0[15:0] 4385h 4285h 4205h 4185h 4105h 4085h 4005h RX[P,N]2 4305h RX[P,N]1 RX[P,N]0 - - - - P_GPP_ 0120h RX[P,N]3 P_UMI_ 0120h - - - - RX[P,N]3 RX[P,N]2 RX[P,N]1 RX[P,N]0 P_GFX_ 0121h RX[P,N]7 RX[P,N]6 RX[P,N]5 RX[P,N]4 RX[P,N]3 RX[P,N]2 RX[P,N]1 RX[P,N]0 P_GFX_ 0221h RX[P,N]15 RX[P,N]14 RX[P,N]13 RX[P,N]12 RX[P,N]11 RX[P,N]10 RX[P,N]9 RX[P,N]8 Table 93: Broadcast Mapping for D0F0xE4_x0[2:1]2[1:0]_[5:4][7:6,3:0][8,0]5 D0F0xE0[31:16] D0F0xE0[15:0] 5705h 5605h 5005h 0120h D0F0xE4_x0120_4[3:2][8,0]5 D0F0xE4_x0120_4[1:0][8,0]5 D0F0xE4_x0120_4[3:0][8,0]5 0121h D0F0xE4_x0121_4[3:2][8,0]5 D0F0xE4_x0121_4[1:0][8,0]5 D0F0xE4_x0121_4[3:0][8,0]5 0221h D0F0xE4_x0221_4[3:2][8,0]5 D0F0xE4_x0221_4[1:0][8,0]5 D0F0xE4_x0221_4[3:0][8,0]5 Bits Description 31 Reserved. 30 JitterInjEn: jitter injection enable. Read-write. Reset: 0. 1=Jitter injection test mode is enabled. 29 JitterInjDir: jitter injection direction. Read-write. Reset: 0. Bit Definition 0 Move clock before the nominal lock position. 1 Move clock after the nominal lock position. 28:23 JitterInjOnCnt: jitter injection on count. Read-write. Reset: 0. 22:16 Reserved. 15:10 JitterInjOffCnt: jitter injection off count. Read-write. Reset: 0. The jitter injection off time count is a 12bit code, this field specifies the most significant 6 bits. The least significant 6 bits are the same as JitterInjOnCnt. 9 8:0 JitterInjHold: jitter injection hold. Read-write. Reset: 0. 1=Jitter injection hold is enabled. Reserved. D0F0xE4_x0[2:1]2[1:0]_[5:4][7:6,3:0][8,0]6 Phy Receiver DFE and DFR Control The processor supports decision feedback restore (DFR), a function that enables on-chip AC coupling on the receiver path, to improve the receiver’s ability to operate over a longer channel. In this mode, the receiver on the processor must be programmed with the expected peak single-ended DC voltage level over the singleended DC common mode voltage level, as seen by the receiver, when a static 1 or 0 is driven. For example, without deemphasis at nominal supply voltage of 1.2V, the peak single ended voltage is expected to be 300mV ideally above the single ended DC common mode voltage level. The value is dependent on the deemphasis setting of the transmitter on the other end of the channel. Table 94: Recommended DCV settings Far-device DCV deemphasis setting 216 42300 Rev 3.12 - July 14, 2015 BKDG for AMD Family 15h Models 10h-1Fh Processors Table 94: Recommended DCV settings No deemphasis -3dB postcursor -6dB postcursor 20h 17h 10h Decision feedback equalization (DFE) can be enabled to enhance link operation. Once enabled, the receiver uses the logic level of the previous data bit to adjust the voltage threshold of the sampler in the direction that causes the sampler to switch sooner when the data bit transitions to the opposite logic level for the next bit. The control and DFE voltage level are included in this register. Table 95: Index Mapping for D0F0xE4_x0[2:1]2[1:0]_[5:4][7:6,3:0][8,0]6 Pin Group D0F0xE0[31:16] D0F0xE0[15:0] 4386h 4306h 4286h 4206h 4186h 4106h 4086h 4006h P_GPP_ 0120h RX[P,N]3 RX[P,N]2 RX[P,N]1 RX[P,N]0 - - - - P_UMI_ 0120h - - - - RX[P,N]3 RX[P,N]2 RX[P,N]1 RX[P,N]0 P_GFX_ 0121h RX[P,N]7 RX[P,N]6 RX[P,N]5 RX[P,N]4 RX[P,N]3 RX[P,N]2 RX[P,N]1 RX[P,N]0 P_GFX_ 0221h RX[P,N]15 RX[P,N]14 RX[P,N]13 RX[P,N]12 RX[P,N]11 RX[P,N]10 RX[P,N]9 RX[P,N]8 Table 96: Broadcast Mapping for D0F0xE4_x0[2:1]2[1:0]_[5:4][7:6,3:0][8,0]6 D0F0xE0[31:16] D0F0xE0[15:0] 5706h 5606h 5006h 0120h D0F0xE4_x0120_4[3:2][8,0]6 D0F0xE4_x0120_4[1:0][8,0]6 D0F0xE4_x0120_4[3:0][8,0]6 0121h D0F0xE4_x0121_4[3:2][8,0]6 D0F0xE4_x0121_4[1:0][8,0]6 D0F0xE4_x0121_4[3:0][8,0]6 0221h D0F0xE4_x0221_4[3:2][8,0]6 D0F0xE4_x0221_4[1:0][8,0]6 D0F0xE4_x0221_4[3:0][8,0]6 Bits Description 31:8 Reserved. 7 DfeEn: DFE enable. Read-write. Reset: 0. 1=Decision feedback equalization is enabled. 6:5 DfeVoltage: DFE offset voltage level. Read-write. Reset: 0. This field specifies the magnitude of the DFE offset voltage. Bits Definition Bits Definition 00b DFE offset voltage=25mV. 10b DFE offset voltage=12.5mV. 01b DFE offset voltage=0mV. 11b DFE offset voltage=31.25mV. 4:0 Reserved. D0F0xE4_x0[2:1]2[1:0]_[5:4][7:6,3:0][8,0]A Phy DLL Test and Control 3 Table 97: Index Mapping for D0F0xE4_x0[2:1]2[1:0]_[5:4][7:6,3:0][8,0]A Pin Group D0F0xE0[31:16] D0F0xE0[15:0] 438Ah 430Ah 428Ah 420Ah 418Ah 410Ah 408Ah 400Ah P_GPP_ 0120h RX[P,N]3 RX[P,N]2 RX[P,N]1 RX[P,N]0 - - - - P_UMI_ 0120h - - - - RX[P,N]3 RX[P,N]2 RX[P,N]1 RX[P,N]0 217 42300 Rev 3.12 - July 14, 2015 BKDG for AMD Family 15h Models 10h-1Fh Processors Table 97: Index Mapping for D0F0xE4_x0[2:1]2[1:0]_[5:4][7:6,3:0][8,0]A P_GFX_ 0121h RX[P,N]7 RX[P,N]6 RX[P,N]5 RX[P,N]4 RX[P,N]3 RX[P,N]2 RX[P,N]1 RX[P,N]0 P_GFX_ 0221h RX[P,N]15 RX[P,N]14 RX[P,N]13 RX[P,N]12 RX[P,N]11 RX[P,N]10 RX[P,N]9 RX[P,N]8 Table 98: Broadcast Mapping for D0F0xE4_x0[2:1]2[1:0]_[5:4][7:6,3:0][8,0]A D0F0xE0[31:16] D0F0xE0[15:0] 570Ah 560Ah 500Ah 0120h D0F0xE4_x0120_4[3:2][8,0]A D0F0xE4_x0120_4[1:0][8,0]A D0F0xE4_x0120_4[3:0][8,0]A 0121h D0F0xE4_x0121_4[3:2][8,0]A D0F0xE4_x0121_4[1:0][8,0]A D0F0xE4_x0121_4[3:0][8,0]A 0221h D0F0xE4_x0221_4[3:2][8,0]A D0F0xE4_x0221_4[1:0][8,0]A D0F0xE4_x0221_4[3:0][8,0]A Bits Description 31:29 Ls2ExitTime: LS2 exit time. Read-write. Reset: 0. BIOS: 001b. This field selects the internal timer that delays the turn-on of the DLL after exit from LS2 state to L0 state. The added delay allows the forwarded input clock to achieve better stability. Bits Definition 000b Delay=10us. 001b Delay=5us. 010b Delay=2.5us. 011b Delay=1.25us. 100b Delay=625ns. 101b Delay=0s. 11xb Reserved. The value specified by Ls2ExitTime must be less than the value specified by T0Time, or it can cause undefined behavior. 28:18 Reserved. 17 DllLockFastModeEn: DLL lock fast mode enable. Read-write. Reset: 0. 1=Enables DLL lock fast mode. 0=DLL lock operates at standard speed. 16:15 Reserved. 14:13 AnalogWaitTime: analog wait time to turn on DLL. Read-write. Reset: 0. The turning on of the DLL circuit after cold reset is delayed by a timer specified by this field. The encodings are as follows: Bits Definition Bits Definition 00b Delay=1.25us. 10b Delay=2.5us. 01b Delay=0.625us. 11b Delay=0.3125us. 12:8 Reserved. 218 42300 Rev 3.12 - July 14, 2015 7 6:5 BKDG for AMD Family 15h Models 10h-1Fh Processors BiasDisInLs2: bias disable in LS2 power state. Read-write. Reset: 0. IF ((REG==D0F0xE4_x0121_4[1:0][8,0]A && D0F0xE4_x0131_8040[OwnSlice]==0) || (REG==D0F0xE4_x0121_4[3:2][8,0]A && D0F0xE4_x0131_8041[OwnSlice]==0) || (REG==D0F0xE4_x0221_4[1:0][8,0]A && D0F0xE4_x0131_8042[OwnSlice]==0) || (REG==D0F0xE4_x0221_4[3:2][8,0]A && D0F0xE4_x0131_8043[OwnSlice]==0) || REG==D0F0xE4_x0120_4[1:0][8,0]A || REG==D0F0xE4_x0120_4[3:2][8,0]A) THEN BIOS: 1. ENDIF. 1=Enables lower power LS2 state; current consumption is lowered by approximately 2.5mA per receive lane when compared to standard LS2 power mode. Setting this bit increases the amount of T0Time needed to relock the DLL.When this bit is set, Ls2ExitTime must be programmed to select a value that is greater than or equal to AnalogWaitTime. 0=Standard LS2 power mode. Reserved. LockDetOnLs2Exit: DLL lock detect on LS2 exit. Read-write. Reset: 0. This field selects the LS2 to L0 power state transition speed. 1=Fast transition mode selected. 0=Slow transition mode selected. 4 3:2 Reserved. 1 RxPcieMode: Receiver PCI Express Mode. Read-write. Reset: 0. Indicates whether the receiver lane is in Gen1 or Gen2 mode. 0=Gen1 mode. 1=Gen2 mode. 0 EnCoreLoopFirst: enable DLL core loop first on LS2 exit. Read-write. Reset: 0. This field selects LS2 to L0 power state transition speed. 1=Fast transition mode selected. 0=Slow transition mode selected. D0F0xE4_x0[2:1]2[1:0]_[5:4][7:6,3:0][8,0]F Phy Receiver DLL Test and Debug 5 Table 99: Index Mapping for D0F0xE4_x0[2:1]2[1:0]_[5:4][7:6,3:0][8,0]F Pin Group D0F0xE0[31:16] D0F0xE0[15:0] 438Fh 430Fh 428Fh 420Fh 418Fh 410Fh 408Fh 400Fh P_GPP_ 0120h RX[P,N]3 RX[P,N]2 RX[P,N]1 RX[P,N]0 - - - - P_UMI_ 0120h - - - - RX[P,N]3 RX[P,N]2 RX[P,N]1 RX[P,N]0 P_GFX_ 0121h RX[P,N]7 RX[P,N]6 RX[P,N]5 RX[P,N]4 RX[P,N]3 RX[P,N]2 RX[P,N]1 RX[P,N]0 P_GFX_ 0221h RX[P,N]15 RX[P,N]14 RX[P,N]13 RX[P,N]12 RX[P,N]11 RX[P,N]10 RX[P,N]9 RX[P,N]8 Table 100: Broadcast Mapping for D0F0xE4_x0[2:1]2[1:0]_[5:4][7:6,3:0][8,0]F D0F0xE0[31:16] D0F0xE0[15:0] 570Fh 560Fh 500Fh 0120h D0F0xE4_x0120_4[3:2][8,0]F D0F0xE4_x0120_4[1:0][8,0]F D0F0xE4_x0120_4[3:0][8,0]F 0121h D0F0xE4_x0121_4[3:2][8,0]F D0F0xE4_x0121_4[1:0][8,0]F D0F0xE4_x0121_4[3:0][8,0]F 0221h D0F0xE4_x0221_4[3:2][8,0]F D0F0xE4_x0221_4[1:0][8,0]F D0F0xE4_x0221_4[3:0][8,0]F Bits Description 31:13 Reserved 12 DllProcessFreqCtlOverride. Read-write. Reset: 0. 1=Enables the override of DLL delay line capacitance settings with the values of DllProcessFreqCtlIndex1 and DllProcessFreqCtlIndex2. 11 Reserved. 219 42300 Rev 3.12 - July 14, 2015 BKDG for AMD Family 15h Models 10h-1Fh Processors 10:7 DllProcessFreqCtlIndex2. Read-write. Reset: 0. This field specifies the DLL delay line capacitance setting in PCIe Gen2 mode. 6:4 Reserved. 3:0 DllProcessFreqCtlIndex1. Read-write. Reset: 0. This field specifies the DLL delay line capacitance setting in PCIe Gen1 mode. 3.3.3.3 Phy Transmitter Lane Control Registers Each transmitter lane has a group of registers for controlling the operation of the lane. The mapping from address register to transmitter lane is shown in Table 101. Multiple transmitter lanes may be written at the same time using per nibble and per byte broadcast write addresses. The mapping from broadcast address to transmitter lanes is shown in Table 102. Table 101: Phy per transmitter lane register addresses Pin Group D0F0xE0[31:16] D0F0xE0[15:0] 638xh 630xh 628xh 620xh 618xh 610xh 608xh 600xh P_GPP_ 0120h RX[P,N]3 RX[P,N]2 RX[P,N]1 RX[P,N]0 - - - - P_UMI_ 0120h - - - - RX[P,N]3 RX[P,N]2 RX[P,N]1 RX[P,N]0 P_GFX_ 0121h RX[P,N]7 RX[P,N]6 RX[P,N]5 RX[P,N]4 RX[P,N]3 RX[P,N]2 RX[P,N]1 RX[P,N]0 P_GFX_ 0221h RX[P,N]15 RX[P,N]14 RX[P,N]13 RX[P,N]12 RX[P,N]11 RX[P,N]10 RX[P,N]9 RX[P,N]8 P_DP0_ 0122h TX[P,N]3 TX[P,N]2 TX[P,N]1 TX[P,N]0 - - - - P_DP1_ 0122h - - - - TX[P,N]3 TX[P,N]2 TX[P,N]1 TX[P,N]0 P_DP2_ 0123h - TX[P,N]6 TX[P,N]5 TX[P,N]4 TX[P,N]3 TX[P,N]2 TX[P,N]1 TX[P,N]0 Table 102: Phy transmitter broadcast register addresses D0F0xE0[31:16] D0F0xE0[15:0] 77[1,0]xh 76[1,0]xh 70[1,0]xh 0120h P_GPP_RX[P,N][3:0] P_UMI_RX[P,N][3:0] P_GPP_RX[P,N][3:0], P_UMI_RX[P,N][3:0] 0121h P_GFX_RX[P,N][7:4] P_GFX_RX[P,N][3:0] P_GFX_RX[P,N][7:0] P_GFX_RX[P,N][15:12] P_GFX_RX[P,N][11:8] P_GFX_RX[P,N][15:8] 0221h D0F0xE4_x0[2:1]2[3:0]_[7:6][7:6,3:0][8,0]0 Phy Tx Deemphasis and Margining Control Table 103: Index Mapping for D0F0xE4_x0[2:1]2[3:0]_[7:6][7:6,3:0][8,0]0 Pin Group D0F0xE0[31:16] D0F0xE0[15:0] 6380h 6300h 6280h 6200h TX[P,N]3 TX[P,N]2 TX[P,N]1 TX[P,N]0 6180h 6100h 6080h 6000h - - - - TX[P,N]3 TX[P,N]2 TX[P,N]1 TX[P,N]0 TX[P,N]2 P_GPP_ 0120h P_UMI_ 0120h P_GFX_ 0121h TX[P,N]7 TX[P,N]6 TX[P,N]5 TX[P,N]4 TX[P,N]3 TX[P,N]1 TX[P,N]0 P_GFX_ 0221h TX[P,N]15 TX[P,N]14 TX[P,N]13 TX[P,N]12 TX[P,N]11 TX[P,N]10 TX[P,N]9 TX[P,N]8 P_DP0_ 0122h TX[P,N]3 TX[P,N]2 TX[P,N]1 TX[P,N]0 P_DP1_ 0122h - P_DP2_ 0123h - - - - - - - - - TX[P,N]3 TX[P,N]2 TX[P,N]1 TX[P,N]0 TX[P,N]6 TX[P,N]5 TX[P,N]4 TX[P,N]3 TX[P,N]2 TX[P,N]1 TX[P,N]0 - - - 220 42300 Rev 3.12 - July 14, 2015 BKDG for AMD Family 15h Models 10h-1Fh Processors Table 104: Broadcast Mapping for D0F0xE4_x0[2:1]2[3:0]_[7:6][7:6,3:0][8,0]0 D0F0xE0[31:16] D0F0xE0[15:0] 7700h 7600h 7000h 0120h D0F0xE4_x0120_6[3:2][8,0]0 D0F0xE4_x0120_6[1:0][8,0]0 D0F0xE4_x0120_6[3:0][8,0]0 0121h D0F0xE4_x0121_6[3:2][8,0]0 D0F0xE4_x0121_6[1:0][8,0]0 D0F0xE4_x0121_6[3:0][8,0]0 0221h D0F0xE4_x0221_6[3:2][8,0]0 D0F0xE4_x0221_6[1:0][8,0]0 D0F0xE4_x0221_6[3:0][8,0]0 0122h D0F0xE4_x0122_6[3:2][8,0]0 D0F0xE4_x0122_6[1:0][8,0]0 D0F0xE4_x0122_6[3:0][8,0]0 0123h D0F0xE4_x0123_6[3:2][8,0]0 D0F0xE4_x0123_6[1:0][8,0]0 D0F0xE4_x0123_6[3:0][8,0]0 Bits Description 31:16 Reserved. 15 DisLoImpIdle: disable low impedance idle. Read-write. Reset: 0. 1= Disables the low impedance electrical idle feature that requires both the true and complement pins of the transmitter to be pulled to VDDP/2 via low impedance termination in the range of 25 to 50 ohm upon entering electrical idle state. Instead, 5k ohm termination is used. 0=Enables low impedance electrical idle mode. 14:8 Reserved. 7 6:4 3 2:0 TxLs23ClkGateEn: LS2/LS3 clock gating enable. Read-write. Reset: 1. 1= Internal phy clock grids are gated during LS2 or PHY OFF states to save power. Reserved. Post2Sign: Post-cursor 2 Sign. Read-write. Cold-reset: 0. 1=Increases output voltage strength. 0=Lowers output voltage strength. Reserved. D0F0xE4_x0[2:1]2[3:0]_[7:6][7:6,3:0][8,0]5 Phy Transmit Link Configuration Table 105: Recommended link configuration Link configuration GangedModeEn 0 0 0 1 x1 (1 lane per sublink) x2 (2 lanes per sublink) x4 (4 lanes per sublink) x8 IsOwnMstr 1 1 0 0 Table 106: Index Mapping for D0F0xE4_x0[2:1]2[3:0]_[7:6][7:6,3:0][8,0]5 Pin Group D0F0xE0[31:16] D0F0xE0[15:0] 6385h 6305h 6285h 6205h 6185h 6105h 6085h 6005h P_GPP_ 0120h RX[P,N]3 RX[P,N]2 RX[P,N]1 RX[P,N]0 - - - - P_UMI_ 0120h - - - - RX[P,N]3 RX[P,N]2 RX[P,N]1 RX[P,N]0 P_GFX_ 0121h RX[P,N]7 RX[P,N]6 RX[P,N]5 RX[P,N]4 RX[P,N]3 RX[P,N]2 RX[P,N]1 RX[P,N]0 P_GFX_ 0221h RX[P,N]15 RX[P,N]14 RX[P,N]13 RX[P,N]12 RX[P,N]11 RX[P,N]10 RX[P,N]9 RX[P,N]8 P_DP0_ 0122h TX[P,N]3 TX[P,N]2 TX[P,N]1 TX[P,N]0 - - - - 221 42300 Rev 3.12 - July 14, 2015 BKDG for AMD Family 15h Models 10h-1Fh Processors Table 106: Index Mapping for D0F0xE4_x0[2:1]2[3:0]_[7:6][7:6,3:0][8,0]5 P_DP1_ 0122h - - - - TX[P,N]3 TX[P,N]2 TX[P,N]1 TX[P,N]0 P_DP2_ 0123h - TX[P,N]6 TX[P,N]5 TX[P,N]4 TX[P,N]3 TX[P,N]2 TX[P,N]1 TX[P,N]0 Table 107: Broadcast Mapping for D0F0xE4_x0[2:1]2[3:0]_[7:6][7:6,3:0][8,0]5 D0F0xE0[31:16] D0F0xE0[15:0] 7705h 7605h 7005h 0120h D0F0xE4_x0120_6[3:2][8,0]5 D0F0xE4_x0120_6[1:0][8,0]5 D0F0xE4_x0120_6[3:0][8,0]5 0121h D0F0xE4_x0121_6[3:2][8,0]5 D0F0xE4_x0121_6[1:0][8,0]5 D0F0xE4_x0121_6[3:0][8,0]5 0221h D0F0xE4_x0221_6[3:2][8,0]5 D0F0xE4_x0221_6[1:0][8,0]5 D0F0xE4_x0221_6[3:0][8,0]5 0122h D0F0xE4_x0122_6[3:2][8,0]5 D0F0xE4_x0122_6[1:0][8,0]5 D0F0xE4_x0122_6[3:0][8,0]5 0123h D0F0xE4_x0123_6[3:2][8,0]5 D0F0xE4_x0123_6[1:0][8,0]5 D0F0xE4_x0123_6[3:0][8,0]5 Bits Description 31 GangedModeEn. Read-write. Reset: 1. BIOS: Table 105. 1=Enables link ganged mode. 0=Disables link ganged mode. 30 Reserved. 29 IsOwnMstr. Read-write. Reset: 0. BIOS: Table 105. 1=Enables the lane to self initialize its own read pointer. 28:0 Reserved. D0F0xE4_x0[2:1]2[3:0]_[7:6][7:6,3:0][8,0]6 Phy Transmit Nominal Deemphasis Control This register specifies the deemphasis, or preemphasis settings in the case of display port mode, and voltage margining settings for the transmit drivers. Table 108: Recommended preemphasis settings Conditions Link Type D0F0xE4_x0[2:1]2[3:0]_[7:6][7:6,3:0][8,0]6 Preemphasis Peak-to-peak Voltage DeemphGen1Nom TxMarginNom 1.2V 0 0 0.8V 0 42 0.6V 0 64 0.4V 0 85 0.8V 42 0 0.6V 32 32 0.4V 21 64 0.6V 64 0 0.4V 42 42 9.5dB 0.4V 85 0 PCIe || HDMI - 1.2V 42 0 DVI - 1.2V 11 0 0dB Display Port 3.5dB 6dB 222 42300 Rev 3.12 - July 14, 2015 BKDG for AMD Family 15h Models 10h-1Fh Processors Table 109: Index Mapping for D0F0xE4_x0[2:1]2[3:0]_[7:6][7:6,3:0][8,0]6 Pin Group D0F0xE0[31:16] D0F0xE0[15:0] 6386h 6306h 6286h 6206h TX[P,N]3 TX[P,N]2 TX[P,N]1 TX[P,N]0 6186h 6106h 6086h 6006h - - - - TX[P,N]3 TX[P,N]2 TX[P,N]1 TX[P,N]0 TX[P,N]2 P_GPP_ 0120h P_UMI_ 0120h P_GFX_ 0121h TX[P,N]7 TX[P,N]6 TX[P,N]5 TX[P,N]4 TX[P,N]3 TX[P,N]1 TX[P,N]0 P_GFX_ 0221h TX[P,N]15 TX[P,N]14 TX[P,N]13 TX[P,N]12 TX[P,N]11 TX[P,N]10 TX[P,N]9 TX[P,N]8 P_DP0_ 0122h TX[P,N]3 TX[P,N]2 TX[P,N]1 TX[P,N]0 P_DP1_ 0122h - 0123h - P_DP2_ - - - - - - TX[P,N]3 TX[P,N]2 TX[P,N]1 TX[P,N]0 TX[P,N]6 TX[P,N]5 TX[P,N]4 TX[P,N]3 TX[P,N]2 TX[P,N]1 TX[P,N]0 - - - - - Table 110: Broadcast Mapping for D0F0xE4_x0[2:1]2[3:0]_[7:6][7:6,3:0][8,0]6 D0F0xE0[31:16] D0F0xE0[15:0] 7706h 7606h 7006h 0120h D0F0xE4_x0120_6[3:2][8,0]6 D0F0xE4_x0120_6[1:0][8,0]6 D0F0xE4_x0120_6[3:0][8,0]6 0121h D0F0xE4_x0121_6[3:2][8,0]6 D0F0xE4_x0121_6[1:0][8,0]6 D0F0xE4_x0121_6[3:0][8,0]6 0221h D0F0xE4_x0221_6[3:2][8,0]6 D0F0xE4_x0221_6[1:0][8,0]6 D0F0xE4_x0221_6[3:0][8,0]6 0122h D0F0xE4_x0122_6[3:2][8,0]6 D0F0xE4_x0122_6[1:0][8,0]6 D0F0xE4_x0122_6[3:0][8,0]6 0123h D0F0xE4_x0123_6[3:2][8,0]6 D0F0xE4_x0123_6[1:0][8,0]6 D0F0xE4_x0123_6[3:0][8,0]6 Bits Description 31:16 Reserved. 15:8 DeemphGen1Nom. Read-write. Reset: 42. BIOS: Table 108.This field specifies the post cursor deemphasis setting. Value must be less than or equal to 104. 7:0 TxMarginNom. Read-write. Reset: 0. BIOS: Table 108.This field specifies the voltage margining setting of the transmit driver. Value must be less than or equal to 104. D0F0xE4_x0[2:1]2[1:0]_[D:C][7:0][8,0]5 Termination Mode Control Table 111: Index Mapping for D0F0xE4_x0[2:1]2[1:0]_[D:C][7:0][8,0]5 Pin Group D0F0xE0[31:16] D0F0xE0[15:0] C385h C305h C285h C205h C185h C105h C085h C005h P_GPP_ 0120h RX[P,N]3 RX[P,N]2 RX[P,N]1 RX[P,N]0 - - - - P_UMI_ 0120h - - - - RX[P,N]3 RX[P,N]2 RX[P,N]1 RX[P,N]0 P_GFX_ 0121h RX[P,N]7 RX[P,N]6 RX[P,N]5 RX[P,N]4 RX[P,N]3 RX[P,N]2 RX[P,N]1 RX[P,N]0 P_GFX_ 0221h RX[P,N]15 RX[P,N]14 RX[P,N]13 RX[P,N]12 RX[P,N]11 RX[P,N]10 RX[P,N]9 RX[P,N]8 Table 112: Broadcast Mapping for D0F0xE4_x0[2:1]2[1:0]_[D:C][7:0][8,0]5 D0F0xE0[31:16] D0F0xE0[15:0] D705h D605h D005h 223 42300 Rev 3.12 - July 14, 2015 BKDG for AMD Family 15h Models 10h-1Fh Processors Table 112: Broadcast Mapping for D0F0xE4_x0[2:1]2[1:0]_[D:C][7:0][8,0]5 0120h D0F0xE4_x0120_C[3:2][8,0]5 D0F0xE4_x0120_C[1:0][8,0]5 D0F0xE4_x0120_C[3:0][8,0]5 0121h D0F0xE4_x0121_C[3:2][8,0]5 D0F0xE4_x0121_C[1:0][8,0]5 D0F0xE4_x0121_C[3:0][8,0]5 0221h D0F0xE4_x0221_C[3:2][8,0]5 D0F0xE4_x0221_C[1:0][8,0]5 D0F0xE4_x0221_C[3:0][8,0]5 Bits Description 31:2 Reserved. 1:0 3.3.4 TermMode. Read-write; updated-by-hardware. Cold-reset: 10b. Receiver termination mode. Bits Description 00b 100 Ohm differential DC, keepers on. 01b 50 Ohm single ended to ground. 10b High impedance 11b 100 ohm differential DC, keepers off. Wrapper Registers Table 113: Mapping for wrapper registers D0F0xE0[31:16] 0130h 0131h 0132h 0133h Wrapper GPPSB Gfx DDI DDI2 D0F0xE4_x013[1:0]_0046 Subsystem and Vendor ID Bits Description 31:16 SubsystemID: subystem id. Read-write. Reset: 1234h. Specifies the value returned by D[8:2]F0xB4[SubsystemID]. 15:0 SubsystemVendorID: subsystem vendor id . Read-write. Reset: 1022h. Specifies the value returned by D[8:2]F0xB4[SubsystemVendorID]. D0F0xE4_x013[1:0]_0080 Link Configuration Bits Description 31:4 Reserved. 3:0 StrapBifLinkConfig. Read-write; strap. Reset: Product-specific. BIOS: See Table 49 and Table 53. Bits Definition Bits Definition 0000b x16 IO Link (Gfx Only) 0100b 4 x1 IO Links (GPPFCH Only) 0001b x4 IO Link (GPPFCH Only) 0101b 2 x8 IO Links (Gfx Only) 0010b 2 x2 IO Links (GPPFCH Only) 011xb Reserved 0011b 1 x2 IO Link, 2 x1 IO Links (GPPFCH Only) 1xxxb Reserved 224 42300 Rev 3.12 - July 14, 2015 BKDG for AMD Family 15h Models 10h-1Fh Processors D0F0xE4_x013[1:0]_0[C:8]00 Link Training Control Table 114: Index address mapping for D0F0xE4_x013[1:0]_0[C:8]00 Index 0130_0800h 0130_0900h 0130_0A00h 0130_0B00h Bits Function GPPFCH Port A GPPFCH Port B GPPFCH Port C GPPFCH Port D Reset Index Function Reset 0000_0000h 0130_0C00h GPPFCH Port E 0000_0001h 0000_0001h 0131_0800h Gfx Port A 0000_0001h 0000_0001h 0131_0900h Gfx Port B 0000_0001h 0000_0001h - Description 31:1 Reserved. 0 HoldTraining: hold link training. Read-write. 1=Hold training on link. D0F0xE4_x013[1:0]_0[C:8]03 Link Deemphasis Control Table 115: Index address mapping for D0F0xE4_x013[1:0]_0[C:8]03 Index 0130_0803h 0130_0903h 0130_0A03h 0130_0B03h Bits Function GPPFCH Port A GPPFCH Port B GPPFCH Port C GPPFCH Port D Index 0130_0C03h 0131_0803h 0131_0903h - Function GPPFCH Port E Gfx Port A Gfx Port B - Description 31:6 Reserved. 5 4:0 StrapBifDeemphasisSel. Read-write; strap. Reset: 1. Controls the default value of D[8:2]F0x88[SelectableDeemphasis]. Reserved. D0F0xE4_x013[3:0]_8011 Link Transmit Clock Gating Control Bits 31 Description StrapBifValid. Read-write. IF (REG==D0F0xE4_x0130_8011) THEN Reset: 0. ELSE Reset: 1. ENDIF. BIOS: 2.11.4.3.1. 0=Straps are latched. 1=Straps are not latched. 30:26 Reserved. 25 DdiDualLinkOverride. Read-write. Reset: 0. 1=Dual link DDI is clocked as a single link. 24 TxclkLcntGateEnable. Read-write. Reset: 0. BIOS: 1. 1=Enable clock gating the lane counter. 23 DebugBusClkEnable. Read-write. Reset: 0. 1=Enable the debug bus clock. 22:17 TxclkPermGateLatency. Read-write. Reset: 3Fh. Specifies the number of clocks to wait after detecting an entry into L1 before gating off the permanent clock branches. 16 RcvrDetClkEnable. Read-write. Reset: 0. 1=Enable the receiver detect clock. 225 42300 Rev 3.12 - July 14, 2015 BKDG for AMD Family 15h Models 10h-1Fh Processors 15:10 TxclkRegsGateLatency. Read-write. Reset: 3Fh. Specifies the number of clocks to wait after idle is signalled before gating off the register clock branch. 9 TxclkRegsGateEnable. Read-write. Reset: 0. BIOS: 1. 1=Enable clock gating the register clock. 8 TxclkPermStop. Read-write. Reset: 0. 1=All transmitter clocks disabled. This bit should only be set if all links associated with the PCIe core are unconnected. 7 TxclkDynGateEnable. Read-write. Reset: 0. BIOS: 1. 1=Dynamic clock gating enabled. 0=Dynamic clock gating disabled. 6 TxclkPermGateEven. Read-write. Reset: 1. 1=Gate the permanent clock branches for an even number of clocks. 5:0 TxclkDynGateLatency. Read-write. Reset: 3Fh. Specifies the number of clocks to wait after idle is signalled before gating off the dynamic clock branch. D0F0xE4_x013[3:0]_8012 Link Idle-Resume Clock Gating Control Bits Description 31:30 Reserved. 29:24 Pif2p5xIdleResumeLatency. Read-write. Reset: 0_0111b. Specifies the number of clocks to wait after enabling TXCLK2P5X_PIF before sending the acknowledge. 23 Pif2p5xIdleGateEnable. Read-write. Reset: 0. BIOS: 1. 1=Enable idle resume gating of TXCLK2P5X_PIF. 22 Reserved. 21:16 Pif2p5xIdleGateLatency. Read-write. Reset: 0_0001b. Specifies the number of clocks to wait before turning off TXCLK2P5X_PIF. 15:14 Reserved. 13:8 Pif1xIdleResumeLatency. Read-write. Reset: 0_0111b. Specifies the number of clocks to wait after enabling TXCLK1X_PIF before sending the acknowledge. 7 Pif1xIdleGateEnable. Read-write. Reset: 0. BIOS: 1. 1=Enable idle resume gating of TXCLK1X_PIF. 6 Reserved. 5:0 Pif1xIdleGateLatency. Read-write. Reset: 0_0001b. Specifies the number of clocks to wait before turning off TXCLK1X_PIF. D0F0xE4_x013[3:0]_8013 Transmit Clock Pll Control Reset: 0000_0001h. Table 116: Reserved field mappings for D0F0xE4_x013[3:0]_8013 Register Bits 20:13 12:11 10 7:6 5 3:2 1 D0F0xE4_x0130_8013 Reserved Reserved Reserved Reserved Reserved Reserved Reserved D0F0xE4_x0131_8013 - - - - - - - D0F0xE4_x0132_8013 - Reserved - Reserved - Reserved - D0F0xE4_x0133_8013 - Reserved - Reserved - Reserved - 226 42300 Rev 3.12 - July 14, 2015 Bits BKDG for AMD Family 15h Models 10h-1Fh Processors Description 31:21 Reserved. 20 TxclkSelDigBOverride. Read-write.1=Override TxclkDigB selection. 19:17 TxclkSelDigB. Read-write. Specifies the source of the dig B clock when TxclkSelDigBOverride=1. Bits Definition Bits Definition 0x0b Phy Clock A 10xb Phy Clock C 0x1b Phy Clock B 11xb Phy Clock D The selected clock with not function correctly if the divider logic is not enabled for the clock. The divider is enabled if (ClkDividerResetOverrideX | MasterPciePllX | D0F0xE4_x013[3:1]_804[3:0][OwnSlice]) =1. 16 TxclkSelDigAOverride. Read-write.1=Override TxclkDigA selection. 15:13 TxclkSelDigA. Read-write. Specifies the source of the dig A clock when TxclkSelDigAOverride=1. Bits Definition Bits Definition 0x0b Phy Clock A 10xb Phy Clock C 0x1b Phy Clock B 11xb Phy Clock D The selected clock with not function correctly if the divider logic is not enabled for the clock. The divider is enabled if (ClkDividerResetOverrideX | MasterPciePllX | D0F0xE4_x013[3:1]_804[3:0][OwnSlice]) =1. 12 TxclkSelPifDOverride. Read-write.1=Override TxclkPifD selection. 11 TxclkSelPifCOverride. Read-write. 1=Override TxclkPifC selection. 10 TxclkSelPifBOverride. Read-write. 1=Override TxclkPifB selection. 9 TxclkSelPifAOverride. Read-write. 1=Override TxclkPifA selection. 8 TxclkSelCoreOverride. Read-write. 1=Override TxclkCore selection. 7 ClkDividerResetOverrideD. Read-write. 1=Force clock divider D enabled. 6 ClkDividerResetOverrideC. Read-write. 1=Force clock divider C enabled. 5 ClkDividerResetOverrideB. Read-write. 1=Force clock divider B enabled. 4 ClkDividerResetOverrideA. Read-write. 1=Force clock divider A enabled. 3 MasterPciePllD. Read-write. 1=Pll D is the master source for all PCIe transmitter clock branches. 2 MasterPciePllC. Read-write. 1=Pll C is the master source for all PCIe transmitter clock branches. 1 MasterPciePllB. Read-write. 1=Pll B is the master source for all PCIe transmitter clock branches. 0 MasterPciePllA. Read-write. 1=Pll A is the master source for all PCIe transmitter clock branches. D0F0xE4_x013[3:0]_8014 Link Transmit Clock Gating Control 2 Reset: 0000_0000h. Table 117: Reserved field mappings for D0F0xE4_x013[3:0]_8014 Register Bits 27:26 19:18 15:14 11:10 5:4 D0F0xE4_x0130_8014 Reserved Reserved Reserved Reserved Reserved D0F0xE4_x0131_8014 - - - - - D0F0xE4_x0132_8014 Reserved Reserved Reserved Reserved Reserved D0F0xE4_x0133_8014 Reserved Reserved Reserved Reserved Reserved 227 42300 Rev 3.12 - July 14, 2015 Bits BKDG for AMD Family 15h Models 10h-1Fh Processors Description 31:28 Reserved. 27 DdiGateDigDEnable. Read-write. BIOS: IF (REG==D0F0xE4_x0130_8014) THEN 0 ELSE 1 ENDIF. 1=Enable gating of the dig d clock branches in DDI mode. 26 DdiGateDigCEnable. Read-write. BIOS: IF (REG==D0F0xE4_x0130_8014) THEN 0 ELSE 1 ENDIF. 1=Enable gating of the dig c clock branches in DDI mode. 25 DdiGateDigBEnable. Read-write. BIOS: IF (REG==D0F0xE4_x0130_8014) THEN 0 ELSE 1 ENDIF. 1=Enable gating of the dig b clock branches in DDI mode. 24 DdiGateDigAEnable. Read-write. BIOS: IF (REG==D0F0xE4_x0130_8014) THEN 0 ELSE 1 ENDIF. 1=Enable gating of the dig a clock branches in DDI mode. 23:21 Reserved. 20 TxclkPermGateOnlyWhenPllPwrDn. Read-write. BIOS: 1. 1=Gating of the permanent clock branch only occurs when the PLL is powered down. 19 PcieGatePifD2p5xEnable. Read-write. BIOS: 1. 1=Enable gating of the PIF D 2.5x clock branches in PCIe mode. 18 PcieGatePifC2p5xEnable. Read-write. BIOS: 1. 1=Enable gating of the PIF C 2.5x clock branches in PCIe mode. 17 PcieGatePifB2p5xEnable. Read-write. BIOS: 1. 1=Enable gating of the PIF B 2.5x clock branches in PCIe mode. 16 PcieGatePifA2p5xEnable. Read-write. BIOS: 1. 1=Enable gating of the PIF A 2.5x clock branches in PCIe mode. 15 PcieGatePifD1xEnable. Read-write. BIOS: 1. 1=Enable gating of the PIF D 1x clock branches in PCIe mode. 14 PcieGatePifC1xEnable. Read-write. BIOS: 1. 1=Enable gating of the PIF C 1x clock branches in PCIe mode. 13 PcieGatePifB1xEnable. Read-write. BIOS: 1. 1=Enable gating of the PIF B 1x clock branches in PCIe mode. 12 PcieGatePifA1xEnable. Read-write. BIOS: 1. 1=Enable gating of the PIF A 1x clock branches in PCIe mode. 11 DdiGatePifD2p5xEnable. Read-write. BIOS: IF (REG==D0F0xE4_x0130_8014) THEN 0 ELSE 1 ENDIF. 1=Enable gating of the PIF D 2.5x clock branches in DDI mode. 10 DdiGatePifC2p5xEnable. Read-write. BIOS: IF (REG==D0F0xE4_x0130_8014) THEN 0 ELSE 1 ENDIF. 1=Enable gating of the PIF C 2.5x clock branches in DDI mode. 9 DdiGatePifB2p5xEnable. Read-write. BIOS: IF (REG==D0F0xE4_x0130_8014) THEN 0 ELSE 1 ENDIF. 1=Enable gating of the PIF B 2.5x clock branches in DDI mode. 8 DdiGatePifA2p5xEnable. Read-write. BIOS: IF (REG==D0F0xE4_x0130_8014) THEN 0 ELSE 1 ENDIF. 1=Enable gating of the PIF A 2.5x clock branches in DDI mode. 7:6 Reserved. 5 DdiGatePifD1xEnable. Read-write. BIOS: IF (REG==D0F0xE4_x0130_8014) THEN 0 ELSE 1 ENDIF. 1=Enable gating of the PIF D 1x clock branches in DDI mode. 4 DdiGatePifC1xEnable. Read-write. BIOS: IF (REG==D0F0xE4_x0130_8014) THEN 0 ELSE 1 ENDIF. 1=Enable gating of the PIF C 1x clock branches in DDI mode. 228 42300 Rev 3.12 - July 14, 2015 BKDG for AMD Family 15h Models 10h-1Fh Processors 3 DdiGatePifB1xEnable. Read-write. BIOS: IF (REG==D0F0xE4_x0130_8014) THEN 0 ELSE 1 ENDIF. 1=Enable gating of the PIF B 1x clock branches in DDI mode. 2 DdiGatePifA1xEnable. Read-write. BIOS: IF (REG==D0F0xE4_x0130_8014) THEN 0 ELSE 1 ENDIF. 1=Enable gating of the PIF A 1x clock branches in DDI mode. 1 TxclkPrbsGateEnable. Read-write. BIOS: 1. 1=Enable gating of the PRBS clock branch. 0 TxclkPermGateEnable. Read-write. BIOS: 1. 1=Enable gating of the permanent clock branch. D0F0xE4_x013[3:0]_8015 IO Link IOC Control Bits Description 31 RefclkBphyGateEnable. Read-write. Reset:0. BIOS: 1. 1=Enable gating of REFCLK_BPHY. 30 Reserved. 29:24 RefclkBphyGateLatency. Read-write. Reset:3Fh. BIOS: 0. Specifies the number of clocks to wait before turning of f REFCLK_BPHY. 23 RefclkRegsGateEnable. Read-write. Reset:0. BIOS: 1. 1=Enable gating of REFCLK_REGS. 22 Reserved. 21:16 RefclkRegsGateLatency. Read-write. Reset:3Fh. Specifies the number of clocks to wait before turning of f REFCLK_REGS. 15:0 Reserved. D0F0xE4_x013[3:0]_8016 Link Clock Switching Control Reset: 003F_001Fh. Table 118: Reserved field mappings for D0F0xE4_x013[3:0]_8016 Register Bits 11:10 9 D0F0xE4_x0130_8016 Reserved Reserved D0F0xE4_x0131_8016 - - D0F0xE4_x0132_8016 Reserved - D0F0xE4_x0133_8016 Reserved - Bits Description 31:24 Reserved. 23 LclkDynGateEnable. Read-write. BIOS: 1. 1=Enable LCLK_DYN clock gating. 22 LclkGateFree. Read-write. BIOS: 1. 1=LCLK gating is controlled independent of TXCLK gating. 21:16 LclkDynGateLatency. Read-write. Specifies the number of clocks to wait before turning off LCLK_DYN. 15:6 Reserved. 5:0 CalibAckLatency. Read-write. BIOS: IF (REG== D0F0xE4_x0130_8016) THEN 0 ELSE 1Fh. ENDIF. Specifies the number of clocks after calibration is complete before the acknowledge signal is asserted. 229 42300 Rev 3.12 - July 14, 2015 BKDG for AMD Family 15h Models 10h-1Fh Processors D0F0xE4_x013[3:0]_8020 Lane Control Reset: 0000_0000h. Bits Description 31:4 Reserved. 3 2:0 PrbsPcieLbSelect. Read-write. 1=Loopback data selected. 0=IO/PRBS data selected. Reserved. D0F0xE4_x013[3:0]_8021 Transmitter Lane Mux Reset: 7654_3210h. Bits Description 31:28 Lanes1514. Read-write. Specifies the controller lanes that are mapped to PIF TX lanes 15 and 14. See: Lanes10. 27:24 Lanes1312. Read-write. Specifies the controller lanes that are mapped to PIF TX lanes 13 and 12. See: Lanes10. 23:20 Lanes1110. Read-write. Specifies the controller lanes that are mapped to PIF TX lanes 11 and 10. See: Lanes10. 19:16 Lanes98. Read-write. Specifies the controller lanes that are mapped to PIF TX lanes 9 and 8. See: Lanes10. 15:12 Lanes76. Read-write. Specifies the controller lanes that are mapped to PIF TX lanes 7 and 6. See: Lanes10. 11:8 Lanes54. Read-write. Specifies the controller lanes that are mapped to PIF TX lanes 5 and 4. See: Lanes10. 7:4 Lanes32. Read-write. Specifies the controller lanes that are mapped to PIF TX lanes 3 and 2. See: Lanes10. 3:0 Lanes10. Read-write. Specifies the controller lanes that are mapped to PIF TX lanes 1 and 0. Bits Definition Bits Definition 0h Controller lanes 1 and 0 5h Controller lanes 11 and 10. 1h Controller lanes 3 and 2. 6h Controller lanes 13 and 12. 2h Controller lanes 5 and 4. 7h Controller lanes 15 and 14. 3h Controller lanes 7 and 6. Fh-8h Reserved. 4h Controller lanes 9 and 8. D0F0xE4_x013[3:0]_8022 Receiver Lane Mux Reset: 7654_3210h. Bits Description 31:28 Lanes1514. Read-write. Specifies the PIF RX lanes that are mapped to controller lanes 15 and 14. See: Lanes10. 27:24 Lanes1312. Read-write. Specifies the PIF RX lanes that are mapped to controller lanes 13 and 12. See: Lanes10. 230 42300 Rev 3.12 - July 14, 2015 BKDG for AMD Family 15h Models 10h-1Fh Processors 23:20 Lanes1110. Read-write. Specifies the PIF RX lanes that are mapped to controller lanes 11 and 10. See: Lanes10. 19:16 Lanes98. Read-write. Specifies the PIF RX lanes that are mapped to controller lanes 9 and 8. See: Lanes10. 15:12 Lanes76. Read-write. Specifies the PIF RX lanes that are mapped to controller lanes 7 and 6. See: Lanes10. 11:8 Lanes54. Read-write. Specifies the PIF RX lanes that are mapped to controller lanes 5 and 4. See: Lanes10. 7:4 Lanes32. Read-write. Specifies the PIF RX lanes that are mapped to controller lanes 3 and 2. See: Lanes10. 3:0 Lanes10. Read-write. Specifies the PIF RX lanes that are mapped to controller lanes 1 and 0. Bits Definition Bits Definition 0h PIF RX lanes 1 and 0 5h PIF RX lanes 11 and 10. 1h PIF RX lanes 3 and 2. 6h PIF RX lanes 13 and 12. 2h PIF RX lanes 5 and 4. 7h PIF RX lanes 15 and 14. 3h PIF RX lanes 7 and 6. Fh-8h Reserved. 4h PIF RX lanes 9 and 8. D0F0xE4_x013[3:0]_8023 Lane Enable Reset: 0000_FFFFh. Bits Description 31:16 Reserved. 15:0 LaneEnable. Read-write. 1=Lane enabled for transmit. Definition Bit [15:0] Lane <BIT> enable D0F0xE4_x013[3:0]_8025 Lane Mux Power Sequence Control Reset: 1F1F_1F1Fh. Bits Description 31:30 Reserved. 29 LMLinkSpeed3. Read-write. 1=5GHz. 0=2.5GHz. 28:27 LMRxPhyCmd3. Read-write. Specifies the receiver state for lanes 15-12. See: LMRxPhyCmd0. 26:24 LMTxPhyCmd3. Read-write. Specifies the transmitter state for lanes 15-12. See: LMTxPhyCmd0. 23:22 Reserved. 21 LMLinkSpeed2. Read-write. 1=5GHz. 0=2.5GHz. 20:19 LMRxPhyCmd2. Read-write. Specifies the receiver state for lanes 11-8. See: LMRxPhyCmd0. 18:16 LMTxPhyCmd2. Read-write. Specifies the transmitter state for lanes 11-8. See: LMTxPhyCmd0. 15:14 Reserved. 13 LMLinkSpeed1. Read-write. 1=5GHz. 0=2.5GHz. 12:11 LMRxPhyCmd1. Read-write. Specifies the receiver state for lanes 7-4. See: LMRxPhyCmd0. 10:8 LMTxPhyCmd1. Read-write. Specifies the transmitter state for lanes 7-4. See: LMTxPhyCmd0. 231 42300 Rev 3.12 - July 14, 2015 7:6 5 BKDG for AMD Family 15h Models 10h-1Fh Processors Reserved. LMLinkSpeed0. Read-write. 1=5GHz. 0=2.5GHz. 4:3 LMRxPhyCmd0. Read-write. Specifies the receiver state for lanes 3-0. Definition Bits Definition Bits 00b On. 10b Standby 2 (L1). 01b Standby 1 (L0s). 11b Off. 2:0 LMTxPhyCmd0. Read-write. Specifies the transmitter state for lanes 3-0. Bits Definition Bits Definition 000b On. 100b Receiver detect. 001b Standby 1 (L0s). 101b Reserved. 010b Standby 2 (L1). 110b Reserved. 011b Reserved. 111b Off. D0F0xE4_x013[1:0]_8031 Lane Counter Status Reset: 0000_0000h. Bits Description 31:17 Reserved. 16 LnCntValid. Read-write. 1=LnCntBandwidth contains a valid bandwidth measurement. 15:10 Reserved. 9:0 LnCntBandwidth. Read-write. Estimated lane bandwidth in 10 MB/s units. D0F0xE4_x013[3:0]_8060 Soft Reset Command 0 Cold reset: 0000_0000h. Bits Description 31:18 Reserved. 17 Bif0CalibrationReset. Read-write. 1=The BIF 0 calibration block reset is asserted. 16 Bif0GlobalReset. Read-write. 1=The BIF 0 global reset is asserted. 15:3 Reserved. 2 ResetComplete. Read-only. 1=Reset cycle is complete. 1 Reserved. 0 Reconfigure. Read-write; Cleared-when-done. 1=Trigger atomic reconfiguration if D0F0xE4_x013[3:0]_8062[ReconfigureEn]=1. D0F0xE4_x013[3:0]_8062 Soft Reset Control 0 Cold reset: 0001_0800h. Bits Description 31:12 Reserved. 11 ConfigXferMode. Read-write. 1=PCIe core strap settings take effect immediately. 0=PCIe core strap settings take effect when the PCIe core is reset. 232 42300 Rev 3.12 - July 14, 2015 BKDG for AMD Family 15h Models 10h-1Fh Processors 10 BlockOnIdle. Read-write. 1=The PCIe core must be idle before hardware initiates a reconfiguration. 0=The PCIe core does not have to be idle before hardware initiates a reconfiguration. 9:5 Reserved. 4:2 ResetPeriod. Read-write. BIOS:0. Specifies the amount of time that resets are asserted during a reconfiguration. 1 Reserved. 0 ReconfigureEn. Read-write. 1=Atomic reconfiguration enabled. D0F0xE4_x013[3:0]_8063 Soft Reset Control 1 Cold reset: 0000_0C00h. Bits Description 31:15 Reserved. 14 ResetSrbmDcEn. Read-write. 1=The reset to the dc register interface block is asserted during an atomic reset or reconfiguration. 13 ResetSrbmGfxEn. Read-write. 1=The reset to the gfx register interface block is asserted during an atomic reset or reconfiguration. 12 ResetSrbmNbEn. Read-write. 1=The reset to the nb register interface block is asserted during an atomic reset or reconfiguration. 11:6 Reserved. 5 ResetSrbm1En. Read-write. 1=The reset to the PCIe® core register interface block is applied during an atomic reset or reconfiguration. 4 ResetSrbm0En. Read-write. 1=The reset to the wrapper register interface block is applied during an atomic reset or reconfiguration. 3:0 Reserved. D0F0xE4_x0130_80F0 BIOS Timer Reset: 0000_0000h. Bits Description 31:0 MicroSeconds. Read-write; updated-by-hardware. This field increments once every microsecond when the timer is enabled. The counter will roll over and continue counting when it reaches its FFFF_FFFFh. A write to this register causes the counter to reset and begin counting from the value written. 233 42300 Rev 3.12 - July 14, 2015 BKDG for AMD Family 15h Models 10h-1Fh Processors D0F0xE4_x0130_80F1 BIOS Timer Control Reset: 0000_0064h. Bits Description 31:8 Reserved. 7:0 ClockRate. Read-write. Specifies the frequency of the reference clock in 1 MHz increments. Bits Definition 00h Timer disabled FFh-01h <ClockRate> MHz D0F0xE4_x013[3:1]_804[3:0] DDI Slice Reset: 0000_0000h. Table 119: Index address mapping for D0F0xE4_x013[3:1]_804[3:0] D0F0xE0[31:16] Bits D0F0xE0[15:0] 8040h 8041h 8042h 8043h 0131h Gfx Lanes 0-3 Gfx Lanes 4-7 Gfx Lanes 8-11 Gfx Lanes 12-15 0132h DDI Lanes 0-3 DDI Lanes 4-7 DDI Lanes 8-11 DDI Lanes 12-15 0133h DDI2 Lanes 0-3 DDI2 Lanes 4-7 DDI2 Lanes 8-11 DDI2 Lanes 12-15 Description 31:1 Reserved. 0 OwnSlice. Read-write. 1=DDI asserts control over the PCIe lanes specified in Table 119. D0F0xE4_x013[3:1]_804[E:8] DDI Dig Reset: 0000_0700h. Table 120: Register Stream mappings for D0F0xE4_x013[3:1]_804[E:8] D0F0xE4_x013[3:1]_804[E:8] D0F0xE4_x013[3:1]_8048 D0F0xE4_x013[3:1]_8049 D0F0xE4_x013[3:1]_804A D0F0xE4_x013[3:1]_804B D0F0xE4_x013[3:1]_804C D0F0xE4_x013[3:1]_804D D0F0xE4_x013[3:1]_804E Bits Function Stream A Stream B Stream C Stream D Stream E Stream F Stream G Description 31:26 Reserved. 25 CntDig. Read-write. 1=Software asserts control over Dig TxPhyCmd and LinkSpeed. 24 CntPhy. Read-write. 1=Software asserts control over the phy state machine. 234 42300 Rev 3.12 - July 14, 2015 BKDG for AMD Family 15h Models 10h-1Fh Processors 23 Reserved. 22 Nxt_Lnspd. Read-write. 1=Set the value for the dig link speed in case of an override. 21:19 Nxt_phycmd. Read-write. 1=Set the value for the dig Tx phy command in case of an override. 18:16 Nxt_State. Read-write. This specifies the next state for the DDI FSM. 15:11 Reserved. 10:8 Dig_pwrdn_value. Read-write. Specifies the phy power state for links associated with dig streams that are not enabled. Bits Definition Bits Definition 000b On. 100b Receiver detect. 001b Standby 1 (L0s). 101b Reserved. 010b Standby 2 (L1). 110b Reserved. 011b Reserved. 111b Off. 7 Reserved. 6 Hbr2Support. Read-write. 1=HBR is supported. 5 Reserved. 4 Hbr2Active. Read-write. 1=If (Hbr2Support==1) THEN HBR2 will be enabled. 3 Reserved. 2 PwrDnCpl. Read-only; updated-by-hardware. 1=PHY state machine is in the powered up state. 1 Reserved. 0 PwrDnCpl. Read-only; updated-by-hardware. 1=PHY state machine is in the power off state. 235 42300 Rev 3.12 - July 14, 2015 3.4 BKDG for AMD Family 15h Models 10h-1Fh Processors Device 0 Function 2 (IOMMU) Configuration Registers See 3.1 [Register Descriptions and Mnemonics]. See 2.7 [Configuration Space]. See 2.12.1 [IOMMU Configuration Space]. D0F2x00 Device/Vendor ID Bits Description 31:16 DeviceId. Read-only. Reset: 1419h. 15:0 VendorId. Read-only. Reset: 1022h. D0F2x04 Status/Command Bits Description 31 ParityErrorDetected. Read; write-1-to-clear. Reset: 0. 30 SignaledSystemError. Read-only. Reset: 0. 29 ReceivedMasterAbort. Read; write-1-to-clear. Reset: 0. 28 ReceivedTargetAbort. Read; write-1-to-clear. Reset: 0. 27 SignalTargetAbort. Read-only. Reset: 0. 26:25 Reserved. 24 MasterDataError. Read; write-1-to-clear. Reset: 0. 23:21 Reserved. 20 CapList. Read-only. Reset: 1. 1=Capability list supported. 19 IntStatus. Read-only. Reset: 0. 1=INTx message pending. 18:11 Reserved. 10 InterruptDis. Read-write. Reset: 0. 1=INTx interrupt message generation disabled. 9 Reserved. 8 SerrEn. Read-only. Reset: 0. 1=Enables reporting of non-fatal and fatal errors detected. 7 Reserved. 6 ParityErrorEn. Read-write. Reset: 0. 1=Enables setting of ParityErrorDetected status bit. 5:3 Reserved. 2 BusMasterEn. Read-write. Reset: 0. 1=Enables DMA request generation. 1 MemAccessEn. Read-only. Reset: 0. 0 IoAccessEn. Read-only. Reset: 0. 236 42300 Rev 3.12 - July 14, 2015 BKDG for AMD Family 15h Models 10h-1Fh Processors D0F2x08 Class Code/Revision ID Reset: 0806_00xxh. Bits Description 31:8 ClassCode: class code. Read-only. Provides the host bridge class code as defined in the PCI specification. 7:0 RevID: revision ID. Read-only. D0F2x0C Header Type Reset: 0080_0000h. Bits Description 31:24 BIST. Read-only. 23:16 HeaderTypeReg. Read-only. 80h=Type 0 multi-function device. 15:8 LatencyTimer. Read-write. 7:0 CacheLineSize. Read-only. D0F2x2C Subsystem and Subvendor ID Bits Description 31:16 SubsystemId. Read-only. Reset: 0. 15:0 SubsystemVendorId. Read-only. Reset: 0. D0F2x34 Capabilities Pointer Bits Description 31:8 Reserved. 7:0 CapPtr. Read-only. Reset: 40h. D0F2x3C Interrupt Line Bits Description 31:16 Reserved. 237 42300 Rev 3.12 - July 14, 2015 BKDG for AMD Family 15h Models 10h-1Fh Processors 15:8 InterruptPin. Read-only. Reset: 01h. This field indicates the INTx line used to generate legacy interrupts. Bits Description 00h Reserved. 01h INTA. 02h INTB. 03h INTC. 04h INTD. FFh-05h Reserved. 7:0 InterruptLine. Read-write. Reset: 0. This field is read/write for software compatibility. It controls no hardware. D0F2x40 IOMMU Capability Bits Description 31:28 Reserved. 27 IommuEfrSup. Read-only. Reset: 1. 1=Indicates IOMMUx30 [Extended Feature Low] is supported. 0=IOMMUx30 is reserved. 26 IommuNpCache. Read-only. Reset: 0. 1=Indicates that the IOMMU caches page table entries that are marked as not present. When this bit is set, software must issue an invalidate after any change to a PDE or PTE. 0=Indicates that the IOMMU caches only page table entries that are marked as present. When this bit is clear, software must issue an invalidate after any change to a PDE or PTE marked present before the change. 25 IommuHtTunnelSup. Read-only. Reset: 0. 24 IommuIoTlbsup. Read-only. Reset: 1. Indicates support for remote IOTLBs. 23:19 IommuCapRev. Read-only. Reset: 1. Specifies the IOMMU interface revision. 18:16 IommuCapType. Read-only. Reset: 3h. Specifies the layout of the Capability Block as an IOMMU capability block. 15:8 IommuCapPtr. Read-only. Reset: 54h. Indicates the location of the next capability block. 7:0 IommuCapId. Read-only. Reset: Fh. Indicates a Secure Device capability block. D0F2x44 IOMMU Base Address Low Bits Description 31:14 IommuBaseAddr[31:14]: iommu base address bits[31:14]. IF (D0F2x44[IommuEnable]) THEN Read-only. ELSE Read-write. ENDIF. Reset: 0. IommuBaseAddr[63:14] = {D0F2x48[IommuBaseAddr[63:32]], IommuBaseAddr[31:14]}. IommuBaseAddr[63:14] specifies the base address of the IOMMU memory mapped control registers. In order to use the IOMMU event counters, IommuBaseAddr[18:14] must be 0_0000b. 13:1 0 Reserved. IommuEnable. Read; write-1-only. Reset: 0. 1=IOMMU accepts memory accesses to the address specified in IommuBaseAddr[63:14]. When this bit is set, all IOMMU RW capability registers in PCI configuration space are locked. 238 42300 Rev 3.12 - July 14, 2015 BKDG for AMD Family 15h Models 10h-1Fh Processors D0F2x48 IOMMU Base Address High Bits Description 31:0 IommuBaseAddr[63:32]: iommu base address bits[63:32]. See: D0F2x44[IommuBaseAddr[31:14]]. D0F2x4C IOMMU Range Bits Description 31:24 IommuLastDevice. Read-only. Reset: 0. Indicates device and function number of the last integrated device associated with the IOMMU. 23:16 IommuFirstDevice. Read-only. Reset: 0. Indicates device and function number of the first integrated device associated with the IOMMU. 15:8 IommuBusNumber. Read-only. Reset: 0. Indicates the bus number that IommuLastDevice and IommuFirstDevice reside on. 7 IommuRngValid. Read-only. Reset: 0. 1=The IommuBusNumber, IommuFirstDevice, and IommuLastDevice fields are valid. Although the register contents are valid, software is encouraged to use I/O topology information.. 0=Software must use I/O topology information. 6:5 Reserved. 4:0 IommuUnitId. Read-only. Reset: 0. D0F2x50 IOMMU Miscellaneous Information Register Bits Description 31:27 IommuMsiNumPpr. Read-only. Reset: 0. This field must indicate which MSI vector is used for the interrupt message generated by the IOMMU for the peripheral page service request log when IOMMUx30[PprSup]=1. This field must be 0 when IOMMUx30[PprSup]=0. For MSI there can be only one IOMMU so this field must be 0. This interrupt is not remapped by the IOMMU. 26:23 Reserved. 22 IommuHtAtsResv. IF (D0F2x44[IommuEnable]) THEN Read-only. ELSE Read-write. ENDIF. Reset: 0. 1=The HyperTransport Address Translation address range for ATS responses is reserved and cannot be translated by the IOMMU. 0=The Address Translation address range can be translated by the IOMMU. 239 42300 Rev 3.12 - July 14, 2015 BKDG for AMD Family 15h Models 10h-1Fh Processors 21:15 IommuVaSize. Read-only. Reset: 40h. This field must indicate the size of the maximum virtual address processed by the IOMMU. The value is the (unsigned) binary log of the maximum address size. Description Bits 19h-00h Reserved. 20h 32 bits. 27h-21h Reserved. 28h 40 bits. 2Fh-29h Reserved. 30h 48 bits. 3Fh-31h Reserved. 40h 64 bits. 7Fh-41h Reserved. 14:8 IommuPaSize. Read-only. Reset: 28h. This field must indicate the size of the maximum physical address generated by the IOMMU. The value is the (unsigned) binary log of the maximum address size. Description Bits 27h-00h Reserved. 28h 40 bits. 7Fh-29h Reserved. 7:5 IommuGvaSize. Read-only. Reset: 010b. Indicates the size of the maximum guest virtual address processed by the IOMMU. 010b=48 bits. All other values are reserved. 4:0 IommuMsiNum. Read-only. Reset: 0. Indicates the MSI vector used for interrupt messages generated by the IOMMU. D0F2x54 IOMMU MSI Capability Register Bits Description 31:24 Reserved. 23 Msi64En. Read-only. Reset: 1. 1=64-bit MSI addressing is supported 22:20 MsiMultMessEn. Read-only. Reset: 0. Specifies the number of MSI messages assigned to this function. 19:17 MsiMultMessCap. Read-only. Reset: 0. Specifies the number of MSI messages requested by this function. 16 MsiEn. Read-write. Reset: 0. 1=Enables MSI for this function and causes legacy interrupts to be disabled. 15:8 MsiCapPtr. Read-only. Reset: 64h. Pointer to the next capability register offset. 7:0 MsiCapId. Read-only. Reset: 5h. Indicates that this is the MSI capability. D0F2x58 IOMMU MSI Address Low Bits Description 31:2 MsiAddr[31:2]. Read-write. Reset: 0. This register specifies the lower address bits used to issue MSI messages. 1:0 Reserved. 240 42300 Rev 3.12 - July 14, 2015 BKDG for AMD Family 15h Models 10h-1Fh Processors D0F2x5C IOMMU MSI Address High Bits Description 31:0 MsiAddr[63:32]. Read-write. Reset: 0. This register specifies the upper address bits used to issue MSI messages. D0F2x60 IOMMU MSI Data Bits Description 31:16 Reserved. 15:0 MsiData. Read-write. Reset: 0. This register specifies the data issued with MSI messages. D0F2x64 IOMMU MSI Mapping Capability Bits Description 31:27 MsiMapCapType. Read-only. Reset: 15h. Indicates the MSI Mapping Capability. 26:18 Reserved. 17 MsiMapFixd. Read-only. Reset: 1. Always set to 1 to indicate that this device only maps MSI interrupts with address 0xFEEx_xxxx onto Hypertransport interrupts and that the mapping range is not programmable 16 MsiMapEn. Read-only. Reset: 1. Always set to 1 to indicate that the MSI Mapping Capability is always enabled 15:8 MsiMapCapPtr. Read-only. Reset: 0. Points to the next capability list item 7:0 MsiMapCapId. Read-only. Reset: 8h. Indicates a Hypertransport capability list item D0F2x6C IOMMU Control Bits Description 31:10 Reserved. 9 EfrSupW. Read-write. Reset: 1. This field sets the value of D0F2x40[EfrSup]. 8 IoTlbsupW. Read-write. Reset: 1. This field sets the value of D0F2x40[IommuIoTlbsup]. 7:4 MinorRevIdW. Read-write. Reset: 0h. This field sets the value of D0F2x08[RevID[3:0]]. 3 2:0 Reserved. InterruptPinW. Read-write. Reset: 001b. This field sets the value of D0F2x3C[InterruptPin]. D0F2x70 IOMMU MMIO Control Low Bits Description 31:12 Reserved. 241 42300 Rev 3.12 - July 14, 2015 BKDG for AMD Family 15h Models 10h-1Fh Processors 11:10 HatsW. Read-write. Reset: 2h. This field sets the value of IOMMUx30[HATS]. 9 PcSupW. Read-write. Reset: 1. BIOS: 0. This field sets the value of IOMMUx30[PcSup]. This bit should not be set. 8 Reserved. 7 Reserved. 6 IaSupW. Read-write. Reset: 1. This field sets the value of IOMMUx30[IaSup]. 5 Reserved. 4 GtSupW. Read-write. Reset: 1. This field sets the value of IOMMUx30[GtSup]. 3 NxSupW. Read-write. Reset: 0. This field sets the value of IOMMUx30[NxSup]. 2 Reserved. 1 PprSupW. Read-write. Reset: 1. This field sets the value of IOMMUx30[PprSup]. 0 PrefSupW. Read-write. Reset: 1. This field sets the value of IOMMUx30[PrefSup]. D0F2x74 IOMMU MMIO Control High Bits Description 31:4 Reserved. 3:0 PasMaxW. Read-write. Reset: 8h. This field sets the value of IOMMUx34[PasMax]. D0F2x78 IOMMU Range Control The fields in this register set the values of the corresponding fields in D0F2x4C. Bits Description 31:24 LastDeviceW. Read-write. Reset: 0. 23:16 FirstDeviceW. Read-write. Reset: 0. 15:8 7 6:0 BusNumberW. Read-write. Reset: 0. RngValidW. Read-write. Reset: 0. Reserved. D0F2xF0 IOMMU L2 Config Index The index/data pair registers D0F2xF0 and D0F2xF4 is used to access the registers D0F2xF4_x[FF:00]. To read or write to one of these register, the address is written first into the address register D0F2xF0 and then the data are read or written by read or write the data register D0F2xF4. See 2.12.1 [IOMMU Configuration Space]. Bits Description 31:9 Reserved. 8 7:0 L2cfgWrEn. Read-write. Reset: 0. 1=Enable writes to D0F2xF4. L2cfgIndex. Read-write. Reset: 0. 242 42300 Rev 3.12 - July 14, 2015 BKDG for AMD Family 15h Models 10h-1Fh Processors D0F2xF4 IOMMU L2 Config Data See D0F2xF0. D0F2xF4_x00 L2_PERF_CNTL_0 Bits Description 31:24 L2PerfCountUpper1. Read-only. Reset: 0. Upper 8 bits of IOMMU L2 performance counter 1 23:16 L2PerfCountUpper0. Read-only. Reset: 0. Upper 8 bits of IOMMU L2 performance counter 0 15:8 L2PerfEvent1. Read-write. Reset: 0. Selects the IOMMU L2 performance counter event for counter 1 7:0 L2PerfEvent0. Read-write. Reset: 0. Selects the IOMMU L2 performance counter event for counter 0 D0F2xF4_x01 L2_PERF_COUNT_0 Bits Description 31:0 L2PerfCount0. Read-only. Reset: 0. Lower 32 bits of IOMMU L2 performance counter 0 D0F2xF4_x02 L2_PERF_COUNT_1 Bits Description 31:0 L2PerfCount1. Read-only. Reset: 0. Lower 32 bits of IOMMU L2 performance counter 1 D0F2xF4_x03 L2_PERF_CNTL_1 Bits Description 31:24 L2PerfCountUpper3. Read-only. Reset: 0. Upper 8 bits of IOMMU L2 performance counter 3 23:16 L2PerfCountUpper2. Read-only. Reset: 0. Upper 8 bits of IOMMU L2 performance counter 2 15:8 L2PerfEvent3. Read-write. Reset: 0. Selects the IOMMU L2 performance counter event for counter 3 7:0 L2PerfEvent2. Read-write. Reset: 0. Selects the IOMMU L2 performance counter event for counter 2 D0F2xF4_x04 L2_PERF_COUNT_2 Bits Description 31:0 L2PerfCount2. Read-only. Reset: 0. Lower 32 bits of IOMMU L2 performance counter 2 243 42300 Rev 3.12 - July 14, 2015 BKDG for AMD Family 15h Models 10h-1Fh Processors D0F2xF4_x05 L2_PERF_COUNT_3 Bits Description 31:0 L2PerfCount3. Read-only. Reset: 0. Lower 32 bits of IOMMU L2 performance counter 3 D0F2xF4_x06 L2_DEBUG_0 Bits Description 31:0 L2DEBUG0. Read-write. Reset: 0. Reserved for to control ECOs D0F2xF4_x07 L2_DEBUG_1 Bits Description 31:0 L2DEBUG1. Read-write. Reset: 0. Reserved for to control ECOs D0F2xF4_x08 L2_STATUS_0 Bits Description 31:0 L2STATUS0. Read-only. Reset: 0. Internal IOMMU L2A status D0F2xF4_x0C L2_CONTROL_0 Bits Description 31:24 IFifoClientPriority. Read-write. Reset: 0. Each bit of this register controls whether the corresponding L1 client is arbitrated as high priority or not. Not all implementations will use all of the priority bits due to a lower number of clients versus the register width 23:20 IFifoBurstLength. Read-write. Reset: 1. Sets the burst length when arbitrating between clients coming into the L2 19 Reserved. 18 FLTCMBPriority. Read-write. Reset: 0. 0=Round-robin arbitration between cache responses and table-walker responses at the fault combiner. . 1=Table-walker responses always win arbitration at the fault combiner. 17:12 IFifoCMBCredits. Read-write. Reset: 4h. Controls the initial number of credits for the ififo to fault/CMB interface. Credits are loaded whenever the register value changes. This register may only be programmed when IOMMU is not enabled to preserve correct operation. 11 SIDEPTEOnAddrTransExcl. Read-write. Reset: 0. 0=Caches return DTE to L1 on an address translation exclusion range access. 1=Caches return PTE to L1 on an address translation exclusion range access 10 SIDEPTEOnUntransExcl. Read-write. Reset: 0. 0=Caches return DTE to L1 on an untranslated exclusion range access . 1=Caches return PTE to L1 on an untranslated exclusion range access 244 42300 Rev 3.12 - July 14, 2015 BKDG for AMD Family 15h Models 10h-1Fh Processors 9:4 IFifoTWCredits. Read-write. Reset: 4h. Controls the initial number of credits for the ififo to TW interface. Credits are loaded whenever the register changes value. This register may only be programmed when IOMMU is not enabled to preserve correct operation. 3 DTCHitVZeroOrIVZero. Read-write. Reset: 0. 0=A DTE is refetched if a DTE with V=0 for a memory request or IV=1 for an interrupt request is hit in the DTC . 1=A DTE is not refetched if a DTE with V=0 for a memory request or IV=1 for an interrupt request is hit in the DTC. This DTE is used 2 AllowL1CacheATSRsp. Read-write. Reset: 0. 0=L2 does not allow L1 to cache responses to ATS address translation requests . 1=L2 allows L1 to cache responses to ATS address translation requests 1 AllowL1CacheVZero. Read-write. Reset: 0. 0=L2 does not allow L1 to cache DTEs where V=0 . 1=L2 allows L1 to cache DTEs where V=1. L1 stores IR and IW as if they are both set to 1 0 PTCAddrTransReqCheck. Read-write. Reset: 0. 0=Address translation requests do not check the PTC . 1=Address translation requests check the PTC D0F2xF4_x0D L2_CONTROL_1 Bits Description 31:24 PerfThreshold. Read-write. Reset: 0. Fifo threshold level used to calculate certain performance counter values. 23:17 Reserved. 16 SeqInvBurstLimitEn. Read-write. Reset: 1. Enable stalling L2 requests to allow invalidation cycles to make forward progress based upon SeqInvBurstLimitInv and SeqInvBurstLimitL2Req 15:8 SeqInvBurstLimitL2Req. Read-write. Reset: 8h. Sets the number of consecutive IOMMU L2 requests to perform when doing sequential invalidation. Regular L2 and invalidation requests will alternate access to the main L2 caches based upon SeqInvBurstLimitInv and SeqInvBurstLimitL2Req. 7:0 SeqInvBurstLimitInv. Read-write. Reset: 8h. Sets the number of consecutive invalidation requests to perform when doing sequential invalidation. Regular L2 and invalidation requests will alternate access to the main L2 caches based upon SeqInvBurstLimitInv and SeqInvBurstLimitL2Req. D0F2xF4_x10 L2_DTC_CONTROL Bits Description 31:28 DTCEntries. Read-only. Reset: 0. The number of entries in the DTC is indicated as 2^DTCEntries 27:24 Reserved. 23:16 DTCWays. Read-only. Reset: 0. Indicates the number of ways in the DTC 15 DTCParitySupport. Read-only. Reset: 0. 0=The DTC does not support parity protection . 1=The DTC supports parity protection 14 Reserved. 13 DTCBypass. Read-write. Reset: 0. When set, all requests bypass the DTC. 12:11 Reserved. 245 42300 Rev 3.12 - July 14, 2015 BKDG for AMD Family 15h Models 10h-1Fh Processors 10 DTCSoftInvalidate. Read-write. Reset: 0. Software may write this register to 1 to invalidate all entries in the DTC 9:8 DTCInvalidationSel. Read-write. Reset: 0. BIOS: 10b. Selects the DTC invalidation algorithm. Bits Description 00b Invalidate the entire DTC. 01b Fast imprecise invalidation. 10b Sequential precise invalidation. 11b Partial sequential precise invalidation. 7:5 Reserved. 4 DTCParityEn. Read-write. Reset: 0. Enable parity protection of the DTC 3 DTCLRUUpdatePri. Read-write. Reset: 0. 0=Reads update replacement state bits when there is a simultaneous read and write to the same DTC index . 1=Writes update replacement state bits when there is a simultaneous read and write to the same DTC index 2 Reserved. 1:0 DTCReplacementSel. Read-write. Reset: 1. Selects the DTC replacement algorithm. Implementation may not support all replacement algorithms D0F2xF4_x11 L2_DTC_HASH_CONTROL Bits Description 31:16 DtcAddressMask. Read-write. Reset: FFFFh. BIOS: 0. This field is a bit-wise AND mask that selects which bits from the untranslated interrupt {MT[2:0],Vector} are used to index into the DTC. 15:11 Reserved. 10 DtcAltHashEn. Read-write. Reset: 0. BIOS: 1. Enable alternative algorithm for generating hash index into the DTC. 9 Reserved. 8:5 DTCBusBits. Read-write. Reset: 3h. Set the number of bus bits to use when using ReqID to form the DTC address. The following equation must be satisified. Func_bits + Dev_Bits + Bus_Bits &lt;= log2(DTC entries / DTC associativity). 4:2 DTCDevBits. Read-write. Reset: 0. Set the number of device bits to use when using ReqID to form the DTC address. 1:0 DTCFuncBits. Read-write. Reset: 2h. Set the number of function bits to use when using ReqID to form the DTC address. D0F2xF4_x12 L2_DTC_WAY_CONTROL Bits Description 31:16 DTCWayAccessDisable. Read-write. Reset: 0. 15:0 DTCWayDisable. Read-write. Reset: 0. Each bit in this register disables a way in the DTC when set to 1. An implementation may have less than 32 ways. The entire cache may be disabled by setting the DTCWays lower bits of this register. 246 42300 Rev 3.12 - July 14, 2015 BKDG for AMD Family 15h Models 10h-1Fh Processors D0F2xF4_x14 L2_ITC_CONTROL Bits Description 31:28 ITCEntries. Read-only. Reset: 0. The number of entries in the ITC is indicated as 2^ITCEntries 27:24 Reserved. 23:16 ITCWays. Read-only. Reset: 0. Indicates the number of ways in the ITC 15 ITCParitySupport. Read-only. Reset: 0. 0=The ITC does not support parity protection . 1=The ITC supports parity protection 14 Reserved. 13 ITCBypass. Read-write. Reset: 0. When set, all requests bypass the ITC. 12:11 Reserved. 10 ITCSoftInvalidate. Read-write. Reset: 0. Software may write this register to 1 to invalidate all entries in the ITC 9:8 ITCInvalidationSel. Read-write. Reset: 0. BIOS: 10b. Selects the ITC invalidation algorithm. Bits Description 00b Invalidate the entire ITC. 01b Fast imprecise invalidation. 10b Sequential precise invalidation. 11b Partial sequential precise invalidation 7:5 Reserved. 4 ITCParityEn. Read-write. Reset: 0. Enable parity protection of the ITC 3 ITCLRUUpdatePri. Read-write. Reset: 0. 0=Reads update replacement state bits when there is a simultaneous read and write to the same ITC index . 1=Writes update replacement state bits when there is a simultaneous read and write to the same ITC index 2 Reserved. 1:0 ITCReplacementSel. Read-write. Reset: 1. Selects the ITC replacement algorithm. Implementation may not support all replacement algorithms D0F2xF4_x15 L2_ITC_HASH_CONTROL Bits Description 31:16 ITCAddressMask. Read-write. Reset: FFFFh. BIOS: 0. This register is a bit-wise AND mask that selects which bits from the untranslated interrupt {MT[2:0],Vector} are used to index into the ITC. 15:11 Reserved. 10 ItcAltHashEn. Read-write. Reset: 0. BIOS: 1. Enable alternative algorithm for generating hash index into the ITC. 9 Reserved. 8:5 ITCBusBits. Read-write. Reset: 3h. Set the number of bus bits to use when using ReqID to form the ITC address. The following equation must be satisified. Func_bits + Dev_Bits + Bus_Bits &lt;= log2(ITC entries / ITC associativity) 247 42300 Rev 3.12 - July 14, 2015 BKDG for AMD Family 15h Models 10h-1Fh Processors 4:2 ITCDevBits. Read-write. Reset: 0. Set the number of device bits to use when using ReqID to form the ITC address. 1:0 ITCFuncBits. Read-write. Reset: 2h. Set the number of function bits to use when using ReqID to form the ITC address. D0F2xF4_x16 L2_ITC_WAY_CONTROL Bits Description 31:16 ITCWayAccessDisable. Read-write. Reset: 0. 15:0 ITCWayDisable. Read-write. Reset: 0. Each bit in this register disables a way in the ITC when set to 1. An implementation may have less than 32 ways. The entire cache may be disabled by setting the ITCWays lower bits of this register. D0F2xF4_x18 L2_PTC_A_CONTROL Bits Description 31:28 PTCAEntries. Read-only. Reset: 0. The number of entries in the PTC A sub-cache is indicated as 2^PTCAEntries 27:24 Reserved. 23:16 PTCAWays. Read-only. Reset: 0. Indicates the number of ways in the PTC A sub-cache 15 PTCAParitySupport. Read-only. Reset: 0. 0=The PTC A sub-cache does not support parity protection . 1=The PTC A sub-cache supports parity protection 14 Reserved. 13 PTCABypass. Read-write. Reset: 0. When set, all requests bypass the PTC A sub-cache. 12 Reserved. 11 PTCA2MMode. Read-write. Reset: 0. When set, the PTC A sub-cache stores 2M pages instead of 4K pages 10 PTCASoftInvalidate. Read-write. Reset: 0. Software may write this register to 1 to invalidate all entries in the PTC A sub-cache 9:8 PTCAInvalidationSel. Read-write. Reset: 0. BIOS: 10b. Selects the PTC A sub-cache invalidation algorithm. Bits Description 00b Invalidate the entire PTC A sub-cache. 01b Fast imprecise invalidation. 10b Sequential precise invalidation. 11b Partial sequential precise invalidation. 7:5 Reserved. 4 PTCAParityEn. Read-write. Reset: 0. Enable parity protection of the PTC A sub-cache 3 PTCALRUUpdatePri. Read-write. Reset: 0. 0=Reads update replacement state bits when there is a simultaneous read and write to the same PTCA index . 1=Writes update replacement state bits when there is a simultaneous read and write to the same PTCA index 248 42300 Rev 3.12 - July 14, 2015 2 1:0 BKDG for AMD Family 15h Models 10h-1Fh Processors Reserved. PTCAReplacementSel. Read-write. Reset: 1. Selects the PTC A sub-cache replacement algorithm. Implementation may not support all replacement algorithms D0F2xF4_x19 L2_PTC_A_HASH_CONTROL Bits Description 31:16 PTCAAddressMask. Read-write. Reset: FFFFh. BIOS: 0. This register is a bit-wise AND mask that selects which virtual address bits are used to index into the PTC A sub-cache. 15:11 Reserved. 10 PtcAltHashEn. Read-write. Reset: 0. BIOS: 1. Enable alternative algorithm for generating hash index into the PTC 9 Reserved. 8:5 PTCABusBits. Read-write. Reset: 3h. Set the number of bus bits to use when using ReqID to form the PTC A sub-cache address. The following equation must be satisified. FuncBits + DevBits + BusBits &lt;= log2(PTC A sub-cache entries / PTC A sub-cache associativity) 4:2 PTCADevBits. Read-write. Reset: 0. Set the number of device bits to use when using ReqID to form the PTC A sub-cache address 1:0 PTCAFuncBits. Read-write. Reset: 2h. Set the number of function bits to use when using ReqID to form the PTC A sub-cache address D0F2xF4_x1A L2_PTC_A_WAY_CONTROL Bits Description 31:16 PTCAWayAccessDisable. Read-write. Reset: 0. 15:0 PTCAWayDisable. Read-write. Reset: 0. Each bit in this register disables a way in the PTC A sub-cache when set to 1. An implementation may have less than 32 ways. The entire cache may be disabled by setting the PTCAWays lower bits of this register. D0F2xF4_x20 L2_CREDIT_CONTROL_2 Bits Description 31:28 Reserved. 27:24 PprLoggerCredits. Read-write. Reset: 4h. PPR log buffer credit override value 23 FCELOverride. Read-write. Reset: 0. Changing this register from 0 to 1 overrides the FCEL credit counter with FCELCredits. This should only be performed when the IOMMU is idle 22 Reserved. 21:16 FCELCredits. Read-write. Reset: 0. FCEL credit override value 15 FLTCMBOverride. Read-write. Reset: 0. Changing this register from 0 to 1 overrides the FLTCMB credit counter with FLTCMBCredits. This should only be performed when the IOMMU is idle 14 Reserved. 249 42300 Rev 3.12 - July 14, 2015 13:8 BKDG for AMD Family 15h Models 10h-1Fh Processors FLTCMBCredits. Read-write. Reset: 0. FLTCMB credit override value 7 QUEUEOverride. Read-write. Reset: 0. Changing this register from 0 to 1 overrides the QUEUE credit counter with QUEUECredits. This should only be performed when the IOMMU is idle 6 Reserved. 5:0 QUEUECredits. Read-write. Reset: 0. QUEUE credit override value D0F2xF4_x22 L2A_UPDATE_FILTER_CNTL Bits Description 31:5 Reserved. 4:1 L2aUpdateFilterRdlatency. Read-write. Reset: 3h. When L2a_Update_Filter_Bypass is 0, assume the invalidation read has completed in the number of clock cycles specified by this field. 0 L2aUpdateFilterBypass. Read-write. Reset: 1. 1=Disable duplicate update filtering. 0=Enable the dropping of updates that are already in the L2aUpdateFilter or in the destination L2a cache. D0F2xF4_x30 L2_ERR_RULE_CONTROL_3 Bits Description 31:4 ERRRuleDisable3. Read-write. Reset: 0. Each bit in this register disables an error detection rule in the IOMMU. 3:1 Reserved. 0 ERRRuleLock. Read-write. Reset: 0. BIOS: 1. This register is write-once. Setting this register bit locks the error detection rule set in ERRRuleDisable3, D0F2xF4_x31[ERRRuleDisable4], and D0F2xF4_x32[ERRRuleDisable5]. D0F2xF4_x31 L2_ERR_RULE_CONTROL_4 Bits Description 31:0 ERRRuleDisable4. Read-write. Reset: 0. Each bit in this register disables an error detection rule in the IOMMU. D0F2xF4_x32 L2_ERR_RULE_CONTROL_5 Bits Description 31:0 ERRRuleDisable5. Read-write. Reset: 0. Each bit in this register disables an error detection rule in the IOMMU. D0F2xF4_x33 L2_L2A_CK_GATE_CONTROL Bits Description 31:8 Reserved. 250 42300 Rev 3.12 - July 14, 2015 BKDG for AMD Family 15h Models 10h-1Fh Processors 7:6 CKGateL2AStop. Read-write. Reset: 01b. 0=allow 2 clock cycles delay before stopping the clocks when clkready deasserts . 1=allow 4 clock cycles delay before stopping the clocks when clkready deasserts. 2=allow 8 clock cycles delay before stopping the clocks when clkready deasserts . 3=allow 16 clock cycles delay before stopping the clocks when clkready deasserts. 5:4 CKGateL2ALength. Read-write. Reset: 01b. 0=allow 128 clock cycles delay before stopping the clocks when idle asserts . 1=allow 256 clock cycles delay before stopping the clocks when idle asserts. 2=allow 512 clock cycles delay before stopping the clocks when idle asserts . 3=allow 1024 clock cycles delay before stopping the clocks when idle asserts. 3 CKGateL2ASpare. Read-write. Reset: 0. 2 CKGateL2ACacheDisable. Read-write. Reset: 1. BIOS: 0. Disable the gating of the l2b upper cache ways. 1 CKGateL2ADynamicDisable. Read-write. Reset: 1. BIOS: 0. Disable the gating of the l2b dynamic clock branch. 0 CKGateL2ARegsDisable. Read-write. Reset: 1. BIOS: 0. Disable the gating of the l2b register clock branch. D0F2xF4_x34 L2_L2A_PGSIZE_CONTROL Bits Description 31:4 Reserved. 3:2 L2aregHostPgsize. Read-write. Reset: 0. BIOS: 10b. 1:0 L2aregGstPgsize. Read-write. Reset: 0. BIOS: 10b. D0F2xF4_x40 L2_PERF_CNTL_2 Bits Description 31:24 L2PerfCountUpper5. Read-only. Reset: 0. Upper 8 bits of IOMMU L2 performance counter 5 23:16 L2PerfCountUpper4. Read-only. Reset: 0. Upper 8 bits of IOMMU L2 performance counter 4 15:8 L2PerfEvent5. Read-write. Reset: 0. Selects the IOMMU L2 performance counter event for counter 5 7:0 L2PerfEvent4. Read-write. Reset: 0. Selects the IOMMU L2 performance counter event for counter 4 D0F2xF4_x41 L2_PERF_COUNT_4 Bits Description 31:0 L2PerfCount4. Read-only. Reset: 0. Lower 32 bits of IOMMU L2 performance counter 4 D0F2xF4_x42 L2_PERF_COUNT_5 Bits Description 31:0 L2PerfCount5. Read-only. Reset: 0. Lower 32 bits of IOMMU L2 performance counter 5 251 42300 Rev 3.12 - July 14, 2015 BKDG for AMD Family 15h Models 10h-1Fh Processors D0F2xF4_x43 L2_PERF_CNTL_3 Bits Description 31:24 L2PerfCountUpper7. Read-only. Reset: 0. Upper 8 bits of IOMMU L2 performance counter 7 23:16 L2PerfCountUpper6. Read-only. Reset: 0. Upper 8 bits of IOMMU L2 performance counter 6 15:8 L2PerfEvent7. Read-write. Reset: 0. Selects the IOMMU L2 performance counter event for counter 7 7:0 L2PerfEvent6. Read-write. Reset: 0. Selects the IOMMU L2 performance counter event for counter 6 D0F2xF4_x44 L2_PERF_COUNT_6 Bits Description 31:0 L2PerfCount6. Read-only. Reset: 0. Lower 32 bits of IOMMU L2 performance counter 6 D0F2xF4_x45 L2_PERF_COUNT_7 Bits Description 31:0 L2PerfCount7. Read-only. Reset: 0. Lower 32 bits of IOMMU L2 performance counter 7 D0F2xF4_x46 L2_DEBUG_2 Bits Description 31:0 L2DEBUG2. Read-write. Reset: 0. Reserved for to control ECOs D0F2xF4_x47 L2_DEBUG_3 Bits Description 31:3 Reserved. 2 TwAtomicFilterEn. Read-write. Reset: 0. BIOS: 1. 1=Enable table walker atomic filtering. 1 TwNwEn. Read-write. Reset: 0. BIOS: 1. 1=Enable NW bit for ATS requests. 0 Reserved. D0F2xF4_x48 L2_STATUS_1 Bits Description 31:0 L2STATUS1. Read-only. Reset: 0. Internal IOMMU L2B status 252 42300 Rev 3.12 - July 14, 2015 BKDG for AMD Family 15h Models 10h-1Fh Processors D0F2xF4_x4C L2_CONTROL_5 Bits Description 31:10 Reserved. Read-write. 9:8 GstPartialPtcCntrl. Read-write. Reset: 0. BIOS: 11b. 7 FC2AltMode. Read-write. Reset: 0. 0=FC2 primary flow-control mode . 1=FC2 alternate flowcontrol mode 6 FC3Dis. Read-write. Reset: 0. 0=FC3 flow-control loop is enabled . 1=FC3 flow-control look is disabled 5 FC2Dis. Read-write. Reset: 0. 0=FC2 flow-control loop is enabled . 1=FC2 flow-control look is disabled 4 DTCUpdateVZeroIVOne. Read-write. Reset: 0. 0=DTEs with V=0 and IV=1 are not cached in the DTC . 1=DTEs with V=0 and IV=1 are cached in the DTC 3 DTCUpdateVOneIVZero. Read-write. Reset: 0. 0=DTEs with V=1 and IV=0 are not cached in the DTC . 1=DTEs with V=1 and IV=0 are cached in the DTC 2 FC1Dis. Read-write. Reset: 0. 0=FC1 flow control loop enabled . 1=FC1 flow control loop disabled 1 PTCAddrTransReqUpdate. Read-write. Reset: 1. 0=PTEs from address translation requests are not cached . 1=PTEs from address translation requests are cached in the L2 according to the Cache bit in the DTE 0 QueueArbFBPri. Read-write. Reset: 1. 0=Requests in the miss queue and the feedback queue are arbitrated in a round-robin manner . 1=Requests in the feedback queue are given priority over requests in the miss queue D0F2xF4_x4D L2_CONTROL_6 Bits Description 31:24 Perf2Threshold. Read-write. Reset: 0. Fifo threshold level used to calculate certain performance counter values. 23:17 Reserved. 16 SeqInvBurstLimitEn. Read-write. Reset: 1. Enable stalling PDC requests to allow invalidation cycles to make forward progress based upon SeqInvBurstLimitInv and SeqInvBurstLimitPDCReq 15:8 SeqInvBurstLimitPDCReq. Read-write. Reset: 8h. Sets the number of consecutive IOMMU PDC requests to perform when doing sequential invalidation. PDC and invalidation requests will alternat e access to the PDC based upon SeqInvBurstLimitInv and SeqInvBurstLimitPDCReq. 7:0 SeqInvBurstLimitInv. Read-write. Reset: 8h. Sets the number of consecutive invalidation requests to perform when doing sequential invalidation. PDC and invalidation requests will alte rnate access to the PDC based upon SeqInvBurstLimitInv and SeqInvBurstLimitPDCReq. 253 42300 Rev 3.12 - July 14, 2015 BKDG for AMD Family 15h Models 10h-1Fh Processors D0F2xF4_x50 L2_PDC_CONTROL Bits Description 31:28 PDCEntries. Read-only. Reset: 0. Indicates the number of entries in the PDC is indicated as 2^PDCEntries 27:24 Reserved. 23:16 PDCWays. Read-only. Reset: 0. Indicates the number of ways in the PDC 15 PDCParitySupport. Read-only. Reset: 0. 0=The PDC does not support parity protection . 1=The PDC supports parity protection 14 Reserved. 13 PDCBypass. Read-write. Reset: 0. When set, all requests bypass the PDC. This prevents the multiple issue of requests and increases maximum rate of requests to the table-walker. 12 PDCSearchDirection. Read-write. Reset: 0. 0=Search PDC from higher levels down . 1=Search PDC from lower levels up 11 Reserved. 10 PDCSoftInvalidate. Read-write. Reset: 0. Software may write this register to 1 to invalidate all entries in the PDC 9:8 PDCInvalidationSel. Read-write. Reset: 0. BIOS: 10b. Selects the PDC invalidation algorithm. Bits Description 00b Invalidate the entire PDC. 01b Fast imprecise invalidation. 10b Sequential precise invalidation. 11b Partial sequential precise invalidation 7:5 Reserved. 4 PDCParityEn. Read-write. Reset: 0. Enable parity protection of the PDC if the device supports parity 3 PDCLRUUpdatePri. Read-write. Reset: 0. 0=Reads update replacement state bits when there is a simultaneous read and write to the same PDC index . 1=Writes update replacement state bits when there is a simultaneous read and write to the same PDC index 2 Reserved. 1:0 PDCReplacementSel. Read-write. Reset: 1. Selects the PDC replacement algorithm. Implementation may not support all replacement algorithms D0F2xF4_x51 L2_PDC_HASH_CONTROL Bits Description 31:16 PDCAddressMask. Read-write. Reset: FFFFh. BIOS: 0. This register is a bit-wise AND mask that selects which virtual address bits are used to index into the PDC. 15:11 Reserved. 10 PdcAltHashEn. Read-write. Reset: 0. BIOS: 1. 1=Enable alternative algorithm for generating hash index into the PDC. 254 42300 Rev 3.12 - July 14, 2015 BKDG for AMD Family 15h Models 10h-1Fh Processors 9 PDCUpperLvlAddrHash. Read-write. Reset: 1. When set to 1, the PDC cache index is partially formed using the xor of the LSBs of virtual address bits for all levels greater than or equal to the stored/searched level. 8 PDCLvlHash. Read-write. Reset: 1. When set to 1, the PDE level is used as part of the hash for the cache index. 7:6 Reserved. 5:0 PDCDomainBits. Read-write. Reset: 7h. Selects the number of domain bits to use as part of the index into the PDC. D0F2xF4_x52 L2_PDC_WAY_CONTROL Bits Description 31:16 PDCWayAccessDisable. Read-write. Reset: 0. 15:0 PDCWayDisable. Read-write. Reset: 0. Each bit in this register disables a way in the PDC when set to 1. An implementation may have less than 32 ways. The entire cache may be disabled by setting the PDCWays lower bits of this register. D0F2xF4_x53 L2B_UPDATE_FILTER_CNTL Bits Description 31:5 Reserved. 4:1 L2bUpdateFilterRdlatency. Read-write. Reset: 3h. When L2b_Update_Filter_Bypass is 0, assume the invalidation read has completed in the number of clock cycles specified by this field. 0 L2bUpdateFilterBypass. Read-write. Reset: 1. 1 - Disable duplicate update filtering;. 0 - Enable the dropping of updates that are already in the l2b_update_filter or in the PDC. D0F2xF4_x54 L2_TW_CONTROL Bits Description 31:18 Reserved. 17 Twfilter64bDis. Read-write. Reset: 0. 16 TwfilterDis. Read-write. Reset: 0. 15 Reserved. 14:12 TWPrefetchRange. Read-write. Reset: 1. Selects the number of pages to prefetch 11 TWPTEOnAddrTransExcl. Read-write. Reset: 0. 0=Table walker returns DTE to L1 on an address translation exclusion range access . 1=Table walker returns PTE to L1 on an address translation exclusion range access 10 TWPTEOnUntransExcl. Read-write. Reset: 0. 0=Table walker returns DTE to L1 on an untranslated exclusion range access . 1=Table walker returns PTE to L1 on an untranslated exclusion range access 9 TWPrefetchOnly4KDis. Read-write. Reset: 0. 1=Allow non-4K pages to be prefetched . 0=Only 4K pages are prefetched 255 42300 Rev 3.12 - July 14, 2015 BKDG for AMD Family 15h Models 10h-1Fh Processors 8 TWPrefetchEn. Read-write. Reset: 0. Enable prefetching in the table-walker 7 Reserved. 6 TWForceCoherent. Read-write. Reset: 0. 1=Table-walker always genereates coherent requests. The DTE SD bit is ignored when this bit is set to 1. 5:0 TWQueueLimit. Read-write. Reset: 10h. Limit the number of outstanding table-walker requests D0F2xF4_x56 L2_CP_CONTROL Bits Description 31:16 CPRdDelay. Read-write. Reset: 0. Command processor read delay 15:3 Reserved. 2 CPFlushOnInv. Read-write. Reset: 1. BIOS: 0. 1=Command processor flushes out old requests on every invalidation command . 0=No flush is performed during invalidations 1 CPFlushOnWait. Read-write. Reset: 0. BIOS: 1. 1=Command processor flushes out old requests on completion wait . 0=No flush is performed on completion wait 0 CPPrefetchDis. Read-write. Reset: 0. 1=Command processor fetches and executes only one command at a time . 0=Command processor prefetches available commands into its internal storage D0F2xF4_x57 L2_CP_CONTROL_1 Bits Description 31:3 Reserved. 2 L1ImuIntGfxDis. Read-write. Reset: 0. BIOS: IF (GpuEnabled) THEN 0 ELSE 1 ENDIF. 1 Reserved. 0 L1ImuPcieGfxDis. Read-write. Reset: 0. BIOS: This bit should be set if there is no external graphics in the system. D0F2xF4_x58 IOMMU_L2_GUEST_ADDR_CNTRL Bits Description 31:24 Reserved. 23:0 IommuL2GuestAddrMask. Read-write. Reset: 0. D0F2xF4_x60 L2_TW_CONTROL_1 Bits Description 31:15 TWDebugMask. Read-write. Reset: 0. Defines the table-walker trace buffer size by masking address bits 31:16 14:3 Reserved. 256 42300 Rev 3.12 - July 14, 2015 BKDG for AMD Family 15h Models 10h-1Fh Processors 2 TWDebugForceDisable. Read-write. Reset: 0. When set to 1, this register disables the TW Debug feature until a cold-reset is performed 1 TWDebugNoWrap. Read-write. Reset: 0. 1=Table-walker trace to stop at the top of the trace buffer . 0=Table-walker trace wraps around at the top of the trace buffer 0 TWDebugEn. Read-write. Reset: 0. Enables table-walker trace debug mode D0F2xF4_x61 L2_TW_CONTROL_2 Bits Description 31:12 TWDebugAddrLo. Read-write. Reset: 0. Base address bits 31:12 for table-walker trace debug 11:0 Reserved. D0F2xF4_x62 L2_TW_CONTROL_3 Bits Description 31:0 TWDebugAddrHi. Read-write. Reset: 0. Base address bits 51:32 for table-walker trace debug D0F2xF4_x6A L2_INT_CONTROL Bits Description 31:3 Reserved. 2 IntPPROrderEn. Read-write. Reset: 1. Enable ordering between interrupts and ppr log writes 1 IntCPOrderEn. Read-write. Reset: 1. Enable ordering between interrupts and command processor writes 0 IntEventOrderEn. Read-write. Reset: 1. Enable ordering between interrupts and event log writes D0F2xF4_x70 L2_CREDIT_CONTROL_0 Bits Description 31 DTEOverride. Read-write. Reset: 0. Changing this register from 0 to 1 overrides the DTE credit counter with DTECredits. This should only be performed when the IOMMU is idle 30 Reserved. 29:24 DTECredits. Read-write. Reset: 0. DTE credit override value 23 FC3Override. Read-write. Reset: 0. Changing this register from 0 to 1 overrides the FC3 credit counter with FC3Credits. This should only be performed when the IOMMU is idle 22 Reserved. 21:16 FC3Credits. Read-write. Reset: 0. FC3 credit override value 15 FC2Override. Read-write. Reset: 0. Changing this register from 0 to 1 overrides the FC2 credit counter with FC2Credits. This should only be performed when the IOMMU is idle 14 Reserved. 13:8 FC2Credits. Read-write. Reset: 0. FC2 credit override value 257 42300 Rev 3.12 - July 14, 2015 BKDG for AMD Family 15h Models 10h-1Fh Processors 7 FC1Override. Read-write. Reset: 0. Changing this register from 0 to 1 overrides the FC1 credit counter with FC1Credits. This should only be performed when the IOMMU is idle 6 Reserved. 5:0 FC1Credits. Read-write. Reset: 0. FC1 credit override value D0F2xF4_x71 L2_CREDIT_CONTROL_1 Bits Description 31:24 Reserved. 23:20 PprMcifCredits. Read-write. Reset: 4h. PPR logger credit override value 19:16 CpPrefetchCredits. Read-write. Reset: 4h. Command processor prefetch credit override value 15 TWELOverride. Read-write. Reset: 0. Changing this register from 0 to 1 overrides the TWEL credit counter with TWELCredits. This should only be performed when the IOMMU is idle 14 Reserved. 13:8 TWELCredits. Read-write. Reset: 0. TWEL credit override value 7 PDTIEOverride. Read-write. Reset: 0. Changing this register from 0 to 1 overrides the PDTIE credit counter with PDTIECredits. This should only be performed when the IOMMU is idle 6 Reserved. 5:0 PDTIECredits. Read-write. Reset: 0. PDTIE credit override value D0F2xF4_x78 L2_MCIF_CONTROL Bits Description 31:29 Reserved. 28:24 MCIFBaseWriteDataCredits. Read-write. Reset: 8h. Sets the number of base-channel write data credits between the IOMMU L2 and the HTIU/ORB. This register requires a warm-reset to take effect 23:21 Reserved. 20:16 MCIFBaseWriteHdrCredits. Read-write. Reset: 8h. Sets the number of base-channel write header credits between the IOMMU L2 and the HTIU/ORB. This register requires a warm-reset to take effect 15:13 Reserved. 12:8 MCIFIsocReadCredits. Read-write. Reset: 8h. Sets the number of isoc-channel read credits between the IOMMU L2 and the HTIU/ORB. This register requires a warm-reset to take effect 7:5 Reserved. 4:0 MCIFBaseReadCredits. Read-write. Reset: 8h. Sets the number of base-channel read credits between the IOMMU L2 and the HTIU/ORB. This register requires a warm-reset to take effect 258 42300 Rev 3.12 - July 14, 2015 BKDG for AMD Family 15h Models 10h-1Fh Processors D0F2xF4_x80 L2_ERR_RULE_CONTROL_0 Bits Description 31:4 ERRRuleDisable0. Read-write. Reset: 0. Each bit in this register disables an error detection rule in the IOMMU. 3:1 Reserved. 0 ERRRuleLock0. Read-write. Reset: 0. BIOS: 1. This register is write-once. Setting this register bit locks the error detection rule set in ERRRuleDisable0, D0F2xF4_x81[ERRRuleDisable1], and D0F2xF4_x82[ERRRuleDisable2]. D0F2xF4_x81 L2_ERR_RULE_CONTROL_1 Bits Description 31:0 ERRRuleDisable1. Read-write. Reset: 0. Each bit in this register disables an error detection rule in the IOMMU. D0F2xF4_x82 L2_ERR_RULE_CONTROL_2 Bits Description 31:0 ERRRuleDisable2. Read-write. Reset: 0. Each bit in this register disables an error detection rule in the IOMMU. D0F2xF4_x90 L2_L2B_CK_GATE_CONTROL Bits Description 31:8 Reserved. 7:6 CKGateL2BStop. Read-write. Reset: 01b. 0=allow 2 clock cycles delay before stopping the clocks when clkready deasserts . 1=allow 4 clock cycles delay before stopping the clocks when clkready deasserts. 2=allow 8 clock cycles delay before stopping the clocks when clkready deasserts . 3=allow 16 clock cycles delay before stopping the clocks when clkready deasserts 5:4 CKGateL2BLength. Read-write. Reset: 01b. 0=allow 128 clock cycles delay before stopping the clocks when idle asserts . 1=allow 256 clock cycles delay before stopping the clocks when idle asserts. 2=allow 512 clock cycles delay before stopping the clocks when idle asserts . 3=allow 1024 clock cycles delay before stopping the clocks when idle asserts 3 CKGateL2BCacheDisable. Read-write. Reset: 0. Disable the gating of the l2b upper cache ways 2 CKGateL2BMiscDisable. Read-write. Reset: 1. Disable the gating of the l2b miscelaneous clock branch 1 CKGateL2BDynamicDisable. Read-write. Reset: 1. BIOS: 0. Disable the gating of the l2b dynamic clock branch 0 CKGateL2BRegsDisable. Read-write. Reset: 1. BIOS: 0. Disable the gating of the l2b register clock branch 259 42300 Rev 3.12 - July 14, 2015 BKDG for AMD Family 15h Models 10h-1Fh Processors D0F2xF4_x92 PPR_CONTROL Bits Description 31:17 Reserved. 16 PprIntcoallesceEn. Read-write. Reset: 0. BIOS: 0. This bit must not be set if D0F2xF4_x90[CKGateL2BMiscDisable]=0. 15:8 PprIntreqdelay. Read-write. Reset: 0. BIOS: 20h. 7:0 PprInttimedelay. Read-write. Reset: 0. BIOS: 15h. D0F2xF4_x94 L2_L2B_PGSIZE_CONTROL Bits Description 31:4 Reserved. 3:2 L2bregHostPgsize. Read-write. Reset: 0. BIOS: 10b. 1:0 L2bregGstPgsize. Read-write. Reset: 0. BIOS: 10b. D0F2xF8 IOMMU L1 Config Index The index/data pair registers D0F2xF8 and D0F2xFC is used to access the registers D0F2xFC_x[FFFF:0000]_L1sel[3:0]. To read or write to one of these register, the address is written first into the address register D0F2xF8 and then the data are read or written by read or write the data register D0F2xFC. See 2.12.1 [IOMMU Configuration Space]. There are four L1s in the IOMMU. Registers in the L1 indexed space have one instance per L1 denoted by _L1sel[x] where x=D0F2xF8[L1cfgSel]. The syntax for this register type is described by example as follows: • D0F2xFC_x00_L1sel[3:0] refers to all instances of the D0F2xFC_x00 registers. • D0F2xFC_x00_L1sel[2] refers to the GBIF instance of D0F2xFC_x00. Bits 31 Description L1cfgEn. Read-write. Reset: 0. 1=Enable writes to D0F2xFC. 30:20 Reserved. 19:16 L1cfgSel. Read-write. Reset: 0. This field selects one of the four L1s to access. Bits Description 0h GFX 1h GPPSB 2h GBIF 3h INTGEN Fh-4h Reserved 15:0 L1cfgIndex. Read-write. Reset: 0. D0F2xFC IOMMU L1 Config Data See D0F2xF8. 260 42300 Rev 3.12 - July 14, 2015 BKDG for AMD Family 15h Models 10h-1Fh Processors D0F2xFC_x00_L1sel[3:0] L1_PERF_CNTL Bits Description 31:24 L1PerfCountHi1. Read-only. Reset: 0. read back of perf counter 1 bits 39:32 23:16 L1PerfCountHi0. Read-only. Reset: 0. read back of perf counter 0 bits 39:32 15:8 L1PerfEvent1. Read-write. Reset: 0. perf counter event 1 7:0 L1PerfEvent0. Read-write. Reset: 0. perf counter event 0 D0F2xFC_x01_L1sel[3:0] L1_PERF_COUNT_0 Bits Description 31:0 L1PerfCount0. Read-only. Reset: 0. read back of perf counter 0 bits 31:0 D0F2xFC_x02_L1sel[3:0] L1_PERF_COUNT_1 Bits Description 31:0 L1PerfCount1. Read-only. Reset: 0. read back of perf counter 1 bits 31:0 D0F2xFC_x07_L1sel[3:0] L1_DEBUG_1 Bits Description 31:18 Reserved. 17 L1NwEn. Read-write. Reset: 0. BIOS: 1. 1=Enable NW bit on ATS requests. 16:15 Reserved. 14 AtsPhysPageOverlapDis. Read-write. Reset: 0. BIOS: 1. 1=Prevent physical page overlap for ATS responses. 13 Reserved. 12 AtsSeqNumEn. Read-write. Reset: 0. BIOS: 1. 1=Enable logging of ATS sequence number. 11 SpecReqFilterEn. Read-write. Reset: 0. BIOS: 1. 1=Filter special requests in L1 work queue. 10:1 0 Reserved. PhantomFuncEn. Read-write. Reset: 0. BIOS: See 2.12.2. 1=Enable phantom function support. D0F2xFC_x08_L1sel[3:0] L1_DEBUG_STATUS Bits Description 31:0 L1debugStatus. Read-only. Reset: 0. HW status bits reserved for ECOs 261 42300 Rev 3.12 - July 14, 2015 BKDG for AMD Family 15h Models 10h-1Fh Processors D0F2xFC_x0C_L1sel[3:0] L1_CNTRL_0 Bits 31 Description Reserved. 30:28 L1VirtOrderQueues. Read-write. Reset: 0. BIOS: 4h. This field controls number of virtual queues in the L1 work queue. Bits Description 0h 1 1h 2 2h 4 3h 8 4h 16 7h-5h Reserved 27:24 L1Entries. Read-only; updated-by-hardware. Reset: 0. This field specifies the number of entries in each L1 cache as 2^L1entries. 23:22 Reserved. 21:20 L1Banks. Read-only; updated-by-hardware. Reset: 1. This field specifies number of caches in L1. 19:14 Reserved1. 13:8 L2Credits. Read-write. Reset: 4h. This field controls credits for L1 to L2 interface. 7:6 Reserved. 5 ReplacementSel. Read-write. Reset: 0. 4 Reserved0. 3 CacheiwOnly. Read-write. Reset: 1. cache write only pages in L1 2 CacheirOnly. Read-write. Reset: 1. cache read only pages in L1 1 FragmentDis. Read-write. Reset: 0. disable variable page size support in L1 cache - only 4K pages 0 UnfilterDis. Read-write. Reset: 0. disable unfiltering in L1 wq of aborted L2 requests D0F2xFC_x0D_L1sel[3:0] L1_CNTRL_1 Bits Description 31 Reserved. 30 L1DebugCntrMode. Read-write. Reset: 0. Mode control for debug bus 29 Untrans2mFilteren. Read-write. Reset: 0. Enable filtering of requests on a 2M boundry instead of 4K 28 PretransNovaFilteren. Read-write. Reset: 0. When set, VA is not used for filtering pretrans requests 27 L1CacheSelInterleave. Read-write. Reset: 0. when set causes cache updates to toggle between multiple caches 26 L1CacheSelReqid. Read-write. Reset: 0. when set will allow the reqid to be used in hashing between multiple L1 caches 262 42300 Rev 3.12 - July 14, 2015 BKDG for AMD Family 15h Models 10h-1Fh Processors 25:23 SelectTimeoutPulse. Read-write. Reset: 0. 22 L1CacheInvAllEn. Read-write. Reset: 0. Enables invalidation of entire cache when invalidation command is sent 21 L1orderEn. Read-write. Reset: 0. Enables strict ordering of all requests through L1 20 SndFilterDis. Read-write. Reset: 0. Disables filtering of requests to L2 19 AtsNobufferInsert. Read-write. Reset: 0. disables buffering of read completion data when inserting ats responses 18:14 WqEntrydis. Read-write. Reset: 0. Value indicates how many cache entries in L1 to disable 13 BlockL1Dis. Read-only. Reset: 0. N/A 12 L1DTEDis. Read-write. Reset: 0. Disables L1 caching of DTE 11 L1ParityEn. Read-write. Reset: 0. 10 L1CacheParityEn. Read-write. Reset: 0. Enables forced miss of L1 cache due to failed parity check 9 CacheByPass. Read-write. Reset: 0. Enables L1 cache bypass 8 VOQXorMode. Read-write. Reset: 0. N/A 7 Reserved. 6:4 3 2:0 VOQFuncBits. Read-write. Reset: 0. N/A Reserved. VOQPortBits. Read-write. Reset: 0. BIOS: See 2.12.2. When not 000b, enables virtual queue hashing using port id. This field controls the number of bits to use from portid for hashing. D0F2xFC_x0E_L1sel[3:0] L1_CNTRL_2 Bits Description 31:28 Reserved. 27:20 MsiHtRsvIntVector. Read-write. Reset: 00h.This field defines the interrupt vector used when an MSI interrupt is received that has a reserved DM field. 19:12 MsiHtRsvIntDestination. Read-write. Reset: FFh.This field defines the interrupt destination used when an MSI interrupt is received that has a reserved DM field. 11 Reserved. 10 MsiHtRsvIntDM. Read-write. Reset: 0.Defines the interrupt destination mode when an MSI interrupt is received that has a reserved DM field. 9 MsiHtRsvIntRqEio. Read-write. Reset: 0.Specifies the RQEOI state when an MSI interrupt is received that has a reserved DM field. 8:6 MsiHtRsvIntMt. Read-write. Reset: 011b.Specfies the message type used when an MSI interrupt is received that has a reserved DM field. 5:3 Reserved. 2 L1AbrtAtsDis. Read-write. Reset: 0. 1=Disable abort of ATS requests when IOMMU is disabled. 1 MsiToHtRemapDis. Read-write. Reset: 0. 1=Disable mapping of MSI to HT interrupts. 0 Reserved. 263 42300 Rev 3.12 - July 14, 2015 BKDG for AMD Family 15h Models 10h-1Fh Processors D0F2xFC_x0F_L1sel[3:0] L1_CNTRL_3 Bits Description 31:0 AtsTlbinvPulseWidth. Read-write. Reset: C350h. sets the pulse width of the ats invalidation counters D0F2xFC_x10_L1sel[3:0] L1_BANK_SEL_0 Bits Description 31:16 Reserved. 15:0 L1cachebanksel0. Read-write. Reset: 1. value is used to determine the virtual address bit that selects between the 2 banks of the L1 cache (if present). The bank is selected by bitwise ANDing this register against virtual address bits 19:12 and XORing the result D0F2xFC_x11_L1sel[3:0] L1_BANK_DISABLE_0 Bits Description 31:14 Reserved. 13:8 L1cachelinedis1. Read-write. Reset: 0. sets the number of cache entries to disable in cache 1 7:6 Reserved. 5:0 L1cachelinedis0. Read-write. Reset: 0. sets the number of cache entries to disable in cache 0 D0F2xFC_x20_L1sel[3:0] L1_WQ_STATUS_0 Table 121: Valid values for D0F2xFC_x20_L1sel[3:0] Value 0h 1h 2h 3h 4h 5h Bits Description Idle. Wait_L1. Wait_L2. Sending special request to L2. Waiting for completion of special request. Done. Description 31:30 Reserved. 29:27 EntryStatus9. See: EntryStatus0. 26:24 EntryStatus8. See: EntryStatus0. 23:21 EntryStatus7. See: EntryStatus0. 20:18 EntryStatus6. See: EntryStatus0. 17:15 EntryStatus5. See: EntryStatus0. 264 42300 Rev 3.12 - July 14, 2015 BKDG for AMD Family 15h Models 10h-1Fh Processors 14:12 EntryStatus4. See: EntryStatus0. 11:9 EntryStatus3. See: EntryStatus0. 8:6 EntryStatus2. See: EntryStatus0. 5:3 EntryStatus1. See: EntryStatus0. 2:0 EntryStatus0. Read-only. Reset: 0. See: Table 121. D0F2xFC_x21_L1sel[3:0] L1_WQ_STATUS_1 Bits Description 31:30 Reserved. 29:27 EntryStatus19. Read-only. Reset: 0. 26:24 EntryStatus18. Read-only. Reset: 0. 23:21 EntryStatus17. Read-only. Reset: 0. 20:18 EntryStatus16. Read-only. Reset: 0. 17:15 EntryStatus15. Read-only. Reset: 0. 14:12 EntryStatus14. Read-only. Reset: 0. 11:9 EntryStatus13. Read-only. Reset: 0. 8:6 EntryStatus12. Read-only. Reset: 0. 5:3 EntryStatus11. Read-only. Reset: 0. 2:0 EntryStatus10. Read-only. Reset: 0. D0F2xFC_x22_L1sel[3:0] L1_WQ_STATUS_2 Bits Description 31:30 Reserved. 29:27 EntryStatus29. Read-only. Reset: 0. 26:24 EntryStatus28. Read-only. Reset: 0. 23:21 EntryStatus27. Read-only. Reset: 0. 20:18 EntryStatus26. Read-only. Reset: 0. 17:15 EntryStatus25. Read-only. Reset: 0. 14:12 EntryStatus24. Read-only. Reset: 0. 11:9 EntryStatus23. Read-only. Reset: 0. 8:6 EntryStatus22. Read-only. Reset: 0. 5:3 EntryStatus21. Read-only. Reset: 0. 2:0 EntryStatus20. Read-only. Reset: 0. 265 42300 Rev 3.12 - July 14, 2015 BKDG for AMD Family 15h Models 10h-1Fh Processors D0F2xFC_x23_L1sel[3:0] L1_WQ_STATUS_3 Bits Description 31:16 Reserved. 15:8 InvalidationStatus. Read-only. Reset: 0. status of invalidation state machine 7:6 Reserved. 5:3 EntryStatus31. Read-only. Reset: 0. 2:0 EntryStatus30. Read-only. Reset: 0. D0F2xFC_x32_L1sel[3:0] L1_CNTRL_4 Bits Description 31:4 Reserved. 3 TlpprefixerrEn. Read-write. Reset: 0. 2 TimeoutPulseExtEn. Read-write. Reset: 0. 1 AtsMultipleL1toL2En. Read-write. Reset: 0. BIOS: 1. 0 AtsMultipleRespEn. Read-write. Reset: 0. BIOS: 1. D0F2xFC_x33_L1sel[3:0] L1_CLKCNTRL_0 Bits 31 Description L1L2ClkgateEn. Read-write. Reset: 0. BIOS: 1. 30:12 Reserved. 11 L1HostreqClkgateEn. Read-write. Reset: 0. BIOS: 0. 10 L1RegClkgateEn. Read-write. Reset: 0. BIOS: 1. 9 L1MemoryClkgateEn. Read-write. Reset: 0. BIOS: 1. 8 L1PerfClkgateEn. Read-write. Reset: 0. BIOS: 1. 7 L1DmaInputClkgateEn. Read-write. Reset: 0. BIOS: 1. 6 L1CpslvClkgateEn. Read-write. Reset: 0. BIOS: 1. 5 L1CacheClkgateEn. Read-write. Reset: 0. BIOS: 1. 4 L1DmaClkgateEn. Read-write. Reset: 0. BIOS: 1. 3:2 Reserved. 1:0 L1ClkgateLen. Read-write. Reset: 0. D0F2xFC_x34_L1sel[3:0] L1_MEMPWRCNTRL_0 Bits Description 31:24 L1MempwrTimer2. Read-write. Reset: Fh. 23:16 L1MempwrTimer1. Read-write. Reset: Fh. 266 42300 Rev 3.12 - July 14, 2015 BKDG for AMD Family 15h Models 10h-1Fh Processors 15:8 L1MempwrTimer0. Read-write. Reset: Fh. 7:1 Reserved. 0 L1MempwrEn. Read-write. Reset: 0. D0F2xFC_x35_L1sel[3:0] L1_MEMPWRCNTRL_1 Bits Description 31:8 Reserved. 7:0 L1MempwrTimer3. Read-write. Reset: Fh. D0F2xFC_x36_L1sel[3:0] L1_GUEST_ADDR_CNTRL Bits Description 31:8 L1GuestAddrMsk. Read-write. Reset: 0. 7:1 Reserved. 0 L1CanonicalErrEn. Read-write. Reset: 0. D0F2xFC_x37_L1sel[3:0] L1_FEATURE_SUP_CNTRL Bits Description 31:2 Reserved. 1 L1PprSup. Read-write. Reset: 1. 0 L1EfrSup. Read-write. Reset: 1. 267 42300 Rev 3.12 - July 14, 2015 3.5 BKDG for AMD Family 15h Models 10h-1Fh Processors Device 1 Function 0 Configuration Registers See 3.1 [Register Descriptions and Mnemonics]. See 2.7 [Configuration Space]. D1F0x00 Device/Vendor ID Register Bits Description 31:16 DeviceID: device ID. Value: Product-specific. 15:0 VendorID: vendor ID. Read-only. Reset: 1002h. D1F0x04 Status/Command Register Reset: 0010_0000h. Bits Description 31 ParityErrorDetected: detected parity error. Read; Write-1-to-clear. 1=Poisoned TLP received. 30 SignaledSystemError: signaled system error. Read; Write-1-to-clear. 1=A non-fatal or fatal error message was sent and SerrEn=1. 29 ReceivedMasterAbort: received master abort. Read; Write-1-to-clear. 1=A completion with an unsupported request completion status was received. 28 ReceivedTargetAbort: received target abort. Read; Write-1-to-clear. 1=A completion with completer abort completion status was received. 27 SignalTargetAbort: Signaled target abort. Read-only. 26:25 DevselTiming: DEVSEL# Timing. Read-only. 24 MasterDataPerr: master data parity error. Read; Write-1-to-clear. 1=ParityErrorEn=1 and either a poisoned completion was received or the device poisoned a write request. 23 FastBackCapable: fast back-to-back capable. Read-only. 22 UDFEn: UDF enable. Read-only. 21 PCI66En: 66 MHz capable. Read-only. 20 CapList: capability list. Read-only. 1=capability list supported. 19 IntStatus: interrupt status. Read-only. 1=INTx interrupt message pending. 18:11 Reserved. 10 IntDis: interrupt disable. Read-write. 1=INTx interrupt messages generation disabled. 9 FastB2BEn: fast back-to-back enable. Read-only. 8 SerrEn: System error enable. Read-write. 1=Enables reporting of non-fatal and fatal errors detected. 7 Stepping: Stepping control. Read-only. 6 ParityErrorEn: parity error response enable. Read-write. 5 PalSnoopEn: VGA palette snoop enable. Read-only. 4 MemWriteInvalidateEn: memory write and invalidate enable. Read-only. 3 SpecialCycleEn: special cycle enable. Read-only. 268 42300 Rev 3.12 - July 14, 2015 BKDG for AMD Family 15h Models 10h-1Fh Processors 2 BusMasterEn: bus master enable. Read-write. 1=Memory and IO read and write request generation enabled. 1 MemAccessEn: IO access enable. Read-write. This bit controls if memory accesses targeting this device are accepted. 1=Enabled. 0=Disabled. 0 IoAccessEn: IO access enable. Read-write. This bit controls if IO accesses targeting this device are accepted. 1=Enabled. 0=Disabled. D1F0x08 Class Code/Revision ID Register Bits Description 31:8 ClassCode. Value: 03_0000h. 7:0 RevID: revision ID. Value: Product-specific. D1F0x0C Header Type Register Reset: 0080_0000h. Bits Description 31:24 BIST. Read-only. 23:16 HeaderTypeReg. Read-only. The header type field indicates a header type 0 and that this is a multifunction device. 15:8 LatencyTimer. Read-only. These bits are fixed at their default value. 7:0 CacheLineSize. Read-write. This field specifies the system cache line size in units of double words. D1F0x10 Graphic Memory Base Address IF (D0F0x64_x1C[F064BarEn]==0) THEN Reset: 0000_0008h. ELSE Reset: 0000_000Ch. ENDIF. Bits Description 31:26 BaseAddr[31:26]: base address. Read-write. The amount of memory requested by the graphics memory BAR is controlled by D0F0x64_x1C[MemApSize]. 25:4 BaseAddr[25:4]: base address. Read-only. 3 2:1 0 Pref: prefetchable. Read-only. 1=Prefetchable memory region. Type: base address register type. Read-only. 00b=32-bit BAR. 10b=64-bit BAR. MemSpace: memory space type. Read-only. 0=Memory mapped base address. IF (D0F0x64_x1C[F064BarEn]==0) THEN D1F0x14 Graphics IO Base Address Reset: 0000_0001h. Bits Description 31:8 BaseAddr: base address. IF (D0F0x64_x1C[IoBarDis]==0) THEN Read-write. ELSE Read-only. ENDIF. 269 42300 Rev 3.12 - July 14, 2015 7:1 0 BKDG for AMD Family 15h Models 10h-1Fh Processors Reserved. MemSpace: memory space type. Read-only. 1=IO mapped base address. ELSE D1F0x14 Graphics Memory Base Address 64 Reset: 0000_0000h. Bits Description 31:0 BaseAddr[63:32]: base address. Read-write. ENDIF. D1F0x18 Graphics Memory Mapped Registers Base Address IF (D0F0x64_x1C[F064BarEn]==0) THEN Reset: 0000_0000h. ELSE Reset: 0000_0004h. ENDIF. Bits Description 31:18 BaseAddr[31:18]: base address. Read-write. 17:16 BaseAddr[17:16]: base address. IF (D0F0x64_x1C[RegApSize]==0) THEN Read-write. ELSE Read-only. ENDIF. 15:4 BaseAddr[15:4]: base address. Read-only. 3 2:1 0 Pref: prefetchable. Read-only. 0=Non-prefetchable memory region. Type: base address register type. Read-only. 00b=32-bit BAR. 10b=64-bit BAR. MemSpace: memory space type. Read-only. 0=Memory mapped base address. IF (D0F0x64_x1C[F064BarEn]==0) THEN D1F0x1C Base Address 3 Reset: 0000_0000h. Bits Description 31:0 Reserved. ELSE D1F0x1C Graphics Memory Mapped Registers Address 64 Reset: 0000_0000h. Bits Description 31:0 BaseAddr[63:32]: base address. Read-write. ENDIF. IF (D0F0x64_x1C[F064BarEn]==0) THEN 270 42300 Rev 3.12 - July 14, 2015 BKDG for AMD Family 15h Models 10h-1Fh Processors D1F0x20 Base Address 4 Reset: 0000_0000h. Bits Description 31:0 Reserved. ELSE D1F0x20 Graphics IO Base Address Reset: 0000_0000h. Bits Description 31:8 BaseAddr: base address. IF (D0F0x64_x1C[IoBarDis]==0) THEN Read-write. ELSE Read-only. ENDIF. 7:1 0 Reserved. MemSpace: memory space type. Read-only. 1=IO mapped base address. ENDIF. D1F0x24 Base Address 5 Reset: 0000_0000h. Bits Description 31:0 Reserved. D1F0x2C Subsystem and Subvendor ID Register Reset: 0000_0000h. This register can be modified through D1F0x4C Bits Description 31:16 SubsystemID. Read-only. 15:0 SubsystemVendorID. Read-only. D1F0x30 Expansion ROM Base Address Reset: 0000_0000h. Bits Description 31:0 Reserved. 271 42300 Rev 3.12 - July 14, 2015 BKDG for AMD Family 15h Models 10h-1Fh Processors D1F0x34 Capabilities Pointer Reset: 0000_0050h. Bits Description 31:8 Reserved. 7:0 CapPtr: capabilities pointer. Read-only. Pointer to PM capability. D1F0x3C Interrupt Line Reset: 0000_01FFh. Bits Description 31:16 Reserved. 15:8 InterruptPin: interrupt pin. Read-only. This field identifies the legacy interrupt message the function uses. 7:0 InterruptLine: interrupt line. Read-write. This field contains the interrupt line routing information. D1F0x4C Subsystem and Subvendor ID Mirror Reset: 0000_0000h. Bits Description 31:16 SubsystemID. Read-write. This field sets the value in the corresponding field in D1F0x2C. 15:0 SubsystemVendorID. Read-write. This field sets the value in the corresponding field in D1F0x2C. D1F0x50 Power Management Capability Bits Description 31:27 PmeSupport. Value: 0_0000b. Indicates that there is no PME support. 26 D2Support: D2 support. Value: 1. D2 is supported 25 D1Support: D1 support. Value: 1. D1 is supported 24:22 AuxCurrent: auxiliary current. Value: 0. 21 DevSpecificInit: device specific initialization. Value: 0. Indicates that there is no device specific initialization necessary. 20 Reserved. 19 PmeClock. Value: 0. 18:16 Version: version. Value: 011b. 15:8 NextPtr: next pointer. Value: 58h. 7:0 CapID: capability ID.Value: 01h. Indicates that the capability structure is a PCI power management data structure. 272 42300 Rev 3.12 - July 14, 2015 BKDG for AMD Family 15h Models 10h-1Fh Processors D1F0x54 Power Management Control and Status Reset: 0000_0000h. Bits Description 31:24 PmeData. Read-only. 23 BusPwrEn. Read-only. 22 B2B3Support. Read-only. B states are not supported. 21:16 Reserved. 15 PmeStatus: PME status. Read-only. 14:13 DataScale: data scale. Read-only. 12:9 DataSelect: data select. Read-only. 8 7:4 PmeEn: PME# enable. Read-only. Reserved. 3 NoSoftReset: no soft reset. Read-only. Software is required to re-initialize the function when returning from D3hot. 2 Reserved. 1:0 PowerState: power state. Read-write. This 2-bit field is used both to determine the current power state of the root port and to set the root port into a new power state. Bits Definition 00b D0 10b-01b Reserved. 11b D3hot D1F0x58 PCI Express Capability Bits Description 31:30 Reserved. 29:25 IntMessageNum: interrupt message number. Value: 0. This field indicates which MSI vector is used for the interrupt message. 24 SlotImplemented: Slot implemented. Value: 0. 23:20 DeviceType: device type. Value: IF (D0F0x64_x1C[RcieEn]==1) THEN 9 ELSEIF (D0F0x64_x1C[F0NonlegacyDeviceTypeEn]==0) THEN 1 ELSE 0 ENDIF. 19:16 Version. Value: 2h. 15:8 NextPtr: next pointer. Value: IF (D0F0x64_x1C[MsiDis]==0) THEN A0h ELSE 00h ENDIF. 7:0 CapID: capability ID. Value: 10h. D1F0x5C Device Capability Bits Description 31:29 Reserved. 28 FlrCapable: function level reset capability. Value: 0. 273 42300 Rev 3.12 - July 14, 2015 BKDG for AMD Family 15h Models 10h-1Fh Processors 27:26 CapturedSlotPowerScale: captured slot power limit scale. Value: 0. 25:18 CapturedSlotPowerLimit: captured slot power limit value. Value: 0. 17:16 Reserved. 15 RoleBasedErrReporting: role-based error reporting. Value: 1. 14:12 Reserved. 11:9 L1AcceptableLatency: endpoint L1 Acceptable Latency. Value: 111b. 8:6 5 L0SAcceptableLatency: endpoint L0s Acceptable Latency. Value: 110b. ExtendedTag: extended tag support. Value: 1. 8 bit tag support. 4:3 PhantomFunc: phantom function support. Value: 0. No phantom functions supported. 2:0 MaxPayloadSupport: maximum supported payload size. Value: 000b. 128 bytes max payload size. D1F0x60 Device Control and Status Reset: 0000_0810h. Bits Description 31:22 Reserved. 21 TransactionsPending: transactions pending. Read-only. 20 AuxPwr: auxiliary power. Read-only. 19 UsrDetected: unsupported request detected. Read; Write-1-to-clear. 1=Unsupported request received. 18 FatalErr: fatal error detected. Read; Write-1-to-clear. 1=Fatal error detected. 17 NonFatalErr: non-fatal error detected. Read; Write-1-to-clear. 1=Non-fatal error detected. 16 CorrErr: correctable error detected. Read; Write-1-to-clear. 1=Correctable error detected. 15 BridgeCfgRetryEn: bridge configuration retry enable. Read-only. 14:12 MaxRequestSize: maximum request size. Read-only. 11 NoSnoopEnable: enable no snoop. Read-write. 1=The device is permitted to set the No Snoop bit in requests. 10 AuxPowerPmEn: auxiliary power PM enable. Read-only. This capability is not implemented. 9 PhantomFuncEn: phantom functions enable. Read-only. Phantom functions are not supported. 8 ExtendedTagEn: extended tag enable. Read-write. 1=8-bit tag request tags. 0=5-bit request tag. 7:5 MaxPayloadSize: maximum supported payload size. Read-only. 000b=Indicates a 128 byte maximum payload size. 4 RelaxedOrdEn: relaxed ordering enable. Read-write. 1=The device is permitted to set the Relaxed Ordering bit. 3 UsrReportEn: unsupported request reporting enable. Read-write. 1=Enables signaling unsupported requests by sending error messages. 2 FatalErrEn: fatal error reporting enable. Read-write. 1=Enables sending ERR_FATAL message when a fatal error is detected. 274 42300 Rev 3.12 - July 14, 2015 BKDG for AMD Family 15h Models 10h-1Fh Processors 1 NonFatalErrEn: non-fatal error reporting enable. Read-write. 1=Enables sending ERR_NONFATAL message when a non-fatal error is detected. 0 CorrErrEn: correctable error reporting enable. Read-write. 1=Enables sending ERR_CORR message when a correctable error is detected. D1F0x64 Link Capability Bits Description 31:24 PortNumber: port number. Value: 0.This field indicates the PCI Express port number for the given PCI Express link. 23:22 Reserved. 21 LinkBWNotificationCap: link bandwidth notification capability. Value: 0b. 20 DlActiveReportingCapable: data link layer active reporting capability. Value: 0b. 19 SurpriseDownErrReporting: surprise down error reporting capability. Value: 0b. 18 ClockPowerManagement: clock power management. Value: 0b. 17:15 L1ExitLatency: L1 exit latency.Value: 010b. 14:12 L0sExitLatency: L0s exit latency. Value: 001b. 11:10 PMSupport: active state power management support. Value: 11b. 9:4 LinkWidth: maximum link width. Value: 0. 3:0 LinkSpeed: link speed. Value: 1. D1F0x68 Link Control and Status Reset: 1001_0000h. Bits Description 31 LinkAutonomousBWStatus: link autonomous bandwidth status. Read-only. 30 LinkBWManagementStatus: link bandwidth management status. Read-only. 29 DlActive: data link layer link active. Read-only. This bit indicates the status of the data link control and management state machine. Reads return a 1 to indicate the DL_Active state, otherwise 0 is returned. 28 SlotClockCfg: slot clock configuration. Read-only. 1=the root port uses the same clock that the platform provides. 27 LinkTraining: link training. Read-only. 1=Indicates that the physical layer link training state machine is in the configuration or recovery state, or that 1b was written to the RetrainLink bit but link training has not yet begun. Hardware clears this bit when the link training state machine exits the configuration/recovery state. 26 Reserved. 25:20 NegotiatedLinkWidth: negotiated link width. Read-only. This field indicates the negotiated width of the given PCI Express link. 19:16 LinkSpeed: link speed. Read-only. 15:12 Reserved. 11 LinkAutonomousBWIntEn: link autonomous bandwidth interrupt enable. Read-only. 275 42300 Rev 3.12 - July 14, 2015 BKDG for AMD Family 15h Models 10h-1Fh Processors 10 LinkBWManagementEn: link bandwidth management interrupt enable. Read-only. 9 HWAutonomousWidthDisable: hardware autonomous width disable. Read-write. 1=Hardware not allowed to change the link width except to correct unreliable link operation by reducing link width. 8 ClockPowerManagementEn: clock power management enable. Read-write. 7 ExtendedSync: extended sync. Read-write. 1=Forces the transmission of additional ordered sets when exiting the L0s state and when in the recovery state. 6 CommonClockCfg: common clock configuration. Read-write. 1=Indicates that the root port and the component at the opposite end of this Link are operating with a distributed common reference clock. 0=Indicates that the upstream port and the component at the opposite end of this Link are operating with asynchronous reference clock. 5 RetrainLink: retrain link. Read-only. This bit does not apply to endpoints. 4 LinkDis: link disable. Read-only. This bit does not apply to endpoints. 3 ReadCplBoundary: read completion boundary. Read-only. 0=64 byte read completion boundary. 2 Reserved. 1:0 PmControl: active state power management enable. Read-write. This field controls the level of ASPM supported on the given PCI Express link. Bits Definition Bits Definition 00b Disabled. 10b L1 Entry Enabled. 01b L0s Entry Enabled. 11b L0s and L1 Entry Enabled. D1F0x7C Device Capability 2 Reset: 0000_0000h. Bits Description 31:5 Reserved. 4 3:0 CplTimeoutDisSup: completion timeout disable supported. Read-only. CplTimeoutRangeSup: completion timeout range supported. Read-only. D1F0x80 Device Control and Status 2 Reset: 0000_0000h. Bits Description 31:5 Reserved. 4 3:0 CplTimeoutDis: completion timeout disable. Read-only. CplTimeoutValue: completion timeout range supported. Read-only. D1F0x88 Link Control and Status 2 Reset: 0000_0001h. Bits Description 31:17 Reserved. 276 42300 Rev 3.12 - July 14, 2015 16 BKDG for AMD Family 15h Models 10h-1Fh Processors CurDeemphasisLevel: current deemphasis level. Read-only. 1=-3.5 dB. 0=-6 dB. 15:13 Reserved. 12 ComplianceDeemphasis: compliance deemphasis. Read-write. This bit defines the deemphasis level used in compliance mode. 1=-3.5 dB. 0=-6 dB. 11 ComplianceSOS: compliance SOS. Read-write. 1=The device transmits skip ordered sets in between the modified compliance pattern. 10 EnterModCompliance: enter modified compliance. Read-write. 1=The device transmits modified compliance pattern. 9:7 XmitMargin: transmit margin. Read-write. This field controls the non-deemphasized voltage level at the transmitter pins. 6 SelectableDeemphasis: selectable deemphasis. Read-only. 5 HwAutonomousSpeedDisable: hardware autonomous speed disable. Read-write. 1=Disables hardware generated link speed changes. 4 EnterCompliance: enter compliance. Read-write. 1=Force link to enter compliance mode. 3:0 TargetLinkSpeed: target link speed. Read-write. This field defines the upper limit of the link operational speed. D1F0xA0 MSI Capability Bits Description 31:24 Reserved. 23 Msi64bit: MSI 64 bit capability. Read-only. Value: D0F0x64_x46[Msi64bitEn]. 1=The device is capable of sending 64-bit MSI messages. 0=The device is not capable of sending a 64-bit message address. 22:20 MsiMultiEn: MSI multiple message enable. Read-write. Reset: 000b. Software writes to this field to indicate the number of allocated vectors (equal to or less than the number of requested vectors). When MSI is enabled, a function is allocated at least 1 vector. 19:17 MsiMultiCap: MSI multiple message capability. Read-only. Reset: 000b. 000b=The device is requesting one vector. 16 MsiEn: MSI enable. Read-write. Reset: 0. 1=MSI generation is enabled and INTx generation is disabled. 0=MSI generation disabled and INTx generation is enabled. 15:8 NextPtr: next pointer. Read-only. Reset:00h 7:0 CapID: capability ID. Read-only. Reset: 05h. 05h=MSI capability structure. D1F0xA4 MSI Message Address Low Reset: 0000_0000h. Bits Description 31:2 MsiMsgAddrLo: MSI message address. Read-write. This register specifies the dword aligned address for the MSI memory write transaction. 1:0 Reserved. 277 42300 Rev 3.12 - July 14, 2015 BKDG for AMD Family 15h Models 10h-1Fh Processors IF (D0F0x64_x46[Msi64bitEn]==0) THEN D1F0xA8 MSI Message Data Reset: 0000_0000h. Bits Description 31:16 Reserved. 15:0 MsiData: MSI message data. Read-write. This register specifies lower 16 bits of data for the MSI memory write transaction. The upper 16 bits are always 0. ELSE D1F0xA8 MSI Message Address High Reset: 0000_0000h. Bits Description 31:8 Reserved. 7:0 MsiMsgAddrHi: MSI message address. Read-write. This register specifies the upper 8-bits of the MSI address in 64 bit MSI mode. D1F0xAC MSI Message Data Reset: 0000_0000h. Bits Description 31:16 Reserved. 15:0 MsiData: MSI message data. Read-write. This register specifies lower 16 bits of data for the MSI memory write transaction. The upper 16 bits are always 0. ENDIF. D1F0x100 Vendor Specific Enhanced Capability Reset: 0061_000Bh. Bits Description 31:20 NextPtr: next pointer. Read-only. 19:16 CapVer: capability version. Read-only. 15:0 CapID: capability ID. Read-only. D1F0x104 Vendor Specific Header Reset: 0101_0001h. Bits Description 31:20 VsecLen: vendor specific enhanced next pointer. Read-only. 278 42300 Rev 3.12 - July 14, 2015 BKDG for AMD Family 15h Models 10h-1Fh Processors 19:16 VsecRev: vendor specific enhanced capability version. Read-only. 15:0 VsecID: vendor specific enhanced capability ID. Read-only. D1F0x108 Vendor Specific 1 Reset: 0000_0000h. Bits Description 31:0 Scratch: scratch. Read-write. D1F0x10C Vendor Specific 2 Reset: 0000_0000h. Bits Description 31:0 Scratch: scratch. Read-write. 279 42300 Rev 3.12 - July 14, 2015 3.6 BKDG for AMD Family 15h Models 10h-1Fh Processors Device 1 Function 1 (Audio Controller) Configuration Registers See 3.1 [Register Descriptions and Mnemonics] for a description of the register naming convention. See 2.7 [Configuration Space] for details about how to access this space. D1F1x00 Device/Vendor ID Bits Description 31:16 DeviceID: device ID. Read-only. Reset: 9902h. 15:0 VendorID: vendor ID. Read-only. Reset: 1002h. D1F1x04 Status/Command Reset: 0010_0000h. Bits Description 31 ParityErrorDetected: detected parity error. Read; Write-1-to-clear. 1=Poisoned TLP received. 30 SignaledSystemError: signaled system error. Read; Write-1-to-clear. 1=A non-fatal or fatal error message was sent and SerrEn=1. 29 ReceivedMasterAbort: received master abort. Read; Write-1-to-clear. 1=A completion with an unsupported request completion status was received. 28 ReceivedTargetAbort: received target abort. Read; Write-1-to-clear. 1=A completion with completer abort completion status was received. 27 SignalTargetAbort: Signaled target abort. Read-only. 26:25 DevselTiming: DEVSEL# Timing. Read-only. 24 MasterDataPerr: master data parity error. Read; Write-1-to-clear. 1=ParityErrorEn=1 and either a poisoned completion was received or the device poisoned a write request. 23 FastBackCapable: fast back-to-back capable. Read-only. 22 UDFEn: UDF enable. Read-only. 21 PCI66En: 66 MHz capable. Read-only. 20 CapList: capability list. Read-only. 1=capability list supported. 19 IntStatus: interrupt status. Read-only. 1=INTx interrupt message pending. 18:11 Reserved. 10 IntDis: interrupt disable. Read-write. 1=INTx interrupt messages generation disabled. 9 FastB2BEn: fast back-to-back enable. Read-only. 8 SerrEn: System error enable. Read-write. 1=Enables reporting of non-fatal and fatal errors detected. 7 Stepping: Stepping control. Read-only. 6 ParityErrorEn: parity error response enable. Read-write. 5 PalSnoopEn: VGA palette snoop enable. Read-only. 4 MemWriteInvalidateEn: memory write and invalidate enable. Read-only. 3 SpecialCycleEn: special cycle enable. Read-only. 280 42300 Rev 3.12 - July 14, 2015 BKDG for AMD Family 15h Models 10h-1Fh Processors 2 BusMasterEn: bus master enable. Read-write. 1=Memory and IO read and write request generation enabled. 1 MemAccessEn: IO access enable. Read-write. This bit controls if memory accesses targeting this device are accepted. 1=Enabled. 0=Disabled. 0 IoAccessEn: IO access enable. Read-write. This bit controls if IO accesses targeting this device are accepted. 1=Enabled. 0=Disabled. D1F1x08 Class Code/Revision ID Reset: 0403_0000h. Bits Description 31:8 ClassCode. Read-only. 7:0 RevID: revision ID. Read-only. D1F1x0C Header Type Reset: 0080_0000h. Bits Description 31:24 BIST. Read-only. These bits are fixed at their default values. 23:16 HeaderTypeReg. Read-only. 80h=Type 0 multi-function device. 15:8 LatencyTimer. Read-only. These bits are fixed at their default value. 7:0 CacheLineSize. Read-write.This field specifies the system cache line size in units of double words. D1F1x10 Audio Registers Base Address Reset: 0000_0000h. Bits Description 31:14 BaseAddr: base address. Read-write. 13:4 Reserved. 3 Pref: prefetchable. Read-only. 0=Non-prefetchable memory region. 2:1 Type: base address register type. Read-only. 00b=32-bit base address register. 0 MemSpace: memory space type. Read-only. 0=Memory mapped base address. D1F1x14 Base Address 1 Reset: 0000_0000h. Bits Description 31:0 Reserved. 281 42300 Rev 3.12 - July 14, 2015 BKDG for AMD Family 15h Models 10h-1Fh Processors D1F1x18 Base Address 2 Reset: 0000_0000h. Bits Description 31:0 Reserved. D1F1x1C Base Address 3 Reset: 0000_0000h. Bits Description 31:0 Reserved. D1F1x20 Base Address 4 Reset: 0000_0000h. Bits Description 31:0 Reserved. D1F1x24 Base Address 5 Reset: 0000_0000h. Bits Description 31:0 Reserved. D1F1x2C Subsystem and Subvendor ID Reset: 0000_0000h. This register can be modified through D1F1x4C. Bits Description 31:16 SubsystemID. Read-only. 15:0 SubsystemVendorID. Read-only. D1F1x30 Expansion ROM Base Address Reset: 0000_0000h. Bits Description 31:0 Reserved. 282 42300 Rev 3.12 - July 14, 2015 BKDG for AMD Family 15h Models 10h-1Fh Processors D1F1x34 Capabilities Pointer Reset: 0000_0050h. Bits Description 31:8 Reserved. 7:0 CapPtr: capabilities pointer. Read-only. Pointer to PM capability. D1F1x3C Interrupt Line Reset: 0000_02FFh. Bits Description 31:16 Reserved. 15:8 InterruptPin: interrupt pin. Read-only. This field identifies the legacy interrupt message the function uses. 7:0 InterruptLine: interrupt line. Read-write. This field contains the interrupt line routing information. D1F1x4C Subsystem and Subvendor ID Mirror Reset: 0000_0000h. Bits Description 31:16 SubsystemID. Read-write. This field sets the value in the corresponding field in D1F1x2C. 15:0 SubsystemVendorID. Read-write. This field sets the value in the corresponding field in D1F1x2C. D1F1x50 Power Management Capability Bits Description 31:27 PmeSupport. Value: 0_0000b. Indicates that there is no PME support. 26 D2Support: D2 support. Value: 1. D2 is supported 25 D1Support: D1 support. Value: 1. D1 is supported 24:22 AuxCurrent: auxiliary current. Value: 0. 21 DevSpecificInit: device specific initialization. Value: 0. Indicates that there is no device specific initialization necessary. 20 Reserved. 19 PmeClock. Value: 0. 18:16 Version: version. Value: 011b. 15:8 NextPtr: next pointer. Value: 58h. 7:0 CapID: capability ID.Value: 01h. Indicates that the capability structure is a PCI power management data structure. 283 42300 Rev 3.12 - July 14, 2015 BKDG for AMD Family 15h Models 10h-1Fh Processors D1F1x54 Power Management Control and Status Reset: 0000_0000h. Bits Description 31:24 PmeData. Read-only. 23 BusPwrEn. Read-only. 22 B2B3Support. Read-only. B states are not supported. 21:16 Reserved. 15 PmeStatus: PME status. Read-only. 14:13 DataScale: data scale. Read-only. 12:9 DataSelect: data select. Read-only. 8 7:4 PmeEn: PME# enable. Read-only. Reserved. 3 NoSoftReset: no soft reset. Read-only. Software is required to re-initialize the function when returning from D3hot. 2 Reserved. 1:0 PowerState: power state. Read-write. This 2-bit field is used both to determine the current power state of the root port and to set the root port into a new power state. Bits Definition 00b D0 10b-01b Reserved 11b D3hot D1F1x58 PCI Express Capability Bits Description 31:30 Reserved. 29:25 IntMessageNum: interrupt message number. Value: 0. This field indicates which MSI vector is used for the interrupt message. 24 SlotImplemented: Slot implemented. Value: 0. 23:20 DeviceType: device type. Value: IF (D0F0x64_x1C[AudioNonlegacyDeviceTypeEn]==0) THEN 1. ELSE 0. ENDIF. 19:16 Version. Value: 2h. 15:8 NextPtr: next pointer. Value: IF (D0F0x64_x1C[MsiDis]==0) THEN A0h. ELSE 00h. ENDIF. 7:0 CapID: capability ID. Value: 10h. D1F1x5C Device Capability Bits Description 31:29 Reserved. 28 FlrCapable: function level reset capability. Value: 0. 284 42300 Rev 3.12 - July 14, 2015 BKDG for AMD Family 15h Models 10h-1Fh Processors 27:26 CapturedSlotPowerScale: captured slot power limit scale. Value: 0. 25:18 CapturedSlotPowerLimit: captured slot power limit value. Value: 0. 17:16 Reserved. 15 RoleBasedErrReporting: role-based error reporting. Value: 1. 14:12 Reserved. 11:9 L1AcceptableLatency: endpoint L1 Acceptable Latency. Value: 111b. 8:6 5 L0SAcceptableLatency: endpoint L0s Acceptable Latency. Value: 110b. ExtendedTag: extended tag support. Value: 1. 8 bit tag support. 4:3 PhantomFunc: phantom function support. Value: 0. No phantom functions supported. 2:0 MaxPayloadSupport: maximum supported payload size. Value: 000b. 128 bytes max payload size. D1F1x60 Device Control and Status Reset: 0000_0810h. Bits Description 31:22 Reserved. 21 TransactionsPending: transactions pending. Read-only. 20 AuxPwr: auxiliary power. Read-only. 19 UsrDetected: unsupported request detected. Read; Write-1-to-clear. 1=Unsupported request received. 18 FatalErr: fatal error detected. Read; Write-1-to-clear. 1=Fatal error detected. 17 NonFatalErr: non-fatal error detected. Read; Write-1-to-clear. 1=Non-fatal error detected. 16 CorrErr: correctable error detected. Read; Write-1-to-clear. 1=Correctable error detected. 15 BridgeCfgRetryEn: bridge configuration retry enable. Read-only. 14:12 MaxRequestSize: maximum request size. Read-only. 0=The root port never generates read requests with size exceeding 128 bytes. 11 NoSnoopEnable: enable no snoop. Read-write. 1=the device is permitted to set the No Snoop bit in requests. 10 AuxPowerPmEn: auxiliary power PM enable. Read-only. This capability is not implemented. 9 PhantomFuncEn: phantom functions enable. Read-only. Phantom functions are not supported. 8 ExtendedTagEn: extended tag enable. Read-write. 1=8-bit tag request tags. 0=5-bit request tag. 7:5 MaxPayloadSize: maximum supported payload size. Read-only. 000b=Indicates a 128 byte maximum payload size. 4 RelaxedOrdEn: relaxed ordering enable. Read-write. 1=The device is permitted to set the Relaxed Ordering bit. 3 UsrReportEn: unsupported request reporting enable. Read-write. 1=Enables signaling unsupported requests by sending error messages. 2 FatalErrEn: fatal error reporting enable. Read-write. 1=Enables sending ERR_FATAL message when a fatal error is detected. 285 42300 Rev 3.12 - July 14, 2015 BKDG for AMD Family 15h Models 10h-1Fh Processors 1 NonFatalErrEn: non-fatal error reporting enable. Read-write. 1=Enables sending ERR_NONFATAL message when a non-fatal error is detected. 0 CorrErrEn: correctable error reporting enable. Read-write. 1=Enables sending ERR_CORR message when a correctable error is detected. D1F1x64 Link Capability Bits Description 31:24 PortNumber: port number. Value: 0.This field indicates the PCI Express port number for the given PCI Express link. 23:22 Reserved. 21 LinkBWNotificationCap: link bandwidth notification capability. Value: 0b. 20 DlActiveReportingCapable: data link layer active reporting capability. Value: 0b. 19 SurpriseDownErrReporting: surprise down error reporting capability. Value: 0b. 18 ClockPowerManagement: clock power management. Value: 0b. 17:15 L1ExitLatency: L1 exit latency.Value: 010b. 14:12 L0sExitLatency: L0s exit latency. Value: 001b. 11:10 PMSupport: active state power management support. Value: 11b. 9:4 LinkWidth: maximum link width. Value: 0. 3:0 LinkSpeed: link speed.Value: 1. D1F1x68 Link Control and Status Reset: 1001_0000h. Bits Description 31 LinkAutonomousBWStatus: link autonomous bandwidth status. Read-only. 30 LinkBWManagementStatus: link bandwidth management status. Read-only. 29 DlActive: data link layer link active. Read-only. This bit indicates the status of the data link control and management state machine. Reads return a 1 to indicate the DL_Active state, otherwise 0 is returned. 28 SlotClockCfg: slot clock configuration. Read-only. 1=the root port uses the same clock that the platform provides. 27 LinkTraining: link training. Read-only. 1=Indicates that the physical layer link training state machine is in the configuration or recovery state, or that 1b was written to the RetrainLink bit but link training has not yet begun. Hardware clears this bit when the link training state machine exits the configuration/recovery state. 26 Reserved. 25:20 NegotiatedLinkWidth: negotiated link width. Read-only. This field indicates the negotiated width of the given PCI Express link. 19:16 LinkSpeed: link speed. Read-only. 0001b: 2.5 Gb/s. 0010b: 5 Gb/s 286 42300 Rev 3.12 - July 14, 2015 BKDG for AMD Family 15h Models 10h-1Fh Processors 15:12 Reserved. 11 LinkAutonomousBWIntEn: link autonomous bandwidth interrupt enable. Read-only. 10 LinkBWManagementEn: link bandwidth management interrupt enable. Read-only. 9 HWAutonomousWidthDisable: hardware autonomous width disable. Read-write. 1=Hardware not allowed to change the link width except to correct unreliable link operation by reducing link width. 8 ClockPowerManagementEn: clock power management enable. Read-write. 7 ExtendedSync: extended sync. Read-write. 1=Forces the transmission of additional ordered sets when exiting the L0s state and when in the recovery state. 6 CommonClockCfg: common clock configuration. Read-write. 1=Indicates that the root port and the component at the opposite end of this Link are operating with a distributed common reference clock. 0=Indicates that the upstream port and the component at the opposite end of this Link are operating with asynchronous reference clock. 5 RetrainLink: retrain link. Read-only. This bit does not apply to endpoints. 4 LinkDis: link disable. Read-only. This bit does not apply to endpoints. 3 ReadCplBoundary: read completion boundary. Read-only. 0=64 byte read completion boundary. 2 Reserved. 1:0 PmControl: active state power management enable. Read-write. This field controls the level of ASPM supported on the given PCI Express link. Bits Definition Bits Definition 00b Disabled. 10b L1 Entry Enabled. 01b L0s Entry Enabled. 11b L0s and L1 Entry Enabled. D1F1x7C Device Capability 2 Reset: 0000_0000h. Bits Description 31:5 Reserved. 4 3:0 CplTimeoutDisSup: completion timeout disable supported. Read-only. CplTimeoutRangeSup: completion timeout range supported. Read-only. D1F1x80 Device Control and Status 2 Reset: 0000_0000h. Bits Description 31:5 Reserved. 4 3:0 CplTimeoutDis: completion timeout disable. Read-only. CplTimeoutValue: completion timeout range supported. Read-only. 287 42300 Rev 3.12 - July 14, 2015 BKDG for AMD Family 15h Models 10h-1Fh Processors D1F1x84 Link Capability 2 Bits Description 31:0 Reserved. D1F1x88 Link Control and Status 2 Reset: 0000_0001h. Bits Description 31:17 Reserved. 16 CurDeemphasisLevel: current deemphasis level. Read-only. Bit Description 1 -3.5 dB 0 -6 dB 15:13 Reserved. 12 ComplianceDeemphasis: compliance deemphasis. Read-write. This bit defines the deemphasis level used in compliance mode. Bit Description 1 -3.5 dB 0 -6 dB 11 ComplianceSOS: compliance SOS. Read-write. 1=The device transmits skip ordered sets in between the modified compliance pattern. 10 EnterModCompliance: enter modified compliance. Read-write. 1=The device transmits modified compliance pattern. 9:7 XmitMargin: transmit margin. Read-write. This field controls the non-deemphasized voltage level at the transmitter pins. 6 SelectableDeemphasis: selectable deemphasis. Read-only. 5 HwAutonomousSpeedDisable: hardware autonomous speed disable. Read-write. 1=Disables hardware generated link speed changes. 4 EnterCompliance: enter compliance. Read-write. 1=Force link to enter compliance mode. 3:0 TargetLinkSpeed: target link speed. Read-write. This fields defines the upper limit of the link operational speed. D1F1xA0 MSI Capability Bits Description 31:24 Reserved. 23 Msi64bit: MSI 64 bit capability. Read-only. Value: D0F0x64_x46[Msi64bitEn]. 1=The device is capable of sending 64-bit MSI messages. 0=The device is not capable of sending a 64-bit message address. 22:20 MsiMultiEn: MSI multiple message enable. Read-write. Reset: 000b. Software writes to this field to indicate the number of allocated vectors (equal to or less than the number of requested vectors). When MSI is enabled, a function is allocated at least 1 vector. 288 42300 Rev 3.12 - July 14, 2015 BKDG for AMD Family 15h Models 10h-1Fh Processors 19:17 MsiMultiCap: MSI multiple message capability. Read-only. Reset:000b. 000b=The device is requesting one vector. 16 MsiEn: MSI enable. Read-write. Reset: 0. 1=MSI generation is enabled and INTx generation is disabled. 0=MSI generation disabled and INTx generation is enabled. 15:8 NextPtr: next pointer. Read-only. Reset: 00h 7:0 CapID: capability ID. Read-only. Reset: 05h. 05h=MSI capability structure. D1F1xA4 MSI Message Address Low Reset: 0000_0000h. Bits Description 31:2 MsiMsgAddrLo: MSI message address. Read-write. This register specifies the dword aligned address for the MSI memory write transaction. 1:0 Reserved. IF (D0F0x64_x46[Msi64bitEn]==0) THEN D1F1xA8 MSI Message Data Reset: 0000_0000h. Bits Description 31:16 Reserved. 15:0 MsiData: MSI message data. Read-write. This register specifies lower 16 bits of data for the MSI memory write transaction. The upper 16 bits are always 0. ELSE D1F1xA8 MSI Message Address High Reset: 0000_0000h. Bits Description 31:8 Reserved. 7:0 MsiMsgAddrHi: MSI message address. Read-write. This register specifies the upper 8-bits of the MSI address in 64 bit MSI mode. D1F1xAC MSI Message Data Reset: 0000_0000h. Bits Description 31:16 Reserved. 15:0 MsiData: MSI message data. Read-write. This register specifies lower 16 bits of data for the MSI memory write transaction. The upper 16 bits are always 0. ENDIF. 289 42300 Rev 3.12 - July 14, 2015 BKDG for AMD Family 15h Models 10h-1Fh Processors D1F1x100 Vendor Specific Enhanced Capability Reset: 0111_000Bh. Bits Description 31:20 NextPtr: next pointer. Read-only. 19:16 CapVer: capability version. Read-only. 15:0 CapID: capability ID. Read-only. D1F1x104 Vendor Specific Header Reset: 0101_0001h. Bits Description 31:20 VsecLen: vendor specific enhanced next pointer. Read-only. 19:16 VsecRev: vendor specific enhanced capability version. Read-only. 15:0 VsecID: vendor specific enhanced capability ID. Read-only. D1F1x108 Vendor Specific 1 Reset: 0000_0000h. Bits Description 31:0 Scratch: scratch. Read-write. D1F1x10C Vendor Specific 2 Reset: 0000_0000h. Bits Description 31:0 Scratch: scratch. Read-write. 290 42300 Rev 3.12 - July 14, 2015 3.7 BKDG for AMD Family 15h Models 10h-1Fh Processors Device [8:2] Function 0 (Root Port) Configuration Registers See 3.1 [Register Descriptions and Mnemonics] for a description of the register naming convention. See 2.7 [Configuration Space] for details about how to access this space. D[8:2]F0x00 Device/Vendor ID Register Table 122: Reset mapping for D[8:2]F0x00. Register D2F0x00 D3F0x00 D4F0x00 D5F0x00 Bits Reset 1412_1022h. 1413_1022h. 1414_1022h. 1415_1022h. Register D6F0x00 D7F0x00 D8F0x00 - Reset 1416_1022h. 1417_1022h. 1418_1022h. - Description 31:16 DeviceID: device ID. Read-only. 15:0 VendorID: vendor ID. Read-only. D[8:2]F0x04 Status/Command Register Reset: 0010_0000h. Bits Description 31 ParityErrorDetected: detected parity error. Read; Write-1-to-clear; updated-by-hardware. 30 SignaledSystemError: signaled system error. Read; Write-1-to-clear; updated-by-hardware. 1=System error signalled. 29 ReceivedMasterAbort: received master abort. Read; Write-1-to-clear; updated-by-hardware. 28 ReceivedTargetAbort: received target abort. Read; Write-1-to-clear; updated-by-hardware. 27 SignalTargetAbort: signaled target abort. Read; Write-1-to-clear; updated-by-hardware. 26:25 DevselTiming: DEVSEL# Timing. Read-only. 24 DataPerr: data parity error. Read; Write-1-to-clear; updated-by-hardware. 23 FastBackCapable: fast back-to-back capable. Read-only. 22 UDFEn: UDF enable. Read-only. 21 PCI66En: 66 MHz capable. Read-only. 20 CapList: capability list. Read-only. 1= Capability list present. 19 IntStatus: interrupt status. Read-only. 1=An INTx interrupt Message is pending in the device. 18:11 Reserved. 10 IntDis: interrupt disable. Read-write. 9 FastB2BEn: fast back-to-back enable. Read-only. 8 SerrEn: system error enable. Read-write. 1=System error reporting enabled. 7 Stepping: Stepping control. Read-only. 6 ParityErrorEn: parity error response enable. Read-write. 5 PalSnoopEn: VGA palette snoop enable. Read-only. 291 42300 Rev 3.12 - July 14, 2015 BKDG for AMD Family 15h Models 10h-1Fh Processors 4 MemWriteInvalidateEn: memory write and invalidate enable. Read-only. 3 SpecialCycleEn: special cycle enable. Read-only. 2 BusMasterEn: bus master enable. Read-write. 1 MemAccessEn: memory access enable. Read-write. This bit controls if memory accesses targeting this device are accepted or not. 1=Enabled. 0=Disabled. 0 IoAccessEn: IO access enable. Read-write. This bit controls if IO accesses targeting this device are accepted or not. 1=Enabled. 0=Disabled. D[8:2]F0x08 Class Code/Revision ID Register Reset: 0604_00xxh. Bits Description 31:8 ClassCode. Read-only. Provides the host bridge class code as defined in the PCI specification. 7:0 RevID: revision ID. Read-only. D[8:2]F0x0C Header Type Register Reset: 0001_0000h. Bits Description 31:24 BIST. Read-only. These bits are fixed at their default values. 23:16 HeaderTypeReg. Read-only. These bits are fixed at their default values. The header type field indicates a header type 1 and that there is only one function present in this device. 15:8 LatencyTimer. Read-only. This field does not control any hardware. 7:0 CacheLineSize. Read-write. D[8:2]F0x18 Bus Number and Secondary Latency Register Reset: 0000_0000h. Bits Description 31:24 SecondaryLatencyTimer: secondary latency timer. Read-only. This field is always 0. 23:16 SubBusNumber: subordinate number. Read-write. This field contains the highest-numbered bus that exists on the secondary side of the bridge. 15:8 SecondaryBus: secondary bus number. Read-write. This field defines the bus number of the secondary bus interface. 7:0 PrimaryBus: primary bus number. Read-write. This field defines the bus number of the primary bus interface. 292 42300 Rev 3.12 - July 14, 2015 BKDG for AMD Family 15h Models 10h-1Fh Processors D[8:2]F0x1C IO Base and Secondary Status Register Reset: 0000_0101h. Bits Description 31 ParityErrorDetected: detected parity error. Read; Write-1-to-clear. A Poisoned TLP was received regardless of the state of the D[8:2]F0x04[ParityErrorEn]. 30 ReceivedSystemError: signaled system error. Read; Write-1-to-clear. 1=A System Error was detected. 29 ReceivedMasterAbort: received master abort. Read; Write-1-to-clear. 1=A CPU transaction is terminated due to a master-abort. 28 ReceivedTargetAbort: received target abort. Read; Write-1-to-clear. 1=A CPU transaction (except for a special cycle) is terminated due to a target-abort. 27 SignalTargetAbort: signaled target abort. Read; Write-1-to-clear. 26:25 DevselTiming: DEVSEL# Timing. Read-only. 24 MasterDataPerr: master data parity error. Read; Write-1-to-clear. 1=The link received a poisoned or poisoned a downstream write and D[8:2]F0x3C[ParityResponseEn]=1. 23 FastBackCapable: fast back-to-back capable. Read-only. 22 UDFEn: UDF enable. Read-only. 21 PCI66En: 66 MHz capable. Read-only. 20 CapList: capability list. Read-only. 19:16 Reserved. 15:12 IOLimit[15:12]. Read-write. Lower part of the limit address. Upper part is defined in D[8:2]F0x30. 11:8 Reserved. 7:4 IOBase[15:12]. Read-write. Lower part of the base address. Upper part is defined in D[8:2]F0x30. 3:0 Reserved. D[8:2]F0x20 Memory Limit and Base Register Reset: 0000_0000h. Bits Description 31:20 MemLimit. Read-write. 19:16 Reserved. 15:4 MemBase. Read-write. 3:0 Reserved. D[8:2]F0x24 Prefetchable Memory Limit and Base Register Reset: 0001_0001h. Bits Description 31:20 PrefMemLimit. Read-write. Lower part of the limit address. Upper part is defined in D[8:2]F0x2C. 19:16 PrefMemLimitR. Read-only. 1=64 bit memory address decoder. 293 42300 Rev 3.12 - July 14, 2015 BKDG for AMD Family 15h Models 10h-1Fh Processors 15:4 PrefMemBase[31:20]. Read-write. Lower part of the base address. Upper part is defined in D[8:2]F0x28. 3:0 PrefMemBaseR. Read-only. 1=64 bit memory address decoder. D[8:2]F0x28 Prefetchable Memory Base High Register Reset: 0000_0000h. Bits Description 31:0 PrefMemBase[63:32]. Read-write. Upper part of the base address. Lower part is defined in D[8:2]F0x24. D[8:2]F0x2C Prefetchable Memory Limit High Register Reset: 0000_0000h. Bits Description 31:0 PrefMemLimit[63:32]. Read-write. Upper part of the limit address. Lower part is defined in D[8:2]F0x24. D[8:2]F0x30 IO Base and Limit High Register Reset: 0000_0000h. Bits Description 31:16 IOLimit[31:16]. Read-write. Upper part of the limit address. Lower part is defined in D[8:2]F0x1C. 15:0 IOBase[31:16]. Read-write. Upper part of the base address. Lower part is defined in D[8:2]F0x1C. D[8:2]F0x34 Capabilities Pointer Register Reset: 0000_0050h. Bits Description 31:8 Reserved. 7:0 CapPtr: capabilities pointer. Read-only. Pointer to PM capability. D[8:2]F0x3C Bridge Control Register Reset: 0000_00FFh. Bits Description 31:24 Reserved. 23 FastB2BCap: Fast back-to-back capability. Read-only. 22 SecondaryBusReset: Secondary bus reset. Read-write. Setting this bit triggers a hot reset on the corresponding PCI Express Port. 21 MasterAbortMode: Master abort mode. Read-only. 294 42300 Rev 3.12 - July 14, 2015 BKDG for AMD Family 15h Models 10h-1Fh Processors 20 Vga16En: VGA IO 16 bit decoding enable. Read-write. 1= Address bits 15:10 for VGA IO cycles are decoded. 0=Address bits 15:10 for VGA IO cycles are ignored. 19 VgaEn: VGA enable. Read-write. Affects the response by the bridge to compatible VGA addresses. When it is set, the bridge decodes and forwards the following accesses on the primary interface to the secondary interface: Memory accesses in the range of A_0000h to B_FFFFh and IO address where address bits 9:0 are in the ranges of 3B0h to 3BBh or 3C0h to 3DFh. For IO cycles the decoding of address bits 15:10 depends on Vga16En. 18 IsaEn: ISA enable. Read-write. 17 SerrEn: SERR enable. Read-write. 16 ParityResponseEn: Parity response enable. Read-write. Controls the bridge's response to poisoned TLPs on its secondary interface. 1=The bridge takes its normal action when a poisoned TLP is received. 0=The bridge ignores any poisoned TLPs that it receives and continues normal operation. 15:11 IntPinR: interrupt pin. Read-only. 10:8 IntPin: interrupt pin. IF (D0F0xE4_x0[2:1]01_0010[HwInitWrLock]==1) THEN Read-only. ELSE Read-write. ENDIF. 7:0 IntLine: Interrupt line. Read-write. D[8:2]F0x50 Power Management Capability Register Reset: C803_5801h. Bits Description 31:27 PmeSupport. Read-only. 26 D2Support: D2 support. Read-only. D2 is not supported 25 D1Support: D1 support. Read-only. D1 is not supported 24:22 AuxCurrent: auxiliary current. Read-only. Auxiliary current is not supported. 21 DevSpecificInit: device specific initialization. Read-only. This field is hardwired to 0 to indicate that there is no device specific initialization necessary. 20 Reserved. 19 PmeClock. Read-only. 0=Indicate that PCI clock is not needed to generate PME messages. 18:16 Version: version. Read-only. 15:8 NextPtr: next pointer. Read-only. 58h=Address of the next capability structure. 7:0 CapID: capability ID. Read-only. 01h=PCI power management data structure. D[8:2]F0x54 Power Management Control and Status Register Bits Description 31:24 PmeData. Read-only. Reset: 0. 23 BusPwrEn. Read-only. Reset: 0. 22 B2B3Support. Read-only. Reset: 0. B states are not supported. 21:16 Reserved. 295 42300 Rev 3.12 - July 14, 2015 15 BKDG for AMD Family 15h Models 10h-1Fh Processors PmeStatus: PME status. Read; Write-1-to-clear. Reset: 0. This bit is set when the root port would issue a PME message (independent of the state of the PmeEn bit). Once set, this bit remains set until it is reset by writing a 1 to this bit location. Writing a 0 has no effect. 14:13 DataScale: data scale. Read-only. Reset: 0. 12:9 DataSelect: data select. Read-only. Reset: 0. 8 7:4 PmeEn: PME# enable. Read-write. Reset: 0. Reserved. 3 NoSoftReset: no soft reset. Read-only. Reset: 0. Software is required to re-initialize the function when returning from D3hot. 2 Reserved. 1:0 PowerState: power state. Read-write. Reset: 0. This 2-bit field is used both to determine the current power state of the root port and to set the root port into a new power state. Bits Definition Bits Definition 00b D0 10b Reserved 01b Reserved 11b D3 D[8:2]F0x58 PCI Express Capability Register Reset: 0042_A010h. Bits Description 31:30 Reserved. 29:25 IntMessageNum: interrupt message number. Read-only. This register indicates which MSI vector is used for the interrupt message. 24 SlotImplemented: Slot implemented. Read-only. 1=The IO Link associated with this port is connected to a slot. 23:20 DeviceType: device type. Read-only. 4h=Root complex. 19:16 Version. Read-only. 2h=GEN 2 compliant. 15:8 NextPtr: next pointer. Read-only. A0h=Pointer to the next capability structure. 7:0 CapID: capability ID. Read-only. 10h=PCIe® Capability structure. D[8:2]F0x5C Device Capability Register Reset: 0000_8020h. Bits Description 31:29 Reserved. 28 FlrCapable: function level reset capability. Read-only. 27:26 CapturedSlotPowerScale: captured slot power limit scale. Read-only. 25:18 CapturedSlotPowerLimit: captured slot power limit value. Read-only. 17:16 Reserved. 15 RoleBasedErrReporting: role-based error reporting. Read-only. 14:12 Reserved. 296 42300 Rev 3.12 - July 14, 2015 BKDG for AMD Family 15h Models 10h-1Fh Processors 11:9 L1AcceptableLatency: endpoint L1 Acceptable Latency. Read-only. 8:6 5 L0SAcceptableLatency: endpoint L0s Acceptable Latency. Read-only. ExtendedTag: extended tag support. Read-only. 1: 8 bit tag supported 0: 5 bit tag supported. 4:3 PhantomFunc: phantom function support. Read-only. 0=No phantom functions supported. 2:0 MaxPayloadSupport: maximum supported payload size. Read-only. 000b=128 bytes max payload size. D[8:2]F0x60 Device Control and Status Register Reset: 0000_2810h. Bits Description 31:22 Reserved. 21 TransactionsPending: transactions pending. Read-only. 0=No internally generated non-posted transactions pending. 20 AuxPwr: auxiliary power. Read-only. 19 UsrDetected: unsupported request detected. Read; Write-1-to-clear. 1=The port received an unsupported request. Errors are logged in this register even if error reporting is disabled. 18 FatalErr: fatal error detected. Read; Write-1-to-clear. 1=The port detected a fatal error. Errors are logged in this register even if error reporting is disabled. 17 NonFatalErr: non-fatal error detected. Read; Write-1-to-clear. T1=The port detected a non-fatal error. Errors are logged in this register even if error reporting is disabled. 16 CorrErr: correctable error detected. Read; Write-1-to-clear. 1=The port detected a correctable error. Errors are logged in this register even if error reporting is disabled. 15 BridgeCfgRetryEn: bridge configuration retry enable. Read-only. 14:12 MaxRequestSize: maximum request size. Read-write. 11 NoSnoopEnable: enable no snoop. Read-write. 1=The port is permitted to set the No Snoop bit in the Requester Attributes of transactions it initiates that do not require hardware enforced cache coherency. 10 AuxPowerPmEn: auxiliary power PM enable. Read-only. 9 PhantomFuncEn: phantom functions enable. Read-only. 8 ExtendedTagEn: extended tag enable. Read-write. 1=8-bit tags generation enabled. 0=5-bit tags are used. 7:5 4 MaxPayloadSize: maximum supported payload size. Read-write. Bits Definition Bits Definition 0h 128B 3h 1024B 1h 256B 4h 2048B 2h 512B 5h 4096B RelaxedOrdEn: relaxed ordering enable. Read-write. 1=The root port is permitted to set the relaxed ordering bit in the attributes field of transactions it initiates that do not require strong write ordering. 297 42300 Rev 3.12 - July 14, 2015 BKDG for AMD Family 15h Models 10h-1Fh Processors 3 UsrReportEn: unsupported request reporting enable. Read-write. 1=Reporting of unsupported requests enabled. 2 FatalErrEn: fatal error reporting enable. Read-write. 1=Enable sending ERR_FATAL messages. 1 NonFatalErrEn: non-fatal error reporting enable. Read-write. 1=Enable sending ERR_NONFATAL messages. 0 CorrErrEn: correctable error reporting enable. Read-write. 1=Enable sending ERR_CORR messages. D[8:2]F0x64 IO Link Capability Register Bits Description 31:24 PortNumber: port number. Value: 00h. This field indicates the port number for the given IO link. 23 Reserved. 22 AspmOptionalityCompliance. Value: 1. 21 LinkBWNotificationCap: link bandwidth notification capability. Value: 0. 20 DlActiveReportingCapable: data link layer active reporting capability. Value: 0. 19 SurpriseDownErrReporting. Value: 0. 18 ClockPowerManagement: clock power management. Value: 0. 0=Indicates that the reference clock must not be removed while in L1 or L2/L3 ready link states. 17:15 L1ExitLatency: L1 exit latency. Value: 010b. 010b=Indicate an exit latency between 2 us and 4 us. 14:12 L0sExitLatency: L0s exit latency. Value: 001b. 001b=Indicates an exit latency between 64 ns and 128 ns. 11:10 PMSupport: active state power management support. Value: 11b. 11b=Indicates support of L0s and L1. 9:4 LinkWidth: maximum link width. Value: 10h. Bits Definition 00h Reserved. 01h 1 lanes 02h 2 lanes 04h 4 lanes 08h 8 lanes 0Ch 12 lanes 10h 16 lanes 3Fh-11h Reserved. 3:0 LinkSpeed: link speed. Value: IF (D[8:2]F0xE4_xA4[LcGen2EnStrap]==0) THEN 1h ELSE 2h ENDIF. Bits Definition 0h Reserved. 1h 2.5 Gb/s 2h 5.0 Gb/s Fh-3h Reserved. 298 42300 Rev 3.12 - July 14, 2015 BKDG for AMD Family 15h Models 10h-1Fh Processors D[8:2]F0x68 IO Link Control and Status Register Reset: 1001_0000h. Bits Description 31 LinkAutonomousBWStatus: link autonomous bandwidth status. IF (D[8:2]F0x64[LinkBWNotificationCap]==0) THEN Read-only. ELSE Read-write. ENDIF. 30 LinkBWManagementStatus: link bandwidth management status. IF (D[8:2]F0x64[LinkBWNotificationCap]==0) THEN Read-only. ELSE Read-write. ENDIF. 29 DlActive: data link layer link active. Read-only. This bit indicates the status of the data link control and management state machine. 1=DL_Active state. 0=All other states. 28 SlotClockCfg: slot clock configuration. Read-only. 1=The root port uses the same clock that the platform provides. 27 LinkTraining: link training. Read-only. This read-only bit indicates that the physical layer link training state machine is in the configuration or recovery state, or that 1b was written to the RetrainLink bit but link training has not yet begun. Hardware clears this bit when the link training state machine exits the configuration/recovery state. 26 Reserved. 25:20 NegotiatedLinkWidth: negotiated link width. Read-only. This field indicates the negotiated width of the given PCI Express link. Bits Definition 01h 1 lanes 02h 2 lanes 04h 4 lanes 08h 8 lanes 0Ch 12 lanes 10h 16 lanes 19:16 LinkSpeed: link speed. Read-only. Bits Definition 0001b 2.5 Gb/s 0010b 5 Gb/s 15:12 Reserved. 11 LinkAutonomousBWIntEn: link autonomous bandwidth interrupt enable. Read-write. 1=Enables the generation of an interrupt to indicate that the Link AutonomousBWStatus bit has been set. 10 LinkBWManagementEn: link bandwidth management interrupt enable. Read-write. 1=Enables the generation of an interrupt to indicate that the LinkBWManagementStatus has been set. 9 HWAutonomousWidthDisable: hardware autonomous width disable. Read-write. 1=Disables hardware from changing the link width for reasons other than attempting to correct unreliable link operation by reducing link width. 8 ClockPowerManagementEn: clock power management enable. Read-write. 7 ExtendedSync: extended sync. Read-write. 1=Forces the transmission of additional ordered sets when exiting the L0s state and when in the recovery state. 299 42300 Rev 3.12 - July 14, 2015 BKDG for AMD Family 15h Models 10h-1Fh Processors 6 CommonClockCfg: common clock configuration. Read-write. 1=Indicates that the root port and the component at the opposite end of this IO link are operating with a distributed common reference clock. 0=Indicates that the root port and the component at the opposite end of this IO Link are operating with asynchronous reference clock. 5 RetrainLink: retrain link. Read-write; cleared-when-done. 1=Initiate link retraining. 4 LinkDis: link disable. Read-write. 1=Disable link. Writes to this bit are immediately reflected in the value read from the bit, regardless of actual link state. 3 ReadCplBoundary: read completion boundary. Read-only. 0=64 byte read completion boundary. 2 Reserved. 1:0 PmControl: active state power management enable. Read-write. This field controls the level of ASPM supported on the given IO link. Bits Definition Bits Definition 00b Disabled. 10b L1 Entry Enabled. 01b L0s Entry Enabled. 11b L0s and L1 Entry Enabled. D[8:2]F0x6C Slot Capability Register Reset: 0004_0000h. Bits Description 31:19 PhysicalSlotNumber: physical slot number. Read-write. This field indicates the physical slot number attached to this port. This field is set to a value that assigns a slot number that is unique within the chassis, regardless of the form factor associated with the slot. This field must be initialized to 0 for ports connected to devices that are on the system board. 18 NoCmdCplSupport: no command completed support. Read-write. 1 =Indicates that this slot does not generate software notification when an issued command is completed by the hot-plug controller. 17 ElecMechIlPresent: electromechanical interlock present. Read-write. 0=Indicates that a electromechanical interlock is not implemented for this slot. 16:15 SlotPwrLimitScale: slot power limit scale. Read-write. Specifies the scale used for the SlotPwrLimitValue. Range of Values: Bits Definition Bits Definition 00b 1.0 10b 0.01 01b 0.1 11b 0.001 14:7 SlotPwrLimitValue: slot power limit value. Read-write. In combination with the SlotPwrLimitScale value, specifies the upper limit on power supplied by slot. Power limit (in Watts) calculated by multiplying the value in this field by the value in the SlotPwrLimitScale field. 6 HotplugCapable: hot-plug capability. Read-write.1=Indicates that this slot is capable of supporting hot-plug operations. 5 HotplugSurprise: hot-plug surprise. Read-write. 1=Indicates that an adapter present in this slot might be removed from the system without any prior notification. 4 PwrIndicatorPresent: power indicator present. Read-write. 0=Indicates that a power indicator is not implemented for this slot. 3 AttnIndicatorPresent: attention indicator present. Read-write. 0=Indicates that a attention indicator is not implemented for this slot. 2 MrlSensorPresent: manual retention latch sensor present. Read-write. 0=Indicates that a manual retention latch sensor is not implemented for this slot. 300 42300 Rev 3.12 - July 14, 2015 BKDG for AMD Family 15h Models 10h-1Fh Processors 1 PwrControllerPresent: power controller present. Read-write. 0=A power controller is not implemented for this slot. 0 AttnButtonPresent: attention button present. Read-write. 0=An attention button is not implemented for this slot. D[8:2]F0x70 Slot Control and Status Register IF (D[8:2]F0x58[SlotImplemented]==0) THEN Reset: 0040_0000h. ELSE Reset: 0000_0000h. ENDIF. Bits Description 31:25 Reserved. 24 DlStateChanged: data link layer state change. Read; Write-1-to-clear. This bit is set when the value reported in the D[8:2]F0x60[DlActive] is changed. In response to a data link layer state changed event, software must read D[8:2]F0x60[DlActive] to determine if the link is active before initiating configuration cycles to the hot plugged device. 23 ElecMechIlSts: electromechanical interlock status. Read-only. 22 PresenceDetectState: presence detect state. Read-only; Updated-by-hardware. This bit indicates the presence of an adapter in the slot based on the physical layer in-band presence detect mechanism. The in-band presence detect mechanism requires that power be applied to an adapter for its presence to be detected. 0=Slot empty. 1=Card present in slot. For root ports not connected to slots (D[8:2]F0x58[SlotImplemented]=0b), this bit returns always 1. 21 MrlSensorState. Read-only. 20 CmdCpl: command completed. Read-only. 19 PresenceDetectChanged: presence detect changes. Read; Write-1-to-clear. This bit is set when the value reported in PresenceDetectState is changed. 18 MrlSensorChanged. Read; Write-1-to-clear. 17 PwrFaultDetected. Read; Write-1-to-clear. 16 AttnButtonPressed: attention button pressed. Read-only. 15:13 Reserved. 12 DlStateChangedEn: data link layer state changed enable. Read-write. 1=Enables software notification when D[8:2]F0x60[DlActive] is changed. 11 ElecMechIlCntl: electromechanical interlock control. Read-only. 10 PwrControllerCntl: power controller control. Read-only. 9:8 PwrIndicatorCntl: power indicator control. Read-only. 7:6 AttnIndicatorControl: attention indicator control. Read-only. 5 HotplugIntrEn: hot-plug interrupt enable. Read-only. 4 CmdCplIntrEn: command complete interrupt enable. Read-only. 3 PresenceDetectChangedEn: presence detect changed enable. Read-only. 2 MrlSensorChangedEn: manual retention latch sensor changed enable. Read-only. 1 PwrFaultDetectedEn: power fault detected enable. Read-only. 0 AttnButtonPressedEn: attention button pressed enable. Read-only. 301 42300 Rev 3.12 - July 14, 2015 BKDG for AMD Family 15h Models 10h-1Fh Processors D[8:2]F0x74 Root Complex Capability and Control Register Reset: 0001_0000h. Bits Description 31:17 Reserved. 16 CrsSoftVisibility: CRS software visibility. Read-only. 1=Indicates that the root port supports returning configuration request retry status (CRS) completion status to software. 15:5 Reserved. 4 CrsSoftVisibilityEn: CRS software visibility enable. Read-write. 1=Enables the root port returning configuration request retry status (CRS) completion status to software. 3 PmIntEn: PME interrupt enable. Read-write. 1=Enables interrupt generation upon receipt of a PME message as reflected D[8:2]F0x78[PmeStatus]. A PME interrupt is also generated if D[8:2]F0x78[PmeStatus]=1 and this bit is set by software. 2 SerrOnFatalErrEn: system error on fatal error enable. Read-write. 1=Indicates that a system error should be generated if a fatal error (ERR_FATAL) is reported by any of the devices in the hierarchy associated with this root port, or by the root port itself. 1 SerrOnNonFatalErrEn: system error on non-fatal error enable. Read-write. 1=Indicates that a system error should be generated if a non-fatal error (ERR_NONFATAL) is reported by any of the devices in the hierarchy associated with this root port, or by the root port itself. 0 SerrOnCorrErrEn: system error on correctable error enable. Read-write. 1=Indicates that a system error should be generated if a correctable error (ERR_COR) is reported by any of the devices in the hierarchy associated with this root port, or by the root port itself. D[8:2]F0x78 Root Complex Status Register Reset: 0000_0000h. Bits Description 31:18 Reserved. 17 PmePending: PME pending. Read-only. This bit indicates that another PME is pending when PmeStatus is set. When PmeStatus is cleared by software; the PME is delivered by hardware by setting the PmeStatus bit again and updating the requestor ID appropriately. PmePending is cleared by hardware if no more PMEs are pending. 16 PmeStatus: pme status. Read; Write-1-to-clear. This bit indicates that PME was asserted by the requestor ID indicated in the PmeRequestorID field. Subsequent PMEs are kept pending until PmeStatus is cleared by writing a 1. 15:0 PmeRequestorId: pme requestor ID. Read-only. This field indicates the PCI requestor ID of the last PME requestor. D[8:2]F0x7C Device Capability 2 Reset: 0000_001Fh. Bits Description 31:6 Reserved. 302 42300 Rev 3.12 - July 14, 2015 BKDG for AMD Family 15h Models 10h-1Fh Processors 5 AriForwardingSupported. Read-only. 4 CplTimeoutDisSup: completion timeout disable supported. Read-only. 3:0 CplTimeoutRangeSup: completion timeout range supported. Read-only. Fh=Completion timeout range is 64s to 50us. D[8:2]F0x80 Device Control and Status 2 Reset: 0000_0000h. Bits Description 31:6 Reserved. 5 AriForwardingEn. Read-only. 4 CplTimeoutDis: completion timeout disable. Read-write. 1=Completion timeout disabled. 3:0 CplTimeoutValue: completion timeout value. Read-write. BIOS: 6h. Bits Timeout Range Bits Timeout Range 0h 50ms-50us 9h 900ms-260ms 1h 100us-50us Ah 3.5s-1s 2h 10ms-1ms Ch-Bh Reserved 4h-3h Reserved Dh 13s-4s 5h 55ms-16ms Eh 64s-4s 6h 210ms-65ms Fh Reserved 8h-7h Reserved D[8:2]F0x84 IO Link Capability 2 Reset: 0000_0000h. Bits Description 31:0 Reserved. D[8:2]F0x88 IO Link Control and Status 2 Bits Description 31:17 Reserved. 16 CurDeemphasisLevel: current deemphasis level. Read-only. Reset: D[8:2]F0xE4_xA4[LcGen2EnStrap]. 1=-3.5 dB. 0=-6 dB 15:13 Reserved. 12 ComplianceDeemphasis: compliance deemphasis. Read-write. Reset: 0. This bit defines the compliance deemphasis level when EnterCompliance is set. Software should leave this field in its default state. 1=-3.5 dB. 0=-6 dB 11 ComplianceSOS: compliance SOS. Read-write. Reset: 0. 1=The device transmits skip ordered sets in between the modified compliance pattern. 10 EnterModCompliance: enter modified compliance. Read-write. Reset: 0. 1=The device transmits modified compliance pattern. Software should leave this field in its default state. 303 42300 Rev 3.12 - July 14, 2015 9:7 BKDG for AMD Family 15h Models 10h-1Fh Processors XmitMargin: transmit margin. Read-write. Reset: 0. This field controls the non-deemphasized voltage level at the transmitter pins. Software should leave this field in its default state. 6 SelectableDeemphasis: selectable deemphasis. Read-only. Reset: D[8:2]F0xE4_xA4[LcGen2EnStrap]. 0=Selectable deemphasis is not supported. 1=Selectable deemphasis supported. 5 HwAutonomousSpeedDisable: hardware autonomous speed disable. Read-write. Cold reset: 0. 1=Support for hardware changing the link speed for device specific reasons disabled. 4 EnterCompliance: enter compliance. Read-write. Cold reset: 0. 1=Force the link to enter the compliance mode. 3:0 TargetLinkSpeed: target link speed. Read-write. Cold reset: {00b, D[8:2]F0xE4_xA4[LcGen2EnStrap],~D[8:2]F0xE4_xA4[LcGen2EnStrap]}. This field defines the upper limit of the link operational speed. Writes of reserved encodings are not valid. Hardware prevents writes of reserved encodings from changing the state of this field. Bits Definition 0h Reserved 1h 2.5GT/s 2h 5.0GT/s Fh-3h Reserved D[8:2]F0x8C Slot Capability 2 Reset: 0000_0000h. Bits Description 31:0 Reserved. D[8:2]F0x90 Slot Control and Status 2 Reset: 0000_0000h. Bits Description 31:0 Reserved. D[8:2]F0xA0 MSI Capability Register Reset: 0000_B005h. Bits Description 31:24 Reserved. 23 Msi64bit: MSI 64 bit capability. Read-only. Value: D0F0x64_x46[Msi64bitEn]. 1=The device is capable of sending 64-bit MSI messages. 0=The device is not capable of sending a 64-bit message address. 22:20 MsiMultiEn: MSI multiple message enable. Read-write. Software writes to this field to indicate the number of allocated vectors (equal to or less than the number of requested vectors). When MSI is enabled, a function is allocated at least 1 vector. 19:17 MsiMultiCap: MSI multiple message capability. Read-only. 000b=The device is requesting one vector. 304 42300 Rev 3.12 - July 14, 2015 16 BKDG for AMD Family 15h Models 10h-1Fh Processors MsiEn: MSI enable. Read-write. 1=MSI generation is enabled and INTx generation is disabled. 0=MSI generation disabled and INTx generation is enabled. 15:8 NextPtr: next pointer. Read-only. 7:0 CapID: capability ID. Read-only. 05h=MSI capability structure. D[8:2]F0xA4 MSI Message Address Low Reset: 0000_0000h. Bits Description 31:2 MsiMsgAddrLo: MSI message address. Read-write. This register specifies the dword aligned address for the MSI memory write transaction. 1:0 Reserved. IF (D0F0x64_x46[Msi64bitEn]==0) THEN D[8:2]F0xA8 MSI Message Data Reset: 0000_0000h. Bits Description 31:16 Reserved. 15:0 MsiData: MSI message data. Read-write. This register specifies lower 16 bits of data for the MSI memory write transaction. The upper 16 bits are always 0. ELSE D[8:2]F0xA8 MSI Message Address High Reset: 0000_0000h. Bits Description 31:8 Reserved. 7:0 MsiMsgAddrHi: MSI message address. Read-write. This register specifies the upper 8-bits of the MSI address in 64 bit MSI mode. D[8:2]F0xAC MSI Message Data Reset: 0000_0000h. Bits Description 31:16 Reserved. 15:0 MsiData: MSI message data. Read-write. This register specifies lower 16 bits of data for the MSI memory write transaction. The upper 16 bits are always 0. ENDIF. 305 42300 Rev 3.12 - July 14, 2015 BKDG for AMD Family 15h Models 10h-1Fh Processors D[8:2]F0xB0 Subsystem and Subvendor Capability ID Register Reset: 0000_B80Dh. Bits Description 31:16 Reserved. 15:8 NextPtr: next pointer. Read-only. 7:0 CapID: capability ID. Read-only. D[8:2]F0xB4 Subsystem and Subvendor ID Register Bits Description 31:16 SubsystemID. Read-only. Value: D0F0xE4_x013[1:0]_0046[SubsystemID]. 15:0 SubsystemVendorID. Read-only. Value: D0F0xE4_x013[1:0]_0046[SubsystemVendorID]. D[8:2]F0xB8 MSI Capability Mapping Reset: A803_0008h. Bits Description 31:27 CapType: capability type. Read-only. 26:18 Reserved. 17 FixD. Read-only. 16 En. Read-only. 15:8 NextPtr: next pointer. Read-only. 7:0 CapID: capability ID. Read-only. D[8:2]F0xBC MSI Mapping Address Low Bits Description 31:20 MsiMapAddrLo. Read-only. Reset: 0. Lower 32-bits of the MSI address. 19:0 Reserved. D[8:2]F0xC0 MSI Mapping Address High Bits Description 31:0 MsiMapAddrHi. Read-only. Reset: 0. Upper 32-bits of the MSI address. D[8:2]F0xE0 Root Port Index Reset: 0000_0000h. The index/data pair registers D[8:2]F0xE0 and D[8:2]F0xE4 is used to access the registers 306 42300 Rev 3.12 - July 14, 2015 BKDG for AMD Family 15h Models 10h-1Fh Processors D[8:2]F0xE4_x[FF:00]. To read or write to one of these registers, the address is written first into the address register D[8:2]F0xE0 and then the data are read or written by read or write the data register D[8:2]F0xE4. Bits Description 31:8 Reserved. 7:0 PcieIndex. Read-write. D[8:2]F0xE4 Root Port Data See D[8:2]F0xE0. D[8:2]F0xE4_x02 Root Port Hardware Debug Reset: 0000_0000h. Bits Description 31:16 Reserved. 15 RegsLcAllowTxL1Control. Read-write. BIOS: 1. 1=Tx can prevent LC from entering L1 when there are outstanding completions. 14:0 Reserved. D[8:2]F0xE4_x20 Root Port TX Control Reset: 0050_8000h. Bits Description 31:16 Reserved. 15 TxFlushTlpDis: TLP flush disable. Read-write. BIOS: ~D[8:2]F0x6C[HotplugCapable]. 1=Disable flushing TLPs when the link is down. 14:0 Reserved. D[8:2]F0xE4_x50 Root Port Lane Status Reset: 0000_0000h. Bits Description 31:7 Reserved. 6:1 0 PhyLinkWidth: port link width. Read-only. Bits Definition Bits 00_0000b disabled 00_1000b 00_0001b x1 01_0000b 00_0010b x2 10_0000b 00_0100b x4 Definition x8 x12 x16 PortLaneReversal: port lane reversal. Read-only. 1=Port lanes order is reversed. 307 42300 Rev 3.12 - July 14, 2015 BKDG for AMD Family 15h Models 10h-1Fh Processors D[8:2]F0xE4_x6A Root Port Error Control Reset: 0000_0500h. Bits Description 31:1 Reserved. 0 ErrReportingDis: advanced error reporting disable. Read-write. BIOS: 0. 1=Error reporting disabled. 0=Error reporting enabled. D[8:2]F0xE4_x70 Root Port Receiver Control Reset: 0000_43F7h. Bits Description 31:20 Reserved. 19 RxRcbCplTimeoutMode: RCB completion timeout mode. Read-write. BIOS: 1. 18:16 RxRcbCplTimeout: RCB completion timeout. Read-write. Bits Definition Bits Definition 000b Disabled 100b 50ms 001b 50us 101b 100ms 010b 10ms 110b 500ms 011b 25ms 111b 1ms 15:0 Reserved. D[8:2]F0xE4_xA0 Per Port Link Controller (LC) Control Reset: 4000_0050h. Bits Description 31:24 Reserved. 23 LcL1ImmediateAck: immediate ACK ASPM L1 entry. Read-write. BIOS: 1. 1=Always ACK ASPM L1 entry DLLPs. 22:16 Reserved. 15:12 LcL1Inactivity: L1 inactivity timer. Read-write. Bits Definition Bits 0h L1 disabled 8h 1h 1us 9h 2h 2us Ah 3h 4us Bh 4h 10us Ch 5h 20us Dh 6h 40us Eh 7h 100us Fh Definition 400us 1ms 40us 10ms 40ms 100ms 400ms Reserved 308 42300 Rev 3.12 - July 14, 2015 BKDG for AMD Family 15h Models 10h-1Fh Processors 11:8 LcL0sInactivity: L0s inactivity timer. Read-write. Bits Definition Bits 0h L0s disabled 8h 1h 40ns 9h 2h 80ns Ah 3h 120ns Bh 4h 200ns Ch 5h 400ns Dh 6h 1us Eh 7h 2us Fh Definition 4us 10us 40us 100us 400us 1ms 4ms Reserved 7:4 Lc16xClearTxPipe. Read-write. BIOS: 3h. Specifies the number of clock to drain the TX pipe. 3:0 Reserved. D[8:2]F0xE4_xA1 LC Training Control Reset: 0400_1080h. Bits Description 31:13 Reserved. 12 LcInitSpdChgWithCsrEn: enable software initialed speed change. Read-write. 1=Enable link speed negotiation when D[8:2]F0x68[RetrainLink]=1 and the target link speed differs from the current link speed. 0=Link speed negotiation will not be performed when D[8:2]F0x68[RetrainLink]=1. 11 LcDontGotoL0sifL1Armed: prevent Ls0 entry if L1 request in progress. Read-write. BIOS: 1. 1=Prevent the LTSSM from transitioning to Rcv_L0s if an acknowledged request to enter L1 is in progress. 10:0 Reserved. D[8:2]F0xE4_xA2 LC Link Width Control Reset: 0020_0006h. Bits Description 31:23 Reserved. 22:21 LcDynLanesPwrState: unused link power state. Read-write. Controls the state of unused links after a reconfiguration. Bits Definition Bits Definition 00b on 10b SB2 01b SB1 11b Off 20 LcUpconfigCapable: upconfigure capable. Read-only. 1=Both ends of the link are upconfigure capable. 0=Both ends of the link are not upconfigure capable. 19:14 Reserved. 13 LcUpconfigureDis: upconfigure disable. Read-write. 1=Disable link upconfigure. 12 LcUpconfigureSupport: upconfigure support. Read-write. 11 LcShortReconfigEn: short re-configuration enable. Read-write. 1=Enable short link re-configuration 10 LcRenegotiateEn: link reconfiguration enable. Read-write. 1=Enable link re-negotiation. 309 42300 Rev 3.12 - July 14, 2015 BKDG for AMD Family 15h Models 10h-1Fh Processors 9 LcRenegotiationSupport: re-negotiation support. Read-only. 1=Link re-negotiation is supported by the downstream device. 8 LcReconfigNow: re-configure link. Read-write; cleared-when-done. 1=Initiate link width change. 7 LcReconfigArcMissingEscape. Read-write. 1=Expedite transition from Recovery.Idle to Detect during a long reconfiguration. 6:4 3 2:0 LcLinkWidthRd: current link width. Read-only. Bits Definition Bits 000b 0 100b 001b 1 101b 010b 2 110b 011b 4 111b Definition 8 12 16 Reserved Reserved. LcLinkWidth: link width required. Read-write. See: LcLinkWidthRd. D[8:2]F0xE4_xA3 LC Number of FTS Control Reset: 00FF_020Ch. Bits Description 31:10 Reserved. 9 8:0 LcXmitFtsBeforeRecovery: transmit FTS before recovery. Read-write. 1=Transmit FTS before recovery. Reserved. D[8:2]F0xE4_xA4 LC Link Speed Control Reset: 0440_0100h. Bits Description 31:28 Reserved. 27 LcMultUpstreamAutoSpdChngEn: enable multiple automatic speed changes. Read-write. 1=Enable multiple automatic speed changes. 26:20 Reserved. 19 LcOtherSideSupportsGen2: downstream link supports gen2. Read-only. 1=The downstream link currently supports gen2. 18:13 Reserved. 12 LcSpeedChangeAttemptFailed: speed change attempt failed. Read-only; updated-by-hardware. 1=LcSpeedChangeAttemptsAllowed has been reached. 11:10 Reserved. 9 8:7 6 LcInitiateLinkSpeedChange: initiate link speed change. Read-write; cleared-when-done. 1=Initiate link speed negotiation. Reserved. LcForceDisSwSpeedChange: force disable software speed changes. Read-write. 1=Force the PCIe core to disable speed changes initiated by private registers. 310 42300 Rev 3.12 - July 14, 2015 5:1 0 BKDG for AMD Family 15h Models 10h-1Fh Processors Reserved. LcGen2EnStrap: gen2 PCIe support enable. Read-write. 1=Gen2 PCIe support enabled. 0=Gen2 PCIe support disabled. D[8:2]F0xE4_xA5 LC State 0 Cold reset: 0000_0000h. Table 123: Link controller state encodings Bits Description Bits Description Bits Description 00h s_Detect_Quiet. 12h Rcv_L0_and_Tx_L0s. 24h s_Rcvd_Loopback. 01h s_Start_common_Mode. 13h Rcv_L0_and_Tx_L0s_FTS. 25h s_Rcvd_Loopback_Idle. 02h s_Check_Common_Mode. 14h Rcv_L0s_and_Tx_L0. 26h s_Rcvd_Reset_Idle. 03h s_Rcvr_Detect. 15h Rcv_L0s_and_Tx_L0_Idle. 27h s_Rcvd_Disable_Entry. 04h s_No_Rcvr_Loop 16h Rcv_L0s_and_Tx_L0s. 28h s_Rcvd_Disable_Idle. 05h s_Poll_Quiet. 17h Rcv_L0s_and_Tx_L0s_FTS. 29h s_Rcvd_Disable. 06h s_Poll_Active. 18h s_L1_Entry. 2Ah s_Detect_Idle. 07h s_Poll_Compliance. 19h s_L1_Idle. 2Bh s_L23_Wait. 08h s_Poll_Config. 1Ah s_L1_Wait 2Ch Rcv_L0s_Skp_and_Tx_L0. 09h s_Config_Step1. 1Bh s_L1. 2Dh Rcv_L0s_Skp_and_Tx_L0_Idle. 0Ah s_Config_Step3. 1Ch s_L23_Stall. 2Eh Rcv_L0s_Skp_and_Tx_L0s. 0Bh s_Config_Step5. 1Dh s_L23_Entry. 2Fh Rcv_L0s_Skp_and_Tx_L0_FTS. 0Ch s_Config_Step2. 1Eh s_L23_Entry. 30h s_Config_Step2b. 0Dh s_Config_Step4. 1Fh s_L23_Ready. 31h s_Recovery_Speed. 0Eh s_Config_Step6. 20h s_Recovery_lock. 32h s_Poll_Compliance_Idle. 0Fh s_Config_Idle. 21h s_Recovery_Config. 33h s_Rcvd_Loopback_Speed. 10h Rcv_L0_and_Tx_L0. 22h s_Recovery_Idle. 11h Rcv_L0_and_Tx_L0_Idle. 23h s_Training_Bit. Bits 3Fh-34h Reserved. Description 31:30 Reserved. 29:24 LcPrevState3: previous link state 3. Read-only. See: Table 123. 23:22 Reserved. 21:16 LcPrevState2: previous link state 2. Read-only. See: Table 123. 15:14 Reserved. 13:8 LcPrevState1: previous link state 1. Read-only. See: Table 123. 7:6 Reserved. 5:0 LcCurrentState: current link state. Read-only. See: Table 123. 311 42300 Rev 3.12 - July 14, 2015 BKDG for AMD Family 15h Models 10h-1Fh Processors D[8:2]F0xE4_xB1 LC Control 2 Reset: 8608_0280h. Bits Description 31:21 Reserved. 20 LcBlockElIdleinL0: block electrical idle in l0. Read-write. BIOS: 1. 1=Prevent electrical idle from causing the receiver to transition from L0 to L0s. 19 LcDeassertRxEnInL0s: deassert RX_EN in L0s. Read-write. 1=Turn off transmitters in L0s. 18:0 Reserved. D[8:2]F0xE4_xB5 LC Control 3 Reset: 2850_5020h. Bits Description 31 Reserved. 30 LcGoToRecovery: go to recovery. Read-write. 1=Force link in the L0 state to transition to the recovery state. 29:4 Reserved. 3 LcRcvdDeemphasis: received deemphasis. Read-only. Deemphasis advertised by the downstream device. 1=3.5dB. 0=6dB. 2:1 LcSelectDeemphasisCntl: deemphasis control. Read-write. Specifies the deemphasis used by the transmitter. Bits Definition 00b Use deemphasis from LcSelectDeemphasis. 01b Use deemphasis advertised by the downstream device. 10b 6dB 11b 3.5dB 0 LcSelectDeemphasis: downstream deemphasis. Read-write. Specifies the downstream deemphasis. 1=3.5dB. 0=6dB. D[8:2]F0xE4_xC0 LC Strap Override Bits Description 31:16 Reserved. 15 StrapAutoRcSpeedNegotiationDis: autonomous speed negotiation disable strap override. Readwrite. Reset: 1. 14 Reserved. 13 StrapForceCompliance: force compliance strap override. Read-write. Reset: 0b. 12:0 Reserved. 312 42300 Rev 3.12 - July 14, 2015 BKDG for AMD Family 15h Models 10h-1Fh Processors D[8:2]F0xE4_xC1 Root Port Miscellaneous Strap Override Reset: 0000_0000h. Bits Description 31:3 Reserved. 2 StrapExtendedFmtSupported: Extended Fmt Supported strap override. Read-write. BIOS: See 2.12.2 [IOMMU Initialization]. 1 StrapE2EPrefixEn: E2E Prefix En strap override. Read-write. BIOS: See 2.12.2 [IOMMU Initialization]. 0 StrapReverseLanes: reverse lanes strap override. Read-write. D[8:2]F0x100 Vendor Specific Enhanced Capability Register Bits Description 31:20 NextPtr: next pointer. Read-only. IF (D0F0xE4_x0[2:1]01_00B0[StrapF0AerEn] == 1) THEN Reset: 150h. ELSE Reset: 000h. ENDIF. 19:16 CapVer: capability version. Read-only. Reset: 1h. 15:0 CapID: capability ID. Read-only. Reset: 000Bh. D[8:2]F0x104 Vendor Specific Header Register Reset: 0101_0001h. Bits Description 31:20 VsecLen: vendor specific enhanced capability structure length. Read-only. Defined the number of bytes of the entire vendor specific enhanced capability structure including the header. 19:16 VsecRev: vendor specific enhanced capability version. Read-only. 15:0 VsecID: vendor specific enhanced capability ID. Read-only. D[8:2]F0x108 Vendor Specific 1 Register Reset: 0000_0000h. Bits Description 31:0 Scratch: scratch. Read-write. This field does not control any hardware. D[8:2]F0x10C Vendor Specific 2 Register Reset: 0000_0000h. Bits Description 31:0 Scratch: scratch. Read-write. This field does not control any hardware. 313 42300 Rev 3.12 - July 14, 2015 BKDG for AMD Family 15h Models 10h-1Fh Processors D[8:2]F0x128 Virtual Channel 0 Resource Status Register Reset: 0002_0000h. Bits Description 31:18 Reserved. 17 VcNegotiationPending: virtual channel negotiation pending. Read-only. 1=Virtual channel negotiation in progress. This bit must be 0 before the virtual channel can be used. 16 PortArbTableStatus: port arbitration table status. Read-only. 15:0 Reserved. D[8:2]F0x150 Advanced Error Reporting Capability Reset: 0001_0001h. Bits Description 31:20 NextPtr: next pointer. Read-only. 19:16 CapVer: capability version. Read-only. 15:0 CapID: capability ID. Read-only. D[8:2]F0x154 Uncorrectable Error Status Cold reset: 0000_0000h. Bits Description 31:26 Reserved. 25 TlpPrefixStatus: TLP prefix blocked status. Read; Write-1-to-clear. 24 AtomicOpEgressBlockedTLPStatus: atomic op egress blocked TLP status. Read; Write-1-toclear. 23 McBlockedTLPStatus: MC blocked TLP status. Read; Write-1-to-clear. 22 UncorrInteralErrStatus: uncorrectable internal error status. Read; Write-1-to-clear. 21 AcsViolationStatus: access control service status. Read; Write-1-to-clear. 20 UnsuppReqErrStatus: unsupported request error status. Read; Write-1-to-clear. The header of the unsupported request is logged. 19 EcrcErrStatus: end-to-end CRC error status. Read; Write-1-to-clear. 18 MalTlpStatus: malformed TLP status. Read; Write-1-to-clear. The header of the malformed TLP is logged. 17 RcvOvflStatus: receiver overflow status. Read-only. 16 UnexpCplStatus: unexpected completion timeout status. Read; Write-1-to-clear. The header of the unexpected completion is logged. 15 CplAbortErrStatus: completer abort error status. Read; Write-1-to-clear. 14 CplTimeoutStatus: completion timeout status. Read; Write-1-to-clear. 13 FcErrStatus: flow control error status. Read-only. 314 42300 Rev 3.12 - July 14, 2015 12 BKDG for AMD Family 15h Models 10h-1Fh Processors PsnErrStatus: poisoned TLP status. Read; Write-1-to-clear. The header of the poisoned transaction layer packet is logged. 11:6 Reserved. 5 SurprdnErrStatus: surprise down error status. Read-only. 0=Detection and reporting of surprise down errors is not supported. 4 DlpErrStatus: data link protocol error status. Read; Write-1-to-clear. 3:0 Reserved. D[8:2]F0x158 Uncorrectable Error Mask Cold reset: 0000_0000h. Bits Description 31:26 Reserved. 25 TlpPrefixMask: TLP prefix blocked mask. Read-only. 24 AtomicOpEgressBlockedTLPMask: atomic op egress blocked TLP mask. Read-only. 23 McBlockedTLPMask: MC blocked TLP mask. Read-only. 22 UncorrInteralErrMask: uncorrectable internal error mask. Read-only. 21 AcsViolationMask: access control service mask. Read-only. 1=ACS violation errors are not reported. 20 UnsuppReqErrMask: unsupported request error mask. Read-write. 1=Unsupported request errors are not reported. 19 EcrcErrMask: end-to-end CRC error mask. Read-only. 18 MalTlpMask: malformed TLP mask. Read-write. 1=Malformed TLP errors are not reported. 17 RcvOvflMask: receiver overflow mask. Read-only. 16 UnexpCplMask: unexpected completion timeout mask. Read-write. 1=Unexpected completion errors are not reported. 15 CplAbortErrMask: completer abort error mask. Read-only. 14 CplTimeoutMask: completion timeout mask. Read-write. 1=Completion timeout errors are not reported. 13 FcErrMask: flow control error mask. Read-only. 12 PsnErrMask: poisoned TLP mask. Read-write.1=Poisoned TLP errors are not reported. 11:6 Reserved. 5 SurprdnErrMask: surprise down error mask. Read-only. 4 DlpErrMask: data link protocol error mask. Read-write. 1=Data link protocol errors are not reported. 3:0 Reserved. 315 42300 Rev 3.12 - July 14, 2015 BKDG for AMD Family 15h Models 10h-1Fh Processors D[8:2]F0x15C Uncorrectable Error Severity Cold reset: 0006_2030h. Bits Description 31:26 Reserved. 25 TlpPrefixSeverity: TLP prefix blocked severity. Read-only. 24 AtomicOpEgressBlockedTLPSeverity: atomic op egress blocked TLP severity. Read-only. 23 McBlockedTLPSeverity: MC blocked TLP severity. Read-only. 22 UncorrInteralErrSeverity: uncorrectable internal error severity. Read-only. 21 AcsViolationSeverity: access control service severity. Read-only. 1=Fatal error. 0=Non-fatal error. 20 UnsuppReqErrSeverity: unsupported request error severity. Read-write. 1=Fatal error. 0=Nonfatal error. 19 EcrcErrSeverity: end-to-end CRC error severity. Read-only. 18 MalTlpSeverity: malformed TLP severity. Read-write. 1=Fatal error. 0=Non-fatal error. 17 RcvOvflSeverity: receiver overflow severity. Read-only. 16 UnexpCplSeverity: unexpected completion timeout severity. Read-write. 1=Fatal error. 0=Nonfatal error. 15 CplAbortErrSeverity: completer abort error severity. Read-only. 14 CplTimeoutSeverity: completion timeout severity. Read-write. 1=Fatal error. 0=Non-fatal error. 13 FcErrSeverity: flow control error severity. Read-only. 12 PsnErrSeverity: poisoned TLP severity. Read-write. 1=Fatal error. 0=Non-fatal error. 11:6 Reserved. 5 SurprdnErrSeverity: surprise down error severity. Read-only. 4 DlpErrSeverity: data link protocol error severity. Read-write. 1=Fatal error. 0=Non-fatal error. 3:0 Reserved. D[8:2]F0x160 Correctable Error Status Cold reset: 0000_0000h. Bits Description 31:14 Reserved. 13 AdvisoryNonfatalErrStatus: advisory non-fatal error status. Read; Write-1-to-clear. 1=A nonfatal unsupported request errors or a non-fatal unexpected completion errors occurred. 12 ReplayTimerTimeoutStatus: replay timer timeout status. Read; Write-1-to-clear. 11:9 Reserved. 8 ReplayNumRolloverStatus: replay. Read; Write-1-to-clear. 1=The same transaction layer packet has been replayed three times and has caused the link to re-train. 7 BadDllpStatus: bad data link layer packet status. Read; Write-1-to-clear. 1=A link CRC error was detected. 6 BadTlpStatus: bad transaction layer packet status. Read; Write-1-to-clear. 1=A bad non-duplicated sequence ID or a link CRC error was detected. 316 42300 Rev 3.12 - July 14, 2015 5:1 0 BKDG for AMD Family 15h Models 10h-1Fh Processors Reserved. RcvErrStatus: receiver error status. Read-only. 1=An 8B10B or disparity error was detected. D[8:2]F0x164 Correctable Error Mask Cold reset: 0000_2000h. Bits Description 31:14 Reserved. 13 AdvisoryNonfatalErrMask: advisory non-fatal error mask. Read-write. 1=Error is not reported. 12 ReplayTimerTimeoutMask: replay timer timeout mask. Read-write. 1=Error is not reported. 11:9 Reserved. 8 ReplayNumRolloverMask: replay. Read-write.1=Error is not reported. 7 BadDllpMask: bad data link layer packet mask. Read-write. 1=Error is not reported. 6 BadTlpMask: bad transaction layer packet mask. Read-write. 1=Error is not reported. 5:1 0 Reserved. RcvErrMask: receiver error mask. Read-only. 1=Error is not reported. D[8:2]F0x168 Advanced Error Control Cold reset: 0000_0000h. Bits Description 31:9 Reserved. 8 EcrcCheckEn: data link protocol error severity. Read-only. 0=Specifies that end-to-end CRC generation is not supported. 7 EcrcCheckCap: data link protocol error severity. Read-only. 0=Specifies that end-to-end CRC check is not supported. 6 EcrcGenEn: end-to-end CRC enable. Read-only. 0=Specifies that end-to-end CRC generation is not supported. 5 EcrcGenCap: end-to-end CRC capability. Read-only. 0=Specifies that end-to-end CRC generation is not supported. 4:0 FirstErrPtr: first error pointer. Read-only. The First Error Pointer identifies the bit position of the first error reported in the Uncorrectable Error Status register. D[8:2]F0x16C Header Log DW0 Cold reset: 0000_0000h. Bits Description 31:0 TlpHdr: transaction layer packet header log. Read-only. Contains the header for a transaction layer packet corresponding to a detected error. The upper byte represents byte 0 of the header. 317 42300 Rev 3.12 - July 14, 2015 BKDG for AMD Family 15h Models 10h-1Fh Processors D[8:2]F0x170 Header Log DW1 Cold reset: 0000_0000h. Bits Description 31:0 TlpHdr: transaction layer packet header log. Read-only. Contains the header for a transaction layer packet corresponding to a detected error. The upper byte represents byte 4 of the header. D[8:2]F0x174 Header Log DW2 Cold reset: 0000_0000h. Bits Description 31:0 TlpHdr: transaction layer packet header log. Read-only. Contains the header for a transaction layer packet corresponding to a detected error. The upper byte represents byte 8 of the header. D[8:2]F0x178 Header Log DW3 Cold reset: 0000_0000h. Bits Description 31:0 TlpHdr: transaction layer packet header log. Read-only. Contains the header for a transaction layer packet corresponding to a detected error. The upper byte represents byte 12 of the header. D[8:2]F0x17C Root Error Command Reset: 0000_0000h. Bits Description 31:3 Reserved. 2 FatalErrRepEn: fatal error reporting enable. Read-write. 1=Enables the generation of an interrupt when a fatal error is reported by any of the devices in the hierarchy associated with this Root Port. 1 NonfatalErrRepEn: non-fatal error reporting enable. Read-write. 1=Enables generation of an interrupt when a non-fatal error is reported by any of the devices in the hierarchy associated with this Root Port. 0 CorrErrRepEn: correctable error reporting enable. Read-write. 1=Enables generation of an interrupt when a correctable error is reported by any of the devices in the hierarchy associated with this Root Port. D[8:2]F0x180 Root Error Status Cold reset: 0000_0000h. Bits Description 31:27 AdvErrIntMsgNum: advanced error interrupt message number. Read-only. 26:7 Reserved. 6 NFatalErrMsgRcvd: fatal error message received. Read; Write-1-to-clear. Set to 1 when one or more fatal uncorrectable error messages have been received. 318 42300 Rev 3.12 - July 14, 2015 BKDG for AMD Family 15h Models 10h-1Fh Processors 5 NonFatalErrMsgRcvd: non-fatal error message received. Read; Write-1-to-clear. Set to 1 when one or more non-fatal uncorrectable error messages have been received. 4 FirstUncorrFatalRcvd: first uncorrectable fatal error message received. Read; Write-1-to-clear. Set to 1 when the first uncorrectable error message received is for a fatal error. 3 MultErrFatalNonfatalRcvd: ERR_FATAL/NONFATAL message received. Read; Write-1-toclear. Set when either a fatal or a non-fatal error is received and ErrFatalNonfatalRcvd is already set. 2 ErrFatalNonfatalRcvd: ERR_FATAL/NONFATAL message received. Read; Write-1-to-clear. Set when either a fatal or a non-fatal error is received and this bit is not already set. 1 MultErrCorrRcvd: multiple ERR_COR messages received. Read; Write-1-to-clear. Set when a correctable error message is received and ErrCorrRcvd is already set. 0 ErrCorrRcvd: ERR_COR message received. Read; Write-1-to-clear. Set when a correctable error message is received and this bit is not already set. D[8:2]F0x184 Error Source ID Cold reset: 0000_0000h. Bits Description 31:16 ErrFatalNonfatalSrcID: ERR_FATAL/ERR_NONFATAL source identification. Read-only. Loaded with the requestor ID indicated in the received ERR_FATAL or ERR_NONFATAL message when D[8:2]F0x180[ErrFatalNonfatalRcvd] is not already set. 15:0 ErrCorlSrcID: ERR_COR source identification. Read-only. Loaded with the requestor ID indicated in the received ERR_COR message when D[8:2]F0x180[ErrCorrRcvd] is not already set. 319 42300 Rev 3.12 - July 14, 2015 3.8 BKDG for AMD Family 15h Models 10h-1Fh Processors Device 18h Function 0 Configuration Registers See 3.1 [Register Descriptions and Mnemonics]. See 2.7 [Configuration Space]. D18F0x00 Device/Vendor ID Bits Description 31:16 DeviceID: device ID. Read-only. Value: 1400h. 15:0 VendorID: vendor ID. Read-only. Value: 1022h. D18F0x04 Status/Command Bits Description 31:16 Status. Read-only. Value: 0010h. Bit[20] is set to indicate the existence of a PCI-defined capability block. 15:0 Command. Read-only. Value: 0000h. D18F0x08 Class Code/Revision ID Bits Description 31:8 ClassCode. Read-only. Value: 060000h. Provides the host bridge class code as defined in the PCI specification. 7:0 RevID: revision ID. Read-only. Value: 00h. D18F0x0C Header Type Read-only. Value: 0080_0000h. Bits Description 31:0 HeaderTypeReg. These bits are fixed at their default values. The header type field indicates that there are multiple functions present in this device. D18F0x34 Capabilities Pointer Bits Description 31:8 Reserved. 7:0 CapPtr: capabilities pointer. Read-only. Value: 00h. D18F0x40 Routing Table Reset: 0004_0201h. Bits Description 31:0 Reserved. 320 42300 Rev 3.12 - July 14, 2015 BKDG for AMD Family 15h Models 10h-1Fh Processors D18F0x60 Node ID Bits Description 31:21 Reserved. 20:16 CpuCnt[4:0]: CPU count bits[4:0]. Read-write. Reset: 0. CpuCnt specifies the number of cores to be enabled in the system (the boot core of all nodes plus those cores enabled through D18F0x1DC[CpuEn]). Bits Description 00h 1 core. 02h-01h <CpuCnt[4:0]>+1 cores. 03h 4 cores. 3Fh-04h Reserved. 15:0 Reserved. D18F0x64 Unit ID Reset: 0000_00E0h. Bits Description 31:8 Reserved. 7:6 HbUnit: host bridge Unit ID. Read-only. Specifies the coherent link Unit ID of the host bridge used by the coherent fabric. 5:4 MctUnit: memory controller Unit ID. Read-only. Specifies the coherent link Unit ID of the memory controller. 3:0 Reserved. D18F0x68 Link Transaction Control Bits Description 31:23 Reserved. 22:21 DsNpReqLmt: downstream non-posted request limit. Read-write. Reset: 00b. BIOS: 01b. This specifies the maximum number of downstream non-posted requests issued by core(s) which may be outstanding on the IO links attached to this node at one time. Bits Description 00b No limit 01b limited to 1 10b limited to 4 11b limited to 8 20 SeqIdSrcNodeEn: sequence ID source node enable. Read-write. Reset: 0. 1=The source node ID of requests is provided in the SeqID field of the corresponding downstream IO link request packets. This may be useful for debug applications, in order to match downstream packets with their originating node. For normal operation, this bit should be cleared. Correct ordering of requests between different nodes is not ensured when this bit is set. Semaphore sharing between differing nodes may not work properly in systems which are capable of processing IO requests with differing non-zero SeqIds out of request order. 321 42300 Rev 3.12 - July 14, 2015 BKDG for AMD Family 15h Models 10h-1Fh Processors 19 ApicExtSpur: APIC extended spurious vector enable. Read-write. Reset: 0. This enables the extended APIC spurious vector functionality; it affects APICF0[Vector]. 0=The lower 4 bits of the spurious vector are read-only 1111b. 1=The lower 4 bits of the spurious vector are writable. 18 ApicExtId: APIC extended ID enable. Read-write. Reset: 0. Enables the extended APIC ID functionality. 0=APIC ID is 4 bits. 1=APIC ID is 8 bits. 17 ApicExtBrdCst: APIC extended broadcast enable. Read-write. Reset: 0. Enables the extended APIC broadcast functionality. 0=APIC broadcast is 0Fh. 1=APIC broadcast is FFh. If ApicExtBrdCst=1 then software must assert ApicExtId. 16 LintEn: local interrupt conversion enable. Read-write. Reset: 0. 1=Enables the conversion of broadcast ExtInt and NMI interrupt requests to LINT0 and LINT1 local interrupts, respectively, before delivering to the local APIC. This conversion only takes place if the local APIC is hardware enabled. LINT0 and LINT1 are controlled by APIC3[60:50]. 0=ExtInt/NMI interrupts delivered unchanged. 15 LimitCldtCfg: limit coherent link configuration space range. Read-write. Reset: 0. BIOS: 1. 14:12 Reserved. 11 RespPassPW: response PassPW. Read-write. Reset: 0. 1=The PassPW bit in all downstream link responses is set, regardless of the originating request packet. This technically breaks the PCI ordering rules but it is not expected to be an issue in the downstream direction. Setting this bit improves the latency of upstream requests by allowing the downstream responses to pass posted writes. 0=The PassPW bit in downstream responses is based on the RespPassPW bit of the original request. 10 DisFillP: disable fill probe. Read-write. Reset: 0. Controls probes for core-generated fills. 0=Probes issued for cache fills. 1=Probes not issued for cache fills. BIOS: 0. BIOS may set if single core. 9 DisRmtPMemC: disable remote probe memory cancel. Read-write. Reset: 0. 1=Only probed caches on the same node as the target memory controller may generate MemCancel coherent link packets. MemCancels are used to attempt to save DRAM and/or link bandwidth associated with the transfer of stale DRAM data. 0=Probes hitting dirty blocks may generate MemCancel packets, regardless of the location of the probed cache. 8 DisPMemC: disable probe memory cancel. Read-write. Reset: 0. Controls generation of MemCancel coherent link packets. MemCancels are used to attempt to save DRAM and/or coherent link bandwidth associated with the transfer of stale DRAM data. 0=Probes hitting dirty blocks of the core cache may generate MemCancel packets. 1=Probes may not generate MemCancel packets. 7 CPURdRspPassPW: CPU read response PassPW. Read-write. Reset: 0. 1=Read responses to coregenerated reads are allowed to pass posted writes. 0=core responses do not pass posted writes. This bit is not expected to be set. This bit may only be set during the boot process. 6 CPUReqPassPW: CPU request PassPW. Read-write. Reset: 0. 1=Core-generated requests are allowed to pass posted writes. 0=Core requests do not pass posted writes. This bit is not expected to be set. This bit may only be set during the boot process. 5 Reserved. 4 DisMTS: disable memory controller target start. Read-write. Reset: 0. 1=Disables use of TgtStart. TgtStart is used to improve scheduling of back-to-back ordered transactions by indicating when the first transaction is received and ordered at the memory controller. 3:0 Reserved. 322 42300 Rev 3.12 - July 14, 2015 BKDG for AMD Family 15h Models 10h-1Fh Processors D18F0x6C Link Initialization Control Bits Description 31 Reserved. 30 RlsLnkFullTokCntImm: release upstream full token count immediately. Read-write. Cold reset: 0. BIOS: 1 after buffer counts have been programmed. 1=Apply buffer counts programmed in D18F0x90 and D18F0x94 immediately without requiring warm reset. Once this bit is set, additional changes to the buffer counts only take effect upon warm reset. 29 Reserved. 28 RlsIntFullTokCntImm: release internal full token count immediately. Read-write. Cold reset: 0. BIOS: 1 after buffer counts have been programmed. 1=Apply buffer counts programmed in D18F3x6C, D18F3x70, D18F3x74, D18F3x78, D18F3x7C, D18F3x140, D18F3x144, D18F3x148, D18F3x17C, and D18F3x1A0 immediately without requiring warm reset. Once this bit is set, additional changes to the buffer counts only take effect upon warm reset. 27 ApplyIsocModeEnNow. Read-write. Cold reset: 0. BIOS: 1 after RlsLnkFullTokCntImm and RlsIntFullTokCntImm have been set. 1=Apply the programmed value in D18F0x84[IsocEn] immediately without requiring warm reset. This bit may only be set if RlsLnkFullTokCntImm and RlsIntFullTokCntImm are set and isochronous buffers have been allocated. IF (ApplyIsocModeEnNow) THEN (D18F3x148[IsocPreqTok0] > 0). 26:11 Reserved. 10:9 BiosRstDet[2:1]: BIOS reset detect bits[2:1]. See: BiosRstDet[0]. 8:7 Reserved. 6 InitDet: CPU initialization command detect. Read-write. Reset: 0. This bit may be used by software to distinguish between an INIT and a warm/cold reset by setting it to a 1 before an initialization event is generated. This bit is cleared by RESET_L but not by an INIT command. 5 BiosRstDet[0]: BIOS reset detect bit[0]. Read-write. Cold reset: 0. BiosRstDet[2:0] = {BiosRstDet[2:1], BiosRstDet[0]}. May be used to distinguish between a reset event generated by the BIOS versus a reset event generated for any other reason by setting one or more of the bits to a 1 before initiating a BIOS-generated reset event. 4 ColdRstDet: cold reset detect. Read-write. Cold reset: 0. This bit may be used to distinguish between a cold versus a warm reset event by setting the bit to a 1 before an initialization event is generated. 3:1 0 Reserved. RouteTblDis: routing table disable. Read-write. Reset: 1. BIOS: 0. D18F0x84 Link Control Cold reset: 7711_0000h. This register is derived from the link control register defined in the link specification. Bits Description 31:13 Reserved. 12 IsocEn: isochronous flow-control mode enable. Read-write. BIOS: 1. This bit is set to place the link into isochronous flow-control mode (IFCM), as defined by the link specification. 1=IFCM. 0=Normal flow-control mode. See D18F0x6C[ApplyIsocModeEnNow]. 323 42300 Rev 3.12 - July 14, 2015 BKDG for AMD Family 15h Models 10h-1Fh Processors 11:6 Reserved. 5 Reserved. 4 LinkFail: link failure. Read; set-by-hardware; write-1-to-clear. This bit is set high by the hardware if a sync flood is received by the link. See 2.15.1.8.1 [Common Diagnosis Information]. 3:0 Reserved. D18F0x90 Upstream Base Channel Buffer Count D18F0x90 and D18F0x94 specify the hard-allocated flow-control buffer counts in each virtual channel; it also provides the free buffers that may be used by any of the virtual channels, as needed, or reallocated by BIOS to the hard-allocated buffer counts. Base channel buffers are specified in D18F0x90; isochronous buffer counts (if in IFCM) are specified in D18F0x94. For all fields that specify buffer counts in D18F0x90 and D18F0x94 the number of buffers allocated is 2 times the value of the field. The reset value is determined by whether the settings are locked by LockBc. IF (LockBc) THEN Cold reset: 0005_01C9h. ELSE (~LockBc) THEN Reset: 0005_01C9h. Out of cold reset, the processor allocates a minimal number of buffers that is smaller than the default values in the register. BIOS must use D18F0x6C[RlsLnkFullTokCntImm] for the values in the register to take effect. This is necessary even if the values are unchanged from the default values. The free list buffers (specified by FreeData and FreeCmd) are used to optimize buffer usage. When a transaction is received, if a free-list buffer is available, it is used for storage instead of one of the hard allocated buffers; as a result, a buffer release (for one of the hard allocated buffers used by the incoming request) can occur immediately without waiting for the transaction to be routed beyond the flow-control buffers. Buffer allocation rules: • The total number of command buffers allocated in the base and isochronous registers of a link cannot exceed 32. • D18F0x90[NpReqCmd] + D18F0x90[PReq] + D18F0x90[RspCmd] + D18F0x90[ProbeCmd] + D18F0x90[FreeCmd] + D18F0x94[IsocNpReqCmd] + D18F0x94[IsocPReq] + D18F0x94[IsocRspCmd] <= 16. • The total number of data buffers allocated in the base and isochronous registers of a link cannot exceed 16. • D18F0x90[NpReqData] + D18F0x90[RspData] + D18F0x90[PReq] + D18F0x90[FreeData] + D18F0x94[IsocPReq] + D18F0x94[IsocNpReqData] + D18F0x94[IsocRspData] <= 8. • The total number of hard allocated command buffers (ProbeCmd, RspCmd, PReq, NpReqCmd, and D18F0x94[IsocRspCmd, IsocRspCmd, IsocPReq, and IsocNpReqCmd]) cannot exceed 32. • D18F0x90[ProbeCmd] + D18F0x90[RspCmd] + D18F0x90[PReq] + D18F0x90[NpReqCmd] + D18F0x94[IsocRspCmd] + D18F0x94[IsocRspCmd] + D18F0x94[IsocPReq] + D18F0x94[IsocNpReqCmd] <= 16. • BIOS must set up non-zero counts (and adjust the base channel counts accordingly) prior to enabling IFCM. • IF (IFCM) THEN (D18F0x94[IsocNpReqCmd]) > 0. • If an IOMMU is present in the system, D18F0x94[IsocNpReqCmd] must be non-zero. • IF (IOMMU) THEN (D18F0x94[IsocNpReqCmd]) > 0. 324 42300 Rev 3.12 - July 14, 2015 BKDG for AMD Family 15h Models 10h-1Fh Processors Bits Description 31 LockBc: lock buffer count register. Read-write. Cold reset: 0. BIOS: 1. 1=The buffer count registers, D18F0x90 and D18F0x94 are locked such that warm resets do not place the registers back to their default value. Setting this bit does not prevent the buffer counts from being updated after a warm reset based on the value of the buffer counts before the warm reset. 0=Upon warm reset, the buffer count registers return to their default value after the link initializes regardless of the value before the warm reset. 30:28 Reserved. 27:25 FreeData: free data buffer count. Read-write. BIOS: 0. 24:20 FreeCmd: free command buffer count. Read-write. BIOS: 0. 19:18 RspData: response data buffer count. Read-write. BIOS: 1. 17:16 NpReqData: non-posted request data buffer count. Read-write. BIOS: 1. 15:12 ProbeCmd: probe command buffer count. Read-write. BIOS: 0. 11:8 RspCmd: response command buffer count. Read-write. BIOS: 2. 7:5 PReq: posted request command and data buffer count. Read-write. Specifies the number of posted command and posted data buffers allocated. BIOS: 5. 4:0 NpReqCmd: non-posted request command buffer count. Read-write. BIOS: 8. D18F0x94 Link Isochronous Channel Buffer Count See D18F0x90. The cold or warm reset is determined by whether the link initializes and whether the settings are locked by LockBc. IF (D18F0x90[LockBc]) THEN Cold reset: 0000_0000h. ELSE Reset: 0000_0000h. Bits Description 31:29 Reserved. 28:27 IsocRspData: isochronous response data buffer count. Read-write. BIOS: 0. 26:25 IsocNpReqData: isochronous non-posted request data buffer count. Read-write. BIOS: 1. 24:22 IsocRspCmd: isochronous response command buffer count. Read-write. BIOS: 0. 21:19 IsocPReq: isochronous posted request command and data buffer count. Read-write. BIOS: 0. This specifies the number of isochronous posted command and posted data buffers allocated. 18:16 IsocNpReqCmd: isochronous non-posted request command buffer count. Read-write. BIOS: 1. 15:8 SecBusNum: secondary bus number. Read-write. Specifies the configuration-space bus number of the IO link. When configured as a coherent link, this register has no meaning. This field should match the corresponding D18F1x[EC:E0][BusNumBase], unless D18F1x[EC:E0][DevCmpEn]=1, in which case this field should be 00h). 7:0 Reserved. 325 42300 Rev 3.12 - July 14, 2015 BKDG for AMD Family 15h Models 10h-1Fh Processors D18F0x98 Link Type Bits Description 31:6 Reserved. 5 4:3 PciEligible. Read-only. Reset: 1. Reserved. 2 Reserved. Reset: 1. 1 Reserved. Reset: 1. 0 Reserved. Reset: 1. D18F0x9C Link Frequency Extension Bits Description 31:16 Reserved. 15:1 Reserved. Reset: 1Fh. 0 Reserved. D18F0x110 Link Clumping Enable Reset: 0000_0000h. This register specifies how UnitIDs of upstream non-posted requests may be clumped per the link specification. The processor does not clump requests that it generates in the downstream direction. Bits Description 31:2 ClumpEn. Read-write. Each bit of this register corresponds to a link UnitID number. E.g., bit 2 corresponds to UnitID 02h, etc. 1=The specified UnitID is ordered in the same group as the specified UnitID - 1. For example if this register is programmed to 0000_00C0h, then UnitIDs 7h, 6h, and 5h are all ordered as if they are part of the same UnitID. This is used to allow more than 32 tags to be assigned to a single stream for the purposes of ordering. 1 Reserved. 0 Reserved. D18F0x16C Link Global Extended Control Reset: 0074_003Ah. Bits Description 31:0 Reserved. D18F0x170 Link Extended Control Reset: 0000_0001h. Bits Description 31:0 Reserved. 326 42300 Rev 3.12 - July 14, 2015 BKDG for AMD Family 15h Models 10h-1Fh Processors D18F0x1A0 Link Initialization Status Reset: 0000_0000h. Bits 31 Description InitStatusValid. Read-only. 1=Indicates that the rest of the information in this register is valid. 30:2 Reserved. 1:0 InitComplete. Read-only. Bits Description 00b Internal northbridge link has not completed initialization. 10b-01b Reserved. 11b Internal northbridge link has completed initialization. D18F0x1DC Core Enable Reset: 0000_0000h. Bits Description 31:8 Reserved. 7:1 0 CpuEn: core enable. Read-write. This field is used to enable each of the cores after a reset. 1=Enable the core to start fetching and executing code from the boot vector. [1]: Core 1 enable; ...; [N]: Core N enable. The most significant bit N is indicated by CpuCoreNum, as defined in section 2.4.3 [Processor Cores and Downcoring]. All bits greater than N are reserved. Reserved. 327 42300 Rev 3.12 - July 14, 2015 3.9 BKDG for AMD Family 15h Models 10h-1Fh Processors Device 18h Function 1 Configuration Registers See 3.1 [Register Descriptions and Mnemonics]. See 2.7 [Configuration Space]. D18F1x00 Device/Vendor ID Bits Description 31:16 DeviceID: device ID. Read-only. Value: 1401h. 15:0 VendorID: vendor ID. Read-only. Value: 1022h. D18F1x08 Class Code/Revision ID Bits Description 31:8 ClassCode. Read-only. Value: 060000h. Provides the host bridge class code as defined in the PCI specification. 7:0 RevID: revision ID. Read-only. Value: 00h. Processor revision. 00h=A0. D18F1x0C Header Type Reset: 0080_0000h. Bits Description 31:0 HeaderTypeReg. Read-only. These bits are fixed at their default values. The header type field indicates that there are multiple functions present in this device. D18F1x[144:140,44:40] DRAM Base/Limit The following sets of registers specify the DRAM address ranges: Base Low/High Limit Low/High F1x040, F1x140 F1x044, F1x144 F1x0XX registers provide the low address bits and F1x1XX registers provide the high address bits. Transaction addresses that are within the specified base/limit range are routed to the DstNode. See 2.8.2 [NB Routing]. DRAM mapping rules: • Transaction addresses are within the defined range if: {DramBase[47:24], 00_0000h} <= address[47:0] <= {DramLimit[47:24], FF_FFFFh}. • DRAM regions must not overlap each other. • Accesses to addresses that map to both DRAM, as specified by the DRAM base and limit registers (F1x[1, 0][7C:40]), and MMIO, as specified by D18F1x[1CC:180,BC:80], are routed to MMIO only. • Programming of the DRAM address maps must be consistent with the Memory-Type Range Registers (MTRRs) and the top of memory registers, MSRC001_001A and MSRC001_001D. CPU accesses only hit within the DRAM address maps if the corresponding MTRR is of type DRAM. Accesses from IO links are routed based on the DRAM base and limit registers (F1x[1, 0][7C:40]) only. • The appropriate RE or WE bit(s) must be set. When initializing a base/limit pair, the BIOS must write the [limit] register before either the RE or WE bit is set. When changing a base/limit pair that is already enabled, the BIOS should clear RE and WE before changing the address range. • See 2.8.2.1.1 [DRAM and MMIO Memory Space]. 328 42300 Rev 3.12 - July 14, 2015 BKDG for AMD Family 15h Models 10h-1Fh Processors Hoisting. When memory hoisting is enabled in a node via D18F1xF0[DramHtHoleValid, DramHoleValid], the corresponding BaseAddr/LimitAddr should be configured to account for the memory hoisted above the hole. See 2.9.7 [Memory Hoisting]. D18F1x40 DRAM Base Low IF (D18F2x118[LockDramCfg]) THEN Read-only. ELSE Read-write. ENDIF. Reset: 0000_0000h. Bits Description 31:16 DramBase[39:24]: DRAM base address register bits[39:24]. DramBase[47:24] = {D18F1x140[DramBase[47:40]], D18F1x40[DramBase[39:24]]}. 15:2 Reserved. 1 WE: write enable. 1=Writes to this address range are enabled. 0 RE: read enable. 1=Reads to this address range are enabled. D18F1x140 DRAM Base High Bits Description 31:8 Reserved. 7:0 DramBase[47:40]: DRAM base address register bits[47:40]. See: D18F1x40[DramBase[39:24]]. D18F1x44 DRAM Limit Low IF (D18F2x118[LockDramCfg]) THEN Read-only. ELSE Read-write. ENDIF. Bits Description 31:16 DramLimit[39:24]: DRAM limit address register bits[39:24]. Reset: 00FCFFh. DramLimit[47:24] = {D18F1x144[DramLimit[47:40]], D18F1x44[DramLimit[39:24]]}. 15:11 Reserved. 10:8 Reserved. 7:3 Reserved. 2:0 DstNode: destination Node ID. Reset: 000b. Specifies the node that a packet is routed to if it is within the address range. D18F1x144 DRAM Limit High Bits Description 31:8 Reserved. 7:0 DramLimit[47:40]: DRAM limit address register bits[47:40]. Reset: 00h. See: D18F1x44[DramLimit[39:24]]. D18F1x[1CC:180,BC:80] MMIO Base/Limit These registers, The memory mapped IO base and limit registers D18F1x[1CC:180,BC:80] specify the map- 329 42300 Rev 3.12 - July 14, 2015 BKDG for AMD Family 15h Models 10h-1Fh Processors ping from memory addresses to the corresponding node and IO link for MMIO transactions. Address ranges are specified by 12 sets of base/limit registers. Range 0 1 2 3 4 5 6 7 8 9 10 11 Base Low,High F1x80, F1x180 F1x88, F1x184 F1x90, F1x188 F1x98, F1x18C F1xA0, F1x190 F1xA8, F1x194 F1xB0, F1x198 F1xB8, F1x19C F1x1A0, F1x1C0 F1x1A8, F1x1C4 F1x1B0, F1x1C8 F1x1B8, F1x1CC Limit Low,High F1x84, F1x180 F1x8C, F1x184 F1x94, F1x188 F1x9C, F1x18C F1xA4, F1x190 F1xAC, F1x194 F1xB4, F1x198 F1xBC, F1x19C F1x1A4, F1x1C0 F1x1AC, F1x1C4 F1x1B4, F1x1C8 F1x1BC, F1x1CC Transaction addresses that are within the specified base/limit range are routed to the node specified by DstNode and the link specified by DstLink. See 2.8.2 [NB Routing]. .MMIO mapping rules: • Transaction addresses are within the defined range if: {MMIOBase[47:16], 0000h} <= address[47:0] <= {MMIOLimit[47:16], FFFFh}. • MMIO regions must not overlap each other. • Accesses to addresses that map to both DRAM, as specified by the DRAM base and limit registers (see D18F1x[144:140,44:40]), and MMIO, as specified by the memory mapped IO base and limit registers (F1x[BC:80]), are routed to MMIO only. • Programming of the MMIO address maps must be consistent with the Memory-Type Range Registers (MTRRs) and the top of memory registers, MSRC001_001A and MSRC001_001D. CPU accesses only hit within the MMIO address maps if the corresponding MTRR is of type IO. Accesses from IO links are routed based on D18F1x[1CC:180,BC:80]. • The appropriate RE or WE bit(s) must be set. When initializing a base/limit pair, the BIOS must write the limit register before either the RE or WE bit is set. When changing a base/limit pair that is already enabled, the BIOS should clear RE and WE before changing the address range. • Scenarios in which the address space of multiple MMIO ranges target the same IO device is supported. • See 2.8.2.1.1 [DRAM and MMIO Memory Space]. D18F1x[1B8,1B0,1A8,1A0,B8,B0,A8,A0,98,90,88,80] MMIO Base Low Table 124: Register Mapping for D18F1x[1B8,1B0,1A8,1A0,B8,B0,A8,A0,98,90,88,80] Register D18F1x80 D18F1x88 D18F1x90 D18F1x98 D18F1xA0 D18F1xA8 D18F1xB0 Function Range 0 Range 1 Range 2 Range 3 Range 4 Range 5 Range 6 330 42300 Rev 3.12 - July 14, 2015 BKDG for AMD Family 15h Models 10h-1Fh Processors Table 124: Register Mapping for D18F1x[1B8,1B0,1A8,1A0,B8,B0,A8,A0,98,90,88,80] D18F1xB8 D18F1x1A0 D18F1x1A8 D18F1x1B0 D18F1x1B8 Bits Range 7 Range 8 Range 9 Range 10 Range 11 Description 31:8 MMIOBase[39:16]: MMIO base address register bits[39:16]. Read-write. Reset: 0. MMIOBase[47:16] = {D18F1x[1CC:1C0,19C:180][MMIOBase[47:40]], MMIOBase[39:16]}. 7:4 Reserved. 3 Lock. Read-write. Reset: 0. 1=the memory mapped IO base and limit registers (D18F1x[1CC:180,BC:80]) are read-only (including this bit). WE or RE in this register must be set in order for this to take effect. 2 Reserved. 1 WE: write enable. Read-write. Reset: 0. 1=Writes to this address range are enabled. 0 RE: read enable. Read-write. Reset: 0. 1=Reads to this address range are enabled. D18F1x[1BC,1B4,1AC,1A4,BC,B4,AC,A4,9C,94,8C,84] MMIO Limit Low Table 125: Register Mapping for D18F1x[1BC,1B4,1AC,1A4,BC,B4,AC,A4,9C,94,8C,84] Register D18F1x84 D18F1x8C D18F1x94 D18F1x9C D18F1xA4 D18F1xAC D18F1xB4 D18F1xBC D18F1x1A4 D18F1x1AC D18F1x1B4 D18F1x1BC Bits Function Range 0 Range 1 Range 2 Range 3 Range 4 Range 5 Range 6 Range 7 Range 8 Range 9 Range 10 Range 11 Description 31:8 MMIOLimit[39:16]: MMIO limit address register bits[39:16]. Read-write. Reset: 0. MMIOLimit[47:16] = {D18F1x[1CC:1C0,19C:180][MMIOLimit[47:40]], MMIOLimit[39:16]}. 331 42300 Rev 3.12 - July 14, 2015 7 BKDG for AMD Family 15h Models 10h-1Fh Processors NP: non-posted. Read-write. Reset: 0. 1=CPU write requests to this MMIO range are passed through the non-posted channel. This may be used to force writes to be non-posted for MMIO regions which map to the legacy ISA/LPC bus, or in conjunction with D18F0x68 [Link Transaction Control][DsNpReqLmt] in order to allow downstream CPU requests to be counted and thereby limited to a specified number. This latter use of the NP bit may be used to avoid loop deadlock scenarios in systems that implement a region in an IO device that reflects downstream accesses back upstream. See the link summary of deadlock scenarios for more information. 0=CPU writes to this MMIO range use the posted channel. This bit does not affect requests that come from IO links (the virtual channel of the request is specified by the IO request). If two MMIO ranges target the same IO device and the NP bit is set differently in both ranges, unexpected transaction ordering effects are possible. In particular, using PCI- and IO-link-defined producer-consumer semantics, if a producer (e.g., the processor) writes data using a non-posted MMIO range followed by a flag to a posted MMIO range, then it is possible for the device to see the flag updated before the data is updated. 6 5:4 3 2:0 DstSubLink: destination sublink. Read-write. Reset: 0. When a link is unganged, this bit specifies the destination sublink of the link specified by the memory mapped IO base and limit registers F1x[BC:80][DstLink]. 0=The destination link is sublink 0. 1=The destination link is sublink 1. If the link is ganged, then this bit must be low. DstLink: destination link ID. Read-write. Reset: 0. For transactions within the this MMIO range, this field specifies the destination IO link number of the destination node. Bits Description 00b Link 0 01b Link 1 10b Link 2 11b Link 3 Reserved. DstNode: destination node ID bits. Read-write. Reset: 0. For transactions within the this MMIO range, this field specifies the destination node ID. D18F1x[1CC:1C0,19C:180] MMIO Base/Limit High Table 126: Register Mapping for D18F1x[D18F1x[1CC:1C0,19C:180] Register D18F1x180 D18F1x184 D18F1x188 D18F1x18C D18F1x190 D18F1x194 D18F1x198 D18F1x19C D18F1x1C0 Function Range 0 Range 1 Range 2 Range 3 Range 4 Range 5 Range 6 Range 7 Range 8 332 42300 Rev 3.12 - July 14, 2015 BKDG for AMD Family 15h Models 10h-1Fh Processors Table 126: Register Mapping for D18F1x[D18F1x[1CC:1C0,19C:180] D18F1x1C4 D18F1x1C8 D18F1x1CC Bits Range 9 Range 10 Range 11 Description 31:24 Reserved. 23:16 MMIOLimit[47:40]: MMIO limit address register bits[47:40]. See: D18F1x[1BC,1B4,1AC,1A4,BC,B4,AC,A4,9C,94,8C,84][MMIOLimit[39:16]]. 15:8 Reserved. 7:0 MMIOBase[47:40]: MMIO base address register bits[47:40]. See: D18F1x[1B8,1B0,1A8,1A0,B8,B0,A8,A0,98,90,88,80][MMIOBase[39:16]]. D18F1x[DC:C0] IO-Space Base/Limit The IO-space base and limit registers, D18F1x[DC:C0], specify the mapping from IO addresses to the corresponding node and IO link for transactions resulting from x86-defined IN and OUT instructions. IO address ranges are specified by 4 sets of base/limit registers. The first set is F1xC0 and F1xC4, the second set is F1xC8 and F1xCC, and so forth. Transaction addresses that are within the specified base/limit range are routed to the node specified by DstNode and the link specified by DstLink. See 2.8.2 [NB Routing]. IO mapping rules: • IO-space transaction addresses are within the defined range if: {IOBase[24:12], 000h} <= address <= {IOLimit[24:12], FFFh} and as specified by the IE bit; or if the address is in the range specified by the VE bits. • IO regions must not overlap each other. • The appropriate RE or WE bit(s) must be set. • See 2.8.2.1.2 [IO Space]. D18F1x[D8,D0,C8,C0] IO-Space Base Table 127: Register Mapping for D18F1x[D8,D0,C8,C0] Register D18F1xC0 D18F1xC8 D18F1xD0 D18F1xD8 Bits Function Range 0 Range 1 Range 2 Range 3 Description 31:25 Reserved. 24:12 IOBase[24:12]: IO base address register bits[24:12]. Read-write. Reset: 0. 11:6 Reserved. 333 42300 Rev 3.12 - July 14, 2015 BKDG for AMD Family 15h Models 10h-1Fh Processors 5 IE: ISA enable. Read-write. Reset: 0. 1=The IO-space address window is limited to the first 256 B of each 1 KB block specified; this only applies to the first 64 KB of IO space. 0=The PCI IO window is not limited in this way. 4 VE: VGA enable. Read-write. Reset: 0. 1=Include IO-space transactions targeting the VGAcompatible address space within the IO-space window of this base/limit pair. These include IO accesses in which address bits[9:0] range from 3B0h to 3BBh or 3C0h to 3DFh (address bits[15:10] are not decoded); this only applies to the first 64 KB of IO space; i.e., address bits[24:16] must be low). 0=IO-space transactions targeting VGA-compatible address ranges are not added to the IOspace window. This bit should only be set in one register. The MMIO range associated with the VGA enable bit in the PCI specification is NOT included in the VE bit definition; to map this range to an IO link, see D18F1xF4 [VGA Enable]. When D18F1xF4[VE] is set, the state of this bit is ignored. 3:2 Reserved. 1 WE: write enable. Read-write. Reset: 0. 1=Writes to this IO-space address range are enabled. 0 RE: read enable. Read-write. Reset: 0. 1=Reads to this IO-space address range are enabled. D18F1x[DC,D4,CC,C4] IO-Space Limit Table 128: Register Mapping for D18F1x[DC,D4,CC,C4] Register D18F1xC4 D18F1xCC D18F1xD4 D18F1xDC Bits Function Range 0 Range 1 Range 2 Range 3 Description 31:25 Reserved. 24:12 IOLimit[24:12]: IO limit address register bits[24:12]. Read-write. Reset: 0. 11:7 Reserved. 6 DstSubLink: destination sublink. Read-write. Reset: 0. When a link is unganged, this bit specifies the destination sublink of the link specified by F1x[DC:C0][DstLink]. 0=The destination link is sublink 0. 1=The destination link is sublink 1. If the link is ganged, then this bit must be low. 5:4 DstLink: destination link ID. Read-write. Reset: 0. For transactions within the this IO-space range, this field specifies the destination IO link number of the destination node. Bits Description 00b Link 0 01b Link 1 10b Link 2 11b Link 3 3 2:0 Reserved. DstNode: destination node ID bits. Read-write. Reset: 0. For transactions within the this IO-space range, this field specifies the destination node ID. 334 42300 Rev 3.12 - July 14, 2015 BKDG for AMD Family 15h Models 10h-1Fh Processors D18F1x[EC:E0] Configuration Map D18F1x[EC:E0] specify the mapping from configuration address to the corresponding node and IO link. Configuration address ranges are specified by 4 pairs of base/limit registers. Transaction addresses that are within the specified base/limit range are routed to the node specified by DstNode and the link specified by DstLink. See 2.8.2 [NB Routing]. Table 129: Register Mapping for D18F1x[EC:E0] Register D18F1xE0 D18F1xE4 D18F1xE8 D18F1xEC Function Range 0 Range 1 Range 2 Range 3 Configuration space mapping rules: • Configuration addresses (to“BusNo” and “Device” as specified by IOCF8 [IO-Space Configuration Address] in the case of IO accesses or 2.7 [Configuration Space] in the case of MMIO accesses) are within the defined range if: ( {BusNumBase[7:0]} <= BusNo <= {BusNumLimit[7:0]} ) & (DevCmpEn==0); or ( {BusNumBase[4:0]} <= Device <= {BusNumLimit[4:0]} ) & (DevCmpEn==1) & (BusNo == 00h). • Configuration regions must not overlap each other. • The appropriate RE or WE bit(s) must be set. • See 2.8.2.1.3 [Configuration Space]. Bits Description 31:24 BusNumLimit[7:0]: bus number limit bits[7:0]. Read-write. Reset: 0. 23:16 BusNumBase[7:0]: bus number base bits[7:0]. Read-write. Reset: 0. 15:3 Reserved. 2 DevCmpEn: device number compare mode enable. Read-write. Reset: 0. 1=A device number range rather than a bus number range is used to specify the configuration-space window (see above). 1 WE: write enable. Read-write. Reset: 0. 1=Writes to this configuration-space address range are enabled. 0 RE: read enable. Read-write. Reset: 0. 1=Reads to this configuration-space address range are enabled. D18F1xF0 DRAM Hole Address IF (D18F2x118[LockDramCfg]) THEN Read-only. ELSE Read-write. ENDIF. See 2.9.7 [Memory Hoisting]. Bits Description 31:24 DramHoleBase[31:24]: DRAM hole base address. Reset: 0. Specifies the base address of the IO hole, below the 4GB address level, that is used in memory hoisting. Normally, DramHoleBase >= MSRC001_001A[TOM[31:24]]. 23:16 Reserved. 15:7 DramHoleOffset[31:23]: DRAM hole offset address. Reset: 0. When DramHoleValid=1, this offset is subtracted from the physical address of certain accesses in forming the normalized address. 335 42300 Rev 3.12 - July 14, 2015 BKDG for AMD Family 15h Models 10h-1Fh Processors 6:3 Reserved. 2 Reserved. 1 DramMemHoistValid: dram memory hoist valid. Reset: 0. 1=Memory hoisting for the address range is enabled. 0=Memory hoisting is not enabled. This bit should be set if DramHoleValid=1. 0 DramHoleValid: dram hole valid. Reset: 0. 1=Memory hoisting is enabled. 0=Memory hoisting is not enabled. This bit should be set in the node(s) that own the DRAM address space that is hoisted above the 4 GB address level. See DramHoleOffset. D18F1xF4 VGA Enable Reset: 0000_0000h. All these bits are read-write unless Lock is set. Bits Description 31:15 Reserved. 14 DstSubLink: destination sublink. Read-write. When a link is unganged, this bit specifies the destination sublink of the link specified by D18F1xF4[DstLink]. 0=The destination link is sublink 0. 1=The destination link is sublink 1. If the link is ganged, then this bit must be low. 13:12 DstLink: destination link ID. Read-write. For transactions within the D18F1xF4[VE]-defined ranges, this field specifies the destination IO link number of the destination node. Bits Description 00b Link 0 01b Link 1 10b Link 2 11b Link 3 11:7 Reserved. 6:4 DstNode: destination node ID. Read-write. For transactions within the D18F1xF4[VE]-defined range, this field specifies the destination node ID. 3 Lock. Read-write. 1=All the bits in this register (D18F1xF4) are read-only (including this bit). 2 Reserved. 1 NP: non-posted. Read-write. 1=CPU write requests to the D18F1xF4[VE]-defined MMIO range are passed through the non-posted channel. 0=CPU writes may be posted. 0 VE: VGA enable. Read-write. 1=Transactions targeting the VGA-compatible address space are routed and controlled as specified by this register. The VGA-compatible address space is: (1) the MMIO range A_0000h through B_FFFFh; (2) IO-space accesses in which address bits[9:0] range from 3B0h to 3BBh or 3C0h to 3DFh (address bits[15:10] are not decoded; this only applies to the first 64 KB of IO space; i.e., address bits[24:16] must be low). 0=Transactions targeting the VGAcompatible address space are not affected by the state of this register. When this bit is set, the state of D18F1x[DC:C0][VE] is ignored. D18F1x10C DCT Configuration Select Reset: 0000_0000h. Bits Description 31:6 Unused. 336 42300 Rev 3.12 - July 14, 2015 BKDG for AMD Family 15h Models 10h-1Fh Processors 5:4 NbPsSel: NB P-state configuration select. Read-write. Read-write. Specifies the set of DCT Pstate registers to which accesses are routed. Description Bits 00b NB P-state 0 01b NB P-state 1 10b NB P-state 2 11b NB P-state 3 The following registers must be programmed for each NB P-state enabled by D18F5x1[6C:60][NbPstateEn]: • D18F2x210_dct[1:0]_nbp[3:0][MaxRdLatency, DataTxFifoWrDly, RdPtrInit]. 3 MemPsSel: Memory P-state configuration select. Read-write. Specifies the set of DCT controller registers to which accesses are routed. This register works independently of NbPsSel. 0=Memory Pstate 0. 1=Memory P-state 1. See 2.5.7.1 [Memory P-states] and 2.9.1 [DCT Configuration Registers]. The following registers must be programmed for each memory P-state enabled by D18F5x1[6C:60][MemPstate]: • D18F2x2E8_dct[1:0]_mp[1:0] • D18F2x2EC_dct[1:0]_mp[1:0] 2:1 0 Reserved. DctCfgSel: DRAM controller configuration select. Read-write. Specifies DCT controller to which accesses are routed. 0=DCT 0. 1=DCT 1. See 2.9.1 [DCT Configuration Registers]. D18F1x120 DRAM Base System Address IF (D18F2x118[LockDramCfg]) THEN Read-only. ELSE Read-write. ENDIF. Cold reset: 0000_0000h. D18F1x120 and D18F1x124 are required to specify the base and limit system address range of the DRAM connected to the local node. DRAM accesses to the local node with physical address Addr[47:0] that are within the following range are directed to the DCTs: {DramBaseAddr[47:27], 000_0000h} <= Addr[47:0] <= {DramLimitAddr[47:27], 7FF_FFFFh}; DRAM accesses to the local node that are outside of this range are master aborted. DRAM may be mapped as continuous regions for each node or it may be interleaved between nodes. If node interleaving is not invoked, as specified by DramIntlvEn, then the address of the DRAM transaction is normalized before passing it to the DCTs by subtracting DramBaseAddr. Bits Description 31:21 Reserved. 20:0 DramBaseAddr[47:27]. D18F1x124 DRAM Limit System Address IF (D18F2x118[LockDramCfg]) THEN Read-only. ELSE Read-write. ENDIF. See D18F1x120 [DRAM Base System Address]. Bits Description 31:21 Reserved 20:0 DramLimitAddr[47:27]. Cold reset: 1F_FFFFh. 337 42300 Rev 3.12 - July 14, 2015 BKDG for AMD Family 15h Models 10h-1Fh Processors 3.10 Device 18h Function 2 Configuration Registers See 3.1 [Register Descriptions and Mnemonics]. See 2.7 [Configuration Space]. D18F2x00 Device/Vendor ID Bits Description 31:16 DeviceID: device ID. Read-only. Value: 1402h. 15:0 VendorID: vendor ID. Read-only. Value: 1022h. D18F2x08 Class Code/Revision ID Reset: 0600_0000h. Bits Description 31:8 ClassCode. Read-only. Provides the host bridge class code as defined in the PCI specification. 7:0 RevID: revision ID. Read-only. D18F2x0C Header Type Reset: 0080_0000h. Bits Description 31:0 HeaderTypeReg. Read-only. These bits are fixed at their default values. The header type field indicates that there multiple functions present in this device. D18F2x[5C:40]_dct[1:0] DRAM CS Base Address IF (D18F2x118[LockDramCfg]) THEN Read-only. ELSE Read-write. ENDIF. Reset: 0000_0000h. See 2.9.1 [DCT Configuration Registers]. These registers along with D18F2x[6C:60]_dct[1:0] [DRAM CS Mask], translate DRAM request addresses (to a DRAM controller) into DRAM chip selects. Supported DIMM sizes are specified in D18F2x80_dct[1:0] [DRAM Bank Address Mapping]. For more information on the DRAM controllers, see 2.9 [DRAM Controllers (DCTs)]. For each chip select, there is a DRAM CS Base Address register. For every two chip selects there is a DRAM CS Mask Register. Table 130: DIMM, Chip Select, and Register Mapping Base Address Registers D18F2x40_dct[1:0] D18F2x44_dct[1:0] D18F2x48_dct[1:0] D18F2x4C_dct[1:0] D18F2x50_dct[1:0] D18F2x54_dct[1:0] Mask Register F2x60 Logical DIMM 0 F2x64 1 F2x68 Reserved Chip Select1 MEMCS[1:0]_L[0] MEMCS[1:0]_L[1] MEMCS[1:0]_L[2] MEMCS[1:0]_L[3] Reserved. Reserved. 338 42300 Rev 3.12 - July 14, 2015 BKDG for AMD Family 15h Models 10h-1Fh Processors Table 130: DIMM, Chip Select, and Register Mapping Base Address Registers Mask Register F2x6C Logical DIMM Reserved D18F2x58_dct[1:0] D18F2x5C_dct[1:0] 1. See 2.9.2 [DDR Pad to Processor Pin Mapping] Chip Select1 Reserved. Reserved. The DRAM controller operates on the normalized physical address of the DRAM request. The normalized physical address includes all of the address bits that are supported by a DRAM controller. See 2.8 [Northbridge (NB)]. Each base address register specifies the starting normalized address of the block of memory associated with the chip select. Each mask register specifies the additional address bits that are consumed by the block of memory associated with the chip selects. If both chip selects of a DIMM are used, they must be the same size; in this case, a single mask register covers the address space consumed by both chip selects. Lower-order address bits are provided in the base address and mask registers, as well. These allow memory to be interleaved between chip selects, such that contiguous physical addresses map to the same DRAM page of multiple chip selects. See 2.9.6.1 [Chip Select Interleaving]. The hardware supports the use of lower-order address bits to interleave chip selects if (1) the each chip select of the memory system spans the same amount of memory and (2) the number of chip selects of the memory system is a power of two. System BIOS is required to assign the largest DIMM chip-select range to the lowest normalized address of the DRAM controller. As addresses increase, the chip-select size is required to remain constant or decrease. This is necessary to keep DIMM chip-select banks on aligned address boundaries, regardless as to the amount of address space covered by each chip select. For each normalized address for requests that enters a DRAM controller, a ChipSelect[i] is asserted if: CSEnable[i] & ( {(InputAddr[38:27] & ~AddrMask[i][38:27]), (InputAddr[21:11] & ~AddrMask[i][21:11])} == {(BaseAddr[i][38:27] & ~AddrMask[i][38:27]), (BaseAddr[i][21:11] & ~AddrMask[i][21:11])} ); Bits 31 Description Reserved. 30:19 BaseAddr[38:27]: normalized physical base address bits [38:27]. 18:16 Reserved. 15:5 BaseAddr[21:11]: normalized physical base address bits [21:11]. 4 Reserved. 339 42300 Rev 3.12 - July 14, 2015 BKDG for AMD Family 15h Models 10h-1Fh Processors 3 OnDimmMirror: on-DIMM mirroring (ODM) enabled. 1=Address and bank bits are swapped by hardware for MRS commands sent to this chip select. This mode accounts for routing on the DIMM. Hardware bit swapping does not occur for commands sent via D18F2x7C_dct[1:0][SendMrsCmd] when D18F2x7C_dct[1:0][EnDramInit] = 0. This bit is expected to be set for the odd numbered rank of unbuffered DDR3 DIMMs if SPD byte 63 indicates that address mapping is mirrored. See 2.9.5.8.1.1 [DDR3 MR Initialization]. The following bits are swapped when enabled: • BA0 and BA1. • A3 and A4. • A5 and A6. • A7 and A8. 2 TestFail: memory test failed. Set by BIOS to indicate that a rank is present but has failed memory training or a memory consistency test, indicating that the memory is bad. BIOS should treat CSEnable=1 and TestFail=1 as mutually exclusive. 1 Reserved. 0 CSEnable: chip select enable. D18F2x[6C:60]_dct[1:0] DRAM CS Mask IF (D18F2x118[LockDramCfg]) THEN Read-only. ELSE Read-write. ENDIF. Reset: 0000_0000h. See 2.9.1 [DCT Configuration Registers]. See D18F2x[5C:40]_dct[1:0]. Bits 31 Description Reserved. 30:19 AddrMask[38:27]: normalized physical address mask bits [38:27]. 18:16 Reserved. 15:5 AddrMask[21:11]: normalized physical address mask bits [21:11]. 4:0 Reserved. D18F2x78_dct[1:0] DRAM Control See 2.9.1 [DCT Configuration Registers]. Bits Description 31:18 Reserved. 17 AddrCmdTriEn: address command tristate enable. Read-write. Reset: 0. BIOS: See 2.9.5.7.1=Tristate the address, command, and bank buses when a Deselect command is issued. 16:0 Reserved. D18F2x7C_dct[1:0] DRAM Initialization Reset: 0000_0000h. See 2.9.1 [DCT Configuration Registers]. See 2.9.5.8.1 [Software DDR3 Device Initialization]. BIOS can directly control the DRAM initialization sequence using this register. BIOS sets EnDramInit to start DRAM initialization. BIOS should then complete the initialization sequence specified in the appropriate JEDEC specification. After completing the sequence, BIOS clears EnDramInit to complete DRAM initializa- 340 42300 Rev 3.12 - July 14, 2015 BKDG for AMD Family 15h Models 10h-1Fh Processors tion. Setting more than one of the command bits in this register (SendControlWord, SendMrsCmd, and SendAutoRefresh) at a time results in undefined behavior. Bits Description 31 EnDramInit: enable DRAM initialization. Read-write. 1=Place the DRAM controller in the BIOScontrolled DRAM initialization mode. The DCT asserts memory reset and deasserts CKE when this bit is set. BIOS must wait until D18F2x98_dct[1:0][DctAccessDone] = 1 before programming AssertCke=1 and DeassertMemRstX=1. BIOS must clear this bit after DRAM initialization is complete. BIOS must not set this bit on a DCT with no attached DIMMs. See 2.9.5.8.1 [Software DDR3 Device Initialization]. 30 SendControlWord: send control word. Read; write-1-only; cleared-by-hardware. 1= The DCT sends a control word to a chip select pair defined in D18F2xA8_dct[1:0][CtrlWordCS]. This bit is cleared by hardware after the command completes. This bit is valid only when D18F2x90_dct[1:0][UnbuffDimm] = 0. 29 SendZQCmd: send ZQ command. Read; write-1-only; cleared-by-hardware. 1=The DCT sends the ZQ calibration command with either all even or all odd chip selects active. The first command targets even chip selects. Subsequent commands alternate between even and odd chip selects. This bit is cleared by the hardware after the command completes. This bit is valid only when EnDramInit=1. 28 AssertCke: assert CKE. Read-write. Setting this bit causes the DCT to assert the CKE pins. This bit cannot be used to deassert the CKE pins. 27 DeassertMemRstX: deassert memory reset. Read-write. Setting this bit causes the DCT to deassert the memory reset. This bit cannot be used to assert the memory reset pin. 26 SendMrsCmd: send MRS command. Read; write-1-only; cleared-by-hardware. 1=The DCT sends the MRS commands defined by the MrsAddress and MrsBank fields of this register. This bit is cleared by hardware after the command completes. See 2.9.5.8.1.1 [DDR3 MR Initialization]. 25 SendAutoRefresh: send auto refresh command. Read; write-1-only; cleared-by-hardware. 1=The DCT sends an auto refresh command. This bit is cleared by hardware after the command completes. 24 Reserved. 23:21 MrsChipSel: MRS command chip select. Read-write. Specifies which DRAM chip select is used for MRS commands. Defined only if (~EnDramInit | ~D18F2x90_dct[1:0][UnbuffDimm]); otherwise MRS commands are sent to all chip selects. Bits Description 000b MRS command is sent to CS0 110b-001b MRS command is sent to CS<MrsChipSel> 111b MRS command is sent to CS7 20:18 MrsBank[2:0]: bank address for MRS commands. Read-write. Specifies the data driven on the DRAM bank pins for MRS commands. 17:0 MrsAddress[17:0]: address for MRS commands. Read-write. Specifies the data driven on the DRAM address pins for MRS commands. D18F2x80_dct[1:0] DRAM Bank Address Mapping IF (D18F2x118[LockDramCfg]) THEN Read-only. ELSE Read-write. ENDIF. Reset: 0000_0000h. See 2.9.1 [DCT Configuration Registers]. These fields specify DIMM configuration information. These fields are required to be programmed per the following table, based on the DRAM device size and with information of the DIMM. Table 131 show the bit numbers for each position. 341 42300 Rev 3.12 - July 14, 2015 Bits BKDG for AMD Family 15h Models 10h-1Fh Processors Description 31:16 Reserved. 15:12 DimmAddrMap3: DIMM 3 address map. 11:8 DimmAddrMap2: DIMM 2 address map. 7:4 DimmAddrMap1: DIMM 1 address map. 3:0 DimmAddrMap0: DIMM 0 address map. Table 131: DDR3 DRAM Address Mapping Device size, Bits CS Size 0000b width Bank 2 1 Address 0 Reserved 15 14 13 12 11 10 9 8 7 6 5 4 3 2 1 0 Row x x x x 17 16 27 26 25 24 23 22 21 20 19 18 Col x x x x x Row Col 0001b 256MB 0010b 512MB 0011b 512Mb, x16 512Mb, x8 15 14 13 15 14 13 AP 12 11 10 9 8 7 6 5 4 3 Row x x x 17 16 28 27 26 25 24 23 22 21 20 19 18 1Gb, x16 Col x x x x Reserved Row Row x x 17 16 29 28 27 26 25 24 23 22 21 20 19 18 Col x x x Row x 17 16 30 29 28 27 26 25 24 23 22 21 20 19 18 Col x x Row 17 16 31 30 29 28 27 26 25 24 23 22 21 20 19 18 Col x Row 17 32 31 30 29 28 27 26 25 24 23 22 21 20 19 18 Col x x AP 12 11 10 9 8 7 6 5 4 3 Col 0100b Reserved Row Col 0101b 1GB 1Gb, x8 15 14 13 2Gb, x16 0110b 0111b 2GB 2Gb, x8 1000b Reserved 1001b Reserved 15 14 13 4GB 4Gb, x8 15 14 13 8Gb, x16 1011b 1100b x AP 12 11 10 9 8 7 6 5 4 3 Reserved 4Gb, x16 1010b x 8GB 8Gb, x8 16 15 14 x x x x x x x x x x AP 12 AP 12 13 AP 12 11 11 11 10 10 10 9 9 9 8 8 8 7 7 7 6 6 6 5 5 5 4 4 4 3 3 3 Reserved D18F2x84_dct[1:0] DRAM MRS Reset: 0000_0005h. See 2.9.1 [DCT Configuration Registers]. Bits Description 31:24 Reserved. 342 42300 Rev 3.12 - July 14, 2015 23 BKDG for AMD Family 15h Models 10h-1Fh Processors PchgPDModeSel: precharge power down mode select. Read-write. BIOS: 1. Specifies how a chip select enters and exits power down mode. This mode is enabled by D18F2x94_dct[1:0][PowerDownEn] and its behavior varies based on the setting of D18F2x94_dct[1:0][PowerDownMode]and MR0[PPD] in Table 38. See 2.9.5.8 [DRAM Device and Controller Initialization]. PowerDownMode PchgPDModeSel MR0[PPD] Description 0b 0b 0b Full channel slow exit (DLL off) 0b 0b 1b Full channel fast exit (DLL on) 0b 1b xb Full channel dynamic fast exit/slow exit 1b 0b 0b Reserved (Full channel slow exit) 1b 0b 1b Reserved 1b 1b xb Partial channel dynamic fast exit/slow exit See D18F2x248_dct[1:0]_mp[1:0][Txpdll, Txp]. In dynamic fast exit/slow exit power down mode, the DCT dynamically issues MRS command(s) to the DRAM to specify the powerdown mode; the DCT specifies fast exit mode when chip selects on one of the two CKEs has recently been active; it specifies deep power down when chip selects on all CKEs have been idle. 22:2 Reserved. 1:0 BurstCtrl: burst length control. Read-write. BIOS: 01b. Specifies the number of sequential beats of DQ related to one read or write command. Requests from the processor are always 64-byte-length. Requests generated by D18F2x250_dct[1:0] are always 64-byte-length. Requests from GMC may be 32-byte or 64-byte-length. Software must ensure that GMC requests are disabled to configure the controller and drams for 8-beat burst length (e.g. during training). If this mode is changed, software must issue a mode-register set command to MR0 of the drams to place them in the same mode. Description Bits 00b 8 beats 01b dynamic 4 or 8 beats 11b-10b Reserved D18F2x88_dct[1:0] DRAM Timing Low See 2.9.1 [DCT Configuration Registers]. Bits Description 31:30 Reserved. 29:24 MemClkDis: MEMCLK disable. Read-write. Reset: 3Fh. 1=Disable the MEMCLK. 0=Enable MEMCLK. BIOS: See 2.9.5.11. All enabled clocks should be 0; all disabled clocks should be 1. Bit Pad [0] MEMCLK[1,0]_H[0] [1] MEMCLK[1,0]_H[1] [2] MEMCLK[1,0]_H[2] [3] MEMCLK[1,0]_H[3] [4] MEMCLK[1,0]_H[4] [5] MEMCLK[1,0]_H[5] 23:0 Reserved. 343 42300 Rev 3.12 - July 14, 2015 BKDG for AMD Family 15h Models 10h-1Fh Processors D18F2x8C_dct[1:0] DRAM Timing High Reset: 0000_0000h. See 2.9.1 [DCT Configuration Registers]. Bits Description 31:19 Reserved. 18 DisAutoRefresh: disable automatic refresh. Read-write. BIOS: See 2.9.5.7. 1=Automatic refresh is disabled. See 2.9.5.9.1 [Write Levelization Training] and 2.9.5.9.6 [Continuous Pattern Generation]. 17:16 Tref: refresh rate. Read-write. This specifies the average time between refresh requests to all DRAM devices. Bits Description 00b Undefined behavior. 01b Reserved 10b Every 7.8 us 11b Every 3.9 us 15:0 Reserved. D18F2x90_dct[1:0] DRAM Configuration Low See 2.9.1 [DCT Configuration Registers]. Bits Description 31:28 Reserved. 27 DisDllShutdownSR: disable DLL shutdown in self-refresh mode. Read-write. Reset: 1. BIOS: See 2.9.5.4.2.1. 1=Disable the power saving features of shutting down DDR phy DLLs during DRAM self refresh and memory P-states. 0=Shutdown DLLs during DRAM self refresh and allow memory Pstate transitions. Setting this bit does not effect the current memory P-state. 26 Reserved. 25 PendRefPaybackS3En: pending refresh payback S3 enable. Read-write. Reset: 0. BIOS: 1. Specifies the S3 refresh payback behavior when PendRefPayback=0. 1=Pending refreshes are paid back on S3 entry. 0=Pending refreshes are not paid back on S3 entry. 24 StagRefEn: Stagger Refresh Enable. Read-write. Reset: 0. BIOS: 11=The DRAM controller arbitrates refreshes among chip selects based on the Tstag value . See D18F2x228_dct[1:0]. 0=DCT arbitrates among chip selects using the Trfc value. See D18F2x208_dct[1:0]. 23 ForceAutoPchg: force auto precharging. Read-write. Reset: 0. See 2.9.5.7. 1=Force auto-precharge cycles with every read or write command. 22:21 IdleCycLowLimit: idle cycle low limit. Read-write. Reset: 0. Specifies the number of MEMCLK cycles a page is allowed to be open before it may be closed by the dynamic page close logic. This field is ignored if D18F2x90_dct[1:0][DynPageCloseEn] = 0. Bits Description 00b 16 clocks 01b 32 clocks 10b 64 clocks 11b 96 clocks 344 42300 Rev 3.12 - July 14, 2015 BKDG for AMD Family 15h Models 10h-1Fh Processors 20 DynPageCloseEn: dynamic page close enable. Read-write. Reset: 0. See 2.9.5.7. 1=The DRAM controller dynamically determines when to close open pages based on the history of that particular page and D18F2x90_dct[1:0][IdleCycLowLimit]. 0=Any open pages not auto-precharged by the DRAM controller are automatically closed after 128 clocks of inactivity. 19 Reserved. 18 PendRefPayback: pending refresh payback. Read-write. Reset: 0. BIOS: 0. 1=The DRAM controller executes all pending refresh commands before entering the self refresh state. 0=The controller enters the self refresh state regardless of the number of pending refreshes; applies to any self refresh entry if PendRefPaybackS3En=0, else any non-S3 self refresh entry. 17 EnterSelfRef: enter self refresh command. Read, write-1-only; cleared-by-hardware. Reset: 0. 1=The DRAM controller places the DRAMs into self refresh mode. The DRAM interface is tristated 1 MEMCLK after the self refresh command is issued to the DRAMs. Once entered, the DRAM interface must remain in self refresh mode for a minimum of 5 MEMCLKs. This bit is read as a 1 while the enter-self-refresh command is executing; it is read as 0 at all other times. See 2.9.5.9 [DRAM Training] for information on how to use this bit. 16 UnbuffDimm: unbuffered DIMM. Read-write or read-only, depending on the product. Reset: Product-specific. 1=The DRAM controller is connected to unbuffered DIMMs. 0=The DRAM controller is connected to registered DIMMs or LR-DIMMs. 15:12 Reserved. 11:9 Reserved. 8 Reserved. 7:2 Reserved. 1 ExitSelfRef: exit self refresh (after suspend to RAM or for DRAM training) command. Read, write-1-only; cleared-by-hardware. Reset: 0. Writing a 1 to this bit causes the DRAM controller to bring the DRAMs out of self refresh mode. It also causes the DRAM controller to issue ZQCL and MRS MR0 commands. This command should be executed by BIOS when returning from the suspend to RAM state, after the DRAM controller configuration registers are properly initialized, or when self refresh is used during DRAM training. See 2.9.5.9 [DRAM Training]. This bit is read as a 1 while the exit-self-refresh command is executing; it is read as 0 at all other times. This bit should not be set if the DCT is disabled. After using this bit during a return from suspend to RAM , BIOS must issue an additional MRS command to set MR0[PPD]=1 if Fast exit precharge powerdown mode is desired. See D18F2x94_dct[1:0][1:0][PowerDownEn] and D18F2x84_dct[1:0][PchgPDModeSel]. 0 Reserved. D18F2x94_dct[1:0] DRAM Configuration High See 2.9.1 [DCT Configuration Registers]. Table 132: Memory Clock Frequency Value Definition Bits 03h-00h 04h 05h 06h 09h-07h 0Ah Description Reserved 333 MHz. (667 MT/s) Reserved 400 MHz. (800 MT/s) Reserved 533 MHz. (1066 MT/s) 345 42300 Rev 3.12 - July 14, 2015 BKDG for AMD Family 15h Models 10h-1Fh Processors Table 132: Memory Clock Frequency Value Definition Bits 0Dh-0Bh 0Eh 11h-0Fh 12h 15h-13h 16h 18h-17h 19h 1Ah 1Eh-1Bh 1Fh Bits 31 Description Reserved 667 MHz. (1333 MT/s) Reserved 800 MHz. (1600 MT/s) Reserved 933 MHz. (1866 MT/s) Reserved 1050 MHz. (2100 MT/s) 1066 MHz. (2133 MT/s) Reserved 1200 MHz. (2400 MT/s) Description DphyMemPsSelEn: Ddr phy MemPsSel enable. Read-write. Reset: 0h. BIOS: 1. 1=The DCT uses D18F1x10C[MemPsSel] to select the memory P-state context of a phy register when accessed by software with D18F2x98_dct[1:0][DctOffset[29:20]]==0D0h. DctOffset[6] is ignored. 0=Software accesses use DctOffset[6] to select context of those registers. If D18F2x98_dct[1:0][DctOffset[29:20]]!=0D0h then this register has no effect. See D18F2x9C_x0D04_E008_dct[1:0][PstateToAccess]. 30:29 Reserved. 28:24 DcqBypassMax: DRAM controller queue bypass maximum. Read-write. Reset: 0h. BIOS: 2.9.5.7. The DRAM controller arbiter normally allows transactions to pass other transactions in order to optimize DRAM bandwidth. This field specifies the maximum number of times that the oldest memory-access request in the DRAM controller queue may be bypassed before the arbiter decision is overridden and the oldest memory-access request is serviced instead. Description Bits 0h No bypass; the oldest request is never bypassed. 1Fh-1h The oldest request may be bypassed no more than <DcqBypassMax> time. 23 ProcOdtDis: processor on-die termination disable. Read-write. Reset: 0h. 1=The processor-side on-die termination is disabled. 0=Processor-side on-die termination enabled. See D18F2x9C_x0000_0000_dct[1:0]_mp[1:0][ProcOdt] for ODT definitions. Changes to this bit must be performed prior to DRAM initialization. 346 42300 Rev 3.12 - July 14, 2015 BKDG for AMD Family 15h Models 10h-1Fh Processors 22 BankSwizzleMode: bank swizzle mode. Read-write. Reset: 0. See 2.9.5.7. 1=Remaps the DRAM device bank address bits as a function of normalized physical address bits. Each of the bank address bits, as specified in Table 131 of D18F2x80_dct[1:0], are remapped as follows: • Define X as a bank address bit (e.g., X=15 if the bank bit is specified to be address bit 15). • Define S(n) as the state of address bit n (0 or 1) and B as the remapped bank address bit. Then, B= S(X) ^ S(X + 3) ^ S(X + 6); for an 8-bank DRAM and D18F2x[6C:60]_dct[1:0][RankDef] = 0xb. B= S(X) ^ S(X + 4) ^ S(X + 7); for an 8-bank DRAM and D18F2x[6C:60]_dct[1:0][RankDef] = 10b. B= S(X) ^ S(X + 5) ^ S(X + 8); for an 8-bank DRAM and D18F2x[6C:60]_dct[1:0][RankDef] = 11b. For example, encoding 02h of Table 131 would be remapped from Bank[2:0]={A15, A14, A13} to the following: Bank[2:0] = {A15^A18^A21, A14^A17^A20, A13^A16^A19}. 21 FreqChgInProg: frequency change in progress. Read-only. Reset: 0. 1=A MEMCLK frequency change is in progress. The DDR phy asserts this bit when it is in the process of locking the PLL. BIOS should not program the phy registers while this bit is set. 0=DRAM-interface commands can be sent to the phy. 20 SlowAccessMode: slow access mode (a.k.a. 2T mode). Read-write. Reset: 0. 1=One additional MEMCLK of setup time is provided on all DRAM address and control signals (not including CS, CKE, and ODT); i.e., these signals are driven for two MEMCLK cycles rather than one. 0=DRAM address and control signals are driven for one MEMCLK cycle. 2T mode may be needed in order to meet electrical requirements of certain DIMM speed and loading configurations. See 2.9.5.6.6 [DRAM Address Timing and Output Driver Compensation Control]. 19 Reserved. 18:17 Reserved. 16 PowerDownMode: power down mode. Read-write. Reset: 0. BIOS: 1. Specifies how a chip select or group of chip selects enters power down mode when enabled by D18F2x94_dct[1:0][PowerDownEn]. A chip select enters power down mode when the DCT deasserts the CKE pin. The command and address signals tristate one MEMCLK after CKE deasserts. The DCT behavior varies based on the setting of D18F2x84_dct[1:0][PchgPDModeSel]. See also Table 130 [DIMM, Chip Select, and Register Mapping]. 0=Channel CKE control mode; the DRAM channel is placed in power down mode when all chip selects associated with the channel are idle; both CKE pins for the channel operate in lock step in terms of placing the channel in power down mode. 1=Chip select CKE control mode; the chip select group controlled by a CKE pinis placed in power down mode when no transactions are pending. 15 PowerDownEn: power down mode enable. Read-write. Reset: 0. BIOS: 2.9.5.7. 1=Power down mode is enabled. Only precharge power down mode is supported, not active power down mode. See PowerDownMode, D18F2x84_dct[1:0][PchgPDModeSel], D18F2xA8_dct[1:0][PrtlChPDEnhEn, AggrPDEn, PDPhyPSDis], and D18F2x248_dct[1:0]_mp[1:0][PchgPDEnDelay]. 14 DisDramInterface: disable the DRAM interface. Read-write. Reset: 0. 1=The DRAM controller is disabled and the DRAM interface is placed into a low power state. This bit must be set if there are no DIMMs connected to the DCT. 13 Reserved. 12 Reserved. 347 42300 Rev 3.12 - July 14, 2015 BKDG for AMD Family 15h Models 10h-1Fh Processors 11:10 ZqcsInterval: ZQ calibration short interval. Read-write. Reset: 00b. BIOS: See 2.9.5.7. This field specifies the programmable interval for the controller to send out the DRAM ZQ calibration short command. Description Bits 00b ZQ calibration short command is disabled 01b 64 ms 10b 128 ms 11b 256 ms 9:8 7 Reserved. MemClkFreqVal: memory clock frequency valid. Read-write. Reset: 0. System BIOS should set this bit after setting up D18F2x94_dct[1:0][MemClkFreq] to the proper value. This indicates to the DRAM controller that it may start driving MEMCLK at the proper frequency. After programming MemClkFreqVal=1, BIOS should poll FreqChgInProg to determine when the DRAM-interface clocks are stable. This bit should not be set if the DCT is disabled. BIOS must change each DCT’s operating frequency in order. See 2.9.5.8 [DRAM Device and Controller Initialization]. 6:5 Reserved. 4:0 MemClkFreq: memory clock frequency. Read-write. Reset: 000b. Specifies the frequency and rate of the DRAM interface (MEMCLK). See: Table 132 [Memory Clock Frequency Value Definition]. The rate is twice the frequency. See D18F5x84[DdrMaxRate] and D18F5x84[DdrMaxRateEnf]. D18F2x98_dct[1:0] DRAM Controller Additional Data Offset Reset: 8000_0000h. See 2.9.1 [DCT Configuration Registers]. Each DCT includes an array of registers that are used primarily to control DRAM-interface electrical parameters. Access to these registers is accomplished as follows: Reads: 1. Write the register number to D18F2x98_dct[1:0][DctOffset] with D18F2x98_dct[1:0][DctAccessWrite]=0. 2. Read the register contents from D18F2x9C_dct[1:0]. Writes: 1. Write all 32 bits of register data to D18F2x9C_dct[1:0] (individual byte writes are not supported). 2. Write the register number to D18F2x98_dct[1:0][DctOffset] with D18F2x98_dct[1:0][DctAccessWrite]=1. • The data will be delivered to the phy similar to a posted memory-write, and the write will complete without any further action. However, to ensure that the contents of the array register write have been delivered to the phy, software issues a subsequent configuration register read or write to any register in the northbridge. For example, reading D18F2x98_dct[1:0] will accomplish this. Registers for which there is one instance per memory P-state (listed with “_mp[1:0]” appended to the register mnemonic) use D18F1x10C[MemPsSel], D18F2x94_dct[1:0][DphyMemPsSelEn], and D18F2x9C_x0D04_E008_dct[1:0][PStateToAccess] for software accesses. BIOS programs these fields appropriately to ensure consistency of registers with controller fields and DDR phy fields. • D18F2x9C_x0000_0[3:0]0[2:1]_dct[1:0]_mp[1:0] refers to all instances of the D18F2x9C_x0000_0[3:0]0[3:1] register. • D18F2x9C_x0000_0[3:0]0[3:1]_dct[1:0]_mp[1] refers to the register for memory P-state 1 of either or both DCTs. 348 42300 Rev 3.12 - July 14, 2015 BKDG for AMD Family 15h Models 10h-1Fh Processors • It is recommended that BIOS program context sensitive registers in batches (training/restoring all registers of a context before selecting a new context). BIOS should do the following prior to a new “batch”: • Program D18F2x94_dct[1:0][DphyMemPsSelEn]=1. • Program D18F2x9C_x0D04_E008_dct[1:0][PStateToAccess]=<target>. • Program D18F1x10C[MemPsSel]=<target>. Writes to any register in this additional address space causes the FIFO pointers to be reset. Therefore, it is recommended that only BIOS write these registers. System software other than BIOS may reliably read certain memory P-state specific DDR Phy registers with DctOffset[29:20]!=0D0h in-context normally, or out-of-context using alternative direct access methods (without the requirement for programming D18F2x9C_x0D04_E008_dct[1:0][PStateToAccess]). BIOS may also use this method after memory traffic is enabled, for reading and saving register state in preparation for S3 support. The following table shows the fields that are supported and the equivalent phy direct access: Table 133: Memory P-state Specific Indirect CSR Access by non-BIOS Software Indirect Register D18F2x9C_x0000_0000_dct[1:0]_mp[1:0] [ProcOdt] D18F2x9C_x0000_0000_dct[1:0]_mp[1:0] [DqsDrvStren] D18F2x9C_x0000_0000_dct[1:0]_mp[1:0] [DataDrvStren] D18F2x9C_x0000_0000_dct[1:0]_mp[1:0] [ClkDrvStren] D18F2x9C_x0000_0000_dct[1:0]_mp[1:0] [AddrCmdDrvStren] D18F2x9C_x0000_0000_dct[1:0]_mp[1:0] [CsOdtDrvStren] D18F2x9C_x0000_0004_dct[1:0]_mp[1:0] [AddrCmdFineDelay] D18F2x9C_x0000_0004_dct[1:0]_mp[1:0] [CsOdtFineDelay] D18F2x9C_x0000_0004_dct[1:0]_mp[1:0] [CkeFineDelay] D18F2x9C_x0000_000D_dct[1:0]_mp[1:0] [RxDLLWakeupTime] D18F2x9C_x0000_000D_dct[1:0]_mp[1:0] [RxCPUpdPeriod] D18F2x9C_x0000_000D_dct[1:0]_mp[1:0] [RxMaxDurDllNoLock] Equivalent Direct Register D18F2x9C_x0D0F_0[F,7:0]0[8,0]_dct[1:0]_mp[1 :0] [ProcOdt] D18F2x9C_x0D0F_0[F,7:0]04_dct[1:0]_mp[1:0] [DqsDrvStren] D18F2x9C_x0D0F_0[F,7:0]0[8,0]_dct[1:0]_mp[1 :0] [DataDrvStren] D18F2x9C_x0D0F_0[F,7:0]0[8,0]_dct[1:0]_mp[1 :0] [ClkDrvStren] D18F2x9C_x0D0F_8[1:0]0[8,4,0]_dct[1:0]_mp[1 :0] [DrvStrength] D18F2x9C_x0D0F_8[1:0]0[8,4,0]_dct[1:0]_mp[1 :0] [DrvStrength] D18F2x9C_x0D0F_8[1:0]20_dct[1:0]_mp[1:0] [Delay] D18F2x9C_x0D0F_8[1:0]20_dct[1:0]_mp[1:0] [Delay] D18F2x9C_x0D0F_C021_dct[1:0]_mp[1:0] [Delay] D18F2x9C_x0D0F_0[F,7:0]11_dct[1:0]_mp[1:0] [RxDLLWakeupTime] D18F2x9C_x0D0F_0[F,7:0]11_dct[1:0]_mp[1:0] [RxCPUpdPeriod] D18F2x9C_x0D0F_0[F,7:0]11_dct[1:0]_mp[1:0] [RxMaxDurDllNoLock] 349 42300 Rev 3.12 - July 14, 2015 BKDG for AMD Family 15h Models 10h-1Fh Processors Table 133: Memory P-state Specific Indirect CSR Access by non-BIOS Software Indirect Register D18F2x9C_x0000_000D_dct[1:0]_mp[1:0] [TxDLLWakeupTime] D18F2x9C_x0000_000D_dct[1:0]_mp[1:0] [TxCPUpdPeriod] D18F2x9C_x0000_000D_dct[1:0]_mp[1:0] [TxMaxDurDllNoLock] Equivalent Direct Register D18F2x9C_x0D0F_0[F,7:0]10_dct[1:0]_mp[1:0] [TxDLLWakeupTime] D18F2x9C_x0D0F_0[F,7:0]10_dct[1:0]_mp[1:0] [TxCPUpdPeriod] D18F2x9C_x0D0F_0[F,7:0]10_dct[1:0]_mp[1:0] [TxMaxDurDllNoLock] System software other than BIOS may read any memory P-state specific DCT register with DctOffset[29:20]!=0D0h in-context normally, or out-of-context by programming D18F1x10C[MemPsSel]. The following table shows the fields that are supported. For indirect registers with both DCT and DDR Phy fields, software should mask off the DDR Phy fields as these may not be correct without programming D18F2x9C_x0D04_E008_dct[1:0][PStateToAccess]. Table 134: Memory P-state Specific DCT CSR In Indirect Space Indirect Register 03h-00h D18F2x9C_x0000_0[3:0]0[2:1]_dct[1:0]_mp[1:0][WrDatGrossDly] (for all bytelanes) D18F2x9C_x0000_0004_dct[1:0]_mp[1:0][AddrCmdSetup, CsOdtSetup, CkeSetup] D18F2x9C_x0000_00[2A:10]_dct[1:0]_mp[1:0][DqsRcvEnGrossDelay] (for all bytelanes) D18F2x9C_x0000_00[4A:30]_dct[1:0]_mp[1:0][WrDqsGrossDly] (for all bytelanes) Reads or writes to any register in this additional address space collide with system self-refresh requests. Once powermanagement is enabled software should temporarily disable powermanagement prior to accessing these registers. Bits Description 31 Reserved. 30 DctAccessWrite: DRAM controller read/write select. RAZ; write. 0=Specifies a read access. 1=Specifies a write access. 29:0 DctOffset: DRAM controller offset. Read-write. D18F2x9C_dct[1:0] DRAM Controller Additional Data Port See D18F2x98_dct[1:0] for register access information. See 2.9.1 [DCT Configuration Registers]. D18F2x9C_x0000_0000_dct[1:0]_mp[1:0] DRAM Output Driver Compensation Control See 2.9.5.6.6 [DRAM Address Timing and Output Driver Compensation Control]. Bits 31 Description Reserved. 30:28 ProcOdt. Read-write. DdrPhy: See D18F2x9C_x0D0F_0[F,7:0]0[8,0]_dct[1:0]_mp[1:0][ProcOdt]. Writes to this field write all byte-lanes and groups 0 and 2. Reads to this field read byte-lane 0 group 0 ProcOdt. 350 42300 Rev 3.12 - July 14, 2015 BKDG for AMD Family 15h Models 10h-1Fh Processors 27:23 Reserved. 22:20 DqsDrvStren. Read-write. DdrPhy: See D18F2x9C_x0D0F_0[F,7:0]04_dct[1:0]_mp[1:0] [DqsDrvStren]. Writes to this field write all byte-lanes. Reads to this field read byte-lane 0 group 1 DqsDrvStren. 19 Reserved. 18:16 DataDrvStren. Read-write. DdrPhy: See D18F2x9C_x0D0F_0[F,7:0]0[8,0]_dct[1:0]_mp[1:0][DataDrvStren]. Writes to this field write all byte-lanes and groups 0 and 2. Reads to this field read byte-lane 0 group 0 DataDrvStren. 15 Reserved. 14:12 ClkDrvStren. Read-write. DdrPhy: See D18F2x9C_x0D0F_2[F,2:0]00_dct[1:0]_mp[1:0][ClkDrvStren]. Writes to this field write all clock chips. Reads to this field read clock chip 0 ClkDrvStren. 11 Reserved. 10:8 AddrCmdDrvStren. Read-write. DdrPhy: See D18F2x9C_x0D0F_8[1:0]0[8,4,0]_dct[1:0]_mp[1:0][DrvStrength] and D18F2x9C_x0D0F_C0[10,0C,08,04]_dct[1:0]_mp[1:0][DrvStrength]. Writes to this field write all groups of Addr/Cmd 1 chip and all groups of Address2 chip. Reads to this field read DrvStrength of group 0 of Addr/Cmd 1 chip. 7 6:4 3 2:0 Reserved. CsOdtDrvStren. Read-write. DdrPhy: See D18F2x9C_x0D0F_8[1:0]0[8,4,0]_dct[1:0]_mp[1:0][DrvStrength]. Writes to this field write all groups of Addr/Cmd 0 chip. Reads to this field read DrvStrength of group 0 of Addr/Cmd 0 chip. Reserved. CkeDrvStren. Read-write. DdrPhy: See D18F2x9C_x0D0F_C000_dct[1:0]_mp[1:0][CkeDrvStren]. D18F2x9C_x0000_0[3:0]0[2:1]_dct[1:0]_mp[1:0] DRAM Write Data Timing BIOS: See 2.9.5.9.4 [DQS Position Training]. Table 135: Index Mapping for D18F2x9C_x0000_0[3:0]0[2:1]_dct[1:0]_mp[1:0] D18F2x98_dct[1:0][31:0] Function1 0000_0001h DIMM/CS 0 Bytes 3-0 0000_0002h DIMM/CS 0 Bytes 7-4 0000_0101h DIMM/CS 1 Bytes 3-0 0000_0102h DIMM/CS 1 Bytes 7-4 0000_0201h CS 2 Bytes 3-0 0000_0202h CS 2 Bytes 7-4 0000_0301h CS 3 Bytes 3-0 0000_0302h CS 3 Bytes 7-4 1. If D18F2xA8_dct[1:0][PerRankTimingEn]=1 then the function is CS. Otherwise the function is DIMM. 351 42300 Rev 3.12 - July 14, 2015 BKDG for AMD Family 15h Models 10h-1Fh Processors Table 136: Byte Lane Mapping for D18F2x9C_x0000_0[3:0]0[2:1]_dct[1:0]_mp[1:0] Register Bits 31:24 23:16 15:8 7:0 D18F2x9C_x0000_0[3:0]01 Byte3 Byte2 Byte1 Byte0 D18F2x9C_x0000_0[3:0]02 Byte7 Byte6 Byte5 Byte4 These registers control the timing of write DQ with respect to MEMCLK and allow transmit DQS to be centered in the data eye. The delay starts 1 UI before the rising edge of MEMCLK corresponding to the CASwrite-latency. See 2.9.5.9 [DRAM Training]. WrDatGrossDly must be programmed for a given DIMM and lane such that WrDatDly - WrDqsDly <= 0.5 MEMCLKs. Bits Description 31:29 WrDatGrossDly3: write data gross delay. See: D18F2x9C_x0000_0[3:0]0[2:1]_dct[1:0]_mp[1:0][WrDatGrossDly]. 28:24 WrDatFineDly3: write data fine delay. See: D18F2x9C_x0000_0[3:0]0[2:1]_dct[1:0]_mp[1:0][WrDatFineDly]. 23:21 WrDatGrossDly2: write data gross delay. See: D18F2x9C_x0000_0[3:0]0[2:1]_dct[1:0]_mp[1:0][WrDatGrossDly]. 20:16 WrDatFineDly2: write data fine delay. See: D18F2x9C_x0000_0[3:0]0[2:1]_dct[1:0]_mp[1:0][WrDatFineDly]. 15:13 WrDatGrossDly1: write data gross delay. See: D18F2x9C_x0000_0[3:0]0[2:1]_dct[1:0]_mp[1:0][WrDatGrossDly]. 12:8 WrDatFineDly1: write data fine delay. See: D18F2x9C_x0000_0[3:0]0[2:1]_dct[1:0]_mp[1:0][WrDatFineDly]. 7:5 WrDatGrossDly: write data gross delay. Read-write. Reset: 0. Bits Description 000b No delay 001b 0.5 MEMCLK delay 110b-010b <WrDatGrossDly>/2 MEMCLK delay 111b 3.5 MEMCLK delay 4:0 WrDatFineDly: write data fine delay. Read-write. Cold reset: 0. Description Bits 00h 0/64 MEMCLK delay 1Eh-01h <WrDatFineDly>/64 MEMCLK delay 1Fh 31/64 MEMCLK delay D18F2x9C_x0000_0004_dct[1:0]_mp[1:0] DRAM Address/Command Timing Control BIOS: 2.9.5.6.6. This register controls the timing of the address, command, chip select, ODT and clock enable pins with respect to MEMCLK as shown in Figure 11. See 2.9.5.4.2 [DRAM Channel Frequency Change] and 2.9.5.4.3 [Phy Fence Programming]. 352 42300 Rev 3.12 - July 14, 2015 BKDG for AMD Family 15h Models 10h-1Fh Processors MEMCLK Command Command 1 MEMCLK ½ MEMCLK Coarse Setup = 1 (setup is one full MEMCLK) – Fine delay is zero. Coarse Setup = 0 (setup is one half MEMCLK) – Fine delay is zero. Command Coarse Setup = 1 (setup is one full MEMCLK) plus a non zero fine delay. Command Coarse Setup = 0 (setup is one half MEMCLK) – A half MEMCLK setup plus a non-zero fine delay. Figure 11: Address/Command Timing at the Processor Pins 2T timing is controlled by D18F2x94_dct[1:0][SlowAccessMode]. If a setup time (course delay) field is changed and D18F2x94_dct[1:0][MemClkFreqVal]=1, then software must toggle MemClkFreqVal for the delay to take effect. Bits Description 31:22 Reserved. 21 AddrCmdSetup: address/command setup time. Read-write. Reset: 0. DCT: Selects the default setup time for the address and command pins versus MEMCLK. 0=1/2 MEMCLK (1 1/2 MEMCLK for 2T timing). 1=1 MEMCLK (2 MEMCLKs for 2T timing). 20:16 AddrCmdFineDelay. Read-write. Cold Reset: 0. DdrPhy: See D18F2x9C_x0D0F_8[1:0]20_dct[1:0]_mp[1:0][Delay] and D18F2x9C_x0D0F_C020_dct[1:0]_mp[1:0][Delay]. Writes to this field write D18F2x9C_x0D0F_8[1:0]20_dct[1:0]_mp[1:0][Delay] of the Cmd/Addr 1 chip and to D18F2x9C_x0D0F_C020_dct[1:0]_mp[1:0][Delay]. Reads to this field read D18F2x9C_x0D0F_8[1:0]20_dct[1:0]_mp[1:0][Delay] from Cmd/Addr 1 chip. 15:14 Reserved. 13 CsOdtSetup: CS/ODT setup time. Read-write. Reset: 0. DCT: Selects the default setup time for the CS and ODT pins versus MEMCLK. 0=1/2 MEMCLK. 1=1 MEMCLK. 12:8 CsOdtFineDelay. Read-write. Cold Reset: 0. DdrPhy: See D18F2x9C_x0D0F_8[1:0]20_dct[1:0]_mp[1:0][Delay]. Writes to this field write Delay of the Cmd/Addr 0 chip. Reads to this field read Delay from Cmd/Addr 0 chip. 7:6 5 4:0 Reserved. CkeSetup: CKE setup time. Read-write. Reset: 0. DCT: Selects the default setup time for the CKE pins versus MEMCLK. 0=1/2 MEMCLK. 1=1 MEMCLK. CkeFineDelay: CKE fine delay. Read-write. Cold reset: 00h. DdrPhy: See D18F2x9C_x0D0F_C021_dct[1:0]_mp[1:0][Delay]. Writes to this field write to D18F2x9C_x0D0F_C021_dct[1:0]_mp[1:0][Delay]. Reads to this field read from D18F2x9C_x0D0F_C021_dct[1:0]_mp[1:0][Delay]. 353 42300 Rev 3.12 - July 14, 2015 BKDG for AMD Family 15h Models 10h-1Fh Processors D18F2x9C_x0000_0[3:0]0[6:5]_dct[1:0]_mp[1:0] DRAM Read DQS Timing Table 137: Index Mapping for D18F2x9C_x0000_0[3:0]0[6:5]_dct[1:0]_mp[1:0] D18F2x98_dct[1:0][31:0] Function1 0000_0005h DIMM/CS 0 Bytes 3-0 0000_0006h DIMM/CS 0 Bytes 7-4 0000_0105h DIMM/CS 1 Bytes 3-0 0000_0106h DIMM/CS 1 Bytes 7-4 0000_0205h Reserved /CS 2 Bytes 3-0 0000_0206h Reserved /CS 2 Bytes 7-4 0000_0305h Reserved /CS 3 Bytes 3-0 0000_0306h Reserved /CS 3 Bytes 7-4 1. If D18F2xA8_dct[1:0][PerRankTimingEn]=1 then the function is CS. Otherwise the function is DIMM. Table 138: Byte Lane Mapping for D18F2x9C_x0000_0[3:0]0[6:5]_dct[1:0]_mp[1:0] Register Bits 29:25 21:17 13:9 5:1 D18F2x9C_x0000_0[3:0]05 Byte3 Byte2 Byte1 Byte0 D18F2x9C_x0000_0[3:0]06 Byte7 Byte6 Byte5 Byte4 These registers control the timing of read (input) DQS signals with respect to DQ. See 2.9.5.9 [DRAM Training]. The actual delay applied to the DQS input signal before sampling data includes an internal part dependent delay plus the nominal register delay specified here. The part dependent (insertion) delay is large in proportion to individual step delay controlled by this register. If the actual delay exceeds 1UI and read bubbles exist in the read data stream then incorrect operation of received DQS may occur. BIOS training allows this scenario temporarily by ensuring no bubbles in read streams exist. The actual delay of the optimally trained RdDqsTime value does not exceed 1UI. Bits Description 31:30 Reserved. 29:25 RdDqsTime3: read DQS timing control. See: D18F2x9C_x0000_0[3:0]0[6:5]_dct[1:0]_mp[1:0][RdDqsTime]. 24:22 Reserved. 21:17 RdDqsTime2: read DQS timing control. See: D18F2x9C_x0000_0[3:0]0[6:5]_dct[1:0]_mp[1:0][RdDqsTime]. 16:14 Reserved. 13:9 RdDqsTime1: read DQS timing control. See: D18F2x9C_x0000_0[3:0]0[6:5]_dct[1:0]_mp[1:0][RdDqsTime]. 8:6 Reserved. 354 42300 Rev 3.12 - July 14, 2015 5:1 0 BKDG for AMD Family 15h Models 10h-1Fh Processors RdDqsTime: read DQS timing control. Read-write. Cold reset: 0Fh. Bits Description 00h 0/64 MEMCLK delay 1Eh-01h <RdDqsTime>/64 MEMCLK delay 1Fh 31/64 MEMCLK delay Reserved. D18F2x9C_x0000_0008_dct[1:0]_mp[1:0] DRAM Phy Control Cold reset: 0208_0000h. See 2.9.5.9 [DRAM Training]. Bits Description 31 Reserved. 30 DisAutoComp: disable automatic compensation. Read-write. 1=Disable the compensation control state machine. 0=The phy automatic compensation engine is enabled. 29 DisablePredriverCal: disable predriver calibration. Read-write. BIOS: See 2.9.5.4.4. 1=Disable the update of predriver codes to all pads. Setting this bit in DCT0 disables calibration engine for DCT0 and DCT1. Setting this bit in DCT1 has no effect. 28:14 Reserved. 13 DqsRcvTrEn: DQS receiver training enable. Read-write. 1=Initiate hardware assisted read DQS receiver training.0=Stop read DQS receiver training. This allows BIOS to reliably read the DQS receiver training data. 12 WrLvOdtEn: write levelization ODT enabled. Read-write. 1=ODT specified by WrLvOdt is enabled during write levelization training. 0=ODT is disabled during write levelization training. 11:8 WrLvOdt: write levelization ODT. Read-write. Specifies the state of the ODT pins when WrLvOdtEn is set. 1=ODT is enabled. 0=ODT is disabled. See 2.9.5.6.5 [DRAM ODT Control]. Tri-state enable for ODT is turned off by the phy when WrLvOdtEn=1. 7:6 FenceTrSel: fence train select. Read-write. Specifies the flop to be used for phy based fence training. See PhyFenceTrEn. This field is shared by D18F2x9C_x0000_0008_dct[1:0]_mp0 and D18F2x9C_x0000_0008_dct[1:0]_mp1. Bits Description 00b PRE flop 01b RxDll flop 10b TxDll flop 11b TxPad flop 355 42300 Rev 3.12 - July 14, 2015 BKDG for AMD Family 15h Models 10h-1Fh Processors 5:4 TrChipSel: training DIMM select. Read-write. Specifies a timing control for a corresponding rank which (see D18F2xA8_dct[1:0][PerRankTimingEn]) is to be trained. IF (D18F2xA8_dct[1:0][PerRankTimingEn]==0) THEN Description Bits 00b chip select 0 and 1 01b chip select 2 and 3 10b Reserved 11b Reserved ELSE Description Bits 00b chip select 0 01b chip select 1 10b chip select 2 11b chip select 3 ENDIF. 3 PhyFenceTrEn: phy fence training enable. Read-write. 1=Initiate phy based fence training. 0=Stop the phy based fence training engine. 2 TrNibbleSel: training nibble select. Read-write. BIOS: 0. Specifies nibbles of each DIMM data byte trained during write levelization training. 0=Lower nibbles. 1=Upper nibbles. 1 Reserved. 0 WrtLvTrEn: write levelization training enable. Read-write. 1=Initiate write levelization (tDQSS margining) training. 0=The phy stops driving DQS and exits write levelization training. D18F2x9C_x0000_000B_dct[1:0] DRAM Phy Status Register Bits Description 31 DynModeChange: dynamic mode change. RAZ; write. Reset: 0. 1=Phy enters the state specified by PhySelfRefreshMode. 30 PhyPSReq: phy pstate request. RAZ; write. Reset: 0. 1=Phy enters the memory P-state specified by PhyPS. 29:27 Reserved. 26 PhyPS: phy pstate. RAZ; write. Reset: 0. 1=M1. 0=M0. See PhyPSReq. 25:24 Reserved 23 PhySelfRefreshMode: phy self refresh mode. RAZ; write. Reset: 0. 1=Enter self refresh mode. 0=Exit self refresh mode. See DynModeChange. 22:0 Reserved D18F2x9C_x0000_000C_dct[1:0] DRAM Phy Miscellaneous This register provides access to the DDR phy to control signal tri-state functionality. See Table 130 for processor pin map. Based on the system configuration, BIOS may tri-state signals with associated chip selects that are unpopulated in an effort to conserve power. The recommendations for tri-state of ODT pins are as follows: • BIOS tri-states ODT pins to unpopulated DIMM slots. • BIOS tri-states ODT pins that are unused on slots with unbuffered DIMMs. • BIOS does not tri-state any ODT pins to slots with registered DIMMs . • For single rank registered DIMMs with address parity capability, BIOS must not tri-state the chip select pin 356 42300 Rev 3.12 - July 14, 2015 BKDG for AMD Family 15h Models 10h-1Fh Processors corresponding to the second chip select of the DIMM. Bits 31 Description Reserved. 30:26 FenceThresholdTxDll: phy fence threshold transmit DLL. Read-write. Cold reset: 13h. BIOS: See 2.9.5.4.3. This field specifies the fence delay threshold value used for DQS receiver valid. See FenceThresholdTxPad. 25:21 FenceThresholdRxDll: phy fence threshold DQS receiver enable. Read-write. Cold reset: 13h. BIOS: See 2.9.5.4.3. This field specifies the fence delay threshold value used for DQS receiver enable. See FenceThresholdTxPad. 20:16 FenceThresholdTxPad: phy fence threshold transmit pad. Read-write. Cold reset: 13h. BIOS: See 2.9.5.4.3. This field specifies the fence delay threshold value used for write data, write DQS, Addr/Cmd, CS, ODT, and CKE. Bits Description 00h 0/64 MEMCLK delay 1Eh-01h <FenceThresholdTxPad>/64 MEMCLK delay 1Fh 31/64 MEMCLK delay 15:12 CKETri: CKE tri-state. Read-write. Cold reset: 0h. 0=The CKE signal is not tri-stated. 1=Tri-state unconnected CKE signal from the processor. See 2.9.2 [DDR Pad to Processor Pin Mapping]. Bit Pad [0] MEMCKE[1:0][0] [1] MEMCKE[1:0][1] [2] MEMCKE[1:0][2] [3] MEMCKE[1:0][3] 11:8 ODTTri: ODT tri-state. Read-write. Cold reset: 0h. 0=The ODT signals are not tri-stated unless directed to by the DCT. 1=Tri-state unconnected ODT signals from the processor. See 2.9.2 [DDR Pad to Processor Pin Mapping]. Bit Pad [0] MEMODT[1:0][0] [1] MEMODT[1:0][1] [2] MEMODT[1:0][2] [3] MEMODT[1:0][3] 7:0 ChipSelTri: chip select tri-state. Read-write. Cold reset: 00h. 0=The chip select signals are not tristated unless directed to by the DCT. 1=Tri-state unpopulated chip selects when motherboard termination is available. See 2.9.2 [DDR Pad to Processor Pin Mapping]. Bit Pad [0] MEMCS[1:0]_L[0] [1] MEMCS[1:0]_L[1] [2] MEMCS[1:0]_L[2] [3] MEMCS[1:0]_L[3] [4] MEMCS[1:0]_L[4] [5] MEMCS[1:0]_L[5] [6] MEMCS[1:0]_L[6] [7] MEMCS[1:0]_L[7] D18F2x9C_x0000_000D_dct[1:0]_mp[1:0] DRAM Phy DLL Control Cold Reset: 0000_0000h. This register defines programmable options for the phy's DLLs for power savings. There are two identical sets of configuration registers: one for the transmit DLLs (those running off of the phy's 357 42300 Rev 3.12 - July 14, 2015 BKDG for AMD Family 15h Models 10h-1Fh Processors internal PCLK which is running at rate of 2*MEMCLK) and receive DLLs (those running off of the DQS from the DIMMs). These values are programmed by BIOS based on programmed DDR frequency. This register must be programmed before DRAM device initialization. Bits Description 31:26 Reserved. 25:24 RxDLLWakeupTime: receive DLL wakeup time. Read-write. DdrPhy: See D18F2x9C_x0D0F_0[F,7:0]11_dct[1:0]_mp[1:0][RxDLLWakeupTime]. 23 Reserved. 22:20 RxCPUpdPeriod: receive charge pump period. Read-write. DdrPhy: See D18F2x9C_x0D0F_0[F,7:0]11_dct[1:0]_mp[1:0][RxCPUpdPeriod]. 19:16 RxMaxDurDllNoLock: receive maximum duration DLL no lock. Read-write. DdrPhy: See D18F2x9C_x0D0F_0[F,7:0]11_dct[1:0]_mp[1:0][RxMaxDurDllNoLock]. 15:10 Reserved. 9:8 7 TxDLLWakeupTime: transmit DLL wakeup time. Read-write. DdrPhy: See D18F2x9C_x0D0F_0[F,7:0]10_dct[1:0]_mp[1:0][TxDLLWakeupTime]. Reserved. 6:4 TxCPUpdPeriod: transmit charge pump DLL wakeup time. Read-write. DdrPhy: See D18F2x9C_x0D0F_0[F,7:0]10_dct[1:0]_mp[1:0][TxCPUpdPeriod]. 3:0 TxMaxDurDllNoLock: transmit maximum duration DLL no lock. Read-write. DdrPhy: See D18F2x9C_x0D0F_0[F,7:0]10_dct[1:0]_mp[1:0][TxMaxDurDllNoLock]. D18F2x9C_x0000_00[2A:10]_dct[1:0]_mp[1:0] DRAM DQS Receiver Enable Timing Table 139: Index Mapping for D18F2x9C_x0000_00[2A:10]_dct[1:0]_mp[1:0] D18F2x98_dct[1:0][31:0] Function1 0000_0010h DIMM/CS 0 Bytes 1-0 0000_0011h DIMM/CS 0 Bytes 3-2 0000_0013h DIMM/CS 1 Bytes 1-0 0000_0014h DIMM/CS 1 Bytes 3-2 0000_0016h CS 2 Bytes 1-0 0000_0017h CS 2 Bytes 3-2 0000_0019h CS 3 Bytes 1-0 0000_001Ah CS 3 Bytes 3-2 0000_0020h DIMM/CS 0 Bytes 5-4 0000_0021h DIMM/CS 0 Bytes 7-6 0000_0023h DIMM/CS 1 Bytes 5-4 0000_0024h DIMM/CS 1 Bytes 7-6 0000_0026h CS 2 Bytes 5-4 0000_0027h CS 2 Bytes 7-6 0000_0029h CS 3 Bytes 5-4 0000_002Ah CS 3 Bytes 7-6 1. If D18F2xA8_dct[1:0][PerRankTimingEn]=1 then the function is CS. Otherwise the function is DIMM. 358 42300 Rev 3.12 - July 14, 2015 BKDG for AMD Family 15h Models 10h-1Fh Processors Table 140: Byte Lane Mapping for D18F2x9C_x0000_00[2A:10]_dct[1:0]_mp[1:0] Register Bits 25:16 9:0 D18F2x9C_x0000_001[9,6,3,0] Byte1 Byte0 D18F2x9C_x0000_001[A,7,4,1] Byte3 Byte2 D18F2x9C_x0000_002[9,6,3,0] Byte5 Byte4 D18F2x9C_x0000_002[A,7,4,1] Byte7 Byte6 Each of these registers control the timing of the receiver enable from the start of the read preamble with respect to MEMCLK. See 2.9.5.9 [DRAM Training]. These delay registers must be programmed such that across all DIMMs and lanes MAX(DqsRcvEnDelay) - MIN(DqsRcvEnDelay) <= 7 UI. Bits Description 31:26 Reserved. 25:21 DqsRcvEnGrossDelay1: DQS receiver enable gross delay. See: D18F2x9C_x0000_00[2A:10]_dct[1:0]_mp[1:0][DqsRcvEnGrossDelay]. 20:16 DqsRcvEnFineDelay1: DQS receiver enable fine delay. See: D18F2x9C_x0000_00[2A:10]_dct[1:0]_mp[1:0][DqsRcvEnFineDelay]. 15:10 Reserved. 9:5 DqsRcvEnGrossDelay: DQS receiver enable gross delay. Read-write. Reset: 01h. Bits Description 00h No delay 01h 0 MEMCLK delay 1Eh-02h <DqsRcvEnGrossDelay>/2 MEMCLK delay 1Fh 15.5 MEMCLK delay 4:0 DqsRcvEnFineDelay: DQS receiver enable fine delay. Read-write. Cold reset: 00h. Bits Description 00h 0/64 MEMCLK delay 1Eh-01h <DqsRcvEnFineDelay>/64 MEMCLK delay 1Fh 31/64 MEMCLK delay D18F2x9C_x0000_00[4A:30]_dct[1:0]_mp[1:0] DRAM DQS Write Timing BIOS: See 2.9.5.9 [DRAM Training]. Each of these registers control the DQS timing delay for write commands relative to MEMCLK. The delay starts at the rising edge of MEMCLK corresponding to the CASwritelatency. Each control includes a gross timing field and a fine timing field, the sum of which is the total delay. Table 141: Index Mapping for D18F2x9C_x0000_00[4A:30]_dct[1:0]_mp[1:0] D18F2x98_dct[1:0][31:0] 0000_0030h 0000_0031h 0000_0033h 0000_0034h 0000_0036h 0000_0037h Function1 DIMM/CS 0 Bytes 1-0 DIMM/CS 0 Bytes 3-2 DIMM/CS 1 Bytes 1-0 DIMM/CS 1 Bytes 3-2 CS 2 Bytes 1-0 CS 2 Bytes 3-2 359 42300 Rev 3.12 - July 14, 2015 BKDG for AMD Family 15h Models 10h-1Fh Processors Table 141: Index Mapping for D18F2x9C_x0000_00[4A:30]_dct[1:0]_mp[1:0] 0000_0039h CS 3 Bytes 1-0 0000_003Ah CS 3 Bytes 3-2 0000_0040h DIMM/CS 0 Bytes 5-4 0000_0041h DIMM/CS 0 Bytes 7-6 0000_0043h DIMM/CS 1 Bytes 5-4 0000_0044h DIMM/CS 1 Bytes 7-6 0000_0046h CS 2 Bytes 5-4 0000_0047h CS 2 Bytes 7-6 0000_0049h CS 3 Bytes 5-4 0000_004Ah CS 3 Bytes 7-6 1. If D18F2xA8_dct[1:0][PerRankTimingEn]=1 then the function is CS. Otherwise the function is DIMM. Table 142: Byte Lane Mapping for D18F2x9C_x0000_00[4A:30]_dct[1:0]_mp[1:0] Register Bits 23:16 7:0 D18F2x9C_x0000_003[9,6,3,0] Byte1 Byte0 D18F2x9C_x0000_003[A,7,4,1] Byte3 Byte2 D18F2x9C_x0000_004[9,6,3,0] Byte5 Byte4 D18F2x9C_x0000_004[A,7,4,1] Byte7 Byte6 Each of these registers control the DQS timing delay for write commands relative to MEMCLK. See 2.9.5.9 [DRAM Training]. Bits Description 31:29 Reserved. 28:24 Reserved. 23:21 WrDqsGrossDly1: DQS write gross delay. See: D18F2x9C_x0000_00[4A:30]_dct[1:0]_mp[1:0][WrDqsGrossDly]. 20:16 WrDqsFineDly1: DQS write fine delay. See: D18F2x9C_x0000_00[4A:30]_dct[1:0]_mp[1:0][WrDqsFineDly]. 15:13 Reserved. 12:8 Reserved. 360 42300 Rev 3.12 - July 14, 2015 BKDG for AMD Family 15h Models 10h-1Fh Processors 7:5 WrDqsGrossDly: DQS write gross delay. Read-write. Reset: 0. Bits Description 000b No delay 001b 0.5 MEMCLK delay 110b-010b <WrDqsGrossDly>/2 MEMCLK delay 111b 3.5 MEMCLK delay 4:0 WrDqsFineDly: DQS write fine delay. Read-write. Cold reset: 0. Bits Description 00h 0/64 MEMCLK delay 1Eh-01h <WrDqsFineDly>/64 MEMCLK delay 1Fh 31/64 MEMCLK delay D18F2x9C_x0000_00[51:50]_dct[1:0] DRAM Phase Recovery Control These registers are used by BIOS for hardware assisted DRAM training. Writes to these registers seed the phase recovery engine prior to training. Reads from the registers indicate how much the phase recovery engine has advanced to align the MEMCLK and DQS edges and is under hardware control. See 2.9.5.9 [DRAM Training]. Table 143: Register Mapping for D18F2x9C_x0000_00[51:50]_dct[1:0] Register D18F2x9C_x0000_0050_dct[1:0] D18F2x9C_x0000_0051_dct[1:0] Function Bytes 3-0 Bytes 7-4 Table 144: Byte Lane Mapping for D18F2x9C_x0000_00[51:50]_dct[1:0] Register Bits 31:24 23:16 15:8 7:0 D18F2x9C_x0000_0050 Byte3 Byte2 Byte1 Byte0 D18F2x9C_x0000_0051 Byte7 Byte6 Byte5 Byte4 Bits 31 Description Reserved. 30:29 PhRecGrossDly: phase recovery gross delay. See: PhRecGrossDly. 28:24 PhRecFineDly: phase recovery fine delay. See: PhRecFineDly. 23 Reserved. 22:21 PhRecGrossDly: phase recovery gross delay. See: PhRecGrossDly. 20:16 PhRecFineDly: phase recovery fine delay. See: PhRecFineDly. 15 Reserved. 14:13 PhRecGrossDly: phase recovery gross delay. See: PhRecGrossDly. 12:8 PhRecFineDly: phase recovery fine delay. See: PhRecFineDly. 7 Reserved. 361 42300 Rev 3.12 - July 14, 2015 BKDG for AMD Family 15h Models 10h-1Fh Processors 6:5 PhRecGrossDly: phase recovery gross delay. Read-write. Reset: X. Gross timing indicates the number of half-MEMCLK periods that the phase recovery engine advanced while aligning edges. Description Bits 00b No delay 01b 0.5 MEMCLK delay 10b 1.0 MEMCLK delay 11b 1.5 MEMCLK delay 4:0 PhRecFineDly: phase recovery fine delay. Read-write. Reset: X. Bits Description 00h 0/64 MEMCLK delay 1Eh-01h <PhRecFineDly>/64 MEMCLK delay 1Fh 31/64 MEMCLK delay D18F2x9C_x0D0F_0[F,7:0]0[8,0]_dct[1:0]_mp[1:0] Data Byte Pad Configuration Cold reset: xxxx_xxxxh. BIOS: See D18F2x9C_x0000_0000_dct[1:0]_mp[1:0]. See 2.9.2.1 for group to pin mapping. Table 145: Index Mapping for D18F2x9C_x0D0F_0[F,7:0]0[8,0]_dct[1:0]_mp[1:0] Bits Address Bits Valid Values Name D18F2x98_dct[1:0][3:2] 2h,0h Pad Group D18F2x98_dct[1:0][6] [1:0] Memory Pstate D18F2x98_dct[1:0][11:8] [7:0] Bytelane D18F2x98_dct[1:0][11:8] Fh Bytelane [7:0] Description 31:16 Reserved. 15:7 Reserved. 6:4 ProcOdt: processor on-die termination. Read-write. Specifies the resistance of the on-die termination resistors. This field is valid only when D18F2x9C_x0D0F_0[F,7:0]04_dct[1:0]_mp[1:0][POdtOff]=0. Bits Description 000b 240 ohms +/- 20% 001b 120 ohms +/- 20% 010b 80 ohms +/- 20% 011b 60 ohms +/- 20% 111b-100b Reserved 362 42300 Rev 3.12 - July 14, 2015 3 2:0 BKDG for AMD Family 15h Models 10h-1Fh Processors Reserved. DataDrvStren: data drive strength. Read-write. This field specifies the drive strength of the DRAM data pins. Bits Description 000b 0.75x 001b 1.0x 010b 1.25x 011b 1.5x 111b-100b Reserved D18F2x9C_x0D0F_0[F,7:0]02_dct[1:0] Data Byte Transmit PreDriver Calibration Cold reset: xxxx_xxxxh. BIOS: See 2.9.5.4.4. Table 146: Index Mapping for D18F2x9C_x0D0F_0[F,7:0]02_dct[1:0] D18F2x98_dct[1:0][31:16] D18F2x98_dct[1:0][15:0] 0702h 0602h 0502h 0402h 0302h 0202h 0102h 0002h 0D0Fh Byte 7 Byte 6 Byte 5 Byte 4 Byte 3 Byte 2 Byte 1 Byte 0 Table 147: Broadcast Mapping for D18F2x9C_x0D0F_0[F,7:0]02_dct[1:0] D18F2x98_dct[1:0][31:16] D18F2x98_dct[1:0][15:0] 0F02h 0D0Fh D18F2x9C_x0D0F_0[7:0]02 Table 148: Valid Values for D18F2x9C_x0D0F_0[F,7:0]02_dct[1:0][TxPreP] Bits 0h 8h-1h 9h 11h-Ah 12h 1Ah-13h 1Bh 23h-1Ch 24h 2Ch-25h 2Dh 35h-2Eh 36h 3Eh-37h 3Fh Description Slew Rate 0 (slowest) Reserved Slew Rate 1 Reserved Slew Rate 2 Reserved Slew Rate 3 Reserved Slew Rate 4 Reserved Slew Rate 5 Reserved Slew Rate 6 Reserved Slew Rate 7 (fastest) 363 42300 Rev 3.12 - July 14, 2015 Bits BKDG for AMD Family 15h Models 10h-1Fh Processors Description 31:16 Reserved. 15 ValidTxAndPre: predriver calibration code valid. Read-write; cleared-by-hardware. 1=Predriver calibration codes are copied from this register and D18F2x9C_x0D0F_0[F,7:0]0[A,6]_dct[1:0] into the associated transmit pad. Hardware clears this field after the copy is complete. 14:12 Reserved. 11:6 TxPreP: PMOS predriver calibration code. Read-write. Specifies the rising edge slew rate of the transmit pad. See: Table 148 [Valid Values for D18F2x9C_x0D0F_0[F,7:0]02_dct[1:0][TxPreP]]. After updating this value, BIOS must program ValidTxAndPre=1 for the change to take effect. 5:0 TxPreN: NMOS predriver calibration code. Read-write. Specifies the falling edge slew rate of the transmit pad. See: Table 148 [Valid Values for D18F2x9C_x0D0F_0[F,7:0]02_dct[1:0][TxPreP]]. After updating this value, BIOS must program ValidTxAndPre=1 for the change to take effect. D18F2x9C_x0D0F_0[F,7:0]04_dct[1:0]_mp[1:0] Data Byte Pad Configuration Cold reset: 0000_0033h. BIOS: See D18F2x9C_x0000_0000_dct[1:0]_mp[1:0]. See 2.9.2.1 for group to pin mapping. Table 149: Index Mapping for D18F2x9C_x0D0F_0[F,7:0]04_dct[1:0]_mp[1:0] Bits Address Bits Valid Values Name D18F2x98_dct[1:0][3:2] 1h Dbyte Pad Group D18F2x98_dct[1:0][6] [1:0] Memory Pstate D18F2x98_dct[1:0][11:8] [7:0] Bytelane D18F2x98_dct[1:0][11:8] Fh Bytelane [7:0] Description 31:13 Reserved. 12 POdtOff: processor ODT off. Read-write. BIOS: See 2.9.5.6.6.1=Phy disables receiver pad termination. 0=Phy enables receiver pad termination. Hardware programs this field for the current M-state (D18F2x9C_x0D0F_E006_dct[1:0][PhyPS]) with the value in D18F2x94_dct[1:0][ProcOdtDis] when the memory frequency is updated. See 2.9.5.4.2. BIOS must reprogram this field after each frequency change if the target value differs from D18F2x94_dct[1:0][ProcOdtDis]. 11:7 Reserved. 6:4 ProcOdt: processor on-die termination. Read-write. Specifies the resistance of the on-die termination resistors. This field is valid only when POdtOff=0. Bits Description 000b 240 ohms +/- 20% 001b 120 ohms +/- 20% 010b 80 ohms +/- 20% 011b 60 ohms +/- 20% 111b-100b Reserved 364 42300 Rev 3.12 - July 14, 2015 3 BKDG for AMD Family 15h Models 10h-1Fh Processors Reserved. 2:0 DqsDrvStren: DQS drive strength. Read-write. Specifies the drive strength of the DQS pins. Description Bits 000b 0.75x 001b 1.0x 010b 1.25x 011b 1.5x 111b-100b Reserved D18F2x9C_x0D0F_0[F,7:0]0[A,6]_dct[1:0] Data Byte Transmit PreDriver Calibration 2 Cold reset: xxxx_xxxxh. BIOS: See 2.9.5.4.4. Table 150: Index Mapping for D18F2x9C_x0D0F_0[F,7:0]06_dct[1:0] D18F2x98_dct[1:0][31:16] D18F2x98_dct[1:0][15:0] 0706h 0606h 0506h 0406h 0306h 0206h 0106h 0006h 0D0Fh Byte 7 Byte 6 Byte 5 Byte 4 Byte 3 Byte 2 Byte 1 Byte 0 Table 151: Index Mapping for D18F2x9C_x0D0F_0[F,7:0]0A_dct[1:0] D18F2x98_dct[1:0][31:16] D18F2x98_dct[1:0][15:0] 070Ah 060Ah 050Ah 040Ah 030Ah 020Ah 010Ah 000Ah 0D0Fh Byte 7 Byte 6 Byte 5 Byte 4 Byte 3 Byte 2 Byte 1 Byte 0 Table 152: Broadcast Mapping for D18F2x9C_x0D0F_0[F,7:0]0[A,6]_dct[1:0] D18F2x98_dct[1:0][31:16] D18F2x98_dct[1:0][15:0] 0F0Ah 0D0Fh Bits D18F2x9C_x0D0F_0[7:0]0A 0F06h D18F2x9C_x0D0F_0[7:0]06 Description 31:12 Reserved. 11:6 TxPreP: PMOS predriver calibration code. Read-write. This field specifies the rising edge slew rate of the transmit pad. See: D18F2x9C_x0D0F_0[F,7:0]02_dct[1:0][TxPreP]. After updating this value, BIOS must program D18F2x9C_x0D0F_0[F,7:0]02_dct[1:0][ValidTxAndPre]=1 for the change to take effect. 5:0 TxPreN: NMOS predriver calibration code. Read-write. This field specifies the falling edge slew rate of the transmit pad. See: D18F2x9C_x0D0F_0[F,7:0]02_dct[1:0][TxPreN]. After updating this value, BIOS must program D18F2x9C_x0D0F_0[F,7:0]02_dct[1:0][ValidTxAndPre]=1 for the change to take effect. D18F2x9C_x0D0F_0[F,7:0]0F_dct[1:0] Data Byte DLL Clock Enable Cold reset: 0000_0013h. 365 42300 Rev 3.12 - July 14, 2015 BKDG for AMD Family 15h Models 10h-1Fh Processors Table 153: Index addresses for D18F2x9C_x0D0F_0[F,7:0]0F_dct[1:0] D18F2x98_dct[1:0][31:16] D18F2x98_dct[1:0][15:0] 070Fh 060Fh 050Fh 040Fh 030Fh 020Fh 010Fh 000Fh 0D0Fh Byte 7 Byte 6 Byte 5 Byte 4 Byte 3 Byte 2 Byte 1 Byte 0 Table 154: Broadcast write index address for D18F2x9C_x0D0F_0[F,7:0]0F_dct[1:0] D18F2x98_dct[1:0][31:16] D18F2x98_dct[1:0][15:0] 0F0Fh 0D0Fh Bits D18F2x9C_x0D0F_0[7:0]0F Description 31:15 Reserved. 14:12 AlwaysEnDllClks: always enable DLL clocks. Read-write. BIOS: 2.9.5.4.3.0=DLL clocks are turned off during periods of inactivity. 1=DLL clocks remain on during inactivity. Prior to programming AlwaysEnDllClks to a value other than 000b, D18F2x9C_x0000_000D_dct[1:0]_mp[1:0][RxMaxDurDllNoLock, TxMaxDurDllNoLock] must both be programmed to 0000b. The bits AlwaysEnDllClks[2:0] are mapped to DLLs as follows: Bit Definition [2] TxDqs DLL [1] TxDq DLL [0] RxEn DLL 11:0 Reserved. D18F2x9C_x0D0F_0[F,7:0]10_dct[1:0]_mp[1:0] Data Byte DLL Power Management Cold reset: 0000_0000h. Table 155: Index Addresses for D18F2x9C_x0D0F_0[F,7:0]10_dct[1:0]_mp[1:0] D18F2x98_dct[1:0][31:0] 0D0F_0010h 0D0F_0110h 0D0F_0210h 0D0F_0310h 0D0F_0410h 0D0F_0510h 0D0F_0610h 0D0F_0710h 0D0F_0F10h Bits Function Byte 0 Byte 1 Byte 2 Byte 3 Byte 4 Byte 5 Byte 6 Byte 7 Byte [7:0] Description 31:13 Reserved. 366 42300 Rev 3.12 - July 14, 2015 12 BKDG for AMD Family 15h Models 10h-1Fh Processors EnRxPadStandby: enable receiver pad standby. Read-write. BIOS: See 2.9.5.7.1=Phy receiver standby mode is enabled to save power when not receiving data. 0=Phy receiver standby mode is disabled. 11:10 Reserved. 9:8 7 TxDLLWakeupTime: transmit DLL wakeup time. Read-write. BIOS: See 2.9.5.11. Specifies the number of PCLK's that the DLL standby signal must deassert prior to a DLL relock event or before write traffic is sent to transmit DLLs. This field is an alias of D18F2x9C_x0000_000D_dct[1:0]_mp[1:0][TxDLLWakeupTime]. Reserved. 6:4 TxCPUpdPeriod: transmit charge pump DLL wakeup time. Read-write. BIOS: See 2.9.5.11. Specifies the number of DLL relocks required to keep the TxDLLs locked for the period where there is no write traffic. This field is an alias of D18F2x9C_x0000_000D_dct[1:0]_mp[1:0][TxCPUpdPeriod]. 3:0 TxMaxDurDllNoLock: dll no lock. Read-write. BIOS: See 2.9.5.7.This field enables DLL standby and specifies the number of PCLK cycles that occur before the phy DLLs relock. A DLL relock occurs every MIN(2^RxMaxDurDllNoLock, 512) clock cycles if there are no accesses during the period. 0=Disable the DLL standby power saving feature. This field is an alias of D18F2x9C_x0000_000D_dct[1:0]_mp[1:0][TxMaxDurDllNoLock]. D18F2x9C_x0D0F_0[F,7:0]11_dct[1:0]_mp[1:0] Data Byte DLL Power Management Cold reset: 0000_0000h. Table 156: Index Addresses for D18F2x9C_x0D0F_0[F,7:0]11_dct[1:0]_mp[1:0] D18F2x98_dct[1:0][31:0] 0D0F_0011h 0D0F_0111h 0D0F_0211h 0D0F_0311h 0D0F_0411h 0D0F_0511h 0D0F_0611h 0D0F_0711h 0D0F_0F11h Bits Function Byte 0 Byte 1 Byte 2 Byte 3 Byte 4 Byte 5 Byte 6 Byte 7 Byte [7:0] Description 31:10 Reserved. 9:8 7 RxDLLWakeupTime: transmit DLL wakeup time. Read-write. BIOS: See 2.9.5.11. Specifies the number of PCLK's that the DLL standby signal must deassert prior to a DLL relock event or before read traffic is sent to receive DLLs. This field is an alias of D18F2x9C_x0000_000D_dct[1:0]_mp[1:0][RxDLLWakeupTime]. Reserved. 367 42300 Rev 3.12 - July 14, 2015 BKDG for AMD Family 15h Models 10h-1Fh Processors 6:4 RxCPUpdPeriod: transmit charge pump DLL wakeup time. Read-write. BIOS: See 2.9.5.11. Specifies the number of DLL relocks required to keep the RxDLLs locked for the period where there is no read traffic. This field is an alias of D18F2x9C_x0000_000D_dct[1:0]_mp[1:0][RxCPUpdPeriod]. 3:0 RxMaxDurDllNoLock: dll no lock. Read-write. BIOS: See 2.9.5.7.This field enables DLL standby and specifies the number of PCLK cycles that occur before the phy DLLs relock. A DLL relock occurs every MIN(2^RxMaxDurDllNoLock, 512) clock cycles if there are no accesses during the period. 0=Disable the DLL standby power saving feature. This field is an alias of D18F2x9C_x0000_000D_dct[1:0]_mp[1:0][RxMaxDurDllNoLock]. D18F2x9C_x0D0F_0[F,7:0]13_dct[1:0]_mp[1:0] Data Byte DLL Configuration Table 157: Index Addresses for D18F2x9C_x0D0F_0[F,7:0]13_dct[1:0]_mp[1:0] D18F2x98_dct[1:0][31:0] 0D0F_0013h 0D0F_0113h 0D0F_0213h 0D0F_0313h 0D0F_0413h 0D0F_0513h 0D0F_0613h 0D0F_0713h 0D0F_0F13h Bits Function Byte 0 Byte 1 Byte 2 Byte 3 Byte 4 Byte 5 Byte 6 Byte 7 Byte [7:0] Description 31:15 Reserved. 14 ProcOdtAdv: ProcOdt advance. Read-write. Cold reset: 1. BIOS: IF(SODIMM && DdrRate <= 1333) THEN 0 ELSE 1 ENDIF. 0=ProcOdt is asserted 1.5 PCLK before DqsEn . 1=If preceding write, ProcOdt is asserted 2.5 PCLK before DqsEn, else, ProcOdt is asserted 5.0 PCLK before DqsEn. 13:9 Reserved. 8 RxSsbMntClkEn: receive channel maintenance clock enable. Read-write. Cold reset: 0. BIOS: See 2.9.5.11. 1=Enable receive channel maintenance clocks to improve internal timing margin at the cost of some extra power. To enable clock generation, BIOS must first enable maintenance clocks in the DCT (see D18F2x248_dct[1:0]_mp[1:0][RxChMntClkEn]). 0=Disable clocks. The maintenance clock period is fixed at 8 PCLKs when no traffic occurs. 7 RxDqsUDllPowerDown: receive DQS upper DLL power down. Read-write. Cold reset: 0. BIOS: See 2.9.5.11. 1=Power down the upper receiver DQS DLL. 6:2 Reserved. 1 DllDisEarlyU: DLL disable early upper. Read-write. Cold reset: 0. BIOS: See 2.9.5.11. 1=Disable upper receiver DQS DLL early timing for power savings. 0 DllDisEarlyL: DLL disable early lower. Read-write. Cold reset: 0. BIOS: See 2.9.5.11. 1=Disable lower receiver DQS DLL early timing for power savings. 368 42300 Rev 3.12 - July 14, 2015 BKDG for AMD Family 15h Models 10h-1Fh Processors D18F2x9C_x0D0F_0[F,7:0]1C_dct[1:0]_mp[1:0] Data Byte DLL Power Management Cold reset: 0000_0000h. Table 158: Index Addresses for D18F2x9C_x0D0F_0[F,7:0]1C_dct[1:0]_mp[1:0] D18F2x98_dct[1:0][31:8] 0D0F_001Ch 0D0F_011Ch 0D0F_021Ch 0D0F_031Ch 0D0F_041Ch 0D0F_051Ch 0D0F_061Ch 0D0F_071Ch 0D0F_0F1Ch Bits Function Byte 0 Byte 1 Byte 2 Byte 3 Byte 4 Byte 5 Byte 6 Byte 7 Byte [7:0] Description 15 RxDllStggrEn: Rx DLL stagger enable. Read-write. BIOS: 0. 1=Insert RxDllStggrDly delay prior to waking the receive DLLs from standby. 0=DLLs wake from standby at the default pending-activity reference point. 14 Reserved. 13:8 RxDllStggrDly[5:0]: Rx DLL stagger delay. Read-write. BIOS: 0. Specifies the delay in PCLK cycles from a CAS command or an internal DLL maintenance lock timing event before the receive DLLs exit standby if the data bus is still idle. If data bus traffic is pending and the DLLs are in standby then they wake immediately regardless of the delay value. 7 TxDllStggrEn: Tx DLL stagger enable. Read-write. BIOS: 0. 1=Insert TxDllStggrDly delay prior to waking the transmit DLL from standby. 0=DLL wakes from standby at the default pending-activity reference point. 6 Reserved. 5:0 TxDllStggrDly[5:0]: Tx DLL stagger delay. Read-write. BIOS: 0. Specifies the delay in PCLK cycles from a CAS command or an internal DLL maintenance lock timing event before the transmit DLL exits standby if the data bus is still idle. If data bus traffic is pending and the DLL is in standby then it wakes immediately regardless of the delay value. D18F2x9C_x0D0F_0[F,7:0]1F_dct[1:0]_mp[1:0] Data Byte Receiver Configuration Cold reset: 0000_2003h. Table 159: Index Addresses for D18F2x9C_x0D0F_0[F,7:0]1F_dct[1:0]_mp[1:0] D18F2x98_dct[1:0][31:8] 0D0F_001Fh 0D0F_011Fh 0D0F_021Fh 0D0F_031Fh 0D0F_041Fh 0D0F_051Fh Function Byte 0 Byte 1 Byte 2 Byte 3 Byte 4 Byte 5 369 42300 Rev 3.12 - July 14, 2015 BKDG for AMD Family 15h Models 10h-1Fh Processors Table 159: Index Addresses for D18F2x9C_x0D0F_0[F,7:0]1F_dct[1:0]_mp[1:0] 0D0F_061Fh 0D0F_071Fh 0D0F_0F1Fh Bits Byte 6 Byte 7 Byte [7:0] Description 31:9 Reserved. 8 Rx4thStgEn: receiver 4th input stage amplifier enable. Read-write. BIOS: 2.9.5.9.4. 1=Enable 4th stage of amplifiers in the DQ receiver. 0=Disable 4th stage. This bit is ignored if RxBypass3rd4thStg=1. 7:5 Reserved. 4:3 RxVioLvl: receiver voltage level. Read-write. BIOS: 2.9.5.4.1. Specifies the VDDIO voltage level. Bits Description 00b 1.5 V 01b 1.35 V 10b 1.25 V 11b Reserved 2 RxBypass3rd4thStg: receiver bypass 3rd and 4th input stage. Read-write. BIOS: 2.9.5.9.4. 1=Bypass the 3rd and 4th stages of amplifiers in the DQ receiver. 0=Disable bypass. This setting adds delay to DQ signals to increase the sampling envelope. 1:0 Reserved. D18F2x9C_x0D0F_0[F,7:0]30_dct[1:0] Data Byte DLL Configuration and Power Down Cold reset: 0000_0001h. Table 160: Index Addresses for D18F2x9C_x0D0F_0[F,7:0]30_dct[1:0] D18F2x98_dct[1:0][31:16] D18F2x98_dct[1:0][15:0] 0730h 0630h 0530h 0430h 0330h 0230h 0130h 0030h 0D0Fh Byte 7 Byte 6 Byte 5 Byte 4 Byte 3 Byte 2 Byte 1 Byte 0 Table 161: Broadcast Write Index Address for D18F2x9C_x0D0F_0[F,7:0]30_dct[1:0] D18F2x98_dct[1:0][31:16] D18F2x98_dct[1:0][15:0] 0F30h 0D0Fh Bits D18F2x9C_x0D0F_0[7:0]30 Description 31:9 Reserved. 8 7:5 4 BlockRxDqsLock: block rx dqs lock. Read-write. BIOS: 2.9.5.9.3. Specifies how the receive DLLs lock. 1=Lock on PCLK. 0=Lock on both PCLK and the received DQS Reserved. PwrDn: power down. Read-write. BIOS: IF (REG==D18F2x9C_x0D0F_0830_dct[1:0] && D18F2x90_dct[1:0][DimmEccEn]==0) THEN 1 ELSE 0 ENDIF. 1=Turn off DLL circuitry. 370 42300 Rev 3.12 - July 14, 2015 3 2:0 BKDG for AMD Family 15h Models 10h-1Fh Processors Reserved. DllGain: Dll relock gain. Read-write. Specifies the gain of DLL relock circuitry. 000b=Fastest lock time...101b= Slowest lock time. Values of 110b and 111b are reserved. D18F2x9C_x0D0F_0[F,7:0]31_dct[1:0] Data Byte Fence2 Threshold BIOS: 2.9.5.4.3. Table 162: Index addresses for D18F2x9C_x0D0F_0[F,7:0]31_dct[1:0] D18F2x98_dct[1:0][31:16] D18F2x98_dct[1:0][15:0] 0731h 0631h 0531h 0431h 0331h 0231h 0131h 0031h 0D0Fh Byte 7 Byte 6 Byte 5 Byte 4 Byte 3 Byte 2 Byte 1 Byte 0 Table 163: Broadcast write index address for D18F2x9C_x0D0F_0[F,7:0]31_dct[1:0] D18F2x98_dct[1:0][31:16] D18F2x98_dct[1:0][15:0] 0F31h 0D0Fh Bits D18F2x9C_x0D0F_0[7:0]31 Description 31:10 Reserved. 9 8:5 4 3:0 Fence2EnableTxDll: phy fence2 enable transmit DLL. Read-write. Cold reset: 0. 1=Enable the use of Fence2ThresholdTxDll for transmit DLL fence threshold. 0=Use D18F2x9C_x0000_000C_dct[1:0][FenceThresholdTxDll]. Fence2ThresholdTxDll: phy fence2 threshold transmit DLL. Read-write. Cold reset: 0. If Fence2EnableTxDll=1, this field specifies the fence delay threshold value used for DQS receiver valid. This field is only used during DQS receiver enable training. See Fence2ThresholdTxPad. Fence2EnableTxPad: fence2 enable transmit pad. Read-write. Cold reset: 0. 1=Enable the use of Fence2ThresholdTxPad for transmit pad fence threshold. 0=Use D18F2x9C_x0000_000C_dct[1:0][FenceThresholdTxPad]. Fence2ThresholdTxPad: phy fence2 threshold transmit pad. Read-write. Cold reset: 0. If Fence2EnableTxPad=1, this field specifies the fence delay threshold value used for write data and write DQS. Bits Definition 0h No delay 1h 1/64 MEMCLK delay Eh-2h <Fence2ThresholdTxPad> MEMCLK delay Fh 15/64 MEMCLK delay D18F2x9C_x0D0F_2[F,2:0]00_dct[1:0]_mp[1:0] Clock Pad Configuration Cold reset: xxxx_xxxxh. BIOS: See D18F2x9C_x0000_0000_dct[1:0]_mp[1:0]. See 2.9.2.1 for clock chip to pin mapping. 371 42300 Rev 3.12 - July 14, 2015 BKDG for AMD Family 15h Models 10h-1Fh Processors Table 164: Index Addresses for D18F2x9C_x0D0F_2[F,2:0]00_dct[1:0]_mp[1:0] D18F2x98_dct[1:0][31:0] 0D0F_2000h 0D0F_2100h 0D0F_2200h 0D0F_2F00h Bits Function Clk chip 0 Clk chip 1 Reserved All clk chips Description 31:3 Reserved. 2:0 ClkDrvStren: MEMCLK drive strength. Read-write. This field specifies the drive strength of the MEMCLK pins. Bits Description 000b 1.0x 001b 1.25x 010b 1.5x 011b 2.0x 111b-100b Reserved D18F2x9C_x0D0F_2[2:0]02_dct[1:0] Clock Transmit PreDriver Calibration Cold reset: xxxx_xxxxh. BIOS: See 2.9.5.4.4. Table 165: Index Mapping for D18F2x9C_x0D0F_2[2:0]02_dct[1:0] D18F2x98_dct[1:0][31:0] 0D0F_2002h 0D0F_2102h 0D0F_2202h Bits Function Clock 0 Pad Group 0 Clock 1 Pad Group 0 Clock 2 Pad Group 0 Description 31:16 Reserved. 15 ValidTxAndPre: predriver calibration code valid. Read-write; cleared-by-hardware. 1=Predriver calibration codes are copied from this register into the associated transmit pad. Hardware clears this field after the copy is complete. 14:12 Reserved. 11:6 TxPreP: PMOS predriver calibration code. Read-write. This field specifies the rising edge slew rate of the transmit pad. See: D18F2x9C_x0D0F_0[F,7:0]02_dct[1:0][TxPreP]. After updating this value, BIOS must program ValidTxAndPre=1 for the change to take effect. 5:0 TxPreN: NMOS predriver calibration code. Read-write. This field specifies the falling edge slew rate of the transmit pad. See: D18F2x9C_x0D0F_0[F,7:0]02_dct[1:0][TxPreN]. After updating this value, BIOS must program ValidTxAndPre=1 for the change to take effect. D18F2x9C_x0D0F_[C,8,2][2:0]1F Receiver Configuration Cold reset: 0000_2000h. 372 42300 Rev 3.12 - July 14, 2015 BKDG for AMD Family 15h Models 10h-1Fh Processors Table 166: Index address mapping for D18F2x9C_x0D0F_[C,8,2][2:0]1F D18F2x98_dct[1:0][31:0] Function 0D0F_201Fh Clock 0 0D0F_211Fh Clock 1 0D0F_221Fh Clock 2 0D0F_801Fh Cmd/Addr 0 0D0F_811Fh Cmd/Addr 1 0D0F_821Fh Reserved 0D0F_C01Fh Address 0D0F_C11Fh Reserved 0D0F_C21Fh Reserved Bits Description 31:5 Reserved. 4:3 RxVioLvl: receiver voltage level. Read-write. BIOS: See 2.9.5.4.1. Specifies the VDDIO voltage level. Bits Description 00b 1.5 V 01b 1.35 V 10b 1.25 V 11b Reserved 2:0 Reserved. D18F2x9C_x0D0F_2[F,2:0]30_dct[1:0] Clock DLL Configuration and Power Down Cold reset: 0000_0001h. Table 167: Index Addresses for D18F2x9C_x0D0F_2[F,2:0]30_dct[1:0] D18F2x98_dct[1:0][31:0] 0D0F_2030h 0D0F_2130h 0D0F_2230h 0D0F_2F30h Bits Function Clk chip 0 Clk chip 1 Reserved All clk chips Description 31:16 Reserved. 15:5 Reserved. 4 PwrDn: power down. Read-write. 1=Turn off DLL circuitry. 3 Reserved. 2:0 DllGain: Dll relock gain. Read-write. See D18F2x9C_x0D0F_0[F,7:0]30_dct[1:0][DllGain] for a description. 373 42300 Rev 3.12 - July 14, 2015 BKDG for AMD Family 15h Models 10h-1Fh Processors D18F2x9C_x0D0F_4003_dct[1] DCT0 M1 Fence Value Cold reset: 0000_3F3Fh. This register is used by BIOS to temporarily store the trained fence values for DCT0 for the M1 memory P-state, and has no effect on the operation of the hardware. See 2.9.5.4.3. This register is accessed through DCT1 only, and is available regardless of D18F2x94_dct[1:0][DisDramInterface]. Bits Description 31:15 Reserved. 14:10 TxDll: Transmit dll fence value. Read-write. BIOS saves the trained TxDll to this field until such time as it is able to write the value to NVRAM. 9:5 RxDll: Receiver dll fence value. Read-write. BIOS saves the trained RxDll to this field until such time as it is able to write the value to NVRAM. 4:0 TxPad: Transmit pad fence value. Read-write. BIOS saves the trained TxPad to this field until such time as it is able to write the value to NVRAM. D18F2x9C_x0D0F_4004_dct[1] DCT1 M1 Fence Value Cold reset: 0000_0000h. This register is used by BIOS to temporarily store the trained fence values for DCT1 for the M1 memory P-state, and has no effect on the operation of the hardware. See 2.9.5.4.3. This register is accessed through DCT1 only, and is available regardless of D18F2x94_dct[1:0][DisDramInterface]. Bits Description 31:15 Reserved. 14:10 TxDll: Transmit dll fence value. Read-write. BIOS saves the trained TxDll to this field until such time as it is able to write the value to NVRAM. 9:5 RxDll: Receiver dll fence value. Read-write. BIOS saves the trained RxDll to this field until such time as it is able to write the value to NVRAM. 4:0 TxPad: Transmit pad fence value. Read-write. BIOS saves the trained TxPad to this field until such time as it is able to write the value to NVRAM. D18F2x9C_x0D0F_4009_dct[1:0] Phy Cmp Configuration Cold reset: 0000_2000h. Bits Description 31:16 Reserved. 15:14 CmpVioLvl: cmp voltage level. Read-write. BIOS: See 2.9.5.4.1. This field specifies the VDDIO voltage level. Setting this field in DCT0 adjusts the VDDIO voltage level for DCT0 and DCT1. Setting this field in DCT1 has no effect. Bits Description 00b 1.5 V 01b 1.35 V 10b 1.25 V 11b Reserved 13:4 Reserved. 374 42300 Rev 3.12 - July 14, 2015 BKDG for AMD Family 15h Models 10h-1Fh Processors 3:2 ComparatorAdjust: comparator adjust. Read-write. BIOS: D18F2x9C_x0D0F_0[F,7:0]1F_dct[1:0]_mp[1:0][RxVioLvl]. This field specifies the adjustment signals for the comparator differential amplifier. Setting this field in DCT0 adjusts the comparator for DCT0 and DCT1. Setting this field in DCT1 has no effect. 1:0 Reserved. 375 42300 Rev 3.12 - July 14, 2015 BKDG for AMD Family 15h Models 10h-1Fh Processors D18F2x9C_x0D0F_8[1:0]0[8,4,0]_dct[1:0]_mp[1:0] CAD1 Pad Configuration Cold reset: xxxx_xxxxh. BIOS: See D18F2x9C_x0000_0000_dct[1:0]_mp[1:0]. See 2.9.2.1 for chip and group to pin mapping. Table 168: Index Mapping for D18F2x9C_x0D0F_8[1:0]0[8,4,0]_dct[1:0]_mp[1:0] Bits Address Bits Valid Values Name D18F2x98_dct[1:0][3:2] [2:0] Pad Group D18F2x98_dct[1:0][6] [1:0] Memory Pstate D18F2x98_dct[1:0][11:8] [1:0] Cmd/Addr chip D18F2x98_dct[1:0][11:8] Fh Cmd/Addr chip[1:0] Description 31:3 Reserved. 2:0 DrvStrength: address/command/csodt drive strength. Read-write. This field specifies the drive strength of the address, CS, ODT, RAS, CAS, WE, bank and parity pins. Bits Description 000b 1.0x 001b 1.25x 010b 1.5x 011b 2.0x 111b-100b Reserved D18F2x9C_x0D0F_[C,8][1:0]02_dct[1:0] Transmit PreDriver Calibration Cold reset: xxxx_xxxxh. BIOS: See 2.9.5.4.4. Table 169: Index Mapping for D18F2x9C_x0D0F_[C,8][1:0]02_dct[1:0] D18F2x98_dct[1:0][31:0] 0D0F_8002h 0D0F_8102h 0D0F_C002h Bits Function Cmd/Addr 0 Pad Group 0 Cmd/Addr 1 Pad Group 0 Address Pad Group 0 Description 31:16 Reserved. 15 ValidTxAndPre: predriver calibration code valid. Read-write; cleared-by-hardware. 1=Predriver calibration codes are copied from this register and D18F2x9C_x0D0F_[C,8][1:0][12,0E,0A,06]_dct[1:0] into the associated transmit pad. Hardware clears this field after the copy is complete. 14:12 Reserved. 376 42300 Rev 3.12 - July 14, 2015 BKDG for AMD Family 15h Models 10h-1Fh Processors 11:6 TxPreP: PMOS predriver calibration code. Read-write. This field specifies the rising edge slew rate of the transmit pad. See: D18F2x9C_x0D0F_0[F,7:0]02_dct[1:0][TxPreP]. After updating this value, BIOS must program ValidTxAndPre=1 for the change to take effect. 5:0 TxPreN: NMOS predriver calibration code. Read-write. This field specifies the falling edge slew rate of the transmit pad. See: D18F2x9C_x0D0F_0[F,7:0]02_dct[1:0][TxPreN]. After updating this value, BIOS must program ValidTxAndPre=1 for the change to take effect. D18F2x9C_x0D0F_[C,8][1:0][12,0E,0A,06]_dct[1:0] Transmit PreDriver Calibration 2 Cold reset: xxxx_xxxxh. BIOS: See 2.9.5.4.4. Table 170: Index Mapping for D18F2x9C_x0D0F_[C,8][1:0][12,0E,0A,06]_dct[1:0] D18F2x98_dct[1:0][31:0] 0D0F_8006h 0D0F_800Ah 0D0F_8106h 0D0F_810Ah 0D0F_C006h 0D0F_C00Ah 0D0F_C00Eh 0D0F_C012h Bits Function Cmd/Addr 0 Pad Group 1 Cmd/Addr 0 Pad Group 2 Cmd/Addr 1 Pad Group 1 Cmd/Addr 1 Pad Group 2 Address Pad Group 1 Address Pad Group 2 Address Pad Group 3 Address Pad Group 4 Description 31:12 Reserved. 11:6 TxPreP: PMOS predriver calibration code. Read-write. This field specifies the rising edge slew rate of the transmit pad. See: D18F2x9C_x0D0F_0[F,7:0]02_dct[1:0][TxPreP]. After updating this value, BIOS must program D18F2x9C_x0D0F_[C,8][1:0]02_dct[1:0][ValidTxAndPre]=1 for the change to take effect. 5:0 TxPreN: NMOS predriver calibration code. Read-write. This field specifies the falling edge slew rate of the transmit pad. See: D18F2x9C_x0D0F_0[F,7:0]02_dct[1:0][TxPreN]. After updating this value, BIOS must program D18F2x9C_x0D0F_[C,8][1:0]02_dct[1:0][ValidTxAndPre]=1 for the change to take effect. D18F2x9C_x0D0F_8[1:0]20_dct[1:0]_mp[1:0] DLL Delay and Configuration Cold reset: 0000_8000h. BIOS: D18F2x9C_x0000_0004_dct[1:0]_mp[1:0]. See 2.9.2.1 for chip to pin mapping. Table 171: Index Mapping for D18F2x9C_x0D0F_8[1:0]20_dct[1:0]_mp[1:0] Address Bits Valid Values Name D18F2x98_dct[1:0][6] [1:0] Memory Pstate D18F2x98_dct[1:0][11:8] [1:0] Cmd/Addr 1 chip D18F2x98_dct[1:0][11:8] Fh Cmd/Addr 1 chip[1:0] Note: Control of certain pins in Cmd/Addr 0 chip is overriden with D18F2x9C_x0D0F_8[1:0]20_dct[1:0]_mp[1:0][DiffTimingEn] 377 42300 Rev 3.12 - July 14, 2015 Bits BKDG for AMD Family 15h Models 10h-1Fh Processors Description 31:8 Reserved. 7 FenceBit: fence bit. Read-write; updated-by-hardware. 1= A 0.5-PCLK-delayed version of the signal from the TxFIFO is used for the input of the output pad. 0= An undelayed version is used for the input of the output pad. 6 Fence2Bit: second fence bit. Read-write; updated-by-hardware. 1= A 1-PCLK-delayed version of the signal from the TxFIFO is used for the input of the output pad. FenceBit should be programmed 1b. 0= FenceBit determines the input to the output pad. 5 Reserved. 4:0 Delay: delay. Read-write. Specifies the time that the signals are delayed from the default setup time. Bits Description 00h 0/64 MEMCLK delay 1Eh-01h <Delay>/64 MEMCLK delay 1Fh 31/64 MEMCLK delay D18F2x9C_x0D0F_8021_dct[1:0]_mp[1:0] DLL CS 6 & 7 Timing Control Cold reset: 0000_0000h. BIOS: 2.9.5.4.3.D18F2x9C_x0000_0004_dct[1:0]_mp[1:0] must be reprogrammed any time new values are written to this register. Bits Description 31:16 Reserved. 15 DiffTimingEn: differential timing enable. Read-write. Enables independent timing controls for CS[7:6]. 1=Timing is specified by Fence and Delay. 0=Timing is specified by D18F2x9C_x0000_0004_dct[1:0]_mp[1:0][CsOdtFineDelay]. 14:8 Reserved. 7 Fence: fence. Read-write. Adjusts the phase relationship between the fifo and the pad when DiffTimingEn=1. 6:5 Reserved. 4:0 Delay: delay. Read-write. Specifies the time that the CS6 & 7 pins are delayed from the default setup time when DiffTimingEn=1. Bits Description 00h 0/64 MEMCLK delay 1Eh-01h <Delay>/64 MEMCLK delay 1Fh 31/64 MEMCLK delay D18F2x9C_x0D0F_812F_dct[1:0] Tristate Configuration Reset: 0000_00A0h. BIOS: See 2.9.5.11. Bits Description 31:1 Reserved. 0 PARTri: MEMPAR tri-state. Read-write. Specifies tri-state control for the memory parity signal. 1=Signal is tri-stated. 0=Signal is not tri-stated. 378 42300 Rev 3.12 - July 14, 2015 BKDG for AMD Family 15h Models 10h-1Fh Processors D18F2x9C_x0D0F_[C,8][F:0]30_dct[1:0] Cmd/Addr DLL Configuration and Power Down Cold reset: 0000_0001h. BIOS: See 2.9.5.4.4. See Table 19 and Table 20 for chip to pin mapping. Table 172: Index Mapping for D18F2x9C_x0D0F_[C,8][F:0]30_dct[1:0] D18F2x98_dct[1:0][31:0] Function 0D0F_8030h Cmd/Addr 0 0D0F_8130h Cmd/Addr 1 0D0F_8F30h Cmd/Addr 0 and Cmd/Addr 1 0D0F_C030h Address 2 Bits Description 15:5 Reserved. 4 PwrDn: power down. Read-write. 1=Turn off DLL circuitry. 3 Reserved. 2:0 DllGain: Dll relock gain. Read-write. See D18F2x9C_x0D0F_0[F,7:0]30_dct[1:0][DllGain] for a description. D18F2x9C_x0D0F_C000_dct[1:0]_mp[1:0] CKE 2.0X Pad Configuration Reset: 0000_0003h. BIOS: See 2.9.5.11. Bits Description 31:9 Reserved. 8 LowPowerDrvStrengthEn: low power drive strength enable. Read-write. 1=CKE driver is forced to low drive strength when phy is in self-refresh mode. 0=CKE is driven with strength specified by D18F2x9C_x0000_0000_dct[1:0]_mp[1:0][CkeDrvStren] when phy is in self-refresh mode. This bit may only be set if CkeDrvStren==010b or 011b. 7:3 Reserved. 2:0 CkeDrvStren: CKE drive strength. Read-write. This field specifies the drive strength of the CKE pins. Bits Description 000b 1.0x 001b 1.25x 010b 1.5x 011b 2.0x 111b-100b Reserved 379 42300 Rev 3.12 - July 14, 2015 BKDG for AMD Family 15h Models 10h-1Fh Processors D18F2x9C_x0D0F_C0[10,0C,08,04]_dct[1:0]_mp[1:0] CAD2 Pad Configuration Cold reset: xxxx_xxxxh. BIOS: See D18F2x9C_x0000_0000_dct[1:0]_mp[1:0]. See 2.9.2.1 for Address2 chip group to pin mapping. Table 173: Index Mapping for D18F2x9C_x0D0F_C0[10,0C,08,04]_dct[1:0]_mp[1:0] Bits Address Bits Valid Values Name D18F2x98_dct[1:0][4:2] [4:1] Pad Group D18F2x98_dct[1:0][6] [1:0] Memory Pstate D18F2x98_dct[1:0][15:12] Ch Address2 chip Description 31:3 Reserved. 2:0 DrvStrength: address drive strength. Read-write. Cold reset: 011b. This field specifies the drive strength of the address, CS, ODT, RAS, CAS, WE, bank and parity pins. Bits Description 000b 1.0x 001b 1.25x 010b 1.5x 011b 2.0x 111b-100b Reserved D18F2x9C_x0D0F_C020_dct[1:0]_mp[1:0] CAD2 Address DLL Delay and Configuration Cold reset: 0000_8000h. BIOS: D18F2x9C_x0000_0004_dct[1:0]_mp[1:0]. This register controls timing for signals on Address[15:0] pins. Table 174: Index Mapping for D18F2x9C_x0D0F_C020_dct[1:0]_mp[1:0] Bits Address Bits Valid Values Name D18F2x98_dct[1:0][6] [1:0] Memory Pstate Description 31:8 Reserved. 7 FenceBit: fence bit. See: D18F2x9C_x0D0F_8[1:0]20_dct[1:0]_mp[1:0][FenceBit]. 6 Fence2Bit: second fence bit. See: D18F2x9C_x0D0F_8[1:0]20_dct[1:0]_mp[1:0][Fence2Bit]. 5 Reserved. 4:0 Delay: delay. See: D18F2x9C_x0D0F_8[1:0]20_dct[1:0]_mp[1:0][Delay]. D18F2x9C_x0D0F_C021_dct[1:0]_mp[1:0] CAD2 CKE DLL Delay and Configuration Cold reset: 0000_0000h. BIOS: D18F2x9C_x0000_0004_dct[1:0]_mp[1:0]. This register controls timing for signals on CKE[3:0] pads. See 2.9.2.1 for pad to pin mapping. 380 42300 Rev 3.12 - July 14, 2015 BKDG for AMD Family 15h Models 10h-1Fh Processors Table 175: Index Mapping for D18F2x9C_x0D0F_C021_dct[1:0]_mp[1:0] Bits Address Bits Valid Values Name D18F2x98_dct[1:0][6] [1:0] Memory Pstate Description 31:8 Reserved. 7 FenceBit: fence bit. See: D18F2x9C_x0D0F_8[1:0]20_dct[1:0]_mp[1:0][FenceBit]. 6 Fence2Bit: second fence bit. See: D18F2x9C_x0D0F_8[1:0]20_dct[1:0]_mp[1:0][Fence2Bit]. 5 Reserved. 4:0 Delay: delay. See: D18F2x9C_x0D0F_8[1:0]20_dct[1:0]_mp[1:0][Delay]. D18F2x9C_x0D0F_E000_dct[1:0]_mp[1:0] Phy Master Configuration Cold reset: 0000_0002h. Bits Description 31:16 Reserved. 15:14 Rate[4:3]. See: Rate[2:0]. 13:4 Reserved. 3 FreqValid: Frequency valid. Read-write. 1=The memory frequency specified with the Rate field is valid. software must set this bit when programming Rate for the setting to take effect. Regardless of this bit, a new rate does not take effect unless the target register context is of the current memory Pstate, or until the memory P-state is changed to the target Rate context. 2:0 Rate[2:0]. Read-write. Rate[4:0] = {Rate[4:3], Rate[2:0]}. Rate[4:0] is the 5-bit DDR data rate that the phy uses to configure the PLL. See Table 132 [Memory Clock Frequency Value Definition]. Software must set FreqValid=1 when programming this field. If memory P-states are enabled then software must program the M1 context. It is not necessary for software to program the M0 context when using D18F2x94_dct[1:0][MemClkFreqVal]. D18F2x9C_x0D0F_E006_dct[1:0] Phy PLL Lock Time Cold reset: 0000_0190h. Bits Description 31:16 Reserved. 15:0 PllLockTime: pll lock time. Read-write. Specifies the number of 5ns periods the phy waits for PLLs to lock during a frequency change. See 2.9.5.4.2 [DRAM Channel Frequency Change]. D18F2x9C_x0D00_E008_dct[1:0] Phy Master Configuration Cold reset: 0000_0013h. Bits Description 31:13 Reserved. 381 42300 Rev 3.12 - July 14, 2015 12 BKDG for AMD Family 15h Models 10h-1Fh Processors PhyPS: Phy memory P-state. Read-only; updated-by-hardware. Indicates the current memory Pstate for the phy. 0=M0, 1=M1. 11:5 Reserved. 4:0 FenceValue: fence value. Read-only; updated-by-hardware. Indicates the fence value used to create the fence bit in the DLL delay registers. See 2.9.5.4.3 [Phy Fence Programming]. D18F2x9C_x0D04_E008_dct[1:0] Phy Master Configuration Cold reset: 0000_0013h. Bits Description 31:9 Reserved. 8 7:0 PStateToAccess: P-state to access. Read-write. Specifies the memory P-state context for context sensitive DDR phy CSR accesses using D18F2x98_dct[1:0] when DctOffset[29:20]!=0D0h. 0=M0, 1=M1. The processor updates this field during a memory P-state change with the target state. This field is only valid in the master channel. See D18F2x9C_x0D0F_E018_dct[1:0][PhyPSMasterChannel]. See D18F2x94_dct[1:0][DphyMemPsSelEn]. Reserved. D18F2x9C_x0D0F_E00A_dct[1:0] Phy Dynamic Power Mode Cold reset: 0000_0000h. Bits Description 31:5 Reserved. 4 3:0 SkewMemClk: skew MEMCLK. Read-write. BIOS: IF ((REG==D18F2x9C_x0D0F_E00A_dct[0]) & ~D18F2x94_dct[0][DisDramInterface] & ~D18F2x94_dct[1][DisDramInterface]) THEN 1 ELSE 0 ENDIF. 1=Skew MEMCLK signals with respect to other channel. SkewMemClk must be 0 if D18F2x94_dct[1:0][DisDramInterface]=1 for either channel. This bit must be set prior to setting D18F2x94_dct[1:0][MemClkFreqVal] during DRAM initialization. See 2.9.5.4.2 [DRAM Channel Frequency Change]. Reserved. D18F2x9C_x0D0F_E010_dct[1:0] DLL Gain Sequencer Timer 0 Cold reset: 0000_00C0h. The gain sequencer allows multiple gain settings during reset/relock of DLLs. Nominally, the DLLs start with the highest gain and remains in that mode for EnDllGain0 amount of time. The sequencer then selects the next highest gain and so on until all gain settings are exhausted, remaining in a low gain tracking mode. A thresholding register allows further tailoring for frequency-independent setup of the sequencer register. . 382 42300 Rev 3.12 - July 14, 2015 Bits BKDG for AMD Family 15h Models 10h-1Fh Processors Description 31:15 Reserved. 14:10 EnDllGain0[4:0]: Enable Dll gain G=0. Read-write. BIOS: 2.9.5.4. Specifies the number of PCLK cycles times 16 that the Dll gain is configured to a value of “G” during a relock event. See D18F2x9C_x0D0F_0[F,7:0]30_dct[1:0][DllGain] for a definition of gain values. 9:5 EnDllGain1[4:0]: Enable Dll gain G=1. Read-write. BIOS: 2.9.5.4. See EnDllGain0. 4:0 EnDllGain2[4:0]: Enable Dll gain G=2. Read-write. BIOS: 2.9.5.4. See EnDllGain0. D18F2x9C_x0D0F_E011_dct[1:0] DLL Gain Sequencer Timer 1 Cold reset: 0000_0020h. Bits Description 31:15 Reserved. 14:10 EnDllGain3[4:0]: Enable Dll gain G=3. Read-write. BIOS: 2.9.5.4. See D18F2x9C_x0D0F_E010_dct[1:0][EnDllGain0]. 9:5 EnDllGain4[4:0]: Enable Dll gain G=4. Read-write. BIOS: 2.9.5.4. See D18F2x9C_x0D0F_E010_dct[1:0][EnDllGain0]. 4:0 EnDllGain5[4:0]: Enable Dll gain G=5. Read-write. BIOS: 2.9.5.4. See D18F2x9C_x0D0F_E010_dct[1:0][EnDllGain0]. D18F2x9C_x0D0F_E012_dct[1:0] DLL Gain Frequency Threshold Cold reset: 0000_33FFh. Bits Description 31:15 Reserved. 14:10 DllGainFreqThresholdG0[4:0]: Dll gain G=0 frequency threshold. Read-write. BIOS: 2.9.5.4. Specifies a minimum FreqCode[4:0] threshold to skip gain = G in the relock sequence. 9:5 DllGainFreqThresholdG1[4:0]: Dll gain G=1 frequency threshold. Read-write. BIOS: 2.9.5.4. See DllGainFreqThresholdG0. 4:0 DllGainFreqThresholdG2[4:0]: Dll gain G=2 frequency threshold. Read-write. BIOS: 2.9.5.4. See DllGainFreqThresholdG0. D18F2x9C_x0D0F_E013_dct[1:0] Phy PLL Regulator Wait Time Cold reset: 0000_00D8h. 383 42300 Rev 3.12 - July 14, 2015 Bits BKDG for AMD Family 15h Models 10h-1Fh Processors Description 31:16 Reserved. 15:0 PllRegWaitTime: PLL regulator wait time. Read-write. BIOS: 2.9.5.4. 1=Specifies the number of 5 ns periods the phy waits for the PLL to become stable when coming out of PLL regulator off power down mode. D18F2x9C_x0D0F_E018_dct[1:0] Fence Cold reset: 0000_0013h. BIOS: 2.9.5.4.3. Bits Description 31:9 Reserved. 8 PhyPSMasterChannel: phy memory P-state master channel. Read-write. 0=Channel 0 updates the PLL during a memory P-state change. 1=Channel 1 updates the PLL during a memory P-state change. BIOS should set this bit only if channel 0 has no memory and is disabled. This field is valid for D18F2x9C_x0D0F_E018_dct[0] only. 7:5 Reserved. 4:0 Fence: fence value. Read-write; updated-by-hardware. This field specifies the fence delay threshold value used for command, address, write data and write DQS. D18F2x9C_x0D0F_E019_dct[1:0] Fence2 Cold reset: 0000_0000h. BIOS: 2.9.5.4.3. Bits Description 31:15 Reserved. 14 Fence2EnableRxDll: phy fence2 enable receive DLL. Read-write. 1=Enable the use of Fence2ThresholdRxDll for DQS receiver enable second fence threshold. 0=Disable second fence. See D18F2x9C_x0000_000C_dct[1:0][FenceThresholdRxDll] for first fence. 13:10 Fence2ThresholdRxDll: phy fence2 threshold DQS receiver enable. Read-write. If Fence2EnableRxDll=1, this field specifies the fence delay threshold value used for DQS receiver enable. See Fence2ThresholdTxPad. 9:5 4 3:0 Reserved. Fence2EnableTxPad: fence2 enable transmit pad. Read-write. 1=Enable the use of Fence2ThresholdTxPad for transmit pad second fence threshold. 0=Disable second fence. See D18F2x9C_x0000_000C_dct[1:0][FenceThresholdTxPad] for first fence. Fence2ThresholdTxPad: phy fence2 threshold transmit pad. Read-write. If Fence2EnableTxPad=1, this field specifies the second fence delay threshold value used for command, address, write data and write DQS. Bits Definition 0h 16/64 MEMCLK delay 1h 17/64 MEMCLK delay Eh-2h {1,<Fence2ThresholdTxPad>}/64 MEMCLK delay Fh 31/64 MEMCLK delay 384 42300 Rev 3.12 - July 14, 2015 BKDG for AMD Family 15h Models 10h-1Fh Processors D18F2xA4 DRAM Controller Temperature Throttle See 2.9.1 [DCT Configuration Registers] and 2.9.9 [DRAM On DIMM Thermal Management and Power Capping]. Bits Description 31:24 Reserved. 23:20 BwCapCmdThrottleMode: bandwidth capping command throttle mode. Read-write. Reset: 0. Specifies the command throttle mode when BwCapEn=1. The DCT throttles commands over a rolling window of 100 clock cycles, maintaining the average throttling as specified by this field. Description Bits 0000b Command throttling is disabled 0001b Throttle commands by 30% 0010b Throttle commands by 40% 0011b Throttle commands by 50% 0100b Throttle commands by 55% 0101b Throttle commands by 60% 0110b Throttle commands by 65% 0111b Throttle commands by 70% 1000b Throttle commands by 75% 1001b Throttle commands by 80% 1010b Throttle commands by 85% 1011b Throttle commands by 90% 1100b Throttle commands by 95% 1101b Reserved 1110b Throttle commands as specified by CmdThrottleMode 1111b Reserved Throttling should not be enabled until after DRAM initialization (D18F2x110[DramEnable]=1) and training (see 2.9.5.9 [DRAM Training]) are complete. 19:15 Reserved. 14:12 CmdThrottleMode: command throttle mode. Read-write. Reset: 0. BIOS: See 2.9.5.7. Specifies the command throttle mode when ODTSEn=1 and the EVENT_L pin is asserted. The DCT throttles commands over a rolling window of 100 clock cycles, maintaining the average throttling as specified by this field. Description Bits 000b Command throttling is disabled. 001b Throttle commands by 30%. 010b Throttle commands by 50%. 011b Throttle commands by 60%. 100b Throttle commands by 70%. 101b Throttle commands by 80%. 110b Throttle commands by 90%. 111b Place the DRAM devices in powerdown mode (see D18F2x94_dct[1:0][PowerDownMode]) when EVENT_L is asserted. This mode is not valid if D18F2x94_dct[1:0][PowerDownEn]=0. Throttling should not be enabled until after DRAM initialization (D18F2x110[DramEnable]=1) and training (see 2.9.5.9 [DRAM Training]) are complete. See also BwCapEn. 385 42300 Rev 3.12 - July 14, 2015 11 BKDG for AMD Family 15h Models 10h-1Fh Processors BwCapEn: bandwidth capping enable. Read-write. Reset: 0. 1=The memory command throttle mode specified by BwCapCmdThrottleMode is applied. This bit is used by software to enable command throttling independent of the state of the EVENT_L pin. If this bit is set, ODTSEn=1, and the EVENT_L pin is asserted, the larger of the two throttle percentages specified by CmdThrottleMode and BwCapCmdThrottleMode is used. 10:9 Reserved. 8 7:0 ODTSEn: on DIMM temperature sensor enable. Read-write. Reset: 0. BIOS: See 2.9.5.7. Enables the monitoring of the EVENT_L pin and indicates whether the on DIMM temperature sensors of the DIMMs on a channel are enabled. 0=Disabled. 1=Enabled. While the EVENT_L pin is asserted, the controller (a) doubles the refresh rate (if Tref=7.8 us), and (b) throttles the address bus utilization as specified by CmdThrottleMode[2:0]. Reserved. D18F2xA8_dct[1:0] DRAM Controller Miscellaneous 2 See 2.9.1 [DCT Configuration Registers]. Bits Description 31 PerRankTimingEn: per rank timing enable. Read-write. Reset: 0. BIOS: 1. Specifies the mapping between chip selects and a set of programmable timing delays. 1=Each chip select is controlled by a set of timing delays. 0=Each chip select pair is controlled by a set timing delays. See D18F2x9C_x0000_0[3:0]0[2:1]_dct[1:0]_mp[1:0], D18F2x9C_x0000_0[3:0]0[6:5]_dct[1:0]_mp[1:0], D18F2x9C_x0000_00[2A:10]_dct[1:0]_mp[1:0], and D18F2x9C_x0000_00[4A:30]_dct[1:0]_mp[1:0]. 30 Reserved. 29 RefChCmdMgtDis: refresh channel command management disable. Read-write. Reset: 0. 1=DCTs issue refresh commands independently. 0=DCTs stagger the issue of refresh commands. This bit must be set the same in all DCTs. 28 FastSelfRefEntryDis: fast self refresh entry disable. Read-write. Reset: 1. BIOS: ~D18F2x1B4[FlushWrOnS3StpGnt]. 1=DCT pushes outstanding transactions to memory prior to entering self refresh. 0=DCT enters self refresh immediately unless instructed to push outstanding transactions to memory by D18F2x11C[FlushWrOnStpGnt] or D18F2x1B4[FlushWrOnS3StpGnt]. This bit must be programmed to the same value on all enabled DCTs. 27 CsMux67: chip select mux 6 and 7. See: CsMux45. See also CSTimingMux67. 26 CsMux45: chip select mux 4 and 5. Read-Write. Reset: 0. Specifies the DCT mode used to drive the associated chip selects. 0=Chip select mode. 1=Extended address mode; A[17:16] are driven on the DRAM CS pins. 25:24 Reserved. 23 Reserved. 22 PrtlChPDEnhEn: partial channel power down enh enable. Read-write. Reset: 0. BIOS: 0. Selects the channel idle hysteresis for fast exit/slow exit mode changes when (D18F2x94_dct[1:0][PowerDownMode] & D18F2x84_dct[1:0][PchgPDModeSel]). 1=Hysteresis specified by D18F2x244_dct[1:0][PrtlChPDDynDly]. 0=256 clock hysteresis. 21 AggrPDEn: aggressive power down enable. Read-write. Reset: 0. BIOS: 1. 1=The DCT places the DRAM devices in precharge power down mode when pages are closed as specified by D18F2x248_dct[1:0]_mp[1:0][AggrPDDelay]. 0=The DCT places the DRAM devices in precharge power down mode when pages are closed as specified by D18F2x90_dct[1:0][DynPageCloseEn]. 386 42300 Rev 3.12 - July 14, 2015 20 BKDG for AMD Family 15h Models 10h-1Fh Processors BankSwap: swap bank address. Read-write. Reset: 0. BIOS: See 2.9.5.7. 1=Swap the DRAM bank address bits. IF (D18F2x110[DctSelIntLvEn]==1 && D18F2x110[DctSelIntLvAddr]==100b) THEN normalized address bits (7+n):8 are swapped with normalized address bits used for bank address (See D18F2x80_dct[1:0]) ELSE normalized address bits (8+n):9 are swapped with normalized address bits used for bank address; n is the number of bank address bits for the chip select. For example, if D18F2x110[DctSelIntLvAddr]==101b and D18F2x80_dct[1:0][DimmAddrMap0]==0001b, then normalized address bits 11:9 are swapped with normalized address bits 15:13, when sent to chip select 0. This swap happens before D18F2x94_dct[1:0][BankSwizzleMode] is applied. 19:18 Reserved. 17:16 MemPhyPllPdMode: memory phy PLL powerdown mode. Read-write. Reset: 00b. BIOS: 10b. These bits control how the memory PLL powers down during self-refresh. If self-refresh is requested for an NB p-state change, then the memory PLL is not powered down. The value of this field should be programmed the same for both DCTs. Memory phy PLL powerdown can only be enabled if NB PLL power down is enabled. See D18F5x128[NbPllPwrDwnRegEn]. Bits Description 00b PLL powerdown is disabled 01b PLL VCO powerdown during SR 10b PLL regulator powerdown during SR 11b PLL VCO and regulator powerdown during SR 15:8 CtrlWordCS[7:0]: control word chip select. Read-write. Reset: 0. Specifies the target DIMM chip selects used for control word programming; Used in conjunction with D18F2x7C_dct[1:0][SendControlWord]. Description Bits 02h-00h Reserved 03h CS0,CS1 is asserted 0Bh-04h Reserved 0Ch CS2,CS3 is asserted 2Fh-0Dh Reserved 30h CS4,CS5 is asserted BFh-31h Reserved C0h CS6,CS7 is asserted FFh-C1h Reserved 7:6 Reserved. 5 SubMemclkRegDly: Sub-one MEMCLK register delay. Read-write. Reset: 0. 1=The delay through the DIMM register and routing delay is less than 1 MEMCLK. The dram controller does not add 1 MEMCLK to calculate write latency for DDR3 registered DIMMs. 0=The delay through the DIMM register and routing delay is at least 1 MEMCLK. The DRAM controller adds 1 MEMCLK to calculate write latency for DDR3 registered DIMMs. 4 Reserved. 3 Reserved. 2 CSTimingMux67: CS timing mux 6 and 7. Read-write. Reset: 0. BIOS: 2.9.5.4.3. Selects the register used for the setup time of the CS[7:6] pins versus MEMCLK. 0=D18F2x9C_x0000_0004_dct[1:0]_mp[1:0][CsOdtSetup]. 1=D18F2x9C_x0000_0004_dct[1:0]_mp[1:0][AddrCmdSetup]. 1:0 Reserved. 387 42300 Rev 3.12 - July 14, 2015 BKDG for AMD Family 15h Models 10h-1Fh Processors D18F2xAC DRAM Controller Temperature Status Cold reset: 0000_0000h. Bits Description 31:3 Reserved. 2 MemTempHot1: Memory temperature hot, DCT1. See: MemTempHot0. 1 Reserved. 0 MemTempHot0: Memory temperature hot, DCT0. Read; Write-1-to-clear. 1=The EVENT_L pin was asserted indicating the memory temperature exceeded the normal operating limit; the DCT may be throttling the interface to aid in cooling. See D18F2xA4. D18F2xF8 P-state Power Information 1 Read-only. Bits Description 31:24 PwrValue3: P3 power value. See PwrValue0. Value: Product-specific. 23:16 PwrValue2: P2 power value. See PwrValue0. Value: Product-specific. 15:8 PwrValue1: P1 power value. See PwrValue0. Value: Product-specific. 7:0 PwrValue0: P0 power value. PwrValue and PwrDiv together specify the expected power draw of a single core in P0 and 1/NumCores of the Northbridge in the NB P-state as specified by MSRC001_00[6B:64][NbPstate]. NumCores is defined to be the number of cores per node at cold reset. Value: Product-specific. PwrDiv Description 00b PwrValue / 1 W, Range: 0 to 255 W 01b PwrValue / 10 W, Range: 0 to 25.5 W 10b PwrValue / 100 W, Range: 0 to 2.55 W 11b Reserved D18F2x104 P-state Power Information 3 Read-only. Bits Description 31:24 Reserved. 23:16 PwrValue7: P7 power value. See D18F2xF8[PwrValue0]. Value: Product-specific. 15:8 PwrValue6: P6 power value. See D18F2xF8[PwrValue0]. Value: Product-specific. 7:0 PwrValue5: P5 power value. See D18F2xF8[PwrValue0]. Value: Product-specific. D18F2x10C Swap Interleaved Region Base/Limit IF (D18F2x118[LockDramCfg]) THEN Read-only. ELSE Read-write. ENDIF. Reset: 0000_0000h. Enables swapping a region below 16G with the same sized region located at the bottom of memory. This register is typically used to map addresses of a graphics frame buffer located below the sub-4GB IO hole to interleaved DRAM in low memory, and is only necessary if the frame buffer normally maps to non-interleaved memory. • The size of the swapped high region must be a integer multiple of 128M, defined to be {IntLvRgnBaseAddr, 388 42300 Rev 3.12 - July 14, 2015 • • • • • • BKDG for AMD Family 15h Models 10h-1Fh Processors 000b, 000000h} to {IntLvRgnLmtAddr, 111b, FFFFFFh}. The size of the swapped region must be less than or equal to the alignment of IntLvRgnBaseAddr. • E.g. If IntLvRgnBaseAddr=2h then size <=256MB and BIOS programs IntLvRgnLmtAddr <= IntLvRgnBaseAddr + 1h. • It is expected that BIOS may program IntLvRgnBaseAddr to a value less than the base address of the graphics frame buffer if realignment is necessary to achieve a larger swap size. The location of the low region is defined to be 0000_0000h to {IntLvRgnLmtAddr - IntLvRgnBaseAddr, 111b, FFFFFFh}. The swapped region must be all DRAM. I.e. No IO hole. Channel interleaving must be enabled and the DCTs must be of unequal size. Swapping must not be enabled on more than one node, and D18F1x[144:140,44:40][DramBase] must be zero. See D18F2x110[DctSelIntLvEn]. See 2.9.7 [Memory Hoisting]. Bits Description 31:27 Reserved. 26:20 IntLvRgnSize[33:27]: Interleave swap region size bits[33:27]. Interleave swap region size [33:27]. 19:18 Reserved. 17:11 IntLvRgnLmtAddr[33:27]: Interleave swap region limit address bits[33:27]. Interleave swap region limit address [33:27]. 10 Reserved. 9:3 IntLvRgnBaseAddr[33:27]: Interleave swap region base address bits[33:27]. Interleave swap region base address [33:27]. 2:1 Reserved. 0 IntLvRgnSwapEn: Interleave region swap enable. 1=Enables swapping a region from the top of memory to the bottom of DRAM space. D18F2x110 DRAM Controller Select Low Reset: 0000_0000h. Bits Description 31:11 DctSelBaseAddr[47:27]: DRAM controller select base address bits[47:27]. IF (D18F2x118[LockDramCfg]) THEN Read-only. ELSE Read-write. ENDIF. Delineates the address range of the two DCTs by specifying the base address of the upper address range. See DctSelHiRngEn. 10 MemCleared: memory cleared. Read-only. 1=Memory has been cleared since the last warm reset. This bit is set by MemClrInit. See MemClrInit. 9 MemClrBusy: memory clear busy. Read-only. 1=The memory clear operation in either of the DCTs is in progress. Reads or writes to DRAM while the memory clear operation is in progress result in undefined behavior. 8 DramEnable: DRAM enabled. Read-only. 1=All of the used DCTs are initialized (see 2.9.5.8 [DRAM Device and Controller Initialization]) or have exited from self refresh (D18F2x90_dct[1:0][ExitSelfRef] transitions from 1 to 0). 389 42300 Rev 3.12 - July 14, 2015 BKDG for AMD Family 15h Models 10h-1Fh Processors 7:6 DctSelIntLvAddr[1:0]: DRAM controller select channel interleave address bit. IF (D18F2x118[LockDramCfg]) THEN Read-only. ELSE Read-write. ENDIF. IF (DctSelIntLvEn) THEN BIOS: 00b. ENDIF. Specifies how interleaving is selected between the DCTs. In all cases, if the select function is low then DCT0 is selected; if the select function is high then DCT1 is selected. DctSelIntLvAddr[2:0] = {D18F2x114[DctSelIntLvAddr[2], D18F2x110[DctSelIntLvAddr[1:0]]}. The select functions are defined as follows: DctSelIntLvAddr DCT Select Function 000b Address bit 6. 001b Address bit 12. 010b Hash: exclusive OR of address bits[20:16, 6]. 011b Hash: exclusive OR of address bits[20:16, 9]. 100b Address bit 8. 101b Address bit 9. 110b Reserved. 111b Reserved. If the internal GPU is enabled, only encodings 001b, 100b, and 101b are supported. 5 DctDatIntLv: DRAM controller data interleave enable. IF (D18F2x118[LockDramCfg]) THEN Read-only. ELSE Read-write. ENDIF. BIOS: Set if ECC is enabled. 1=DRAM data bits from every two consecutive 64-bit DRAM lines are interleaved in the ECC calculation such that a dead bit of a DRAM device is correctable. 4 Reserved. 3 MemClrInit: memory clear initialization. IF (D18F2x118[LockDramCfg]) THEN Read-only. ELSE Read-write. ENDIF. 1=The node writes 0’s to all locations of system memory attached to the node and sets the MemCleared bit. The status of the memory clear operation can be determined by reading the MemClrBusy and MemCleared bits. This command is ignored if MemClrBusy=1 when the command is received. DramEnable must be set before setting MemClrInit. The memory prefetcher must be disabled by setting D18F2x11C[PrefIoDis] and D18F2x11C[PrefCpuDis] before memory clear initialization and then can be re-enabled when MemCleared=1. 2 DctSelIntLvEn: DRAM controller interleave enable. IF (D18F2x118[LockDramCfg]) THEN Read-only. ELSE Read-write. ENDIF. BIOS: See 2.9.5.7. 1=Channel interleave is enabled; DctSelIntLvAddr specifies which address bit is used to select between DCT0 and DCT1; this applies from the base system memory address of the node (specified by D18F1x[144:140,44:40] [DRAM Base/Limit]) to DctSelBaseAddr (if enabled). If the amount of memory connected to each of the DCTs is different, then channel interleaving may be supported across the address range that includes both DCTs, the top of which is specified by DctSelBaseAddr; the remainder of the address space, above DctSelBaseAddr, would then be allocated to only the DCT connected to the larger amount of memory, specified by DctSelHi. 1 DctSelHi: DRAM controller high select. IF (D18F2x118[LockDramCfg]) THEN Read-only. ELSE Read-write. ENDIF. If DctSelHiRngEn is set, this specifies which DCT receives accesses with addresses in the high range (greater than or equal to DctSelBaseAddr). 0=High addresses go to DCT0. 1=High addresses go to DCT1. 0 DctSelHiRngEn: DRAM controller select high range enable. IF (D18F2x118[LockDramCfg]) THEN Read-only. ELSE Read-write. ENDIF. 1=Enables addresses greater than or equal to DctSelBaseAddr[47:27] to be used to select between DCT0 and DCT1; DctSelHi specifies which DCT occupies the high range. 390 42300 Rev 3.12 - July 14, 2015 BKDG for AMD Family 15h Models 10h-1Fh Processors D18F2x114 DRAM Controller Select High IF (D18F2x118[LockDramCfg]) THEN Read-only. ELSE Read-write. ENDIF. Reset: 0000_0000h. Bits Description 31:10 DctSelBaseOffset[47:26]: DRAM controller select base offset address bits[47:26]. When D18F2x110[DctSelHiRngEn]=1, this value is subtracted from the physical address of certain transactions before being passed to the DCT. See 2.9.7.2 [DctSelBaseOffset Programming]. 9 8:0 DctSelIntLvAddr[2]: DRAM controller select channel interleave address bit. IF (D18F2x110[DctSelIntLvEn]) THEN BIOS: 1. ENDIF. See D18F2x110[DctSelIntLvAddr[1:0]]. Reserved. D18F2x118 Memory Controller Configuration Low Fields in this register (bits[17:0]) indicate priority of request types. Variable priority requests enter the memory controller as medium priority and are promoted to high priority if they have not been serviced in the time specified by MctVarPriCntLmt. This feature may be useful for isochronous IO traffic. If isochronous traffic is specified to be high priority, it may have an adverse effect on the bandwidth and performance of the devices associated with the other types of traffic. However, if isochronous traffic is specified as medium priority, the processor may not meet the isochronous bandwidth and latency requirements. The variable priority allows the memory controller to optimize DRAM transactions until isochronous traffic reaches a time threshold and must be serviced more quickly. Bits Description 31:28 MctVarPriCntLmt: variable priority time limit. Read-write. Reset: 0000b. BIOS: 0000b. Description Bits Description Bits 0000b 80ns 1000b 720ns 0001b 160ns 1001b 800ns 0010b 240ns 1010b 880ns 0011b 320ns 1011b 960ns 0100b 400ns 1100b 1040ns 0101b 480ns 1101b 1120ns 0110b 560ns 1110b 1200ns 0111b 640ns 1111b 1280ns 27 MctEccDisLatOptEn: memory controller ecc disabled latency optimization. Read-write. Reset: 0. BIOS: 1. 1=Enable bypass of memory ECC checking logic. BIOS must program this bit before any memory accesses are issued from the processor. 26:24 McqHiPriByPassMax: memory controller high priority bypass max. Read-write. Reset: 100b. Specifies the number of times a medium- or low-priority DRAM request may be bypassed by highpriority DRAM requests. 23 Reserved. 22:20 McqMedPriByPassMax: memory controller medium bypass low priority max. Read-write. Reset: 100b. Specifies the number of times a low-priority DRAM request may be bypassed by medium-priority DRAM requests. 391 42300 Rev 3.12 - July 14, 2015 BKDG for AMD Family 15h Models 10h-1Fh Processors 19 LockDramCfg. Read; write-1-only. Reset: 0. BIOS: See 2.9.8 [DRAM CC6/PC6 Storage], 2.5.3.2.3.3 [Core C6 (CC6) State]. The following registers are read-only if LockDramCfg=1; otherwise the access type is specified by the register: • D18F1xF0 [DRAM Hole Address] • D18F2x[5C:40]_dct[1:0] [DRAM CS Base Address] • D18F2x[6C:60]_dct[1:0] [DRAM CS Mask] • D18F2x80_dct[1:0] [DRAM Bank Address Mapping] • D18F2x10C [Swap Interleaved Region Base/Limit] • D18F2x110 [DRAM Controller Select Low] • D18F2x114 [DRAM Controller Select High] • D18F4x128[CoreStateSaveDestNode] • D18F1x[144:140,44:40] [DRAM Base/Limit] • D18F1x120 [DRAM Base System Address] • D18F1x124 [DRAM Limit System Address] • D18F2x118[CC6SaveEn] 18 CC6SaveEn. IF (D18F2x118[LockDramCfg]) THEN Read-only. ELSE Read-write. ENDIF. Reset: 0. 1=CC6 save area is enabled. See 2.5.3.2.7 [BIOS Requirements for Initialization]. BIOS: (D18F4x118[PwrGateEnCstAct0] | D18F4x118[PwrGateEnCstAct1] | D18F4x11C[PwrGateEnCstAct2]). 17:16 MctPriScrub: scrubber priority. Read-write. Reset: 00b. Bits Description 00b Medium 01b Reserved 10b High 11b Variable 15:14 MctPriTrace: trace-mode request priority. See: MctPriCpuRd. Read-write. Reset: 10b. This must be set to high. 13:12 MctPriIsoc: display refresh read priority. See: MctPriCpuRd. Read-write. Reset: 10b. 11:10 MctPriWr: default write priority. See: MctPriCpuRd. Read-write. Reset: 01b. 9:8 MctPriDefault: default non-write priority. See: MctPriCpuRd. Read-write. Reset: 00b. 7:6 MctPriIsocWr: IO write with the isoch bit set priority. See: MctPriCpuRd. Read-write. Reset: 00b. This does not apply to isochronous traffic that is classified as display refresh. 5:4 MctPriIsocRd: IO read with the isoch bit set priority. See: MctPriCpuRd. Read-write. Reset: 10b. This does not apply to isochronous traffic that is classified as display refresh. 3:2 MctPriCpuWr: CPU write priority. See: MctPriCpuRd. Read-write. Reset: 01b. 1:0 MctPriCpuRd: CPU read priority. Read-write. Reset: 00b. Bits Description 00b Medium 01b Low 10b High 11b Variable D18F2x11C Memory Controller Configuration High The two main functions of this register are to control write bursting and memory prefetching. 392 42300 Rev 3.12 - July 14, 2015 BKDG for AMD Family 15h Models 10h-1Fh Processors Write bursting. DctWrLimit and MctWrLimit specify how writes may be burst from the MCT into the DCT to improve DRAM efficiency. When the number of writes in the MCT reaches the value specified in MctWrLimit, then they are all burst to the DCTs at once. Prior to reaching the watermark, a limited number of writes can be passed to the DCTs (specified by DctWrLimit), tagged as low priority, for the DCTs to complete when otherwise idle. Rules regarding write bursting: • Write bursting mode only applies to low-priority writes. Medium and high priority writes are not withheld from the DCTs for write bursting. • If write bursting is enabled, writes stay in the MCQ until the threshold specified by MctWrLimit is reached. • Once the threshold is reached, all writes in MCQ are converted to medium priority. • Any write in MCQ that matches the address of a subsequent access is promoted to either medium priority or the priority of the subsequent access, whichever is higher. • DctWrLimit only applies to low-priority writes. Memory prefetching. The MCT prefetcher detects stride patterns in the stream of requests and then, for predictable stride patterns, generates prefetch requests. A stride pattern is a pattern of requests through system memory that are the same number of cachelines apart. The prefetcher supports strides of -4 to +4 cachelines, which can include alternating patterns (e.g. +1, +2, +1, +2), and can prefetch 1, 2, 3, 4, or 5 cachelines ahead, depending on the confidence. In addition, a fixed stride mode (non-alternating) may be used for IO requests which often have fixed stride patterns. This mode bypasses the stride predictor such that CPU-access stride predictions are not adversely affected by IO streams. The MCT tracks several stride patterns simultaneously. Each of these has a confidence level associated with it that varies as follows: • Each time a request is received that matches the stride pattern, the confidence level increases by one. • Each time a request is received within +/- 4 cachelines of the last requested cacheline in the pattern that does not match the pattern, then the confidence level decreases by one. • When the confidence level reaches the saturation point specified by PrefConfSat, then it no-longer increments. Each request that is not within +/- 4 cachelines of the last requested cacheline line of all the stride patterns tracked initiates a new stride pattern by displacing one of the existing least-recently-used stride patterns. The memory prefetcher uses an adaptive prefetch scheme to adjust the prefetch distance based upon the buffer space available for prefetch request data. The adaptive scheme counts the total number of prefetch requests and the number of prefetch requests that cannot return data because of buffer availability. After every 16 prefetch requests, the prefetcher uses the following rules to adjust the prefetch distance: • If the ratio of prefetch requests that cannot return data to total prefetch requests is greater than or equal to D18F2x1B0[AdapPrefMissRatio] then the prefetch distance is reduced by D18F2x1B0[AdapPrefNegativeStep]. • If the ratio of prefetch requests that cannot return data to total prefetch requests is less than D18F2x1B0[AdapPrefMissRatio] then the prefetch distance is increased by D18F2x1B0[AdapPrefPositiveStep]. • If the adjusted prefetch distance is greater than the prefetch distance defined for the current confidence level, the prefetch distance for the current confidence level is used. The adaptive prefetch scheme supports fractional prefetch distances by alternating between two whole number prefetch distances. For example a prefetch distance of 1.25 causes a prefetch distance sequence of: 1, 1, 1, 2, 1, 1, 1, 2. Bits 31 Description Reserved. 393 42300 Rev 3.12 - July 14, 2015 BKDG for AMD Family 15h Models 10h-1Fh Processors 30 FlushWr: flush writes command. Read; write-1-only; cleared-by-hardware. Reset: 0. Setting this bit causes write bursting to be canceled and all outstanding writes to be flushed to DRAM. This bit is cleared when all writes are flushed to DRAM. 29 FlushWrOnStpGnt: flush writes on stop-grant. Read-write. Reset: 0. BIOS: ~D18F2x1B4[FlushWrOnS3StpGnt]. 1=Causes write bursting to be canceled and all outstanding writes to be flushed to DRAM when in the stop-grant state. 28 PrefDramTrainMode: prefetch DRAM training mode. Read-write; cleared-by-hardware. Reset: 0. 1=Enable DRAM training mode. Hardware clears this bit when the prefetch request limit is reached. Writing a zero to this bit clears the prefetch buffer and disables the DRAM training mode prefetcher. BIOS must write a zero to this bit after training is complete. See 2.9.5.9.6 [Continuous Pattern Generation]. 27:25 PrefThreeConf: prefetch three-ahead confidence. Read-write. Reset: 100b. BIOS: 110b. Confidence level required in order to prefetch three cachelines ahead (same encoding as PrefTwoConf below). 24:22 PrefTwoConf: prefetch two-ahead confidence. Read-write. Reset: 011b. BIOS: 011b. Confidence level required in order to prefetch two cachelines ahead. Bits Description 000b 0 110b-001b <PrefTwoConf*2> 111b 14 21:20 PrefOneConf: prefetch one-ahead confidence. Read-write. Reset: 10b. BIOS: 10b. Confidence level required in order to prefetch one ahead (0 through 3). 19:18 PrefConfSat: prefetch confidence saturation. Read-write. Reset: 00b. BIOS: 00b. Specifies the point at which prefetch confidence level saturates and stops incrementing. Bits Description 00b 15 01b 7 10b 3 11b Reserved 17:16 PrefFixDist: prefetch fixed stride distance. Read-write. Reset: 00b. Specifies the distance to prefetch ahead if in fixed stride mode. 00b=1 cacheline; 01b=2 cachelines; 10b=3 cachelines; 11b=4 cachelines. 15 PrefFixStrideEn: prefetch fixed stride enable. Read-write. Reset: 0. 1=The prefetch stride for all requests (CPU and IO) is fixed (non-alternating). 14 PrefIoFixStrideEn: Prefetch IO fixed stride enable. Read-write. Reset: 0. 1=The prefetch stride for IO requests is fixed (non-alternating). 13 PrefIoDis: prefetch IO-access disable. Read-write. Reset: 1. BIOS: 0. 1=Disables IO requests from triggering prefetch requests. 12 PrefCpuDis: prefetch CPU-access disable. Read-write. Reset: 1. BIOS: 0. 1=Disables CPU requests from triggering prefetch requests. 11:7 MctPrefReqLimit: memory controller prefetch request limit. Read-write. Reset: 1Eh. Specifies the maximum number of outstanding prefetch requests allowed. See D18F3x78 for restrictions on this field. 394 42300 Rev 3.12 - July 14, 2015 BKDG for AMD Family 15h Models 10h-1Fh Processors 6:2 MctWrLimit: memory controller write-burst limit. Read-write. Reset: 1Fh. BIOS: 0Ch. Specifies the number of writes in the memory controller queue before they are burst into the DCTs. Description Bits 00h 16 0Fh-01h <16-MctWrLimit> 1Fh-10h Write bursting disabled 1:0 DctWrLimit: DRAM controller write limit. Read-write. Reset: 00b. BIOS: 01b. Specifies the maximum number of writes allowed in the DCT queue when write bursting is enabled, prior to when the number of writes in MCQ exceeds the watermark specified by MctWrLimit. Bits Description 00b 0 01b 2 10b 4 11b Reserved D18F2x1B0 Extended Memory Controller Configuration Low The main function of this register is to control the memory prefetcher. See D18F2x11C [Memory Controller Configuration High] about the adaptive prefetch scheme. Table 176: D18F2x1B[4:0] Recommended Settings Condition DdrRate 667 800 1066 1333 1600 1866 Bits D18F2x1B0 DcqBwThrotWm 0h D18F2x1B4 DcqBwThrotWm1 DcqBwThrotWm2 3h 4h 3h 5h 4h 6h 5h 8h 6h 9h 7h Ah Description 31:28 DcqBwThrotWm: dcq bandwidth throttle watermark. Read-write. Reset: 3h. BIOS: Table 176. Specifies the number of outstanding DRAM read requests before new DRAM prefetch requests and speculative prefetch requests are throttled. 0h=Throttling is determined by D18F2x1B4[DcqBwThrotWm1, DcqBwThrotWm2]. Legal values are 0h through Ch. 27:25 PrefFiveConf: prefetch five-ahead confidence. Read-write. Reset: 110b. BIOS: 111b. Confidence level required in order to prefetch five cachelines ahead. Description Bits 000b 0 110b-001b <PrefFiveConf*2> 111b 14 395 42300 Rev 3.12 - July 14, 2015 BKDG for AMD Family 15h Models 10h-1Fh Processors 24:22 PrefFourConf: prefetch four-ahead confidence. Read-write. Reset: 101b. BIOS: 111b. Confidence level required in order to prefetch four cachelines ahead. Bits Description 000b 0 110b-001b <PrefFourConf*2> 111b 14 21 Reserved. 20 DblPrefEn: double prefetch enable. Read-write. Reset: 0. 1=The memory prefetcher only generates prefetch requests when it is able to generate a pair of prefetch requests to consecutive cache lines. BIOS must clear D18F2x11C[PrefDramTrainMode] prior to setting this bit. 19:18 Reserved. Reset: 00b 17:13 Reserved. Reset: 11100b. 12 EnSplitDctLimits: split DCT write limits enable. Read-write. Reset: 0. BIOS: 1. 1=The number of writes specified by D18F2x11C[DctWrLmt and MctWrLmt] is per DCT. 0=The number of writes specified by D18F2x11C[DctWrLmt and MctWrLmt] is total writes independent of DCT. Setting this bit also affects the encoding of D18F2x11C[DctWrLmt]. 11 DisIoCohPref: disable coherent prefetched for IO. Read-write. Reset: 0. 1=Probes are not generated for prefetches generated for reads from IO devices. 10:8 CohPrefPrbLmt: coherent prefetch probe limit. Read-write. Reset: 000b. BIOS: 000b. Specifies the maximum number of probes that can be outstanding for memory prefetch requests. Bits Description 000b Probing disabled for memory prefetch requests 001b 4 outstanding probes 010b 8 outstanding probes 011b Reserved. 1xxb Reserved 7:6 Reserved. 5:4 AdapPrefNegativeStep: adaptive prefetch negative step. Read-write. Reset: 00b. BIOS: 00b. Specifies the step size that the adaptive prefetch scheme uses when decreasing the prefetch distance. Bits Description 00b 2/16 01b 4/16 10b 8/16 11b 16/16 396 42300 Rev 3.12 - July 14, 2015 BKDG for AMD Family 15h Models 10h-1Fh Processors 3:2 AdapPrefPositiveStep: adaptive prefetch positive step. Read-write. Reset: 00b. BIOS: 00b. Specifies the step size that the adaptive prefetch scheme uses when increasing the prefetch distance. Bits Description 00b 1/16 01b 2/16 10b 4/16 11b 8/16 1:0 AdapPrefMissRatio: adaptive prefetch miss ratio. Read-write. Reset: 00b. BIOS: 01b. Specifies the ratio of prefetch requests that do not have data buffer available to the total number of prefetch requests at which the adaptive prefetch scheme begins decreasing the prefetch distance. Bits Description 00b 1/16 01b 2/16 10b 4/16 11b 8/16 D18F2x1B4 Extended Memory Controller Configuration High Register Bits 31 Description FlushOnMmioWrEn: flush on mmio write enable. Read-write. Reset: 0. 1=Any CPU-sourced MMIO write that matches D18F1x[1CC:180,BC:80] causes the memory controller data buffers to be flushed to memory. 30:28 S3SmafId: S3 SMAF id. Read-write. Reset: 100b. SMAF encoding of D18F3x[84:80] corresponding to the ACPI S3 state when FlushWrOnS3StpGnt=1. Reserved when FlushWrOnS3StpGnt=0. 27 FlushWrOnS3StpGnt: flush write on S3 stop grant. Read-write. Reset: 0. BIOS: 1. 1=Write bursting is canceled and all outstanding writes are flushed to DRAM when in the stop-grant state and the SMAF code is equal to S3SmafId, indicating entry into the ACPI S3 state. See D18F2xA8_dct[1:0][FastSelfRefEntryDis], D18F2x11C[FlushWrOnStpGnt]. 26 EnSplitMctDatBuffers: enable split MCT data buffers. Read-write. Reset: 0. BIOS: 1. 1=Enable resource allocation into the split buffer resources BIOS must program this bit before any memory accesses are issued from the processor. 25:23 Reserved. 22 SpecPrefDisWm1: speculative prefetch disable watermark 1. Read-write. Reset: 0. 0=Disable speculative prefetches at the DcqBwThrotWm2 limit. 1=Disable speculative prefetches at the DcqBwThrotWm1 limit. See also D18F2x1B0[SpecPrefDis]. 21 RegionAlloWm2: region prefetch allocate watermark 2. Read-write. Reset: 0. See DemandAlloWm2. 20 RegionPropWm2: region prefetch propagate watermark 2. Read-write. Reset: 0. See DemandPropWm2. 19 StrideAlloWm2: stride prefetch allocate watermark 2. Read-write. Reset: 1. See DemandAlloWm2. 18 StridePropWm2: stride prefetch propagate watermark 2. Read-write. Reset: 1. See DemandPropWm2. 397 42300 Rev 3.12 - July 14, 2015 BKDG for AMD Family 15h Models 10h-1Fh Processors 17 DemandAlloWm2: demand request allocate watermark 2. Read-write. Reset: 1. Specifies the behavior from the DcqBwThrotWm1 limit to the DcqBwThrotWm2 limit. 0=Requests do not allocate a new entry. 1=Requests allocate a new entry; defined only if (DemandAlloWm1 & DemandPropWm2). 16 DemandPropWm2: demand request propagate watermark 2. Read-write. Reset: 1. Specifies the behavior from the DcqBwThrotWm1 limit to the DcqBwThrotWm2 limit. 0=Requests do not update existing entries. 1=Requests update existing entries; defined only if (DemandPropWm1=1). 15 RegionAlloWm1: region prefetch allocate watermark 1. Read-write. Reset: 0. See DemandAlloWm1. 14 RegionPropWm1: region prefetch propagate watermark 1. Read-write. Reset: 1. See DemandPropWm1. 13 StrideAlloWm1: stride prefetch allocate watermark 1. Read-write. Reset: 1. See DemandAlloWm1. 12 StridePropWm1: stride prefetch propagate watermark 1. Read-write. Reset: 1. See DemandPropWm1. 11 DemandAlloWm1: demand request allocate watermark 1. Read-write. Reset: 1. Specifies the behavior prior to the DcqBwThrotWm1 limit. 0=Requests do not allocate a new entry. 1=Requests allocate a new entry; defined only if (DemandPropWm1=1). 10 DemandPropWm1: demand request propagate watermark 1. Read-write. Reset: 1. Specifies the behavior prior to the DcqBwThrotWm1 limit. 0=Requests do not update existing entries. 1=Requests update existing entries. 9:5 DcqBwThrotWm2: DCQ bandwidth throttle watermark 2. Read-write. Reset: 06h. BIOS: Table 176. Specifies a prefetch throttling watermark based on the number of outstanding DRAM read requests. This field is reserved when D18F2x1B0[DcqBwThrotWm] != 0. When throttling is enabled, if the number of outstanding DRAM read requests exceeds DcqBwThrotWm2 both request allocate and propagate are blocked and new prefetches are disabled. When throttling is enabled, DcqBwThrotWm2 should be programmed to a value greater than DcqBwThrotWm1. 0h=Throttling is disabled. Legal values are 0h through 18h. 4:0 DcqBwThrotWm1: DCQ bandwidth throttle watermark 1. Read-write. Reset: 03h. BIOS: Table 176. Specifies a prefetch throttling watermark based on the number of outstanding DRAM read requests. This field is reserved when D18F2x1B0[DcqBwThrotWm] != 0. 0h=Throttling is disabled. Legal values are 0h through 18h. D18F2x200_dct[1:0]_mp[1:0] DRAM Timing 0 Reset: 0F05_0505h. See 2.9.1 [DCT Configuration Registers]. Bits Description 31:30 Reserved. 29:24 Tras: row active strobe. Read-write. Specifies the minimum time in memory clock cycles from an activate command to a precharge command, both to the same chip select bank. Bits Description 07h-00h Reserved 2Ah-08h <Tras> clocks 3Fh-2Bh Reserved 398 42300 Rev 3.12 - July 14, 2015 BKDG for AMD Family 15h Models 10h-1Fh Processors 23:21 Reserved. 20:16 Trp: row precharge time. Read-write. Specifies the minimum time in memory clock cycles from a precharge command to an activate command or auto refresh command, both to the same bank. Bits Description 04h-00h Reserved 13h-05h <Trp> clocks 1Fh-14h Reserved 15:13 Reserved. 12:8 Trcd: RAS to CAS delay. Read-write. Specifies the time in memory clock cycles from an activate command to a read/write command, both to the same bank. Bits Description 01h-00h Reserved 13h-02h <Trcd> clocks 1Fh-14h Reserved 7:5 Reserved. 4:0 Tcl: CAS latency. Read-write. Specifies the time in memory clock cycles from the CAS assertion for a read cycle until data return (from the perspective of the DRAM devices). Bits Description 04h-00h Reserved 10h-05h <Tcl> clocks 1Fh-11h Reserved D18F2x204_dct[1:0]_mp[1:0] DRAM Timing 1 Reset: 0400_040Bh. See 2.9.1 [DCT Configuration Registers]. Bits Description 31:28 Reserved. 27:24 Trtp: read CAS to precharge time. Read-write. Specifies the earliest time in memory clock cycles a page can be closed after having been read. Satisfying this parameter ensures read data is not lost due to a premature precharge. Description Bits 3h-0h Reserved Bh-4h <Trtp> clocks Fh-Ch Reserved 23:22 Reserved. 21:16 FourActWindow: four bank activate window. Read-write. Specifies the rolling tFAW window in memory clock cycles during which no more than 4 banks in an 8-bank device are activated. Bits Description 00h No tFAW window restriction 05h-01h Reserved 2Ch-06h <FourActWindow> clocks 3Fh-2Dh Reserved 15:12 Reserved. 399 42300 Rev 3.12 - July 14, 2015 BKDG for AMD Family 15h Models 10h-1Fh Processors 11:8 Trrd: row to row delay (or RAS to RAS delay). Read-write. Specifies the minimum time in memory clock cycles between activate commands to different chip select banks. Bits Description 0h Reserved 9h-1h <Trrd> clocks Fh-Ah Reserved 7:6 Reserved. 5:0 Trc: row cycle time. Read-write. Specifies the minimum time in memory clock cycles from and activate command to another activate command or an auto refresh command, all to the same chip select bank. Bits Description 09h-00h Reserved 3Ah-0Ah <Trc> clocks 3Fh-3Bh Reserved D18F2x208_dct[1:0] DRAM Timing 2 See 2.9.1 [DCT Configuration Registers]. Bits Description 31:27 Reserved. 26:24 Trfc3: auto refresh row cycle time for CS 6 and 7. See: Trfc0. 23:19 Reserved. 18:16 Trfc2: auto refresh row cycle time for CS 4 and 5. See: Trfc0. 15:11 Reserved. 10:8 Trfc1: auto refresh row cycle time for CS 2 and 3. See: Trfc0. 7:3 Reserved. 2:0 Trfc0: auto refresh row cycle time for CS 0 and 1. Read-write. Reset: 100b. Specifies the minimum time from a refresh command to the next valid command, except NOP or DES. The recommended programming of this register varies based on DRAM density and speed. Bits Description 000b Reserved 001b 90 ns (all speeds, 512 Mbit) 010b 110 ns (all speeds, 1 Gbit) 011b 160 ns (all speeds, 2 Gbit) 100b 300 ns (all speeds, 4 Gbit) 101b 350 ns (all speeds, 8 Gbit) 111b-110b Reserved D18F2x20C_dct[1:0]_mp[1:0] DRAM Timing 3 See 2.9.1 [DCT Configuration Registers]. Bits Description 31:18 Reserved. 400 42300 Rev 3.12 - July 14, 2015 BKDG for AMD Family 15h Models 10h-1Fh Processors 17:16 WrDqDqsEarly: DQ and DQS write early. Read-Write. Reset: 0. Specifies the DQ and DQS launch timing for write commands relative to the Tcwl MEMCLK. Bits Description 00b 0 MEMCLK early 01b 0.5 MEMCLK early 10b Reserved 11b Reserved 15:12 Reserved. 11:8 Twtr: internal DRAM write to read command delay. Read-write. Reset: 4h. Specifies the minimum number of memory clock cycles from a write operation to a read operation, both to the same chip select. This is measured from the rising clock edge following last non-masked data strobe of the write to the rising clock edge of the next read command. Bits Description 3h-0h Reserved Bh-4h <Twtr> clocks Fh-Ch Reserved 7:5 Reserved. 4:0 Tcwl: CAS write latency. Read-write. Reset: 5h. Specifies the number of memory clock cycles from internal write command to first write data in at the DRAM. Bits Description 04h-00h Reserved 11h-05h <Tcwl> clocks 1Fh-12h Reserved D18F2x210_dct[1:0]_nbp[3:0] DRAM NB P-state See 2.9.1 [DCT Configuration Registers]. For D18F2x210_dct[1:0]_nbp[x], x=D18F1x10C[NbPsSel]; see D18F1x10C[NbPsSel]. Table 177: BIOS Recommendations for RdPtrInit Condition D18F2x210_dct[1:0]_nbp[3:0] NCLK:MCLK ratio DDR rate (MT/s) RdPtrInit < 2:1 < 2133 0011b < 2:1 2133 <= rate <= 2400 0011b or 0010b See Note 1. >=2:1 < 1866 0110b >=2:1 1866 <= rate < 2400 0101b >=2:1 2400 0100b 1. For each NB P-state, IF any D18F2x9C_x0000_0[3:0]0[2:1]_dct[1:0]_mp[MemPstate][WrDatGrossDly] ==0 THEN RdPtrInit=0010b ELSE RdPtrInit=0011b. 401 42300 Rev 3.12 - July 14, 2015 Bits BKDG for AMD Family 15h Models 10h-1Fh Processors Description 31:22 MaxRdLatency: maximum read latency. Read-write. Reset: 000h. BIOS: 2.9.5.9.5. Specifies the maximum round-trip latency in the system from the processor to the DRAM devices and back. The DRAM controller uses this to help determine when the first two beats of incoming DRAM read data can be safely transferred to the NCLK domain. The time includes the asynchronous and synchronous latencies. Description Bits 000h 0 NCLKs 3FEh-001h <MaxRdLatency> NCLKs 3FFh 1023 NCLKs 21:19 Reserved. 18:16 DataTxFifoWrDly: data transmit FIFO write delay. Read-write. Reset: 0. BIOS: 0h. Specifies the DCT to phy write data FIFO delay. BIOS must program this field for the current or target NB P-state prior to a frequency change or NB P-state change. See also 2.9.5.2 [NB P-state Specific Configuration]. Bits Description 000b 0 MEMCLK 001b 0.5 MEMCLK 010b 1.0 MEMCLK 011b 1.5 MEMCLKs 100b 2.0 MEMCLKs 101b 2.5 MEMCLKs 110b 3.0 MEMCLKs 111b Reserved 15:4 Reserved. 3:0 RdPtrInit: read pointer initial value. Read-write. Reset: 6h. BIOS: Table 177. There is a synchronization FIFO between the NB clock domain and memory clock domain. Each increment of this field positions the read pointer one half clock cycle closer to the write pointer thereby reducing the latency through the FIFO. BIOS must program this field for the current or target NB P-state prior to a frequency change or NB P-state change. See also 2.9.5.2 [NB P-state Specific Configuration]. Bits Description 0001b-0000b Reserved 0010b 3 MEMCLKs 0011b 2.5 MEMCLKs 0100b 2 MEMCLKs 0101b 1.5 MEMCLKs 0110b 1 MEMCLK 1111b-0111b Reserved D18F2x214_dct[1:0]_mp[1:0] DRAM Timing 4 Reset: 0001_0202h. Bits Description 31:20 Reserved. 402 42300 Rev 3.12 - July 14, 2015 BKDG for AMD Family 15h Models 10h-1Fh Processors 19:16 TwrwrSdSc: write to write timing same DIMM same chip select. Read-write. BIOS: See 2.9.5.6.2 [TwrwrSdSc, TwrwrSdDc, TwrwrDd (Write to Write Timing)]. Specifies the minimum number of cycles from the last clock of virtual CAS of the first write-burst operation to the clock in which CAS is asserted for a following write-burst operation. Description Bits 0h Reserved 1h 1 clock Ah-2h <TwrwrSdSc> clocks Bh 11 clocks Fh-Ch Reserved 15:12 Reserved. 11:8 TwrwrSdDc: write to write timing same DIMM different chip select. See: TwrwrDd. 7:4 Reserved. 3:0 TwrwrDd: write to write timing different DIMM. Read-write. BIOS: See 2.9.5.6.2 [TwrwrSdSc, TwrwrSdDc, TwrwrDd (Write to Write Timing)]. Specifies the minimum number of cycles from the last clock of virtual CAS of the first write-burst operation to the clock in which CAS is asserted for a following write-burst operation. Bits Description 1h-0h Reserved Bh-2h <TwrwrDd> clocks Fh-Ch Reserved D18F2x218_dct[1:0]_mp[1:0] DRAM Timing 5 Reset: 0103_0203h. See 2.9.1 [DCT Configuration Registers]. Bits Description 31:28 Reserved. 27:24 TrdrdSdSc: read to read timing same DIMM same chip select. Read-write. BIOS: See 2.9.5.6.1 [TrdrdSdSc, TrdrdSdDd, and TrdrdDd (Read to Read Timing)]. Specifies the minimum number of cycles from the last clock of virtual CAS of a first read-burst operation to the clock in which CAS is asserted for a following read-burst operation. Bits Description 0h Reserved Bh-1h <TrdrdSdSc> clocks Fh-Ch Reserved 23:20 Reserved. 19:16 TrdrdSdDc: read to read timing same DIMM different chip select. See: TrdrdDd. 15:12 Reserved. 403 42300 Rev 3.12 - July 14, 2015 BKDG for AMD Family 15h Models 10h-1Fh Processors 11:8 Twrrd: write to read DIMM termination turnaround. Read-write. BIOS: See 2.9.5.6.3 [Twrrd (Write to Read DIMM Termination Turn-around)]. Specifies the minimum number of cycles from the last clock of virtual CAS of the first write-burst operation to the clock in which CAS is asserted for a following read-burst operation, both to different chip selects. Bits Description 0h Reserved Bh-1h <Twrrd> clocks Fh-Ch Reserved 7:4 Reserved. 3:0 TrdrdDd: read to read timing different DIMM. Read-write. BIOS: See 2.9.5.6.1 [TrdrdSdSc, TrdrdSdDd, and TrdrdDd (Read to Read Timing)]. Specifies the minimum number of cycles from the last clock of virtual CAS of a first read-burst operation to the clock in which CAS is asserted for a following read-burst operation. Bits Description 1h-0h Reserved Bh-2h <TrdrdDd> clocks Fh-Ch Reserved D18F2x21C_dct[1:0]_mp[1:0] DRAM Timing 6 Reset: 0004_0300h. See 2.9.1 [DCT Configuration Registers]. Bits Description 31:21 Reserved. 20:16 TrwtWB: read to write turnaround for opportunistic write bursting. Read-write. BIOS: TrwtTO + 1. Specifies the minimum number of clock cycles from the last clock of virtual CAS of a first readburst operation to the clock in which CAS is asserted for a following write-burst operation. Bits Description 02h-00h Reserved 17h-03h <TrwtWB> clocks 1Fh-18h Reserved 15:13 Reserved. 12:8 TrwtTO: read to write turnaround. Read-write. BIOS: See 2.9.5.6.4 [TrwtTO (Read-to-Write Turnaround for Data, DQS Contention)]. Specifies the minimum number of clock cycles from the last clock of virtual CAS of a first read-burst operation to the clock in which CAS is asserted for a following write-burst operation. Bits Description 01h-00h Reserved 16h-02h <TrwtTO> clocks 1Fh-17h Reserved 7:0 Reserved. D18F2x220_dct[1:0] DRAM Timing 7 Reset: 0000_0C04h. See 2.9.1 [DCT Configuration Registers]. 404 42300 Rev 3.12 - July 14, 2015 Bits BKDG for AMD Family 15h Models 10h-1Fh Processors Description 31:13 Reserved. 12:8 Tmod: mode register command delay. Read-write. BIOS: See 2.9.5.5. Specifies the minimum time in memory clock cycles from an MRS command to another non-MRS command (excluding NOP and DES), all to the same chip select. Description Bits 1h-0h Reserved 14h-2h <Tmod> clocks 1Fh-15h Reserved 7:4 Reserved. 3:0 Tmrd: mode register command cycle time. Read-write. BIOS: See 2.9.5.5. Specifies the minimum time in memory clock cycles from an MRS command to another MRS command, all to the same chip select. Bits Description 1h-0h Reserved 8h-2h <Tmrd> clocks Fh-9h Reserved D18F2x224_dct[1:0] DRAM Timing 8 Reset: 0000_0408h. See 2.9.1 [DCT Configuration Registers]. Bits Description 31:11 Reserved. 10:8 Tzqcs: Zq short cal command delay. Read-write. BIOS: See 2.9.5.5. Specifies the minimum time in memory clock cycles from a ZQCS command to any other command (excluding NOP and DES) on the channel. Description Bits Description Bits 000b Reserved 100b 64 clocks 001b 16 clocks 101b 80 clocks 010b 32 clocks 110b 96 clocks 011b 48 clocks 111b Reserved 7:4 Reserved. 3:0 Tzqoper: Zq long cal command delay. Read-write. BIOS: See 2.9.5.5. Specifies the minimum time in memory clock cycles from a ZQCL command to any other command (excluding NOP and DES) on the channel. Bits Description Bits Description 0000b Reserved 1000b 256 clocks 0001b 32 clocks 1001b 288 clocks 0010b 64 clocks 1010b 320 clocks 0011b 96 clocks 1011b 352 clocks 0100b 128 clocks 1100b 384 clocks 0101b 160 clocks 1111b-1101b Reserved 0110b 192 clocks 0111b 224 clocks 405 42300 Rev 3.12 - July 14, 2015 BKDG for AMD Family 15h Models 10h-1Fh Processors D18F2x228_dct[1:0] DRAM Timing 9 See 2.9.1 [DCT Configuration Registers]. Bits Description 31:24 Tstag3: auto refresh stagger time for logical DIMM 3. See: Tstag0. 23:16 Tstag2: auto refresh stagger time for logical DIMM 2. See: Tstag0. 15:8 Tstag1: auto refresh stagger time for logical DIMM 1. See: Tstag0. 7:0 Tstag0: auto refresh stagger time for logical DIMM 0. Read-write. Reset: 00h. BIOS: 14h. Specifies the number of clocks between auto refresh commands to different ranks of a DIMM when D18F2x90_dct[1:0][StagRefEn]=1 or D18F2x[6C:60]_dct[1:0][RankDef] != 0. BitsDescription 00h 0 clocks FEh-01h <Tstag0> clocks FFh 255 clocks D18F2x22C_dct[1:0]_mp[1:0] DRAM Timing 10 Reset: 0000_000Ch. See 2.9.1 [DCT Configuration Registers]. Bits Description 31:5 Reserved. 4:0 Twr: write recovery. Read-write. BIOS: See 2.9.5.5. Specifies the minimum time from the last data write until the chip select bank precharge. Description Bits 4h-0h Reserved 8h-5h 8 to 5 clocks 9h Reserved Ah 10 clocks Bh Reserved Ch 12 clocks Dh Reserved Eh 14 clocks Fh Reserved 10h 16 clocks 11h Reserved 12h 18 clocks 1Fh-13h Reserved D18F2x[234:230]_dct[1:0] DRAM Read ODT Pattern [High:Low] Reset: 0000_0000h. See 2.9.1 [DCT Configuration Registers]. This register is used by BIOS to specify the state of the ODT pins during DDR reads. F2x230 is used to control chip selects 0-3. F2x234 is used to control chip selects 4-7. See section 2.9.5.6.5 [DRAM ODT Control] for more information. Bits Description 31:28 Reserved. 406 42300 Rev 3.12 - July 14, 2015 BKDG for AMD Family 15h Models 10h-1Fh Processors 27:24 RdOdtPatCs73: read ODT pattern chip select [7,3]. See: RdOdtPatCs40. 23:20 Reserved. 19:16 RdOdtPatCs62: read ODT pattern chip select [6,2]. See: RdOdtPatCs40. 15:12 Reserved. 11:8 RdOdtPatCs51: read ODT pattern chip select [5,1]. See: RdOdtPatCs40. 7:4 Reserved. 3:0 RdOdtPatCs40: read ODT pattern chip select [4,0]. Read-write. Specifies the state of ODT[3:0] pins when a read occurs to the specified chip select. D18F2x[23C:238]_dct[1:0] DRAM Write ODT Pattern [High:Low] Reset: 0000_0000h. See 2.9.1 [DCT Configuration Registers]. This register is used by BIOS to specify the state of the ODT pins during DDR writes. F2x238 is used to control chip selects 0-3. F2x23C is used to control chip selects 4-7. See section 2.9.5.6.5 [DRAM ODT Control] for more information. Bits Description 31:28 Reserved. 27:24 WrOdtPatCs73: write ODT pattern chip select [7,3]. See: WrOdtPatCs40. 23:20 Reserved. 19:16 WrOdtPatCs62: write ODT pattern chip select [6,2]. See: WrOdtPatCs40. 15:12 Reserved. 11:8 WrOdtPatCs51: write ODT pattern chip select [5,1]. See: WrOdtPatCs40. 7:4 Reserved. 3:0 WrOdtPatCs40: write ODT pattern chip select [4,0]. Read-write. Specifies the state of ODT[3:0] pins when a write occurs to the specified chip select. D18F2x240_dct[1:0]_mp[1:0] DRAM ODT Control Reset: 0000_0000h. See 2.9.1 [DCT Configuration Registers]. Bits Description 31:15 Reserved. 14:12 WrOdtOnDuration: write ODT on duration. Read-write. BIOS: 6. Specifies the number of memory clock cycles that ODT is asserted for an eight-beat write burst. The controller will shorten the ODT pulse duration by two clock cycles if the burst is chopped. Bits Description 5h-0h Reserved 7h-6h <WrOdtOnDuration> clocks 11 Reserved. 10:8 WrOdtTrnOnDly: Write ODT Turn On Delay. Read-write. BIOS: 0. Specifies the number of memory clock cycles that ODT assertion is delayed relative to write CAS. Bits Description 0h 0 clocks 7h-1h <WrOdtTrnOnDly> clocks, Reserved if (WrOdtOnDuration=0) 7 Reserved. 407 42300 Rev 3.12 - July 14, 2015 BKDG for AMD Family 15h Models 10h-1Fh Processors 6:4 RdOdtOnDuration: Read ODT On Duration. Read-write. BIOS: 6. Specifies the number of memory clock cycles that ODT is asserted for an eight-beat read burst. The controller will shorten the ODT pulse duration by two clock cycles if the burst is chopped. Description Bits 5h-0h Reserved 7h-6h <RdOdtOnDuration> clocks 3:0 RdOdtTrnOnDly: Read ODT Turn On Delay. Read-write. BIOS: MAX(0, D18F2x200_dct[1:0]_mp[1:0][Tcl] - D18F2x20C_dct[1:0]_mp[1:0][Tcwl]). Specifies the number of clock cycles that ODT assertion is delayed relative to read CAS. Bits Description 0h 0 clocks Fh-0h <RdOdtTrnOnDly> clocks, Reserved if (RdOdtOnDuration=0) D18F2x244_dct[1:0] DRAM Controller Miscellaneous 3 Reset: 0000_0000h. See 2.9.1 [DCT Configuration Registers]. Bits Description 31:4 Reserved. 3:0 PrtlChPDDynDly: partial channel power down dynamic delay. Read-write. BIOS: IF(D18F2xA8_dct[1:0][PrtlChPDEnhEn]) THEN 4h ELSE 0 ENDIF. Specifies the channel idle hysteresis for fast exit/slow exit mode changes when D18F2xA8_dct[1:0][PrtlChPDEnhEn]=1. Bits Description 0h 0 clocks 7h-1h <PrtlChPDDynDly*32> clocks 8h 256 clocks Fh-9h Reserved D18F2x248_dct[1:0]_mp[1:0] DRAM Power Management 0 Reset: 0000_0A03h. See 2.9.1 [DCT Configuration Registers]. Bits Description 31 RxChMntClkEn: Receive channel maintenance clocks. Read-Write. BIOS: See 2.9.5.11. 1=Enable receive channel maintenance clocks to improve internal timing margin at the cost of some extra power. 0=Disable clocks. To disable clocks, BIOS must first disable clock generation in the phy (see D18F2x9C_x0D0F_0[F,7:0]13_dct[1:0]_mp[1:0][RxSsbMntClkEn]). 30 Reserved. 29:24 AggrPDDelay: aggressive power down delay. Read-Write. BIOS: 20h. Specifies a hysteresis count from the last DRAM activity for the DCT to close pages prior to precharge power down. Reserved if ~D18F2xA8_dct[1:0][AggrPDEn]. See PchgPDEnDelay and D18F2x94_dct[1:0][PowerDownEn]. Bits Description 00h 64 clocks 01h 1 clock 3Eh-02h <AggrPDDelay> clocks 3Fh 63 clocks 23:22 Reserved. 408 42300 Rev 3.12 - July 14, 2015 BKDG for AMD Family 15h Models 10h-1Fh Processors 21:16 PchgPDEnDelay: precharge power down entry delay. Read-write. BIOS: MAX(D18F2x200_dct[1:0]_mp[1:0][Tcl] + 5, D18F2x20C_dct[1:0]_mp[1:0][Tcwl] + D18F2x22C_dct[1:0]_mp[1:0][Twr] + 5, D18F2x220_dct[1:0][Tmod]). Specifies the power down entry delay. If D18F2xA8_dct[1:0][AggrPDEn] = 0, this delay behaves as a hysteresis. This field must satisfy the minimum power down entry delay requirements. See also D18F2x94_dct[1:0][PowerDownEn]. Description Bits 00h 64 clocks 01h 1 clock 3Eh-02h <PchgPDEnDelay> clocks 3Fh 63 clocks 15:13 Reserved. 12:8 Txpdll: exit DLL and precharge powerdown to command delay. Read-write. BIOS: See 2.9.5.5. Specifies the minimum time that the DCT waits to issue a command after exiting precharge powerdown mode if the DLL was also disabled. Bits Description 09h-00h Reserved 1Dh-0Ah <Txpdll> clocks 1Fh-1Eh Reserved 7:4 Reserved. 3:0 Txp: exit precharge PD to command delay. Read-write. BIOS: See 2.9.5.5. Specifies the minimum time that the DCT waits to issue a command after exiting precharge powerdown mode. Bits Description 2h-0h Reserved 8h-3h <Txp> clocks Fh-9h Reserved D18F2x24C_dct[1:0] DRAM Power Management 1 Reset: 0214_0803h. See 2.9.1 [DCT Configuration Registers]. Bits Description 31:30 Reserved. 29:24 Tcksrx: clock stable to self refresh exit delay. Read-write. BIOS: See 2.9.5.5. Specifies the minimum time in memory clock cycles that the DCT waits to assert CKE after clock frequency is stable. Bits Description 01h-00h Reserved 0Eh-02h <Tcksrx> clocks 3Fh-0Fh Reserved 23:22 Reserved. 21:16 Tcksre: self refresh to command delay. Read-write. BIOS: See 2.9.5.5. Specifies the minimum time in memory clock cycles that the DCT waits to remove external clocks after entering self refresh or powerdown. Bits Description 04h-00h Reserved 27h-05h <Tcksre> clocks 3Fh-28h Reserved 15:14 Reserved. 409 42300 Rev 3.12 - July 14, 2015 BKDG for AMD Family 15h Models 10h-1Fh Processors 13:8 Tckesr: self refresh to command delay. Read-write. BIOS: See 2.9.5.5. Specifies the minimum time in memory clock cycles that the DCT waits to issue a command after entering self refresh. Bits Description 01h-00h Reserved 2Bh-02h <Tckesr> clocks 3Fh-2Ch Reserved 7:4 Reserved. 3:0 Tpd: minimum power down entry to exit. Read-write. BIOS: See 2.9.5.5. Bits Description 0h Reserved Ah-1h <Tpd> clocks Fh-Bh Reserved D18F2x250_dct[1:0] DRAM Loopback and Training Control Reset: 0000_0000h. See 2.9.1 [DCT Configuration Registers]. See 2.9.5.9.6.1 [DRAM Training Pattern Generation]. Bits Description 31:13 Reserved. 12 CmdSendInProg: command in progress. Read-only. 0=DCT is idle. 1=DCT is busy. 11 SendCmd: send command. Read-write. 0=Stop command generation. 1=Begin command generation as specified in CmdTgt, CmdType, and D18F2x260_dct[1:0][CmdCount]. BIOS must set this field to a 0 after a command series is completed. Reserved if ~CmdTestEnable. 10 TestStatus: test status. Read-only. 0=Command generation is in progress. 1=Command generation has completed. Reserved if ~(SendCmd & (D18F2x260_dct[1:0][CmdCount] > 0 | StopOnErr)). 9:8 CmdTgt: command target. Read-write. Specifies the SendCmd command target address mode. See D18F2x25[8,4]_dct[1:0]. Bits Description 00b Issue commands to address Target A 01b Issue alternating commands to address Target A and Target B 11b-10b Reserved 7:5 CmdType: command type. Read-write. Specifies the SendCmd command type. Description Bits 000b Read 001b Write 010b Alternating write and read 111b-011b Reserved 4 StopOnErr: stop on error. Read-write. Specifies the DCT behavior if a data comparison error occurs. 1=Stop command generation. 0=Continue command generation. If StopOnErr=1, BIOS must program ResetAllErr=1 when programming SendCmd=1. 3 ResetAllErr: reset all errors. Read; write-1-only; cleared-by-hardware. 1=Clear error status bits and error counters in D18F2x264_dct[1:0], D18F2x268_dct[1:0], and D18F2x268_dct[1:0]. 2 CmdTestEnable: command test enable. IF (D18F2x118[LockDramCfg]) THEN Read-only. ELSE Read-write. ENDIF. 0=Disable the command generation mode. 1=Enable the command generation mode. See SendCmd. 1:0 Reserved. 410 42300 Rev 3.12 - July 14, 2015 BKDG for AMD Family 15h Models 10h-1Fh Processors D18F2x25[8,4]_dct[1:0] DRAM Target [B, A] Base Reset: 0000_0000h. See 2.9.1 [DCT Configuration Registers]. See 2.9.5.9.6.1 [DRAM Training Pattern Generation]. Bits Description 31:27 Reserved. 26:24 TgtChipSelect: target chip select. Read-write. Specifies the chip select. 0=CS0, 1=CS1, …, 7=CS7. 23:21 TgtBank: target bank [2:0]. Read-write. Specifies the bank address. 20:10 Reserved. 9:0 TgtAddress[9:0]: target address [9:0]. Read-write. Specifies the column address bits [9:0].The corresponding address sequence in a command series is as follows: TgtAddress[9:3] is incremented by one, with wrap around, after each command if D18F2x250_dct[1:0][CmdType] = 00xb or if D18F2x250_dct[1:0][CmdType] = 010b and D18F2x250_dct[1:0][CmdTgt] = 01b. TgtAddress[9:3] is incremented by one, with wrap around, after each command pair if D18F2x250_dct[1:0][CmdType] = 010b and D18F2x250_dct[1:0][CmdTgt] = 00b. D18F2x260_dct[1:0] DRAM Command 1 Reset: 0000_0000h. See 2.9.1 [DCT Configuration Registers]. See 2.9.5.9.6.1 [DRAM Training Pattern Generation]. Bits Description 31:21 Reserved. 20:0 CmdCount: command count. Read-write. Specifies the maximum number of commands to generate when D18F2x250_dct[1:0][SendCmd]=1. See also D18F2x250_dct[1:0][StopOnErr]. Bits Description 0h Infinite commands 1F_FFFFh-1h <CmdCount> commands D18F2x264_dct[1:0] DRAM Status 0 Reset: 0000_0000h. See 2.9.1 [DCT Configuration Registers]. See 2.9.5.9.6.1 [DRAM Training Pattern Generation]. Bits Description 411 42300 Rev 3.12 - July 14, 2015 BKDG for AMD Family 15h Models 10h-1Fh Processors 31:25 ErrDqNum: error DQ number. Read-only. Indicates the DQ bit of the first error occurrence when D18F2x264_dct[1:0][ErrCnt] > 0. Cleared by D18F2x250_dct[1:0][ResetAllErr]. Bits Description 00h Data[0] 3Eh-01h Data[<ErrDqNum>] 3Fh Data[63] 40h ECC[0] 46h-41h ECC[<ErrDqNum>-40h] 47h ECC[7] 7Fh-48h Reserved 24:0 ErrCnt: error count. Read; set-by-hardware; write-1-to-clear. Specifies a saturating counter indicating the number of DQ bit errors detected. Counts a maximum of 72 errors per bit-time. Status is accumulated until cleared by D18F2x250_dct[1:0][ResetAllErr]. Errors can an be masked on per-bit basis by programming D18F2x274_dct[1:0] and D18F2x278_dct[1:0]. Description Bits 0h 0 errors 1FF_FFFDh-1h <ErrCnt> errors 1FF_FFFEh 1FF_FFFEh errors 1FF_FFFFh 1FF_FFFFh or more errors D18F2x268_dct[1:0] DRAM Status 1 Reset: 0000_0000h. See 2.9.1 [DCT Configuration Registers]. See 2.9.5.9.6.1 [DRAM Training Pattern Generation]. Bits Description 31:18 Reserved. 17:0 NibbleErrSts: nibble error status. Read-only. Indicates error detection status on a per nibble basis when D18F2x264_dct[1:0][ErrCnt] > 0. Status is accumulated until cleared by D18F2x250_dct[1:0][ResetAllErr]. Bit Description [0] Data[3:0] [1] Data[7:4] [14:2] Data[(<NibbleErrSts>*4)+3:<NibbleErrSts>*4] [15] Data[63:60] [16] ECC[3:0] [17] ECC[7:4] D18F2x26C_dct[1:0] DRAM Status 2 Reset: 0000_0000h. See 2.9.1 [DCT Configuration Registers]. See 2.9.5.9.6.1 [DRAM Training Pattern Generation]. Bits Description 412 42300 Rev 3.12 - July 14, 2015 BKDG for AMD Family 15h Models 10h-1Fh Processors 31:18 Reserved. 17:0 NibbleErr180Sts: nibble error 180 status. Read-only. Indicates error detection status on a per nibble basis when D18F2x264_dct[1:0][ErrCnt] > 0, comparing read data against data shifted 1-bit time earlier. Status is accumulated until cleared by D18F2x250_dct[1:0][ResetAllErr]. Bit Description [0] Data[3:0] [1] Data[7:4] [14:2] Data[(<NibbleErr180Sts>*4)+3:<NibbleErr180Sts>*4] [15] Data[63:60] [16] ECC[3:0] [17] ECC[7:4] D18F2x270_dct[1:0] DRAM PRBS See 2.9.1 [DCT Configuration Registers]. See 2.9.5.9.6.1 [DRAM Training Pattern Generation]. Bits 31 Description Reserved. 30:24 Reserved. 23:19 Reserved. 18:0 DataPrbsSeed: data PRBS seed. Read-write. Reset: 7FFFFh. Specifies the seed value used for creating pseudo random traffic on the data bus. This register must be written with a non-zero seed value. Recommended BIOS values for DRAM training: CmdCount DataPrbsSeed 32 62221h 64 66665h 128 26666h 25644443h D18F2x274_dct[1:0] DRAM DQ Mask Low See D18F1x10C[DctCfgSel]. See 2.9.1 [DCT Configuration Registers]. See 2.9.5.9.6.1 [DRAM Training Pattern Generation]. Bits Description 31:0 DQMask[31:0]: DQ mask. Read-write. DQMask[63:0] = {D18F2x278_dct[1:0][DQMask[63:32]], DQMask[31:0]}. Reset: 0000_0000_0000_0000h. 1=The corresponding DQ bit will not be compared. 0=The corresponding DQ bit will be compared. See D18F2x264_dct[1:0][ErrCnt]. Description Bit [0] Data[0] [62:1] Data[<DQMask>] [63] Data[63] D18F2x278_dct[1:0] DRAM DQ Mask High Bits Description 31:0 DQMask[63:32]: DQ mask. See: D18F2x274_dct[1:0][DQMask[31:0]]. 413 42300 Rev 3.12 - July 14, 2015 BKDG for AMD Family 15h Models 10h-1Fh Processors D18F2x28C_dct[1:0] DRAM Command 2 Reset: 0000_0000h. See 2.9.1 [DCT Configuration Registers]. See 2.9.5.9.6.1 [DRAM Training Pattern Generation]. This register may only be used when D18F2x250_dct[1:0][CmdTestEnable]=1. Bits Description 31 SendActCmd: send activate command. Read; write-1-only; cleared-by-hardware. 1=The DCT sends an activate command as specified by ChipSelect, Bank, and Address. This bit is cleared by hardware after the command completes. 30 SendPchgCmd: send precharge all command. Read; write-1-only; cleared-by-hardware. The DCT sends a precharge command based on CmdAddress[10]. This bit is cleared by hardware after the command completes. 0=Command has completed. 1=If (CmdAddress[10]=1) then send a precharge all command as specified by CmdChipSelect; If (CmdAddress[10]=0) then send a precharge command as specified by CmdChipSelect, CmdBank. 29:22 CmdChipSelect: command chip select. Read-write. Specifies the chip select. For LR-DIMMs, BIOS programs this field with the logical rank CS. Bit Description [0] CS0 [6:1] CS<CmdChipSelect> [7] CS7 21:19 CmdBank[2:0]: command bank [2:0]. Read-write. Specifies the bank address. 18 Reserved. 17:0 CmdAddress[17:0]: command address [17:0]. Read-write. Specifies the row address. D18F2x290_dct[1:0] DRAM Status 3 Reset: 0000_0000h. See 2.9.1 [DCT Configuration Registers]. See 2.9.5.9.6.1 [DRAM Training Pattern Generation]. Bits Description 31:27 Reserved. 26:24 ErrBeatNum: error beat number. Read-only. Indicates the data beat of the first error occurrence in the command reported by ErrCmdNum when D18F2x264_dct[1:0][ErrCnt] > 0 and D18F2x260_dct[1:0][CmdCnt] > 0. Cleared by D18F2x250_dct[1:0][ResetAllErr]. Bits Description 7h-0h <ErrBeatNum> beat 23:21 Reserved. 20:0 ErrCmdNum: error command number. Read-only. Indicates the command number of the first error occurrence when D18F2x264_dct[1:0][ErrCnt] > 0 and D18F2x260_dct[1:0][CmdCnt] > 0. Cleared by D18F2x250_dct[1:0][ResetAllErr]. Bits Description 0 No error 1F_FFFFh-1h <ErrCmdNum> command D18F2x294_dct[1:0] DRAM Status 4 See 2.9.1 [DCT Configuration Registers]. See 2.9.5.9.6.1 [DRAM Training Pattern Generation]. 414 42300 Rev 3.12 - July 14, 2015 Bits BKDG for AMD Family 15h Models 10h-1Fh Processors Description 31:0 DQErr[31:0]: DQ error. Read-only. DQErr[63:0] = {D18F2x298_dct[1:0][DQErr[63:32]], DQErr[31:0]}. Reset: 0000_0000_0000_0000h. Indicates error detection status on a per bit basis when D18F2x264_dct[1:0][ErrCnt] > 0. Status is accumulated until cleared by D18F2x250_dct[1:0][ResetAllErr]. Bit Description [0] Data[0] [62:1] Data[<DQErr>] [63] Data[63] D18F2x298_dct[1:0] DRAM Status 5 Bits Description 31:0 DQErr[63:32]: DQ error. See: D18F2x294_dct[1:0][DQErr[31:0]]. D18F2x2E0_dct[1:0] Memory P-state Control and Status See 2.9.1 [DCT Configuration Registers]. Bits Description 31 Reserved. 30 FastMstateDis: fast M-state change disable. Read-write. Reset: 0. 1=The DCT changes MEMCLK frequency only after the NCLK frequency has changed. 0=The DCT changes MEMCLK frequency while the northbridge changes NCLK. 29 Reserved. 28:24 M1MemClkFreq: M1 memory clock frequency. Read-write. Reset: 00h. Specifies the frequency of the DRAM interface (MEMCLK) for memory P-state 1. See Table 132 [Memory Clock Frequency Value Definition]. The hardware enforces D18F5x84[DdrMaxRateEnf] when writes to this field occur. See D18F5x84[DdrMaxRate] and D18F5x84[DdrMaxRateEnf]. BIOS must also program D18F2x9C_x0D0F_E000_dct[1:0]_mp[1:0][Rate] for M1. 23 Reserved. 22:20 MxMrsEn: Mx Mrs enable. Read-write. Reset:0h. 1=The DCT writes to the DRAM MR after a memory P-state change. 0=The DCT does not write to the DRAM MR. Bit Description, MR value [0] MR0, D18F2x2E8_dct[1:0]_mp[1:0][MxMr0] [1] MR1, D18F2x2E8_dct[1:0]_mp[1:0][MxMr1] [2] MR2, D18F2x2EC_dct[1:0]_mp[1:0][MxMr2] 19:1 Reserved. 0 CurMemPstate: current memory P-state. Read-only; updated-by-hardware. Specifies the current memory P-state. 0=M0. 1=M1. D18F2x2E8_dct[1:0]_mp[1:0] MRS Buffer • See 2.9.1 [DCT Configuration Registers]. 415 42300 Rev 3.12 - July 14, 2015 Bits BKDG for AMD Family 15h Models 10h-1Fh Processors Description 31:16 MxMr1: Mx MR1. Read-write. Reset: 0000h. Specifies the value written to DRAM MR1 after a memory P-state change. If the M1 value is the same as the M0 value, then BIOS should optimize Pstate switching latency by programming D18F2x2E0_dct[1:0][MxMrsEn]=0 15:0 MxMr0: Mx MR0. Read-write. Reset: 0000h. Specifies the value written to DRAM MR0 after a memory P-state change. Bit 8 of MxMr0, corresponding to “DLL Reset”, should be programmed with 0b. If the M1 value is the same as the M0 value, then BIOS should optimize P-state switching latency by programming D18F2x2E0_dct[1:0][MxMrsEn]=0 D18F2x2EC_dct[1:0]_mp[1:0] MRS Buffer See 2.9.1 [DCT Configuration Registers]. Bits Description 31:16 Reserved. 15:0 MxMr2: Mx MR2. Read-write. Reset: 0000h. Specifies the value written to DRAM MR2 after a memory P-state change. If the M1 value is the same as the M0 value, then BIOS should optimize Pstate switching latency by programming D18F2x2E0_dct[1:0][MxMrsEn]=0 D18F2x2F0_dct[1:0]_mp[1:0] DRAM Controller Misc 3 See 2.9.1 [DCT Configuration Registers]. Bits Description 31:1 Reserved. 0 EffArbDis: Efficient arbitration disable. Read-write. Reset: 0. BIOS: 0. 0=The DCT optimizes the arbitration phases to improve performance under certain traffic conditions whenever the NCLK to MEMCLK ratio is less than 2:1. 1=The DCT arbitrates normally, at all NCLK:MEMCLK ratios. D18F2x400_dct[1:0] GMC to DCT Control 0 See 2.9.1 [DCT Configuration Registers]. The GMC to DCT interface controls how DRAM bus resources are allocated and arbitrated between the MCT and the GMC. A token is the unit of available resource and is equivalent to a DCQ entry. A minimum count guarantees a number of available DCQ entries. A token limit for MCT or GMC guarantees resources are not all allocated to the GMC, or MCT respectively. Limits are configured bimodal: for normal GMC traffic and for when urgent (nominally display refresh) GMC traffic is occurring. Bits Description 31:16 Reserved. 15:12 Reserved. 11:8 GmcTokenLimit: GMC token limit. Read-write.Reset: 0h. BIOS: 4h. Limit of outstanding GMC tokens. 7:4 Reserved. 3:0 MctTokenLimit: MCT token limit. Read-write. Reset: 0h. BIOS: 4h. Limit of outstanding MCT tokens. 416 42300 Rev 3.12 - July 14, 2015 BKDG for AMD Family 15h Models 10h-1Fh Processors D18F2x404_dct[1:0] GMC to DCT Control 1 Reset: 0000_0000h. See 2.9.1 [DCT Configuration Registers]. Bits Description 31:17 Reserved. 16 UrgentTknDis: Urgent token disable. Read-write. BIOS: 0. 0=When urgent GMC traffic is requested , override the programmed values in D18F2x400_dct[1:0] and force the token scheme to heavily weight towards graphics by using the programmable token limits in D18F2x404_dct[1:0]. 1=Token scheme remains at the previously programmed non-urgent token limits in D18F2x400_dct[1:0] regardless of urgent GMC traffic. 15:12 UrGmcMinTokens: Display refresh GMC minimum tokens. Read-write. BIOS: 4h. Urgent mode minimum number of tokens assigned to the GMC. 11:8 UrGmcTokenLimit: Display refresh GMC token limit. Read-write. BIOS: 4h. Urgent mode limit of outstanding GMC tokens. 7:4 UrMctMinTokens: Display refresh MCT minimum tokens. Read-write. BIOS: 4h. Urgent mode minimum number of tokens assigned to the MCT. 3:0 UrMctTokenLimit: Display refresh MCT token limit. Read-write. BIOS: 4h. Urgent mode limit of outstanding MCT tokens. IF (!UrgentTknDis) THEN UrMctTokenLimit <= D18F2x400_dct[1:0][MctTokenLimit]. D18F2x408_dct[1:0] GMC to DCT Control 2 See 2.9.1 [DCT Configuration Registers]. Bits Description 31:2 Reserved. 1 TokenAllocSelect: Token allocation select. Read-write. Reset: 0. BIOS: 0. 0=When both the MCT and GMC have less than their maximum outstanding tokens, garlic allocates tokens by alternating between each. 1= When both the MCT and GMC have less than their maximum outstanding tokens, garlic allocates tokens to whichever has less (DCQ entries + current outstanding). 0 CpuElevPrioDis: Cpu elevate priority disable. Read-write. Reset: 0. 1=Reads from MCT arbitrate with GMC traffic normally. 0=Elevate the priority of an MCT read to high. This can alleviate CPU stalls during very long graphics requests. D18F2x420_dct[1:0] GMC to DCT FIFO Config 1 See 2.9.1 [DCT Configuration Registers]. Bits Description 31:12 Reserved. 11:8 SbRdPtrInit: Sideband signal read pointer init value. Read-write. Reset: 4h. BIOS: 4h. Specifies the read pointer for the sideband signal FIFO between GMC and DCT. 7:4 Reserved. 3:0 CmdRdPtrInit: Command read pointer init value. Read-write. Reset: 4h. BIOS: 4h. Specifies the read pointer for command channel between GMC and DCT. 417 42300 Rev 3.12 - July 14, 2015 BKDG for AMD Family 15h Models 10h-1Fh Processors 418 42300 Rev 3.12 - July 14, 2015 BKDG for AMD Family 15h Models 10h-1Fh Processors 3.11 Device 18h Function 3 Configuration Registers See 3.1 [Register Descriptions and Mnemonics]. See 2.7 [Configuration Space]. D18F3x00 Device/Vendor ID Bits Description 31:16 DeviceID: device ID. Read-only. Value: 1403h. 15:0 VendorID: vendor ID. Read-only. Value: 1022h. D18F3x04 Status/Command Bits Description 31:16 Status. Read-only. Reset: 0000h, except bit[20]. Bit[20] is set to indicate the existence of a PCIdefined capability block, if one exists. 15:0 Command. Read-only. Reset: 0000h. D18F3x08 Class Code/Revision ID Bits Description 31:8 ClassCode. Read-only. Reset: 060000h. Provides the host bridge class code as defined in the PCI specification. 7:0 RevID: revision ID. Read-only. Reset: 00h. D18F3x0C Header Type Reset: 0080_0000h. Bits Description 31:0 HeaderTypeReg. Read-only. These bits are fixed at their default values. The header type field indicates that there are multiple functions present in this device. D18F3x34 Capability Pointer Bits Description 31:8 Reserved. 7:0 CapPtr. Read-only. Value: 00h. D18F3x40 MCA NB Control Bits Description 31:0 MSR0000_0410[31:0] is an alias of D18F3x40. See MSR0000_0410[31:0]. 419 42300 Rev 3.12 - July 14, 2015 BKDG for AMD Family 15h Models 10h-1Fh Processors D18F3x44 MCA NB Configuration See D18F3x180 [Extended NB MCA Configuration]. It is expected that all fields of this register are programmed to the same value in all nodes, except for the fields used for link error injection: GenLinkSel, GenSubLinkSel, GenCrcErrByte1, GenCrcErrByte0. Bits Description 31 NbMcaLogEn: northbridge MCA log enable. Read-write. Reset: 0. 1=Enables logging (but not reporting) of NB MCA errors even if MCA is not globally enabled. 30 SyncFloodOnDramAdrParErr: sync flood on DRAM address parity error. Read-write. Reset: 0. BIOS: 1. 1=Enables sync flood on detection of a DRAM address parity error. 29 DisMstAbortCpuErrRsp: master abort CPU error response disable. Read-write. Reset: 0. 1=Disables master abort reporting through the CPU MCA error-reporting banks; Suppresses sending of RDE to CPU; Does not log any MCA information in the NB. 28 DisTgtAbortCpuErrRsp: target abort CPU error response disable. Read-write. Reset: 0. 1=Disables target abort reporting through the CPU MCA error-reporting banks; Suppresses sending of RDE to CPU; Does not log any MCA information in the NB. 27 NbMcaToMstCpuEn: machine check errors to master CPU only. Read-write. Reset: 0. BIOS: 1. 1=NB MCA errors in CMP device are only reported to the node base core (NBC); the following registers are only accessible from the NBC, non-NBC writes are ignored and reads are RAZ: MSR0000_0410, MSR0000_0411, MSR0000_0412. This allows machine check handlers running on different cores to avoid coordinating accesses to the NB MCA registers. This field does not affect PCI-defined configuration space accesses to these registers, which are accessible from all cores. See 3.1 [Register Descriptions and Mnemonics] for a description of MSR space and 3 [Registers] for PCIdefined configuration space. 0=NB MCA errors may be reported to the core that originated the request, if applicable and known, and the NB MCA registers in MSR space are accessible from any core. Note: • When the CPU which originated the request is known, it is stored in D18F3x4C[ErrCoreId], regardless of the setting of NbMcaToMstCpuEn. See Table 224 for errors where ErrCoreId is known. • If IO originated the request, then the error is reported to the NBC, regardless of the setting of NbMcaToMstCpuEn. 26 FlagMcaCorrErr: correctable error MCA exception enable. Read-write. Reset: 0. 1=Raise a machine check exception for correctable machine check errors which are enabled in D18F3x40. 25 DisPciCfgCpuErrRsp: PCI configuration CPU error response disable. Read-write. Reset: 0. 1=Disables generation of an error response to the core on detection of a master abort, target abort, or data error condition, and disables logging and reporting through the MCA error-reporting banks for PCI configuration accesses. For NB WDT errors on PCI configuration accesses, this prevents sending an error response to the core, but does not affect logging and reporting of the NB WDT error. See D18F3x180[DisPciCfgCpuMstAbtRsp], which applies only to master aborts. 24 IoRdDatErrEn: IO read data error log enable. Read-write. Reset: 0. 1=Enables MCA logging and reporting of errors on transactions from IO devices upon detection of a target abort, master abort, or data error condition. 0=Errors on transactions from IO devices are not logged in MCA, although error responses to the requesting IO device may still be generated. 23:22 Reserved. 420 42300 Rev 3.12 - July 14, 2015 BKDG for AMD Family 15h Models 10h-1Fh Processors 21 SyncFloodOnAnyUcErr: sync flood on any UC error. Read-write. Reset: 0. BIOS: 1. 1=Enables flooding of all links with sync packets on detection of any NB MCA error that is uncorrectable, including northbridge array errors and link protocol errors. 20 SyncFloodOnWDT: sync flood on watchdog timer error. Read-write. Reset: 0. BIOS: 1. 1=Enables flooding of all links with sync packets on detection of a watchdog timer error. 19:18 GenSubLinkSel: sublink select for CRC error generation. Read-write. Reset: 0. Selects the sublink of a link selected by GenLinkSel to be used for CRC error injection through GenCrcErrByte0 and GenCrcErrByte1. When the link is ganged, GenSubLinkSel must be 00b. When the link is unganged, the following values indicate which sublink is selected: Bits Description 00b Sublink 0 01b Sublink 1 10b Reserved 11b Reserved 17 GenCrcErrByte1: generate CRC error on byte lane 1. Read-Write. Reset: 0. 1=For ganged links (see GenSubLinkSel), a CRC error is injected on byte lane 1 of the link specified by GenLinkSel. For ganged links in retry mode or unganged links, this field is reserved, and GenCrcErrByte0 must be used. The data carried by the link is unaffected. This bit is cleared after the error has been generated. 16 GenCrcErrByte0: generate CRC error on byte lane 0. Read-Write. Reset: 0. 1=Causes a CRC error to be injected on byte lane 0 of the link specified by GenLinkSel and the sublink specified by GenSubLinkSel. The data carried by the link is unaffected. This bit is cleared after the error has been generated. 15:14 Reserved. 13:12 WDTBaseSel: watchdog timer time base select. Read-write. Reset: 0. Selects the time base used by the watchdog timer. The counter selected by WDTCntSel determines the maximum count value in the time base selected by WDTBaseSel. Bits Description 00b 1.31 ms 01b 1.28 us 10b Reserved 11b Reserved 421 42300 Rev 3.12 - July 14, 2015 BKDG for AMD Family 15h Models 10h-1Fh Processors 11:9 WDTCntSel[2:0]: watchdog timer count select bits[2:0]. Read-write. Reset: 0. Selects the count used by the watchdog timer. WDTCntSel = {D18F3x180[WDTCntSel[3]], D18F3x44[WDTCntSel[2:0]]}. The counter selected by WDTCntSel determines the maximum count value in the time base selected by WDTBaseSel. WDTCntSel is encoded as: Bits Description 0000b 4095 0001b 2047 0010b 1023 0011b 511 0100b 255 0101b 127 0110b 63 0111b 31 1000b 8191 1001b 16383 1111b-1010b Reserved Because WDTCntSel is split between two registers, care must be taken when programming WDTCntSel to ensure that a reserved value is never used by the watchdog timer or undefined behavior could result. 8 WDTDis: watchdog timer disable. Read-write. Cold reset: 0. 1=Disables the watchdog timer. The watchdog timer is enabled by default and checks for NB system accesses for which a response is expected and where no response is received. If such a condition is detected the outstanding access is completed by generating an error response back to the requestor. An MCA error may also be generated if enabled in D18F3x40 [MCA NB Control]. 7 IoErrDis: IO error response disable. Read-write. Reset: 0. 1=Disables setting either Error bit in link response packets to IO devices on detection of a target or master abort error condition. 6 CpuErrDis: CPU error response disable. Read-write. Reset: 0. BIOS: 1. 1=Disables generation of a read data error response to the core on detection of a target or master abort error condition. 5 IoMstAbortDis: IO master abort error response disable. Read-write. Reset: 0. 1=Signals target abort instead of master abort in link response packets to IO devices on detection of a master abort error condition.When IoMstAbortDis and D18F3x180[ChgMstAbortToNoErr] are both set, D18F3x180[ChgMstAbortToNoErr] takes precedence. 4 SyncPktPropDis: sync packet propagation disable. Read-write. Reset: 0. 1=Disables flooding of all outgoing links with sync packets when a sync packet is detected on an incoming link. Sync packets are propagated by default. 3 SyncPktGenDis: sync packet generation disable. Read-write. Reset: 0. 1=Disables flooding of all outgoing links with sync packets when a CRC error is detected on an incoming link. By default, sync packet generation for CRC errors is controlled through D18F0x84 [Link Control]. 2 SyncFloodOnDramUcEcc: sync flood on uncorrectable DRAM ECC error. Read-write. Reset: 0. 1=Enables flooding of all links with sync packets on detection of an uncorrectable ECC error. 1 CpuRdDatErrEn: CPU read data error log enable. Read-write. Reset: 0. 1=Enables reporting of read data errors (master aborts and target aborts) for data destined for the CPU on this node. This bit should be clear if read data error logging is enabled for the remaining error reporting blocks in the CPU. Logging the same error in more than one block may cause a single error event to be treated as a multiple error event and cause the CPU to enter shutdown. 0 Reserved. 422 42300 Rev 3.12 - July 14, 2015 BKDG for AMD Family 15h Models 10h-1Fh Processors D18F3x48 MCA NB Status Low Bits Description 31:0 MSR0000_0411[31:0] is an alias of D18F3x48. See MSR0000_0411. D18F3x4C MCA NB Status High Bits Description 31:0 MSR0000_0411[63:32] is an alias of D18F3x4C. See MSR0000_0411. D18F3x50 MCA NB Address Low Bits Description 31:0 MSR0000_0412[31:0] is an alias of D18F3x50. See MSR0000_0412[31:0]. D18F3x54 MCA NB Address High Bits Description 31:0 MSR0000_0412[63:32] is an alias of D18F3x54. See MSR0000_0412[63:32]. D18F3x64 Hardware Thermal Control (HTC) See 2.10.4.1 [PROCHOT_L and Hardware Thermal Control (HTC)]. If D18F3xE8[HtcCapable]=0 then this register is reserved. Bits 31 Description Reserved. 30:28 HtcPstateLimit: HTC P-state limit select. Read-write. Reset: Product-specific. Specifies the P-state limit of all cores when in the HTC-active state. This field uses hardware P-state numbering and is not changed on a write if the value written is greater than D18F3xDC[HwPstateMaxVal]. No P-state limit is applied if the value written is less than D18F4x15C[NumBoostStates]. See 2.10.4.1 [PROCHOT_L and Hardware Thermal Control (HTC)] and 2.5.3.1.2.2 [Hardware P-state Numbering]. 27:24 HtcHystLmt: HTC hysteresis. Read-write. Reset: Product-specific. The processor exits the HTCactive state when Tctl is less than HTC temperature limit (HtcTmpLmt) minus HTC hysteresis (HtcHystLmt). Bits Description Fh-0h <HtcHystLmt*0.5> 23 HtcSlewSel: HTC slew-controlled temperature select. Read-write. Reset: 0. 1=HTC logic is driven by the slew-controlled temperature, Tctl, specified in D18F3xA4 [Reported Temperature Control]. 0=HTC logic is driven by the measured control temperature with no slew controls. 22:16 HtcTmpLmt: HTC temperature limit. Read-write. Reset: Product-specific. The processor enters the HTC-active state when Tctl reaches or exceeds the temperature limit defined by this register. Bits Description 7Fh-00h <HtcTmpLmt*0.5 + 52> 423 42300 Rev 3.12 - July 14, 2015 BKDG for AMD Family 15h Models 10h-1Fh Processors 15:8 Reserved. 7 PslApicLoEn: P-state limit lower value change APIC interrupt enable. Read-write. Reset: 0. PslApicLoEn and PslApicHiEn enable interrupts using APIC330 [LVT Thermal Sensor] of each core when the active P-state limit in MSRC001_0061 [P-state Current Limit][CurPstateLimit] changes. PslApicLoEn enables the interrupt when the limit value becomes lower (indicating higher performance). PslApicHiEn enables the interrupt when the limit value becomes higher (indicating lower performance). 1=Enable interrupt. 6 PslApicHiEn: P-state limit higher value change APIC interrupt enable. Read-write. Reset: 0. See PslApicLoEn. 5 HtcActSts: HTC-active status. Read; set-by-hardware; write-1-to-clear. Reset: 0. This bit is set by hardware when the processor enters the HTC-active state. 4 HtcAct: HTC-active state. Read-only, updated-by-hardware. Reset: X. . 1=The processor is currently in the HTC-active state. 0=The processor is not in the HTC-active state. 3:1 0 Reserved. HtcEn: HTC enable. Read-write. Reset: 0. BIOS: IF (D18F3x64[HtcTmpLmt]==0) THEN 0. ELSE 1. ENDIF. 1=HTC is enabled; the processor is capable of entering the HTC-active state. D18F3x68 Software P-state Limit See 2.10.4.3 [Software P-state Limit Control]. If D18F3xE8[HtcCapable]=0 then this register is reserved. Bits 31 Description Reserved. 30:28 SwPstateLimit: software P-state limit select. Read-write. Reset: Product-specific. Specifies a Pstate limit for all cores. Uses hardware P-state numbering; see 2.5.3.1.2.2 [Hardware P-state Numbering]. Not changed on a write if the value written is greater than D18F3xDC[HwPstateMaxVal]. No Pstate limit is applied if the value written is less than D18F4x15C[NumBoostStates]. See SwPstateLimitEn. 27:6 Reserved. 5 4:0 SwPstateLimitEn: software P-state limit enable. Read-write. Reset: 0. 1=SwPstateLimit is enabled. Reserved. D18F3x6C Data Buffer Count Out of cold reset, the processor allocates a minimal number of buffers that is smaller than the default values in the register. BIOS must use D18F0x6C[RlsLnkFullTokCntImm] for the values in the register to take effect. This is necessary even if the values are unchanged from the default values. • To ensure deadlock free operation the following minimum buffer allocations are required: • D18F3x6C[UpRspDBC] >= 1. • D18F3x6C[DnReqDBC] >= 1. • D18F3x6C[UpReqDBC] >= 1. • D18F3x6C[DnRspDBC] >= 1. • If D18F0x84[IsocEn]=1: IsocRspDBC >= 1. • The total number of data buffers allocated in this register and D18F3x7C must satisfy the following equation: • D18F3x6C[UpReqDBC] + D18F3x6C[UpRspDBC] + D18F3x6C[DnReqDBC] + D18F3x6C[DnRsp- 424 42300 Rev 3.12 - July 14, 2015 BKDG for AMD Family 15h Models 10h-1Fh Processors DBC] + D18F3x6C[IsocRspDBC] + (IF (D18F3x7C[Sri2XbarFreeRspDBC]==0) THEN (D18F3x7C[Sri2XbarFreeXreqDBC]*2) ELSE D18F3x7C[Sri2XbarFreeXreqDBC] ENDIF) + D18F3x7C[Sri2XbarFreeRspDBC] <= 16. Bits 31 Description Reserved. 30:28 IsocRspDBC: isochronous response data buffer count. Read-write. Cold reset: 3. BIOS: 1. 27:19 Reserved. 18:16 UpRspDBC: upstream response data buffer count. Read-write. Cold reset: 2. BIOS: 1. 15:8 Reserved. 7:6 DnRspDBC: downstream response data buffer count. Read-write. Cold reset: 2. BIOS: 1. 5:4 DnReqDBC: downstream request data buffer count. Read-write. Cold reset: 1. BIOS: 1. 3 2:0 Reserved. UpReqDBC: upstream request data buffer count. Read-write. Cold reset: 2. BIOS: 2. D18F3x70 SRI to XBAR Command Buffer Count Out of cold reset, the processor allocates a minimal number of buffers that is smaller than the default values in the register. BIOS must use D18F0x6C[RlsLnkFullTokCntImm] for the values in the register to take effect. This is necessary even if the values are unchanged from the default values. • To ensure deadlock free operation the following minimum buffer allocations are required: • D18F3x70[UpRspCBC] >= 1. • D18F3x70[UpPreqCBC] >= 1. • D18F3x70[DnPreqCBC] >= 1. • D18F3x70[UpReqCBC] >= 1. • D18F3x70[DnReqCBC] >= 1. • D18F3x70[DnRspCBC] >= 1. • IF (D18F0x84[IsocEn]) THEN (D18F3x70[IsocReqCBC] >= 1). • IF (D18F0x84[IsocEn]) THEN (D18F3x70[IsocRspCBC] >= 1). • If D18F0x84[IsocEn]=1 and isochronous posted requests may be generated by the system: IsocPreqCBC >= 1 • The total number of SRI to XBAR command buffers allocated in this register and D18F3x7C must satisfy the following equation: • D18F3x70[IsocRspCBC] + D18F3x70[IsocPreqCBC] + D18F3x70[IsocReqCBC] + D18F3x70[UpRspCBC] + D18F3x70[DnPreqCBC] + D18F3x70[UpPreqCBC] + D18F3x70[DnReqCBC] + D18F3x70[DnRspCBC] + D18F3x70[UpReqCBC] + D18F3x7C[Sri2XbarFreeRspCBC] + D18F3x7C[Sri2XbarFreeXreqCBC] <= 48. Bits 31 Description Reserved. 30:28 IsocRspCBC: isoc response command buffer count. Read-write. Cold reset: 4. BIOS: 1. 27 Reserved. 26:24 IsocPreqCBC: isoc posted request command buffer count. Read-write. Cold reset: 1. BIOS: 0. 23 Reserved. 22:20 IsocReqCBC: isoc request command buffer count. Read-write. Cold reset: 5. BIOS: 1. 425 42300 Rev 3.12 - July 14, 2015 19 BKDG for AMD Family 15h Models 10h-1Fh Processors Reserved. 18:16 UpRspCBC: upstream response command buffer count. Read-write. Cold reset: 3. BIOS: 7. 15 Reserved. 14:12 DnPreqCBC: downstream posted request command buffer count. Read-write. Cold reset: 3. BIOS: 1. 11 Reserved. 10:8 UpPreqCBC: upstream posted request command buffer count. Read-write. Cold reset: 3. BIOS: 1. 7:6 DnRspCBC: downstream response command buffer count. Read-write. Cold reset: 2. BIOS: 1. 5:4 DnReqCBC: downstream request command buffer count. Read-write. Cold reset: 2. BIOS: 1. 3 2:0 Reserved. UpReqCBC: upstream request command buffer count. Read-write. Cold reset: 3. BIOS: 7. D18F3x74 XBAR to SRI Command Buffer Count Out of cold reset, the processor allocates a minimal number of buffers that is smaller than the default values in the register. BIOS must use D18F0x6C[RlsLnkFullTokCntImm] for the values in the register to take effect. This is necessary even if the values are unchanged from the default values. Table 178: XBAR Definitions Term SpqSize SrqSize PrbRsp Definition Probe command queue size. SpqSize = 12. SRQ (XBAR command and probe response to SRI) queue size. SrqSize = 44. SRQ entries hard allocated to probe responses. PrbRsp = 4. • To ensure deadlock free operation the following minimum buffer allocations are required: • D18F3x74[ProbeCBC] >= 2 • D18F3x74[UpReqCBC] >= 1 • D18F3x74[UpPreqCBC] >= 1 • IF (D18F0x84[IsocEn]) THEN (D18F3x74[IsocReqCBC] >= 1). • If D18F0x84[IsocEn]=1 and isochronous posted requests may be generated by the system: IsocPreqCBC >= 1 • The total number of XBAR to SRI command buffers allocated in this register and D18F3x7C must satisfy the following equation: D18F3x74[UpReqCBC] + D18F3x74[UpPreqCBC] + D18F3x74[DnReqCBC] + D18F3x74[DnPreqCBC] + D18F3x74[IsocReqCBC] + D18F3x74[IsocPreqCBC] + D18F3x74[DRReqCBC] + D18F3x7C[Xbar2SriFreeListCBC] + D18F3x7C[SrqExtFreeListBufCnt] + (D18F3x1A0[CpuCmdBufCnt] * NumOfCompUnits) + PrbRsp <= SrqSize • The total number of SPQ (probe command) buffers allocated must satisfy the following equation: • (D18F3x17C[SPQPrbFreeCBC] + D18F3x74[ProbeCBC]) <= SpqSize. Bits Description 31:28 DRReqCBC: display refresh request command buffer count. Read-write. Cold reset: 0. BIOS: 0. 27 Reserved. 26:24 IsocPreqCBC: isochronous posted request command buffer count. Read-write. Cold reset: 0. BIOS: 1. 426 42300 Rev 3.12 - July 14, 2015 BKDG for AMD Family 15h Models 10h-1Fh Processors 23:20 IsocReqCBC: isochronous request command buffer count. Read-write. Cold reset: 0. BIOS: 1. 19:16 ProbeCBC: probe command buffer count. Read-write. Cold reset: Ch. BIOS: 8. Description Bits 0h 0 buffers Ch-1h <ProbeCBC> buffers Fh-Dh Reserved. 15 Reserved. 14:12 DnPreqCBC: downstream posted request command buffer count. Read-write. Cold reset: 0. BIOS: 0. 11 Reserved. 10:8 UpPreqCBC: upstream posted request command buffer count. Read-write. Cold reset: 1. BIOS: 1. 7 6:4 3 2:0 Reserved. DnReqCBC: downstream request command buffer count. Read-write. Cold reset: 0. BIOS: 0. Reserved. UpReqCBC: upstream request command buffer count. Read-write. Cold reset: 1. BIOS: 1. D18F3x78 MCT to XBAR Buffer Count Out of cold reset, the processor allocates a minimal number of buffers that is smaller than the default values in the register. BIOS must use D18F0x6C[RlsLnkFullTokCntImm] for the values in the register to take effect. This is necessary even if the values are unchanged from the default values. • To ensure deadlock free operation the following minimum buffer allocations are required: ProbeCBC >= 1 RspCBC >= 1 RspDBC >= 2 RspDBC >= D18F2x11C[MctPrefReqLimit]+2 • To ensure deadlock free operation when online spare is enabled (D18F2x[5C:40]_dct[1:0][Spare] = 1) the following minimum buffer allocation is required: RspCBC >= Dh • The total number of command buffers allocated in this register must satisfy the following equation: (D18F3x78[ProbeCBC] + D18F3x78[RspCBC]) <= 32 Bits Description 31:22 Reserved. 21:16 RspDBC: response data buffer count. Read-write. Cold reset: 20h. Bits Description 01h-00h Reserved 02h 2 Buffers 1Fh-03h <RspDBC> Buffers 20h 32 Buffers 3Fh-21h Reserved 15:13 Reserved. 12:8 ProbeCBC: probe command buffer count. Read-write. Cold reset: Ch. 7:5 Reserved. 4:0 RspCBC: response command buffer count. Read-write. Cold reset: 14h. 427 42300 Rev 3.12 - July 14, 2015 BKDG for AMD Family 15h Models 10h-1Fh Processors D18F3x7C Free List Buffer Count Out of cold reset, the processor allocates a minimal number of buffers that is smaller than the default values in the register. BIOS must use D18F0x6C[RlsLnkFullTokCntImm] for the values in the register to take effect. This is necessary even if the values are unchanged from the default values. See D18F3x6C and D18F3x70. • To ensure deadlock free operation the following minimum buffer allocations are required: • IF (D18F3x7C[Sri2XbarFreeRspCBC]==0) THEN (D18F3x7C[Sri2XbarFreeXreqCBC]>2). • IF (D18F3x7C[Sri2XbarFreeRspCBC]!=0) THEN (D18F3x7C[Sri2XbarFreeRspCBC]>2). • IF (D18F3x7C[Sri2XbarFreeRspDBC]==0) THEN (D18F3x7C[Sri2XbarFreeXreqDBC]>2). • IF (D18F3x7C[Sri2XbarFreeRspDBC]!=0) THEN (D18F3x7C[Sri2XbarFreeRspDBC]>2). • D18F3x7C[Xbar2SriFreeListCBC] >= (D18F3x1A0[CpuToNbFreeBufCnt] * NumOfCompUnits) + 2. Bits 31 Description Reserved. 30:28 Xbar2SriFreeListCBInc: XBAR to SRI free list command buffer increment. Read-write. Cold reset: 0. This field is used to add buffers to the free list pool if they are reclaimed from hard allocated entries without having to go through warm reset. This field may only be programmed after buffers have been allocated and released via D18F0x6C[RlsLnkFullTokCntImm]. 27 Reserved. 26:23 ExtSrqFreeList: extend SRQ freelist tokens. Read-write. Cold reset: 8h. BIOS: 8h. Can only be used by requests from cores to DRAM. 22:20 Sri2XbarFreeRspDBC: SRI to XBAR free response data buffer count. Read-write. Cold reset: 0. BIOS: 0. 19:16 Sri2XbarFreeXreqDBC: SRI to XBAR free request and posted request data buffer count. Readwrite. Cold reset: 3h. BIOS: 5. If Sri2XbarFreeRspDBC=0h, then these buffers are shared between requests, responses and posted requests and the number of buffers allocated is two times the value of this field. 15:12 Sri2XbarFreeRspCBC: SRI to XBAR free response command buffer count. Read-write. Cold reset: Bh. BIOS: 0. 11:8 Sri2XbarFreeXreqCBC: SRI to XBAR free request and posted request command buffer count. Read-write. Cold reset: Bh. BIOS: Eh. If Sri2XbarFreeRspCBC=0h, then these buffers are shared between requests, responses and posted requests and the number of buffers allocated is two times the value of this field. 7:5 Reserved. 4:0 Xbar2SriFreeListCBC: XBAR to SRI free list command buffer count. Read-write. Cold reset: Product-specific. BIOS: 18h. D18F3x[84:80] ACPI Power State Control This block consists of eight identical 8-bit registers, one for each System Management Action Field (SMAF) code associated with STPCLK assertion commands from the link. Refer to the descriptions below for the associated ACPI state and system management actions for each of the 8 SMAF codes. The SmafAct fields specify the system management actions taken when the corresponding SMAF code is received. For instance, a SMAF code of 5 results in the power management actions specified by SmafAct5. Some ACPI states and associated SMAF codes may not be supported in certain conditions. See 2.5 [Power Management] for which states are supported. 428 42300 Rev 3.12 - July 14, 2015 BKDG for AMD Family 15h Models 10h-1Fh Processors When a link STPCLK assertion command is received by the processor, the power management commands specified by the register with the corresponding SMAF code are invoked. When the STPCLK deassertion command is received by the processor, the processor returns into the operational state. In multi-node systems, these registers should be programmed identically in all nodes. Table 179: SMAF Action Definition Register SmafAct D18F3x84[31:24] SmafAct7 ACPI state C1 D18F3x84[23:16] SmafAct6 S4/S5 D18F3x84[15:8] D18F3x84[7:0] SmafAct5 SmafAct4 S3 D18F3x80[31:24] SmafAct3 S1 D18F3x80[23:16] SmafAct2 - D18F3x80[15:8] SmafAct1 D18F3x80[7:0] SmafAct0 C1E, or Link init. C2 Description Initiated when a Halt instruction is executed by processor. This does not involve the interaction with the SMC, therefore the SMC is required to never send STPCLK assertion commands with SMAF=7h. Initiated by a processor access to the ACPI-defined PM1_CNTa register. Reserved. Initiated by a processor access to the ACPI-defined PM1_CNTa register. Initiated by a processor access to the ACPI-defined PM1_CNTa register. Initiated by a processor for NB P-state changes. See 2.5.4.1 [NB P-states]. Initiated by an access to the ACPI-defined P_LVL3 register. Initiated by a processor access to the ACPI-defined P_LVL2 register. D18F3x80 ACPI Power State Control Low Reset: 0000_0000h. Read-write. Bits Description 31:29 ClkDivisorSmafAct3. See: ClkDivisorSmafAct0. 28:27 Reserved. 26 NbGateEnSmafAct3. See: NbGateEnSmafAct0. 25 NbLowPwrEnSmafAct3. See: NbLowPwrEnSmafAct0. 24 CpuPrbEnSmafAct3. See: CpuPrbEnSmafAct0. 23:21 ClkDivisorSmafAct2. See: ClkDivisorSmafAct0. 20:19 Reserved. 18 NbGateEnSmafAct2. See: NbGateEnSmafAct0. 17 NbLowPwrEnSmafAct2. See: NbLowPwrEnSmafAct0. 16 CpuPrbEnSmafAct2. See: CpuPrbEnSmafAct0. 15:13 ClkDivisorSmafAct1. See: ClkDivisorSmafAct0. 12:11 Reserved. 10 NbGateEnSmafAct1. See: NbGateEnSmafAct0. 9 NbLowPwrEnSmafAct1. See: NbLowPwrEnSmafAct0. 429 42300 Rev 3.12 - July 14, 2015 8 BKDG for AMD Family 15h Models 10h-1Fh Processors CpuPrbEnSmafAct1. See: CpuPrbEnSmafAct0. 7:5 ClkDivisorSmafAct0: clock divisor. Read-write. Specifies the core clock frequency while in the low-power state. This divisor is relative to the current FID frequency, or: • 100 MHz * (10h + MSRC001_00[6B:64][CpuFid]) of the current P-state specified by MSRC001_0063[CurPstate]. If MSRC001_00[6B:64][CpuDid] of the current P-state indicates a divisor that is deeper than specified by this field, then no frequency change is made when entering the low-power state associated with this register. Bits Description Bits Description 000b /1 100b /16 001b /2 101b /128 010b /4 110b /512 011b /8 111b Turn off clocks 4:3 Reserved. 2 NbGateEnSmafAct0: northbridge gate enable. Read-write. This bit does not control hardware. NbLowPwrEn is required to be set if this bit is set. 1 NbLowPwrEnSmafAct0: Northbridge low-power enable. Read-write. 1=The NB clock is ramped down to the divisor specified by D18F3xD4[NbClkDiv] and DRAM is placed into self-refresh mode when LDTSTOP_L is asserted while in the low-power state. 0 CpuPrbEnSmafAct0: CPU direct probe enable. Read-write. Specifies how probes are handled while in the low-power state. 0=When the probe request comes into the NB, the core clock is brought up to the COF (based on the current P-state), all outstanding probes are completed, the core waits for a hysteresis time based on D18F3xD4[ClkRampHystSel], and then the core clock is brought down to the frequency specified by ClkDivisor. 1=The core clock does not change frequency; the probe is handled at the frequency specified by ClkDivisor; this may only be set if: • ClkDivisor specifies a divide-by 1, 2, 4, 8, or 16 and NbCof <= 3.2 GHz • ClkDivisor specifies a divide-by 1, 2, 4, or 8 and NbCof >= 3.4 GHz This bit also specifies functionality of the timer used for cache flushing during halt. See D18F3xDC[CacheFlushOnHaltTmr]. • If D18F3x[84:80][CpuPrbEnSmafAct7]=0 and D18F3xDC[IgnCpuPrbEn]=0, only the time when the core is halted and has its clocks ramped up to service probes is counted. • If D18F3x[84:80][CpuPrbEnSmafAct7]=1 or D18F3xDC[IgnCpuPrbEn]=1, all of the time the core is halted is counted. D18F3x84 ACPI Power State Control High Reset: 0000_0000h. Read-write. Bits Description 31:29 ClkDivisorSmafAct7. See: D18F3x80[ClkDivisorSmafAct0]. 28:27 Reserved. 26 NbGateEnSmafAct7. See: D18F3x80[NbGateEnSmafAct0]. 25 NbLowPwrEnSmafAct7. See: D18F3x80[NbLowPwrEnSmafAct0]. 24 CpuPrbEnSmafAct7. See: D18F3x80[CpuPrbEnSmafAct0]. 23:21 ClkDivisorSmafAct6. See: D18F3x80[ClkDivisorSmafAct0]. 20:19 Reserved. 18 NbGateEnSmafAct6. See: D18F3x80[NbGateEnSmafAct0]. 430 42300 Rev 3.12 - July 14, 2015 BKDG for AMD Family 15h Models 10h-1Fh Processors 17 NbLowPwrEnSmafAct6. See: D18F3x80[NbLowPwrEnSmafAct0]. 16 CpuPrbEnSmafAct6. See: D18F3x80[CpuPrbEnSmafAct0]. 15:13 ClkDivisorSmafAct5. See: D18F3x80[ClkDivisorSmafAct0]. 12:11 Reserved. 10 NbGateEnSmafAct5. See: D18F3x80[NbGateEnSmafAct0]. 9 NbLowPwrEnSmafAct5. See: D18F3x80[NbLowPwrEnSmafAct0]. 8 CpuPrbEnSmafAct5. See: D18F3x80[CpuPrbEnSmafAct0]. 7:5 ClkDivisorSmafAct4. See: D18F3x80[ClkDivisorSmafAct0]. BIOS: 111b. 4:3 Reserved. 2 NbGateEnSmafAct4. See: D18F3x80[NbGateEnSmafAct0]. 1 NbLowPwrEnSmafAct4. See: D18F3x80[NbLowPwrEnSmafAct0]. BIOS: 1. 0 CpuPrbEnSmafAct4. See: D18F3x80[CpuPrbEnSmafAct0]. D18F3x88 NB Configuration Low Bits Description 31:0 MSRC001_001F[31:0] is an alias of D18F3x88. See MSRC001_001F. D18F3x8C NB Configuration High Bits Description 31:0 MSRC001_001F[63:32] is an alias of D18F3x8C. See MSRC001_001F. D18F3xA0 Power Control Miscellaneous Bits 31 Description CofVidProg: COF and VID of P-states programmed. Read-only. Reset: Product-specific. 1=Out of cold reset, the VID, FID, and DID values of the P-state registers specified by MSRC001_0071[StartupPstate] and D18F5x174[StartupNbPstate] have been applied to the processor. 0=Out of cold reset, the boot VID is applied to all processor power planes, the NB clock plane is set to 800 MHz (with a FID of 04h=800 MHz and a DID of 0b) and core CPU clock planes are set to 800 MHz (with a FID of 00h=1.6 GHz and a DID of 1h). Registers containing P-state information such as FID, DID, and VID values are valid out of cold reset independent of the state of D18F3xA0[CofVidProg]. BIOS must transition the processor to a valid P-state out of cold reset when D18F3xA0[CofVidProg]=0. See 2.5.3.1.7 [BIOS Requirements for Core P-state Initialization and Transitions]. 30:28 Reserved. 27:16 ConfigId: Configuration identifier. Read-only. Reset: Product-specific. Specifies the configuration ID associated with the product. 15 Reserved. 431 42300 Rev 3.12 - July 14, 2015 14 BKDG for AMD Family 15h Models 10h-1Fh Processors Svi2HighFreqSel: SVI high frequency select. Read-write. Cold reset: 0. BIOS: 1. 0=3.4 MHz. 1=20 MHz. Writes to this field take effect at the next SVI command boundary. If 20 MHz is supported by the VRM, BIOS should program this to 1 prior to any VID transitions. Once this bit is set, it should not be cleared until the next cold reset. See 2.5.1.1 [Serial VID Interface]. 13:11 PllLockTime: PLL synchronization lock time. Read-write. Reset: 0. BIOS: 001b. If a P-state change occurs that applies a new FID to the PLL, this field specifies the time required for the PLL to lock to the new frequency. Bits Description Bits Description 000b 1 us 100b 8 us 001b 2 us 101b 16 us 010b 3 us 110b Reserved 011b 4 us 111b Reserved 10:9 Reserved. 8 PsiVid[7]. Read-write. Reset: 0. BIOS: 2.5.1.3.1.1. See PsiVid[6:0]. 7 PsiVidEn: PSI_L VID enable. Read-write. Reset: 0. BIOS: 2.5.1.3.1.1. This bit specifies how PSI_L is controlled. This signal may be used by the voltage regulator to improve efficiency while in reduced power states. 1=Control over the PSI_L signal is as specified by the PsiVid field of this register. 0=PSI_L is always high. See 2.5.1.3.1 [PSIx_L Bit]. 6:0 PsiVid[6:0]: PSI_L VID threshold. Read-write. Reset: 0. BIOS: 2.5.1.3.1.1. PsiVid[7:0] = {PsiVid[7], PsiVid[6:0]}. When enabled by PsiVidEn, PsiVid[7:0] specifies the threshold value of the VID code generated by the processor, which in turn determines the state of PSI0_L. When the VID code generated by the processor is less than PsiVid[7:0] (i.e., the VID code is specifying a higher voltage level than the PsiVid-specified voltage level), then PSI0_L is high; when the VID code is greater than or equal to PsiVid[7:0], PSI0_L is driven low. See 2.5.1.3.1 [PSIx_L Bit]. D18F3xA4 Reported Temperature Control See 2.10.1 [The Tctl Temperature Scale] and 2.10.2 [Temperature Slew Rate Control]. Bits Description 31:21 CurTmp: current temperature. IF (D18F3xA4[CurTmpTjSel]==11b) THEN Read-write. ELSE Read-only, updated-by-hardware. ENDIF. Reset: X. Provides the current control temperature, Tctl, after the slew-rate controls have been applied. IF (D18F3xA4[CurTmpTjSel]!=11b THEN Bits Description 000h 0 001h 0.125 7FEh-002h <CurTmp*0.125> 7FFh 255.875 ELSE Bits Description 000h -49 001h -48.875 7FEh-002h <(CurTmp*0.125)-49> 7FFh 206.875 ENDIF. 432 42300 Rev 3.12 - July 14, 2015 20 BKDG for AMD Family 15h Models 10h-1Fh Processors TcenPwrDnCc6En: TCEN power down CC6 Enable. Read-write. Cold reset: 0. 1=A TCEN remains powered down if the associated compute-unit is in CC6 when the TCEN power down delay timer expires. 0=TCEN powers up immediately after TCEN power down delay timer expires. Field has no effect if CC6-based TCEN power down feature is disabled by hardware. 19:18 Reserved. 17:16 CurTmpTjSel: Current temperature select. Read-write. Reset: 00. These bits may be used for diagnostic software. Bits Description 00b CurTmp provides the read-only Tctl value. 01b Reserved. 10b Reserved. 11b CurTmp is a read-write register that specifies a value used to create Tctl. The two LSB’s are read-only zero. 15:13 Reserved. 12:8 PerStepTimeDn: per step time down. Read-write. Cold reset: 18h. BIOS: 0Fh. Specifies the time that measured temperaturemust remain below Tctl before applying a 0.125 downward step. See PerStepTimeUp for encodings. 7 TmpSlewDnEn: temperature slew downward enable. Read-write. Cold reset: 0. BIOS: 1. 1=Downward slewing enabled. 0=Downward slewing disabled. 6:5 TmpMaxDiffUp: temperature maximum difference up. Read-write. Cold reset: 00b. BIOS: 11b. Specifies the maximum difference, (measured temperature - Tctl), when Tctl immediatly updates to the measured temperature. Bits Description 00b 0.0 (disable upward slew) 01b 1.0 10b 3.0 11b 9.0 4:0 PerStepTimeUp: per 1/8th degree step time up. Read-write. Cold reset: 00h. BIOS: 0Fh. Specifies the time that measured temperaturemust remain above Tctl before applying a 0.125 upward step.. It is encoded as follows: Definition Bits 1Fh-00h <(PerStepTimeUp[2:0] + 1) * 10^PerStepTimeUp[4:3]> ms, ranging from 1 ms to 8000 ms. D18F3xA8 Pop Up and Down P-states Bits Description 31:29 PopDownPstate. Read-write. Reset: D18F3xDC[HwPstateMaxVal]. BIOS: D18F3xDC[HwPstateMaxVal]. Specifies the pop-down P-state number. This field uses hardware P-state numbering. See 2.5.3.2.3.3 [Core C6 (CC6) State]. 28:0 Reserved. 433 42300 Rev 3.12 - July 14, 2015 BKDG for AMD Family 15h Models 10h-1Fh Processors D18F3xD4 Clock Power/Timing Control 0 Bits 31 Description NbClkDivApplyAll. Read-write. Cold reset: 0. BIOS: 1. See NbClkDiv. 30:28 NbClkDiv: NB clock divisor. Read-write. Cold reset: Product-specific. BIOS: 100b. Specifies the NB CLK divisor associated with D18F3x[84:80] [ACPI Power State Control][NbLowPwrEn]. This divisor is applied while LDTSTOP is asserted if the corresponding core CLK divisor, D18F3x[84:80][ClkDivisor], is set to “turn off clocks” or if NBClkDivApplyAll=1; otherwise, the divisor specified by D18F3x[84:80][ClkDivisor] is applied. This divisor is relative to the current NB FID frequency, or: • 100 MHz * (4 + D18F5x1[6C:60][NbFid]). If D18F5x1[6C:60][NbDid] of the current P-state indicates a divisor that is lower than specified by this field, then no NB frequency change is made when entering the low-power state associated with this register (i.e., if this field specifies a divide-by 1 and the DID is divide-by 2, then the divisor remains 2 while in the low-power state). This field is encoded as follows: Bits Description Bits Description 000b Divide-by 1 100b Divide-by 16 001b Divide-by 2 101b Reserved 010b Divide-by 4 110b Reserved 011b Divide-by 8 111b Reserved 27:24 PowerStepUp. Read-write. Cold reset: 0000b. BIOS: 1000b. Specifies the rate at which blocks of compute unit and NB logic are gated on while the processor transitions from a quiescent state to an active state as part of a power management state transition. There are about 15 steps in this transition for each compute unit and about 5 steps for the NB for the PowerStepDown and PowerStepUp transitions. So the total transition time for a single compute unit is about 15 times the time specified by PowerStepDown and PowerStepUp and the transition time for the NB is about 5 times the time specified by PowerStepDown and PowerStepUp. Use of longer transition times may help reduce voltage transients associated with power state transitions. The bits for PowerStepUp and PowerStepDown are encoded as follows: Description Bits Description Bits 0000b Reserved. 1000b 50 ns 0001b Reserved. 1001b Reserved. 0010b Reserved. 1010b Reserved. 0011b 100 ns 1011b Reserved. 0100b 90 ns 1100b Reserved. 0101b 80 ns 1101b Reserved. 0110b 70 ns 1110b Reserved. 0111b 60 ns 1111b Reserved. • If PowerStepDown or PowerStepUp are programmed to greater than 50 ns, then the value applied to the NB is clipped to 50 ns. The compute unit steps are not clipped. 23:20 PowerStepDown. Read-write. Cold reset: 0000b. BIOS: 1000b. This specifies the rate at which blocks of compute unit and NB logic are gated off while the processor transitions from an active state to a quiescent state as part of a power management state transition. 19:15 Reserved. 14 CacheFlushImmOnAllHalt: cache flush immediate on all halt. Read-write. Cold reset: 0. BIOS: 0. 1=Flush the caches immediately when all cores in a package have halted. The following condition must be true in order for the caches to be flushed: • D18F4x11[C:8][CacheFlushEn]=1 for the corresponding C-state action field on all cores. 434 42300 Rev 3.12 - July 14, 2015 BKDG for AMD Family 15h Models 10h-1Fh Processors 13 Reserved. 12 ClkRampHystCtl: clock ramp hysteresis control. Read-write. Cold reset: 0. Specifies the time base for ClkRampHystSel when D18F3x[84:80][CpuPrbEn]=0. 0=320 ns. 1=1.28 us. 11:8 ClkRampHystSel: clock ramp hysteresis select. Read-write. Cold reset: 0h. BIOS: Fh. When the core(s) are in the stop-grant or halt state and a probe request is received, the core clock may need to be brought up to service the probe. • If D18F3x[84:80][CpuPrbEn]=0 for the low-power state, then this field specifies how long the core clock is left up to service additional probes before being brought back down. Each time a probe request is received, the hysteresis timer is reset such that the period of time specified by this field must expire with no probe request before the core clock is brought back down. The hysteresis time is encoded as (the time base specified by D18F3xD4[ClkRampHystCtl]) * (1 + ClkRampHystSel). • If D18F3x[84:80][CpuPrbEn]=1 for the low-power state, and for requests to change core P-states, then this field specifies a fixed amount of time to allow for probes to be serviced after completing the transition of each core. If, for example, two cores enter stop-grant or halt at the same time, then (1) the first core would complete the transition to the low power state, (2) probe traffic would be serviced for the time specified by this field, (3) the second core would complete the transition to the low power state, and (4) probe traffic would be seviced for the time specified by this field (and afterwards, until the next power state transition). For this purpose, values range from 0h=40 ns to Fh=640 ns, encoded as 40 ns * (1 + ClkRampHystSel). 7:6 Reserved. 5:0 MaxSwPstateCpuCof:maximum software P-state core COF. Read-only. Cold reset: Product-specific. Specifies the maximum CPU COF supported by the processor in a software P-state. The maximum frequency is 100 MHz * MaxSwPstateCpuCof, if MaxSwPstateCpuCof is greater than zero; if MaxSwPstateCpuCof = 00h, then there is no frequency limit. Any attempt to change a software P-state CPU COF to a frequency greater than specified by this field is ignored. See2.5.3.1.2.1 [Software P-state Numbering] . D18F3xD8 Clock Power/Timing Control 1 See 2.5.1.4 [Voltage Transitions]. Bits Description 31:7 Reserved. 6:4 VSRampSlamTime. Read-write. Cold reset: 000b. BIOS: Voltage Ramp Time1. Specifies the time the processor waits for voltage transitions to complete before beginning an additional voltage change or a frequency change. Wait time = (VSRampSlamTime / 12.5mV) * ABS(destination voltage - current voltage). Wait time = (VSRampSlamTime / 15mV) * ABS(destination voltage - current voltage).BitsDescriptionBits Description 000b 5.00 us 100b 2.00 us 001b 3.75 us 101b 1.50 us 010b 3.00 us 110b 1.20 us 011b 2.40 us 111b 1.00 us 1. Voltage Ramp Time = The maximum time to change VDD or VDDNB 15mV rounded to the next higher encoding. 3:0 Reserved. 435 42300 Rev 3.12 - July 14, 2015 BKDG for AMD Family 15h Models 10h-1Fh Processors D18F3xDC Clock Power/Timing Control 2 Bits Description 31:27 Reserved. 26 IgnCpuPrbEn: ignore CPU probe enable. Read-write. Cold reset: 0. BIOS: 0. See D18F3x[84:80][CpuPrbEn] and D18F4x11[C:8][CpuPrbEn]. 25:19 CacheFlushOnHaltTmr: cache flush on halt timer. Read-write. Cold reset: 00h. BIOS: 14h. Specifies how long each core needs to stay in a C-state before it flushes its caches. See CacheFlushOnHaltCtl, D18F3x[84:80][CpuPrbEn], and D18F4x11[C:8][CacheFlushTmrSel]. Bits Description 00h 5.12 us 7Fh-01h (<CacheFlushOnHaltTmr> * 10.24us) - 5.12us <= Time <= <CacheFlushOnHaltTmr> * 10.24 us 18:16 CacheFlushOnHaltCtl: cache flush on halt control. Read-write. Cold reset: 000b. BIOS: 111b.If CacheFlushOnHaltCtl != 0, then cache flush on halt is enabled. CacheFlushOnHaltCtl also specifies what core clock divisor is used after the caches have been flushed. See D18F4x11[C:8][CacheFlushTmrSelCstAct0]. Description Bits 000b /1. 001b /2 010b /4 011b /8 100b /16 101b Reserved 110b Reserved 111b Turn off clocks See D18F3x[84:80] and D18F4x11[C:8] for clock divisor specifications that are in effect during a Cstate before the caches have been flushed. See 2.5.3.2.3.1 [C-state Probes and Cache Flushing]. 15 Reserved. 14:12 NbsynPtrAdj: NB/core synchronization FIFO pointer adjust. Read-write. Cold reset: 000b. BIOS: 101b. Changes to this field take effect after any of the following events: • Warm reset. • All compute units perform a P-state transition. • An NB P-state transition. There is a synchronization FIFO between the NB clock domain and core clock domains. At cold reset, the read pointer and write pointer for each of these FIFOs is positioned conservatively, such that FIFO latency may be greater than is necessary. This field may be used to position the read pointer and write pointer of each FIFO closer to each other such that latency is reduced. Each increment of this field represents one clock cycle of whichever is the slower clock (longer period) between the NB clock and the core clock. Values less than the recommended value are allowed; values greater than the recommended value are illegal. Description Bits 5h-0h Position the read pointer <NbsynPtrAdj> clock cycles closer to the write pointer 7h-6h Reserved 11 Reserved. 436 42300 Rev 3.12 - July 14, 2015 BKDG for AMD Family 15h Models 10h-1Fh Processors 10:8 HwPstateMaxVal: P-state maximum value. Read-write. IF ((D18F3xE8[HtcCapable]==1) && (D18F3x64[HtcTmpLmt]!=0) && (D18F3x64[HtcPstateLimit] > HwPstateMaxVal)) THEN BIOS: D18F3x64[HtcPstateLimit]. ENDIF. Cold reset: specified by the reset state of MSRC001_00[6B:64][PstateEn]; the cold reset value is the highest P-state number corresponding to the MSR in which PstateEn is set (e.g., if MSRC001_0064 and MSRC001_0065 have this bit set and the others do not, then PstateMaxVal=1; if MSRC001_0064 has this bit set and the others do not, then PstateMaxVal=0). This specifies the highest P-state value (lowest performance state) supported by the hardware. This field must not be written to a value less (higher performance) than MSRC001_0071[CurPstateLimit]. See MSRC001_0061[PstateMaxVal]. This field uses hardware Pstate numbering. See 2.5.3.1.2.2 [Hardware P-state Numbering]. 7:0 Reserved. D18F3xE4 Thermtrip Status Bits 31 Description SwThermtp: software THERMTRIP. Write-1-only; cleared-by-hardware. Reset: 0. Writing a 1 to this bit position induces a THERMTRIP event. This bit returns 0 when read. This is a diagnostic bit, and it should be used for testing purposes only. 30:6 Reserved. 5 ThermtpEn: THERMTRIP enable. Read-only. Reset: Product-specific. 1=The THERMTRIP state is supported. See 2.10.4.4 [THERMTRIP]. 4 Reserved. 3 ThermtpSense: THERMTRIP sense. Read-only. Cold reset: 0. 1=The processor temperature exceeded the THERMTRIP value (regardless as to whether the THERMTRIP state is enabled). This bit is also set when the diagnostic bit SwThermtp = 1. 2 Reserved. 1 Thermtp: THERMTRIP. Read-only. Cold reset: 0. 1=The processor has entered the THERMTRIP state. 0 Reserved. D18F3xE8 Northbridge Capabilities Read-only. Value: Product-specific. Unless otherwise specified, 1=The feature is supported by the processor; 0=The feature is not supported. Bits Description 31:29 Reserved. 28:26 Reserved. 25 L3Capable. 1=Specifies that an L3 cache is present. See CPUID Fn8000_0006_EDX. 24 MemPstateCap: memory P-state capable. 23:20 Reserved. 19 x2Apic: x2APIC capability. 18:16 Reserved. 15 Reserved. 437 42300 Rev 3.12 - July 14, 2015 14 BKDG for AMD Family 15h Models 10h-1Fh Processors MultVidPlane: multiple VID plane capable. Value: 1. 13:12 Reserved. 11 Reserved. 10 HtcCapable: HTC capable. This affects D18F3x64 and D18F3x68. 9 SvmCapable: SVM capable. 8 MctCap: memory controller (on the processor) capable. Value: 1. 7:5 Reserved. 4 ChipKill: chipkill ECC capable. 3 ECC: ECC capable. 2 EightNode: Eight-node multi-processor capable. 1 DualNode: Dual-node multi-processor capable. 0 Reserved. D18F3xF0 DEV Capability Header Register The DEV secure loader function is configured through D18F3xF4 and D18F3xF8. The register number (i.e., the number that follows F8_x in the register mnemonic) is specified by D18F3xF4[DevFunction]. Access to this register is accomplished as follows: • Reads: Write the register number to D18F3xF4[DevFunction]. Read the register contents from D18F3xF8. • Writes: Write the register number to D18F3xF4[DevFunction]. Write the register contents to D18F3xF8. IF (D18F3xE8[SvmCapable]==0) THEN Bits Description 31:0 Reserved. ELSE Bits Description 31:19 Reserved. Reset: 0002h. 18:16 CapType: DEV capability block type. Read-only. Reset: 000b. Specifies the layout of the Capability Block. 15:8 CapPtr: capability pointer. Read-only. Reset: 00h. Indicates that this is the last capability block. 7:0 CapId: capability ID. Read-only. Reset: 0Fh. Indicates a DEV capability block. ENDIF. D18F3xF4 DEV Function Register Reset: 0000_0000h. IF (D18F3xE8[SvmCapable]==0) THEN Bits Description 31:0 Reserved. 438 42300 Rev 3.12 - July 14, 2015 BKDG for AMD Family 15h Models 10h-1Fh Processors ELSE Bits Description 31:16 Reserved. 15:8 DevFunction. Read-write. See D18F3xF0. The valid value for this field is 4h. Writing invalid values may result in undefined behavior. 7:0 Reserved. ENDIF. D18F3xF8 DEV Data Port Reset: 0000_0000h. See D18F3xF0 for details about this port. IF (D18F3xE8[SvmCapable]==0) THEN Bits Description 31:0 Reserved. ELSE Bits Description 31:0 DevData. Read-write. See D18F3xF8_x4. ENDIF. D18F3xF8_x4 DEV Secure Loader Control Register Reset: 0000_0000h. Bits Description 31:9 Reserved. 8 7:6 5 4:3 2 1:0 SecureGfxMode: secure graphics mode. RAZ; write-0-only. Secure graphics mode is entered when an SKINIT instruction is executed. In this mode, all accesses to memory except for accesses to the frame buffer from the GPU can be checked by the IOMMU if IOMMU is enabled. This bit always reads as a zero so software should check if SlDev or IoDis is set to determine whether the processor is in secure graphics mode. Reserved. SlDev: secure loader DEV protection enable. Read; write-0-only; set-by-hardware. This bit is set when an SKINIT instruction is executed. 1=The memory region associated with the SKINIT instruction is protected from DMA access. Reserved. IoDis: upstream IO disable. Read-write; set-by-hardware. This bit is set when an SKINIT instruction is executed. 1=Upstream IO-space accesses are regarded as DEV protection violations. Reserved. D18F3xFC CPUID Family/Model CPUID Fn0000_0001_EAX, CPUID Fn8000_0001_EAX are an alias of D18F3xFC. 439 42300 Rev 3.12 - July 14, 2015 Bits BKDG for AMD Family 15h Models 10h-1Fh Processors Description 31:28 Reserved. 27:20 ExtFamily: extended family. Read-only. Reset: 06h. 19:16 ExtModel: extended model. Read-only. Value: 1h. 15:12 Reserved. 11:8 BaseFamily. Read-only. Reset: Fh. 7:4 BaseModel. Read-only. Value: Product-specific. 3:0 Stepping. Read-only. Value: Product-specific. D18F3x140 SRI to XCS Token Count Out of cold reset, the processor allocates a minimal number of buffers that is smaller than the default values in the register. BIOS must use D18F0x6C[RlsLnkFullTokCntImm] for the values in the register to take effect. This is necessary even if the values are unchanged from the default values. D18F3x140, D18F3x144, and D18F3x148 specify the number of XCS (XBAR command scheduler) entries assigned to each virtual channel within each source port. See 2.8 [Northbridge (NB)]. The totals of SRI, MCT and the links must not exceed the number of XCS entries. The default totals are: • SRI: 22 • MCT: 10 • Upstream channel: 8. • Total: 40, which is the total number of entries supported by XCS. The defaults for D18F3x140, D18F3x148, do not allocate any tokens in the isochronous channel. If isochronous flow control mode (IFCM) is enabled (D18F0x84[IsocEn]), then the XCS token counts must be changed. • If IFCM is enabled, then D18F3x140[IsocReqTok and IsocRspTok] must each be non-zero. If isochronous posted requests may be generated in the system, then D18F3x140[IsocPreqTok] must also be non-zero. • If an IOMMU is present, D18F3x148[IsocReqTok] must be non-zero. Bits Description 31:24 Reserved. 23:20 FreeTok: free tokens. Read-write. Cold reset: Ah. BIOS: Ah. The number of free tokens must always be greater than or equal to 2 to ensure deadlock free operation. 19:18 Reserved. 17:16 IsocRspTok: isochronous response tokens. Read-write. Cold reset: 0. BIOS: 1. 15:14 IsocPreqTok: isochronous posted request tokens. Read-write. Cold reset: 0. BIOS: 0. 13:12 IsocReqTok: isochronous request tokens. Read-write. Cold reset: 0.BIOS: 1. 11:10 DnRspTok: downstream response tokens. Read-write. Cold reset: 1. BIOS: 1. 9:8 UpRspTok: upstream response tokens. Read-write. Cold reset: 3. BIOS: 1. 7:6 DnPreqTok: downstream posted request tokens. Read-write. Cold reset: 1. BIOS: 1. 5:4 UpPreqTok: upstream posted request tokens. Read-write. Cold reset: 1. BIOS: 1. 3:2 DnReqTok: downstream request tokens. Read-write. Cold reset: 1. BIOS: 1. 1:0 UpReqTok: upstream request tokens. Read-write. Cold reset: 3. BIOS: 1. 440 42300 Rev 3.12 - July 14, 2015 BKDG for AMD Family 15h Models 10h-1Fh Processors D18F3x144 MCT to XCS Token Count See D18F3x140. Bits Description 31:8 Reserved. 7:4 ProbeTok: probe tokens. Read-write. Cold reset: 7h. BIOS: 7h. 3:0 RspTok: response tokens. Read-write. Cold reset: 7h. BIOS: 7h. D18F3x148 Link to XCS Token Count See D18F3x140. Bits Description 31:30 FreeTok[3:2]: free tokens. See: FreeTok[1:0]. 29 Reserved. 28 IsocRspTok1: isochronous response tokens sublink 1. Read-write. Cold reset: 0. BIOS: 0. 27 Reserved. 26 IsocPreqTok1: isochronous posted request tokens sublink 1. Read-write. Cold reset: 0. BIOS: 0. 25 Reserved. 24 IsocReqTok1: isochronous request tokens sublink 1. Read-write. Cold reset: 0. BIOS: 0. 23:22 ProbeTok1: probe tokens sublink 1. Read-write. Cold reset: 0. BIOS: 0. 21:20 RspTok1: response tokens sublink 1. Read-write. Cold reset: 0. BIOS: 0. 19:18 PReqTok1: posted request tokens sublink 1. Read-write. Cold reset: 0. BIOS: 0. 17:16 ReqTok1: request tokens sublink 1. Read-write. Cold reset: 0. BIOS: 0. 15:14 FreeTok[1:0]: free tokens. Read-write. Cold reset: 0000b. FreeTok[3:0] = {FreeTok[3:2], FreeTok[1:0]}. BIOS: 0. 13:12 IsocRspTok0: isochronous response tokens sublink 0. Read-write. Cold reset: 0. BIOS: 0. 11:10 IsocPreqTok0: isochronous posted request tokens sublink 0. Read-write. Cold reset: 0. BIOS: 1. See D18F0x6C[ApplyIsocModeEnNow]. 9:8 IsocReqTok0: isochronous request tokens sublink 0. Read-write. Cold reset: 0. BIOS: 1. 7:6 ProbeTok0: probe tokens sublink 0. Read-write. Cold reset: 2. BIOS: 0. 5:4 RspTok0: response tokens sublink 0. Read-write. Cold reset: 2. BIOS: 2. 3:2 PReqTok0: posted request tokens sublink 0. Read-write. Cold reset: 2. BIOS: 2. 1:0 ReqTok0: request tokens sublink 0. Read-write. Cold reset: 2. BIOS: 2. D18F3x17C Extended Freelist Buffer Count Out of cold reset, the processor allocates a minimal number of buffers that is smaller than the default values in the register. BIOS must use D18F0x6C[RlsLnkFullTokCntImm] for the values in the register to take effect. This is necessary even if the values are unchanged from the default values. Bits Description 441 42300 Rev 3.12 - July 14, 2015 BKDG for AMD Family 15h Models 10h-1Fh Processors 31:4 Reserved. 3:0 SPQPrbFreeCBC: XBAR to SRI Probe command buffer freelist. Cold reset: 0h. BIOS: 4h. Readwrite. D18F3x180 Extended NB MCA Configuration Reset: 0000_0000h. This register is an extension of D18F3x44 [MCA NB Configuration]. Bits Description 31:27 Reserved. 26 ChgUcToCeEn: change uncorrectable error to correctable error enable. Read-write. 1=The status of uncorrectable errors is changed to appear as correctable errors; D18F3x4C[UC, PCC] are cleared and a machine check exception will not be raised. For uncorrectable ECC errors, D18F3x4C[UECC] is cleared and D18F3x4C[CECC] is set.This field is intended for debug observability. 25 Reserved. 24 McaLogErrAddrWdtErr: log error address on WDT errors. Read-write. BIOS: 1. 1=When a watchdog timeout error occurs (see D18F3x40[WDTRptEn]), the associated address is logged and D18F3x4C[AddrV] is set. 0=When a watchdog timeout error occurs, NB state information is saved and D18F3x4C[AddrV] is cleared. See D18F3x50 for details on saved information. 23 Reserved. 22 SyncFloodOnTblWalkErr: sync flood on table walk error. Read-write. BIOS: 1. 1=A sync flood is generated when the GART table walker encounters an uncorrectable error. A machine check exception is generated independent of the state of this bit. 21 SyncFloodOnCpuLeakErr: sync flood on CPU leak error. Read-write. BIOS: 1. 1=A sync flood is generated when one of the cores encounters an uncorrectable error which cannot be contained to the process on the core. 20 SyncFloodOnL3LeakErr: sync flood on L3 cache leak error. Read-write. BIOS: 1. 1=A sync flood is generated when the L3 cache encounters an uncorrectable error which cannot be contained to the process on one core. 19 PwP2pDatErrRmtPropDis: posted write for remote peer-to-peer data error propagation disable. Read-write. 1= A peer-to-peer posted write with a data error is not propagated to the target IO link chain if the target IO link chain is not attached to the local node (the same node as the source IO link chain). Instead, the write is dropped by the host bridge. This bit can be used in conjunction with DatWrErrDeferEn to cause a machine check exception and SyncFloodOnDeferErrToIO to cause a sync flood. 18 PwP2pDatErrLclPropDis: posted write for local peer-to-peer data error propagation disable. Read-write. 1=A peer-to-peer posted write with a data error is not propagated to the target IO link chain if the target IO link chain is attached to the local node (the same node as the source IO link chain). Instead, the write is dropped by the host bridge. This bit can be used in conjunction with DatWrErrDeferEn to cause a machine check exception and SyncFloodOnDeferErrToIO to cause a sync flood. 17:11 Reserved. 10 Reserved. 9 SyncFloodOnUCNbAry: sync flood on uncorrectable NB array error. Read-write. BIOS: 1. 1=Enables sync flood on detection of an uncorrectable error in a NB array. 442 42300 Rev 3.12 - July 14, 2015 BKDG for AMD Family 15h Models 10h-1Fh Processors 8 SyncFloodOnProtocolErr: sync flood on link protocol error. Read-write. BIOS: 1. 1=Enables sync flood on detection of a protocol error. 7 SyncFloodOnTgtAbortErr. Read-write. BIOS: 1. 1=Enable sync flood on generated or received link responses that indicate target aborts. 6 SyncFloodOnDatErr. Read-write. BIOS: 1. 1=Enable sync flood on generated or received link responses that indicate data error. 5 DisPciCfgCpuMstAbortRsp. Read-write. BIOS: 1. 1=For master abort responses to CPU-initiated configuration accesses, disables MCA error reporting and generation of an error response to the core. It is recommended that this bit be set in order to avoid MCA exceptions being generated from master aborts for PCI configuration accesses, which are common during device enumeration. 4 ChgMstAbortToNoErr. Read-write. 1=Signal no errors instead of master abort in link response packets to IO devices on detection of a master abort condition. When ChgMstAbortToNoErr and D18F3x44[IoMstAbortDis] are both set, ChgMstAbortToNoErr takes precedence. 3 ChgDatErrToTgtAbort. Read-write. 1=Signal target abort instead of data error in link response packets to IO devices (for Gen1 link compatibility). 2 WDTCntSel[3]: watchdog timer count select bit[3]. Read-write. See D18F3x44[WDTCntSel]. 1:0 Reserved. D18F3x190 Downcore Control Cold reset: 0000_0000h. See 2.4.3 [Processor Cores and Downcoring] and 2.4.3.1 [Software Downcoring using D18F3x190[DisCore]]. Bits Description 31:0 DisCore. Read-write; reset-applied. 1=Disable core. 0=Core enabled. [0]=Core 0; ...; [N]=Core N. • Once a core has been removed by D18F3x190[DisCore]=1, it cannot be added back without a cold reset. E.g. Software may only set DisCore bits, never clear them. • The most significant bit N is the number of cores enabled at cold reset and is indicated by the value of CPUID Fn8000_0008_ECX[NC] at cold reset. The most significant bit N and the core ID significance of DisCore is not affected by the value of DisCore followed by a warm-reset. • E.g. If core 2 is disabled by DisCore[3:0]=0100b followed by a warm reset, then the new core 2 is the old core 3. If the new core 2 needs to then be disabled then DisCore[3:0]=1100b followed by a warm reset. • All bits greater than bit N are reserved. D18F3x1A0 Core to NB Buffer Count Out of cold reset, the processor allocates a minimal number of buffers that is smaller than the default values in the register. BIOS must use D18F0x6C[RlsLnkFullTokCntImm] for the values in the register to take effect. This is necessary even if the values are unchanged from the default values. Table 180: Buffer Count Definitions Term NumOfCompUnits Definition The number of compute units for which at least 1 core is enabled. NumOfCompUnits = COUNT(D18F5x80[Enabled]). • The following buffer allocations rules must be satisfied: 443 42300 Rev 3.12 - July 14, 2015 BKDG for AMD Family 15h Models 10h-1Fh Processors • CpuCmdBufCnt >= 2. Bits Description 31:18 Reserved. 17:16 CpuToNbFreeBufCnt. Read-write. Cold reset: 01b. BIOS: 11b. Provides the number of tokens which can released to each compute unit from the freelist pool. This field can be updated at any time by BIOS and does not require a warm reset to take effect. 15 Reserved. 14:12 Reserved.Cold reset: 4h. 11:9 Reserved. 8:4 3 2:0 Reserved. Cold reset: Product-specific. Reserved. CpuCmdBufCnt: CPU to SRI command buffer count. Read-write; reset-applied. Each compute unit is allocated the number of buffers specified by this field. Cold reset: 2. D18F3x1CC IBS Control Reset: 0000_0000h. MSRC001_103A is an alias of D18F3x1CC. D18F3x1CC is programmed by BIOS; The OS reads the LVT offset from MSRC001_103A. Bits Description 31:9 Reserved. 8 LvtOffsetVal: local vector table offset valid. Read-write. BIOS: 1. 1=The offset in LvtOffset is valid. 0=The offset in LvtOffset is not valid and IBS interrupt generation is disabled. 7:4 Reserved. 3:0 LvtOffset: local vector table offset. Read-write. BIOS: 0h. Specifies the address of the IBS LVT entry in the APIC registers. See APIC[530:500]. Bits Description 3h-0h LVT address = <500h + LvtOffset<<4> Fh-4h Reserved D18F3x1FC Product Information Register 1 Bits Description 31 Reserved. 30 Fp2PcieGen2Sup. IF (CPUID Fn8000_0001_EBX[PkgType]!=0h) THEN Reserved. ELSE Value: Product-specific. 0=The FP2 processor link only supports Gen1 mode. 1=The FP2 processor link supports Gen2 mode. ENDIF. 29 EnDcqChgPriToHigh. Value: Product-specific. 28:25 DllProcessFreqCtlIndex2Rate50[3:0]. Value: Product-specific. 24 SWDLLCapTableEn. Value: Product-specific. 23 ForceSmcCheckFlwStDis. Value: Product-specific. See MSRC001_102D[ForceSmcCheckFlwStDis]. 22 DiDtCfg5. Value: Product-specific.See MSRC001_1028[DiDtCfg5]. 444 42300 Rev 3.12 - July 14, 2015 BKDG for AMD Family 15h Models 10h-1Fh Processors 21 Reserved. 20 EnCstateBoostBlockCC6Exit.Value: 0. See D18F5x88[EnCstateBoostBlockCC6Exit]. 19:17 DiDtCfg4. Value: Product-specific. See MSRC001_1028[DiDtCfg4]. 16 Reserved. 15:14 DiDtCfg2. Value: Product-specific. See MSRC001_1028[DiDtCfg2]. 13:6 DiDtCfg1. Value: Product-specific. See MSRC001_1028[DiDtCfg1]. 5:1 0 DiDtCfg0. Value: Product-specific. See MSRC001_1028[DiDtCfg0]. DiDtMode. Value: Product-specific. See MSRC001_1028[DiDtMode]. 445 42300 Rev 3.12 - July 14, 2015 BKDG for AMD Family 15h Models 10h-1Fh Processors 3.12 Device 18h Function 4 Configuration Registers See 3.1 [Register Descriptions and Mnemonics]. See 2.7 [Configuration Space]. D18F4x00 Device/Vendor ID Bits Description 31:16 DeviceID: device ID. Read-only. Value: 1404h. 15:0 VendorID: vendor ID. Read-only. Value: 1022h. D18F4x04 Status/Command Bits Description 31:16 Status. Read-only. Reset: 0000_0000_000X_0000b. Only Status[4] may be set to indicate the existence of a PCI-defined capability block. 0=No supported links are unganged. 1=At least one link may be unganged, in which case there is a capability block associated with sublink one of the link in this function. 15:0 Command. Read-only. Value: 0000h. D18F4x08 Class Code/Revision ID Reset: 0600_0000h. Bits Description 31:8 ClassCode. Read-only. Provides the host bridge class code as defined in the PCI specification. 7:0 RevID: revision ID. Read-only. D18F4x0C Header Type Reset: 0080_0000h. Bits Description 31:0 HeaderTypeReg. Read-only. These bits are fixed at their default values. The header type field indicates that there are multiple functions present in this device. D18F4x34 Capabilities Pointer Bits Description 31:8 Reserved. 7:0 CapPtr: capabilities pointer. Read-only. Value: 00h. 446 42300 Rev 3.12 - July 14, 2015 BKDG for AMD Family 15h Models 10h-1Fh Processors D18F4x108 TDP Limit 1 Bits Description 31:0 Reserved. D18F4x10C TDP Limit 2 Updated-by-hardware. Bits Description 31:12 Reserved. 11:0 NodeTdpLimit. Read-write. Reset: Product-specific. Specifies the maximum allowed sum of TDPs from all cores on a node. If the consumed power exceeds the NodeTdpLimit, a P-state limit is applied to all cores on the processor to reduce the power consumption so that it remains within the TDP limit. If D18F4x15C[BoostLock]=1, NodeTdpLimit can only be written with values that are less than or equal to the reset value.Attempts to write an invalid value are ignored. See . D18F4x110 Sample and Residency Timers Bits Description 31:21 Reserved. 20:13 MinResTmr: minimum residency timer. IF D18F4x15C[BoostLock] THEN Read-only. ELSE Read-write. ENDIF. Cold reset: Product-specific. Specifies the minimum amount of time required between TDP-initiated P-state transitions. The minimum amount of time is defined as MinResTmr * CSampleTimer * 5.12us. 12 Reserved. 11:0 CSampleTimer. Read-write. Cold reset: 0. BIOS: 002h. Specifies the value that the internal CSampleTimer counter must increment to before expiring. When the internal CSampleTimer counter expires, it is reset to 0. See 2.5.3.1.1 [Application Power Management (APM)]. D18F4x11[C:8] C-state Control D18F4x11[C:8] consist of three identical 16-bit registers, one for each C-state Action Field (CAF) associated with an IO address that is read to enter C-states. Refer to 2.5.3.2 [Core C-states]. • D18F4x118[15:0] specifies the actions attempted by the core when software reads from the IO address specified by MSRC001_0073[CstateAddr]. • D18F4x118[31:16] specifies the actions attempted by the core when software reads from the IO address specified by MSRC001_0073[CstateAddr+1]. • D18F4x11C[15:0] specifies the actions attempted by the core when software reads from the IO address specified by MSRC001_0073[CstateAddr+2]. D18F4x118 C-state Control 1 Reset: 0000_0000h. Read-write, updated-by-hardware. 447 42300 Rev 3.12 - July 14, 2015 Bits BKDG for AMD Family 15h Models 10h-1Fh Processors Description 31:30 Reserved. 29 SelfRefrEarly1. See: SelfRefrEarly0.BIOS: 0. 28 SelfRefr1. See: SelfRefr0. BIOS: 1. 27 NbClkGate1. See: NbClkGate0. BIOS: 1. 26 NbPwrGate1. See: NbPwrGate0. BIOS: 1. 25 PwrOffEnCstAct1. See: PwrOffEnCstAct0. BIOS: 1. 24 PwrGateEnCstAct1. See: PwrGateEnCstAct0. BIOS: 1. 23:21 ClkDivisorCstAct1. See: ClkDivisorCstAct0. BIOS: 000b. 20 Reserved. 19:18 CacheFlushTmrSelCstAct1. See: CacheFlushTmrSelCstAct0. BIOS: 01b. 17 CacheFlushEnCstAct1. See: CacheFlushEnCstAct0. BIOS: 1. 16 CpuPrbEnCstAct1. See: CpuPrbEnCstAct0. BIOS: 1. 15:14 Reserved. 13 SelfRefrEarly0: allow early self-refresh. BIOS: 0. 1=Allow self-refresh while cores in CC1 are waiting for the cache flush timer to expire. 0=Wait for cache flush timer to expire before allowing self-refresh. See 2.5.7.2 [DRAM Self-Refresh] and 2.5.3.2.3.1 [C-state Probes and Cache Flushing]. 12 SelfRefr0: self-refresh. BIOS: 1. 1=Allow DRAM self-refresh while in NB C-states. 0=Prevent DRAM self-refresh while in NB C-states. NbClkGate0 must be programmed to 1 if this field is programmed to 1. See 2.5.7.2 [DRAM Self-Refresh] and 2.5.4.2 [NB C-states]. 11 NbClkGate0: NB clock-gating. BIOS: 1. 1=Allow clock-gating of the NB. 0=Prevent clock-gating of the NB. See 2.5.4.2 [NB C-states]. 10 NbPwrGate0: NB power-gating. BIOS: 1. 1=Allow power-gating of the NB. 0=Prevent power-gating of the NB. See 2.5.4.2 [NB C-states]. 9 PwrOffEnCstAct0: power off enable. BIOS: 1. 1=Package power off enable. CacheFlushEnCstAct0 is required to be set if this bit is set. PwrGateEnCstAct0 is required to be set if this bit is set. See 2.5.3.2.3.4 [Package C6 (PC6) State]. 8 PwrGateEnCstAct0: power gate enable. BIOS: 1. 1=Core power gating is enabled. CacheFlushEnCstAct0 is required to be set if this bit is set. See2.5.3.2.3.3 [Core C6 (CC6) State]. 448 42300 Rev 3.12 - July 14, 2015 7:5 4 3:2 BKDG for AMD Family 15h Models 10h-1Fh Processors ClkDivisorCstAct0: clock divisor. BIOS: 000b. Specifies the core clock frequency while in the lowpower state before the caches are flushed. This divisor is relative to the current FID frequency, or: • 100 MHz * (10h + MSRC001_00[6B:64][CpuFid]) of the current P-state specified by MSRC001_0063[CurPstate]. If MSRC001_00[6B:64][CpuDid] of the current P-state indicates a divisor that is deeper than specified by this field, then no frequency change is made when entering the low-power state associated with this register. Bits Description Bits Description 000b /1 100b /16 001b /2 101b /128 010b /4 110b /512 011b /8 111b Turn off clocks. See CacheFlushTmrSelCstAct0. Reserved. CacheFlushTmrSelCstAct0: cache flush timer select. BIOS: 10b. Specifies the timer to use for cache flush if (D18F4x128[CoreCstateMode]==0). Bits Cache flush timer 00b 0 us 01b D18F3xDC[CacheFlushOnHaltTmr] 10b D18F4x128[CacheFlushTmr] 11b Reserved Each compute unit has one timer that is shared by both cores. D18F3xDC[CacheFlushOnHaltCtl] specifies the core clock divisor to use after the caches are flushed. Writing values greater than 10b result in 10b. See CacheFlushEnCstAct0 and CpuPrbEnCstAct0. 1 CacheFlushEnCstAct0: cache flush enable. BIOS: 1. 1=Cache flush enable if (D18F4x128[CoreCstateMode]==0). The cache flush timer starts counting when the C-state is entered. See CacheFlushTmrSelCstAct0 and 2.5.3.2.3.1 [C-state Probes and Cache Flushing]. 0 CpuPrbEnCstAct0: core direct probe enable. BIOS: 1. Specifies how probes are handled while in the low-power state. 0=When the probe request comes into the NB, the core clock is brought up to the COF (based on the current P-state), all outstanding probes are completed, the core waits for a hysteresis time based on D18F3xD4[ClkRampHystSel], and then the core clock is brought down to the frequency specified by ClkDivisorCstAct0. 1=The core clock does not change frequency; the probe is handled at the frequency specified by ClkDivisorCstAct0; this may only be set if: • ClkDivisorCstAct0 specifies a divide-by 1, 2, 4, 8, or 16 and NbCof <= 3.2 GHz • ClkDivisorCstAct0 specifies a divide-by 1, 2, 4, or 8 and NbCof >= 3.4 GHz This bit also specifies functionality of the timer used for cache flushing. See CacheFlushTmrSelCstAct0. • If CpuPrbEnCstAct0=0 and D18F3xDC[IgnCpuPrbEn]=0, only the time when the core is in a nonC0 state and has its clocks ramped up to service probes is counted. • If CpuPrbEnCstAct0=1 or D18F3xDC[IgnCpuPrbEn]=1, all of the time the core is in a non-C0 state is counted. D18F4x11C C-state Control 2 Reset: 0000_0000h. Read-write. 449 42300 Rev 3.12 - July 14, 2015 Bits BKDG for AMD Family 15h Models 10h-1Fh Processors Description 31:14 Reserved. 13 SelfRefrEarly2. See: D18F4x11[C:8][SelfRefrEarly0]. 12 SelfRefr2. See: D18F4x11[C:8][SelfRefr0]. 11 NbClkGate2. See: D18F4x11[C:8][NbClkGate0]. 10 NbPwrGate2. See: D18F4x11[C:8][NbPwrGate0]. 9 PwrOffEnCstAct2. See: D18F4x11[C:8][PwrOffEnCstAct0]. 8 PwrGateEnCstAct2. See: D18F4x11[C:8][PwrGateEnCstAct0]. 7:5 ClkDivisorCstAct2. See: D18F4x11[C:8][ClkDivisorCstAct0]. 4 3:2 Reserved. CacheFlushTmrSelCstAct2. See: D18F4x11[C:8][CacheFlushTmrSelCstAct0]. 1 CacheFlushEnCstAct2. See: D18F4x11[C:8][CacheFlushEnCstAct0]. 0 CpuPrbEnCstAct2. See: D18F4x11[C:8][CpuPrbEnCstAct0]. D18F4x124 C-state Interrupt Control Reset: 0000_0000h. Read-write. See 2.5.3.2.4.2 [Timer Tick Monitor]. Bits Description 31:27 Reserved. 26:23 IntMonPC6Limit: interrupt monitor package C6 limit. Read-write. Specifies the threshold for disallowing access to the PC6 state. Bits Threshold 0000b PC6 entry disallowed 1001b-0001b IntMonPC6Limit * 1ms 1110b-1010b ((5 * IntMonPC6Limit) - 40) * 1ms 1111b 40ms 22 IntMonPC6En: interrupt monitor package C6 enable. Read-write.Specifies whether the interrupt monitor is enabled for the PC6 state. 0=Disabled. 1=Enabled. 21:0 Reserved. D18F4x128 C-state Policy Control 1 Reset: 0000_0000h. Bits 31 Description CstateMsgDis: C-state messaging disable. Read-write. BIOS: 1. Specifies whether any messages are sent to the FCH when a core enters or exits a C-state. 0=Messages are sent. 1=Messages are not sent. See 2.5.3.2.4.1 [FCH Messaging]. 30:21 Reserved. 20:18 CacheFlushSucMonThreshold: cache flush success monitor threshold. Read-write. BIOS: 111b. Flush the caches immediately if cache flushing is enabled and the cache flush success monitor count > CacheFlushSucMonThreshold. A value of 0 disables the cache flush success monitor. See D18F4x11[C:8][CacheFlushEn]. 450 42300 Rev 3.12 - July 14, 2015 BKDG for AMD Family 15h Models 10h-1Fh Processors 17:12 Reserved. 11:5 CacheFlushTmr: cache flush timer. Read-write. BIOS: 14h. Specifies how long each core needs to stay in a C-state before it flushes its caches. See D18F4x11[C:8][CacheFlushTmrSel]. Bits Description 00h <= 5.12 us 7Fh-01h (<CacheFlushTmr> * 10.24us) - 5.12us <= Time <= <CacheFlushTmr> * 10.24 us 4:2 HaltCstateIndex. Read-write. BIOS: 0. Specifies the IO-based C-state that is invoked by a HLT instruction. 1 CoreCstatePolicy. Read-write. BIOS: 0. Specifies how the processor arbitrates voltage and frequency when different non-C0 C-state requests are received on each core in a compute unit. 0=Transition both cores to the shallower C-state request. 1=Transition both cores to the deeper C-state request. For instance, if core 0 gets a request to go to C2 and core 1 gets a request to go to C1, hardware looks at the setting of CoreCstatePolicy. If CoreCstatePolicy is programmed to 0, the processor sends both cores to C1. If CoreCstatePolicy is programmed to 1, the processor sends both cores to C2. BIOS should program this field to the same value in all nodes of a multi-node processor. See also 2.5.2.2 [Dependencies Between Subcomponents on VDDNB]. 0 Reserved. D18F4x13C SMU P-state Control Updated-by-hardware. Bits Description 31:4 Reserved. 3:1 0 SmuPstateLimit. Read-only; updated-by-hardware. Reset: 0. SmuPstateLimitEn. Read-only; updated-by-hardware. Reset: 0. D18F4x15C Core Performance Boost Control Updated-by-hardware. Bits 31 Description BoostLock. Read-only. Reset: Product-specific. Specifies whether the following registers are readwrite, read-only, or have special requirements related to writability. See individual register definitions for details. • MSRC001_00[6B:64][CpuFid, CpuDid, CpuVid]. • D18F4x10C[NodeTdpLimit]. • D18F4x110[MinResTmr]. • D18F4x15C[NumBoostStates]. • D18F4x16C[CstateCnt, CstateBoost]. 30:8 Reserved. 7 6:5 ApmMasterEn: APM master enable. Read-write. Reset: 0. BIOS: IF(D18F4x15C[NumBoostStates]==0) THEN 0. ELSE 1. ENDIF. 1=Enables the ability to turn on features associated with APM when used in conjunction with the individual feature enable bits. See 2.5.3.1.1 [Application Power Management (APM)]. Reserved. 451 42300 Rev 3.12 - July 14, 2015 BKDG for AMD Family 15h Models 10h-1Fh Processors 4:2 NumBoostStates: number of boosted states. IF (D18F4x15C[BoostLock] | ApmMasterEn) THEN Read-only. ELSE Read-write. ENDIF. Reset: Product-specific. Specifies the number of P-states that are considered boosted P-states. See 2.5.3.1.1 [Application Power Management (APM)]. 1:0 BoostSrc: boost source. Read-write. Reset: 0. BIOS: 2.5.3.1.7. Specifies whether CPB is enabled or disabled. Bits Description 00b Boosting disabled 01b Boosting enabled 10b Reserved 11b Reserved D18F4x164 Fixed Errata Bits Description 31:0 FixedErrata. Value: Product-specific.. See the Revision Guide for AMD Family 15h Models 10h-1Fh Processors for the definition of this field. D18F4x16C APM TDP Control Bits Description 31:12 Reserved. 11:9 CstateCnt: C-state count. IF D18F4x15C[BoostLock] THEN Read-only. ELSE Read-write. ENDIF. Reset: Product-specific. Specifies the number of compute units that must be in CC6 before a transition can occur to a boosted P-state that is higher performance than the P-state specified by CstateBoost. A value of 0 disables access to P-states above CstateBoost. 8:6 CstateBoost. Read-write. Reset: Product-specific. Specifies the P-state which requires the number of compute units specified in CstateCnt to be in CC6 before a transition to a higher performance (lower numbered) boosted P-state is allowed. CstateBoost must be less than or equal to D18F4x15C[NumBoostStates] otherwise undefined behavior results. If D18F4x15C[BoostLock]=1, CstateBoost can only be written with values that are greater than or equal to the reset value. Attempts to write values less than the reset value are ignored. A value of 0 indicates that the C-state boost feature is not supported. This field uses hardware P-state numbering. See 2.5.3.1.2.2 [Hardware P-state Numbering]. 5 Reserved. 4 ApmTdpLimitIntEn: APM TDP limit interrupt enable. Read-write. Reset: 0. 1=Enables the generation of an interrupt using APIC330 of each core when TDP changes. 3:0 Reserved. 452 42300 Rev 3.12 - July 14, 2015 BKDG for AMD Family 15h Models 10h-1Fh Processors 3.13 Device 18h Function 5 Configuration Registers See 3.1 [Register Descriptions and Mnemonics]. See 2.7 [Configuration Space]. D18F5x00 Device/Vendor ID Bits Description 31:16 DeviceID: device ID. Read-only. Value: 1405h. 15:0 VendorID: vendor ID. Read-only. Value: 1022h. D18F5x04 Status/Command Bits Description 31:16 Status. Read-only. Value: 0000h. 15:0 Command. Read-only. Value: 0000h. D18F5x08 Class Code/Revision ID Bits Description 31:8 ClassCode. Read-only. Value: 060000h. Provides the host bridge class code as defined in the PCI specification. 7:0 RevID: revision ID. Read-only. Value: 00h. D18F5x0C Header Type Reset: 0080_0000h. Bits Description 31:0 HeaderTypeReg. Read-only. These bits are fixed at their default values. The header type field indicates that there are not multiple functions present in this device. D18F5x34 Capabilities Pointer Bits Description 31:8 Reserved. 7:0 CapPtr: capabilities pointer. Read-only. Value: 00h. D18F5x[70,60,50,40] Northbridge Performance Event Select Low Bits Description 31:0 MSRC001_024[6,4,2,0][31:0] is an alias of D18F5x[70,60,50,40]. 453 42300 Rev 3.12 - July 14, 2015 BKDG for AMD Family 15h Models 10h-1Fh Processors D18F5x[74,64,54,44] Northbridge Performance Event Select High Bits Description 31:0 MSRC001_024[6,4,2,0][63:32] is an alias of D18F5x[74,64,54,44]. D18F5x[78,68,58,48] Northbridge Performance Event Counter Low Bits Description 31:0 MSRC001_024[7,5,3,1][31:0] is an alias of D18F5x[78,68,58,48]. D18F5x[7C,6C,5C,4C] Northbridge Performance Event Counter High Bits Description 31:0 MSRC001_024[7,5,3,1][63:32] is an alias of D18F5x[7C,6C,5C,4C]. D18F5x80 Compute Unit Status Read-only. See 2.4.3 [Processor Cores and Downcoring]. Reset: Product-specific. Software associates core ID to the cores of the compute units according to the following table. All combinations not listed are reserved. Table 181: D18F5x80[Enabled, DualCore] Definition Enabled 1h 3h Bits DualCore Definition 1h 1 compute unit is enabled; both cores of the compute unit are enabled. 3h 2 compute units are enabled; both cores of each compute unit are enabled. Description 31:20 Reserved. 19:16 DualCore: both cores of a compute unit are enabled. 1=Both cores of a compute unit are enabled. See Table 181 [D18F5x80[Enabled, DualCore] Definition]. Bit Description [0] Compute unit 0 [1] Compute unit 1 [2] Reserved [3] Reserved 15:4 Reserved. 3:0 Enabled: at least one core of a compute unit is enabled. 1=At least one core is enabled in a compute unit. See Table 181 [D18F5x80[Enabled, DualCore] Definition]. Bit Description [0] Compute unit 0 [1] Compute unit 1 [2] Reserved [3] Reserved 454 42300 Rev 3.12 - July 14, 2015 BKDG for AMD Family 15h Models 10h-1Fh Processors D18F5x84 Northbridge Capabilities 2 Read-only. Value: Product-specific. Unless otherwise specified, 1=The feature is supported by the processor; 0=The feature is not supported. Bits Description 31:29 Reserved. 28:24 DdrMaxRateEnf: enforced maximum DDR rate. See: DdrMaxRate. Specifies the maximum DRAM data rate that the processor is designed to support. Writes to D18F2x94_dct[1:0][MemClkFreq] that specify a frequency greater than specified by DdrMaxRateEnf will result in the D18F2x94_dct[1:0][MemClkFreq] being set to DdrMaxRateEnf. 23:21 Reserved. 20:16 DdrMaxRate: maximum DDR rate. Specifies the maximum DRAM data rate that the processor is designed to support. See: Table 132 [Memory Clock Frequency Value Definition]; except 00h is defined as no limit. See D18F2x94_dct[1:0][MemClkFreq] and D18F5x84[DdrMaxRateEnf]. 15:14 Reserved. 13:12 DctEn: DCT enabled. Specifies which DCT controllers are enabled. [0]=DCT0. [1]=DCT1. 1=Enabled. 0=Disabled. 11:8 Reserved. 7:0 CmpCap: CMP capable. Number of enabled cores on the node is CmpCap+1. CmpCap does not reflect cores disabled by D18F3x190[DisCore]. D18F5x88 Northbridge Configuration 4 (NB_CFG4) Bits Description 31:25 Reserved. 24 DisHbNpReqBusLock. Read-write. Reset: 0. BIOS: 1. 0=While bus locks are in progress, all nonposted commands from I/O, including atomics, are blocked until the core has completed the locked transaction and releases the bus. 1=All non-posted commands except atomics do not honor bus locks and are allowed to proceed. This bit may be set to achieve better DMA performance in the presence of bus locks. 23:20 Reserved. 19 EnCpuInSWP0DctHint. Read-write. Reset: 0. The processor allocates additional memory bandwidth to the GPU when all cores are not in P0. This bit controls whether software P0 or hardware P0 is used. 1=Use software P0. 0=Use hardware P0. See 2.5.3.1.2 [Core P-state Naming and Numbering]. 18 EnCstateBoostBlockCC6Exit. Read-write. Reset: 0. BIOS: D18F3x1FC[EnCstateBoostBlockCC6Exit]. If (EnCstateBoostBlockCC6Exit==1), cores cannot exit CC6 until VDD is less than or equal to the voltage of the P-state indexed by D18F4x16C[CstateBoost]. 17:3 Reserved. 455 42300 Rev 3.12 - July 14, 2015 2 1:0 BKDG for AMD Family 15h Models 10h-1Fh Processors IntStpClkHaltExitEn. Read-write. Reset: 0. BIOS: 1. 1=The Northbridge sends out HALT_EXIT messages received from the core due to internal StpClk interrupts. 0=The Northbridge ignores HALT_EXIT messages received from the core due to internal StpClk interrupts. This bit should always be set. Reserved. D18F5xE0 Processor TDP Running Average Bits Description 31:4 Reserved. 3:0 RunAvgRange: running average range. Read-write. Reset: 0. BIOS: 2h. Specifies the interval over which the processor averages power consumption estimates from the cores for boosting. Time interval = 2^(RunAvgRange + 1) * FreeRunSampleTimer rate. A value of 0 disables the TDP running average accumulator capture function. See 2.5.3.1.1 [Application Power Management (APM)]. D18F5x128 Clock Power/Timing Control 3 Bits Description 31:23 Reserved. 22 NbPllPwrDwnRegEn: NB PLL power down . Read-write. Cold reset: Product-specific.1=The NB PLL is powered down when the NB is power gated and DRAM is placed into self-refresh (see 2.5.4.2 [NB C-states]). 0=The NB PLL is not powered down during NB C-states. 21 PC6Vid[7]. Read-write. Cold reset: Product-specific. See PC6Vid[6:0]. 20:16 Reserved. 15 CC6PwrDwnRegEn: CC6 power down regulator enable. Read-write. Cold reset: Product-specific. 1=Power down the VDDA regulator on CC6 entry. See PllRegTime. 14 PC6PwrDwnRegEn: PC6 power down regulator enable. Read-write. Cold reset: Product-specific. 1=Power down the VDDA regulator on PC6 entry. See PllRegTime. 13:12 PwrGateTmr: power gate timer. Read-write. Cold reset: 01b. BIOS: 01b. Specifies the minimum delay time required from the power gating or ungating of one compute unit to the power gating or ungating of the same compute unit or another compute unit. Bits Description Bits Description 00b 500 ns 10b 5 us 01b 1 us 11b 10 us 11:10 PllVddOutUpTime. Read-write. Cold Reset: 0. BIOS: 11b. The VDD regulator may be powered down when the processor transitions to PC6. If the regulator is powered down, this field specifies the time required to initialize the core PLL logic once the regulator is powered back up. Bits Description Bits Description 00b 100 ns 10b 400 ns 01b 200 ns 11b 800 ns 9 FastSlamTimeDown. Read-write. Cold Reset: 0. BIOS: 1. Specifies the time the processor waits for downward voltage transitions to complete. This field only effects transitions from D18F4x16C[CstateBoost] or lower performance P-states. 0=D18F3xD8[VSRampSlamTime] . 1=10 us. 456 42300 Rev 3.12 - July 14, 2015 BKDG for AMD Family 15h Models 10h-1Fh Processors 8:7 PllRegTime: Pll regulator time. Read-write. Cold Reset: 10b. The VDDAregulator may be powered down when the processor transitions to PC6 or CC6. See PC6PwrDwnRegEn and CC6PwrDwnRegEn. If CC6PwrDwnRegEn=1, the VDDA regulator is powered down during CC6. If PC6PwrDwnRegEn=1, the VDDA regulator is powered down during PC6. If the VDDA regulator is powered down during CC6 and the core transitions from CC6 to PC6, the regulator remains powered down during PC6 regardless of the PC6PwrDwnRegEn setting. This field specifies the time required for the VDDA regulator to power back up and initialize the core PLL logic that is powered by the VDDA regulator. Description Bits Description Bits 00b Reserved. 10b 1.5 us 01b Reserved. 11b 2.0 us 6:0 PC6Vid[6:0]: package C6 vid. Read-write. Cold reset: Value varies by product. PC6Vid[7:0] = {PC6Vid[7], PC6Vid[6:0]}. PC6Vid[7:0] specifies the VID driven in the PC6 state. See 2.5.3.2.3.4 [Package C6 (PC6) State] and 2.5.1.3.2 [Low Power Voltages]. Required to be programmed within the range specified by MSRC001_0071[MaxVid and MinVid]. D18F5x12C Clock Power/Timing Control 4 Cold reset: 0000_000Eh. See the AMD Serial VID Interface 2.0 (SVI2) Specification. Bits Description 31 Svi2CmdBusy. Read-only, updated-by-hardware. 1=SVI2 command in progress. This bit is set by hardware when any SVI2 command is sent to the voltage regulator. Software must wait for this bit to clear to 0 before writing any of the below fields: • D18F5x12C[CorePsi1En, CoreLoadLineTrim, CoreOffsetTrim] • D18F5x188[NbPsi1, NbLoadLineTrim, NbOffsetTrim] This bit is cleared by hardware when the SVI2 command is complete. On a voltage change, this bit is cleared when the voltage transition is completed. See 2.5.1.4.1 [Hardware-Initiated Voltage Transitions]. On a PSIx_L change, this bit is cleared as soon as the SVI2 command is sent to the voltage regulator. See 2.5.1.1.1 [SVI2 Features] and 2.5.1.3.1 [PSIx_L Bit]. 30 WaitVidCompDis: wait VID completion disable. Read-write. 0=Hardware waits for the VOTF complete indicator from the voltage regulator before clearing Svi2CmdBusy or making additional voltage change requests. 1=Hardware clears Svi2CmdBusy 500us after changes to CoreLoadLineTrim, CoreOffsetTrim, or D18F5x188[NbLoadLineTrim, NbOffsetTrim] are made. Hardware clears Svi2CmdBusy and additional voltage changes are allowed after the time specified by 2.5.1.4 [Voltage Transitions]. 29:6 Reserved. 5 CorePsi1En: Core PSI1_L enable. Read-write. BIOS: 1. 0=PSI1_L for VDD is deasserted. 1=PSI1_L for VDD is asserted when all cores are in CC6. See 2.5.3.2.3.4 [Package C6 (PC6) State], 2.5.1.3.1 [PSIx_L Bit], and Svi2CmdBusy. 4:2 CoreLoadLineTrim: Core load line trim. Read-write. BIOS: D0F0xBC_xE0104184[SviLoadLineTrimVdd]. CoreLoadLineTrim and NbLoadLineTrim specify a percentage change relative to the initial load line slope for VDD and VDDNB, respectively. Bits Description Bits Description 000b Load line disabled 100b +20% 001b -40% 101b +40% 010b -20% 110b +60% 011b 0% 111b +80% See Svi2CmdBusy. 457 42300 Rev 3.12 - July 14, 2015 BKDG for AMD Family 15h Models 10h-1Fh Processors CoreOffsetTrim: Core offset trim. Read-write. BIOS: D0F0xBC_xE0104184[SviLoadLineOffsetVdd]. CoreOffsetTrim and NbOffsetTrim specify a voltage offset relative to the initial load line offset for VDD and VDDNB, respectively. Description Bits Description Bits 00b Load line offset disabled 10b 0mV 01b -25mV 11b +25mV See Svi2CmdBusy. 1:0 D18F5x1[6C:60] Northbridge P-state [3:0] Cold reset: Product-specific. Each of these registers specify the frequency and voltage associated with each of the NB P-states. Table 182: Register Mapping for D18F5x1[6C:60] Register D18F5x160 D18F5x164 D18F5x168 D18F5x16C Function NB P-state 0 NB P-state 1 NB P-state 2 NB P-state 3 The NbVid field is allowed to be different between processors in a multi-processor system. All other fields are required to be programmed to the same value for all processors in the coherent fabric. See 2.5.4.1 [NB P-states] for more information about these registers. Table 183: NB P-state Definitions Term NBCOF NBCOF[0] NBCOF[1] NBCOF[2] NBCOF[3] Bits Definition NB current operating frequency in MHz. NBCOF = 100 * (D18F5x1[6C:60][NbFid] + 4h) / (2^D18F5x1[6C:60][NbDid]). NB current operating frequency in MHz for NB P-state 0. NBCOF[0] = (100 * (D18F5x160[NbFid] + 4h) / (2^D18F5x160[NbDid])). NB current operating frequency in MHz for NB P-state 1. NBCOF[1] = (100 * (D18F5x164[NbFid] + 4h) / (2^D18F5x164[NbDid])). NB current operating frequency in MHz for NB P-state 2. NBCOF[2] = (100 * (D18F5x168[NbFid] + 4h) / (2^D18F5x168[NbDid])). NB current operating frequency in MHz for NB P-state 3. NBCOF[3] = (100 * (D18F5x16C[NbFid] + 4h) / (2^D18F5x16C[NbDid])). Description 31:24 NbIddValue: northbridge current value. Read-write. See NbIddDiv. 458 42300 Rev 3.12 - July 14, 2015 BKDG for AMD Family 15h Models 10h-1Fh Processors 23:22 NbIddDiv: northbridge current divisor. Read-write. After reset, NbIddDiv and NbIddValue combine to specify the expected maximum current drawn on the VDDNB power plane at a given VDDNB voltage. These values are intended to be used by 2.5.1.3.1.1 [BIOS Requirements for PSI0_L]. These values are not intended to convey final product power levels. These fields may be subsequently altered by software; they do not affect the hardware behavior. NbIddDiv Description 00b IddValue / 1 A, Range: 0 to 255 A. 01b IddValue / 10 A, Range: 0 to 25.5 A. 10b IddValue / 100 A, Range: 0 to 2.55 A. 11b Reserved. 21 NbVid[7]. Read-write. See NbVid[6:0]. 20:19 Reserved. 18 MemPstate: Memory P-state. Read-write. 1=The Northbridge P-state specified by this register maps to memory P-state 1. 0=The Northbridge P-state specified by this register maps to memory P-state 0. Memory P-states may be globally disabled by programming D18F5x170[MemPstateDis]. See 2.5.7.1 [Memory P-states]. 17 Reserved. 16:10 NbVid[6:0]: Northbridge VID. Read-write. NbVid[7:0] = {NbVid[7], NbVid[6:0]}. NbVid[7:0] specifies the Northbridge voltage. Writes outside the D18F5x17C[MaxVid, MinVid] range are ignored. Per AMD Serial VID Interface 2.0 (SVI2) Specification, effective VDDNB voltage should consider NbVid[7:0] and D18F5x188[NbOffsetTrim]. 9:8 7 Reserved. NbDid: Northbridge divisor ID. Read-write. Specifies the Northbridge frequency divisor; see NbFid. 6:1 NbFid: Northbridge frequency ID. Read-write. Specifies the Northbridge frequency multiplier. The NB COF is a function of NbFid and NbDid, and defined by NBCOF. NbFid and NbDid are not changed on a write if the value written results in a frequency greater than MSRC001_0071[MaxNbCof]. See 2.5.5 [P-state Bandwidth Requirements]. 0 NbPstateEn: Northbridge P-state enable. Read-write. 1=The Northbridge P-state specified by this register is valid. 0=The Northbridge P-state specified by this register is not valid. This bit must be set to 1 in order for the Northbridge P-state specified by this register to be programmed in D18F5x170[NbPstateHi, NbPstateLo]. D18F5x170 Northbridge P-state Control Updated-by-hardware. See also 2.5.4.1 [NB P-states]. Bits Description 31 MemPstateDis: memory P-state disable. IF (D18F3xE8[MemPstateCap]) THEN Read-write, updated-by-hardware. ELSE Read-only, updated-by-hardware. ENDIF. Reset: Value varies by product.. 1=Memory P-state transitions are disabled. The current P-state is not changed by programming this bit. The memory P-state will be forced to M0 on the next NB P-state transition. On processors where memory P-states are enabled, programming this bit may result in a violation of bandwidth requirements stated in 2.5.3.1.6 . Software must ensure that NB P-states which violate those requirements are forced disabled. 0=Memory P-state transitions are enabled if D18F2x90_dct[1:0][DisDllShutdownSR]=0. 30 Reserved. 459 42300 Rev 3.12 - July 14, 2015 BKDG for AMD Family 15h Models 10h-1Fh Processors 29:27 NbPstateHiRes: NB P-state high residency timer. Read-write. Reset: 0. Specifies the minimum time the processor must spend in the high NB P-state before transitions to the low NB P-state are allowed. See 2.5.4.1 [NB P-states]. Description Bits Description Bits 000b 0us 100b 1ms 001b 10us 101b 5ms 010b 100us 110b 10ms 011b 500us 111b 50ms 26:24 NbPstateLoRes: NB P-state low residency timer. Read-write. Reset: 0. Specifies the minimum time the processor must spend in the low NB P-state before transitions to the high NB P-state are allowed. See: NbPstateHiRes. 23 NbPstateGnbSlowDis. Read-write. Reset: 0. Specifies whether NBP-state transitions take GPU activity into account. 0=Take GPU activity into account. 1=Ignore GPU activity . 22:15 Reserved. 14 SwNbPstateLoDis: software NB P-state low disable. IF (MSRC001_0071[NbPstateDis]) THEN Read-only. ELSE Read-write. ENDIF. BIOS: See 2.5.4.1.1.Reset: 0. 1=Transition to NbPstateHi and disable transitions to NbPstateLo. 13 NbPstateDisOnP0: NB P-state disable on P0. IF (MSRC001_0071[NbPstateDis]) THEN Readonly. ELSE Read-write. ENDIF. Reset: 0. 1=Transition to NbPstateHi and disable transitions to NbPstateLo if any compute unit is in P0 or a boosted P-state. This field uses software P-state numbering. See 2.5.3.1.2.1 [Software P-state Numbering]. 12 Reserved. 11:9 NbPstateThreshold: NB P-state threshold. Read-write. Reset: Product-specific. BIOS: COUNT(D18F5x80[Enabled]). Specifies the minimum number of compute units that must be in a Pstate with MSRC001_00[6B:64][NbPstate]=1 before transitions to lower performance NB P-states are allowed. See NbPstateLo and NbPstateHi. 8 7:6 5 4:3 2 Reserved. NbPstateHi: NB P-state high. Read-write, updated-by-hardware. Cold reset: Product-specific. If NB P-states are enabled, this field specifies the NB P-state that is used when the number of compute units in a P-state with MSRC001_00[6B:64][NbPstate]=1 is less than NbPstateThreshold. This field must be programmed to the same value for all processors in the coherent fabric. This field is not changed on a write if the value written is greater than the NbPstateMaxVal value written or greater than the current NbPstateLo value. See also NbPstateDisOnP0, SwNbPstateLoDis, NbPstateLo, D18F5x174[NbPstateDis], and D18F5x1[6C:60][NbPstateEn]. Reserved. NbPstateLo: NB P-state low. Read-write, updated-by-hardware. Cold reset: Product-specific. If NB P-states are enabled, this field specifies the NB P-state that is used when the number of compute units in a P-state with MSRC001_00[6B:64][NbPstate]=1 is greater than or equal to NbPstateThreshold. NbPstateLo must be greater than or equal to NbPstateHi. This field must be programmed to the same value for all processors in the coherent fabric. This field is not changed on a write if the value written is greater than the NbPstateMaxVal value written or less than the current NbPstateHi value. See also NbPstateDisOnP0, SwNbPstateLoDis, D18F5x174[NbPstateDis], and D18F5x1[6C:60][NbPstateEn]. Reserved. 460 42300 Rev 3.12 - July 14, 2015 1:0 BKDG for AMD Family 15h Models 10h-1Fh Processors NbPstateMaxVal: NB P-state maximum value. Read-write. Cold reset: specified by the reset state of D18F5x1[6C:60][NbPstateEn]; the cold reset value is the highest NB P-state number corresponding to the register in which NbPstateEn is set (e.g., if D18F5x160 and D18F5x164 have this bit set and the others do not, then NbPstateMaxVal=1; if NbPstateEn is only set in D18F5x160, then NbPstateMaxVal=0). This specifies the highest NB P-state value (lowest performance state) supported by the hardware. D18F5x174 Northbridge P-state Status Bits Description 31:25 Reserved. 24 CurMemPstate: current memory P-state. Read-only; updated-by-hardware. Reset: 0. Specifies the current memory P-state. 1=Memory P-state 1. 0=Memory P-state 0. See 2.5.7.1 [Memory P-states]. 23 CurNbVid[7]: current northbridge voltage ID[7]. Read-only; updated-by-hardware. Reset: 0. MSRC001_0071[CurNbVid[7]] is an alias of D18F5x174[CurNbVid[7]]. . 22:21 Reserved. 20:19 CurNbPstate: current northbridge P-state. Read-only; updated-by-hardware. Cold reset: Productspecific. Provides the NB P-state that corresponds to the current frequency component of the NB. The value of this field is updated when the COF transitions to a new value associated with an NB Pstate. Bits Description 00b NB P0 01b NB P1 10b NB P2 11b NB P3 18:12 CurNbVid[6:0]: current northbridge voltage ID. MSRC001_0071[CurNbVid[6:0]] is an alias of D18F5x174[CurNbVid[6:0]]. Per AMD Serial VID Interface 2.0 (SVI2) Specification, effective VDDNB voltage should consider CurNbVid[7:0] and D18F5x188[NbOffsetTrim]. 11:10 Reserved. 9 CurNbDid: current northbridge divisor ID. Read-only, updated-by-hardware. Reset: 0. 8:3 CurNbFid: current northbridge frequency ID. Read-only, updated-by-hardware. Reset: 0. 2:1 StartupNbPstate: startup northbridge P-state number. Read-only. Cold reset: Product-specific. Specifies the cold reset VID, FID and DID for the Northbridge based on the NB P-state number selected. 0 NbPstateDis: northbridge P-state disable. Read-only. Value: Product-specific. MSRC001_0071[NbPstateDis] is an alias of D18F5x174[NbPstateDis]. D18F5x178 Northbridge Fusion Configuration Bits Description 31:20 Reserved. 19 SwGfxDis. Read-write. Reset: 1. BIOS: IF (GpuEnabled) THEN 0. ELSE 1. ENDIF. This register is write once. 1=Hardware handshakes for NB P-state transitions and DRAM self-refresh entry are ignored. See 2.5.4.1.1 [NB P-state Transitions] and 2.5.7.2 [DRAM Self-Refresh]. 461 42300 Rev 3.12 - July 14, 2015 BKDG for AMD Family 15h Models 10h-1Fh Processors 18 CstateFusionHsDis: C-state fusion handshake disable. Read-write. Reset: 0. BIOS: 1. 1=Ignore the FCH handshake response for PC6 transitions. 0=Use the FCH handshake response for PC6 entry. See 2.5.3.2.4.1 [FCH Messaging]. 17 Dis2ndGnbAllowPsWait. Read-write. Reset: 0. BIOS: 1. 0=When doing an NB P-state transition wait for the display buffers to fill after quiescing cores. See 2.5.4.1.1 [NB P-state Transitions]. 16:12 Reserved. 11 AllowSelfRefrS3Dis: allow self-refresh S3 disable. Read-write. Reset: 0. BIOS: 1. 1=The NB does not wait for handshake before placing DRAM into self-refresh (see 2.5.7.2 [DRAM Self-Refresh]) on S3 entry (see 2.5.8.1.1 [ACPI Suspend to RAM State (S3)]). 0=The NB waits for handshake before placing DRAM into self-refresh on S3 entry. 10 InbWakeS3Dis: InbWake S3 disable. Read-write. Reset: 0. BIOS: 1. 1= The NB does not wait for handshake before placing DRAM into self-refresh (see 2.5.7.2 [DRAM Self-Refresh]) on S3 entry (see 2.5.8.1.1 [ACPI Suspend to RAM State (S3)]). 0=The NB waits for handshake before placing DRAM into self-refresh on S3 entry. 9:4 Reserved. 3 CstateThreeWayHsEn: C-state three way handshake disable. Read-write. Reset: 0. 1=Enable the three way handshake with the FCH when entering a C-state. 0=Only a two way handshake with FCH is used. There is no message about the resulting package state sent to FCH. See 2.5.3.2.4.1 [FCH Messaging]. 2 CstateFusionDis: C-state fusion disable. Read-write. Reset: 0. BIOS: 1. 1=All HALT or C-state requests are forwarded to the FCH. 0=HALT and C-state requests are forwarded to the FCH when each core has made a request. See 2.5.3.2.4.1 [FCH Messaging]. 1:0 Reserved. D18F5x17C Miscellaneous Voltages Bits 31 Description NbPsi0VidEn: Northbridge PSI0_L VID enable. Read-write. Reset: 0. BIOS: 2.5.1.3.1.1. This bit specifies how PSI0_L is controlled for VDDNB. See D18F3xA0[PsiVidEn] and 2.5.1.3.1 [PSIx_L Bit]. 30:23 NbPsi0Vid[7:0]: Northbridge PSI0_L VID threshold. Read-write. Reset: 0. BIOS: 2.5.1.3.1.1. When enabled by NbPsi0VidEn, NbPsi0Vid specifies the threshold value of the VID code generated by the Northbridge, which in turn determines the state of PSI0_L. See D18F3xA0[PsiVid[6:0]] and 2.5.1.3.1 [PSIx_L Bit]. 22:18 Reserved. 17:10 MinVid: minimum voltage. Read-only. Reset: Product-specific. Specifies the VID code corresponding to the minimum voltage (highest VID code) that the processor drives. 00h indicates that no minimum VID code is specified. See 2.5.1 [Processor Power Planes And Voltage Control]. 9:8 Reserved. 7:0 MaxVid: maximum voltage. Read-only. Reset: Product-specific. Specifies the VID code corresponding to the maximum voltage (lowest VID code) that the processor drives. 00h indicates that no maximum VID code is specified. See 2.5.1 [Processor Power Planes And Voltage Control]. 462 42300 Rev 3.12 - July 14, 2015 BKDG for AMD Family 15h Models 10h-1Fh Processors D18F5x188 Clock Power/Timing Control 5 Cold reset: 0000_000Eh. Updated-by-hardware. See the AMD Serial VID Interface 2.0 (SVI2) Specification. Bits Description 31:6 Reserved. 5 NbPsi1: Northbridge PSI1_L. Read-write, updated-by-hardware. Specifies how PSI1_L is controlled for VDDNB. 1=PSI1_L is low. 0=PSI1_L is high. See 2.5.1.3.1 [PSIx_L Bit]. 4:2 NbLoadLineTrim: Northbridge load line trim. Read-write. BIOS: D0F0xBC_xE0104184[SviLoadLineTrimVddNb]. See D18F5x12C[CoreLoadLineTrim]. 1:0 NbOffsetTrim: Northbridge offset trim. Read-write, updated-by-hardware. BIOS: D0F0xBC_xE0104184[SviLoadLineOffsetVddNb]. For encoding, see D18F5x12C[CoreOffsetTrim]. D18F5x194 Name String Address Port D18F5x194 and D18F5x198 provide BIOS with a read-only name string that may be copied to MSRC001_00[35:30] at warm reset. Each of D18F5x198_x[B:0] is read as follows: 1. Write D18F5x194[Index]. 2. Read D18F5x198. Bits Description 31:4 Reserved. 3:0 Index: name string register index. Read-write. Reset: 0. Bits Description Bh-0h Name String Registers. See D18F5x198_x[B:0]. Fh-Ch Reserved D18F5x198 Name String Data Port See D18F5x194 for register access information. D18F5x198_x[B:0] Name String Data Read-only. Value: Product-specific. Bits Description 31:24 NameStringByte3: name string ASCII character 3. 23:16 NameStringByte2: name string ASCII character 2. 15:8 NameStringByte1: name string ASCII character 1. 7:0 NameStringByte0: name string ASCII character 0. 463 42300 Rev 3.12 - July 14, 2015 BKDG for AMD Family 15h Models 10h-1Fh Processors 3.14 GPU Memory Mapped Registers GMMx63C CFG_VOLTAGE_CONTROL See 2.5.1.4.2.1 [Software-Initiated NB Voltage Transitions]. Bits Description 31:16 Reserved. 15:8 VoltageLevel. Read-write. Reset: 0. Specifies the voltage requested when software toggles GMMx63C[VoltageChangeReq]. Values outside the D18F5x17C[MaxVid, MinVid] range are invalid and result in undefined behavior. 7:3 Reserved. 2 VoltageChangeReq. Read-write. Reset: 0. Software toggles this field to make VDDNB voltage requests. 1 VoltageChangeEn. Read-write. Reset: 0. Specifies whether changes to GMMx63C[VoltageChangeReq] causes voltage change requests. 1=Requests occur. 0=Requests do not occur. 0 VoltageForceEn. Read-write. Reset: 0. If GMMx63C[VoltageChangeEn]==1, this field specifies whether changes to GMMx63C[VoltageChangeReq] caused forced voltage changes. 1=Voltage changes are forced. 0=Voltage changes are not forced. GMMx640 CFG_VOLTAGE_STATUS See 2.5.1.4.2.1 [Software-Initiated NB Voltage Transitions]. Bits Description 31:9 Reserved. 8:1 CurrentVoltageLevel. Read-only. Reset: 0. Specifies the current voltage level requested by GMMx63C. See GMMx63C[VoltageLevel]. To determine the current voltage level, software must poll on this field until two consecutive reads return the same value. 0 VoltageChangeAck. Read-only. Reset: 0. Specifies whether the voltage change requested by GMMx63C[VoltageChangeReq] is complete. GMMx770 CG Voltage Control Reset: 0000_0006h. See 2.5.1.4.2 [Software-Initiated Voltage Transitions]. Bits Description 31:11 Reserved. 10 VoltageForceEn. Read-write. See GMMx63C[VoltageForceEn]. 9 VoltageChangeEn. Read-write. See GMMx63C[VoltageChangeEn]. 8:1 0 VoltageLevel. Read-write. See GMMx63C[VoltageLevel]. VoltageChangeReq. Read-write. See GMMx63C[VoltageChangeReq]. GMMx774 CG Voltage Status Reset: 0000_0000h. See 2.5.1.4.2 [Software-Initiated Voltage Transitions]. 464 42300 Rev 3.12 - July 14, 2015 Bits BKDG for AMD Family 15h Models 10h-1Fh Processors Description 31:9 Reserved. 8:1 0 CurrentVoltageLevel. Read-only. See GMMx640[CurrentVoltageLevel]. VoltageChangeAck. Read-only. See GMMx640[VoltageChangeAck]. GMMx7A0 LCLK_DEEP_SLEEP_CNTL Bits Description 31 EnableDs. Read-write. Reset: 0. Bits Definition 0 Deep Sleep Disabled. 1 Deep Sleep Enabled. 30 DmaactiveMask. Read-write. Reset: 1. Bits Definition 0 Allow deep sleep on DMAACTIVE_L 1 DMAACTIVE_L has to be deasserted to enter deep sleep 29 OnOutbWakeAckMask. Read-write. Reset: 1. Bits Definition 0 Allow deep sleep on OnOutbWakeAck 1 OnOutbWakeAck has to be deasserted to enter deep sleep 28 OnOutbWakeMask. Read-write. Reset: 1. Bits Definition 0 Allow deep sleep on OnOutbWake 1 OnOutbWake has to be deasserted to enter deep sleep 27 OnInbWakeAckMask. Read-write. Reset: 1. Bits Definition 0 Allow deep sleep on OnInbWakeAck 1 OnInbWakeAck has to be deasserted to enter deep sleep 26 OnInbWakeMask. Read-write. Reset: 1. Bits Definition 0 Allow deep sleep on OnInbWake 1 OnInbWake has to be deasserted to enter deep sleep 25 OrbIdleMask. Read-write. Reset: 1. Bits Definition 0 Do not wait for ORB_SMU_idle to enter deep sleep 1 ORB_SMU_idle has to be asserted to enter deep sleep 24 L2imuIdleMask. Read-write. Reset: 1. Bits Definition 0 Do not wait for L2IMU_SMU_idle to enter deep sleep 1 L2IMU_SMU_idle has to be asserted to enter deep sleep 23 L1imuintgenIdleMask. Read-write. Reset: 1. Bits Definition 0 Do not wait for L1IMUINTGEN_SMU_idle to enter deep sleep 1 L1IMUINTGEN_SMU_idle has to be asserted to enter deep sleep 465 42300 Rev 3.12 - July 14, 2015 BKDG for AMD Family 15h Models 10h-1Fh Processors 22 L1imubifIdleMask. Read-write. Reset: 1. Bits Definition 0 Do not wait for L1IMUBIF_SMU_idle to enter deep sleep 1 L1IMUBIF_SMU_idle has to be asserted to enter deep sleep 21 L1imugppsbIdleMask. Read-write. Reset: 1. Bits Definition 0 Do not wait for L1IMUGPPSB_SMU_idle to enter deep sleep 1 L1IMUGPPSB_SMU_idle has to be asserted to enter deep sleep 20 L1imugfxIdleMask. Read-write. Reset: 1. Bits Definition 0 Do not wait for L1IMUGFX_SMU_idle to enter deep sleep 1 L1IMUGFX_SMU_idle has to be asserted to enter deep sleep 19 PcieLclkIdle2Mask. Read-write. Reset: 1. Bits Definition 0 Do not wait for PCIE_SMU_lclk_idle2 to enter deep sleep 1 PCIE_SMU_lclk_idle2 has to be asserted to enter deep sleep 18 PcieLclkIdle1Mask. Read-write. Reset: 1. Bits Definition 0 Do not wait for PCIE_SMU_lclk_idle1 to enter deep sleep 1 PCIE_SMU_lclk_idle1 has to be asserted to enter deep sleep 17 SmuBusyMask. Read-write. Reset: 1. Bits Definition 0 Do not wait for smu stop to enter Deep Sleep 1 SMU stop has to be asserted to enter Deep Sleep. 16 SclkRunningMask. Read-write. Reset: 1. Bits Definition 0 Do not wait for GFX Idle to enter Deep Sleep 1 GFX should be idle to enter Deep sleep. 15:4 3 2:0 Hysteresis. Read-write. Reset: 0. RampDis. Read-write. Reset: 0. Bits Definition 0 Clock will be ramped up and down. 1 Clock ramping disabled. DivId. Read-write. Reset: 5h. Bits Definition 000b Clock OFF 001b Divide by 2. 010b Divide by 4. 011b Divide by 8. 100b Divide by 16. 101b Divide by 32. 110b Reserved. 111b Reserved. 466 42300 Rev 3.12 - July 14, 2015 BKDG for AMD Family 15h Models 10h-1Fh Processors 3.15 IOMMU Memory Mapped Registers See 3.1 [Register Descriptions and Mnemonics] for a description of the register naming convention. See 2.12.1 [IOMMU Configuration Space]. IOMMUx00 Device Table Base Address Low Bits Description 31:12 DevTblBase[31:12]: device table base address bits[31:12]. Read-write. Reset: 0. DevTblBase[51:12] = {IOMMUx04[DevTblBase[51:32], DevTblBase[31:12]}. DevTblBase[51:12] specifies the 4Kbyte-aligned base address of the first level device table. 11:9 Reserved1.Reserved for future use. This register controls no hardware 8:0 DevTblSize: device table size. Read-write. Reset: 0. This field contains 1 less than the length of the device table, in multiples of 4K bytes. A minimum size of 0 corresponds to a 4K byte device table and a maximum size of 1FFh corresponds to a 2M byte device table. IOMMUx04 Device Table Base Address High Bits Description 31:20 Reserved. 19:0 DevTblBase[51:32]: device table base address bits[51:32]. See: IOMMUx00[DevTblBase[31:12]]. IOMMUx08 Command Buffer Base Address Low Bits Description 31:12 ComBase[31:12]: command buffer base address bits[31:12]. Read-write. Reset: 0. ComBase[51:12] = {IOMMUx0C[ComBase[51:32], ComBase[31:12]]}. ComBase[51:12] specifies the 4Kbyte-aligned base address of the command buffer. 11:0 Reserved. IOMMUx0C Command Buffer Base Address High Bits Description 31:28 Reserved. 27:24 ComLen: command buffer length. Read-write. Reset: 8h. Specifies the length of the command buffer in power of 2 increments. The minimum size is 256 entries (4K bytes); values less than 8h are reserved. Bits Description 7h-0h Reserved. Fh-8h 2^ComLen entries (2^ComLen*16 bytes). 467 42300 Rev 3.12 - July 14, 2015 BKDG for AMD Family 15h Models 10h-1Fh Processors 23:20 Reserved. 19:0 ComBase[51:32]: command buffer base address bits[51:32]. See: IOMMUx08[ComBase[31:12]]. IOMMUx10 Event Log Base Address Low Bits Description 31:12 EventBase[31:12]: event log base address bits[31:12]. Read-write. Reset: 0. EventBase[51:12] = {IOMMUx14[EventBase[51:32], EventBase[31:12]}. EventBase[51:12] specifies the 4K-byte aligned base address of the event log. 11:0 Reserved. IOMMUx14 Event Log Base Address High Bits Description 31:28 Reserved. 27:24 EventLen: event log length. Read-write. Reset: 8h. Specifies the length of the event log in power of 2 increments. The minimum size is 256 entries (4K bytes); values less than 8h are reserved. Bits Description 7h-0h Reserved. Fh-8h 2^EventLen entries (2^EvemtLen*16 bytes). 23:20 Reserved. 19:0 EventBase[51:32]: event log base address bits [51:32]. See: IOMMUx10[EventBase[31:12]]. IOMMUx18 Control Low Bits Description 31:22 Reserved. 21:18 Tlpt. Read-write. Reset: 0. Tlpt contains the 4-bit value matched to the PCIe TLP Type field when the PCIe TLP Fmt value indicates the field carries a prefix. 17 GaEn. Read-write. Reset: 0. Guest APIC enable. 1 = loose. 0 = prohibited 16 GtEn. Read-write. Reset: 0. 1=Guest translation may be enabled for a peripheral by programming DTE[GV]. This bit must be programmed to zero when IOMMUx30[GtSup]=0. 15 PprEn. Read-write. Reset: 0. 1=Peripheral page service requests are processed. 0=Peripheral page service requests are treated as invalid device requests. This bit must be programmed to zero when IOMMUx30[PprSup]=0. 14 PprIntEn. Read-write. Reset: 0. 1=An interrupt is generated when IOMMUx2020[PprInt]=1 or IOMMUx2020[PprOverflow]=1. The interrupt vector used is indicated in D0F2x50[IommuMsiNumPpr].This bit must be programmed to zero when IOMMUx30[PprSup]=0. 468 42300 Rev 3.12 - July 14, 2015 BKDG for AMD Family 15h Models 10h-1Fh Processors 13 PprLogEn. Read-write. Reset: 0. 1=Peripheral page service request events are written to the peripheral page service request log when IommuEn=1. 0=Peripheral page service request logging is not enabled. Peripheral page service requests are discarded when PprLogEn=0 or IOMMUx30[PprSup]=0. When IommuEn=1 and software sets PprLogEn, the IOMMU clears IOMMUx2020[PprOverflow] and sets IOMMUx2020[PprRun]. The IOMMU can then write new entries to the event log if there are usable entries available. Software can read IOMMUx2020[PprRun] to determine the status of the peripheral page service request log. Note the peripheral page service request and event logs are independent. IOMMUx38, IOMMUx2030, and IOMMUx2038 must be programmed prior to setting PprLogEn. 12 CmdBufEn. Read-write. Reset: 0. 1=Start or restart command buffer processing. When CmdBufEn=1 and IommuEn=1, the IOMMU starts fetching commands and sets IOMMUx2020[CmdBufRun]. 0=Halt command buffer processing. Writing a 0 to this bit causes the IOMMU to cease fetching new commands although commands previously fetched are completed. The IOMMU stops fetching commands upon reset and after errors. See IOMMUx2020[CmdBufRun]. Writing of event log entries is independently controlled by EventLogEn. IOMMUx08, IOMMUx0C, IOMMUx2000, and IOMMUx2008 must be programmed prior to setting CmdBufEn. 11 Isoc. Read-write. Reset: 0. This bit controls the state of the isochronous bit in the HyperTransport read request packet when the IOMMU issues I/O page table reads and device table reads on the HyperTransport link. 1=Request packet to use isochronous channel. 0=Request packet to use standard channel. This bit must be set if IommuEn==1. 10 Coherent. Read-write. Reset: 1. This bit controls the state of the coherent bit in the HyperTransport read request packet when the IOMMU issues device table reads on the HyperTransport link. 1=Device table requests are snooped by the processor. 0=Device table requests are not snooped by the processor. 9 ResPassPw. Read-write. Reset: 0. This bit controls the state of the ResPassPW bit in the HyperTransport read request packet when the IOMMU issues I/O page table reads and device table reads on the HyperTransport link. 1=Response may pass posted requests. 0=Response may not pass posted requests. 8 PassPw. Read-write. Reset: 0. This bit controls the state of the PassPW bit in the HyperTransport read request packet when the IOMMU issues I/O page table reads and device table reads on the HyperTransport link.. 1=Request packet may pass posted requests. 0=Request packet may not pass posted requests. 7:5 InvTimeout. Read-write. Reset: 0. This field specifies the invalidation timeout for IOTLB invalidation requests. Bits Description 000b No timeout. 001b 1 ms. 010b 10 ms. 011b 100 ms. 100b 1 sec. 101b 10 sec. 111b-110b Reserved 4 ComWaitInten. Read-write. Reset: 0. 1=An interrupt is generated when IOMMUx2020[ComWaitInt]=1. 3 EventIntEn. Read-write. Reset: 0. 1=An interrupt is generated when IOMMUx2020[EventLogInt]=1 or IOMMUx2020[EventOverflow]=1. 469 42300 Rev 3.12 - July 14, 2015 BKDG for AMD Family 15h Models 10h-1Fh Processors 2 EventLogEn. Read-write. Reset: 0. 1=All events detected are written to the event log when IommuEn=1. 0=Event logging is not enabled. Events are discarded when the event log is not enabled. When IommuEn=1 and software sets EventLogEn, the IOMMU clears IOMMUx2020[EventOverflow] and sets IOMMUx2020[EventLogRun]. IOMMUx10, IOMMUx14, IOMMUx2010, and IOMMUx2018 must be programmed prior to setting EventLogEn. 1 HtTunEn. Read-write. Reset: 0. 1= Upstream traffic received by the HyperTransport tunnel is translated by the IOMMU. 0=Upstream traffic received by the HyperTransport tunnel is not translated by the IOMMU. The IOMMU ignores the state of this bit while IommuEn=0. See D0F2x40[IommuHtTunnelSup]. 0 IommuEn. Read-write. Reset: 0. 1=IOMMU enabled. All upstream transactions are translated by the IOMMU. IOMMUx00 [Device Table Base Address Low] and IOMMUx04 [Device Table Base Address High] must be configured by software before setting this bit. 0=IOMMU is disabled and no upstream transactions are translated or remapped by the IOMMU. When disabled, the IOMMU does not read any commands or create any event log entries. IOMMUx20 Exclusion Range Base Low Bits Description 31:12 ExclBase[31:12]: exclusion range base address bits[31:12]. Read-write. Reset: 0. ExclBase[51:12] = {IOMMUx20[ExclBase[51:32]], ExclBase[31:12]}. Specifies the 4Kbyte-aligned base address of the exclusion range. 11:2 Reserved. 1 ExAllow: exclusion allow. Read-write. Reset: 0. 1=All accesses to the exclusion range are forwarded untranslated. 0=The EX bit in the device table entry specifies if accesses to the exclusion range are translated. 0 ExEn: exclusion enable. Read-write. Reset: 0. 1=The exclusion range is enabled. IOMMUx24 Exclusion Range Base High Bits Description 31:20 Reserved. 19:0 ExclBase[51:32]: exclusion range base address bits[51:32]. See: IOMMUx20[ExclBase[31:12]]. IOMMUx28 Exclusion Range Limit Low Bits Description 31:12 ExclLimit[31:12]: exclusion range limit address bits[31:12]. Read-write. Reset: 0. ExclLimit[51:12] = {IOMMUx2C[ExclLimit[51:32]], ExclLimit[31:12]}. ExclLimit[51:12] specfies the 4Kbyte-aligned limit address of the exclusion range. 11:0 Reserved. 470 42300 Rev 3.12 - July 14, 2015 BKDG for AMD Family 15h Models 10h-1Fh Processors IOMMUx2C Exclusion Range Limit High Bits Description 31:20 Reserved. 19:0 ExclLimitHi. See: IOMMUx28[ExclLimit[31:12]]. IOMMUx30 Extended Feature Low Bits Description 31:15 Reserved. 14:13 GlxSup. Read-only. Reset: 00b. 01b=Two-level GCR3 base address table is supported in hardware. 0=single-level GCR3 base table address translation is supported. 00b=GLX in the DTE is ignored and the IOMMU performs only single-level guest CR3 lookups. This value is not meaningful when GtSup=0. 12 Reserved. 11:10 HATS: host address translation size. Read-only. Reset: 10b. The maximum number of host address translation levels supported. This value is not meaningful when GtSup=0. IOMMU behaviour is undefined if Next Level in a page directory entry exceeds the limit set by HATS. Bits Description 00b 4 levels. 01b 5 levels. 10b 6 levels. 11b Reserved. 9 PcSup: performance counters supported. Read-only. Reset: 1. 1=performance counters are supported. 8 HeSup: hardware error registers supported. Read-only. Reset: 0. 0=Hardware error registers do not report error information.. 1=Error information is reported in hardware error registers. 7 GaSup: guest APIC supported. Read-only. Reset: 0. 1=Guest APIC supported. 6 IaSup: INVALIDATE_IOMMU_ALL supported. Read-only. Reset: 1. 1=The INVALIDATE_IOMMU_ALL command is supported. 5 Reserved. 4 GtSup: guest translation supported. Read-only. Reset: 1. 1=Guest address translation is supported. 0=Only nested address translation is supported. When GtSup=0, the following values in the DTE must be zero: GV, GLX and GCR3 Table Root Pointer. See IOMMUx18[GtEn]. 3 NxSup: no execute supported. Read-only. Reset: 0. 1=No-execute protection is supported.. 0=No-execute protection is not supported. 2 XtSup: x2 apic supported. Read-only. Reset: 0. 1=The interrupt remapping table is expanded to support x2APIC interrupt information.. 0=x2APIC support is disabled. 1 PprSup: peripheral page service request supported. Read-only. Reset: 1. 1=Indicates that IOMMU handles page service request events from peripherals, the IOMMU supports the page service request queue, and that the second IOMMU interrupt can be used to signal peripheral page service request events. 0 PrefSup: prefetch support. Read-only. Reset: 1. 1=Indicates that IOMMU will accept PREFETCH_IOMMU_PAGES commands. 471 42300 Rev 3.12 - July 14, 2015 BKDG for AMD Family 15h Models 10h-1Fh Processors IOMMUx34 Extended Feature High Bits Description 31:4 Reserved. 3:0 PasMax: PASID maximum. Read-only. Reset: 8h. This specifies the maximum PASID value supported. This field is not meaningful when IOMMUx34[GtSup]=0. Bits Description 2h-0h Reserved. Fh-3h 2^(PasMax+1)-1. IOMMUx38 PPR Log Base Address Low Bits Description 31:12 PprBase[31:12]: ppr base address bits[31:12]. Read-write. Reset: 0. PprBase[51:12] = {IOMMUx3C[PprBase[51:32]], PprBase[31:12]}. PprBase[51:12] specifies the 4Kbyte-aligned base address of the PPR log. 11:0 Reserved. IOMMUx3C PPR Log Base Address High Bits Description 31:28 Reserved. 27:24 PprLen: ppr length. Read-write. Reset: 8h. Specifies the length of the PPR log in power of two increments. Bits Description 7h-0h Reserved. Fh-8h 2^PprLen entries (2^PprLen*16 bytes). 23:20 Reserved. 19:0 PprBase[51:32]: ppr base address bits[51:32]. See: IOMMUx38[31:12]. IOMMUx40 Hardware Error Upper Low Bits Description 31:0 FirstEvCode[31:0]: first event code bits[31:0]. Read-write. Reset: 0. FirstEvCode[59:0] = {IOMMUx44[FirstEvCode[59:32]], FirstEvCode[31:0]}. IOMMUx44[EvCode] and FirstEvCode[59:0] specify the upper 64 bits of the most recent hardware error detected by the IOMMU. IOMMUx44 Hardware Error Upper High Bits Description 31:28 EvCode: event code. Read-write. Reset: 0. Event code for the type of error logged. 27:0 FirstEvCode[59:32]: first event code bits[59:32]. See: IOMMUx40[FirstEvCode[31:0]]. 472 42300 Rev 3.12 - July 14, 2015 BKDG for AMD Family 15h Models 10h-1Fh Processors IOMMUx48 Hardware Error Lower Low Bits Description 31:0 SecondEvCode[31:0]: second event code bits[31:0]. Read-write. Reset: 0. SecondEvCode[63:0] = {IOMMUx4C[SecondEvCode[63:32]], SecondEvCode[31:0]}. SecondEvCode[63:0] specifies the lower 64 bits of the most recent hardware error detected by IOMMU. IOMMUx4C Hardware Error Lower High Bits Description 31:0 SecondEvCode[63:32]: second event code bits[63:32]. See: IOMMUx48[SecondEvCode[31:0]]. IOMMUx50 Hardware Error Status Bits Description 31:2 Reserved. 1 HEO: hardware error overflow. Read-write. Reset: 0. Defines the contents of the IOMMU hardware error registers as having beeing overwritten.. 0=not overwritten.. 1=contents overwritten by new information.. HEO is not meaningful when HEV=0. 0 HEV: hardware error valid. Read-write. Reset: 0. 1=Contents of the IOMMU hardware error registers are valid. IOMMUx2000 Command Buffer Head Pointer Bits Description 31:19 Reserved. 18:4 CmdHdptr: command buffer head pointer. Read-write. Reset: 0. Specifies the 128-bit aligned offset from the command buffer base address register of the next command to be fetched by the IOMMU. The IOMMU increments this register, rolling over to zero at the end of the buffer, after fetching and validating the command in the command buffer. After incrementing this register, the IOMMU cannot re-fetch the command from the buffer. If this register is written by software while IOMMUx2020[CmdBufRun]=1, the IOMMU behavior is undefined. If this register is set by software to a value outside the length specified by IOMMUx08[ComLen], the IOMMU behavior is undefined. 3:0 Reserved. IOMMUx2008 Command Buffer Tail Pointer Bits Description 31:19 Reserved. 473 42300 Rev 3.12 - July 14, 2015 BKDG for AMD Family 15h Models 10h-1Fh Processors 18:4 CmdTailptr: command buffer tail pointer. Read-write. Reset: 0. Specifies the 128-bit aligned offset from the command buffer base address register of the next command to be written by the software. Software must increment this field, rolling over to zero at the end of the buffer, after writing a command to the command buffer. If software advances the tail pointer equal to or beyond the head pointer after adding one or more commands to the buffer, the IOMMU behavior is undefined. If software sets the command buffer tail pointer to an offset beyond the length of the command buffer, the IOMMU behavior is undefined. 3:0 Reserved. IOMMUx2010 Event Log Head Pointer Bits Description 31:19 Reserved. 18:4 EventHdptr: event log head pointer. Read-write. Reset: 0. Specifies the 128 bit aligned offset from the event log base address register that will be read next by software. Software must increment this field, rolling over at the end of the buffer, after reading an event from the event log. If software advances the head pointer beyond the tail pointer, the IOMMU behavior is undefined. If software sets the event log head pointer to an offset beyond the length of the event log, the IOMMU behavior is undefined. 3:0 Reserved. IOMMUx2018 Event Log Tail Pointer Bits Description 31:19 Reserved. 18:4 EventTailptr: event log tail pointer. Read-write. Reset: 0. Specifies the 128-bit aligned offset from the event log base address register that will be written next by the IOMMU when an event is detected. The IOMMU increments this register, rolling over at the end of the buffer, after writing an event to the event log. If this register is written while IOMMUx2020[EventLogRun]=1, the IOMMU behavior is undefined. If this register is set by software to a value outside the length specified by IOMMUx10[EventLen], the IOMMU behavior is undefined. 3:0 Reserved. IOMMUx2020 Status Bits Description 31:8 Reserved1.Reserved for future use. This register controls no hardware 7 PprRun: peripheral page service request running. Read-only. Reset: 0. 1=PPR requests are logged as they occur. 0=PPR requests are discarded without logging.. When PprOverflow=1, the IOMMU does not write new PPR log entries even when PprRun=1. When halted, PPR request logging is restarted by using IOMMUx18[PprLogEn]. 6 PprInt: peripheral page service request interrupt. Read-write; Write-1-to-clear. Reset: 0. 1=PPR request entry written to the PPR log by the IOMMU. 0=No PPR entry written to the PPR log by the IOMMU. See IOMMUx18[PprIntEn]. 474 42300 Rev 3.12 - July 14, 2015 BKDG for AMD Family 15h Models 10h-1Fh Processors 5 PprOverflow: peripheral page service request overflow. Read-write; Write-1-to-clear. Reset: 0. 1=IOMMU PPR log overflow has occured. This bit is set when a new peripheral page service request is to be written to the PPR log and there is no usable entry in the PPR log, causing the new information to be discarded. No new PPR log entries are written while this bit is set. See IOMMUx18[PprIntEn]. 4 CmdBufrun: command buffer running. Read-only. Reset: 0. 1=Commands may be fetched from the command buffer. 0=IOMMU has stopped fetching new commands. The IOMMU freezes command processing after COMMAND_HARDWARE_ERROR or ILLEGAL_COMMAND_ERROR errors. When frozen, command fetching is restarted by using IOMMUx18[CmdBufEn]. 3 EventLogrun: event log running. Read-only. Reset: 0. 1=Events are logged as they occur.. 0=Event reports are discarded without logging. When EventOverflow=1, the IOMMU does not write new event log entries even when EventLogRun=1. When halted, event logging is restarted by using IOMMUx18[EventLogEn]. 2 ComwaitInt: completion wait interrupt. Read-write; Write-1-to-clear. Reset: 0. 1=COMPLETION_WAIT command completed. This bit is only set if the i bit is set in the COMPLETION_WAIT command. See IOMMUx18[ComWaitIntEn]. 1 EventLogInt: event log interrupt. Read-write; Write-1-to-clear. Reset: 0. 1=Event entry written to the event log by the IOMMU. See IOMMUx18[EventIntEn]. 0 EventOverflow. Read-write; Write-1-to-clear. Reset: 0. 1=IOMMU event log overflow has occurred. This bit is set when a new event is to be written to the event log and there is no usable entry in the event log, causing the new event information to be discarded. No new event log entries are written while this bit is set. See IOMMUx18[EventIntEn]. IOMMUx2030 PPR Log Head Pointer Bits Description 31:19 Reserved. 18:4 PprHdptr: ppr head pointer. Read-write. Reset: 0. Specifies the 128-bit aligned offset from the PPR log base address register that will be read next by software. Software must increment this field, rolling over at the end of the buffer, after reading a PPR request entry from the PPR event log. If software advances the head pointer beyond the tail pointer, the IOMMU behavior is undefined. If software sets the PPR log head pointer to an offset beyond the length of the PPR log, the IOMMU behavior is undefined. 3:0 Reserved. IOMMUx2038 PPR Log Tail Pointer Bits Description 31:19 Reserved. 475 42300 Rev 3.12 - July 14, 2015 BKDG for AMD Family 15h Models 10h-1Fh Processors 18:4 PprTailptr. Read-write. Reset: 0. Specifies the 128-bit aligned offset from the PPR log base address register that will be written next by the IOMMU when a PPR request is detected. The IOMMU increments this register, rolling over at the end of the buffer, after writing a PPR request to the PPR log. If this register is written while IOMMUx2020[PprRun]=1, the IOMMU behavior is undefined. If software sets the PPR log tail pointer to an offset beyond the length of the PPR log, defined by IOMMUx3C[PprLen], the IOMMU behavior is undefined. 3:0 Reserved. IOMMUx4000 Counter Configuration Bits Description 31:18 Reserved. 17:12 NCounterBanks: number of counter banks. Read-only. Reset: 2h. The number of counter banks supported by the IOMMU. Each bank contains two or more counter and control registers as specified by NCounter. For each counter bank, a corresponding control bit is in IOMMUx4008, IOMMUx4010, and IOMMUx4018. Each supported event counter bank is in a distinct, consecutive 4K byte page. Bits Description 00h No counter banks supported. 3Fh-01h NCounterBanks event counter banks are supported. Note: IOMMU event counter banks are numbered starting with 0. 11 Reserved. 10:7 NCounter: number of counters per bank. Read-only. Reset: 4h. Reports the number of individual counters in each IOMMU counter bank. Each counter bank contains the same number of counters. Bits Description 0h No counters supported. 1h Reserved. Fh-2h NCounter counters in each bank. 6:0 Reserved. IOMMUx4008 Counter PASID Bank Lock Low Bits Description 31:0 PasidLock[31:0]: pasid lock enable bits[31:0]. Read-write. Reset: 0. PasidLock[63:0] = {IOMMUx400C[PasidLock[63:32]], PasidLock[31:0]}. For each bit in PasidLock[63:0], if the bit is set then writes to the corresponding bank in IOMMUx4[1,0][3:0]10 and IOMMUx4[1,0][3:0]14 are ignored. Bit positions above the value reported in IOMMUx4000[NCounterBanks] are ignored when written and return zero when read. IOMMUx400C Counter PASID Bank Lock High Bits Description 31:0 PasidLock[63:32]. See: IOMMUx4008[PasidLock[31:0]]. 476 42300 Rev 3.12 - July 14, 2015 BKDG for AMD Family 15h Models 10h-1Fh Processors IOMMUx4010 Domain Bank Lock Low Bits Description 31:0 DomainLock[31:0]: domain lock enable bits[31:0]. Read-write. Reset: 0. DomainLock[63:0] = {IOMMUx4014[DomainLock[63:32]], DomainLock[31:0]}. For each bit in DomainLock[63:0], if the bit is set then writes to the corresponding bank in IOMMUx4[1,0][3:0]18 and IOMMUx4[1,0][3:0]1C are ignored. Bit positions above the value reported in IOMMUx4000[NCounterBanks] are ignored when written and return zero when read. IOMMUx4014 Domain Bank Lock High Bits Description 31:0 DomainLock[63:32]. See: IOMMUx4010[DomainLock[31:0]]. IOMMUx4018 DeviceID Bank Lock Low Bits Description 31:0 DevIDLock[31:0]: deviceID lock enable bits[31:0]. Read-write. Reset: 0. DevIDLock[63:0] = {IOMMUx401C[DevIDLock[63:32]], DevIDLock[31:0]}. For each bit in DevIDLock[63:0], if the bit is set then writes to the corresponding bank in IOMMUx4[1,0][3:0]20 and IOMMUx4[1,0][3:0]24 are ignored. Bit positions above the value reported in IOMMUx4000[NCounterBanks] are ignored when written and return zero when read. IOMMUx401C DeviceID Bank Lock High Bits Description 31:0 DevIDLock[63:32]. See: IOMMUx4018[DevIDLock[31:0]]. IOMMUx4[1,0][3:0]00 Counter Low Table 184: Block to register mapping for IOMMUx4[1,0][3:0]00 Register IOMMUx40000 IOMMUx40100 IOMMUx40200 IOMMUx40300 Function Bank 0 Counter 0 Bank 0 Counter 1 Bank 0 Counter 2 Bank 0 Counter 3 Register IOMMUx41000 IOMMUx41100 IOMMUx41200 IOMMUx41300 Function Bank 1 Counter 0 Bank 1 Counter 1 Bank 1 Counter 2 Bank 1 Counter 3 Bits Description 31:0 Icounter[31:0]. Read-write. Reset: 0. Icounter[47:0] = {IOMMUx4[1,0][3:0]04[Icounter[47:32]], Icounter[31:0]}. Icounter[47:0] reports the counter value. The counter counts up continuously, wrapping at the maximum value. There is no overflow indicator. 477 42300 Rev 3.12 - July 14, 2015 BKDG for AMD Family 15h Models 10h-1Fh Processors IOMMUx4[1,0][3:0]04 Counter High Table 185: Block to register mapping for IOMMUx4[1,0][3:0]04 Register IOMMUx40004 IOMMUx40104 IOMMUx40204 IOMMUx40304 Bits Function Bank 0 Counter 0 Bank 0 Counter 1 Bank 0 Counter 2 Bank 0 Counter 3 Register IOMMUx41004 IOMMUx41104 IOMMUx41204 IOMMUx41304 Function Bank 1 Counter 0 Bank 1 Counter 1 Bank 1 Counter 2 Bank 1 Counter 3 Description 31:16 Reserved. 15:0 Icounter[47:32]. See: IOMMUx4[1,0][3:0]00[Icounter[31:0]]. IOMMUx4[1,0][3:0]08 Counter Source Table 186: Block to register mapping for IOMMUx4[1,0][3:0]08 Register IOMMUx40008 IOMMUx40108 IOMMUx40208 IOMMUx40308 Bits Function Bank 0 Counter 0 Bank 0 Counter 1 Bank 0 Counter 2 Bank 0 Counter 3 Register IOMMUx41008 IOMMUx41108 IOMMUx41208 IOMMUx41308 Function Bank 1 Counter 0 Bank 1 Counter 1 Bank 1 Counter 2 Bank 1 Counter 3 Description 31 Cac: counter source architectural or custom. Read-write. Reset: 0. 0=Architectural counter input group. 1=Custom input group. 30 CountUnits. Read-write. Reset: 0. 0=Counter counts events (level). 1=Counter counts clocks (edges). 29:8 Reserved. 7:0 Csource: counter source. Read-write. Reset: 0. Counter source. Selects event counter input from the choices provided. IOMMUx4[1,0][3:0]10 PASID Match Low See IOMMUx4008. Table 187: Block to register mapping for IOMMUx4[1,0][3:0]10 Register IOMMUx40010 IOMMUx40110 IOMMUx40210 IOMMUx40310 Function Bank 0 Counter 0 Bank 0 Counter 1 Bank 0 Counter 2 Bank 0 Counter 3 Register IOMMUx41010 IOMMUx41110 IOMMUx41210 IOMMUx41310 Function Bank 1 Counter 0 Bank 1 Counter 1 Bank 1 Counter 2 Bank 1 Counter 3 478 42300 Rev 3.12 - July 14, 2015 Bits 31 BKDG for AMD Family 15h Models 10h-1Fh Processors Description PasMEn: PASID match enable. Read-write. Reset: 0. 0=PASID is ignored. 1=Filtered PASID must match to count an event. An event with no PASID tag is only counted when Pasmen=0. 30:16 Reserved. 15:0 PasidMatch. Read-write. Reset: 0. This value is compared with the masked (filtered) value of the incoming PASID of the transaction to decide if the corresponding event is counted. The event is counted if PasidMatch is exactly equal to the masked incoming PASID. IOMMUx4[1,0][3:0]14 PASID Match High See IOMMUx4008. Table 188: Block to register mapping for IOMMUx4[1,0][3:0]14 Register IOMMUx40014 IOMMUx40114 IOMMUx40214 IOMMUx40314 Bits Function Bank 0 Counter 0 Bank 0 Counter 1 Bank 0 Counter 2 Bank 0 Counter 3 Register IOMMUx41014 IOMMUx41114 IOMMUx41214 IOMMUx41314 Function Bank 1 Counter 0 Bank 1 Counter 1 Bank 1 Counter 2 Bank 1 Counter 3 Description 31:16 Reserved. 15:0 PasidMask. Read-write. Reset: 0. This bit-mask is ANDed with the PASID of the transaction to decide to count the corresponding event. 0=Count events for all values of incoming PASID. 0001h-FFFFh=Bit-wise mask ANDed with incoming PASID. IOMMUx4[1,0][3:0]18 Domain Match Low Table 189: Block to register mapping for IOMMUx4[1,0][3:0]18 Register IOMMUx40018 IOMMUx40118 IOMMUx40218 IOMMUx40318 Bits 31 Function Bank 0 Counter 0 Bank 0 Counter 1 Bank 0 Counter 2 Bank 0 Counter 3 Register IOMMUx41018 IOMMUx41118 IOMMUx41218 IOMMUx41318 Function Bank 1 Counter 0 Bank 1 Counter 1 Bank 1 Counter 2 Bank 1 Counter 3 Description DomMEn: domain match enable. Read-write. Reset: 0. 0=Domain is ignored. 1=Filtered Domain must match DomainMatch to count an event. 30:16 Reserved. 15:0 DomainMatch. Read-write. Reset: 0. This value is compared with the masked (filtered) value of the incoming Domain of the transaction to decide to count the corresponding event. The event is counted if DomainMatch is exactly equal to the masked incoming Domain. 479 42300 Rev 3.12 - July 14, 2015 BKDG for AMD Family 15h Models 10h-1Fh Processors IOMMUx4[1,0][3:0]1C Domain Match High Table 190: Block to register mapping for IOMMUx4[1,0][3:0]1C Register IOMMUx4001C IOMMUx4011C IOMMUx4021C IOMMUx4031C Bits Function Bank 0 Counter 0 Bank 0 Counter 1 Bank 0 Counter 2 Bank 0 Counter 3 Register IOMMUx4101C IOMMUx4111C IOMMUx4121C IOMMUx4131C Function Bank 1 Counter 0 Bank 1 Counter 1 Bank 1 Counter 2 Bank 1 Counter 3 Description 31:16 Reserved. 15:0 DomainMask. Read-write. Reset: 0. This bit-mask is ANDed with the Domain of the transaction to decide to count the corresponding event. 0000h=Count events for all values of incoming Domain. 0001h-FFFFh=Bit-wise mask ANDed with incoming Domain. IOMMUx4[1,0][3:0]20 DeviceID Match Low Table 191: Block to register mapping for IOMMUx4[1,0][3:0]20 Register IOMMUx40020 IOMMUx40120 IOMMUx40220 IOMMUx40320 Bits 31 Function Bank 0 Counter 0 Bank 0 Counter 1 Bank 0 Counter 2 Bank 0 Counter 3 Register IOMMUx41020 IOMMUx41120 IOMMUx41220 IOMMUx41320 Function Bank 1 Counter 0 Bank 1 Counter 1 Bank 1 Counter 2 Bank 1 Counter 3 Description DidMEn: deviceID match enable. Read-write. Reset: 0. 0=DeviceID is ignored. 1=Filtered DeviceID must match to count an event. 30:16 Reserved. 15:0 DeviceidMatch. Read-write. Reset: 0. This value is compared with the masked (filtered) value of the incoming DeviceID of the transaction to decide to count the corresponding event. The event is counted if DeviceidMatch is exactly equal to the masked incoming DeviceID. IOMMUx4[1,0][3:0]24 DeviceID Match High Table 192: Block to register mapping for IOMMUx4[1,0][3:0]24 Register IOMMUx40024 IOMMUx40124 IOMMUx40224 IOMMUx40324 Function Bank 0 Counter 0 Bank 0 Counter 1 Bank 0 Counter 2 Bank 0 Counter 3 Register IOMMUx41024 IOMMUx41124 IOMMUx41224 IOMMUx41324 Function Bank 1 Counter 0 Bank 1 Counter 1 Bank 1 Counter 2 Bank 1 Counter 3 480 42300 Rev 3.12 - July 14, 2015 Bits BKDG for AMD Family 15h Models 10h-1Fh Processors Description 31:16 Reserved. 15:0 DeviceidMask. Read-write. Reset: 0. This bit-mask is ANDed with the DeviceID of the transaction to decide to count the corresponding event. 0=Count events for all values of incoming DeviceID. 0001h-FFFFh= Bit-wise mask ANDed with incoming DeviceID. IOMMUx4[1,0][3:0]28 Counter Report Low Table 193: Block to register mapping for IOMMUx4[1,0][3:0]28 Register IOMMUx40028 IOMMUx40128 IOMMUx40228 IOMMUx40328 Function Bank 0 Counter 0 Bank 0 Counter 1 Bank 0 Counter 2 Bank 0 Counter 3 Register IOMMUx41028 IOMMUx41128 IOMMUx41228 IOMMUx41328 Function Bank 1 Counter 0 Bank 1 Counter 1 Bank 1 Counter 2 Bank 1 Counter 3 Bits Description 31:0 EventNote[31:0]. Read-write. Reset: 0. EventNote[51:0] = {IOMMUx4[1,0][3:0]2C[EventNode[51:32]], EventNode[31:0]}. When IOMMUx4[1,0][3:0]2C[CERE]=1 and the corresponding counter is incremented and wraps to zero, EventNote[51:0] is reported in the EVENT_COUNTER_ZERO e
advertisement
Related manuals
advertisement