Digital Equipment Corporation MicroVAX II MicroVAX II System, MicroVAX 78032, MicroVAX 78132, VAX-11 780 VAX-11/780, VAX 8600 VAX 8600 processor, VAX 8200, MicroVMS, ULTRIX 32 ULTRIX-32, VAX ELN VAXE LN, TK 50 TK50 Technical Journal
Hirdetés
Hirdetés
o
MicroVAX
II System
Digital Technical Journal ofDigital Equipment Corporation
Number 2
March
1986
Cover Design
Hardware, software, and peripheral devices for the
Micro
VAX 11
system are featured in this issue. Two
VLSI
devices, the
78032
CPU chip and the
78132
FPU chip, form the core of this system. Our cover shows the input programmable logic array for the FPU chip.
The cover was designed by Deborah Falck of the Graphic
Design Department.
Editorial Staff
Editor - Richard W. Beane
Production Staff
Production Editor- M. Terri Autieri
Designer - Charlotte Bell
Typesetting Programmer -James K. Scarsdale
Advisory Board
Samuel H. Fuller, Chairman
Robert M. Glorioso
John W . McCredie
John
F.
Mucci
Mahendra R. Patel
Grant F. Saviers
William D. Strecker
Maurice V. Wilkes
The Digital Technical journal is published by Digital
Equipment Corporation, 77 Reed Road, Hudson,
Massachusetts 01749.
Comments on the content of any paper are welcomed.
Write to the editor at Mail Stop HL02-3/KI1 at the published-by address.
Comments can also be sent on the ENET to
RDVAX::BEANE or on the ARPANET to
BEANE%RDVAX.DEC@DECWRL.
Copyright
©
1986 Digital Equipment Corporation.
Copying without fee is permitted provided that such copies are made for use in educational institutions by faculty members and are not distributed for commer cial advantage. Abstracting with credit of Digital
Equipment Corporation's authorship is permitted.
Requests for other copies for a fee may be made to the
Digital Press of Digital Equipment Corporation. All rights reserved.
The information in this journal is subject to change without notice and should not be construed as a com mitment by Digital Equipment Corporation. Digital
Equipment Corporation assumes no responsibility for any errors that may appear in this document.
ISBN 932376-89-4
Documentation Number EY-3474E-DP
The following are trademarks of Digital Equipment
Corporation: CompacTape, DEC, the Digital logo,
MicroVAX, MicroVAX I, MicroVAX II, MicroVMS,
PDP-7, PDP-I I , Q-BUS, RSTS, TK50, ULTRIX,
ULTRIX-32, UN113US, VAX, VAX-ll/730, VAX-ll/750,
VAX-ll/780, VAX 8600, VAX 8200, VAXELN,
VAXstation, VMS, VT.
Apple II is a trademark of Apple Computer, Inc.
AT&T is a trademark of American Telephone & Tele graph Company.
IBM is a registered trademark of International Business
Machines, Inc.
Mylar is a trademark of E. I. duPont deNemours &
Company.
Tek is a registered trademark of Tektronix, Inc.
UNIX and System V are trademarks of AT&T Bell
Laboratories.
Xerox is a registered trademark of Xerox Corporation.
68000 is a trademark of Motorola, Inc.
8086 and Intel are trademarks of Intel Corporation.
The manuscript for this book was created using generic coding and, via a translation program, was automatically typeset. Book production was done by
Educational Services Media Communications Group in
Bedford,
MA.
Contents
8
Foreword
jeffrey
C.
Kalb
12
The MicroVAX 78032 Chip, A 32-Bit Microprocessor
Daniel W. Dobberpuhl. Robert M. Supnik, Richard
T.
Witek
24
The MicroVAX 78132 Floating Point Chip
William R. Bidermann, Amnon Fisher, Burton M. Leary,
Robert J Simcoe, William R. Wheeler
3 7
Developing the Micro VAX II CPU Board
Barry A. Maskas
48
The Evolution of the Custom CAD Suite Used on the
Micro VAX
II
System
Anthony F. Hutchings
56
The Making of a Micro VAX Workstation
Rick Spitz, Peter George. Stephen Zalewski
66
The RQDX3 Design Project
Nicholas
A.
Warchol, Stephen
F.
Shirron
76
The Evolution of Instruction Emulation for the
MicroVAX Systems
Kathleen D. Morse, Lawrence). Kenah
86
The TK50 Cartridge Tape Drive
Steven
E.
Doone, Guemer
E.
Schneider
99
Porting ULTRIX Software to the Micro VAX System
Raymond J. Lanza
New Products
Editor,s Introduction
Richard W. Beane
F.ditor
This issue of the journal is the second pub lished by Digital's engineering organization.
Our first issue ( August 198 5 ) fe atured papers about the technologies us ed i n designing the VAX 8600 processor. The jour nal presents papers written by the technical contributors who design Digital's products .
The information is directed at engineering faculty members, Digital's own engi neers, and customers.
This issue features the M icroV AX ll system , which imp lements the VAX architecture o n a single CPU chip, the 780 3 2 . Another chip, the 7813 2 , executes fast floating point oper ations ; a si ngle board holds both those chips, plus one m egabyte of memory. New per i pherals have been designed, and the VMS a n d ULTRIX s o ftware a d a p ted to the
MicroVAX II syste m . This col l ecti o n of papers, by authors from different engineer ing groups, presents a wide spectrum of the
MicroVAX II hardware and software.
The first paper, by Dan Dobberpuhl , Bob
Supnik, and Rich Witek, is a description of the 780 3 2 CPU chip, which imp le ments a subset of the full VAX i nstruction set. The decisions a bout w hi c h i n s truc t i o n s to microcode are discussed , along with hard ware simplifications needed to fit functions on one chi p . The chip's various operations are explained, with emphasis on paral le l execution .
The CPU chip can use a coprocessor, the
78 1 3 2 FPU chip , to perform fast floating p oint o p e ra t i o n s . T h e p a p e r by B i ll
Bidermann, Amnon Fisher, Mike Leary, Bob
Simcoe , and Bill Wheeler relates t he 7813 2's arch itecture and algori thms. The protocol between the two chips is discussed and a description is given of the wiring and signal i n t e g r i ty i s sues a n d how t h e y were addressed.
Both chips are mounted on a single board containing one megabyte of me mory. Barry
Maskas' paper exp lains how the CPU board had to be designed as a linked sequential machine with dual ports . The development process is interesting because the board and the chips were designed in para l lel.
The paper on CAD tools , by Tony Hutch i ngs, relates the large role they played in the chi p and board designs . The various levels of
CAD support , from be havioral modeling, t hrough logic and circuit simulation, to wire list generation is described.
The software gra p hics t hat turn t he
MicroVAX II system into a single-user work station are reported in the paper by Rick
Spitz , Peter George, and Steve Zalewski . The control of windowing software and virtual displays is discussed, as are the implementa tion details.
The RQDX3 disk controller provides fast data transfers between a CPU and disk stor age devices. Nick Warc hol and Stephen Shir ron explain the top-down development pro cess that lead to unique solutions to difficult problems. Their descri ption of the fi nal architecture shows how the original goa ls were met in the eventual design.
With a subset architecture , those instruc t ions not in the set have to be executed another way. The paper by Kathy Morse and
Larry Kenah describes the macrocode emula tion of the VMS changes required to do that.
The testi ng techniques are i nteresting since they were done without M icroVAX hardware .
The paper by Steve Boone and Guenter
Schne ider describes the TKSO, a streaming cartridge tape drive providing fast data trans fer. The authors d iscuss the unique cartridge, tape transport , and controller designs, high
Iighting the self-thread i ng technique and the serpentine readjwrite process .
The fi nal paper, by Ray Lanza , describes porting the ULTRIX- 3 2 software to the
MicroVAX processor. Ray explains the cross development environment and the mapping techniques that allowed the heart of the
ULTRIX software to fit on a small system .
3
4
Biographies
William
R.
Bidermann Bill Bidermann is the engineering manager of the Advanced Development Memory Group. He consu lted on the float ing point chips for both the VAX 8 2 0 0 and MicroVAX II processors .
Before joining Digital in 198 4, he was a consultant for Tenex and
Rampower. Previously, he worked as a project manager at Hewlett
Packard LaboratOries in Palo Alto, Cal ifornia, and as a design engineer at
Texas Instru ments Central Research Labs . Bill received his S.B. and S . M . degrees i n electrical engineering and computer science from M . I .T. i n
1978
Steven
E.
Boone Steve Boone graduated from Michigan State Univer sity (B.S. E . E . , 1974) and the University of Mich igan (M.S. E .C . E ., 197 5).
He has also done advanced graduate work at Southern Methodist Un iver sity. Before joining Digital in 1984, Steve worked as a principal hard ware engineer for Sequoia Systems, and as a senior design engineer at
Prime and Raytheon. For two years, he was an engineering supervisor working on the TK 50 controller design . Steve is currently the technical
�ngineering manager for TK Cartridge Tape Su bsystem Engineering.
Daniel W. Dobberpuhl Dan Dobberpu hl is a senior consulting engi neer and manager of the Processor Advanced Deve lopment Group. On the MicroVAX II project, he led the implementation of the 78032 CPU chip. Previously, he consulted on CMOS, ZMOS, and TIPI technology development, and worked on the Til and F l l projects . Dan joined
Digital in 1976 from General Electric Company. He received a B .S . E . E . degree from the University o f I l linois in 1967. A member o f IEEE, he holds fou r patents and is the coauthor of The Design and Analysis of
VLS! Circuits.
Amnon Fisher
Ed ucated at Israel Institute of Techno logy (B.S. E . E . ,
1973) and City College o f New York (M.S.E . E . , 1975), Amnon Fisher worked as both a contribu tor and project leader on the
3 20 16
CPU at
National Semiconductor. join ing Digital in 1983, he was a project leader of the Vl l jSCORPIO floating point chip (VAX 8 2 0 0 system) , and a contrib u tOr tO the MicroVAX I I 78132 chip. Amnon is currently an engineering manager in the SemiconductOr Engineering Grou p , working on t h e design and deve lopment of a fou r-chip s e t VAX implementation .
Peter
C.
George Earning h is bachelors and masters degrees in com puter science and engineering from M.I .T. in 1980, Peter George joined the VMS Development Group in that year. He first worked on VMS user interfaces, then on the workstation software as a principal engineer on the VAXstation project. Peter is currently a project leader, working on advanced workstation software projects . Peter is a member of ACM, and the nationa l honor societies Tau Beta Pi, and Eta Kappa N u .
Anthony F. Hutchings Tony Hutchings received his B.S. degree from the University of Newcastle On Tyne in 19 6 5 . At ICL in the U.K. for 16 years, he designed operating systems and was one of the VME· system architects on the 2900 series . He later became corporate man· ager of CAD . Tony joined Digital in 1 982 as the project manager for the proprietary DECSIM software and then became manager of the VLSI CAD
Group. Tony, a member of IEEE and the British Computer Society, is currently chairman of the CAD section of the ICCD.
Lawrence J. Kenah Larry Kenah , a consulting software engineer in the VMS Development Group, wrote the decimaljstring emu lator for the MicroVA.X project. Since joining engineering i n 1 980, Larry has worked on the VMS nucleus in the areas of memory management, pro· cess scheduling, and image activation. He came to Digital in 1 9 7 5 as an instructor and course deve loper in Educational Services. Larry received his B.S. degree ( 1 968) from Boston College and his M . S. ( 1 970) and
Ph . D . ( 1 977) degrees in high-energy physics from Northwestern Uni· versity. He is coauthor of VAXjVMS Internals and Data Structures.
Raymond J. Lanza
Ray Lanza is currently the project leader for the
ULTRIX- 3 2 system. After joining Digital in 1 98 3 , he ported the ULTRIX system to t he MicroVA.X I processor. As project leader, he ported the system to the Micro VAX I I processor in 1 9 8 4 . Ray received h is
B . S .E.E.jC.E. degree from the University of New Hampshire in 1 980, then became the lead engineer in a UNIX group a t AT&T. Later he was a senior software engineer a t Wa ng Laboratories, Inc . , researc hing wi ndowing systems and UNIX distribu ted systems.
Burton M. Leary
In 1 980, M ike Leary joined Digital after receiving his B.S. degree i n electrical engineering from the University of Massa· chusetts . In semiconductor engineering, he worked on chip designs and helped to develop the floating point chip for the MicroVA.X II system .
Mike did behavioral model ing, wrote microcode , and designed the main sequencer for that chip. He is now a senior engineer in the
Advanced Development Memory Group, designing the internal cache for an advanced chip project .
Barry A. Maskas
Barry Maskas is a principal engineer cu rrently speci fying and designing an integrated circuit, and fiber-optic boards for future systems . As a senior engineer on the MicroVA.X I I project , he was co-designer of the CPU board and the memory boards. Barry came to
Digital in 1 9 79 after receiving his B. S.E.E. degree from Pennsylvania
State University. He also holds a n associate's degree from the Commu· nity College of Allegheny Cou nty and did u ndergraduate work at LSU .
Barry is a member of Eta Kappa Nu ; he has a patent pending for a seJf. configurable memory su bsystem .
5
------------�----
Biographies
Kathleen D. Morse As a consulting software engineer, Kathy Morse is responsible for VMS support on all low-end CPUs and peripherals. Ear lier, she did the VMS support for both MicroVAX systems , the VAX
1 1 /782 system , and the MA780 multiport memory. Kathy joined Digital in 1 9 76 after receiving her B . S . C . S . degree from Worcester Polytechnic
Institute, where she also earned her M . S . C .S . degree in 1 98 5 . Kathy is a member of IEEE, the Professional Council, and ACM, as wel l as Tau Beta
Pi and Upsilon Phi Epsilon . She has published in the Computer Mea surement Group ' s 1 985 Conference Proceedings , and Datamation.
Guenter E. Schneider
Guenter Schneider joined the Mass Storage
Group in 1 97 0 , when it had only about 2 5 people . He has worked on the designs for the RX0 5 , RL0 1 , RX0 2 , TU5 8 , RX5 0 , and RD50j5 1 storage devices .
As a consulting engineer, he helped to design the TK5 0 cartridge tape drive . Guenter received a Diplom lngenieur from the
Technische Hochschu le Aachen in West Germany and his M . S . M . E . degree from M . I .T. in 1 969 . H e holds two patents , with a third pending, and is a member of the e ngineering society Verein D e u tscher
I ngenieure .
Stephen F. Shirron
Educated at Catholic University of America (13 . S .,
1 980 and M . S . , 1 98 1 ) , Stephen Shirron came to Digita l after graduating
Summa Cum Laude .
As a senior software engineer, he developed an interpreter for VAXjSmalltalk-80 and designed the V�Xstation 1 00 firmware . Currently a principal software engineer, Stephen designed and implemented the firmware for the RQDX3 disk controller. He is a member of Phi Beta Kappa and has written a chapter in Sma/ltalk-80:
Bits of History, Words of A dvice.
Robert J. Simcoe
13ob Simcoe is a technical manager currently work· ing on serial interconnect products . He was the technical manager for the floating point chips in both the MicroVAX I I and VAX 8 2 0 0 systems.
Before joining Digital in 1 98 2 , Bob worked for the Department of
Defense and General Electric Company. His du ties i nvolved MOS design , process deve lopment, and product design using custom res.
Bob holds seven patents on IC circuitry and systems. He graduated from the University of llli nois (B.S . E . E . , 1 966) .
Rick Spitz Rick Spitz manages VAXjVMS software deve lopment for
CPUs and peripherals. As a consu lting software engineer, he was a primary member of the architectural design team on the MicroVAX workstation project. Rick designed the VMS graphics hardware interface architecture and, for six years , has specialized in VAXjVMS hardware software interfaces . He joined Digital in 1 977 as a senior software specialist and received Digital's Software Excellence Award. Previously,
Rick developed m icroprocessor software for Inco , Inc. He earned a
B . S . E . E . degree from Clemson University i n 1 974 and his M . S . C . E . degree from the U niversity o f Lowell in 19 83 .
6
Robert M. Supnik
Bob Supnik is a corporate consultant and group manager in semiconductor engineering . On the MicroVAX CPU chip project, he was project leader and lead m icroprogrammer. Bob was the project manager for the J 11, a contributor to the F11, and supervised advanced development on the HSC50 and U DA5 0 . Before joining
Digital in
1977, he worked at Applied Data Research . Bob received his
S . B . degrees (1967) in math and history from M . I .T. and his M .A. degree
(1972) in history from Brandeis U niversity. He received Science
Digest's"
1 00 Top I nnovators of
1985" award.
)
Nicholas A. Warchol
In
1977,
Nick Warchol joined Digital after receiving his B . S . E . E . degree (cum laude) from the New Jersey I nstitute of Technology. Later he earned his M . S . E . E . degree from Worcester
Polytechnic I nstitute in 1984. He is a member of Tau Beta Pi and Eta
Kappa Nu . Nick has worked on the advanced development of charged couple device memories , bubble memories, and laser video disks . I n his present position a s a principal engineer, h e worked on the design of the RQDX3 disk controller.
William R. Wheeler
After earning his B .S . E . E . degree in
1982 and his
M . S . E . E . degree i n 1983 from Cornell University, Bill Wheeler came to
Digital as a junior engineer. On the Micro VAX I I project, he designed the exponent datapath and control for the
78132 floating point chip.
Later he designed the exponent section of the floating point chip in the
VAX 8 2 0 0 system . Bill is currently working on the instruction box and bus interface unit for a new m icroprocessor chip.
RichardT. Witek Rich Witek is a consulti ng engineer working on the architecture and implementation of new m icroprocessors . He helped to develop and debug the MicroVAX 78032 CPU chi p . Rich a lso worked on implementing DECnetjE and on the DECnet Architecture Review
Group during Phases 2 and 3. He also worked in the VLSI CAD group .
Before joining Digital i n
1977,
Rich was a senior technical associate at
AT&T Bell Laboratories and an engineering assistant at Argonne National
Laboratory . He received his B .A . degree in computer science from
Aurora Coll ege , and is a member of ACM and I E E E .
Stephen
H.
Zalewski Steve Zalewski i s a senior software engineer working on the graphics execution routines for the VAXstation IIjGPX system . He joined Digital i n 1981 after receiving his B . S . degree in computer engineering from \Vorcester Polytechnic I nstitute . Steve developed the graphics device driver for the VAXstation I and II sys tems. His earlier work involved writing RMS file-sharing internals and impl ementing RMS fi le sharing and global buffers for VAXcluster software .
7
----------
Foreword
Jeffrey
C. Kalb
Vice President and Group Manager
Large Scale Integration
The roots of the MicroV AX program go back to the summer of 1 98 1 . To understand why this program was initiated and the thinking behind it, one has to look at the events of that time.
Many developments were taking place, sug gesting that a whole new class of systems capa bilities could emerge before long .
The VAX-I 1 /780 system was in its heyday. It was recognized as the standard against which a l l other c o m p u t e rs w e re c o mpared a n d benchmarked . And true t o fashion , everyone seemed to find some way to benchmark his machine in some particu lar niche against the l l j780's capabilities . That was particu larly t r u e of t h e u p c o m i n g g e n e r a t i o n o f microprocessors and microprocessor-based sys tems. The universities were busily benchmark ing Intel Corporation's latest generations of
8 0 86s, 80 1 86s, and the early 8 0 2 86s on spe cific jobs. The same was true of the 68000based system . Many companies were starting tO come to market with engineering workstations a n d s i m i l a r p ro d u c t s b a s e d o n t h e s e m icroprocessor chips . I n fact i f one believed the trade press , the VAX - l lj780 system had actu a l ly been eclipsed i n performance and capabil ities by these "upstarts . "
Needless to say, these events caused some degree of consternation and soul -searc h i ng within Digital Equipment Corporation. More over, another factor was becom ing painfully obvious: the emergence of the independent software vendors . Hoards of small companies were springing up everywhe re to generate software for various personal computers that e ither had already been i ntroduced to the mar ketplace, like the Apple I I , or shortly wou ld be , like the IBM PC . These small vendors wanted to write software for the systems that had the high est market volum e . Their reasoning was clear.
To sell as many of their software packages as possible required i mplementing their ideas on the highest volume hardware . It was also clear that the highest volume hardware was going to b e m i c r o p r o c e s s o r b a s e d a n d q u i t e inexpensive .
8
Meanwhile , within Digital, the Semiconduc tor Engineering Group (SEG) was busy devel oping a multichi p implementation of the VAX archi tecture . Built with a midrange , multiuser, high-performance system in m ind, this chip set and its attendant system implementations were aimed at the marketplace for systems above
SSO thousand. CAD tools were being developed and ma nufactu r i n g p rocesses de velo p e d a n d refined . The module and system concepts were then i n the definition stage .
Discussi ons began a t this t i me , centered around what was later known as the MicroVAX system . There was a perceived need to counter the rising tide of encroachment on our systems business by microprocessors. We wanted tO cre ate systems with volumes high enough to war rant the attention of the i ndependent software vendors. I n general, we wanted to establish the
VAX architecture as one of the preferred archi tectures at all potential price levels in the entire i ndustry.
These discussions and strategic thinking con verged after receiving an unsolicited proposal from a semiconductor manufacturer. This firm had approached us during that summer, want i ng tO implement the VAX archi tecture in one or two high-performance chips. This set of chips could be used in our systems and sold as standalone products. The firm wanted tO use the VAXjVMS architecture (and primarily the software associated with it) to get a jump in the marketplace by establi s h i n g a h i g h-volume architectural standard at the 3 2-bit level. We were concerned from the beginning that the capabi lities and resources of this smaller firm would not be sufficient tO execute such a for midable program. But the notion t hat building a single-chip VAX implementation and using it to counter-attack the emerging m icroprocessor based systems had struck a responsive chord .
Until that time, our thinking had been in terms of our traditional pricejperformance learning curves . Our strategies did not i nclude extraor dinarily low-priced VAX systems.
As indicated above , the Semiconductor Engi neeri ng Group i n Hudson, Massachusetts, was already heavily com m itted to the multichip
VAX syste m . A number of other major chip projects were i n development as wel l . There fore , we searched for a larger semiconductOr vendor who could bring additional design and manufacturing resources to bear on this con cept . Such a vendor could also make available additional distribution channels for sales of high-volume chips tO the general marketplace .
This line of thinking was pursued with various vendors throughout the fall and winter of 1981, until April 1982
Interestingly enough , there was less than wholehearted enthusiasm on the part of the various vendors who were approached. Each of them had already decided on an approach tO the problem and were unwilling to make the development of the MicroVAX chip a priority i t e m. That com m itment was an extremely i mportant issue tO us . Experience had shown that complex projects of this nature always exceeded the schedules and the budgets antici pated when they received second-class atten tion within the merchant semiconductOr indus try . Thus one criteria for working with a vendor was that he commit to the M icroVAX architec ture as a pri mary market thrust. No one was willing tO do that .
At the same time, other issues had to be worked. I t was clear that the full VAX architec ture as implemented in the multichip set could not easily be put on a single chip. That would have taken over 1 million transistOrs, a capabil ity t hat would not be avai lable until the end of the decade . Therefore , early in the project, we recognized that there was a need to subset the architecture to make it implementable on a sin gle chip. By December 1981, the idea of devel oping a single-chip VAX implementation was begi nning to get some positive re-enforcement within Digital. As a result , in that month ,
Gordon Bell, then vice-president of Engineer ing, chartered a subcommittee tO investigate what s hould be i ncluded in a M icroVAX architecture .
The key people involved were Roy Moffa , who had been leading the strategic th inking about a single-chip VAX system; Bob Supnik, representing semiconductor technology; Dick
9
Fo reword
H u s tvedt a n d Dave C u t l e r , r e p r e s e n t i ng software technology; and Bill Strecker, repre senting VAX architectu re technology. After a few intensive meetings, they proposed a subset of the VAX architecture in January 198 2 . Bob
Supnik and the sem iconductor technologists thought that this subset could be implemented i n a single chip. This new archi tecture would be modified slightly later in the year, but it is essentially the architecture that exists today.
The only sign ificant modification was in the memory management capabi l i ty, and in some sense , this change actually simplified the devel opment o f the chip.
I n para l le l with these other activities, Bob
Supnik and other members of SEG had been studying ways to get the chip developed inter na l ly . They were h o p i ng to l ev e rage t h e existing i nvestments in process technology, chip mode ling, CAD tools, and the various other e lements that were necessary. Further more, and highly signi ficant to the whole pro gram, they developed ways of re-using some of the i nvestments be ing made in the multichip
VAX i m p l e me n tation and o t h e r p rograms already in progress .
As a res u l t the fl oating point c h i p being dev e loped for a PDP
-
1 1 microprocessor was used as the building block for the MicroVA.X implementat ion. Not only that but the chip was a lso retrofitted back into the existing m u ltichip set to minimize the workload . Moreover, the datapath was lifted from the instructionjexecution u n i t of the mul tichip s e t to fo rm t h e backbone o f the
MicroVA.X CPU. Tools and techniques were bor rowed whenever it was possible .
I n this sense the Mi croVA.X program was u n iq u e . There were al most n i ne months of strategy discussion and evaluations of various ways of implementing and executing before any real design actua l ly started. While many of the p roposed b u s i n e ss strategies were never adopted, they a t least received a hearing. In any case the die was cast.
The real i mplementation of the MicroVA.X chip did not get started until June
198 2, the official start date being July 6, 1982 . (Some work had been done prior to that for recru iting and staffing.) It was soon evident that there were some key e l e m e n ts t h a t had to be addressed. The first was CAD tools. There was no question that this device had to be simulated extensively at a l l leve ls of implementa tion .
There was no other way ro get the quality of design and performance l evels being planned .
At the time the program started, these tools were mostly experi menta l . Some techn iques had been tested, but the real i ty was that CAD rools "broke" on numerous occasions during the deve lopment of the system . Crisis-ori ented
SWAT teams had to be put in place to bridge over or break through barriers that threatened to bring the entire program to a halt.
There was another equa l ly i mportant ele ment. The entire program was an extremely complicated one, with many elements on para l lel paths . Process technology had to be devel oped,
CAD tools deve loped and refined, chip designs done, systems imple mentations exe cuted, and test tech niques and equ ip ment deve loped. Each of t hose elements was inti mately entwined with the others . Therefore the possibili ty clearly existed that, upon reaching the end of the design, we wou ld be faced with debugging a new process technology, a new manufactu ring line, new testers, a new chip design, new packages, and a new system , a l l s i m u l taneously. A real possibil ity existed that we cou ldn 't separate the variables in a suffi ciently clear and timely manner to a llow the chip debugging and system eva luation to take place. This phase could last for months or per haps even years , something that has happened before on many such programs i n the merchant i ndustry.
To avoid that, we segmented the major risks in the program and put plans in place to mini mize as many of those as possi ble in paralle l before the new chip arrived. F o r instance , rather than debugging an entirely new manufac tu ring line while trying to build this new chip, we combined the existing two wafer fabrication l ines in to one . The sma l ler li ne was t h e n retrofitted r o provide a pi lot l ine capability.
That gave us a trai ned staff, a debugged faci l i ty, and a l l the other e lements necessary to mi ni-
10
m ize the interaction of the process and facility .
Additionally, a test vehicle was designed s o that manufacturing cou ld run wafers, debug process steps, and i mprove the basic yields of the pro cess well before the new chip arrived. In the test area, test programs were implemented on older, proven testers on which the engineers had experience . That worked even though we knew that, for the eventual production, an entire ly new generation of testers wou ld be necessary tO precisely test such a complicated device at its full speed .
Simi larly, other areas , such as packaging,
CAD cool development, and parts of the system evaluation , were exam ined and improved i n parallel long before they had tO work together.
A major program was put in place to u ncouple risks and tO hire and train the workforce wel l in advance of the completion of the MicroVAX chip design . This effort was quite expensive; some people thought that much of the money was being thrown out with the materials that were made experimentally. But the end result was one of the smoothest debugs and introduc tions i ntO chip manufacturing that I have ever witnessed for a complex device . While there were problems and a lthough thi ngs didn ' t always work right, there were almost always independent ways of separating the variables in the proble m . In that way it cou ld be properly analyzed and corrections put in place . This example should serve us wel l with comp lex development programs in the future .
One other thing done tO enhance the debug and ensure the qual ity at the system level was to co-locate the CPU modu le designers with the chip designers . I n that way their interaction was enhanced and the rate of problem resolu tion greatly accelerated. The module team itself was exceptionally sma ll for such a major pro gra m , consisting of only three primary engi neering people. But this unique program envi ronment featured a high degree of simu lation , close proximity of the engineers (the Micro VAX chip team had only 20 people) , and heavy rel i ance on thorough evaluation at every step.
The end result was very, very few bugs in either the chip or the system. In fact there were fewer than 20 bugs that had tO be corrected before the integrated chip and system were able tO boot the operating system . It should be noted that this quality has continued tO manifest itself in the rapid manufacturing ramp-up and the qual ity of the systems that have been generated .
There were more engineering changes to the parts and the system tO enhance our margin and ease of manufacture than there were tO make the system functional in the first p lace . That is evidence of a fundamentally different approach to building systems.
As noted above , the MicroVAX program is quite u nique, from its initial conception to the continuing efforts to enhance quality and pro ductivity. From the i nitial conception of the strategy, through the organization of the people and problems, to the ongoing e ngineering activity around quality and ease of manufac ture , this program has provided a new paradigm for program execution and management. Our hope is that, with this knowledge, people can emulate the success of this program while elim i nating the errors . I n so doing, Digita l can greatly enhance its abil ity to build and manu factu re high-quality systems in i ncreasingly shorter periods of time.
1 1
Daniel
W.
Dobberpuhl
Robert
M.
Supnik
Richard
T.
Witek
I
The Micro VAX 78032 Chip,
A 32-Bit Microprocessor
The Micro
VAX
78032 implements the
VAX
architecture on one chip. To do that, the instruction set was repartitioned to reduce the number of tran sistors. The instructions used most frequently are in microcode; others, notably floating point, are emulated in macrocode. Hardware was sim plified by having a small address translation cache and no memory cache; however, full
VAX
memory manag
emen
t is supported. Afast 200nanosecond microcycle allows instructions to execute in parallel. The
CPU
chip
is made
using a 3-micron, double-metal NMOS process. The control store ROM has X-shaped cells, which help to reduce its size.
The MicroVAX 780 3 2 chip is the latest exten sion of the VAX architecture and the first in the form of a single-chip microprocessor.
As the
CP of the MicroVAX II computer system , the
7 8 0 3 2 p e rfo r m s n e a r l y as fa s t as t h e
VAX- 1 1 / 7 8 0 superminicomputer, b u t i n a microcomputer package .
Origins and Goals
Digital began the MicroVAX CPU chip project in late 1981 in anticipation of increasing com p e t it i ve pressu res from i n d u s t ry-standa rd microprocessors . The original intent of the pro gram was to license a semiconductor vendor to design and manufacture a MicroVAX single-chip microprocessor. However, the l eading semicon ductor companies were unable tO meet the h i g h - pe rformance req u i r e m ents a n d t i g h t schedules that the project requi red . I n May
1 98 2 , an i nternal development project was chartered tO design the MicroVAX CPU chip .
From a designer's viewpoint, the develop ment of this CPU was a challenging exercise i n shrinking the VAX computer architecture with out changing i ts functio n . There were five major goals that governed the design.
1 . The kernel architecture was tO be imple mented on a single chip . Other chips or hardware could be used to i mprove per formance or to provide additional functional ity, but the basic VAX functions had to be i ncorporated in the base CPU design .
2 . The chip had to be compatible with all
VAX application programs . I t had t o exe cute any application program, whatever its size or complexity, written for any computer in the VAX fam ily. And it had tO execute without alterations to the pro gram code . That meant that the chip had to run the M i croVMS and ULTRIX- 3 2 m
( D igita l ' s e n h a n c e d U N I X software) operating systems, and the VAXE LN real time kernel .
3 . The chip had tO perform at or near the speed of the VAX - 1 1 /780 processor . This goal implied that the chip had to have a highly parallel internal implementation, a high-performance external interface , and a fast microcycle . Accordingly, the internal microcycle of the chip was set at the same 200 nanoseconds (ns) as the l l j780's microcycle .
4 . The price of the chip had to be competi t i v e w i t h c o m m e r c i a l 3 2 -b i t m icroprocessors of comparable com plexity. This required a relatively con servative d i e size and an i nexpensive package . I t also required the i mplemen-
1 2
Digital Technical journal
No. 2 March
1986
tation of an external interface that was compatible with standard VLSI periph eral chips and demanded minimal sup port from the hardware on the C PU board .
5 . The chip had to be designed and buil t quickly. T o meet or beat competitive products, the chip had to be in produc tion less than 2
Yz years after the start of development.
With these goals gu iding the chip design team, the major problem was quickly identi fied: to reduce the number of transistors. That, in t u rn, re q u ired reparti t i o n i n g the VAX instruction set and simplifying hardware func tions wherever possible .
Reducing the Number of Transistors
The pri ncipal problem in designing the 7 8 0 3 2 was how t o implement the complexity o f the
VAX architecture on a single chip . There are
3 0 4 instructions in the ful l instru ction set, w i th
1 4 data types and 2 1 addressing modes. I nstruc tions vary in length from 1 byte to 54 bytes.
1
Dema n d - paged virtu a l memory s u pport is requ ired to guarantee compatibility with the operating system software . To accommodate this complexity in a ful l-scale VLSI VAX imple mentation requ ires about 1 . 2 5 million transis tor sites2 However, the semiconductor tech nologies available at the time of design could support only about one-tenth that number in a single-chip microprocessor.3
The architectural functions i n all VAX sys t e m s a r e p a r t i t i o n e d a m o n g h a r d w a r e , microcode, and the operating system . Al l previ ous VAX implementations have similar bounda ries between these three. The hardware pro vides the registers and memory, the microcode provides the instruction set, and the operating system provides the program services . A large contro l store-a minimum of 4 00 kilobits (Kb) is requ i re d to c o n t a i n t h e i ns t r u c t i o n microcode . The console function is handled in either microcode or a support processor. More over, the control logic needed to support mem ory manage ment and the variable instruc tion format is quite complex4
Two d i fferent approaches were taken to reduce the transistor count in the microproces sor chip. First, the VAX instruction set was repartitioned to cut the size of the control store to 62Kb . Second , the amount of on-chip hard ware was reduced by simplifying some func tions, placing others elsewhere, or omi tting some altogether.
Repartitio n ing the Instructio n Set
As the first repartitioning step, the design team assumed that all VAX instructions had to be implemented in order to execute all VAX appli cation software . However, there are several classes of i nstructions that involve a good deal of m ic rocode and yet are infrequently exe cuted. For example, a typical timesharing work load is handled by base instructions, scientifi cally oriented i nstructions , and commercially oriented i nstructions . Analyses of more than 70 mill ion executed i nstructions showed that the commercially oriented ones represented less than 0 . 2 percent of the total executed5·6 Stud ies of scientific and engineering workloads showed even lower percentages. Even in com mercial applications, the commercially ori ented instructions represented less than 4 per cent of the total executed , the majority being base instru ctions. Therefore , emu lating the commercially oriented instructions in the oper ating syste m rather than u s i ng m i crocode wou ld significantly reduce the size of the con trol store, but would have l ittle effect on over a l l performance because these i nstructions were seldom executed .
On the other hand, floating point instructions require a good deal of microcode and are exe cuted more frequently. Even with m icrocode, i nstruction execution is relatively slow unless a separate floating point accelerator (FPA) is used . Therefore, although existing VAX imple mentations offered both m icrocoded (warm) a nd hardware (hot) floating point , the design team decided not to implement these instruc tions in microcode . I nstead, floating point instructions wou ld be executed in an optional floating point chip, or by e m u lation using macrocode .
6
I n tota l , 1 75 of the 3 0 4 VAX instructions and of the 14 data types are implemented in on chi p microcode . Those include integer and log ical instructions, variable-bit fie l d , contro l , queue, procedure calls, character string moves, and operating system support . This microcoded subset comprises over 9 8 percent of the instructions that are used to execute a typical program . H owever, the requ i red microcode
Digital Technical journal
No. 2
Mm·cb 1986
1 3
New Products
.
-------
The Micro VAX 78032 Chip, A 32- Bit Microprocessor occupies only one-fifth the control store space of a fu l l VAX impl ementation. Seventy t1oating point instructions and three data types (F. D, and G t1oating) are implemented in the t1oating poinr chip, when it is present. If that chip is absent, the instru ctions are emu l ated i n macrocode . T h e remaining
59 instructions and
5 data types are a lways emu lated in macrocode.
Those are main ly decima l string, character string, and H floating point operations . The
CPU chip provides some microcode support for the emu lated instructions . Table
1 summarizes the instruction set arch itecture of the 78032 chip.
The decis ion to emu late instructions in macrocode has an effect on speed because emu lated instructions take three to ten times longer to execute than microcoded instructions . How ever, the instructions in this group of
59 are
Table 1 Instruction Set Architecture
Implemented i n
CPU Chip
Im plemented i n
Floating Point Chip normally used so infreq uently that the execu tion speed of a typical program is reduced by no more than four percent. Tab le
2 i l lustrates the division of instructions between the CPU chip , the FPU chip, and the macrocode . All in all, the fivefold reduction in the size of the control srore ha lved what wou ld have been the active area of the chip .
Simplifying the Hardware Fu nctio ns
The principal hardware simp lifications in the
78032 are the reduced size of the address trans lation cache (translation buffer) , and the e limi nation of a memory cache in favor of tightly coupled local memory.
As mentioned earlier, demand-paged virtual memory management was requ i red for compati bi lity with the
VAX architecture . Consequently, the design team decided that the 78032 woul d
Implemented in
Macrocode
In structions:
I nteger and
Logical
Address
Va riable Bit
Field
Control
Procedure Call
M iscellaneous
Queue
Operati ng System
Support
Character Move
Total
Data Types:
Byte I nteger
Word Integer
Longword I nteger
Quadword Integer
Variable Bit Fie ld
Variable Character
String
89
8
7
39
3
10
6
11
2
175
F floati ng
D floating
G floating
F floating
D floating
G floati ng
24
23
23
70
H floating
Octaword
Character
String
Decimal String
Edit
CRC
H floating
Octaword
Leading Separate
Numeric String
Trai ling N u meric String
Packed Decimal
59
28
4
9
16
14
Digital Technical journal
No. 2 March /')86
Table
2 Division of I nstructions
Instructions
Implemented in
CPU Chip
Percent by I n struction
Count
Percent by Microword
Count
Percent by Typical
Execution Frequency
57.6%
20.0%
98. 1 %
Instructions
Implemented in
Floating Point Chip
23.0%
20.0%
1 .7%
Instructions
Implemented in
Macrocode
1 9.4%
60.0%
0.2% be the first single-chip CPU with fu l l demand paged virtual me mory support right on the chip. At first the design team proposed to use a simp lified version o f VAX memory manage ment. During the course of the design , how ever, the software engineers reported that not providing fu l l memory management was quite expensive in terms of the use of p hysical mem ory. Therefore, the design team i mpleme nted fu ll VAX double-mapped compatibil i ty in the chip.
As the design progressed, it became evi dent that the incrementa l cost of providing this capabi l i ty was much lower than orig i n a l l y anticipated .
Al l existing VAX processors implement mem ory management with a large address transla tion cache (at least 1 2 8 entries) , with system and process addresses in separate halves. A translation cache must have a high hit rate to be effective . Si nce most caches are direct mapped, many entries are requ ired to achieve a high cache rate 7·H Implementing a comparable num ber of translation cache entries in the 780 3 2 was out o f the question, due to die size con straints. However, the VLSI technology in the
780 3 2 is very amenable to using a fu l ly associa tive translation cache with least-recently-used
(LRU) replacement.
Such a cache needs many fewer entries to achieve the same hit rate as the d irect-mapped version. In add ition, the tight coupling to loca l memory, as explained i n the next paragraph, made i t poss ible to reduce drastica l l y the amount of time requ ired to process a transla tion cache miss . Thus the translation cache in the chip has only eight entries , but the cache is fu lly associative , uses true LRU replacement, a nd is su p p o r t e d b y h i g h l y o p t i m i z e d microcode for fast processing o f m isses. Moreover , simu lation studies showed that the best use of the eight entries was with a homogene ous structure . Therefore, the system and pro cess addresses are cached together .
The team also decided to forgo the use of an external memory cache, which requ ired a com plex external interface . Use of an internal mem ory cache had already been ruled out due to die size constraints. Accordi ngly, the speed of memory access is 400 ns, or two m icrocycles, which is the speed of local memory. Thus the chip encounters no wait states, and its average time to access memory is approximately the same as the 1 1 j780's. In a typical program , there is little difference between the i nteger instruction performance of the two CPUs .
Additional s i mplifications i ncluded the elim i nation of warm (microcoded) floating point in favor of a floating point accelerator , e l i m i na tion of wri table control store capabi l ity, and e l i m i nation of on-chip console su pport.
Design Narrative
The starting point for the chip design was the i nstruction execution chip of a multichip VLSI
VAX processor already in design . This chip would provide a general floorplan and a base m i croarchi tecture , and might even provide complete design sections that cou ld be used for t h e M i c r o VAX 7 8 0 3 2 . As t h e p r o j e c t progressed, the designs o f the VLSI VAX proces sor and the MicroVAX 780 3 2 tended to d iverge under the pressure of differing constraints : chip set and system functionality for the former ; die size, power, and time to market for the latter .
Ultimately, only part of the main datapath was s hared betw e e n the two ; the rest o f t he
M icroVAX 7 80 3 2 design and i ts m icrocode were unique.
Digital Tecbni<:al jourual
No. 2
/lfarch 1 ')86 l 5
New Products
-------
The Micro
VAX
7803 2 Chip, A 3 2-Bit Microprocessor
The MicroVAX 780 3 2 project took 20 momhs from start to fi rst-pass mask generat i o n : 6 months for specification and general design, and 1 4 months for physical implementation .
Eighteen people worked on the design ream .
Project Design Tools
The design ream was aided by a h ierarchica l
CAD rool suite that ran on a VAX system. The use of these tools was one of the primary rea sons that the project was completed on sched u le . The principal components of t h is rool su ite are as fol lows :
1 . A proprietary chi p-database manager and tool i nterface called the CHAS system
2 . A schematic capture progra m , QUICK
D RAW, t hat uses simple term inals
3 . A proprietary h i e rarch ical s i m u lator ca l l e d t h e DECS I M system , used for behavioral simulation
4 . A switch- leve l MOS logic s i m u lator,
RSI M , used for u n i t-delay l ogic s i m u la tion
5 . A modified version of the standard SPICE circu i t simu lator that incorporates new ana lytica l, rather than empi rica l, MOS transistOr mode Is
6 . Design-rule checking programs, DRC and
DRACULA I I
7 . A n i n te rcon nect verifica t i o n program cal led the IV system , which performs both layout extraction and wiring verifi cation9
8 . A cross-reference progra m , XREF, that ana lyzes c o u p l i n g , boots trap ra t i os , dynamic node stabi l i ty, and other c i rcuit problems
The chip layout was done on Calma GDS n systems. Three dedicated VAX- 1 1 /780 systems and five Ca lma stations were used throughout the project. The back-end verification of cir· cu i ts and the layout required as many as eight
VAX systems .
Final Chip Design
The fina l product of this design process is a microprocessor that conta i ns 1 2 5 , 000 transis tOr s i tes in a 3 - micron , double-metal NMOS chip that measures 8 .7 by 8 . 6 m m . It requires only 5 Vdc and a maxi mum o f 3 watts of power; it is pac kaged in a 68-p i n , su rface-mounted leaded chip carrier. The chip operates at 2 0
MHz and has fu ll 3 2 -bit i n terna l and external datapaths. The 780 3 2 is mounted on a s i ngle board , quad-si zed ( 8 . 5 by
I
0 . 5 i n . ) CPU mod u le having a Q 2 2 1/0 bus and 1 megabyte (MB) o f loca l m e m o r y . An o p t i o n a l FPA, t h e
M icroVAX 78 1 3 2 chip, can also be mou nred o n the CPU board .
The measu red speeds of integer and floating point operations of the 780 3 2 represent a breakthrough in 3 2-bit m icroprocessors . System evaluations of M icroVAX 780 3 2 modu les ind i cate that t h e i r performance in processi ng i nte gers is approxi mately equal to that of the
VAX- 1 1 /780 system . With the floating point chip, the performance is between those of the
VAX- 1 1 /7 5 0 and VAX- 1 1 /780 systems w i t h
FPAs .
The remainder of this paper expla i ns the fu nctional organization of the chip and i ts phys ical i mplementation in silicon .
Functional Organization
The diagram i n Figure
1 and the photomicro graph in Figure 2 out l i ne the various subsec tions, or fu nctional boxes, of the M icroV AX
780 3 2 chip. They are organi zed i n to three sec tions . At the left of Figure 2 are the darapaths for decod ing and executing i nstructions and for memory management. At the center is the con trol logic for i nternal operations and the proto col signal logic for external operations. At the right is the sequencing l ogic for both internal and external operations.
The left sect i o n i n t h e p hotom icrograph
( Figure 2 ) , comprising the data paths, consists of the I Box, the E Box, and the M Box.
•
The l
Box prefetches and decodes instruc tions. I ts main function is to parse the cur re n t macro i nstruct i o n in the i nstructi o n stream and work in conjunction with the mi crosequencer to generate t h e m icroad dress for the next mi cro instruction . This microaddress is a function of the current macroinstru ction . A prefetcher, which works i n para l l e l w i t h other c h i p opera t i o n s , accesses a n d stores instruction data i n a n eight-byte prefetch q u e u e . The prefetcher acts autonomously by atte mpting to keep that queue fu l l at all ti mes, using any free
I/O-bus cyc les to access the i nstru ction
1 6
Digital Technical journal
No.
2
March
1 986
I BOX
E B OX
EXTERNAL
DATA AND
ADDRESS BUS
INPUT
DRIVERS
M
:
•r r+
''N rw-::R:::,T::-E---, I
AND
� g��e��
I
I �I
1 r
I
I
I
I
I
I
- - - - - - - - � -
-rl
H
--
,-----
�
L...1
I
I
INTE RNAL
DATA AND
ADDRESS
B�
ALU ,;;;ITH
B
SHIFTER
I
•
LENGTH COMPARATOR
J
..___.j
I
1
L �
I
I
I
I
-
B BUS
I I
---1
TB/AW BUS
....--J
L.. _ _
I
\
BARREL
\
REAO OATA LATCH
ALIGNMENT Mux
PREFETCH STACK
INSTRUCTION
STREAM ROTATOR
INSTRUCTION
DATA REGISTER
•
n�·
GPRS AND
SCRATCH PADS t
WITH
LATCH AND
I
BIT SHIFT
LENGTH REGISTERS
ADDRESS REGISTERS
I
� l
VIRTUAL
ADDRESS BUS
TRANSLATION
BUFFER, TAGS
PAGE TABLE
ENTRIES
_
•
t
MUX
X/
J
PHYS!CAL
ADDRESS BUS
R I-
-.-
J
I
1
I
I
I
.---1 n
_;-
I
I
I
I
!BOX
Aw
Bus
I oecooe
I
INTERRUPT
N
;,;
R�-
PRIORITY
LEVEL
Lr
INTERRUPT
CONTROL
INSTRUCTION
LOGIC
CONTROL
.
�
S Y l AN
� - - - - �- - -
NCHRONIZER
D
PRIOA1T1ZE R
E
(ENTAY·POINT .uADDAESSES)
,----'-....1...-,
�
��ODE
MICROSEOUENCER
CONTROL
- - - - - - - - l�J_,
J
I
I
I
EBOX
CONTROL
MBOX
CONT ROL
_ _ r -
J
I
CONTROL
I
NSTRUCT
r
I
I
I
I
I
I
I
I
I
REO
- -
:
U
HA
LT
ON
MtCROtNSTAUCTtON
14---
· - - - - _ ...J
1---- -
I r--- - - -,
I
I
I
I
I
I
II
I t
EXTERNAL CONTROLS
AND STROBES
INTERRuPT
I j
I'
ES�
I I
(�-tBRANCH oFFsET)
!
'--1
I
I
I
I
I
I
1
S p
M I CROSEOUENCER
,---
�
1,--L-----'.__,L...
�
MUX
':::
LATCH
I
ES:-:
L
.. pc
JtAOO RESS
STACK
I
1
_j
,
E
-r--8-l ----,
ADDER
L___..
Lr------
UX p
OOR s
51
SS BUS
CONTROL STORE
I
�-
CLOCK GEN ERATOR
TEST
,uADDRESS
Figure 1 Block Diagram of the CPU Chip strea m . Even i f t h e q u e u e is fu l l , the prefetcher will start to read data if the queue will be at least half-empty after the current microcycle.
The l
Box a lso decodes i nstructions and vari able-length operand specifiers in para l lel with other chip operati ons . That avoids requ iring explicit decode cycles to execu te successive macroinstructions . Due to the constraints on the size of the control store, most of the address-specific microcode had to be shared among a l l instructions. The i nstruction-decode PLA (lPLA) generates 1 9 bits of opcode-specific data for controlling other chip operations related to a giv e n instruction . That a l l ows many microcode sequences to be table driven and shared.
•
The E Box is the instruction execu tion unit and contains the main datapath of the chip.
This box holds
1 6 VAX-specified general pu rpose registers (GPRs) , 20 microcode reg isters, a 3 2-bit arithmetic logic unit (ALU) , and a 3 2 -b i t barrel shifter. The E Box a lso maintains condition codes for t he process
DigiUll Technical journal
No. 2 March
1986
1 7
New Products
The Micro VAX 78032 Chip, A 3 2-Bit Microp1·ocessor t.
-
:...
-
� .,.
:--
..
-
I
t
I
'" �
•
.'
"'
I
I
-
I t
.
•
I
.
•
.,
·
,_
I
..
I
�
Figure 2 Photomicrograph of the CPU Chip status longword (PSL) and determi nes VAX branch conditions at the macrocode level . In a 200-ns cycle , the E I3ox can read two regis ters , perform an ALU operation or shift, and write the result into a register. Si nce reading a n d writing to regis ters are performed sequentially, the ALU result bus is mul tiplexed with an i nput bus. thus saving vert i cal interconnect . The ALU employs a 4 -bit lookahead carry sche me, with ripple carries across the nibbles . The carry chain uses dual rail logic for maxi mum speed . The barre l shifter is a pass-transistor network, which is very compact and fast enough for this task.
•
The M I3ox serves as the memory ma nage ment unit and translates vi rtual addresses to physical addresses. The address translation cache , which is ful ly associative, stores t he most rece ntly referenced address transla tions. The M I3ox maintains three virtual address registers , one for instruction data and two for program data. This unit also detects cross-page accesses and includes a separate compararor for length checking. A dedicated adder generates the next vi rtual address for seque ntial data and instruction addresses .
The time to perform an add ress translation is less than 2 5 ns when the virtua l address is i n
1 8
Digital Technical journal
No. 2 March 1 986
the translation cache . This short translation time a l lows memory management to be trans parent to the external chip timing.
The center section of the photom icrograph is composed mostly of random control logic. That logic translates the highly vertical ( 3 9 - bit) microcode into the many discrete control sig na ls requ ired to operate the datapath .
The right section of the photomicrograph, comprising the sequencing and clock ing logic, consists of the inte rrupt logic, the control store , t h e DAL interfa c e , and t h e c l oc k generatOr.
•
The interrupt logic accepts , synchronizes, and prioritizes external interrupt requ ests, compares them with the current interrupt priority leve l (IPL) , and determines if the request w i l l be s e rv i ce d . T h e interrupt requ ests are c hecked at the beginning of each microcycle and the interrupt update is forwarded to the I Box . That all happens through the central control logic before the next microcycle begins .
Exte rna l interrupt processing has been implemented on-chip in the 78032 to avo id the complex ity that results from having the interrupt priorities arbitrated outside the chip. Since these priorities are an integral part of the processor state, an off-chip design wou ld invo lve broadcasting the interrupt priority level eac h time it changed . More over, off-chip interrupt processing wou l d also require additional hardware o n t h e C PU board .
•
The microsequencer accepts inputs from various points on the chip and generates the next microaddress to access the contro l store . The microsequencer logic pe rforms such operations as microsubroutine calls and returns, microcode traps, n-way (or case) branches , and s i gned offset cond it ional branc h e s . I m pl emented in the m i c rose quencer is an e ight- level m icroprogram stack.
•
The control store is a 3 9 -bit ROM with 1 6 00 entries . It receives microaddresses and status s ignals and generates the next set of microin structions . The control store transfers those microinstructions to the control section in the center area . That section , in turn, generates control signa ls for the three principal functions in the main datapath : the I Box , the
E Box , and the M Box. The access time of the control store is less than
I
00 ns .
•
The DAL interface handles a l l control signals and transfers data and addresses between the chip and local memory, peri pherals, and other devices outside the c h i p . The DAL interface transparently processes variable length operands and al igns data references that cross natural 3 2 -bit memory boundaries.
It also causes the microprocessor to sta l l dur ing r;o referenc e s , so t h at add i t i ona l microcode is not needed tO test for I/0 com pletion. The DAL interface contro ls transac tions involving the C PU chip, the FPU chip, and externa l devices. It also a rbitrates direct memory access (DMA) requests .
•
The clock generatOr receives an externa l 40 -
MHz c lock reference and produces the eight
2 5 -ns clock phases that time functions on the chip . The control logic of the chip makes extensive use of bootstrapped drivers . For that reason, certain clock phases have to drive very high capac itances, as much as 2 5 0 picofarads . To assist in that task, a special driver c ircu it with cu rrent- l i m iting resistors is used to provide fast edges without using excessive power or silicon area. These resis tors control the overlap cu rrent drawn d u r ing bootstrapping and provide a voltage drop during the overlap.
External Interface
A principal goa l in designing the chip's exter nal interface (Figure 3) was to demand as few support functions as possible from the CPU board . The 7 8 0 3 2 chip provides seven hard ware interrupt inputs . Fou r of these inp uts
(IRQ<3 : 0> L) correspond to standard VAX I/0 interru pts and res u l t in vectored interrupt transacti ons . Three others (I NTTIM L, PWRFL L,
HALT L) have preassigned interpretations and the corresponding vectors are generated inside the chip. The 7 8 0 3 2 takes in a double-fre qu ency clock input from a standard oscil lator.
The chip produces a normal-frequ ency clock output, which can be used to drive or synchro nize external logic. The functions between the chip and the Q-bus can be implemented in off the-shelf discrete logic .
New Products
Di�ital Tecbnica/ journal
No.
2
Manb I ')8()
1 9
-------
The Micro VAX 78032 Chip, A 32-Bit Microprocessor
I N T E R R U P T
C O N T R O L
D M A
C O N T R O L
{
{ H A LT E R R
-
R OY PWR F L
--
I N T T I M
I R Q < 3 0>
--
D M R
--
D M G
B M < 3 :0>
-
D S
-
AS
D A L < 3 100>
�
/"
� r-V
A D D R E S S
LAT C H
M i c ro V A X
78032
C E N T R A L P R O C E S S I N G
U N I T
� r-V
DATA
T R A N S C E I VE R S
-
E R R
R O Y
B M < 3 O>
D S
AS
BA<3 1 : 00>
�
8 0 < 3 1 : 00>
--
R E S E T
C L K I
-
O B E
-
W R
-
-
EPS
-
�
�
M 1 cr0VAX 7 8 1 3 2
F L O A T I N G
PO I N T
U N I T
C S < 2 0 >
C L K O r--
D B E
W R
E P S
C S < 2 : 0 >
C L K O
Figure 3 External Interface
E xcept for the 3 2-bit DAL bus , the external interface closely resembles t hose for existing
1 6-bit microprocessors . Specifical ly, its timing and signal complement are quite similar tO those in current machines . The addresses and data on the DAL are time division multi plexed, with separate timing strobes (AS and DS, respec tively, in Figure 3) . The data direction and the data buffer signals (WR and DBE in Figure 3) are used to control exte rnal transcei ve rs directly. The cycle status signals differentiate among the various types of bus transactions .
Four-byte mask signals , one for each group of eight bits on the DAL bus, allow straightfor ward manipulation of bytes within longwords
(four bytes) .
The
ROY signal a l l ows slower peri p heral devices on the I/0 bus to stretch the me mory access time beyond 4 0 0 ns until they are ready tO respond .
Parallel Opera tio n
Besides giving the 780 3 2 opti mized microcode and a fast microcycle time, the design team en hanced the chip's performance by allowing parallel operations between and within func ti onal subsections. This parallel flow is actually a form of pipel ining in which the operations happen independently a nd concurre ntly. For e xa m p l e , w h i l e the E Box is execut i n g a datapath operation, the control store can access the next microinstruction . At the same time , the
2 0
Digital Technical journal
No. 2 March 1 ')8()
microsequencer can be calculating the add ress of the microinstruction after that one, and the
M Box can be trans lating a virtual ad dress .
Meanw h i l e , the I Box can be decoding an instruction or operand specifier and prefetch ing more instruction data . And the DAL inter fa ce can be i n i ti a t i n g or c o m p l e t i n g a n external bus operation .
For example, assume that the chip is tO exe cute the fo l l owing two t h r e e - m icro cyc l e macroinstructions i n sequence :
ADDL3 RO, R l , R2
SUI3L3 R4 , R S , R6
Withi n the thi rd 2 0 0-ns microcyc!e, some operations associated with these two macroin structions are performed in parallel by several subsections. The E Box will write the resul t of
ADDL3 into R2 in the register fil e , set the PSL condition codes, and check for arithmetic exceptions, such as an overflow trap. Mean while, the I Box w i l l decode the next macroi n struction , SUBL3 , and its first specifier, R4 .
Concurrently, the prefe tcher in the I Box will determine i f the decode of the instruction and specifier w i l l c le a r e n o u g h space i n t h e prefetch stack t O warrant another longword transfer. If so, the I Box will then initiate the transfer and fetch another macroinstruction, which also involves the DAL interface .
Within each subsection, there are a lso a num ber of parallel operations that reduce the over all execution speed significantly. In addition tO simultaneous prefetch and decode actions in the I Box (as described above) , the microcode access in the control stare is pipeline d : The next microaddress is accessed while the current microinstruction at the current microaddress is being executed . In the M Box, length checks against referenced addresses take place simulta neously with the translation cache l ookups . If a lookup misses , therefore, the length check will have already determined whether or not the ref erenced page is within range . In the E Box, a separate program counter ( PC) adder maintains the PC so that the ALU can be dedicated to its primary tas k.
Some typical execution times for instructions u nder normal operating conditions (al igned operands, no memory management exceptions) are as fol lows :
In struction
MOVL
ADDL2
MOVL
A D D L2
MOVL
A D D L2
Conditional
Branch , not taken
Conditional
Bra nch, taken
Operands
Reg , Reg
Reg , Reg
Mem, Reg
Mem, Reg
Reg , Mem
Reg , Mem
Typical
Execution Time
(Nanoseconds)
400
400
800
800
600
1
2
00
200
800
Physical Implementation
The MicroVAX 780 3 2 chip is made using a
3-micron, double -meta l NMOS process that a llows power savings and superior circuit flexi bility . Until the MicroVAX 7 8 0 3 2 chip design, single metal was a standard for NMOS technol ogy. The use of a second layer on the 7 8 0 3 2 c h i p was a significant departure for NMOS design . There are two main advantages of a double-metal implementation . First, it is easier to place logic circu its in the interconnect layer, where there are more circuits per u nit area of silicon . Second, the metal interconnect has lower resistance than polysilicon, thus avoiding wire delays that are difficu l t to elim inate in design .
The double-metal process provided the chip design team with two layers of alu minum inter connect and fou r types of devices (N, E, L, and
D) . The four types allow some savings in power and a substantial i ncrease in circuit flexibility.
However , the E device ( l ight enhancement) is typical l y used only in source-follower circuits, and the L device ( l ight depletion) only in latches and stati c memories . The second layer of aluminum interconnect manages the com plexity associated with 3 2 -bit microprocessors .
That p e r m i t s g l o b a l c o m m u n i c a t i o ns a n d a llows local control or routing to share the same chip area. H owever, second metal can only contact first meta l , and then only through an offset, or staggered, contact.
New Products
Digital Technical journal
No. 2 March
1986
2 1
The Micro VAX 7803 2 Chip, A 3 2-Bit Microprocessor
Figure
4
X-shaped Cells
The control store is a 1 600-entry by 39-bit
RO M . Although its size was decreased mostly t h r o u g h r e p a r t i t i o n i n g a n d o p t i m i z e d microcode , abou t ten percent of the reduction was gai ned through the cell structure chose n .
X-shaped ce l ls w i t h a virtual -ground design were used (Figure 4 ) . This ROM has no physi cal grou n d , whereas standard ROMs with
H-sha ped ce lls have one ground li ne for every two data lines. The X-shaped cell, which is 9 5 microns square, is also more dense t han the standard ce l l . Moreover, in the X-shaped ce lls, second metal is stra pped across the top of the array to min imize the row propagation time .
The cell access time is 1 00 ns .
The ROM bit l i nes are precharged tO Vee using depletion pull ups. Sensing is done with a cross -cou ple d s tage u s i ng local deplet ion divider voltage references set at 0.6
X
V00 . Col umn access occurs in 2 5 ns.
The control c i rcu its ( at the center i n
Figure 1 ) are imp lemented in dynamic logic so that the total power dissipation is kept be low three watts. That also al lows a low-cost packag ing design . The eight clock p hases provi de refresh timing references to the dynamic logic .
Due to tight silicon constraints, the test fea tu res built i nto the design had to be limited in scope . The principal ones used are as follows :
•
•
Serial s h ift registers with fe edback for observing the control srore , I PLA, and mi cro sequencer outputs
Special test mode for overriding normal sequencing with external microaddresses
•
Dedicated mi crocode for opti mizing state observations in the special test mode
Summary
The M icroVA.,"'{ 7803 2 represents a major break through both in semiconductor technology and in the VAX fam i ly. From a technology perspec tive, it is the first imple mentation of a success fu l 3 2 -bit superm inicomputer on a single chip.
I t is the first chip to provide integral demand paged virtual memory management. And it is the first chip to provide system performance comparable tO the l l j780 . From a VAX per spective , the 78032 is the key tO the downward extension of the industry-standard VAX family i nto the realm of small systems and worksta tions .
Acknowledgements
The authors acknowledge the technical contri but ions of John Beck, Sandy Carrol l , Gerry Che ney, Mary )o Doherty, John Glynn, Jim Gorr,
Bob Grondalski , Dave Grondalsk i , Pat Hart ,
Ernie Hohengasser, Taan Lee , Steve Morris,
Tony Pasqui to, Steve Thierauf, Tim Thrush,
Janet Vitello, and Barry Worster.
References
1 . VAX A rchitecture Handbook
(Maynard :
Digital Equi pment Corporat ion, Order
No. EB- 1 9 5 8 0 , 1 98 1 ) .
2 . W . N . Johnson, " A
V LSI
Superminicom puter CPU,"
IEEE In ternatio nal Solid
State Circuits Co nference Digest of
Technical Papers
( 1 984) : 1 7 4- 1 75 .
3 . J . Slager et a ! . , "A 1 6-bit M icroprocessor w i t h O n - c h i p Memory Protec t i o n , "
International Solid- State Circuits Con fe re nce Digest of Techn ical Pape1·s
( 1 983) : 2 4 - 2 5
2 2
Digital Technical ]out-nat
No. 2 March 1986
4 . H . M . Levy and R . H . Eckhouse, Computer
Programming a n d A rchitecture: The
VAX- 1 1
(Bedford : Digital Press, 1 9 80) .
5 . D .W. Clark and ].S. Emer, "Measurement and Analysis of I nstruction Use in the
VAX- 1 1 /78 0 , " IEEE Proceedings of the
9th A nn ual Symposium on Comp uter
A rchitecture
( 1 982) : 9- 1 7 .
6 . j .S . Emer and D .W. Clark, "A Characteri zation of Processor Performance in the
VAX- 1 1 /780," IEEE Proceedings of the l i th A n n ual Symposium on Comp u ter
A rchitecture
( 1 98 4 ) : 3 0 1 - 3 1 0 .
7 . W . O . Strecker, "Transient Behavior of
Cache M emories , " A CM Transactions on Comp u ter Systems, val . 1 , no . 4
(November 1 98 3 ) : 2 8 1 -2 9 3 .
8 . D .W. Clark, "Cache Performance on the
VAX- 1 1 /7 80 , " A CM T1-ansactions on
Compu ter Systems, val .
1 , no. 1 (Febru ary 1 98 3 ) 2 4 - 3 7 .
9 . G . M . Tarolli and W.j . Herma n , "Hierar chical Circuit Extraction with Detai led
Parasitic Capacitances, " A CM IEEE 20th
Design A u tomation Conference Pro ceedings
( 1 98 3 ) : 3 3 7-34 5 .
New Products
Digital Tecbnical]ournal
No. 2
March
1986
2 3
William R. Bidermann
Amnon Fisher
Burton M. Leary
Robert]. Simcoe
William R. Wheeler
The Micro VAX
78132
Floating Point Chip
A separate chip, the 78132, in the Micro VAX II system petforms fast floating point calculations. Three datapaths, each controlled by microcode, work in parallel to yield a 1 00-nanosecond microcycle. The wide datapaths accommodate a large variety of instructions, using microwords of only 35 bits for control. The 78132 is a 3-micron NMOS chip connecting to the CPU chip of the Micro VAX
II system via a general purpose protocol and a limited set of lines. Crosstalk and resistivity posed particular design problems, as did the routing of signals and power. The 78132's electrical integrity was carefully checked to ensure high reliability.
Scientific and engi neering applications req uire strong floating point su pport from their com· pute rs . All VAX i m plementations offer both microcoded (warm) and hardware ( hot) capa bilities to execute the 95 floating point instruc tions in the fu l l VAX i nstruction s e t . The
MicroVAX II processor also supports floating point instructions, b u t in a slightly different fas h i o n . S i n c e t h e c o n tro l s t o re i n t h e mi croprocessor, the CPU chip, has a lim ited size, these instructions are not execu ted in m i crocode ; instead they are e m u l a t e d i n macrocode . 1 · 2 Emulation i s relatively slow and does not provide the fast speeds requ ired for i ntensive mathemat ical applications . Therefore , a separate floating point acceleratOr (FPA) , the
MicroVAX 78 1 3 2 chip, has been deve loped as a companion tO the CPU chip, the MicroVA.,'{
7803 2 chip.
The 78 1 3 2 , or FPU chip, i s designed tO pro vide fast floating point calcu lations on a single chip. It executes 6 1 of the 70 floating point instructions i n the MicroVAX i nstruction set.
Nine of the 70 i nstructions simply move data, and the CPU chip does not need the FPU chip tO handle them. The FPU chip also accel erates cal culations for 9 integer i nstructions, which are associated with integer m u ltiplies and divides .
The FPU chip executes i nstructions about 1 00 ti mes faster than macrocoded emulation .
The FPU chip (Figure 1 ) contains 3 2 , 1 4 1 transistors i n a 3-m icron, double-metal NMOS chip, which req u i res just u n der 2 watts of power at 5 Vdc . It measures 8.4 by 6 . 6 mm and is packaged in a 68-pin leaded chip carrier. The chip has a 1 DO -nanosecond (ns) microcycl e , d ivided i nto four 2 5 -ns c lock phases generated from a 4 0-MHz i nput clock. The CPU chip, which also operates on a 4 0-MHz input clock, has a microcycl e of 200 ns . The faster m icro cyc le and wide datapaths enable the FPU chip tO perform fl oating point operations much faster than the CPU chip with its general data path.
This paper discusses the implementation of floating point i n the MicroVAX II's FPU chip and the u n ique constrai nts of a s i ngle-chip floating point accelerator. These constra ints are not l i m ited only to architecture but include interface design, wiring, and signa l integrity, all areas where design trade-otis are important.
At the highest leve l , the FPU chip imple ments the F, D, and G fl oating point instruc tions in the VAX instruction set . The chip is constrained by the requ i rements of the VA.,'\ architecture-data formats, accu racy require ments, and i nstruction vaga ries-and by the characteristics of the technology-lim ited num ber of pins, lim ited die size , and l i m ited in ter-
24
Digital Technical Journal
No. 2 March J 986
Digital Technical journal
No.
2
March 1 986
Figure 1 Photom icrog raph of the FPU chip
2 5
New Products
The Micro VAX 781 3 2 Floating Po int Chip connect. These constraints di ctated many of the design considerations in the FPU chip.
FPU Chip Architecture
The mai n el ements of the FPU chip, shown in the block diagram i n Figure 2, are similar to those in most floating point devices . � Three separate processors-a 67-bit fraction processor, a 1 3 -bi t exponent processor, and a single-bit sign processor-operate in parallel . The bus interface unit handles data transfers over the external bus to the CPU chip and data move ment into a nd out of the three datapaths . The mi crosequencer controls the parallel opera tions of the processors.
Each e lement i n the FPU chip operates i n parallel t o speed u p instruction processing. The microsequencer steps through the m icrocode for an i nstruction and determines which opera tion is to be performed by each processor for the current cycle. The m icrosequencer also ta kes inputs from each of the processors to determine which microword is to be executed next. The datapath of the fraction processor performs a l l the arithmetic computations on the mantissa of a floating point number. This datapath is designed to be flexible enough to handle the many different operations required in a general-purpose FPA. The datapath is also segmented to hand le the F, D, and G data types, and is optimized to provide the maximum pos sible performance from the N-channel MOS technology.
The datapath of the exponent processor han dles only the exponent portion of a floating point nu mber. The exponent datapath is a lso used as a counter during certa in operat ions such as m u ltiply and divide. This datapath does all the exception and bounds checking for operations l i ke addition and subtraction. The sign processor is incorporated into the expo-
SIGN
PROCESSOR
I
EXPONENT
PROCESSOR
I
B U S
INTERFACE
U N IT
FRACTION
PROCESSOR
I
I
MICROSEQUENCER
Figure 2 Block Diagram of the FPU chip nent datapath and handles a l l operations per taining to the sign bit. During an addi tion or subtraction, the sign bit determi nes which case is performed by checking the signs of the two operands and the opcode of the i nstruction .
The bus i nterface unit (BIU) is responsible for handl ing all the FPU portions of the bus traffic between the FPU and CPU chips. The BIU decodes the opcode sent to the FPU chip and tells the m icrosequencer which instruction to execute . That allows the FPU and CPU chips to coordinate their actions without a lot of proto col or pins. Since many different data types are processed, the BJU is responsible for u npacking the operands and steering them to the appropri ate datapath . O nce the instruction is com plete d , the BIU takes the u npacked result from each datapath and formats the result i nto the specified data type . Figure 3 contains a more detai led block d iagram for the entire floating point unit.
Algorithms
To keep the FPU chip at a size that cou ld be produced, we decided not to use special -pur pose hardware to implement instructions like a d d i t i o n or m u l ti p l i c a t i o n . I n s t e a d , t h e datapaths are designed t o b e general-purpose ones to accommodate the needs of a wide vari ety of instructions.
Addition and Subtraction
The datapaths are u nder microcode control and work in para l l e l . Wit h i n eac h , the steps requ ired for either addition or su btraction are done serially. First, the exponents of the two operands are compared to see if they are of equal magnitude . If not, the larger exponent is stored in a register, and the exponent differ ence is used to control the a l ignment. The shifter on the output of the fraction arithmetic logic unit (ALU shifter) al lows the fraction with the smaller exponent to be aligned five bits at a t i me . During each alignment step, the exponent difference is reduced by u p tO a magnitude of five until the exponents are equal . Once equal, the fractions are added. (In subtraction, the fraction tO be al igned is complemented before alignment.)
The resu lting fraction is then normal ized.
The normalize s hift is accomp lished by a single left shift i n the fraction ALU and two left shifts i n the ALU shifter. I f the addition of the
2 6
Digital Technical Journal
No. 2 March 1986
CLKI
I
CLOCKS
RTs'('l
I
1 0
8U F-H R 1J
�·AIN
SE;QUF;�CE R
CZOO X 351
13
CS2 CSI CSO [P'S Wi'i
Figure 3 Block Diagram of the FPU Processor fractions results in an overflow i nto the top guard bit, a single right shift i n the ALU shifter is requi red to normalize the result. During nor mal ization , a 3-bit code is sent to the exponent datapath, which determi nes the amount the exponent must be adjuste d .
After normalization , the fraction i s rounded using a rou nding constant appropriate for the data type of the floating point operation being performed. If the round resu lts i n an overflow in t h e fraction datapath, the exponent is incremented by one and the fraction is normal ized. The exponent datapath then checks the resulting exponent for any error conditions . If no errors are found, the final fraction and expo nent values are loaded into the output register and the sequencer signals the BIU that the oper ation is complete .
Multzply
The mu ltiply operation in the FPU chip is based on a 3-bit retirement algori t h m . The 3 -bit retirement, or octal multiply, m ust generate the requ ired mu ltiple , 0-7, of the m u ltiplicand to be added into the partial produ ct for each step .
The multi pl es must be generated by simply shifting the multipl icand and adding or sub tracting them from t he partial product. The multiples 0, 2, 4 , and 8 are easy to generate in this way. The multiple
6 can be formed by tak ing three-quarters of the multipl icand and stor ing that in a register at the beginning of the mu ltiply (� X 8
=
6) .
As shown in Table 1 , all the even multiples can be generated . To gener ate all the odd m u ltiples, a -1 multiple is added to achieve the final exact multiple for each retired group of three bits .
Digital Tecbntcaljournal
No. 2 Mm·ch 1986
2 7
New Products
The Micro VA X 78132 Floating Point Chip
Table 1 Multiply Operation - Booth Encodings
Multiplier
Group
Required
Multiple
Data
Used
Multiple
Shift
000
001
01 0
01 1
1 00
1 0 1
1 1 0
1 1 1
2
3
4
0
1
5
6
7
0 mult mult mult mult
3.4 mult
3/4 mult mult
0
2
2
3
3
3
Mu ltiple
Added
6
8
4
6
2
4
0
2
Mu ltiple
Owed
0
-1
0
-1
0
- 1
0
-1
The key to making this scheme work is that this - 1 multiple must be generated from the previous group of three bits . To that group, the
- 1 multiple for the next group is equ ivalent to a -8 mu ltipl e . To know whether or not the next group will need the - 1 multiple, it is sufficient to examine the least significant bit (lsb) of the next group of bits. If the lsb is a 1 , then the gro u p w i l l be odd a n d wi l l n e e d t h e
- 1 multiple. This process is started by examin ing the lsb of the multiplier and initializing the partial product register to either zero or minus the m u l tipl icand . If the lsb is a 0, the - 1 multi ple will not be neede d . The operation always term inates i n the case not requ iring compensa tion because the numbers are a l l norma lized.
Table I shows the Booth encodi ngs for each multiplier group.
These Booth encodi ngs translate into the frac tion data path controls depicted in Table 2 .
A multipl ication in the FPU chip is begun by load ing the mu ltiplier into the Q Register (quo tient register) and loading the m u l t i p l icand into register 0 in the scratch RAM
.
Three quarters of the multiplicand i s then calculated during two ALU cycles and is stored in register
I of the scratch RAM . Subseq u e n t l y , the
A Register is initialized to store the partial products.
During each cycle of the m u ltiply loop, the four least significant bits of the Q Register are latched to control each multiply step. Based on these fou r b i ts , the m u ltiply control loads either the mu ltipl icand or three-quarters of the multipl icand from the scratch
RAM into the
8
Register . The control then adds or subtracts the 8 Register from the A Register. The resu lting new partial product is shifted right by the ALU s hifter and relatched in the A Register. The
Q Register is then shifted three bits to the right to retire the c u rrent set of multiplier bits and to set up for the next iteration .
The exponent datapath i s used to control the nu mber of iterations that should occur for each multiply operation and to calcu late the resu lt ing exponent. The number of iterations that take place for a multiply depends on the length of the mantissa . For example, an F format num ber with a 2 3 - b i t mantissa req u i res eight iterations .
Dit•ision
The floating point unit performs a I . 5 - bit, non restoring division . This algorithm is similar to a
1 -bit, non-restoring d ivisio n , b u t takes advan tage of the fact that long strings of zeros or ones in the partial remainder can be skipped over without doing an addition or subtraction. The
FPU chip handles double precision through i ts normal datapath .
Within the FPU chip, the partial remaind ers will always be < + Y2 and > -Yz because both floating point nu mbers are norma lized . If the partial remainder is small relative to the nor mal ized d ivisor, a 1 w i l l not be shifted into the quotient over the next few cycles (The oppo site is true if an addition is performed . ) Know ing this fact and whether the previous opera tion was an addition, subtraction, or a shift wi
I I determine how the quotient bits are deve loped.
If the previous operation was a shift, the pro cess is in the middle of a long string of zeros or ones and no addition or subtraction has to be performed. If the partial remainder is not small re lative to the normalized divisor, the quotient bits are deve loped as they wou ld be in a 1 -bit
28
Digital Technical journal
No. 2
March
I
'J86
0
0
0
0
0
0
0
0
Table
2
Mu ltiply Operation - Fraction Datapath Controls
Next Group
Look Ahead
Actual
Multiple
Present
Group
Group
Multiple
Group
Multiple
4
5
0
1
2
3
6
7
0
1
2
3
4
5
6
7
000
001
01 0
01 1
1 00
1 0 1
1 1 0
1 1 1
000
001
01 0
01 1
1 00
1 0 1
1 1 0
1 1 1
4
6
6
2
4
0
2
4
6
6
8
2
4
0
2
8
-8
-8
-8
-8
-8
-8
-8
-8
0
0
0
0
0
0
0
0
Multiple
Generated
ALU
Operation
6
8
-8
4
6
2
4
0
2
-6
-6
-4
-4
-2
-2
0
A - A
A - A+B; B=RO
A ,_ A+B; B=RO
A ,_ A+B; B=RO
A ,_ A+B; B=RO
A - A+B; B= R 1
A ,_ A+B; B= R 1
A
<----
A+B; B=RO
A - A-B; B=RO
A <---- A-B; B=R 1
A
<----
A-B; B=R1
A - A-B; B=RO
A ,_ A-B; B=RO
A - A-B; B=RO
A - A-B; B=RO
A <- A where: RO contains the multiplicand
R 1 contai ns 314 multiplicand
New Products division algorithm . Tabl e 3 s u mmarizes the
1 . 5-bit, non-restoring division.
The implementation of this algorithm i n the
FPU chip is straightforward . To start, the divisor is loaded into the B Register and the dividend into the A Register. The Q Register is initialized to
0 and will become the location where the quotient is deve loped .
During each step of the d ivision, quotient bits are inserted at the least significant end of the Q Register. The register contents are then shifted left either 1 or 2 as required to develop the new quotient for that step. If necessary, the divisor is added to or subtracted from the par tial remainder. The result is then shifted left by the appropriate number of places.
When bit 65 in the Q Register becomes a 1 , the division is stopped . Since these numbers are normalized, the resu l t will fal l in the range of greater than Yz but less t han 2 . The contents of the Q Register, already norma l ized, are then read back into the A Register. However, if the initial subtraction resulted in a positive partial remainder, then one must be added to the expo nent to accou nt for the fact that the result has a whole part ( i . e . ,
� 1 ) .
Integer Division
The FPU chip a lso performs a 1 -bit, non-restor ing divide algorithm , which is used to acceler ate the execution of the DIVL and EDIV instruc t i o ns . I n a l l case s , the integer divide is accomplished with a 3 2 -bit d ivisor and a 64-bit d ividen d .
Polynomial Calculations
The polynomial evaluation a lgorithm , POLY, uses Horner's Method to calculate all trigono metric functions . Because execution time can be so long, POLY is the only VAX floating point i nstruction that can be interru pted by the CPU chip. The algorithm performs a series of ax+ b operations once d u r i ng each cycle. I n each operation, x is treated as a constant, the value of b is provided by the CPU chip, and the value of ax+ b in the current cycle becomes a in the next cycle.
The FPU chip first m u ltipl ies a by x with the
MUL algorithm and then adds b with the ADD algorithm. The main sequencer tel ls the I/0 control ler that the first POLY cycle has been completed and that the result is ready in the
Digital Technical journal
No. 2 March
1 986
2 9
The Micro VAX 78132 Floating Point Chip
Table
3
1 .5-Bit Division Operation
66
Most Significant Bits of Partial Remainder
65 64 63
Value of bits
66-63
Shift
Left
ALU
Operation
Add/Sub
Quotient
Shift
Quotient
0
0
0
0
0
0
0
0
0
0
1
0
0
1
0
1
0
0
'/e
1f4
3fe
>-'12
_J/a
-%
-1/a
2
2 none subt subt subt add add add none
1 0
1 0
00
00
0
0
1
0
0
1
2
2
1
0
0
01
01
1 1
1 1
Add/Sub Quotient: Bits sh ifted into the quotient if previous operation was an addition or subtraction.
S h i ft Quotient:
Bits shifted into the qu otient if the previous operation was a pure shift (no ALU operation). ljO registers for transfer to the CPU chip. The s e q u e n c e r e x e c u t e s t h e s e c o n d M U L ,
(ax + b )x, during the time that the CPU chip is reading the first resu lt, storing it in a register, and transferring the next value of b tO the FPU c h i p . T h e s e c o n d A D D o p e r a t i o n ,
(ax + b)x + b, then takes place to complete the second cycle, and the process continues. The
CPU chip's register is updated with the new resul t at the end of each cycle. This pipelining a llows fast generation of trigonometric and transcendental functions. Both the CPU and
F P U chips are working to i m p le m e n t the i nstruction, and t he actual m u l tiply time is overlapped by the operand fetch time.
The Microsequencer
The microcode for the FPU chip is contained in a large programmable logic array (PLA) , which is the heart of the microsequencer. I nputs to the PLA are received from all major sections of the FPU chip. A microword of 35 bits is all that is needed to control the two main datapaths
(the sign processor is part of the exponent datapath) and to communicate with the bus i nterface unit. Each field in the microword is encoded to reduce the number of wires routed to the other sections. Two hundred microwords are required to implement the sixty-one float ing point and n i ne accelerated integer instru c tions executed by the FPU chip. The block dia gram for the microseq uencer is s hown in
Figure 4 .
Inputs to the PLA are comprised of five next address bits, three ded icated inputs, and forty signal s from the three major processors on the chip. Three bits from the next-address field are used tO select five of the forty signals for the next FPU cycle. These five multiplexed i nputs, in conjunction with the eight direct inputs, are used to address the next microword . The thi rty five outputs, or signa ls, from the PLA are used to com municate with the rest of the fl oating point u ni t . These signals determine which operation is to be performed by each of the three data paths (exponent, fraction and sign processor) .
Interface Between Chips
Inte1jace L ines
The communication between the CPU and FPU chips is done through a very limited set of lines: a write (W) strobe, three cycle status
(CS) l ines , an external processor strobe (EPS) , and the 3 2-bit data and address li nes (DAL) .
( Th i s a p p r o a c h was used t o red u c e t h e pincount o n both chips .)
In t he MicroVAX II processor, the chip proto col is designed as a genera l-purpose one so that other coprocessors cou ld take the place of the
FPU chip. Each interface l i ne has a specific pur pose, as explained below.
•
The W strobe sends a signal from the CPU chip to indicate the direction of data flow over the DAL. For the FPU chip, the write
30
Digital Technical journal
No 2 March 1986
DATAPATH
STATUS l3 BITS
40 1 NPUTS
+
40:5
MUX
NEXT ADDR<9:7> is BITS
AND PLANE
(200 TERMS)
3 BITS
NEXT ADDR<6:5> i2 BITS
I
MICROCODE PLA
30 ADDITIONAL
OUTPUTS
OR PLA N E
(200 TERMS)
Fig ure
4
Block Diagram of the Microsequencer signal indicates that data is being transferred from the CPU chip.
•
The EPS is used by the CPU chip to qualify all communication between itself and the
FPU chip or other non-memory devi ce.
•
The three CS l ines provide status about the current bus cycle. Two of the l ines indicate the type of information be ing transferred; they are "val id" when the external processor strobe is asserted. The third line is an open drain output (fu nctionally similar to an open collector in TTL) , which will be active when the bus cycle is a response enable and the
FPU chip has completed the current com manded operation.
•
The DAL is a 3 2 -bit, bidirectional bus that exchanges data between the CPU and FPU chips. The CPU chip is always the bus master and controls the transfer of operands to the
FPU chip and results back to itself.
The information exchanged between the C PU and FPU chips could be of different types : write external processor command, read or wri te external processor data, command to other external processors (not the FPU c hip) , and externa l processor response enable. The exter nal processor strobe (EPS) is used by the CPU chip to qualify a l l commu nication between itself and the FPU chip.
Figure 5 i l l ustrates all the interface l ines between the two chips.
Co m m u n icatio ns Pro tocol
The commun ications protocol permi ts the FPU and CPU chips to communica te efficiently.
Every interchip operation will be assoc iated with the following sequence of bus activities :
1 . The CPU chip initiates an interaction by placing a command onto the DAL bus, a status code on two CS lines, a write sig nal of " low , " and an E PS of " l ow . " The
FPU chip recognizes this sequence as a command-write cyc le and aborts any instruction being executed. The FPU chip then decomposes the command to determine the req uired operation and the nu mber and size of the operands .
2 . The CPU chip fetches the required oper ands and executes one or more data write cycles to transfer them to the FPU chip.
3. After transferring the last operand , the
C PU chip asserts a response-enable sig nal on the CS l ines and pulses the EPS
"low . " The chip does that once for each m icrocycle that it has control of the bus in order to determine if the FPU chip has finished processing the data .
4 . To signal the completion of operations, the FPU chip asserts the CS<2 > l i ne
" low" when the response-enable signal is on the two CS lines and the E PS is
" l ow . " At the same time, the FPU chip asserts the status of the just-completed operation.
5 . The CPU chip recognizes the "low" sig nal from the FPU chip and reads the sta tus i n formation . The C P U c h i p w i l l repeat this transaction t o compensate for
New Products
Digital Technical journal
No. 2 March J 986
3 1
The Micro
VAX
78132 Floating Point Chip
I R 0 <3:0> L
Vss
�
C L K I
!
I
G N O ,
AS L
Vee
I NTTI M L
OAL<31 :00>
+
LATCH
OMR L
O M G L
B M <3:0> L
M I C ROVAX
CPU C H I P
- TRANSCEIVER
OS L
PWRFL L
WR L
O B E L
H A LT L
E R R L
R O Y L
R ESET L
'-
EPS L
FLOATI N G
P O I N T
C H I P
CS<2:0>
I t
CS<2:0>
Figure
5 Interfaces Between the CPU and FPU Chips
AS L
BA<29:00>
8 0 <3 1 :00>
WR L
O B E
L
C L K I
Vss
G N O , Vee its microcoded pipeline, capturing the status information the second time.
6 . The CPU chip executes zero or more data-read cycles to read the results, if there are any, from the FPU chip. Both chips are now free to perform the next transaction in the instruction stream .
(The FPU chip w i l l respond unpredictably to other nonstandard protocols and relies on the sequence of interactions described above for proper operation . )
Performance A nalysis
The performance of the FPU chip is very sensi tive to the 1/0 bandwidth. Every floating point opera t i o n i s assoc iated w i t h a spec ified sequence of events that must occur berween the chips before the execution can start. There is another sequence of events that must take place when the computation is completed.
These sequences happen without any parallel ism or pipelining.
The protocol affects the performance o f the
FPU chip because cycles must be expended for sending and reading status signals, and transfer ring data . Table 4 i l lustrates the i nd ividual steps that occur for three types of operations :
ADDF, MULF, and MULD . For these examples, assume that no time is spent on i nstruction fetch and decode , and that the memory subsys tem has an unlimited bandwidth and buffering capability for reads and outstanding wri tes. The performance is measured from the completion of the initial instruction decode to the final result store in the memory (or a register) .
The total execution time for other instruc tions can be derived in the same manner using the fol lowing internal execution times :
Add in D format
-
700 ns
Division in F format
-
2 2 0 0 ns
Division in D format 4 4 0 0 ns
3 2
Digital Technica/Journal
No. 2 March 1 986
Table 4 Steps for Add and Multiply Operations
Instruction: ADDF
Specifier decode and data transfer for first operand
Specifier decode and d ata transfer for second operand
Internal transfer (first operand)
Execution
Status read
Status read
Result transfer on DAL bus
Total
Total Execution Time:
Instruction: MULF
Register Mode
Protocol
Time
Execute
Time
(nanoseconds)
300
200
1 00
600
200
200
200
1 1 00 700
1 .8 microseconds
Specifier decode and data transfer for first operand
Specifier decode and data transfer for second operand
Internal transfer (first operand)
Execution
Status read
Status read
Result transfer on DAL bus
Total
Total Execution Time:
Instruction: MULD
Register Mode
Protocol
Time
Execute
Time
(nanoseconds)
300
200
1 00
1 900
200
200
200
1 1 00 2000
3.1 microseconds
Specifier decode and data transfer for first operand
Specifier decode and data transfer for second operand
Internal transfer (first operand)
Execution
Status read
Status read
Result transfer on DAL bus
Total
Total Execution Time:
Register Mode
Protocol Execute
Time Time
(nanoseconds)
400
300
1 00
2700
200
200
400
1 500 2800
4.3 mi croseconds
Byte Displacement
Protocol
Time
Execute
Time
(nanoseconds)
500
500
1 00
600
200
200
400
1 800
700
2.5 microseconds
Byte Displacement
Protocol
Time
Execute
Time
(nanoseconds)
500
500
1 00
1 900
200
200
400
1 800 2000
3.8 microseconds
Byte Displacement
Protocol
Execute
Time
Time
(nanoseconds)
600
600
1 00
2700
200
200
800
2400 2800
5.2 microseconds
New Products
Digital Technical journal
No. 2 March
I
')86
3 3
The Micro VAX 78132 Floating Point Chip
Wiring and Signal Integrity in the
FPU
Signal integrity in a large VLSI chip such as the 7 8 1 3 2 is fundamental to ensure correct functionality and good yield, given the varia tions in manufacturing. The one- to two-micron proximity of signal lines on an integrated cir c u i t ( IC) can cause s i g n ificant cou p l i ng problems . Moreover, there are p roblems i n terms o f clock distribution and power-supply noise . The design of the logic must allow suffi cient noise margin to permit correct operation in spite of the noise present in the system . The use of charge as the signal (used in many cir cuits in an NMOS design) , rather than voltage or cu rre n t , created s o m e s p e c i a l d e s i g n problems for the FPU chip tea m .
IC Wiring Characteristics
The FPU chip has four layers-two of metal, one of polysilicon, and one of diffusion-that are used to interconnect and form devices . The wir i ng in an IC is conceptually similar to the wir ing on a printed circuit board. Although the total wiring length on the FPU chip is only about four meters, the interconnected nodes and elements number in the tens of thousands.
Placing and routing the logic functions i nevita bly affects the estimates of loading and system performance . Thus an iterative process of first routing a design , then simu lating the subse quent performance is needed to identify a workable routing plan . Once this workable routing-performance trade-off is identified, the final routing and loadings can be made .
The wiring considerations for a VLSI design are different from those for conventional sys tems in several ways. First, the dimensions are smaller. In the NMOS process the horizontal metal separation is about three microns and the vertical separation is from one to two m icrons .
Even with the smaller size of the wiring in the
MicroVAX II chips, crosstal k can become a seri ous prob l e m . On a M O S c h i p , c ross t a l k between poorly designed nodes can approach fifty percent. The capacitance on many of the critical nodes in the FPU chip is only about 1 00 femtofarads (0 . 1 picofarad) . Any cou piing at all on these nodes becomes quite significant. The largest capacitance on the chip is the clock l ines at around 1 1 0 picofarads. On dynamic nodes , which rely on a charge stored on a capacitor t o represent a logic leve l , this coup ling is particularly troublesome .
To eliminate this problem on the FPU chip, the design team checked each of the over
1 2, 5 0 0 nodes for crosstalk from all other nodes in the chip. This data was then used to change the layout , where appropriate , to minimize or in some critical cases, eliminate intolerable levels of crosstalk. These checks took about three man-months to complete .
Another difference in the wiring of a VLSI chip is the resistivity of the wiring. The metal layers in the FPU chip have resistivities on the order of 1 00 m i ll iohms per square . However, the resistivities of the polysilicon and diffusion interconnect layers are about 40 ohms per square , or 4 00 times that of the metal layers.
The interaction of this parasitic resistance with the on-chip capacitive loads can cause serious p er fo r m a n ce l i m i t a t i o n s if not carefu l ly monitored.
I n fact, these two layers are so resistive that they were u nusable for u nconditional routing of either signals or power; they could be used only for very local routing.
As a special precau tion , a hand-check of those layers was made at pattern generation time to verify that no long, speed-critical paths utilized these layers as part of the routing network.
Power and Signal Routing
A minimum-width wire routed the length of the
FPU chip has a resistance of about 2 0 0 ohms .
The use of metal layers with noticeable resis tance therefore begins to set system perform ance limits through RC delays as well as I R drops , which happens in larger systems . The clock distribution introduces a delay of about one nanosecond across the FPU chip, due solely to the resistance of the metal interconnect and the distributed load capacitance . This de lay amounts to about fou r percent of the length of a single p hase in the chi p . A well-monitored clock distribution system is a requ irement in any semiconductor chip. The problem is that the performance of the u nderlying semiconduc tor device is beginning to outstrip the capabil ity of the chip wiring to distribute the clock.
RC delays become the limiting speed factor of the wiring in an IC, while the speed of light across transmission J ines is the limiting factor in a larger syste m . These resistances can also
3 4
Digital Technical Journal
No.
2 March 1986
seriously affect the power and ground supply as it is distribu ted throughout the FPU chip.
We used several techniques to keep the sup ply noise under 200 mV as power is distributed throughout the chip. First, the tota l de current was calcu lated by summ ing the current used in each power and ground line as it joined other branches on the route to the actual su pply pad .
At this point in the net , rwo factors had tO be analyzed so that the width of the power bus cou ld be sized correctly. That sizing kept the equ ivalent resistance low enough so that the overall drop from a pad to the most remote logic cou ld be kept under 200 mV. Unfortu nately, that sometimes requ ired large (on an IC scale) power buses in which a significant frac tion of an ampere must be provided by one supply line.
The second prob lem, and the more difficult one , associated with the power and ground wir ing is the large ac voltage transients that can occur when large portions of the system switch at the same time. That problem is especially significant with the V55 lines. And i t is particu larly difficu l t when driving wide buses or large datapaths as wide as the 8 1 bits in the FPU chip.
In these cases, large transients (one ampere or more) flow in ground and power lines for a few nanoseconds. In a large system environment, decoupl ing capacitors can be used to supply these currents locally. Unfortunately, that is not possible in an IC environment where such large capacitors are not practical .
As a resul t certain ground l i nes in the FPU chip are allowed to have significant noise on the m . In some cases this noise spike can be as much as rwo volts.
This noise is handled by r� mn ing these " dirty" grounds in a separate metal line all the way back to the pad on the chip.
However, even when the li ne is taken back to the pad to prevent local IR drops from u pset ting the logic, parasitic inductance in the pack aging can still cause problems . The most strik ing exam ple is that of off-chip bus drivers . Here a typica l 3 2-bit bus is driven over 4- or 5 -volt swings in as little as fou r or five nanoseconds.
With each bus load being on the order of 1 00 pf, the large dijdt that the chip imposes on the power pins causes inductive ringing. Solving this problem by placing a decoupl ing capacitor on the externa l pins is of l ittle value s ince the package inductance effect ively isolates t he capac i t o r from t h e a c t u a l nodes i t m u s t decouple inside the chip. Therefore, the FPU chip, like most chips that drive wide buses, has separate power pins going only to the output transistors. The subsequ ent ringing is tOlerated since it does not affect any internal logic. (The ringing can become even more of a prob lem on c hips with several buses with different timings, since separate supp l i es must be used fo r each bus . That drastically increases the number of supply pins required on the chip.) The FPU chip devotes 1 9 of its 68 pins to V55 and V00 distribution.
Electro migration
A final wiring consideration in design ing the
FPU chip was e lectrom igration. Electromigra tion is a reliabil ity issue in IC wiring because high current density in the metal interconnect can cause the metal to m igrate , thinning sec t ions of wiring until they fi nally fa i l . Current densities much higher than 1 05 amperes per square centimeter can cause increases in wiring res istance and eventually, open c i rcu its or increased interlevel leakage , and s hort circu its.
Clock lines, power and ground buses , as well as some global wiring, are susceptible to this fai l ure mechanism.
As a resu lt, all lines o n t h e F P U chip have an a d d i t i o n a l cu rrent constraint imposed by electromigration . When the chip was designed, these lines all had to be checked to eliminate the problem .
Wiring In tegrity
Considerable time was spent checking the elec trical integrity of the wiring in the FPU chip.
The fol lowing list contains the most important wiring integrity checks made of the intercon nect on the chip:
1 . Transistor SourcejDrain Integrity - This check assured that the silicon intercon nect resistance caused less than five per cent degradation .
2 . RC Delays - Al l RC de lays greater than one nanosecond were analyzed .
3 . Cou p l i n g - Al l i nternodal cou p l i ng capacitors were checked to verify that there would be less than 2 0 0 mV of noise injected into the node.
4. V00 and V55 Nets - Three checks were performed. First, al l IR drops were mea su red tO ensure that ac and de voltage
New Products
Digital Technical journal
No. 2 March
1 986
3 5
The Micro VA X 78 132 Floating Po in t Chip sources were kept under 200 mV. Sec ond, all buses were sized to verify their r e l i a b i l i ty for el ectromigration resis tance. This check included contact elec trom igration . Th ird , a check ensured that sufficient isola ted power pins existed to guarantee that clean and dirty grounds were isolated.
5 . Clock
-
An analysis identical to that for
V
00 and V55 nets was done on all eight clock l i nes.
Although there were significant CAD tools to perform most of the checking, this task atone required approximately ten percent of the total engineering time for the entire project.
Summary
The VLSI chips we are now designing are as complex as several boards of TTL used in past implementations of the VAX architecture . The
FPU chip performs the same functions at about the same speed as five boards containing ICs in the VAX- 1 1 /780 system. The designs of these complex systems on chips present a set of con straints and considerations similar to and yet different from those encou ntered by board level system designers. We hope that this paper captu res t h e com p l e x i ty a n d u n i q u e ness i nvolved in the MicroVAX FPU chip.
Acknowledgements
The FPU chip team completed the design of two VAX floating point chips, the MicroVAX
FPU and the 8200 chip, in e ighteen months .
That was possible only because another design team working on the J - 1 1 FPA had established the basic arch itecture and took the time to he lp our team to understand that work . This c lose worki ng relationship allowed us to complete the MicroVAX FPU design in step with the CPU chip tea m , which was our major challenge .
References
1 . D .W. Dobberpu h l et a ! , "The MicroVAX
7 8 0 3 2 Chip, A 3 2 -bit Microprocessor,"
Digital Technicaljournal
(March 1 986, this issue) : 1 2- 2 3 .
2 . R.J . Simcoe e t a l , "A Floating Point Un it for a 3 2 -bit M i croprocessor Syste m , "
Proceedings of the 1984 IEEE Custom
In tegrated Circ u it Co nference
( May
1 98 4 ) : 4 78 -4 8 1 .
3 . G . Wolrich et a l , "A H igh Performance
Floating Point Coprocessor , " IEEE jour nal of Solid State Circuits, vol . SC- 1 9 , no. 5 (October 1 98 4 ) : 690-696
3 6
Digital Technical jounlal
No. 2 March 1 986
Barry A. Maskas
I
Developing the
Micro VAX II
CPU Board
Within the Micro VAX
II
system, the CPU board provides an environment to optimize the performance of the CPU and floating point processor chips. The board is designed as a linked sequential machine to accom modate the sequential control of the CPU chip. A Q-bus handles I ;o for the system. The memory access path is dual ported, allowing the memory and the CPU chip to run synchronously without wait states. A scatter gather map provides Q-bus address translations. To minimize p1·oduct delivery time, the CPU board was developed in parallel with the chips.
Using CAD tools helped to go from first-pass chips to running the
Micro VMS system in only two weeks.
The CPU board i n the MicroVA.X I I system
(Figu re 1) holds two chips: a microprocessor, ca l led the CPU c h i p , and a fl oating point coprocessor, cal led the FPU chip. The board also integrates a synchronous memory su bsys tem, a synchronous I/O-bus controller, and a synchronous on-board 1/0 subsyste m . The pro ject tO develop the CPU board was governed primarily by t i m e -tO-market consi derations .
Figure
I
The Micro VAX
II
CPU Board
Other factOrs, such as VMS and ULTRIX compat i b i l i ty , performance, re liability, cost, and ease of high-volume production were a lso important criteria . The end resu lt is a su ccessfu l balance between a l l these factors .
Development Goals
The im portance of the primary goal governed how the project team orga nized itself to make decisions and tO execu te tasks. Rapid decision making, and parallel and overlapping act ivities were the norms for this deve lopment effort .
Unfortu nately, para l l e l activi ties can cause communication problems, thus increasing the r i sk s of prod u c t fa i l u re . Howeve r , these problems were anticipated and mechanisms put in place tO reduce the risks to an acceptable leve l .
The CPU board was designed around the specifications of the CPU and FPU chips, which were bei ng developed at the same time . There fore, one deve lopment goal was to m i n i m ize the dependency of the board design and layout on the first-pass designs for these chips. The team aimed at provid ing a fu lly fu nctional sys tem environment into which the first-pass chips cou ld drop. This aggressive approach lead the team to leap-frog over events rather than to take a conventional steppi ng-stone progression. The overal l project manager encou raged the taking
Digital Technical journal
No .
2
March I
')8(i
37
of pru dent risks because he was respons ible for m e e t i n g t h e deve l o p m e n t sch e d u l e . The acceptance of these risks eventually paid off in an on-time delivery of the CPU-board design .
Single-board Design
Deve loping the CPU board aro u nd the two chips required us to provide a specific system environment. That environment had to ba lance the memory bandwidth of the CPU chip against its I/0 bandwidth requirements. The real iza tion of that balance is the key to the board's success. Having either a slower memory or a slower ljO subsystem would degrade system performance by at l east twenty-five percent.
The environment also had to su pport the
Mi croVMS , ULTRIX, and VAXE L N operating systems.
Our goal was to provide the hardware speci fied by the three operating systems on one
Digital-standard quad-sized board
(8-!12 by
I O-Y2 inch) . The single-board goal was a consequence of technology improvements balanced by the costs of replacing the unit in the fie l d . In this case , needing fewer pieces to build the system wou ld reduce manufacturing costs, i mprove reliabi lity, and ease maintainability costs . The objective of operating at the ful l bandwidths of the chip and the IjO bus was especially chal lenging when so l i ttle board space was avai la ble for the necessary functions .
Most new chips do not run at their ful l speed immediately; they take some time to debug.
Our design objective was to run the CPU chip at an operating frequency lower than i ts maxi mum during the first-pass debug. Of course , ru nning at a slower clock rate was never an acceptable compromise for the fi nal product.
(Two versions of the CPU board were devel oped with minimal component differences, one running at the ful l 2 0 0-nanosecond (ns) micro cycle speed and the other at a slower 24 2 -ns microcycle speed.) However, if the first-pass chip had missed its performance target , the deve lopment of the CPU board could still have continued. It is a tribute to the chip designers that the first-pass chips did run at ful l speed , which was quite unusual in so compl icated a product.
The bus c hosen to meet the 1/0 needs of the system was the Q2 2-bus. This 2 2 -bit bus has sufficient bandwidth to hand le traffic from the system disk, the Ethernet
LAN, and other ljO sources, such as other processors. The risk of using this bus was low due to its proven design, and the deve lopment cost for this application was reasonable. The Q22-bus is also supported by many disk, tape, and other ljO products from both D i g i ta l and t h i rd-party add-on manufacturers .
CPU Board Functions
We ruled out using the Q 2 2 -bus for accessing memory directly, since the bus cou l d not meet the memory cycle time of 400 nanoseconds for the CPU chip .
1
Therefore , a new memory archi tecture had to be devel oped. We i nvestigated rwo alternative schemes, the first being the widely used d irect me mory access (DMA) with a s i n g l e port. Unfort u natel y , DMA forces addresses and data to cross the microproces sor bus on their way to memory. The usual pro cedure is to halt the microprocessor with a
D MA request or grant while the DMA device uses the m icroprocessor's data and address paths . In this case the CPU chip, having no cache, wou l d waste time by exercising the memory request and memory grant signals.
Therefore, we chose the second sche me, a dual ported memory control ler. Figure 2 depicts the single- and dual-ported memory control lers that were considered .
This dual-ported control ler requires that the
CPU chip have different address and datapaths for the Q2 2-bus and the memory control ler.
While a DMA access is taki ng place , the CPU chip can continue operating on i ts 3 2 -bit exter nal datapath, primarily comm unicating with memory and the FPU chip. I n this context, memory cycles can be pictured as strings of
4 00-ns time slots controlled by a central arbi ter. This memory contro l ler m i n imizes the i mpact on the CPU chip's performance by DMA accesses to memory on the Q 2 2 -bus. This organization is not locked u p by asynchronous
Q2 2 -bus cycles, whose transactions are three to four times slower than the CPU chip's memory cycles. It also al lows the Q 2 2 -bus protocol to operate autonomously with the CPU chip and memory, except when the buffered bus proto col and the memory system exchange buffered data .
The memory controller also serves as an alter native to one based on a cache . The CPU chip does not implement an i nternal cache due to power and chip-size constraints.
1
Cycles for
3 8
Digital Technical Journal
No. 2 March 1986
CPU
C H I P
'----
FPU
CHIP f--
132
LATCH
+24
M U LTIPLEXER
+10
M EMORY
ARRAY
+32
SING LE-PORTED ORGANI ZATION
-
LATCH
SCATTER/
GATHER
MAP r--
-
DATA
TRANSCEIVER
1 6
(f)
:::>
(lJ
0J
0
CPU
C H I P
FPU
C H I P r-f-f----12--
,---
LATCH f24
MULTIPLEXER
+o
MEMORY
ARRAY
+2
LATCH
SCATTER/
GATHER
M A P
-
'-----
DATA
TRANSCEIVER
DATA
TRANSCEIVER
DUAL-PORTED ORGA N I ZATION
Figure 2
Block Diagrams of the Proposed Con trollers
�
(f)
:::>
(lJ
0J
0
DMA, refresh i ng memory, and CPU-chip access are interleaved in time .
The MicroVAX I I system is designed to be used in a multicomputing e nvironment. There fore , the bus interface logic has to accommo date the role of e i ther bus arb i ter or auxiliary processor. To that end , a doorbel l register facil itates an i nterprocessor i nterrupt mechanism.
The datapath of the Q 2 2 -bus interface has to provide the address translations from the virtual memory space of the bus to the address space of memory.
We defi ned several other elements as being essential for supporting an operating system on a s i ngle board . Those are the t i me-of-year
(TOY) clock, the console serial line, the VAX console command program , and the console i nterface-boot and self-test ROM . These ele ments, a long with some status and error regis ters, comprise the on-board 1/0 subsystem .
Digital Technical journal
No. 2 March I 986
39
N e w Products
The fu nctional organization of the CPU board is depicted in Figu re 3
Linked Sequential Machines
Optimizing the overa l l computer performance means that data transfers between the CPU chip and memory have tO be as fast as the chip can operate . Withou t a cache me mory , the CPU chip has a relatively long memory cycle time of
400 ns (two m icrocycJes) . Thus CPU chip-tO memory data transfers can take place without wai t states .
The 4 00-ns 1/0 cycle is nevertheless fast enough that the CPU board had tO be designed as a l inked sequential machine rather than as flow-through logic. The control function in the
MicroVAX II system receives signals, interprets the m , and generates control outputs , a l l in a defined sequence . This mode of contro l cannot be satisfied using a combinational logic syste m .
In addition t O perm itting 4 00 -ns mem o ry cycles without wait states, sequential machine design requi res less random logic and board space than a flow-through design . The des ign process is si mpl ifi ed because the machines are implemented in easily changeable FPLS (fuse progra mmable l ogic sequencer) l ogi c. Moreo ver, design changes can be read i ly documented and less time is needed for de bugging and trac ing events. Sequential circuitry is more easily s i m u l ated than random logi c , i n which a l l events must b e sampled. And , since the CPU board 's logic compone nts run on the same clock, it is possible tO debug them at faster or slower operating speeds .
When the C PU-board project starte d , this sequential mac h i ne approach had nor been widely used in microcomputer design . Off-the shelf hardware and adequate CAD tools were not available . This project shows that designing with commerc i a l PALs and F P LS l ogic can reduce the chip cou nt, as well as cost and deve lopment time .
The overa l l control logic of t h i s l i nked sequ entia l machine is d ivided into partitions .
The events inside i ndividual partitions are gov-
EXPANSION MEMORY
DATAPATH
�
CONSOLE
SERIAL L I N E
CONN ECTOR/
DISTRIBUTION PANEL
SCATTER-GATHER MAP
1 M B OR
256 KB DRAM
GATE
ARRAY
022-BUS
INTER FACE
GATE
ARRAY
M ICROVAX 78032
INTERFACE
BOOT/
DIAGNOSTIC ROMS
MICROVAX 78032
M I CROPROCESSOR
4 0
MI CROVAX 78132
FLOATING POINT U N I T
C/D INTERCONNECT
EXPANSION MEMORY
CONTROL PATH
D c B
Figure
3
Functional Partitions of the CPU Board
A
Digital
Technical journal
No. 2
March
I
Y86
erned by i ndependent sequential mach ines, called contro llers. The logic within a partition goes through a fixed, repetitive sequence of operations , or states, duri ng the fou r quarters, or phases , of a microcycle. The operations of the various partitions are coordinated i n two ways . First, all sequential machines run from the same clock so that their timing is based on the same stream of clock edges . Second, the sequential machines are constantly exchanging signals, providing each other with the protocol i nformation needed for coordi nating their flow sequences .
The sequential machines can be classified as modified Mealy mach ines 2 The outputs are determined by the present input condi tions and the present state of the machine. However, the state register is separated from the output regis ter, with the AND programmable logic array fed by both the state register and the i np uts tO gen erate OR p l a ne terms for the cl ocked S R latches . The advantage o f clocked S R latches is that the past state need not be regenerated by every clock edge; only changes need activate an
OR term. Using D-type latches would require that regeneration .
The block diagram i n Figure 4 depicts a sequential machi ne representation of the CPU board's functional configuration in Figure 3 .
Under the on-board control partition at the left, the control function for t he memory su bsystem is distributed among three sequential devices : the memory sequencer, the memory arbiter, and the auxiliary device controller. Under Q 2 2 bus control, there are also three sequential devices : the sl ave, arbitra t i o n , and master machines . These machines exchange request, acknowledge , and status signals to control operations .
Memory Subsystem
Our market research data suggested that the on board memory should be either 2 5 6 kilobytes
(KB) or 1 megabyte (MB) . The amount depends on whether 6 4 K DRAMs or 2 5 6K DRAMs are used . At the time the design was started, 2 5 6 K parts were i n short supply. Therefore, using
64
K
D RAMs was a strategy to cou nter that s hortage .
The function of the memory controller is to carry out 4 00-ns read and write operations and to refresh its RAM chips. This control ler con-
New Products f-.-----
ON-BOARD CONTROL
--------to>-------
022-BUS CONTROL -----a:
0 rJ)
UJ
() >o w < a. < CI: o "- < a: CI: w o w >-
1
MEMORY
SEQUENCER i
MEMORY
ARBITER
I
I
I
I
l u--
I
I
I
SLAVE
MACHINE
ARBITRATION
MACHINE z
I u
BUS
INTERFACE
GATE ARRAY
I
I
AUXILIARY DEVICE
CONTROLLER
I
�
MASTER
MACHINE
I
I
I
I
I
Fig ure
4
Block Diagram of the Control A rchitecture
Digital Tecbnicaljournal
No. 2 March 1 986
4 1
.
-------
Developing the Micro VA X II CPU Board tains a Q 2 2-bus scatter-gather map that handles transfers between the Q2 2 -bus virtual memory and on-board physical memory.
Memory access is controlled by the memory arbiter. This arbi ter checks for outstanding memory access requests i n a fixed-priority sequence at the ends of 200-ns idle cycles and
4 0 0 -ns memory cycl es . It also c h ecks for requests from the Q22 -bus slave machine , the memory-refresh counter, and the CPU chip, in that order. The fixed-priority sequence resolves col l ision requests for memory usage . If the arbi ter requires exclusive control of the memory subsystem, a locking mechanism built into the subsystem prevents contention.
When the CPU chip requ ires a memory-read lock, the memory arbiter will stall the chip and direct the Q 2 2 -bus arbitration machine to sus pend other bus activity. Those actions will hap pen only after any pending memory cycles of the slave machine have been completed. The arbitration machine will retai n Q 2 2 -bus master ship until the writejunlock cycle of the CPU c h i p frees t h e b u s . Unti l t h e arbi tra t i o n machine becomes Q22 -bus master and while the CPU chip i s stalled, the memory arbiter will perform the demand-driven refresh cycles and resolve slave-deadlock cycles from the Q 2 2 bus . As each memory cycle is completed, the memory arbiter c hecks t hese requests again, and either the Q 2 2-bus or the refresh-memory cycle can begin at the next clock edge . If no
Q22-bus or refresh requests are pending, the arbiter anticipates that a CPU-chip cycle will be next.
That anticipation and the fix e d -priori ty sequence save a lot of program execution time .
The CPU chip makes about seventy percent of all memory references. Slave machine accesses by the 1/0 bus devices occur twenty percent of the time (a maximum burst rate , not the aver age rate) , and those by the refresh counter, two perce n t . (The remainder are i d l e cycles . )
Therefore the controller, b y anticipating that the CPU chip-rather than the I/0 bus or the memory-refresh counter-wi ll make the next memory access, allows a memory cycle of 4 0 0 ns, instead o f 6 0 0 ns. (The 600-ns cycle would be necessary because the address strobe of the
CPU chip wou ld have to assert before the mem· ory cyc l e c o u l d start , t h u s was t i ng o n e microcycle .)
When timing microcycles, the memory arbi ter enables the memory sequencer at phases coincident with the CPU chip's entry to a new microcycle . This enabl ing happens even though the sequencer does not yet know whether or not there will actually be a memory access by the CPU chip. Not until three phases later can the sequencer determine whether or not the address strobe has been asserted for a memory reference . If so, the sequencer enables the con tinuation of the anticipated memory access .
After that cycle completes, the next memory access wi ll be enabl ed , and the procedure repeated. If not, the sequencer " k i l l s " the cycle and runs another poll loop after checking for Q 2 2 -bus slave or refresh requests. Not antic ipating a memory access would reduce per for m a n c e by a p p r o x i m a t e l y t h i rty - t h re e percent.
The memory sequencer generates the row and column address strobes, sets up reads and writes on each byte , and handles pariry genera tion and detection . The auxiliary device con troller can "stretch " the memory cycle of the
C PU chip to synchronize its timing with slower devices, such as the TOY clock and the boot
ROM .
The scatter-gather map converts between the
22 -bit virtual addresses of the Q 2 2 -bus ( 4 MB addressable) and the 2 4 -bit physical addresses of the memory (up tO 1 6MB addressable) . As defined by VAX memory management, the 4MB is divided into 8 1 9 2 pages of 5 1 2 bytes eac h .
The 2 2 -bit virtual address consists of a 1 3 -bit page number and a 9-bit offset to the addressed byte in that page . The 24 -bit physical address consists of a 1 5 -bit page number and a 9-bit offset. An entry in the map for each 5 1 2 -byte page and offset points to a location in physical memory. Each physical address has four byte masks that select which bytes are inactive on any memory reference.
There are, of course, other ways tO map addresses ber�'l'een the 1/0 bus and memory.
One way is one-ro-one address trans lation, which in this case would have restricted physi cal memory to 4 MB . Another way is first tO map one-ro-one into the lowest 4 M B of memory.
Then, the CPU chip can perform the transla tions and data transfers to the proper pages in the address space of the remaining memory.
Unfortunately, this approach is unacceptable due to its effect on performance. A third way is to have fewer than 8 1 9 2 mapped pages. In this case , programmers might have tO provide the ir own mapping software for many real-time IjO
4 2
Digital Technical journal
No.
2 March 1986
app l i cat ions . That typ i c a l l y i nvo lves D MA access to large numbers of RAM locations . None of these methods proved as satisfactory as the use of the scatter-gather map.
Interface Co ntrol Signals
The interface control signals to the CPU chip include the fol l ow i ng:
•
Cloc k- i n (40 MHz) , clock-out (20 MHz; used to time the sequential machines) , and reset signals
•
Address, data , externa l-processor, and tim ing-strobes-out signals
•
Three chip-status , four byte-mask, and the readjwrite signals
•
DMA-request and DMA-gran t signals
•
Four interrupt - li n e signals and one HALT s ignal
•
Ready and error signals
The pu lse of the design is a four-state grey code bi nary counter, which is c locked from the synchronous clock-out signal of the CPU chip.
The first edge assertion of the clock-out signal after power-up puts the CPU chip in the first
5 0-ns phase of the four-phase m icrocycle . The grey code allows the memory arbi ter and auxil iary device controller to track the state of the microcycles. The 28-bit address of the CPU chip is decoded to select the accessed device and then encoded into a series of 3 -bit cycle cod es . The a u x il iary device controll e r , the me mory a rb i t e r , and the master m a c h i n e decode those cycle codes t o identify what type of timing cyc les to sequence through . The two key signa l s , apart from the cycle codes, are those for the address strobe and the readjwrite .
They direct the auxi liary device controller, the memory sequencer, and the master machine to perform the read or write operations with the device specifi e d in the cycle codes .
Those three e lements control the CPU chip's cycles and any system exceptions via the ready and error s ignals. The DMA request s ignal is used only during a reset operation to delay the
CPU chip until the system has finished reset ting . The byte-mask signals si mply direct the control logic to perform certai n operations.
Those i nc l u d e masked ( byte or word) or unmasked ( longword) memory cycles and data fu nneling operations on the Q 2 2 -bus. (Data fu nne l i ng converts 3 2-b i t Iongwords tO 1 6-bit words and v ice-versa . ) The unmasked cycles are required since the Q 2 2 -bus is 1 6 bits wide, whereas the me mory and CPU-chip buses are
3 2 bits wide . The on-board ljO time can be extended tO accommodate s l ower external devices. The memory controller a l l ows t he memory cycle to end only when a device has asserted a ready (ROY) signal , indicating the completion of i ts task.
Add- o n Memory
The syste m's memory can be expanded with one or two memory boards, each containing e ither 1 , 2 , 4 or 8 megabytes . Thus total mem ory can be as large as 1 6MB and stil l offer a fixed 4 00-ns access time with no wait-states .
Each board is linked to the CPU board by means of a local i nterconnect. This interconnect con s ists of special control signals on the C and D rows of the Q 2 2 backplane and a 5 0-pin mod u l e -header and ribbon cab le for data . Each i nterconnect l inks a board d i rectly to the one just below it in the board cage of the system enclosure . Thus control signals and addresses can pass directly between the chips and mem ory without using the Q 2 2 -bus. The diagram in
Figure 5 shows the functional organization of the memory boards.
For ease of installation and maintainabil ity, t h e a d d - o n m e m o ry boards are s e l f- c o n figurab le; there are n o user-settable switches or j umpers on the CPU board or memory boards.
This design requires a logic fu nction that com bines active addresses with static configuration data tO generate the proper control strobes accord i n g to the configu rat i o n . Therefore , a lthough the add-on memory boards are posi tion i ndependent, they "recogn ize" which expansion slots they occupy. (To get the fu l l
1 6MB configuration, t h e memory control ler design su pports 1 MB-by- 1 DRAM chips.)
On- board IjO Su bsystem
The serial line interface in the on-board ljO su bsystem provides the CPU board with a fu 1 1 duplex, RS-4 2 3 EIA conso le terminal interface .
The console i nterface program is imp lemented in macrocode in the boot ROM . The console mode functions include general booting, user computer interface , se l f-test and HALT. The boot ROM also includes special su pport fu nc-
New Products
Digital Technical journal
No. 2 March
1986
4 3
MEMORY OATAPATH
( 1 0 MB/S)
PR IVATE MEMORY
1
A
C/D I N TERCONNECT
022-BUS
BLOCKMODE DMA
(3.3 MB/S)
1/0
MASS
STORAGE
Figure
5
Fu nctional Partitions of Memory Modules tions for the software i n the MicroVMS, ULTRIX and VAXELN systems .
As the boot ROM goes t hrough a self- test sequence , programmable LEDs display the test status, identifying any board subsystem that conta ins a fail ure. By analyzing this sequence for effect iveness , we found that it provided a confidence level of eighty-six percent in the functional integrity of the CPU board and add on memory boards. Although some Q2 2-bus l ogic functions cou ld not be tested with this method , it he lped to reduce significantly the ti mes to do manufacturing and field service tests .
To emulate a CPU-halted condition, the CPU chip can be directed by either software or hard ware switches to transfer program control to a firmware routine at a fixed PROM address. The
HALT function retains the board state . The CPU chip traps to the boot ROM when there is a
HALT, masking it u ntil there is an i nstruction fetch outside the ROM . While i n this emu lated
HALT, the firmware will perform the specified operations only after receiving either console commands or a signal from the AUTO-REBOOT switch .
The CPU chip does not have a RESET i nstruc tion; t he chip simply sets a RESET request flag.
The UNJAM command in the console mode initial izes the bus by forcing the CPU chip to the
D MA grant state . UNJAM then transfers control to RESET in the interface gate array of the CPU chip . After that, the logic resets the board's functions and the arbitration machine resets the
Q 2 2 -bus. Any auxiliary processors are reset from the Q 2 2 -bus reset signal .
Exceptions , which may originate in the con sole, the on-board 1/0, the Q 2 2 -bus, or the memory subsystem , are reported to the CPU chip for a machine check. This process invo lves setting an error-register flag i n the i nterface gate array of the CPU chip. The chip then treats the exception as either fatal (HALT or AUTO
REBOOT) or non-fata l (abort the process) .
Board Components
Logic hardware for the CPU board was selected by balancing the need for minimum power and board space against the use of low-cost, off-the shelf components. The gate arrays for the CPU board and the bus interface, for i nstance, are more expensive than discrete logic; however, they are necessary to fit all support functions on one quad-sized board . Due to a conductivity connectivity l i m i tation through the board's edge fi ngers , the maximum allowable power consumption is 45 watts for a 1 MB on-board memory configu ration . We were also con-
44 Digital Tecbnicaljournal
No. 2 March 1986
strained by the watts per square inch that had to be conducted from the board surface to the environment. That was important given that the enclosure is cooled by the flow of forced air.
The gate array for the CPU-chip i nterface decodes add resses a n d latches boot-ROM words. This gate array also contains registers for booti ng, diagnostics, and memory subsystem errors; the on-board 1/0 datapath ; and the inter rupt-acknowledge decode and control .
The gate array for the bus interface includes such components as the doorbe ll register, the memory-refresh cou nter, the holding latches for byte and word packing and u npacking, and timeout counters . This gate array also generates the bus addresses.
The memory subsystem i ncludes a number of discrete components . The memory arbiter and auxiliary device controller are both commer cial programmable sequencers . The memory sequencer consists of 1 2 d iscrete l ogic chips.
However, we had to design our own memory control lers . The ava i lable commercial ones cou l d not handle both the speed and the higher- level arbitration fu nction requ ired to anticipate memory accesses.
Previous board designs used an eight-layer construction tech nology (two power, fou r sig nal, two covers, and top and bottOm solder masks) . However, to reduce the board's cost, a six-layer technology had tO be developed (two power, fou r signal, and top and bottom dry-film solder masks) . Six-layer construction costs less than eight-layer due to alignment and dri ll ing problems with the stacked layers of the latter.
We used a CAD system to evaluate the chip interconnects on the board layout . The system showed that the signals could not be routed on two signal layers, but could on four. The two additional layers provide the 5V power and gro u n d p l anes . D i g i ta l ' s C o m p u te r -A i d e d
Design (CAD) Group in Maynard, Massachu setts, designed a custom software tool to he lp i n developing the board layout. With this tool , i t was possible to fit all functions on the board with 8-mil lines and spaces, and 60-mil pads.
Having the lines and pads as wide as possible offers satisfactory yield in production and good signal quality due to strip-line characteristics.
Enclosures
Two enclosures were considered to house the boards, the BA2 3 and the BA1 2 3 boxes. At the time, the BA2 3 box was an active product; only minor modifications were needed to accommo date it to the MicroVAX II system , a nice, l ow risk plan . In contrast, the BA1 2 3 box was still being developed. Using it represented a greater risk; however, it cou l d support more mass stor age . The backplane cages of either box could accept add-on memory and peripheral device interfaces on eit her quad-sized or dual-sized
( 5 - ll.; by 8-Y2 inch) boards. However, the BA1 23 box accepted more quad-sized and dual-sized boards . That was a distinct advantage because there would be different n umbers of board slots in the board cages i n different packages of the
MicroVA.X I I system . Moreover, each enclosure had a different thermal environment that had to be considered in the layout of the CPU and memory boards . Based on these considerations, we chose to use both the BA2 3 and BA 1 2 3 boxes as the enclosures for the boards.
CAD Tools
The tight schedule dictated that separate design teams had to develop each of the chips and the
CPU board as parallel projects . These separate efforts were made possible by the extensive use of CAD tools and computer simulation. Simu la tion was used extensively to design the CPU and FPU chips, the on-board memory and ljO subsyste ms, the gate arrays , the sequential machine controllers, and the Q 2 2 -bus. A board development tool set was selected from CAD packages avai lable i n the industry. Since these packages were genera l ly i ncompatible , we developed a process that transported wire lists between these various CAD tools. The process linked inp uts and outputs between the sche matic-capture work stations, the PC-board lay out system, the simulator, the gate-array ven dor, and the documentation control group. One key to the rapid development of schematics was to let the designers reta in control by perform ing their own drawings and edits.
We planned to use gate arrays right from the start of the project. Therefore, a hierarchical schematic-capture system was needed to fac ili tate the representation of devices at a number of levels. To verify the schematics, we selected a mixed-mode logic simulator that had l ibrary support for most of the off-the-shelf devices used in PC-board design . That minimized the develop ment time to construct the simulat ion libraries . A complete simu lation model of the
CPU board was also constructed to expedite the
Digital Tecbnical]ountal
No. 2 March 1986
4 5
New Products
-------
Developing the Micro VAX ll
CPU Board design veri fication process . This mode l pro vided a "soft" test bed for design changes before they were c o m m itted tO hardwa re .
Behavioral mode ls were used to simulate the signals from the CPU chip, as wel l as any device attached to the Q 2 2 -bus . No attempt was made to emu late the VAX i nstruction set. I nstead, the goal was to verify the sequences for read s , writes, interrupt acknowledgements , and the cycle flows for the block and non-block modes of the Q 2 2 -bus.
Several CAD packages developed by Digital were also employed to expedite the board design process. Figure 6 shows the CAD flow process that was asse mbled. (For more deta i ls on the CAD tool suite, see reference 3 . )
USER
HAN D·DRAWN
CONTROL FLOW
DIAGRAMS
WIRE LISTS:
•
•
• sac
2 GATE ARRAYS
2
MEM BOARDS
SEQUENCER DESIGN
(FPLS)
46
MICROPROCESSOR
HARDWARE
SIMULATORS
PROTOTYPE
SINGLE BOARD
COMPUTER
LOGIC
ANALYZER
Figure
6
CAD To ols Used in the CPU Board Development Process
Digital Technical journal
No. 2 March 1 986
Two CAD tools were used to help in the deci sion process for selecting re liable components.
The CPU board was mode led with the reliabil ity prediction program PREDI C, which is based on MIL STD 2 1 7 . PREDIC utilizes component thermal data from the second tool, the THUDS analysis progra m . Using these tools helped us to avoid the creation of hot spots on the board layout and the use of low-rel iabil ity compo nents.
These
CAD tools were so successful that the
CPU board was ready by the time the first-pass
CPU and FPU chips were ready. It then took only two weeks of debug to go from the func tional chips to running the MicroVMS operating system . In a l l , the deve lopment of the CPU board took less than one year from initial speci fication to operational prototypes .
Summary
The CPU board was designed as part of a larger project with formidable time constraints . Such an environment demanded that the design of any one component rely on the proposed speci fications for other, interlocking components, rather than on actual pieces of deve loped hard ware . That environment requ ired a cooperative team spirit that was goal oriented and fostered the assumption of rational risks . Both inter group and intra-group communication became extremely i mportant . The achievement of these factors was largely responsible for the su ccess of the MicroVAX II project.
Especial ly important was the fact that com munication was aided by the CAD tool suite used to support the overa l l project. In the case of the MicroVAX If system , we started from a wel l organized datapath and employed sequen tial machine architectures for controlling it. In that way, the design documentation, simula tion, verificat ion, and su pport were a l l made more manageabl e . In fu ture projects these tool suites wi l l mature and be havioral component models wil l begin to serve as design specifica tions. The ability to solid ify the design early in a project means that board designers can fash ion sil icon systems on boards that are func tional on the first pass .
References
1 . D .W Dobberpuhl et a l , "The MicroVAX
78032 Chip, A 3 2 -Bit Microprocessor , "
Digital Techn ical jo urnal
(March 1 986, this issue) : 1 2 - 2 3 .
2 . W . I . Fletcher, A n Engineering Approach to Digital Desig n
( Englewood Cliffs :
Prentice-Ha l l , Inc . , 1 980) .
3 . A . F . Hutchings, "The Evolution of the
C u s t o m C A D S u i t e U s e d on t h e
MicroVAX II Syste m , " Digital Technical jo urnal
(March 1 986, this issu e) : 48- 5 5 .
New Products
Digital Technical JounJal
No.
2 March 1986
4 7
Anthony F. Hutchings
I
The Evolution of the
Custom CAD Suite Used on the Micro VAX II System
The Micro VAX II chips were designed in only
20
months, due in part to simulation on
CAD
systems. Digital has a long history of using
CAD
.
Much of the Micro VAX If's CAD suite evolved from tools used on an earlier VLSI VAX design. The higher-level chip junctions were debugged using behavioral simulation, after which the circuits were modeled using the reliable SPICE and GRAPES systems. The
IV
system verified all inter connects and extracted wirelists, while other tools controlled the databases and checked design rules. The next generation of
CAD
tools must deal with a threefold increase in chip complexity.
The factors that must be considered when initi· ating and committing to a new VLSI design are qu ite complex . They are related in the fol low i ng way:
Market Requirements/Chip Definition
+
Technology Status
+
CAD Status
+
Engineering Talent Avai lable
Products w i th long l ead - t imes can accept h igher risks in the process c hosen for chip fabrication and CAD technology . However , products with short lead-ti mes, s u c h a s the
M icroVAX 7803 2 chip, can tolerate virtually no risk in this domain.
One way to reduce these r isks i s to test the chip designs by simu lating their performance before fabrication; another way is to check for a l l possible, known fabrication process viola tions before submitting the mask data for manu facture . CAD systems and tools have been devel oped for this purpose : to discover prob lems so they can be corrected at min imal cost, both in time and resources. Digita l Equipment Corpora tion was an early user of CAD to decrease the time-to-market for its VLSI products .
The M icroVAX I I project needed tO re ly o n a stab le CAD system and set of too ls while design i ng the 780 3 2 CPU chip (and i ts companion floating point coprocessor, the 78 1 3 2 FPU chip) . Much of the stab i l ity of the
CAD system was derived from work done to develop a mul tichip set for another VAX m icroprocessor. 1 We were able to both rational ize and simplify the resu lts of this p ioneering effort to s u i t the needs of the MicroVAX project. Let's begin by discussing this earlier CAD system to see how its use affected decisions made on the 78032 and 7 8 1 3 2 projects .
CAD System for Earlier VLSI VAX
Design
In many ways, the design process for the earlier
VLSI VAX microcomputer set the tone for all subsequent VLSI designs at Digital Equ ipment
Corporation . This process was characterized by the extensive use of simul ation, espec ially high l eve l , or behavioral , si mulation . The commit ment to high-level simu lation was particu larly innovative at that time.
Two types of simulation models were used for this earlier microcomputer. The first rype was d e s i g n e d a s a h i g h - l e v e l s o ft w a r e breadboard used to develop a n d check out the
4 8
Digital Technical journal
No. 2 March 1986
microcode before the chip hardware was avai la b l e . The second type was developed as a rela tively deta i led register transfer level ( RTL) model of the actual p hysical partitions and design concepts of the chips themselves. This model was used directly by the logic and cir cuit designers to develop the switch and cir cuit- level representations of the design.
One problem with using two models is that the output test vectors have to be checked con tinually to ensure compatibility between the microcode and chip designs . Thus, a lthough each was optimized to a specific task, the mod els proved to be somewhat cumbersome to use.
The hub, or kernel, of the data management system was called CHAS 2·3 This proprietary sys tem was developed at D igital's semiconductor faci lity in Hudso n , Massachusetts, expressly to form the nucleus of an integrated MOS custom design suite. The CHAS system performs the necessary data management functions on chip design databases and was originally intended to control all the design activities of a chip pro ject. The system embodies many of the "struc tured top-down design" principles of Carver
Mead.4
The C HAS system manages the data collected from c i rcu i t and l ogic s i m u l a t i on s, layout designs and syntheses, layout verifications, and schematics entry. This central system also pro vides data protection and conversion functions, as well as generating simulation wire l ists.
Decisions Derived from the Earlier
Project
From the outset , the CPU and FPU design teams made a number of important decisions based on the experience gained from the earlier project.
One driving factor in these decisions was the short time-to-market, which dictated that sim plifying the design process was a primary goa l .
•
The first decision was that there would be only one behaviora l , or functional , h i gh level simulation model of the chip rather than the two used earlier. Thus the func tional model was more complicated than the earlier one, but avoided the very time-con suming task of checking the output test vec tors . Using one model guaranteed that the microcode deve lopment would be in step with the chip design , since both teams had to use the same model .
•
The next decision was to carefu lly control the evolution of the CAD system that was used . Any experimentation with e nhance ments to existing CAD tools or with brand new CAD tools would be done only in a con trolled e nvironment. One project engineer, trained i n software and with CAD experi ence, was to be responsible for re-verifying the new functional ity and " robustness" of all new CAD releases. This approach enabled the team to acquire a vastly superior design ru le c hecker (DRC) , which considerably enhanced productivity during the physical design phase of the project.
This approach also differed greatly from that of the earlier project, although the lessons learned fro m that project conside rably shaped the team's attitudes. For example, the earlier project suffered-for a while-from attempting to use a first-generation layout editor that had too many bugs . (This tool was not in fact used on any part of the final design .) It also experimented with early ver sions of the CHAS system . These versions did not perform as well as desired for some func tions ( e . g . , the Assembled Block Wire l ister) .
I n contrast, the M icroVAX design teams decided to perform all layout on the indus try-standard
CALMA
GDS I I layout system, a robust and proven too l .
•
The third decision i nvolved the data manage ment of the design databas e . Rather than use a l l the features of the C HAS system, we decided to manipulate the design data using the s i mpler VMS fi le-management system with its loose but adequate version-control mechanisms. The CHAS system was used, but in the role of tool integrator, l i nking, for example, the QUI CKDRAW schematic editor to the SPICE circuit simulator.5 The CHAS system a lso provided a variety of val u able format conversion utilities .
•
The final decision was to use one proven tool for interconnection verification . This layout extraction; verification tool, called IV, per formed all the e lectrical connectivity check ing in a very efficient manner.6 The earlier project had used a combination of bought out tools and although that verification was very thorough, it was more costly than the s i ng l e -t o o l p rocess ( IV ) u s e d on t h e
MicroVAX project.
Digital Tecbnicaljournal
No.
2 March
1986
49
New Products
50
The Evolution of the Custom CAD Suite Used on the Micro VAX
II
System
The Design Methodology and CAD
Tool System
Having made these simplifications, the design team established a fixed defi n ition of the i r design methodology a n d CAD tool mapping.
This definition was fol l owed faithfu l ly through out the life of the project.
Figure I shows all the activities i n the design phase that were su pported by CAD tools . The middle column l ists each activity; the left col umn shows the type of data used in this activity and manipulated by the CAD tools which are shown in the right-hand col u mn alongside the actual activity and data they supportjuse.
The arrows indicate i teration paths where feedback is sent to a h igher leve l . That is, where results are obtained from a checking or verifica tion activity, it may be necessary to go back and modi fy an earlier set of assumptions and design decisions. For example, in ru nning the DRC, it is highly l i kely that we will find design rule violations that require us to correct the physi cal chip layout.
A nu mber of very i m portant paradigms should be noted .
1 . The behavioral model of the design was kept current with the logic design of the chip to guarantee the accuracy of the microcode with the chip design .
2 . The critica l hurdle for the functiona l correctness of the design was the correct execution of a certain number of VAX macroinstructions under an automated checking process . (The tool used for this process was called AXE, an architectura l test-case generator and execution too l , working in conju nction with the DECS I M system , Digita l ' s proprietary multi-leve l , m ixed-mode s i mu lation syste m ) . The minimum number of cases was 1 00 , 000 tests for each VAX instruction group . ln all, more than 1 m i l l ion tests were exe cuted before the chip was fabricated .
3 . The nu mber o f iterations during the lay o u t - d e s i g n p h a s e was m i n i m i z e d .
DESIGN REPRESENTATION USED
OECSIM Behavioral Modeling
Language )BOS)
Schematics
Switch Level C1rcuit Wirelist
Output Test Vectors
Circuit Netlist
Chop Floorplan
Sized Chip Schematocf
CALMA GOSII
Stream Format
CCF ) La yout
· Format)
CCF )Layout Format)
CCF )Layout Format)
Fairchild Sentry/Tektronix
Tester fnpul Formal
CALMA GOSII·E Beam Format
)MEBES]
DESIGN PHASE
FunciiOnal r esignfVeril ·cation
�
Logic Design Capture
+
Logic Design Verification
�
Circuit Design Verification
�
Electrical Con nectivity Checking/
Parasitic Capacitance Extraction lor Celts
•
Design Rule Checking for Celts
t
Electrical Connectivity Checking/
ParasitiC Capacitance Extraction for
Sub Chips and Full Chi j
Test Vector
'
Preparation
Mask Data Preparat1on
CAD TOOL USED
DECSIM-Behavior/AXE/HCORE
OUICKORAW
RSIM
)output panern comparison]
SPICE-GRAPES
CALMA GOSII
CALMA GOSII
IV/XREF
DRC
IVfXREF
Ad hoc Project Tools Plus OECSIM
MOP
Fig uTe 1
CAD To ols Used in tbe Design Phase
Digital Technical journal
No. 2 March 1 986
C hanges d u ri n g this phase are very expensive and the number was kept sma l l by having the design team submit only sized schematics ( i . e . , ones with transistor width and length specifica tions that were verified using the logic and circu it s i mu lators) to the layout design tea m .
4 . The mask data was not submitted to the mask shop (or even generated) until a11 sections on the whole chip were free of design-ru l e and e lectrical -connectivity errors .
The Value of the NMOS CAD Suite on the Micro VAX
II
Project
Figure 2 illustrates the entire CAD suite used on the 7 8 0 3 2 and 78 1 3 2 chip designs .
Use of the CHAS System
As mentioned earlier, the final use of the CHAS system was pared down considerably by the
M icroVAX project as compared with i ts use in the earlier project. The functions used most fre quently were
•
Schematic wirelisting
•
Layout format conversion
•
Copying files out of the CHAS database
•
Plotting
•
I nvoking the SPICE circuit simulator and the
GRAPES graphical post-processor
Behavioral Modeling and Simulation
A simulation system called DECSIM was used tO simulate the behavioral definition of the chip design .3·7•8 The DECSIM system works interac tive ly and was used to debug the high-level functional design . This system is very reliable and proved to be a vital i ngredient in achieving the high degree of accuracy of the microcode .
New Products
ROM/PLA
LAYOUT ASSEMBLY
�
SPICE
GRAPES
QUICKDRAW
H I ERARCH ICAL BEHAVIORAL/
LOGIC S I M U LATION a
GATE SIMULATION c::J
Figure 2 CAD Suite Used on the Micro VAX /1 VLSI Design
Digital Technical]ournal
No. 2 March 1986
5 1
Schematic Capture
A drawing system called QUICKDRAW was used as a schematic editor. QUICKDRAW's greatest assets were its architectural simplicity , reliabil ity, and ease of use. The system permitted sche matic entry on low-performance graphics ter minals (VT 1 2 5s) . Of course, keyboard entry is not always totally practical for bulk schematics e n t ry , or even good for s m a l l s c h e m a t i c changes . However , QUIC KDRAW cou l d be accessed from any terminal, was easy to learn i n a few hours, and could be used b y t h e whole chip team.
Logic Simulatio n
As in the earlier project, the MicroVAX team decided that they needed the accu racy of switch-level logic simulation . At this level of representati o n , the transistors are l i terally treated as "switches, " but with resistance and capacitance attributes. The models can also represent both bidirectionality and charge shar ing. At the time, the MOS (switch-level) capa bility of the DECSIM software was still matur ing; t herefore, the team decided to use a switch-level simulator called RSI M , developed at the Massachusetts Institute of Technology.
RSIM was sufficiently accurate to enable the complete design to be simulated at this level, although its timing aspects could not be used.
RSIM's usage , therefore, resembled that of a logic simulation syste m . The prime role of this stage of the process was to prove equivalence with the higher-level behavioral model, thus gaining functional completeness at a l ower, more accurate level of representation . That equivalence was achieved by supplying the same test vectors used in the behavioral phase to the RSIM runs.
Circuit Simulation
An industry-standard system , SPICE, was used for circuit simulation. SPICE was the most accu rate mechanism of its kind available for simu lat· ing the electrical performance of circuits on the chips. This simulator was used extensively for circui ts containing up to 1 000 transistors .
There were two major advantages of Digital's version of the SPICE system .
1 . The device models encoded into SPICE were a very accurate representation of the devices made i n Digital's NMOS pro cess . The device equations built i n to these models were derived in two ways: first, by extracting the operating charac t e r i s t i c s o f N M O S d e v i c e s f r o m fabricated test chips; and second, from the results of experi ments performed by another team at Digita l . That team cre ated models for devices and processes by using a battery of sophisticated simula tors , such as MINIMOS, SUPRE M , and
SEDAN .
2 . Throughout the pre- and post-processing stages, all voltage values over time from a SPICE run could be saved and later graphically analyzed by the designers in a proprietary graphical post-processing system called GRAPES. Using this system avoided having to make multiple runs of
SPICE and permitted much easier inter pretation of the o u t p u t waveforms .
Figure 3 is a sample circuit simulation waveform from the GRAPES system .
In terco nnect Verification and
Wirelist Extractio n
The IV syste m , partially proven on previous chip design projects , was a major boon to this design team . The system performed several functions .
1 . I t extracted a wirelist (in SPICE format) from the actual layout database .
2 . I t calculated the parasitic capacitances for devices and nodes and fed those into the extracted wirelist. That automatic input permitted the final simulations in
SPICE to be very accurate .
3 . I t detected any open and short circuits in the electrical network of the wirelist.
4 . I t compared the extracted wire list with the original wire list ( created via the schematics editor, QUICKDRAW) and reported any m ismatches in signal or node names, device sizes, and other ele ments .
This verification and extraction tool per formed all these fu nctions much faster and more accurately than any of the connectivity checkers or extractors that were available com· mercially. The IV system is generally recog nized as one of the best in the industry for this purpose .
5 2
Digital Tecbnical]ournal
No. 2 March 1986
F OO 5 5 1
1 - >
A : V ( J ) ? - >
R
: V (
J
)
3 - > ( : V ( 5 )
1 - >
1 . 2 5 . ������ � � � �� · � 1 - � 7 5 .��·• ·�- ���-- - 2 · 5 � ��---��--- - - J . I 2 5 � - -� - - - -- - 3 . 7 5
( V )
2 - > z . · - ·· - � � � · � ·�•• • • • • z . 3 1 � • •••••• • ••••· - - - ••z . 7 5 �-·�·�·�- · · •·-� 3 , 1 2 5 .� � - - � - -� 3 . 5 ( V )
( n s ) 3 - > p���A� � � �����A � � A � A � A t . 2 5 £ A���A£AA4£��4A�4A� 2 - � 4 4AAA .. � �� .. � ���A � . 7 5��A������A�
� - ���•'0
I � X I
.. A5 ( V )
I
1 . �-���
? . ��M�
3 . � ·� "0
•
•
••���
S . A��\'"
I b , 0M0P I
I
I
C
C
C
C
C
I
I
I
I
I
I
I
I
A B
X
7 , 00 0 U 0 I e . o��·� I
'
C I
I
A 8
Q .
0 0 0 0 0 I l � . � � 0 A
I C A I t - - - - - - - - - - - - - - - - - - - - - - - - • - - - - - - - B - - - - - - - - - - - - - - - - X • • • • • • • • • • • • • • • • • • • • • • • • + · · · · · · · · · - - - - - - - - - · - · - - -
•
� � � � I 8
1 2 . � 0 0 0 B
I
I
A
C
I
I
I
I
1 J . 0 0 0 0 8
1 0 , 0 0 0 �
X
A I
I C
C I
I
I
I
1 5 , 0 0 0 0 X
I b . A 0 0 0 X
1 7 . 0 0 � 0 B A
I
I
I
C c
C
I
I
I
I
I
I
1 8 , 0 0 � 0 B
1 q , 0 0 00
I 8
I
I
I
I
I
I
Figure 3 Sample Outp u t fro m the GRAPES System
The system has some u nique data structures and algorithms.
1 . It simplifies circuit extraction by con verting all shapes into trapezoids . These are very convenient representations that permi t IV to thoroughly analyze lateral node and vertical-device connections .
2 . It calculates the parasitic capaci tances for both area and periphery, taking into account cell-capacitance effects coming from ever-shrinking device geometries.
The system also calcu lates coupling capacitances .
3 . It performs very fast wirelist compari sons (layout to logical) , using a u nique graph-isomorphism algorithm that iso lates e rrors rath e r than propagating them.
System Verification
The fi nal syste m - l ev e l verifi c a t i o n of the
MicroVAX chips was performed using the AX E test-case generator in conjunction w i t h the
DECSIM behavioral models. In this way, test cases (which were in fact VAX macroinstruc tions generated by AXE) were passed to the sim u lation model for execution. The execution resu lts were then compared automatically with those obtained from running the same test cases on an operational VAX system . The MicroVAX team used AXE in a particularly novel way. Via
Digital's Ethernet network, they searched for in house VAX- 1 1 /780 systems with spare capacity on non-prime shifts . Then the team activated
AXE on those systems, which generated a tre mendous number of test case s . This same approach was used (and continues to be used today on subsequent projects) for running CPU intensive SPICE circuit simu lations on many processors in remote locations.
VLSI CAD Beyond the Micro VAX II
Project
Digita l ' s use o f t he NMOS VLSI CAD su ite reached a peak of maturity with the 7 8 0 3 2 and
7 8 1 3 2 projects . We have been able to make a major process-technology step to CMOS 1 with little cost by exploiting the same basic set of tools . That has enabled us to develop a whole new set of VLSI chip products in very quick successio n .
However, fol lowing Moore's Law, it is time to face the challenge of a two- to threefo l d increase i n complexity for the next generation of chip designs. This complexity means that design teams for new custom chips must be able to design parts with twice the transistor count as the 7 8 0 3 2 , yet take the same or less time to do it. Figure 4 illustrates the complex i ty that w i l l be experienced in future chip design projects .
Major productivity improvements in
CAD sys tems must be made to accomplish this doubling of the transistor coun t . Digita l 's VLSI CAD
Group is now making the following improve ments in its custom tool suite:
Digital Technical journal
No. 2 March 1986
5 3
New Products
The Evolution of the Custom CAD Suite Used o n the Micro VAX
II
System
1 280K
Ui
Ui c
.::
0 n
E
_
Q)
640K
�
320K
0
:::!.-
40K
1 25K
MicroVAX II chip
•
•
70K
(V1 1 IE chip)
•
35K
(BIIC chip)
20K
•
1 7K
(F1 1 chip)
•
1 0K
1 979 1 980 1 981 1 9821 983 1 984 1 985 1 9861 987 1 988
Year
Figure 4 Chip Complexity Projections
•
A new system for tool integration and data base management, cal led KATI E , is being developed to replace the CHAS system . The
KATIE system has a simpler, more modular
CAD kernel than has the CHAS system , but with much higher performance .
•
The DECSIM software is being improved to provide true mixed-mode modeling and sim u lation ( b e havioral -gate-switch) . I n i t i a l results i ndicate a doubling o f simulation pro ductivity, and our aim is to gai n equivalent performance in the separate switch and behavioral areas.
•
A variety of techniques is now providing up to ten times the performance of the tradi tional SPICE system for circuit simulation.
For example:
1 . An event-driven circuit simu lation sys tem called SAMSON,9 which exploits the temporal sparseness of d igital networks, has been developed. SAMSON offers from five to fifty times the performance of
SPICE for direct current and transient analyses.
2 . SPICE can be made to run much faster on vector processors and multiprocessors.
3 . A timing verification system called
TV can analyze critical paths at the rate of
I
0 0 0 tran s istors p e r m i n u te of CPU time. 1 0 TV performs within fifteen per cent of the accuracy of SPICE, but its speed is several orders of magnitude faster.
•
Schematic entry can be improved by running
QUICKDRAW on high-performance , high resolution graphics workstations . The system will support multiwindowing, menus , and pointing devices, as wel l as provide high performa nce wirelisting, with at least a doubling of speed over the version used on the 780 3 2 chip design .
•
H igh-resolution, VAX-based graphics work stations will also be used for custom layout editing, using the in-house developed editor,
MEGAN.
Summary and Conclusions
The MicroVAX I I project demonstrated a num ber of valuable lessons about CAD in general and VLSI CAD in particu lar.
1 . The second and subsequent projects that use a particular
CAD technology benefit enormously from the experience gained during the first use.
2 .
As a corollary to the point above, it is imperative that CAD tools and systems be built to endure at least two generations of projects. Otherwise , the cost and diffi culties of using these tools will far Out· weigh the benefits .
3 . The
CAD teams should use the period of stability during these later uses of the tools to develop the next generation of more powerful tools.
4 . Much conservatism exists in the IC industry around the need to archive com plete images of all too ls (layered prod ucts, operating systems, etc . ) used in the design of an I C , along with its final mask database . F u t u re c h i p teams p l a n tO m igrate their mask databases to contem porary CAD systems. This process will use the same exhaustive checks and tools used on the original design to ensure that the conversion is thorough . In this way, there will be no need to revert tO old copies of o u tdated systems and rools when making engineering change orders late in the product's life cycle .
54
Digital Technical jounull
No. 2 March
1986
5 . The close coupli ng between chip design teams and CAD developers is an invalua ble i ngredient in the successfu l comple tion of chip projects .
References
1 . W . N . Johnson, "A VLSI Superminicom puter CPU , " IEEE International Solid
State Circu its Conference Digest of
Technical Papers
( 1 984) : 1 74 - 1 7 5 .
2 . J . C . Mudge , C . Peters, and G . M . Taro l l i ,
"A VLSI C h i p Assembler, " in Desig n
Methodologies fo r VLSI Circu its, ed .
P G . Jespers ( Rockvi l l e : Sijthoff a nd
Noordhoff, 1 98 2 ) , 3 29 - 3 5 6 .
3 . A . F . H u tchings, R J . Bonneau , and W . M .
Fisher, " Integrated VLSI CAD Systems At
Digital Equipment Corporation , " Pro ceedings of the 22nd A CMjiEEE Design
A u to mation Co nference
( 1 985) : 5 4 3-
5 4 8 .
4 . C . Mead and L . Conway, Introduction To
VLSI Systems
(Readi ng: Addison-Wesley,
1 980) .
5 . SPICE was developed by Lawrence Nage l and Ellis Cohen of the Department of
Electrical Engineering and Computer Sci ences, University of California, Berkeley.
6. W.J . Herman and G . M . Taro l l i , " Hierar chical Circuit Extraction With Detailed
Parasitic Capacitance, " A CM IEEE 20th
Desig n A u to m a tion Co nference Pro ceedings
( 1 98 3 ) : 3 3 7- 3 4 5 .
7 . M.A. Kearney, " DECS I M : A M u lti-level
Simulation System For Digital Design, "
Proceedings of the ICCD Conference o n
Computer Desig n
( 1 984 ) : 206-209 .
8 . R . R. Rezac and LT. Smith, " Methodology for and Results from the Use of a Hard ware Logic Simu lation E ngine , " Proceed ings of the ICCD Co nference on Com puter Design
( 1 984) : 4 5 7-46 1 .
9 . K .A. Sakallah and S .W. Director, "SAM
SON :
An
Event Driven VLSI Circuit Simu lator," Proceedings of the Custom Inte grated Circu i ts Conferen ce
( 1 9 8 4 ) :
2 26-23 1 .
1 0 . N . P . )ouppi , "TV: An NMOS Timing Veri fi e r , " (Thes i s , Stanford U n ivers ity,
1 982) .
Other References
Panel Discussion , R .J . Camoin , Moder ator, " Central DA and its Rol e : An Execu tive View , " A CM IEEE 20th Design
A u to ma tio n Conference Proceedings
( 1 98 3 ) : 3 - 1 1 .
R . H . Katz, " Managing the Chip Design
Database , " IEEE Computer, vo! . 1 6 , no.
1 2 (December 1 98 3 ) : 26-3 5 .
W . M . v a n C l e e m p u t a n d H . O fe k ,
" Design Automation for Systems," IEEE
Co mpu ter, vol . 1 7 , no. 1 0 ( October
1 984) : 1 1 4 - 1 2 2 .
J .C. Foster, "A Unified CAD System for
E lectronic Design , " A CM IEEE 2 1 st
Design A u to matio n Conference Pro ceedings
( 1 984) : 3 6 5 - 3 6 9 .
K. Sherhart, M . Vershel, a n d
J. Owen,
"The Engineering Design E nvironment, "
A CM IEEE 2 1 st Design A u to mation
Conference Proceedings
( 1 9 84 ) : 4 66-
4 7 2 .
B .W. Lampson , " Hints for Computer
System Design , " IEEE Software, vol . 1 , no. 1 Qanuary 1 984) : 1 1 -2 8 .
New Products
Digital Technical journal
No. 2
March 1986
5 5
Rick Spitz
Peter George
Stephen Zalewski
I
The Making of a
Micro VAX Workstation
Developing a Micro VAX workstation required that graphics hardware and software be designed. The pmject team kept the hardware simple by using VAX instructions for most of the work. Extensive graphics software bridges the hardware and the graphics applications. The graphics and windowing software, UIS, is the key to that process. UJS supports trans parent multitasking with a distributed method for managing regions on the screen.
A
video device driver manages lists of region descriptors, keeping track of keyboard and mouse changes. The UIS system normally executes in user mode, thus minimizing overhead and utilizing the full performance of the VMS system.
When Digital decided to develop the MicroVAX series, we a lso began to consider how to build them into a family of low-cost VAX engineering workstations. Experience with the VAXstation
1 00 provided us with a great deal of knowledge related to workstation requ irements. However, its architecture required extensive grap hics hardware . This architectural approach was not considered viable for a low-cost, high-volume engineering workstation intended for a single user. Another approach placing greater empha sis on software was i l lustrated by Xerox's Star workstations, which were in use within Digita l .
We decided that combining t h e MicroVAX processor with a low-cost graphics controller, the VMS operating syste m , and a good human interface would resul t in a powerfu l worksta t i o n . The VAX/VMS e nv i r o n m e n t a l ready allowed any VMS appl ication program to ru n on every member of the VAX family. The Micro VAX system woul d extend the family to include lower-cost VAX systems . A MicroVAX worksta tion, in addition to running all existing VMS software, would now provide a base for graph ics applications .
I n the spring of 1 98 3 , a joint task force of hardware and software e ngineers was formed to determine how this workstation shou l d be built. Our strategy was to design a product based on the MicroVAX I system and evolve it to a mature workstation using the MicroVAX I I system .
The task force 's objective was to set the over a l l goals of the project and tO make sure that the graphics hardware and software were wel l inte grated . Guided by a strong focus on time to market , the workstation hardware group had the responsibi lity of building an i n itial graph ics controller. They were also chartered to initiate design work on future hardware graphics con trollers with more features and h igher perform ance . The VMS software group took on the role of developing the sofn:vare components . This paper is written by members of the VMS Devel opment Group; therefore , its primary emphasis is on the software aspects of this proj ect .
Our first task was to make sure that the graph i cs hardware being defined was s u i table for effi cient use by the software . Having l i m ited expe rience with low-cost graphics contro llers and workstations, we proposed a strategy of using a very basic Q-bus controller and doing most of the work with VAX instructions . This approach was viable because the VAX instruction set is rich and versatile in the area of character and bit manipu lation . It also mini mized the risk i n developing hardware and provided maximum flexibi lity for the graphics capab ilities . With greater freedom in the software design, we could ga in experience and provide better d i rec-
56
Digital Technical journal
No. 2 March 1986
tion for hardware features needed in future graphics controllers .
Since no MicroVAX CPU had yet been devel oped, we built a breadboard hardware configur ation to do hardware and software evaluations .
MicroVAX systems execute a subset of the fu ll
VAX i nstruction set i n hardware ; however, software emulation of the other instructions allows all VAX software to run transparently.
For cost and space reasons, MicroVAX systems were targeted to use the Q-bus for ljO , while most existing VAX systems used the UNIBUS for most peripherals .
The breadboard configuration consisted of a
VAX- 1 1 / 7 5 0 system with a UNIBUS-to-Q-bus adapter. We obtained some experimental Q-bus graphics controllers used in the development of the graphics interface for the PR0350 hard ware . Using this configuration, we evaluated the performance of text and graphics by imple menting a number of software algorithms .
1
This technique treated display memory as standard
VAX program memory, and VAX character and bit instructions were used to generate text and graphics . Evaluation of our results showed that this approach was reasonable and the basic per formance was acce p tabl e ; however , some assists were stil l needed in hardware.
The VCBOl Hardware Graphics
Controller
Taking our resu l ts back to the the joint task force, we settled , after several iterations , on a hardware design . The hardware graphics con troller was named the VCBO 1 , known i n ternally as the Q-bus video subsystem , or QVSS. Due to space and power constraints in MicroVAX packages, the controller had to fit on a single-quad
Q-bus modu le. It contained 2 5 6K of bitmap memory that was ful ly addressable by any VAX instruction. That amount of memory was more than was needed to fill a full-screen v ideo mon itor. The extra memory would al low software gra p h i cs ro u ti n e s to operate d i re ctl y o n occluded areas o f windows in the video display memory.
Based on inputs from the software eval uation, the hardware would also contain a scan-line map to allow mapping any scan l i ne in d isplay memory onto the physical scree n . This tech nique allows much better scroll ing perform ance, facilitates the management of occ.luded window areas, and allows the s i m u l taneous s u p p o rt of d i ffe r e n t w i n d o w i ng sys t e m s .
A 1 6 X 1 6-pixel cursor plane , a separate hard ware c o m p o ne n t , g r e a t l y s i m p l i fi e d t h e software logic required to manage the mouse cursor. The pattern is programmable to allow dynamic changes to the cursor pattern , depend i ng on its screen location and the state of the workstation. In addition, a mouse interface and dual UART are provided to connect to a mouse , a keyboard and an optional tablet. The inherent simplicity of the hardware allowed the hard ware team to produ ce the first protOtype by the early summer of 1 98 3 .
Figure 1 shows a block diagram of the VCBO 1 configuration.
Software A rchitecture
The software team was chartered to develop a general software workstation architecture . Our goal was to al low the evo lution of future
Mi croVAX workstations that would address
2048
L I N ES
BITMAP
SCAN MAP
1 024
ENTRIES
1 6
16 BITS
CURSOR
1 024
BITS
Fig ure 1 Block Diagram of the VCBO 1
PHYSICAL SCREEN
864
LINES
960
BITS
New Products
Digital Technical journal
No. 2
March 1986
57
cost-sensltlve markets with basic, inexpensive hardware . We a lso wanted to improve perform ance and take advantage of features to be pro vided by more-intelligent hardware graphics controllers in the future .
Our performance evaluation of the VA.Xsta tion 1 00 architecture poin ted out that the cen tral dispatcher needed to manage the window i ng activities on the physical screen was a real bottleneck . Therefore , we elected to pursue an approach that used a distributed method to manage regions on the physical scree n . In most cases this approach wou ld a l low an individual job, called a process in the VMS system, to oper ate directly on bitmap memory. There is much less overhead than context switching between processes, as requ ired in a centralized screen manager design .
The software architecture that we defined was implemented by a loadable set of VMS sys tem services know as the User Interface Ser vices , or UIS 2 UIS provides fundamental graph ics services and d i s p l ay l i s t capa b i l it i e s .
App l i cation programs , h i g h - l evel gra phics packages, and VMS's VT 1 00 and TEK4 0 1 4 emu lation drivers a l l utilize UIS to construct indi vidual windows, as wel l as for text and graphics functions.
3
A VCBO 1 device driver is used to manage the physical hardware 4 The driver is responsible for controlling the keyboard , the mouse (pointer) , and the scan-line map .
VCBO 1 Video Device Driver
The video device-driver software has one pri mary fu n c t io n : to ma nage l ists of region descriptors . I n particular, it keeps three main lists; one each for keyboard input, button tran sitions, and pointer (mouse) movement .
To be notified about a particular event, an application p rogram posts a request to the driver. The request specifies the type of event desired and the region on the scree n . The driver then places this request on the appropriate list.
For example, if pointer movement requests are active and mouse movement occurs, the driver will search the l is t for the entry that has speci fied a region that the pointer is currently within. The driver then notifies the application that was the last one to specify this area. The notification mechanism used is a software inter rupt, known in the VMS system as an asynchro nous system trap . This trap interrupts the flow of the specified user process and invokes a userdefined action routine. This technique provides a low-cos t , res ponsive notification to the application .
The keyboard is connected to the device driver by a dual UART on the video controller. A hardware interrupt is delivered to the driver each time a key is pressed . The driver then searches the keyboard l ist and delivers the char acter to the process associated wit h the top entry on the l ist. All keys are "soft , " which means that any key on the main keypad can be defined as any of the possible ASC I I character codes . I t is also poss i b l e to defi ne m u l ticharacter sequences for a given key. The sec ond half of the dual UART is used to support a bit tablet or a serial mouse . These devices need to send several bytes of data for each pointer or button transition . The driver buffers this data u ntil it receives enough to decode an event .
Then it searches the appropriate event list and, if necessary, del ivers a software interrupt to the application .
The driver su pports the capability to specify cursor patterns for a region. When cursor move ment is detected , the driver searches a list to determine what the cursor pattern should be for the current location of the pointing device .
Once located , the pattern is l oaded into the hardware . The video controller hardware then superimposes the pattern onto the appropriate screen area by merging the pattern with the video signal from the bitmap memory. This pro cedu re e l i m i nates the need for a save-and restore operation in the physical bitmap each time the cursor moves or a write to bitmap memory occurs . The hardware also has the abil i ty to specify two logical operations, NAND and
XOR, on the cursor pattern . This abil ity pre vents a white cursor from being lost on a white screen , or a b lack cursor on a black screen. The driver tests the physical bitmap location that is overlaid by the cursor to determine which logi cal operation shou ld be used to maxim ize the cursor's visibility.
A p r o p o rt i o n a l - a c c e l e r a t i o n m ove m e n t a lgorithm i s used t o minimize the desktop area required for a mouse pointer. The driver accel erates the cursor's movement if the mouse 's rate of movement exceeds any of a series of thresholds in a given screen refresh interva l . If no acce leration were to occur, it woul d take a desktop space of approximately
1 3 by 1 1 inches to move the mouse both horizontally
5 8
Digital Tecbnicaijournal
No.
2 March 1986
and vertica lly respectively across the screen .
With acceleration, a mouse movement of only 2 inches is needed to move across . The accelera tion values used are as fol lows : 1 to 2 pixels of l i near mouse movement per screen refresh interva l , no acceleration needed; 3 to 4 pixels, accelerate by a factor of 2; 5 to 8 pixels, accel erate by a factor of 4; greater than 8 pixels, accelerate by a factor of 6 .
The driver provides an optional console win dow to a l l ow system-level debugging. The
MicroVAX CPU can communicate directly with the video control ler during booting and debug ging. If this feature is enabled , the top 2 4 0 scan lines of video memory will be allocated for the console window. When the CPU wants to com municate with the console, the VMS console driver will map directly to those 2 4 0 scan l i nes.
Thus, the console driver emu lates a "dumb" terminal in this region . When a function key is pressed on the keyboard, the video driver will map this special console memory onto the top
2 4 0 entries of the physical scan-line map, and the operator console wi ll appear. When the key toggles again, the top 2 4 0 entries of the scan line map will be restored.
UIS Graphics and Windowing
Software
The decision to use simple hardware meant that software had to be developed to bridge the gap between that hardware and the app lications .
This software was of c r i t i c a l i m po rtance because the hardware designers assu med that a software layer would be needed to su pport even the most basic graphics functions .
Early in the design process, we decided that this software would provide more than just basic IjO support through the video controller.
Like the VMS operating system i t was built on, the worksta t i o n gra p h i cs and wi ndowing software , UIS, wou ld support transparent mul titasking. That meant being able to handle simultaneous demands by m u l tiple indepen dent appl ications on the shared VCBO 1 hard ware resou rce s . Therefore , U I S shoul.d be designed to provide two capabilities . First, i t should have a library o f genera l -purpose proce dures that app l i cations could use to easily access the hardware resources . Second , U I S should contain transparent management and synchronization mechanisms . I n that way, independent applications cou ld share both screen space and the use of the system's input devices. This design wou ld also a l l ow the development of UIS appl ication programs on any VAX system, whether it was a workstation or not.
For the initial release of the MicroVMS work station on the VAXstation I , these objectives were broken down i nto the fol l owing specific design goals :
•
Provide routines for creating and manipulat ing viewports on the video display.
•
Su pport multiple overlappi ng viewports and manage viewport occlusion transparently for applications .
•
Allow simultaneous graphics operations into all viewports .
•
Prov i d e a u s e r i n te rfa ce fo r viewport manipu lations .
•
Provide ro u t i nes for crea t i ng gra p hi cs objects.
•
Provide d isplay- li s t backup for graph ics operations so that applications can easi ly perform operations like " pan" and "zoom . "
•
Support shared access t o the mouse and key board and provide routines to notify applica tions of i np u t even ts occurring on these devices .
The fo llowing sections describe the architec ture of UIS and the mechanisms that were used to rea l ize these goals. Figure 2 is a block dia gram show i ng the functions of UIS.
Virtual Displays
The fundamental presentation object manipu lated by applications to construct images is the virtual display. All UIS output functions are per formed within a virtual display.
The coord i nate system of a virtual disp lay is defined in "world coordinates . " The world coordinate system uses the coordinate system of an application as a means of expressing d isplay locations . For exa m p l e , an application that draws a graph s howing popu lation growth versus time may find it conven i e n t to use
"Time" and " Nu mber of People" as x and y coordinates . The range of world-coordi nate val ues is specified to t he graphics subsystem when the virtual d isplay is created . The coord i nates are specified as signed F-floating VAX data types
New Products
Digital Technical Journal
No. 2 March 1 986
59
The Making of a Micro VAX Workstatio n
U S E R A P P L ICATION P R O G R A M
U I S S H A R E A B L E I M A G E
I
UIS
G R A P H I CS/W I N DOWI N G
SYSTEM SERVICES
(BINARY E
�
C O D I N G )
I
E N C O D I N G D I S PATC H E R f----
-
G R A P H ICS EXECUTION R O U TI N ES
&
R O U T I N E S TO U P DATE D I S PLAY LIST
I
BITMAP
G R A P H I CS R O U T I N ES
(GER)
I
D I S PLAY VI EWPORT
S E R V I C ES
(VPS)
I I
D I S P LAY M E M O R Y (VCB01 O R VAX) r---
'l
V C B 0 1
DEVICE
D R I V E R
Figure 2 VIS Fu nctional Block Diagram for reasons of precision and ease of calcu lation in high-level languages .
A display l ist is an encoding of the exact con tents of a virtual display, independent of the device . Display lists are maintained and used by
UIS to achieve the fo llowing short- and long term goals:
•
Al low the automatic management of pan n i ng, zooming, resizing, and duplicating dis play windows
•
Al low high-resolution printing of virtual displays
•
Allow the structuring and manipu lation of virtual-display objects
•
Allow an application to select an arbitrary output from a virtual display, give it to an
" i ntelligent" cooperat i ng appl icat i o n , or simply store it in a file as generic encod ing, and then later replay the generic encoding into a new virtual display
Display lists consist o f the fol l owing basic objects :
•
Output primitives
•
Attribute primitives
•
Structural primitives
Output primitives map directly onto the U l S output operations ( e . g . , plot some l ines , write a text, draw a circle) and the modifications that they make to a virtual display.
Attri bute primitives change the current value of an attri bute in an attribute block i n order to affect subsequent output primitives . Attribute blocks are used by UIS to specify a set of attri bute values for a l l UIS graphics objects ( l ines , text. circles) . Typical attribu tes include the writing mode ( replace , complement, erase) , line style (sol i d , dashed) , and font to use when writing text.
There may be up to 2 5 6 attribute blocks a d d ress a b l e a t o n e ti m e . Att r i b u te b l o c k
6 0
Digital Technical journal
No. 2 March I
<JB(,
nu mbers are used and assigned only by the application, except for attribute block 0. This block is a special one that cannot be modified.
It provides a set of attributes used as a standard default for text and graph ics. Block 0 also pro vides a template for creating alternate attribute blocks .
Structural prim itives al low the hierarchical grouping of attribute and output pri mitives into graph ical begin and end blocks, called seg ments . Segments al low applications to have access to many more than 2 5 6 attribute blocks.
While segments inherit current attribute blocks from higher- level segments , modifications to attribute blocks from within a segment cause local copies of the modified attribute blocks to be created. For example, if a particu lar attri bute block is referenced within a segment, then that segment is first searched for the b lock. If the block isn ' t found, the search is made in successive outer segments.
The coord inate system, called normalized coordinates, is used both within the d isplay l ist and when creating generic encoding. Normal ized coordinates are used to defer the mapping of a set of world coord inates to specific device coord inates until the actual output device is known. As described in the fol lowing section, this mapping tO the physical device does not occur u ntil a display viewport is create d . This delay is important since output devices have different resolu tions. For example , printers typ ically have much higher resolutions than video moni tors.
Since floating point calcu lations are typically slower than integer ones , normal ized coordi n a t e s a r e e x p r e s s e d i n u n i t s c a l l e d
"Gutenbergs , " which are stored as 3 2 -bit inte gers . A Gutenberg, the same u n i t used in UIS font definitions, is defined to be 1 /7200 inch
( . 0 1 points) Their use as normal ized coordi nates is we ll suited because they minimize the nu mber of coordinate transformations that must be performed when writing text. Gutenbergs have the desirabl e characteristics of being both reasonably smal l-and therefore amenable to good graphics resolution-and very efficient for text operations .
The conversion between world and normal ized coord inates is based on the desired physi cal size and world-coordinate size of the virtual display as specified by the appl ication . When a virtual d isplay is create d , the appl ication expresses the desired size of the virtual display in both p hysical and virtual uni ts. That estab l ishes the relationship between the physical size of the fonts and the arbitrary size of a vir tual display's world-coordinate system.
Display Windo ws and Viewpo rts
A display window is the object used by app lica tions to control how much of a virtual disp lay i s avai lable for viewing by the user. This control is accomplished by defining a rectangle speci fying the v i ewable port ion of the virtua l display.
A display viewport is the area of the physical screen into which a display wi ndow is mapped.
Display viewports vary in size and may be placed anywhere in the physical screen area .
Display viewports a lways occlude when they o v e r l a p . The order of o c c l u s i o n u s u a l l y depends o n the order i n which t h e display viewports were created . However, the order may be a ltered by t he user through the UIS user interface or by applications using the U I S windowing services.
A display window is created , mapped, and automatically sca led to a display viewport when the application makes a singl e , routine call tO UIS. Note that at the time of the cal l , the output of the UIS app li cation is directed to a specific physical output device , usually the screen. Scal ing can be avoided if the applica tion directs UIS to use the physical size sup plied by the application when the virtua l dis play was created . That al lows text and graphics to appear in exactly the size and aspect ratio that an application considers ideal .
The a m o u n t a n d size of the i mage t hat appears in a display viewport can be controJled by altering the size and position of the display window or the size of the display viewport. The image can be managed by either the applica tion, through UIS, or the user, through the user interface functions. The fo llowing rules govern the image:
•
To magnify the image , either the size of the window is decreased without a ltering the viewport , or the size of the viewport is increased without altering the window .
•
To reduce the image , either the size of the wi ndow is increased without altering the viewport , or the size of the viewport is decreased without altering the window .
New Products
Digital Technical joun1al
No. 2 March 1 986
6 1
The Making of a Micro VA X Workstation
•
To change the amount of the virtual display being viewed without scaling, both the win dow and the viewport size are expanded or contracted by the same amou nt.
•
To pan the image , the window around the virtual display is moved without altering the viewport size or location.
Figure 3 i l lustrates the mapping that takes place when going directly from a virtual display to a physical display. The left column shows the transformations between the coordinate spaces.
The two colu mns on the right show the way the virtua l d isplay is scaled to the final output device .
Virtual Keyboards
App l ications use a concept ca lled virtual key boards to share and individua l ly manipulate the physical workstation keyboard . Virtual key boards allow an application to get input from the physical keyboard and to modify its charac teristics, both in a synchronized manner. I nput can be received in either of two forms. First, applications can specify that they be del ivered a software interrupt whenever keyboard input occurs . Second , they can periodically poll the v i rt u a l keyboard to see if new i n p u t has occurred . Certain characteristics can be man aged for each virtual keyboard , such as keyc lick volumes and keyboard key mappings.
The connection between the physical key board and the various virtual keyboards availa ble on the workstation is generally managed by the user. An application could force the physi cal keyboard to be bound to a virtual keyboard.
Typica lly, however, the application will associ are the keyboard with some d isplay viewport and al low the user to manage that connection through the user interface.
Mo use Input
Applications can both solicit and manage input from a mouse with respect to rectangles within display viewpons . To do that, an app lication must specify a world-coordi nate rectangle and
6 2
(WORLD COO R D I N ATES)
(
D I S PLAY LIST E N C O D I N G I N
N O R M A L I Z E D COO R D I NATES
\
OUTPUT
P R I M IT I V E
EXECUTION
ROUTI N E S
I
(DEVICE SPECIFIC COO R D I N ATES)
WO RLD COORDI NATES
(WO R L D COO R D I NATES)
I
( S I Z E D TO . . . )
I
V I R TUAL DIS PLAY
( C L I P P E D TO .. )
I
I
D I S PLAY W I N DOW
I
( S I Z E D TO .. )
I
D I S PLAY V I EWPORT
�
(WRITTEN TO .. )
(
7
�
PHYSICAL DISPLAY
VI RTUAL DISPLAY
D I S P LAY W I N DOW
Dl
7
V I EWPORT
EN TO . . . )
Figure
3
Mapping from Virtual- to-Physical Display
Digital Technical]ountaJ
No. 2 March 1986
the disp lay viewport to which the rectangle applies. The application then directs the UIS to
•
C hange the cursor pattern or position when the cursor moves within the rectangle
•
Send a software interrupt whenever the cur sor moves within or out of the rectangle
•
Send a software interrupt whenever a mouse button is depressed or released within the rectangle
App lications can also check the cu rre nt mouse position or button state at any time .
Implementation Details
UIS was designed with two primary implemen tation goals in mind. Of course, the first goal was to implement the architecture described in the previous sections. just as important was the belief that the cost of using UIS had to be as small as possible. The overhead associated with a routine call had to be m i nimized , and the algorithms and architecture employed by UJS had tO be as efficient as possible. UIS also had to be fast because the simple graphics hardware relied upon UIS software to take the place of sophisticated grap hics hardware . To meet these goals, t h e software team made some basic design decisions right at the start. The effect of these decisions on how the design operates are discussed in the following section .
UIS operates in the caller's mode (usually user mode) because the cost involved in c hang ing tO kernel mode wou ld be prohi b i tive .
Because UIS operates in user mode, a l l data structures used by U I S are given user-write pro tection . This d es i g n d e c ision means t h a t timesharing use of the graphics package is pos sible, but without any security considerations .
Most of the UIS code resides in system space, and UIS rou tines exist as system services within the VMS operating system . That gives UIS all the desirable performance characteristics of oper ating system code ( i . e . , minimal image activa tion cost, max i m u m s hareabil ity , separately managed paging, etc . ) .
Fonts are stored in files and treated as system resources. Since several applications are likely to use the same fonts at the same time, UIS font management was designed to optimize font sharing. Fonts currently in use are kept i n a font pool in system memory. Upon beginning a text drawing operation, a process accesses the sys tem font pool to find the required font. If not found in the poo l , a font can be loaded into the font pool by searching the disk for the proper font file and then reading it into system mem ory. Similarly, fonts can be removed from the font pool because they can always be retrieved from disk .
Each virtual display is managed by only one process . That synchronizes the access to virtual displays and disp lay l ists and minimizes the effect that graphics applications have on each other. I f a second process wants to manipu late the virtual display of another process, then the applications running in the two processes must communicate . The process that created the vir tual display must then make modifications to it.
This concept is enforced by the fact that the contexts for a l l virtual d isplays reside in pro cess address space.
Data structures for display viewports, on the other han d , are kept in system space. That a.llows a process to change the topology of the viewpons on the video display. For example, a viewport bound to a display window that i t owns c a n be " popped" w i t h o u t having to notify every other process of the necessary screen changes. The storage for viewport data structures is allocated from paged pool. How ever, the storage protection must be changed to user write to allow access by the process-based graphics routines.
Access tO those data structures by UIS rou tines i s synchronized using t h e VMS lock man ager. Multiple processes are granted shared readjwrite access to the physica l display as long as they are simply reading from or writing tO their own viewports. If a process needs to change the relationships between the display view ports on the screen ( e . g . , create a new viewport or pop an existing viewport) , it must request exclusive readjwrite access to the phys ical display. Thus, no sync hronization overhead is incu rred in the steady state.
Figure 4 depicts the basic use of storage by
UIS.
As s hown in Figure 4, UIS software is organ ized into five basic parts.
The first piece of U I S that app! i'cations encounter is the UIS shareable image . UIS rou tines a r e accessed b y applica tions through transfer vectors in a VMS-protected shareable image . That allows UIS code to increase in size and tO change location within the operating sys tem without affecting the appl ications that use the code . Also , UIS application development can occur on machines where UIS has not been
Digital Technical journal
No. 2
March 1 986
6 3
New Products
The Making of a Micro VAX Workstation
SYSTEM
SPACE
P R O C ESS ( P 1 )
SPACE
PROCESS (PO)
SPACE
•
•
M A P P E D VCB01 M EM O R Y
I N C L U D I N G P H Y S I C A L B I T M A P
•
•
P A G E D POOL
•
•
•
BACKUP VIEWPORT B I TM A P S
VIEWPORT D A T A STRUCTU R E S
FONTS
•
•
•
•
PROCESS- P E R M A N E N T
D I S PLAY CONTEXT
(DELETED AT PROCESS R U N DOWN)
•
•
•
•
N O N - P R O C E SS - P E R M A N E N T
D I S PLAY CONTEXT
( D E LETE D AT I MA G E R U N DOWN)
•
•
Figure
4
VIS Storage installed. The UIS shareable image can be used to resolve UIS references at link and image acti vation time, even if the UIS system services are not present on the system . Finally, because the shareable image is protected, UIS can get con tro l during image rundown and perform some necessary clean-up activities .
The shareable image performs the requested operation by cal l i ng the :�ppropriate UIS system service . At this point, user requests are trans lated into calls tO internal UIS routines, and the relevant internal data structures are located . For exampl e , for a typical keyboard operation , UIS would locate the right virtual keyboard and make the appropriate calls to the VCBO
1 device driver.
For a typical output operation, such as draw i ng a l ine, UIS first creates a display l ist entry.
UIS then calls the display list management rou tines to update the display list and a l l windows i nto the virtual display. These routines, in turn , wil l check with the viewport service routines
(VPS) to find the right area of the physical screen in which to draw. Finally, the manage ment rou tines direct the bitmap graphics exe cution rou ti nes (GER) to draw to those areas .
VPS is more than a simple screen rectangle manager. Its tasks are
•
To present the rest of UIS with the " i l lusion" that viewports are always u noccluded and are contiguous pieces of hardware video controller memory
•
To take advantage of VCBO 1 scan-line scroll ing whenever possible
•
To provide bitmap backup for occluded win dows so that app lications are free from the complex ities of occl usion management
VPS does this by judiciously using and mixing three different types of video memory : on screen VCBO 1 memory, off-screen VCBO 1 mem ory, and off-screen VAX m e mory . VPS a lso manipu lares the entries in the VCBO
1 video scan-line map to present UIS with a virtual scan line map, or virtual viewport, for each physica l display viewport .
If the physical display has only one viewport,
VPS will simply allocate a set of p hysical VCBO
1 scan lines and set up the viewport data struc tures to direct GER tO that set. In this case, the physical and virtual viewports will be the same.
H o w e v e r , if t h e d i s p l ay h a s oc c l u d i n g viewports , VPS w i l l create a virtual viewport i n off-screen memory for each physical viewport .
The n , at 8 0 - m i .l l isecond i nterva l s , VPS wi l l copy the mod ified contents o f the virtual viewports to the physical viewports .
I f changes must be made to the VCBO
1 video scan- l ine m a p , then VPS wil l update the m .
These changes could be caused b y ei ther a viewport that needs to be hardware scrolled or a change in the layout of the viewports on the p hysica l scree n . VPS then merges all the virtual scan-line maps and requests an update of the physical scan-line map. Those actions are done in synchronization with the 60-Hz video verti cal -retrace interva l .
64
Digital Technical ]out·nal
No.
2
March l 'JB(i
Summary
Our initial goals were to design a workstation prod u ct with t he MicroVAX I system, thus pro viding a stabl e , mature product available for the
MicroVAX II syste m . The joint engineering task force was i n itiated in the spring of 1 98 3 ; proto type graphics hardware was available in the early summer. Once that preliminary hardware was ready, the VMS team entered into fu ll -scale development. The VAXjVMS workstation (VWS) product was developed during the fal l and win ter of 1 983, and into the spring of 1 984 . VWS underwent customer field test with the VCBO 1 graphics controller, the MicroVMS system , and the MicroVAX I system in the summer and early fa l l of 1 9 8 4 . T h e f i r s t r e l e a s e of t h e
VA.Xstation I was available i n late 1 984 . This i n i t i a l pro d u ct a l l owed t h i r d - p a rty VAX software vendors to take advantage of the VWS architecture .
La t e r , t h e VAX s t a t i o n I I rep l a ced t h e
MicroVAX I CPU with a MicroVAX I I engine, thus gaining much higher performance . The
MicroVAX I I processor entered customer field test in the early spring of 1 98 5 , with shipments to customers by early s u mmer. A new VWS s o ftware r e l e a s e t h a t s u p p o r t e d t h e
VA.Xstation I I was made available short ly after wards . That VMS software was the fu lfillment of this project's long-term goal .
Acknowledgements
We woul d like to acknowledge the contribu tion made by Dick Hustvedt to the MicroVAX workstation effort. Dick was instru mental in spearhead ing this u ndertaking. The contribu tions of Cathy Learoyd , Tom Furlong, Rob Scott, john DiMack, Mike Rosenblu m , jake Vannoy, and the rest of the VMS workstation team were also invaluable.
References
1 . J . D . Foley and A. van Dam , Fundamen tals of Interactive Computer Graphics
(Reading: Addison-Wesley, 1 9 8 2 ) .
2 . Micro VMS Workstation Graphics Pro g ra m m ing Guide
( Maynard : D i g ital
Equipment Corporation , Order No. AA
G 1 1 0B-TN, 1 9 85) .
3 . Micro VMS Workstation User's Guide
( Maynard : Digital Equ ipment Corpora tion, Order No. AA-EZ24 C-TN , 1 9 8 5 ) .
4 . Micro VMS Workstation Video Device
Driver Manual
(Maynard: Digital Equip ment Corporation, Order No . AA-DY65C
TE , 1 9 85) .
New Products
Digital Technical journal
No. 2 March 1986
6 5
Nicholas A. Warchol
Stephen F. Shirron
I
The RQDX3
Design Project
The RQDX3 is a Winchester and floppy disk controller aimed specifically for use
on
Micro
VAX II
systems. The designers foUowed a top-down development process to meet their goals. Trade-o.ffs, some requiring hardware and firmware to be built and tested for reliability, were identi fied and evaluated early in the project. The RQDX3
bas
a three-port data buffer to smooth data transfers between the host processor, the control ler's microprocessor, and the disks. Four internal subsystems work in parallel to allow maximum system performance.
Design Goals
The project team set a nu mber of specifi c goals at the start of the RQDX3 design. The greatest need was to improve the performance of the
MicroVAX I I system over that available with existing controllers, yet greatly reduce the man ufacturing costs of the disk subsystem. The fol lowing list contains the goals that governed the design of the module:
•
Cost-Obtain a manufacturing cost less than half of the best current disk controller. t he
RQDX2 .
•
Performance-The contro l l e r s hou ld n o t force an interleave o f data sectors on t h e sur face of the hard disk drives or limit the per· formance of the Winchester disk drives. The controller should also avoid wasting system bus bandwidth on the Q-bus . The controller architecture had therefore to be chosen to al low the highest performance possible while meeting the other design goals.
•
Dual Mod u le-The controller shou ld be designed so that it will fit on one Q-bus dual module. This form factor will allow the most flexible system configurations .
•
Schedule-First customer shipment would be approximately one yea r from the project start . Meeting this goal would a!low the phase-out of the higher cost and lower per formance RQDX l and RQDX2 modules .
•
Testable Design-A high percentage of this module would be testable by providing extra hardware , m icroprocessor code, and test strategies . This design would help to reduce both manufactu ri ng and maintenance costs .
The Design Philosophy
The team members decided that a top-down approach to the problem was the only way that the design goals could be met. A wel l struc tured , well docu mented design would allow the maximum communication between team members, and it would allow trade-offs to be made early in the design cycle.
The design process u sed in the project adhered to the following form :
•
Set the goals and assign priorities to deter· mine how flexible each one i s ; that will al low tradeoffs to be made if a goal is not attainable .
•
ColleCt and study any overall system specifi cations and requirements that apply. This is the time to write the prel iminary engineer ing specifi cation and define the i nterfaces
(both hardware and software) that must be adhered to. Any impu lse to go back and change these specifications shou ld be vehe· mently resisted.
66
Digital Technical Journal
No. 2 March I 986
•
Analyze the problem and determine the sys tem architecture based on the flow of infor mation and the complexity of the req uired control functions. If the problem appears too large or is not easy to document or describe , then it should be divided into smaller, more manageable functions. During this phase , operational descriptions are created . Those can be flow diagrams, timing diagrams, state transition diagrams, or anything that w i l l help t o explain how the controller should work. These descriptions should be included as part of the documentation package .
•
Look for the solution to each problem while weighing it against the design goals. Itera tions between this step and the previous one can be expected in order to meet the goals .
This part of the process involves looking at the available technologies and other designs to determine what is or is not usable. If other designs have followed the same documenta tion strategy, then this task is much easier; if they have not, then do not waste too much time tryi ng to " reverse engineer" those designs. The risk of using new technologies must be assessed tO determine what impact they would have on the design's cost and schedule .
The hardware design is documented using drawings called functional partitions . These drawings are a hierarchy showing the inter connection of fu nct ional , not physical , p ieces of the design. All datapaths and con trol signals are named at this time. The draw i ngs will be the reference point of the design team and make up a major portion of the design package . Because of the functional nature of these drawings, simulation of the design can be accomplished in a structured form.
At this time, a technical description docu ment is written to allow others outside the design team to u nderstand the operation of the design . This document is especially use ful in training new groups about the design as it progresses from the design phase to the manufacturing phase .
•
" Paper debug" the design . This is an in depth review by the design team before any hardware is built. The process begins with the operational descriptions and fol lows the documentation hierarchy down to the lowest level of the design. Normal operations and error conditions are checked , and each ele ment is analyzed for test and diagnostic coverage .
M istakes found at this stage are much easier to fix on paper than in circuit boards, gate arrays, or software debugging.
•
Build a prototype . This process includes the drawing of schematics to show the i ntercon nection of the physical pieces , the layout of circuit boards, the development of gate arrays , and the writing of software routines that interface to the hardware.
•
Debug the protOtype. If the paper debug was done correctly, this stage shou ld not uncover any disasters . The individual fu nctional p i eces of the design can be tested and checked off using the functional partitions as a guide . That systematic method will ensure that the entire design is tested.
The design process is the solution to a mul tidi mensional p roblem . Therefore , there is probably more than one design that will meet the goals. There is also the probability that it may be impossible to meet all the goals. In this case , some compromise in the goals must be made in order to make a solution possible.
This design problem is like those encoun tered in most other designs: Make it fast, cheap, smal l , reliable, and don't take too much time .
With each goal being constrained by others, the need for a structured method of finding a solu tion becomes more i mportant. The way to solve a set of simultaneous equations is not to try a solution and see if it fits , but to use some proven techniques tO determine the correct solution . Dividing the overal l problem i nto smaller ones and then determining a solution is probably the most powerful technique that can be applied .
Design Implementation and Testing
A ttacking the Goals
Each goal placed some unique restrictions on the design . Thus, it was important to u nder stand the effect of each goal and how flexible the achievement of that goal was. By keeping a constam watch on how the goals were being met, trade-offs cou ld be made very quickly .
Digital Technical journal
No. 2 March 1 986
67
New Products
The RQDX3 Design Project
The following discussion details each goal and how it was handled:
•
Cost-This was the original goal that caused the creation of the RQDX3 project. The cost/performance relationship was higher than desirable for the current disk control lers . A project l i ke the MicroVAX II system, in order to obta i n a good market share , needed to i mprove this relationship by reducing the cost of the disk subsystem .
Therefore , i t was very important for us to attain our cost goa l . To do that we placed a restriction on which components or technol ogies could be used, and what the assembly cost of the module cou ld be. Maximizing the number of machine-insertable parts there fore became an important consideration.
•
Performance-The MicroVAX
I f system wou ld support the ful l VAXjVMS operating system .
Since i t supports virtual memory, the VMS system uses large data transfers in the d isk subsystem. We therefore chose to optimize the performance of the controller around these large transfers to improve total system performance . By making the physical disk drive the l imiting factor, we evolved an architecture that would a llow simu ltaneous operations in the controller. In contrast, the curren t RQDX I and RQDX2 disk controllers l im i t the data transfer rate between the host memory and the disk drive because of their archi tecture . The single thread of control in these modules, though adequate for PDP- 1 1 systems, forced an interleave of logical data blocks on the disk surface. That interleaving wou l d h i n d e r t h e p e rfor m a n ce of t h e
MicroVAX I I system .
There are also many techniques for reducing the average seek time of the disk drives .
These methods include overlapped seeking on multiple drives, rotational optimi zations, improved seek algorithms, and various data buffering techniques . We wanted to include as many of these optimizations as possible and , since the goals were driven by the design team , the trade-offs were a little more flexible .
•
Dual module-This goal more than any other caused the most problems in the design of the hardware. Many times a solution seemed to meet all the goals but, when a detai led parts count a n d mock-up were created , there were a few components that just didn't fit on the board . Meeting this goal led to the exten sive use of CMOS gate-array technology to meet this size restriction.
•
Schedule-We did not have the luxury of set ting the date for the project's completion.
Because the disk controller was so important tO the overa l l MicroVAX I I project, we were given a completion date based on the availa b i l i ty of the Mi croVAX II hardware . Of course, this procedure i nvolved a manage ment factor that certainly kept the design team on its roes by being cold to see if we could do it. In response, we developed a schedule that wou ld maximize the work that could be done in parallel while keeping the risks at an acceptable l eve l .
•
Testable Design-T h i s goal became more i mportant as the details of the design were completed . The module, being driven by an onboard microprocessor, would be capable of self-diagnosis. Therefore , where possible , a l l i nternal ly addressable registers were made to be writejread registers and extra datapaths were added to maximize the a m o u n t o f l o g i c a va i l a b l e to t h e microprocessor for testing. This goal had to be weighed against the need for limiting the design complexity, cost, and size.
Task Partition ing
The short project schedu le forced us to adopt a development strategy that would maximize par allelism in the deve lopment of the RQDX3 . The first division was made between the hardware development and the microprocessor firmware development. Each major task was fu rther reduced tO smaller design functions . In many cases we had to create a model or emu latOr of some other undeveloped part of the design in order to allow tasks to continue.
Hardware Developmen t
Once the functional partition drawings were created , we had a solution that met the per formance and functionality that were required .
However, we still did not know if the cost and board area requ irements would be met. The design team quickly determined that some cus tom i n tegrated circuits would be needed tO help us meet these goals. Previous experience ,
6 8
Digital Tecbnical]out-nal
No.
2 March 1 986
a known process , and qu i c k tu rnarou nd made
CMOS gate array technology the key to our solution .
Two gate-array devices would be needed, but we had only one gate-array design team on our project . We decided that one gate array would be developed first and a TTL emulator of the second device would be created and used for the module-level testi ng . In that way, the inte gration of the firmware under development with the hardware cou ld begin early in the schedu l e .
The key area i n al most any disk controller ce nters around the design of the phase locked loop and the data separator logic used in recov ering the encoded data from the disk surface .
We knew at the beginning of this project that our team did not have the experience to design this section. Therefore, we employed the ser vices of outside consultants to this proj ect.
They contributed not only their previous expe rience in data separator design, but also rein forcement and management of the design phi losophy taught to us in the past .
Firmware Development
To meet our schedule goal , it was necessary to begin development and testing of the firmware for the onboard m icroprocessor well before any hardware was ready. The firmware consisted of many modu les, the majori ty of which were independent of the hardware . These modules cou ld be d e s i gned , cod e d , debugge d , and tested in parallel with the design, implementa tion, and debugging of the hardware . Then at a later date , the few remain ing hardware-depen dent modules cou ld be developed and inte grated tO form the complete RQDX3 firmware.
Thus, the target system first used for develop ing the firmware was not the prototype RQDX3 w i t h i ts o n board m i crop rocesso r , b u t a
VA.XfVMS system with two software emu latOrs
(one for the Q-bus subsystem and one for the disk subsystem) . The VMS system was chosen for several reasons : first, it has an extremely nice set of program development tools; second, the VMS d isk driver could be adapted to pro duce a ste ady stream of stimu l i ( d isk 1/0 requ ests) to ver i fy the correctn ess of the firmware's responses. With only a small amount of " trickery," the VMS system cou ld be "con vinced" to use a disk controller built not out of hardware , but out of software ; the two emulators mentioned above provided the necessary glue . The emerging RQDX3 firmware coul d be deve loped in the context of a normal VMS process, taking ful l advantage of VMS compil ers, l inkers, and debuggers . Although it took a lot of time (and many system crashes) to get this technique to work , it greatly speeded up the job of building all the hardware-indepen dent modu les . This stage took about fifty per cent of the total time spent to deve lop the firmware .
The next target system was the actual proto type RQDX3 with an in-circu i t emulatOr (ICE) for the microprocessor and a TTL emu lator for one of the gate arrays . Hardware debugging was accompl ished first by special code written to perform repetitive actions on particu lar por tions of t h e hardwa r e . T h e n , t h e a c t u a l fi rmware , which h a d been previously devel oped and was, in a sense , known to work, was loaded into the hardware . The ICE was a great help here since it allowed RAM to be substi tuted for ROM; that allowed a l evel of symbol i c debugging. A t this point in t h e process , the hardware-dependent modu les were built. This stage took abo u t thirty percent of the total firmware devel opment time .
The fi nal target system was the " bare"
RQDX 3 , with no emulators and real ROM . This configuration proved to be identical tO the pre vious one ( i . e . , no problems were fou nd in replacing the emulators with real devices) , but allowed prototype boards to be shipped i nter nally. The firmware of the RQDX3 could now be tested by different operating system groups , and bugs appropriately located and fixed. This stage took about twenty percent of the tOta l firmware devel opment time.
Design Verificatio n Testing
The p urpose of design verification testing
(DVT) is to assess at an early stage whether a d e s i gn has any particu lar i m p l e mentation problems . To do that , the board i s tested agai nst a l l Digital's applicable standards. First, the lay out of the board (the etch) is checked by l ook ing for noise radiation and pickup , and for undershoot or overshoot on clock l i nes. Then, the board is checked thermally to see if it can withstand both operating and nonoperating environmental stresses. Next, FCC testing is d o n e to m easure the rad i a ted fre q u e ncy spectru m . Finally, the module is shaken and
New Products
Digital Technical Journal
No. 2
March
I
'J86
69
dropped tO ensure that no chip fal ls out of its socket under normal hand ling conditions . Feed back from DVT can resu l t in physical changes tO the mod u l e , perhaps as severe as a new etch layout.
In the case of the RQDX3 , a recommendation was made to add resistors to a pair of clock l ines in order to dampen undershoot . Fortu nately, this a l teration did not have much impact on the schedul e .
Relia bility a n d Quality Testing
The purpose of reliabil ity and quality testing
(RQT) is to demonstrate that the product meets ce rtai n m i nimum reliabil ity standards , mea su red as mean time between failures (MTBF) .
The design team specifies the MTBF and also other measures of quality, such as hard and soft error rates, both of which affect the perceived quality of a disk controller product. Then, the
RQT team designs a test that wi l l demonstrate whether or not the product meets or exceeds these measura b l e q u a n t i t i es . Usu a l l y t h a t involves b u i lding a system (CPU , memory , serial line interface) t hat includes the product under test . The system runs some level of host software that exercises the product for a large number of hours under various temperature and humidity extremes. Designing these tests is not an easy task, and indeed the RQDX3 had major problems during RQT because of this dif ficu lty. Feedback from RQT can result in hard ware changes, or firmware changes, or both.
Ideally, if the product is changed, RQT should start again from the beginning. However, sched u les will often not al low that and compromises must be mad e .
A decision affecting all of RQT must b e made near the beginning: whether to test the product at the system level or at the module leve l. Test ing at the system l evel implies that the system
MTBF and error rates must be met, and a l l fai l ures, whether related t o t h e product under test or not, shou ld be counted . Testing at the mod u le leve l impl ies that the module MTBF and error rates must be met, and only fai l u res that can be attribu ted to components under test should be counted . Clearly, module-level test ing is preferred since it gives the most informa tion about the new product. However, modu le l evel testing is more difficu l t because each error has to be investigated tO determ ine its cause and whether or not it shou ld be cou nted .
Furthermore, the burden of proof is on the design team to verify that the error was not caused by their module. (Guilty u n til proven innocent')
Weighing all these factOrs, we decided to test the RQDX3 at the mod ule leve l ; that caused most of our RQT problems . A sealed chamber was used to control the tests of cycling over t e m p e r a t u re a n d h u m i d i ty e x t r e m es . The
RQDX3 modu les were placed in this c hamber, along with the systems into which the modu les were plugge d . Part of the testing included read ing and writing from both floppy disks and
Winchester disks . S ince these disks cou ld not withstand the environmental extremes inside the chamber, they were p laced outs id e . Early testing showed that this setup did not work, since the disk drives had to be connected to the control lers with lengthy cables, which were susceptible to noise pickup . This configuration was modified to bring the disk drives inside the chamber where they were connected tO the controllers with normal cables. That elim inated the noise problem, but now d ictated a reduced environmental stress on the RQDX3 module
(from class C to class A) .
At first, we encou ntered a higher-than-normal rate of soft errors on the floppy disks . A search for the cause of this problem showed that a com bination of two separate but contributing problems were responsible. First, a rare combi nation of events coul d cause the data separator for the floppy disk to temporarily fail to lock to the data stream. Second , most if not all the floppy disk drives themselves were not per forming correctly. The former problem was fixed by a component c ha nge to the data separator; the latter, by testing and repairing those drives that showed the greatest nu mber of soft errors . These two changes reduced the soft error rate for the floppy disks to a l evel we ll withi n the range specified by the design team.
The extensive , and lengthy, RQT also uncov ered one b u g in the error handli ng of the
RQDX3 firmware that had never been seen in our development lab. The problem cou ld only have been experienced by running many, many modules in paralleL Of cou rse, the pu rpose of
RQT is tO catch such problems then instead of at custOmers' sites.
The RQDX3 Architecture
The mass storage controller protocol (MSCP) defines the commu nication between the host processor and the d isk controller. Communica-
70
Digital Technical journal
No.
2 Mm·ch 1986
tion occurs using sequences of command pack ets, generated by the host, and response pack e t s , g e n e r a t e d by t h e c o n t ro l l e r . T h e transmission o f the packets and logical data blocks that are to move between the host and the controller is defined in the U/Q Storage
Systems Port (UQSSP) specification . These two spec i fications place the fo llow i ng req u i re ments on the contro l ler:
•
•
Two sequential-word register locations on the Q-bus are required . Those are referred to as the status and address (SA) register and the initialization and pol l ( I P) register. These registers must be able to be assigned at any longword boundary within the Q-bus 1/0 page .
The controller must have the abil ity to inter rupt the host processor using a previously loaded vector address .
•
The control ler must contain enough intelli gence to initialize itself, perform i nternal diagnostics, decode command packets , per form all disk control functions, transfer data, and encode response packets . These tasks are accomplished on the RQDX3 through the use of a DCT l l microprocessor.
•
The controller must be able to perform DMA data transfers on the Q-bus. These transfers will be for command and response packets, as we il as for disk data.
The diagram in Figure
1 shows the flow of information in an MSCP control.ler. MSCP com mand and response packets flow between the memory in the host processor and the on-board m icroprocessor. Disk data flows between the memory of the host processor and the disk sur face . Information dealing with the format of
New Products
HOST
MEMORY
MICROPROCESSOR
Digital Technical journal
No. 2 March 1986
DISK DRIVE
Figure I Information Flow in the RQDX3
DATA
7 1
7 2
The RQDX3 Design Project data on the disk surface (revector tables, format tables, etc . ) must be transferred between the disk surface and the m icroprocessor.
Figure
1 shows a centralized data buffer ele ment . It is used for temporary storage and as a means for smoothing the differences i n data transfer rates between the host memory, the mi croprocessor, and the disk surface .
It was decided to implement this centralized data bu ffer as a three-port memory system .
Three control elements are provided for the transfer of data between each memory port and the appropriate source or destination. These elements are the Q-bus DMA controller, the microprocessor w i th i ts internal bus-interface controller, and a VlSI d isk controller with an internal DMA interface . The i nterconnection of these subsystems is shown in Figure 2. Each control element assu mes that it has the memory system for its own dedicated use . The arbitra tion between these elements for access to the memory devices is handled within the memory subsystem .
The Memory Subsystem
The m e m o ry s u bsys t e m con t a i ns a fi n i t e sequential-state machine that receives requests for memory cycles from the three ports and per forms the memory cycle for the highest-priority requesting port . It is required that any port reques t i ng a me mory cycle must have i ts address and any required data avai lable before
A
�
0 - B U S
4� v
0 - B U S
I NT E R FACE
SU BSYSTEM t
C O N T R O L
I N FO R M AT I O N t
DATA
�
/)>. v
MULTIPORT
�
M EM O R Y
SU BSYSTEM
�
-----
---, ----,v//
M I CR O PROCESSOR
SUBSYSTEM
M S C P PACKETS t ,().
A D D R ESS
I
D I S K
,( t
A N D W O R K S PA C E t
DATA C O N T R O L v �
I NT E R F A C E
S U BSYSTEM
I
I
------
--'
I N FO R M AT I O N
�
�ll--------::=-:��--'
/
TO DISTR I B UTION BOA R D
'-
Figure 2 RQDX3 Subsystems
Digital Technical journal
No.
2
March 1986
posting the request to the memory controller state machine . The principle function of the memory system is twofold : first, it allows the controller attached to a specific port to deposit data to be written to the memory in a ho lding register; second, it allows the me mory control ler ro write that data to the
RAL\1 devices some time later. For most read requests, the me mory control ler performs a prefetch operation when there is an empty output register in one of the ports. This operation is possible because the accesses by both the d isk and Q-bus controllers are known to be sequential , with the next a d d ress always ava i l a b l e to t h e m e mory controller.
The port of the microprocessor is an excep tion to this prefetch operation. The memory controller cannot prefetch the data si nce mem ory accesses by a microprocessor are not a lways sequential . When requesting a cycle from the memory, the microprocessor will be " cycle slipped" (i . e . , wait states added to i ts micro cycle) until the memory controtler determines that the m icroprocessor is the highest-priority requesting device .
The highest priority for memory cyc les is given to the d isk controller port. Failure to ser vice this port first wil l cause overrun or u nder run errors in the disk controller chip, which has l i ttle buffering. These error cond i tions wou ld cause serious degradation of system per formance , since full disk revolutions would be wasted retrying the operations.
The midd le priority is given to the Q-bus
DMA con troller port . This port requires the h ighest service rate from the system (approxi mately 700 nanoseconds per request) . How ever, the port is capable of slowing itself if it cannot be serviced in time by the memory con troller. Of cou rse , to ach ieve the highest system performance and most effi c i e n t use of the
Q-bus, it is desirable that the Q-bus contro l ler never slow down.
The microprocessor is given the lowest prior ity for memory cycles . That allows the normal
.operation of data transfer between the d is k and host (both disk control ler and Q-bus DMA con troller active) to be completed as fast as possi ble. The microprocessor can use any remaining memory band w i d t h for its operation . The microprocessor uses the s hared memory for both temporary storage and its operational stack. Since its use of that memory will be infre quent, the microprocessor will not be affected by any loss in memory response .
A prototype of the memory subsystem was bui l t to measure the amount of bandwidth available tO the individual ports and tO deter m ine the effect of arbitration between the ports . A worst-case condition of requests from a l l ports was created and the bandwidth used by each was measured . With any two ports oper ating at their fu l l speed , there was no measura ble reduction in service rate from that of the ports running independently . When all three ports were operating, the disk port lost no memory bandwidth, the Q-bus port lost only one percent of its requested bandwidth, and the m i c rop rocessor lost e ig h t percent of i ts requested bandwidth.
These observations dur i ng worst-case condi tions indicated that all three ports are capable of operating at fu ll speed with their normal request patterns. This feature of the RQ DX3 allows it ro overlap disk data transfers, Q-bus
DMA transfers, and microprocessor operations to achieve maximu m performance .
The memory controller is implemented using a field programmable logic sequencer (FPLS) and an external inp u t sychronizer. Even though gate-array technology was used for the majority of the datapath on this mod u l e , it was felt that b u i lding the state mach ine in the gate array was too risky for the project schedule. The state machine was therefore placed outside the gate array. Only a few gate array pins connect it to the datapath elements that it controls .
The memory control l e r also incorporates some features to aid in the test and repair of the module. After module initia li zation , an input signal is asserted to force the memory control ler tO honor only those requests coming from the microprocessor. Without that, a hardware failure in e ither the disk controller or the Q-bus
DMA controller could constantly requ est mem ory cycles and cause the microprocessor to
" hang" on its first access to memory. With this signal asserted , the microprocessor can initiate the module d iagnostics in a small, isolated envi ronment that enables the microprocessor, ROM and RAM devices, and I/0 page registers to be tested . The microprocessor can then clear the signal later in its diagnostics, thus completing the module testing.
New Products
Digital Technical }ow-n a/
No.
2
Marcb
I
')8(j
73
The RQDX3 Design Project
The Micropro cessor Su bsystem
The microprocessor subsystem of the RQDX3 module is made up of a DCT l l m icroprocessor,
1 6K words of EPROM memory, a front-panel interface , and a prioritizing interrupt circu it.
Although many d i fferent m icroprocessors could have been used, the choice of the DCT1 1 was made with the fol lowing criteria in mind :
•
A 1 6-bit microprocessor cou ld handle the
MSC P requ irements adequately, while an
8-bit microprocessor would be strained and a
3 2 -bit microprocessor might be an overkil l .
•
A multiplexed address-and-data bus would reduce t h e n u m ber of gate array pins requ ired .
•
A rich, orthogonal i nstruction set (PDP- 1 1 system) that cou ld be easily understood should be used .
•
The m icroprocessor should be able to be programmed i n a high-level language . Much of the code for this modul e wou ld be written in the C programming language .
•
Relatively fast execution speed is desired.
•
Available hardware and software develop ment tools shou ld be used .
•
Our past design experience shou l d b e exploited t o improve the product's time to market.
The Q- bus Su bsystem
The Q-bus subsystem of this module is made up of the programmed 1/0 section , the Q-bus D MA controller section and the Q-bus interrupt sec tion. The Q-bus DMA controller i s composed of a finite sequential-state machine and associated datapath elements that are used to perform both block-mode and nonblock-mode Q-bus cycles.
The state machine is implemented in a field programmable logic sequencer rather than a gate array to eliminate the risk of schedu le delays due to coding errors . H oweve r, the datapath elements needed to support the state machine are contained within the gate array devices. Some of the features of this controller are
•
Ful l 2 2 -bit Q-bus addressing
•
A 1 6-bit DMA word counter
•
Q-bus memory parity detection
•
Ful l , efficient i mp le menta t i o n of Q-bus block-mode transfers
•
A programmable holdoff timer to regu late the Q-bus activity
The Disk Controller Su bsystem
The disk controller su bsystem had to provide the control and datapath fu nctions for both floppy and hard d isk drives i n the smal lest space and for the least cost. This requirement was satisfied by using a VlSI disk controller device .
The RQDX3 data separator is designed to receive the encoded data stream from the disk and convert it into a binary data stream and clock, both of which are then fed to the disk controller chip. The data separator is designed to operate at three different data frequencies tO be compatible with the available range of
Winchester and floppy disk drives. The fre quencies for each type of drive are as fol lows :
•
•
5 -MHz MFM encoded data recovery from
ST4 1 2 Winchester disks (RDSX type)
5 0 0-KHz M FM encoded data from high speed , h igh-dens i ty floppy d isks ( RX 3 3 type)
•
2 50 -KHz MFM encoded data from standard double-density floppy disks ( RX50 type)
The data recovery system for the RQDX3 is a unique MFM data recovery circuit that is very close to ideal . I n short, with proper matching of the device delays, the recovery window is
+50 nanoseconds, or one hundred percent of the window . This a lmost ideal data recovery is made possible by the following conditions :
•
•
•
•
A solid and precise phase l ocked loop is used.
The MFM encoding rules specify a 1 00 nanosecond " n u l l " period after each flux transition . This period is used to reset the edge store and compensation flip-flops of the circuit.
The VCO output has a fifty percent d u ty cycle.
The logic delay paths in the data separatOr circu its are carefu l ly matched. This matching was accom p l i s h e d by device match in g within the gate array that implements this function. Carefu l simulation of this logic was carried out to prove this operation .
74
Digital Tecbnicaljournal
No. 2
March
1986
The Structure of the Firmware
The firmware had to be designed to take fu ll advantage of the paral lelism provided by the chosen hardware architectu re. Therefore, the
RQDX3 firmware consists of a set of cooperat ing routines, or jobs, each of which performs a dedicated fu nction. Each job has its own stack and thus its own context and state informat ion .
Any operations that could possibly run in para l lel have been separated and are control led by separate jobs . A small operating system kernel provides facil ities for creating new jobs, sus pending and resu ming execution of a given job, acqu iring exclusive access to shared resources and later re leasing those resources, and sched u l ing jobs tO ru n based upon priority and resource contention criteria . This kernel pro vides a controlled way of overlapping opera tions . That effectivel y means that the RQDX3 can be simu ltaneously seeking on one or more drives, reading or writing from another drive, and transferring data to or from the host, a l l while perform ing calculations relat ing either ro the current transfer or tO a pending transfer.
Performance Tests
The main performance goal was to be able ro sustain a high data-transfer rate for large trans fers . In a typical situation, the VMS system uses the disk to swap, page , and load images. The
RQDX3 is tuned so that these operations are completed as rapidly as possible. Maximum sus tained data transfer rates of 4 2 0 KB per second have been measured, compared ro 1 7 0 KB per second on the RQDX2 . Such workloads are atypical , though, and do not give a good indica tion of overa l l system performance . When tested with a workload of from one to fifteen users on a MicroVAX I I syste m , the RQDX3 is faster than the RQDX2, but sl ightly slower than the KDAS O . This relationship is more in li ne with the performance based on theoretical cal culations. A user work load generates a lot of seeking, and the RD-class disks control led by the RQDX2 and RQDX3 seek more slowly than the RA-class disks control led by the KDASO .
Higher performance can be ga ined by split ting the disk activity among two, three , or even four disks . The RQDX3 has the ability to keep all fou r drives seeking at the same time. For small transfers, seek time dom inates , and an increase in system throughput of thirty-five ro forty percent can be rea lized . For large trans· fers, seek time is sti l l important but decreases in s i g n i fi c a n c e ; t h e i n c rease in s y s t e m throughput may only be twenty percent. The
RQ DX2 does not take advantage of separate sys tem and user disks; however, the RQDX3 wil l .
Higher performance o n a single drive can be achieved by queuing mu ltiple requests to the
RQDX 3 . The MSCP prorocol allows these m u lti ple requests to be automatically reordered by the controller to reduce the average seek time.
F o r exa m p l e , the contro l ler cou ld always choose the request with the shortest seek time instead of the first request in its queue . An increase in system throughput of thirty to forty percent occurs when the nu mber of outstand ing 1/0 requests increases from one tO twelve .
Summary
The RQDX3 design project carne close to meet ing all its design goa ls. There were 40 working un its exactly one year after the project began .
However, problems in the re liabili ty test setup, which de layed the manufacturing start u p , caused our first customer shipment t o s l i p . The cost, performance , and module-size goa ls were a l l met ro the satisfaction of the design tea m .
The high yields i n manufacturing can be attrib uted to the quality of both the design and the manufacturing process . Without the structu red design process and the team's adherence to it, this project wou ld not have been successfu l .
References
1 .
W. I . Fletcher, An Engineering Approach to Digital Desig n ( Englewood C l i ffs :
Prentice-Hal l , 1 980 ) .
New Products
Digital Technical journal
No.
2 March I ')86
7 5
Kathleen
D. Morse
Lawrence ]. Kenah
I
The Evolution of
Instruction Emulation for the Micro VAX Systems
The Micro VAX
CPU,
the 78032 chip, implements a subset of the VAX instruction set, yet the operating system must support the full set. To accomplish that, the Micro VMS developers decided to emulate the miss ing instructions-floating point, packed decimal, and character string instructions-in software. Since hardware and software were developed in paraUel, a VAX-11/730 system, with its microcode rewritten to make it act like Micro VAX hardware, was used as a test vehicle. The perfonnance measurements indicated excessively long execution times. The hardware design was extended to assist the software emulation task. The final emulator was also used in the UL TRIX-32 and VAXELN systems.
When Digital Equipment Corporation decided to implement the VAX architecture ' in sil icon, it was clear that the entire instruction set coul d not b e implemented o n a single chip. To deter m ine what could be i mplemented , a team of software and hardware engineers was formed to identify the best subset of the VAX instructi ons that woul d fit . As a consequence, the software engineers had to find ways to provide su pport in the operating system for those i nstructions removed from the base machine . This paper dis cusses how t ha t e m u l a t i o n s u p port was provided .
Micro VAX A rchitecture
The amount of microcode needed to implement an instruction is a good measure of the amount of space needed on a chip tO implement the same instruction. Microcode size thus became one m e a s u re used in determ i n i n g w h i c h instructions t O move off the chip. A second cri terion was the frequency with which particu lar instructions are used . For example, integer and logical instructions are used very heavily and their frequency of use is independent of the app l i cation area . Floating point instructions appear most frequently in scientific and engi neering computations . Packed decimal instruc tions are more common in certain commercial appl ications. Eventually, by balancing these considerations , the engineers i dentified a sub set of the VAX instruction set that would fit on one chip. That subset became the definition of the MicroVAX architecture . (The subset archi tecture also differed from the ful l VAX arch i tec ture in such areas as the console subsyste m . )
Once t h e MicroVAX architecture was com p leted, the hardware and software teams began independent development efforts . Since a major project goal was tO minimize the time to mar k e t , o n e h a r d w a re tea m i nv e s t i ga t e d a
MicroVAX implementation (the MicroVAX I sys tem) that used semicustom logic instead of a single chip. A second hardware team started the design of the MicroVAX chip itself , and a third team initiated the design of the implementation
(the MicroVAX I I system) that would incorpo rate that chip. At the same time, the softv.rare teams began their investigations of how to enhance the VMS, ULTRIX-3 2 , and VAXE LN operating systems in order to ru n these new mach i nes. The software designs were infl u enced in part by the need to implement and test the m issing- instruction software emu lation before any hardware was available.
Operating System Support
The major d i ffe rence between the software architectures of the MicroVAX and the fu ll VAX systems is the group of instructions that were
76 Digital Technical journal
No. 2 March I
986
not implemented i n the chip hardware . This group consists of
•
Floating point i nstructions
•
Packed decimal instructions
•
Character string instructions
(The MicroVAX architecture included the
MOVC3 and MOVC5 instructions because they were heavily used in fu ndamental rou t i nes, such as copying or fi lling memory arrays .)
Each of the three operating systems was sup ported by a d i ffe re nt design grou p . These groups had to decide which course of action to take to accommodate the reduced number of instructions that wou l d be implemented in microcode. The fol lowing alternatives were the most realistic courses to tak e :
1 . All compilers and assemblers could be changed to eliminate all uses of the miss ing i nstructions .
2 . Emulation subrouti nes that appl ications could link into their programs could be supplied. (VMS used this method on early VAX models that did not include h a r d w a re s u p p o r t for t h e G a n d
H floating point data types.)
3 . The e m u lation subro u t i nes could be implemented so that their use would be invisible to application programs and even to most of the operating system.
The VMS Decisio n Pro cess
The VMS design team began a study to deter mine the extent to which the m issing instruc tions were used in the operating system cod e , including all the various VMS u t i l ity programs.
As expected, the character string i nstruct ions were used most frequently and, in fact, were more widely used than expected. The CMPC 3 ,
CMPC5 , and LOCC instructions were the most frequently used string i nstructions, occurring almost everywhere that ASC I I text was manipu lated ( for exam p l e , in device names , fi le names, and DCL commands) . All software that included some kind of bi tmap (about six to ten different areas, ranging from the file system to memory management) used the SCANC and
SPANC instructions. A large nu mber of table lookup designs (including DCL and u til ity com mand parsers) used the MATCHC, MOVTC, and
MOVTUC instructions . Finally, the CRC instru ction was used by the BACKUP utility and by the
DECnet code .
Very few data types were used outside their realms and only a few u nexpected sequences were found that used the missing instructions.
One example was the use of the CVTLF instruc tion in the VMS kernel to determ ine the small est power of 2 larger than a given integer. A second example was the use of the CVTLP instruction i n the FORTRAN run-time support library as a quick method for converting binary representations to text.
Once the extent of the missing instruction usage was determined, the design team consid ered the number of compilers that were sup ported by the VMS operating system. I n all, over fifteen different languages are supported .3 The first alternative, c hanging t he compi lers and assemblers, would requi re that the code gener ators for each product be changed. Moreover, new versions of the VMS operating system and all its layered products would have to be gener a ted using these new compilers. That would involve a significant i nvestment of manpower, not just to enhance the compilers, but to pro vide ongoing support to maintain each product.
In addition, two variants of each new version of each product would have to be produced . A l i kely side effect was that these changes wou l d probably cause other development groups to l imit most layered produ cts to the MicroVAX subset on all VAX machines . I n that way, each group wou ld have to maintain only one version of their product .
Another consideration was the effect that the first or second alternatives wou ld have on the marketing of MicroVAX systems. Customers and
Digital's software engineers had become accus tomed to developing software on one machine and executing it transparently on any other machine in the VAX family. That wou ld not have been possible u nder either of the first two alternatives.
Through this reasoning process, it became obvious that the correct choice was the third alternative, to design for software emulation and make it transparent to both applications and operating system code . While requ iring a concentrated effort to write the emulation sup port , the overall effort for software emulation was much smaller than removing the use of the missing instructions from existing software and com piler code generators . The effort was also
New Products
Digital Technical journal
No. 2 March 1986
77
isolated. While some new code was needed, the number of changes to existing components was minimized . These changes were confined to the exception handler and the startup routi nes for the operating system . Finally, transparent emu lation of all missing instructions would guaran tee that systems implementing the MicroVAX architecture would be fully compatible with the VAX family of machines.
Implementation
As mentioned earlier, the MicroVAX program was geared to a tight time-to-market schedule .
That made i t highly desirable to develop the hardware and software in parallel as much as possible . The VMS design team decided to implement the emulation code and debug i t long before the hardware design specifications for a particular MicroVAX implementation were written. In this way, the emu lation code would be finished and working by the time the first
MicroVAX hardware was ready to be debugged .
Design of the Emulator
At this point in the project, several decisions were made relating to the design and imple mentation of the MicroVMS instruction emula tor. The emulation routines would be devel oped and tested by the VMS Deve lopment
Group. These routines would attempt to avoid features or coding techniques specific to the
VMS operating system . Thus the same emulation source code for the instructions could be used later by the ULTRI X - 3 2 and VAXELN Develop ment Groups.
The emu lation support was divided into two pieces. The first supported character string and packed decimal instructions (including CRC and EDITPC) ; the other, floating point data types. From the beginning of the MicroVAX effort, system configurations would be offered that provided some sort of floating point sup port in hardware . 4 That fact i nfluenced the design of the two pieces in the emulator.
Software support for floating point was viewed as a technique for ru nning programs that contained small amounts of floating point computation. Applications that depended heav ily on floating point operations would likely be run on systems that had floating point support in the hardware. Conversely, applications that depended heavily on packed decimal or charac ter operations did not have a hardware option at their disposa l . The decimaljstring emulator reflects that i n several places where space is sacrificed i n an effort to speed up the emula tion subroutines.
Struc ture of the Em ulator
Once the two pieces were designed, the actual coding began. Each of the two emulation com ponents was further divided into an operand decode piece and an i nstruction execution piece .
The operand decoder was a straightforward finite-state machine. It parsed the instruction stream one operand at a time, placing results into registers " appropriate" to each instruc tion . The register assignments were usually made by examining the expected register con tents after each i nstruction had completed its execution. For example, the final state of a
CMPC5 instruction suggests that R
1 and R3 be used as pointers to the two character strings, while RO and R2 contain the initial sizes of the strings .
The instruction execution routines were sim ple subroutines that accepted input parameters in registers and produced output conforming to the architectural specification of the instruc tions. For example, after the execution of an
ADDP4 instruction , RO and R2 contain zero, Rl and R3 locate the addend and sum strings, and the other registers are preserved .
A t t h e outset, several other decisions were made that simplified the design and implemen tation of the emulator.
•
Emulation support was provided transpar ently by being implemented at a very low level in the operating system.
•
Emulation subroutines were executed in the access mode of the missing instruction.
•
The existing emulation support for G and
H floating point data types woul d serve as a base for fu l l fl o a t i ng p o i n t e m u l a t i o n support.
Tra nsparent Support
To emu late the missing instructions transpar ently, the emulators had to become an integral part of the operating sys te m . They were loaded into system space during the system bootstrap and connected directly to the reserved-opcode exception vector in the system control block.
W h e n e v e r a r e s e rv e d - o p c o d e e x c e p t i o n occurred, t h e emulator woul d distingu ish the
78
Digital TecbntcaJ ]ournm
No. 2 March 1986
execution of a mJssmg instruction from other ill ega l opcodes. Missing i nstructions would cause a control transfer tO the appropriate emu lation subroutines. Other lllegal opcodes were passed on to the operating system as excep tions. Since the host operating system provided support in a transparent fashion, existing pro grams cou ld execute on a M icroVAX system without being changed.
A ccess Mode of Execution
The reserved-opcode exception handler had to begin its execution in kernel mode , as defined by the VAX architecture . However, if the emula tOr routines c o n t i n u e d in that m od e , the address val idation rules demanded that not only each operand but also each byte in a character string be probed for read or write access before that operand could be used. Because of the excessive cost of these operations, we decided that the emulator routines would execute in the access mode in which the missing instruction was used . If an operand or string was not acces sible, an access violation exception would occur, which could be intercepted for special processing by the emulator.
The Use of Existing Routines
An emulator for G and H floating point data types already existed . I nstead of completely rewrit i ng this emu lator to accommodate a l l four data types, it was restructured to separate its operand packi ng and u npacking routines from the arithmetic and conversion operations .
The n , additional packing and u npacking rou tines were added for F and D floating point data types. Also, the overall structure of the floating point emu lator was changed from a condition handler to an i n tegral piece of the operating system . (A condition handler executes only within user programs, while an integral compo nent wou ld receive control whenever a missing floating point i nstruction is executed .)
In itial Testing
I t was obvious that a testbed was needed to enable the design team to debug the emu lation software. Some method was needed to force the emulation software to gain control in order to execu te the missing i nstructions. Since the VMS macro assembler can substitute a macro for an instruction opcode , macros cou ld be used to cause the assembler to take speci a l action whenever it encountered any of the missing instructions .
A set of macros was written that caused spe cial object code to be generated whenever any of the missing i nstructions was encountered.
This special object code consisted of a byte containing the i l l egal opcode FE (hex) , the opcode for the i nstruction, and all the operand specifiers . When one of these i nstructions was executed, a reserved-opcode exception was generated. A special exception handler wou ld then advance the PC from the byte containing the FE opcode to the actual opcode . Control was then passed to the instruction emulator.
One of these macros is l isted in Figure l .
Using these macros, programs written i n assembly language coul d be reassembled and executed using software emu lation for the miss ing instructions . Thus any existing VAX proces sor, such as a VAX- 1 1 /7 3 0 system, could be used as a testbed for the software emulation.
Results of In itial Tests
One key factor to determine was the i ncrease in execution time requ ired by software emulation for different parts of the operating system and for appl ication programs. To determine these differences, the VMS Performance Grou p at
D igital ran standard i nstruction-timing tests against the emulation code . Because these tests were run on an existing VAX processor, the exe cution times for emulated instructions could be compared to those done in hardware on the same VAX processor. These test results showed that it took about ten times longer to emulate character string i nstructions than to execute them in hardware .
To determine the reasons for this disparity, the design team performed a close i nspection of the emu lation code . Quite quickly it became obvious that, for the simpler string i nstructions, the operand decode required as much time as the i nstruction execution . To s peed up the emulated instructions, hardware support was requested by the MicroVMS tea m .
To support this request , w e made a l ist o f the operand types for the missing character string and packed decimal instructions. There were only 5 operand types in all 27 i nstructions .
These operand types were already being used by i ns t r u c t i o n s t h a t were a part of t h e
MicroVAX subset, such a s M OVC3 a n d MOVC5 .
A meeting of the hardware and software teams
New Products
Digital Technical journal
No.
2
March 1 986
79
The Evolution of Instruction Emulation fo r the Micro VAX Systems
. t i t l e
$ o p d e f l o c c t s t
R e d e f i n e t h e
L O C C o p c o d e w i t h a n e w
L O C C ma c r o
. o p d e f l o c c _ f e < < o p $ _ l o c c @ 8 > ! A x f e > , r b , r w , a b
. ma c r o l o c c _ f e
. e n d m l a c e l a c e c h a r . r b , l e n . r w , a d d r . a b c h a r . r b , l e n . r w , a d d r . a b d e s c : . a s c i d " Th i s i s a t e s t "
; T e s t d a t a f o r
L O C C
; T e s t p r o g r a m t o t r y a
L O C C i n s t r u c t i o n .
. e n t r y l a c e m ov z w l r e t
. e n d s t a r t _ h e r e , O
< H A a " " > , d e s c , @d e s c + 4
H 1 , r 0 s t a r t _ h e r e
E n t r y p o i n t f o r t e s t p r o g r a m
G e n e r a t e a n e m u l a t e d
L O C C
S t a n d a r d e x i t s t a t u s c o d e
E x i t f r o m p r o g r am
E n d o f t e s t p r o g r am
Figure 1 Test Program with Macro for L O CC Instruction concluded that there wou l d be little cost to the underlying hardware if these operands were decoded before a missing instruction exception was signaled .
Design of New Em ulatio n Exceptio ns
The result of that meeting was that two new exceptions were added to the MicroVAX archi tecture as emulation assists. Since the hardware cou ld easily decode the operands for the char acter string and decimal stri ng instructions, they were defined as the ones that the new exceptions woul d support . Thus, two of the three i nstruction types not implemented in hardware cou l d now be handled effective ly.
The third type , floating point instructions, woul d continue to cause reserved-opcode exceptions, s ince their operands cou ld not be decoded without significant additional hard ware support . (A separate floating point unit, the MicroVAX 78 1 3 2 chip, provides this hard ware support for three of the four floating point data types.)
4
The first exception is generated whenever a character string or decimal string instruction that is not in the hardware subset is executed .
The process causes the hardware to decode the operands and push the exception parameters onto the current stac k . The exception parame ters are depicted in Figure 2 .
The second exception occurs only when one of the emulated instructions is executed and the first-part-done (FPD) bit is set in the program status longword ( PSL) . The VAX architec ture allows many instructions (including a l l the decimal and character string instructions) to be interrupted after partial execution. The original operand specifiers cannot be decoded again because the register contentS may have been altered to store the intermediate results . When this second exception occurs, the exception handler u npacks the intermediate resu lts and resumes execution at the point where the i nstruction was interrupted.
�
1---
OPCODE
PC OF INSTRUCTION
DECODED FIRST OPERAND
OTH ER
DECODED
OPERANDS
UPDATED PC
PSL OF EXCEPTION
-
'"'
-
Figure
2
Exceptio n Param eters for
Emulation A ssist Exception
80 Digital Technical journal
No. 2 March I 986
Note that this second exception can occur only when an access violation has a l ready occurred during instruction emulation . I n that case , the operating system's access violation handler transfers control to the emu lator.
Enough intermediate state is stored i n the regis ters to allow restarting the instru ction, at which time the stack is restored to its state when the instruction began execution . Then the excep tion PC is changed from a PC i nside the emula tor to the PC of the original instruction that triggered emulation. Finally, control is passed back to the operat i n g sys t e m ' s exception reporting mechanism . ( Page fau lts, device interrupts, and the like are i nvisible to the user and require no special handling. That is, there is no need to pack the state i nto the registers and alter the saved PC.)
Final Design of the Instruction
Emulators
The final design produced emulation support in two pieces: one for the m issing floating point i nstructions; the other for packed decimal and character string i nstructions. Although the two emulatOr programs supported different data types, their overall design contained many com mon threads . This section describes the com mon design philosophy, as well as the step-by step operation of each emulator.
Common Design Philosophy
Nearly all the emulation code executes in the access mode in which each missing i nstruction was origina l ly executed. The stack associated with that access mode is used as a working stor age area for the emulation routines .
The emulation of m issing instructions is nearly invisible to programs in the sense that memory and register contents are identical co those obtained on ful l VAX implementations .
The only difference between the emu lated and hardware i m p l e m e ntations i s in the time requ ired to complete an instruction and i n the stack remnants from the emulator's temporary storage area. (Memory locations at small nega tive offsets from the top of the stack are speci fi e d as U N P R E D I CTAB L E i n t h e VAX architecture .)
The two emulator p ieces share a common phi losophy, if not common code , in regards to the two memory management faults . One fault is made in response to an inva lid page and the other when a reference is made to a page that is not readable or writable as requ ired .
No special treatment is requ ired for page faults (translation-not-valid faults) . If an invalid page is referenced by the emulator, a page-fault exception is reported to the operating system .
The P C in the page-fault frame points at the instruction within the emu lator that referenced the invalid page . After the operating system makes the page valid, execution resumes with the faulting instruction .
References to pages that are not accessible
(access-violation fau lts) are more complicated than the page fau l ts . Access-vio lation fau Its, unlike references to i nvalid pages, are visible at t h e p r og r a m l ev e l . W h e n t h e e m u l a t o r intercepts the exception, the fau lting P C points at the emulator instruction that references the inaccessible page . The stack contains working storage that must be removed and saved regis ters that must be restored. I n that way, the exception looks like an access violation gener ated on a fu ll VAX implementation . For most floating point instructions, an access violation implies that the state of the machine will be reset to its state when the instruction began. For the decimal, string, and POLYx i nstructions, the i nstruction can be left in a partially com pleted state . The intermediate context is scored in the registers and the FPD bit is set in the saved PSL . This bit a llows the emu lator to resu me these instructions at the point where they left off, rather than restarting them from the beginning (assuming that the access viola tion can be resolved) .
Floating Point Emulation Support
The program that emulates the missing floati ng point instructions in software differs in several details from the decimaljstring emulation rou tines . In floating point emulation , the functions are performed in the following order:
1 . Execution begins i n kernel mode as a resu lt of a reserved-opcode exception.
2 . I f the exception occurs i n a mode other than kernel , the exception parameters are copied to the stack of that access mode . Further emulation takes place in that access mode.
3 . Each operand i s decoded .
New Products
Digital Technical Journal
No. 2
March 1 986
8 1
The Evolution of Instruction Emulation fo r the Micro VA X .�ystems
4 . Floating point operands are u npacked into exponent and mantissa.
5 . The operation (arithmetic or conver sion) is performed .
6 . I f the result is a floating point nu mber, the resulting exponent and mantissa are packed into a single number.
7 . The resu It is stored and the exception dism issed .
Before the exception is dism issed, the float ing point emulatOr exami nes the opcode of the next instruction. If it is also a floating point instruction, then control is passed back co the beginning of the emu lator tO begin the operand decode for the next instruction . This technique saves the overhead of dismissing one exception a n d i m m e d i a t e l y generati ng an i d e n t i c a l reserved-opcode exception .
The nature of floating point operations allows many instructions to accomplish their resu lts by sharing different routines. There are routines that can u npack and pack each of the four floating point data types. There are also routines that perform the various arith metic and conversion operations. Because these rou tines operate on unpacked numbers, the rou tines are independent of the initial data type.
The floating point emu lation routines sup port all four floating point data types . Thus the routines can be used with all MicroVAX systems and other VAX systems that do not implement all fou r floating point data types in firmware or hardware .
Decimal/String Emulation
Support
The emulation of a character string or packed decimal instruction proceeds as fol lows :
1 . Execution begins in the access mode in which the missing instruction was origi naily used .
2 . Operands are moved from the stack into registers and control is passed to a n instruction-specific routine_
3. Some instruction results (for example, from MOVTC, MOVTUC , and packed decimal arithmetic and conversions) are s t o r e d w h i l e t h e s e r o u t i n e s a r e executing.
4 . The rou tine executes until an input or ou tpu t string is used up, at which time it completes the storage of resu lts . Execu t i o n i s r e s u m e d w i t h t h e n e x t instruction .
Because the decimal/string emulator relies on hardware for its operand decode stage , the lookahead technique used by the floating point emu lator cannot be used for decimal and string instructions . I f the i nstruction fol l owing an emulated instruction also requ ires emulation support, the following sequence takes place :
1 . The first exception is dismissed.
2 . The next instruction is execu ted.
3 . The operands of t hat i nstruction are decoded and stored on the stac k.
4 . The d e c i m al/string e m ulatOr regains contro l .
Since these instructions perform many unre lated operations, t here is little code that can be shared between their emulation rou tines.
Testing and Debugging
The main problem i n testing the emu lation s o ftware i n i t ia l ly was t h a t t h e re was no
MicroVAX hardware available du ring most of the implementation cyc l e . Thus we had to deve lop techniques to simu late the hardware in order to begin the tests . There were two chief techniques used to test and debug the emula tor. First, instru ction-specific routines were tested as user-mode programs in a normal pro gram development environ ment . Secon d , the exception handler front-end was tested on a
VAX - 1 1 /7 3 0 system that was m od ifie d , by rewriting some of the 1 1 /7 3 0 mi crocode, to act l i ke a MicroVAX system .
Instructio n-Specific Testing
Microcode written for a particular implementa tion (both VAX and MicroVAX systems) can be used only on that particu lar machine or a simu lation of that machine . However, macro-level code can be executed on any VAX processor.
Therefore , since the emu lation routines were written in macro-level code that executes on any VAX p rocessor , " no r m a l " d e b ugging
82
Digital Tecl:mical journal
No. 2 March 1986
techniques could be used for part of the debug effort .
A set of test programs was constructed that would run on other VAX processors ( 1 1 /7 3 0 ,
1 1 /7 5 0 , a n d 1 1 /7 8 0 ) . These test programs would call each instruction-specific subroutine and compare the resu lts (memory contents, reg ister contents, and settings of the condition codes) with the output from the corresponding instru c tions executed on those processors .
These tests allowed the basic a lgorithms to be debugged even before they were plugged i nto the emulator. The set of tests was limited only by the choice of input data for each instruction.
The first set o f tests u nc overed m o s t algorithmic problems b u t d i d not exercise the error paths (such as inaccessible source or des tination strings) . The code to handle these error conditions was written later in the develop ment cycle . Neither the absence of these error paths nor errors in edge conditions (such as zero-length strings) prevented the VMS system from executing.
Another benefit of a macrocode implementa tion was seen during the debug of the edge condition problems. Since the instruction emu lation routines were just an extension of the operating system, the debugging tools used for other operating system code cou ld be used to debug the emulator.
Testing the VAX- 1 1/73 0 Breadboard
Implementation
The availab i l i ty of the two new emu lation exceptions changed the strategy for debugging the emulation code . The software solution used to obtain preliminary resu lts was u nable to mimic the new exceptions invented to assist the emulatio n . Therefore, a new testbed was needed to accommodate the debugging pro cess . The testbed had to decode the operands and generate the appropriate exceptions to pass control to the software emulation code . One way to perform these functions was to alter an existing VAX system, such as the VAX- 1 1 /7 3 0 processor.
The 1 1 /730 is an entirely " soft" machine; that is, all its microcode is loaded at powerup rather than being resident in ROM. By altering that microcode, the design team could make the 1 1 / 7 3 0 look l ike the architecture in a
M icroVAX system . The required changes were simply a matter of removing the microcode for instruction execution while leaving that for operand decode . To finish the alterations, the design team had to write a new "exception gen erator" to create the emulation exceptions.
At this time in the project, the first real
MicroVAX hardware wou ld still not be available for nine months. Therefore , the VMS design team decided to undertake the modifications to the 1 1 /7 3 0 's m icrocode and to b u i l d t he testbed. We estimated that this effort would take one to two months, since the VMS devel oper had to learn to write microcode . That meant that the software emulation code would s t i l l be c o m p l et e d l on g before t h e first
MicroVAX hardware was ready.
The m i crocode s o u rc e programs w e re acquired from the 1 1 /730 m icrocode team and assembled u s i n g the l a test vers i o n of the microcode assembler. The 1 1 /7 3 0 microcode was structured as separate modu les for different functions (for example , floating point, compat ibility mode, exceptions, memory management, and so on) . Due to the lack of a " linker," label files that allowed routines to be called across modules had to be created. To speed the devel opment, the design team wrote several FOR
TRAN tools that automatical ly generated new label files . In addition , command files were built that correctly created a new set of binary microcode files from a set of modified sources .
The next step was to change the 1 1 j730's microcode . Since it had to exist in a limited amount of RAM space, the new code could not be added without removing some existing code . Therefore , we decided to replace the compatibi lity mode microcode with a new rou tine to generate the emulation exception. Some new flags were added that, at the developer's c h o ice , w o u l d a l low d i fferen t c l asses of instructions to be e m u lated ( i . e . , decimal string, c haracter string, or floating point) .
F i n a l ly, to boot the VMS syst e m o n t h is
MicroVAX version of an 1 1 /730 , we had to enhance the VMS bootstrap code to load the emu lation exception hand lers and connect them to the appropriate exception vectors.
Now the software emulation code , from the exception handler all the way down to instruc tion execution, could be debugged . The best measure of the success of this venture was made when M icroVAX hardware was finally available. The customized VAX - 1 1 /7 3 0 system was such a good testbed , not o n ly for the
New Products
Digital Technical journal
No.
2
March 1 986
8 3
instruction emulator but also the rest of the
MicroVAX I support, that it took a mere four days to get the VMS system running.
Other Test Mechanisms
The i nitial testing of the instruction emulator consisted of a set of programs and sample input data for each of the missing instructions. While providing routines that worked in almost a l l cases, these tests d i d not exercise some o f the more exotic edge conditions . Those inclu ded very long or very short strings, i l legal operands, or strings that were not readable or writabl e .
Once MicroVAX hardware was available, several new testing techniques cou ld be used to exer cise the emulator.
Operating System Code
More testing was provided by running the oper ating system code with the emulator providing character-string and packed-decimal support .
The VMS Development Group has a large set of regression tests that exercise most success and error paths within the operating system . These tests plus normal daily use by the VMS develop ment community ensured that extensive testing of the instructions used by the VMS operating system was performed .
Once t h e VMS system was r u n n i n g , the
ULTRIX- 3 2 and VAXELN Deve lopment Groups requested the source code for i ncorporation into their systems . These systems exercised parts of the emulator that the VMS system did not use. The ULTRIX kernel uses a sma l l num ber o f packed decimal i nstructions (AS H P ,
ADDP4 , SUBP4 , a n d EDITPC) for some o f its arithmetic and formatting support. When the
ULTRIX- 3 2 operating system first exercised the e m u l a tor, several bugs were detected a n d corrected.
Compiler-Gen erated Code and Associated
Tests
The base operating systems used packed-deci mal and floating point instructions in a smal l number of cases . These instructions received better testing using programs written in COBOL and FORTRAN . The compilers and their valida tion tests were used to test the emulator rou tines from the time they were first written u ntil they finally shippe d .
Architectural Co nformance
Even such continual testing is no guarantee that each instruction executes according to the VAX archi tecture specification. Most of the testing described so far exercised the su ccess paths of the emu lation subrouti nes. The error paths, especially the code that intercepted and modi fied access violations, required a different set of tests .
CPU Diagnostics
For each CPU designed by Digita l , a set of CPU diagnostics is written that exercises as much of the central processor as possibl e . Included in these diagnostics is an instruction-set exerciser that tests for proper behavior in at l east some of the interesting error cases. The CPU diagnostics for the MicroVAX I served as the primary test for the access violation handler in the deci maljstring emulator.
AXE Ve rification Program
All new VAX computers at Digital are tested with an architectural verification tool known as
AX E . AXE p rograms are used to determ i ne whether or not the machine conforms to the
VAX architectural specification . AXE accom plishes this testing by subjecting each VAX instruction, with many combinations of oper ands, to a variety of error conditions . These conditions i n c l u d e i naccess i b l e operands , instructions or operands that cross page bound aries, and u nusual operands.
When the MicroVAX instruction emulator was subjected to AXE testing, the only bugs that remained involved an instruction restart fol low ing an access violation .
Results
As a result of this strategy, the software emula tion code was completed and fully debugged before the first real MicroVAX hardware was finishe d . The ULTRIX- 3 2 and VAXELN oper ating system groups were able to take the VMS emulation code and convert it to work under their operating systems. That took much l ess effort than was requ ired for the VMS develop ment team to implement that code. With this technique , bugs found in the instruction-execu tion logic in one system cou ld be corrected i n all three operating systems.
8 4
Digital Technical journal
No. 2 March 1986
A second benefit of this engineering effort was seen by the hardware designers . The revised VAX- 1 1 /7 3 0 m icrocode sources and microcode tools were further modified to cre ate a MicroVAX CPU chip simu lator. The simu lator allowed the MicroVAX CPU boards to be tested before any MicroVAX chips were actually available.
The biggest gain of a l l was that no applica tion software, compilers , or operating system code had to be rewritten to avoid the use of the m issing instructions .
References
1 . VAX A rchitecture Reference Man ual
(Bedford: Digital Equ ipment Corpora tion , Order N o . EK-VAXAR- RM- 0 0 2 ,
1 983) .
2 . D.W. Dobberpuhl et a!, "The MicroVAX
78032 Chip: A 3 2 -bit Microprocessor,"
Digital Techn ical journal
(March 1 986, this issue) : 1 2- 2 3 .
3 . VAX Software, Lang uages and Tools
Handbook (Maynard: Digital Equ ipment
Corporation , Order No . E B- 2 7 24 0- 4 8 ,
1 985) .
4 . W.R. Bidermann et a ! , "The MicroVAX
7 8 1 3 2 Floating Point C h i p , " Digital
Technical journal (March 1 98 6 , this issue) : 2 4 - 3 6 .
New Products
Digital Tecbnicaljournal
No. 2 March
1 986
85
Steven
E.
Boone
Guenter
E.
Schneider
I
The TK50 Cartridge
Tape Drive
A streaming tape drive, the TK50 subsystem, provides fast backup and data transfer for small computers like the Micro
VAX II
system. A single reel cartridge, using half-inch magnetic tape, stores 100 megabytes of data. A unique tape transport system automatically threads the tape when the cartridge
is
inserted. The drive reads and writes data in a serpentine manner, going the entire tape length first on one track, then another. For high data integrity, the TK50 subsystem employs a sophisti cated error-recovery algorithm, reading data after writing it and rewrit ing any corrected data farther
down
on the tape. The Q-bus controller, the TQK50, contains complex firmware conforming to Digital's Storage
Architecture and controlling data transfers between the
CPU
and the tape.
As the performance of computer syste m s expands whi le their size shrinks, many factors demand special attention. One major factor is storage systems. Over the past few years , disk drives have made dramatic advances, providing storage capacity of hundreds of megabytes i n very small and relatively inexpensive packages.
Since the predomi nant technology for today's disk drive is based on the fixed-media concept, some means of providing system backup and data transfer capabi l ities is required . Magnetic tape systems are still the most viable way of providing these capabilities.
Ease-of-use considerations require that a backup/transfer device be matched in capacity to the supported disk systems. It should also be extremely reliable , fast, and very cost effective .
This paper describes a peripheral subsystem , t h e TK5 0 m a g n e t i c cartridge tape drive
(Figure 1 ) , that meets a l l these requirements.
Design Goals of the
TK50
Subsystem
The TK5 0 cartridge tape subsystem was con ceived to meet the needs of the MicroVAX I I and similar computer systems. A study o f tape products then available indicated that existing quarter-inch cartridge drives did not provide either the performance or the capacity required to back up the large capacity disk drives sup ported by these systems. Existing drives also l a c ked t h e re l ia b i l i ty and data i n te g r i ty required to complement the designs of our new microsystems. Therefore , Digital designed the
TK5 0 cartridge tape subsystem to meet the needs of the MicroVAX II system and other sma l l to mid-range computers .
A wide variety of factors defined the design goals of the TK5 0 subsystem . It had to fit into a standard 5 l;.:i -inch form factor and provide high capaci ty with high data i ntegrity. The desire for mechanical simplici ty , rel iability, and l ow cost, while maintain i ng good performance, dictated a streaming tape design . The TK50 subsystem had to be compatible with the Q-bus, and the
TK50 controller had to support the Tape Mass
Storage Control Protocol ( TMSCP) of t he
Digital Storage Architecture .
Our i nvestigations led to the concept of an automatic-threading, single-reel cartridge that utilized the established medium of i nstrumen tation tape. This tape supports high bit densi ties and fast tape speeds, allowing great latitude in specifying the performance and capacity of the TK50 subsystem . We also decided to use
86
Digital Tecbnicaljournal
No.
2
March 1986
Figure 1 The TK50 Tape Drive half-inch tape , rather than quarter-inch, to max imize capacity.
The requirement of the MicroVAX
II system, as well as our desire to minimize risks i n a first generat ion product, d ictated that the tape capacity should be 1 00 megabytes (MB) .
System Design
The TK50 cartridge tape subsystem was devel oped with three major components:
•
•
•
A tape cartridge , cal led the CompacTape Car tridge, that houses 600 feet of half-inch tape and supports the auto-threading feature of the transport mechanism
A unique streaming tape transport featuring auto-threadi n g and a microprocessor-con trolled servo-system
An i ntelligent, microprocessor-based Q-bus controller that supports TMSCP
Compac Tape Cartridge
The CompacTape Cartridge is unique in many ways. First, it provides a large amount of data recording surface for its volume. The cartridge has approximately two hundred and fifty times the recordi ng surface area of a single-sided
5 � -inch floppy disk. Moreover, compared to the only commercial tape product then availa ble to fit the 5 � -inch form factor, the Com pacTape Cartridge is four times as efficient in utilizing tape volume i n relation to cartridge volume . The cartridge is designed to maximize the volume of tape in the standard form factor of the 5 � -inch drive . The cartridge , shown i n
Figure 2 , contains a s i ngle reel with the tape occupying forty percent of the cartridge's vol ume . The tape is Y2 inch wide, .00 1 inch thick, and 600 feet long.
Second, the CompacTape Cartridge is a com p letely e nclosed device that never exposes the media to the environment, thus greatly enhanc i ng the data reliability of the entire subsystem .
Third , the CompacTape Cartridge a l lows automatic tape threading once it is inserted i nto the TK50 tape drive . This auto-threading func tion is a key feature of the mechanical design of the tape transport.
Digital Tecbnical journal
No. 2 March
1 986
87
New Products
The TK50 Cartridge Tape Drive
D R I V E
H U B
Figure
2
The TK50 Tape Cartridge
The auto-threading works i n the fol lowing way. When a cartridge is i nserted into the drive, the tape must be threaded around the tape guides, over the readjwri te head, around the take-up reel , and then fastened to the reel hub.
Two leaders are used ro accomplish the thread ing, as shown in Figure 3 . One, made of . 0 0 7 i nch Mylar, is attached r o the BOT e n d o f the tape in the cartridge ; the second is attached to the hub of the take-up reel i n the drive . This second leader has an arrow-shaped tip that reaches from the reel , through the tape path , and i nto the area that w i l l be occupied by the tip of the first leader when the cartridge is i nserted . During the insertion process , the arrow-shaped tip is moved by a cam into the opening of the cartridge leader. Tension is then a p p l ied to lock the leaders toge t her. This
"buckle " is now ready to be pulled t hrough the tape path and wound onto the take-up ree l .
This buckl i ng process is accompl ished by two l i nks in the drive , in conjunction with a constant tension applied by the motor to the take-up leader. One link uses a cam to move the two leader tips into each other. The other l i nk holds the take-up leader in the correct pos ition and retreats at the right instant, allowing the motor to cinch the buckle . The entire process
OPENING
0
CARTRIDGE
LEADER
ARROW
SHAPED
TIP
R E T AINING
NOTCH
TAPE DRIVE
LEADER
0
BUCKLE
RELATIVE
MOTION
0
0
STEP
1
STEP 2
STEP 3
Figure 3 Engagement of Drive Leader to Cartridge Leader
88
Digital Technical journal
No. 2 March 1986
happens du ring the last half-inch of insertion as the cartridge enters the drive . (See Figure 4 . )
This l inking takes place withou t any tape being spool ed out of the cartridge.
When the tape is rewound into the cartridge for removal from the drive, the two ears on the cartridge leader come to rest in a pocket in the cartridge shell . When the cartridge is removed from the drive , two opposing locks hold the reel in this position. The toothed locks engage with rhe teeth on the outer diameter of the reel flange . Thus locked, the tape stays tightly wound and the leader tip is kept in the correct position for a subsequent buckling process.
Tape Transport
The TK50 tape transport (Figure 5) consists of two major components: the tape drive and a single printed circu it board assembly.
The tape drive encom passes the mechanical and electromechanical components to read data from and write data to the magnetic tape . The drive 's major components incl ude
•
The magnetic readjwrite head and i ts linear positioner
PROPER LOCATION OF LEADER
LEADER UNHOOKED
LEADER HIDDEN
LEADER DISPLACED ABOVE LINK
Figw·e
4
View of Leader Shown in Fo ur
Positio ns
New Products
LEADER. TAKE-UP
EXTERNAL
SHIELD
ASSEMBLY
TACHOMETER
ASSEMBLY
BRACKET
&
H E A D ASSEMBLY
SPRINGS
REEL. TAKE-UP
I
LATCH
LINK. BUCKLING
LINEAR
ACTUATOR
Digital Teclmical]ournal
No. 2 March I 986
INSULATOR
ASSEMBLY
Figure 5 TK50 Tape Drive Tra nsport
BEZEL
ASSEMBLY
89
WR ITE
READ
90
TAPE
ISLANDS
READ
WRITE
Figure
6 TK50 ReadjWrite Head (Top View)
•
The cartridge threading mechanism
•
The take-up reel and i ts motor
•
The drive hub mechanis m , which interfaces tO the CompacTape Cartridge, and its motor
•
The tachometer, which provides feedback to a microprocessor, the 805 1 , for tape speed control
•
Various sensing devices that monitor and control the hand ling of the tape as it passes over the readjwrite head
Read/Write Head
The readjwrite head is designed with fou r islands t h a t a r e in contact with the tape
(Figure 6) . The tape forms a polygon as it con tacts these four areas. Each island bends the tape by an angle of six degrees . Over its width, each island is curved by an amount correspond ing to the radius of the natural curvature of the tape under working tension, thus assuring good surface contact (Figure 7) . The narrow islands limit any temporary liftoff (due to contam ina tion) to very short sections of tape , and they clean the tape as wel l .
HEAD ---1--
1/2" TAPE
I
1�1
I
ISLAN DS
Figure 7
TK50 ReadjWrite Head (Side View)
Digital Technicaljournal
No. 2 March 1986
Except for the ferrite cores, the entire head b lock is made of ceramic material to ensu re long life. The two inner islands contain the readjwrite cores; the two o u ter ones direct the tape to the inner ones so that uniform contact between the tape and the head is provided . On the upper part of the head assembly are two gaps, a write gap ( .0 1 8 inch wide) fol lowed by a read gap ( . 008 inch wide) , that read and write data when the tape is moving forward .
Two corresponding lower gaps read and write data during reverse tape motion. The lower gaps cover the odd tracks and the upper gaps cover the even tracks ; thus, the head has to traverse only half the tape width , helping greatly to keep the height of the drive within l imits . The track spaci ng is . 0 1 9 inch.
A utaThreading
As the cartridge i s inserted, its door opens, expos i n g t h e cartri dge l e a d e r . T h e n , as described earlier, two plastic arms i n the drive act to buckle the cartridge's supply l eader to the drive's take-up leader. The rest of the auto threadi ng process is handled by the drive's motOrs, sensors and m icroprocessor.
Tape motion and tension control is accom plished through two microprocessor-control led brushless direct-current motors. One of these motors is connected directly to the take-up hub; the other to a drive hub designed tO inter face to the CompacTape Cartridge .
The engagement of the cartridge hub with the drive motor shaft is accomplished by a pair of gears that transmit torque and simul tane ously center the reel (Figure 8) . A plastic hub with one set of teeth is attached to the spindle; another set of teeth is molded on the underside of the cartridge reel hub. A clutch gear engages both sets of teeth to drive the reel . To facilitate the insertion or removal of the cartridge, the clutch gear is axially retracted out of engage-
New Products
Digital Technical journal
No. 2 March 1 986
Fig ure 8 TK50 Door A ssem bly
9 1
The TK50 Cartridge Tape Drive ment. The clutch gear is activated by the opera tor's lowering or raising the hand le. When the handle is lowered, the spring-loaded lower gear engages the reel and lifts it slightly into the cartri dge to e l i m i nate contact between the rotating reel and the stationary she l l ( Figure 8) .
This clutching arrangement has a big advan tage because it a llows mechan ical simplicity and easy operation of the drive . The cartridge is i nse rted by t h e operator i n to a c h a n n e l
( receiver) that puts t h e two leaders i n to a coplanar relationship. The entire l i nking pro cess is thus accomplished by merely sl iding the cartridge i nto the receiver slot. A solenoid-acti vated interposer locks the cartri dge in place when i t reac hes the end pos i t i o n i n the receiver. When the front handle is then low ered, the drive gear rises to mate with the car tridge ree l . A set of fingers s i m u l taneously enters the bottom of the cartridge to release the reel locks, thus al lowing the tape to move . The operator accompl ishes all these actions with one hand .
After a tape cartridge is i nserted i nto the
TKSO drive, the operatOr presses a buttOn and the 80 5 1 m icroprocessor on the printed circuit assembly i nitiates the thread ing process . The ree l motors , under m i c roprocessor contro l , slowly p u t tension o n t h e tape t o accomplish the process . The buckled leaders and a length of tape are pulled through the drive and onto the ta ke-up ree l . Auto-threading is complete when the BOT hole in the tape is detected by a photo-transistor. When the auto-threading oper a t i o n ends, the m i c roprocessor w i l l have received pulses from a tachometer attached to one of the rotating tape gu ides. Through the information derived from the tachometer, the microprocessor can mai ntain proper tens ion and tape speed.
After the tape is positioned at BOT, the con troller requests a cal ibration procedure. This procedure sets up the d r ive to ensure that proper va lues for the read circu itry gai n and head stepper alignment are obtained. This cal i bration provides o n e o f the key features o f the
TKSO su bsyste m : the a b i l ity of a user to exchange media between different TKSO tape drives without the need for adjustments .
Once calibrated and at BOT, the TKS O drive is ready to read or write data . The drive writes data i n a serpentine fas hion over the entire length of the tape . The upper part of the readjwrite head writes data o n one track down the entire length of tape until it reaches a logi cal EOT marker . (The logical EOT marker is a preset tachometer count; the phys ical EOT marker is a hole in the tape .) The tape d i rection i s then reversed and the other lower write core will write data i n the other d i rection for the entire length of the tape u nt il a logical I30T is reached . The d i rection of the tape is then changed to forward, the head is stepped u p by
1 9 mils, and the upper wri te core is again used to write data . Figure 9 i l l ustrates the physical tape configuration .
Dri1 ·e Circuil!J>
The printed c i rcuit board assembly is bu il t around an 805 1 m icroprocessor. The 805 1 and associated circui try provide the i ntelligence to i nterpret commands, provide servo control for the ree l motors , perform tape calibration proce du res, and monitor various status inputs. The readjwrite c i rcuits necessary to translate data to and from the tape ' s MFM format a lso res ide on the board . F igure 1 0 i l lustrates a simpl ified b lock diagram of the TKSO drive board .
Write data comes i nto the drive' s logic board via the d ifferential signa l cable from the con troller board . The data enters the shift register, which accepts the serial data and outputs a five-bit parallel data pattern into a program mable array logic ( PAL) devi ce. The data is c locked through the shift register by a 500-KHz clock. ( 5 00 KHz is the write pu lse rate , or data rate.)
The PAL first accepts the five parallel bits from the shift register. Then the PAL generates the pre-compensation, as requ ired, and trans lates the data i nto the MFM format recorded on the tape . A constant current sou rce of 1 5 m i l l iamps i s applied a lternately t o each core o f the active write head, resu lting in the flux transi tions necessary t o write data on the tape .
To enhance data re liabil i ty , the TKSO subsys tem reads data just after writing i t . This tech nique uses the read head (pos i t ioned i mmedi ately behind the write head) to read the data from the tape as soon as it has been written .
(See Figu re 7 . )
The read data i s sent back t o the controller, where the com munications interface performs
CRC processing. If an error is detected, the con trol ler rewrites the block that conta ined the error. The rewritten block is placed farther
9 2
Digital Technical journal
No.
2
March
/ 986
LOG I C A L
B E G I N N I N G
O F T R A C K
( F O RWA R D)
BOT
H O L E
B
R
A
T
I c
A
I
L
0
N
B
A
N
D
G u
A
R
D
I
9 1 4
I
MM
1 1 3 FT)
I
I
I
I M I N
I
L E A D E R
E N D
6 1 0
MM
1 2 FT)
M I N
I G E
I l o
I
I
I
A
I N
I D
T
E
N s
I
0
N
1 305 j M M I
I
1 1 1 F T l 1
I
I M I N
I
DATA A R EA
1 83 M ET E R S
1600 F E ET )
N O M I NA L
R E F E R EN C E E D G E
LOG I CA L
B EG I N N I N G
O F T RA C K
( R E V E RS E )
1 2 1 9
MM
( 4 FT)
M I N
1
1 305
I
I M M I
1
1 1 FT)I
I M I N I
I 9 1 4
I
I MM I
I
I
I
G u
E
I u
A T l A
R E
D N
I
G
R
D
B
A s l
0 1
N
N l
D
I
B
A
N
D
EOT
H O L E
9
7
1 7
1 5
1 3
1 1
5
3
4
2
21
1 9
8
6
1 2
1 0
20
1 8
1 6
1 4
N u
M
B
E
R
T
R
A c
K
H UB
E N D
Figure 9 Physical Tape Configuration
Digital Technical journal
No. 2 March 1986
93
New Products
94
The TK50 Cartridge Tape Drive
M
ISCELLANEOUS
¢:>
SE NSE co
NTROL
TACHOMETER
PLS-L
SERIAL COMMAND
GAP
-L
--
8 7 5 1
INTO
VCO ENABLE
R E A D ENABLE vco
+
DATA
SEPARATOR t
MICROPROCESSOR SENSE
...
(AMPLITUDE. TRACKING)
ENABLE
--
INTI
MICRO-
PROCESSOR
�
"
ENABLE
:
�
B U S h
D U A L
DAC
8 X 2 5 6
RAM
NRZ READ DATA
R E A D CLOCK
�
R E A D
FIL TEA'S
G A I N
CONTROL
K:J
S E R I A L
M A
1--
I-r-
L[>-8' H E A D
600+ F E E T
OF TAPE
S T A T U S
(ECHO)
-
FORWARD
+
-B A C K W A RD
CHANNEL
WRITE GATE
MISCELLANEOUS
SENSE
CONTROL
(22)
WRITE DATA (NRZ)
WRITE/ERASE ENABLE
SERVO
TIMER f
)--
ENCODER
WRITE
COM PEN-
SATION
(PAL'S)
ERASE GAP
24 MHZ
__j
Figure 1 0
Block Diagram of the Drive Board
WRITE CLOCK
Digital Tecbnical]ournal
No.
2 March 1986
down on the tape tO avoid the performance loss resu lting from the drive' s having to move the tape back and rewrite over the data block con taining the error. The controller firmware is able to detect these rewritten blocks duri ng a subsequent read pass for data recovery proce dures, thereby e nhancing system integrity.
Read data signals from the read head are fed tO the differential preamplifier circuit and in turn to the read amplifier. The gain of the preamplifier is automatically set during cal ibra tion to maintain an optimum signal level . The signal from the read ampl ifier is then passed to a differentiated , l inear-phase, low-pass filter. A zero-crossing detection c ircu i t produces a digital signal , consisting of a single pulse for each detected zero crossing, that represents data read from the tape .
The digital data is then sent to the phase lock loop (PLL) circuit where the clock signal is recovered and the MFM data is decoded . The
PLL consists of two PALs , a voltage-controlled oscillator, and some analog circuitry.
The read-data pu lse from the read amplifier circuit is used in conjunction with the 5 00-KHz write clock to optimize the " lock time" for the
PLL. Whenever there is a gap (no signal) going i nto the PLL, it will lock onto the 50 0-KHz clock signal . This locking is done so that the loop-fi l ter integrating capacitor is kept at a con stant voltage . This process minimizes the phase lock time duri ng the preamble .
When the READ ENABLE signal is asserted, the
PLL waits for the synchronization (sync) bit.
When the PLL detects the transition, it clocks the sync bit and data onto the serial line to the controller and starts sending back the read clock. The sync bit signals the communications processor on the controJler to start processing the following data and the CRC c heck-word , and to check for a matching CRC.
Q- bus Co n troller
The inte l ligent interface between the TK50 tape transport and the Q-bus is designated as the TQK5 0 . Figure 1 1 is a block diagram of the
New Products
0- B U S
I
I
0- B U S
,.-----
-
M 7 546 0 - B US C O N T R O L L E R
I N T E R F A C E
I
I p�F-;;---- -
M I C R O P R O C E SSO R
I I
R A M
I
--
R O M
I
I
!"
R I V E I N T E R F A C E
U A R T
I
I
I
T l
�,
I
I
I
D R I V E
T R A N S C E I V E R S
.4 �
I
I
I
1 ...
I I ...
T O
D R I V E
F R O M
D R I V E l__
_ _ j
Figm·e 1 1 Block Diagram of the TQK50
Digital Technical journal
No. 2 March 1986
9 5
The TK50 Cartridge Tape Drive
TQK 5 0 . The interface is a Q-bus-compatible dual board based on the 80 1 86 microprocessor.
In conjunction with 32 kilobytes ( KB) of highly complex firmware , the 80 1 86 and its associ ated hardware perform the following functions :
•
Interface the controller to the Q-bus (via sin gle-word and DMA transfers)
•
•
Translate and process TMSCP command packets and responses
Provi d e data format and error recovery processing
•
Control the general operation of the tape transport mechanism
•
Support the serial data link between the con troller and drive
Hardware
The Q-bus interface is controlled primarily by an 80 1 86 microprocessor and an 8 2 S I 0 5 field programmed logic sequencer (FPLS) , which is a high-performance LSI device capable of per forming complex logic functions . Using the
82S 1 05 FPLS sequencer allowed us to create an efficient, flexible design in a very small space.
The FPLS and microprocessor are responsible for maintaining the strict Q-bus protocol du ring
DMA transfers to and from the controller. The
DMA transfe rs and interface i nterrupts are processed very qu ickly due to the high per formance of the microprocessor and FPLS. This h igh performance makes possible the data rates needed to su pport tape streaming and lessens the criticality of the DMA latencies in the host system.
Assisting the FPLS is an 80 1 86 m icroproces sor operating at 6 MHz. The 80 1 86 is a highly complex, 1 6-bit microprocessor; it is responsi ble for all the command, control , and data processing for the TQK50 . A microprocessor with the 80 1 86's performance is required due to the large number of complex tasks that must be performed within very short time frames
(e. g . , ECC processing during inter-block gaps on tape) . The high level of integration available with the 80 1 86 was a key factor in its selection .
I n addition to the CPU, the 80 1 86 contains three onboard timers, an interrupt controller, address decoding, and two DMA channels. Also important in the selection of the 80 1 86 was the availability of sophisticated development tools and efficient software su pport packages.
The 80 1 86 microprocessor is su pported by nu merous components that include SSI , MSI and PAL devices . Furthermore , the program store and the workspacejdata buffers are pro vided by 1 2 8-kilobit (Kb) EPROMs and 64Kb static RAMS . A total of 3 2 KB of program store and 1 6KB of bu ffer is available to the 80 1 8 6 .
Communications between t h e TQK50 con troller and the TK50 tape transport take place over a pair of fu ll-duplex, differential, serial lines . A multiprotocol commu nications proces sor (NEC 7 2 0 1 ) is used to process the serial-to parallel and parallel-to-serial conversions. One fu l l - d u pl e x c h a n ne l , o p e ra t i n g a t 1 8 7 . 5 kilobaud , communicates the commandjstatus information between the controller and the transport. The other channel provides the data communications path, su pported by data-link error checking via CRC- 1 6 . This second chan nel operates synchronously at 5 00Kb per sec ond . The NEC 720 1 communications chip sup ports DMA transfers to and from the 80 1 86 and operates in a priority-interrupt mode.
Firmware
The most complex component of the TK50 sub system is i ts firmware . The 32KB of firmware contained in EPROM are partitioned into five major functions:
•
The PORTjQ 2 2 (Q-bus) for data transfer control
•
The SERVER for TMSCP command processing
•
T h e TOS fo r t a p e t r a n s p o r t c o n t r o l and formatting
•
The ECC for error detection and correction
•
The ROD for resident onboard diagnostics
The PORTjQ 2 2 firmware controls data trans fers between the controller and CPU, and a lso maintains t he command queue processing. Up tO fou r TMSCP commands can be q u e u e d , allowing t h e host t o s e t u p a series o f opera tions for execution whi l e it continues with other processing. DMA transfers of up to 64K- 1 bytes can be made, allowing an effective, low overhead data transfer between the subsystem and CPU memory space.
The SERVER firmware is responsible for trans lating and execu ting the wide variety of TMSCP commands . These commands provide a very structured environment within which control,
96 Digital Technical]ournal
No. 2 March
1986
status, and data transfers are accomplished .
TMSCP is a packet protocol that uses a com mand-response sequence . Each pair of com mand-response packets conta ins information pertaining to the internal command as wel l as various command modifiers, status fields, and subsystem parameters . All levels of information , from the command sequence number to com mand status to hardware and firmware revision levels, are provided in TMSCP. In addition to assembling and processing this information, the
SERVER firmware uses values, such as physical and logical record numbers, to val idate infor mation being processed from the tape .
SERVER has an additional mode that supports the Diagnostic Uti l ity Protocol (DUP) . DUP provides a set of commands that allow detailed tests of the subsystem to be performed . DUP operates in conjunction with the resident on board diagnostic modu le .
The TOS (tape operation support) firmware controls the transfer of data between the tape transport and the buffers al located by SERVER.
This control is accomplished through format ting operations and through physical control of the tape transport mechanism.
The TK50 subsystem is a streaming tape drive that was designed to operate in an efficient block-mode environment. The TK50 subsystem relies on logical information written on the tape to determine the tape's physical and logi cal positions. The physical and logical contexts are maintained by the TOS firmware and writ ten into special control fields embedded in the
TK5 0 tape format. I nformation contained i n these fields includes physical object number, logical object number, tape-mark number, byte count, sequence control number, track num ber, and block typ e . This i nformation is processed by TOS to maintain the physical and logical contexts between the subsystem and the data on the tape .
D u r i n g s t r e a m i n g o p e ra t i o n s , c o n t e x t processing i s the primary function o f TOS .
However, when the host system i s u nable to process data at a sufficient rate to maintain the streaming operation ( 4 5 KB per second) , TOS must provide complex positioning contro l .
Whenever t h e host system fal ls below the required data transfer rate , TOS must stop the tape. Since the TK50 subsystem was mechani cally optimized for streaming, any stoppi ng and starting of the tape is a time-consuming and imprecise operation. Moreover, the TK5 0 subsystem lacks the inter-record gaps that are used for positional information in traditional 9-track tape drives. The TK5 0 subsystem must rely on data read from the tape to locate its position.
When the host system resumes data process i n g , TOS m u st repos i t i o n the tape by a sequence of reverse, stop, forward , and read.
After locating the last data block processed on the tape , TOS continues with the host's request .
The host's failure to process data at a sufficient rate is costly in terms of system throughput.
This situation requires i ncreased complexity in the subsystem design .
TOS provides a padding function to help compensate for insufficient host processi ng power. With padding, TOS allows data latencies of up to 63 milliseconds before reverting to the repositioning mode . During this data latency period, pad blocks are written to the tape in
9-millisecond increments . That allows the tape to c o n t i n u e strea m i n g . T h e t r a d e - off i s i m p roved performance a t t h e expe nse of slightly reduced tape capacity ( 5 1 2 bytes per pad block) . If the 6 3 -m i l l isecond period is exceeded, TOS stops and performs a reposition to the point of the last data block. When addi tional data arrives, TOS overwrites any previ ously written pad blocks . In practice, this pad function enhances performance and seldom reduces tape capacity by more than ten percent.
The ECC firmware provides the means to detect and correct errors . To provide a high level of reliability, the TK50 subsystem is designed to allow only one unrecoverable error in every 1 X 1 0 1 1 bits read . This is equivalent to one u nrecoverable error i n every 1 2 5 cartridge reads . To achieve this goal , ECC implements error-detection and error-correction schemes.
Error detection is based on the CRC- 1 6 method , which is supported by the hardware commun i cations device . This i ndustry-standard method has been proven to be very efficient in this environment.
To implement the error-correction function,
ECC processes serial-formatted data to and from the tape . Data is written to and read from the tape in 5 1 2 -byte blocks. Each block is grouped into 8-block units, called data entities . Within an entity, the four even-numbered data blocks
(0 , 2 ,4 ,6) and the four odd-numbered blocks
( 1 , 3 , 5 , 7) are protected by longitudinal check sum blocks .
An entity, therefore , consists of ten blocks : data blocks 0 through 7 and ECC blocks
Digital Technicaljournal
No. 2 March
1 986
97
New Products
·
-------
The TK50 Cartridge Tape Drive
(5 1 2 BYTE BLOCKS)
Figure 1 2 Entity of Ten Blocks
8 and 9 . Figure 1 2 shows the arrangement of the ten blocks .
This technique, coupled with record-level checking by SERVER and the host operating sys tem, i nsures the complete i ntegrity of the user's data .
The ROD ( resident on board diagnostics) firmware provides additional support for the
TQK5 0 . When the subsystem is initialized , the firmware executes a series of gojno-go tests that validate the functionality of the controller.
Ninety-eight percent of the TQK50 ' s function ality is covered by these tests, excluding the
Q-bus and drive-interface logic circu its . More extensive diagnostics that fully test the TK5 0 subsystem are available under the DUP. Having the diagnostics resident in firmware allows the running of i ntegrated tests that i nteract at levels not permitted from the system interface . That avoids the difficulties in supporting down-line l o a d a b l e c o d e i n v a r i o u s r u n - t i m e environments .
Summary
Designing the TK5 0 cartridge tape subsystem and turning it i nto a product was a significant challenge . The effort proves that good perform ance , h i g h r e l i a b i l i t y , ease of u s e , a n d extraordinary data integrity can b e achieved in a cost-effective manner. These qualities will continue to be required as computer systems increase in performance and capacity.
To that end, the TK5 0 cartridge tape subsys tem is but the first of a family of cartridge tape products. Work is continu i ng on the develop ment of subsystems with h igher performance and greater capacities. Interfaces to computer systems other than those based on the Q-bus have been or are being developed to meet the expanding needs for greater storage capacity.
Acknowledgements
Designing the TK5 0 cartridge tape subsystem required a multitude of disciplines involving scores of individuals. Each member of the TK5 0 program team contributed time, energy, and personal commitment to yield a successfu l product. The authors wish to acknowledge those contributions here .
98
Digital Technicaljournal
No. 2 March
1 986
Raymond]. Lanza
I
Porting UL TRIX
Software to the
Micro VAX System
The ULTRIX system, written in the
C
programming language, was ported to the Micro VAX II processor by a multistep process. This involved estab lishing a cross-development environment, building a bootpath, porting the ULTRIX kernel, and writing special device drivers. The remaining software was ported after those steps were completed. To minimize
ULTRIX design changes, the system's IjO architecture was mapped into the Micro VAX physical address space so as to mirror the equivalent mapping on larger VAX systems. Some Micro VAX instructions must be emulated in macrocode. The emulator used in the Micro VMS software was adapted for use in this ULTRIX software.
The UNIX system came i nto existence in 1 969 at the AT&T Bell Laboratories in Murray H i l l ,
New Jersey. The initial system was written i n assembler and ran on a PDP· 7 system that was l oaded from paper tapes. From l ate 1 970 to early 1 9 7 1 , the UNIX software was reimple mented for the PDP- 1 1 system using a cross assembler running on the original PDP- 7 sys tem. In 1 9 7 3 , the kernel was rewritten in the C programming language . Since that time the sys tem has undergone many changes and is stil l the subject of much research. 1 Today, there are two major 3 2 -bit variants of the original software : 4 BSD, developed at the University of
California at Berkeley; and System V , from
AT&T Corporation . Digita l Equipment Corpora tion's original U LTRIX-3 2 product is a direct descendant of 4 . 2BSD.
In 1 9 83, Digital decided to develop and dis tribute a UNIX software product. At that time,
4 . 2 BSD was the only virtual-me mory UNIX operating system ru nning on VAX processors. I t i s Sti l l the only UNIX software derivative to pro vide network support. These features were the key factors i n deciding to use 4 . 2BSD as the basis of the ULTRIX- 3 2 system.
Deve lopment started in the fal l of 1 9 83 on one of the first 4 . 2 BSD distributions, and the final product was re leased in April 1 9 84 as
U LTRIX- 3 2 V l . O . In the current version of the product, we have combined the two UNIX sys tem derivatives by adding the system services of the AT&T version to the original U LTRIX- 3 2 sys tem . To that base we have added reliability and maintainability features , as wel l as new-proces sor support . The resu lting system , one of the industry's most powerful and versatile UNIX software versions, spans the ful l VAX system pricejperformance range .
Porting the
UNIX
System
" Porting" is the process of implementing an operating system on a new processor. The UNIX system has been ported to more processors than any other system in existence . It runs on all c lasses of machines , from 8086 m icroproces sors to the CRAY- 2 . For VMS and RSX systems and the like , port ing normal ly means a major rewrite because significant parts of them are written in low-level languages, usually macro assembler. Rewriting one of these systems is so expensive that either the effort wou ld not be undertaken or the new system would be written from scratch .
The UNIX system is different. I t i s written in a single high-level language, C ,2 and has been
Digital Technical }ourn .
al
No. 2 March
1986
99
structured to be as processor independent as possible. However, vestiges of i ts PDP- 1 1 heri tage are still apparent.
All 3 2 -bit versions of the ULTRI X- 3 2 system are built from a common set of sou rce files. The kernel fi les are organized into machine-depen dent and mac hine-independent parts. The dif ferences between the VAX and MicroVAX ver sions of the system are resolved through the use of conditional compilation and l inking. The present kernel sources for the MicroVAX ver sion are as fol lows :
Files
209
31 5
21
Language
C headers
C source
Assembler source
The 2 1 assembler source files can be further broken down as follows:
Files
1 4
3
3
Purpose
Mi croVAX su bset and floating point emu lator
Te mplates for rpb,scb, spt
Macro defin itions
I n itial startup code
(locore .s)
The l ast and most significant file is locore .s, which contains the i nitial startup code and a few critical rou t i n e s needed for process management.
Bri nging the UNIX system up on a new processor is normally done i n multiple steps by a sma l l team . The difficulty and extent of the work i nvolved is directly related to the archi tectural differences between the versions for the existing and target processors. Our team consisted of three people, later joined by a fourth . The first was responsible for the com piler and subset emulator. The second did the software i nstallation and verification for the first version of the product. Later , he was responsible for some device drivers. The author of this paper d id the kernel port and other device drivers . The fou rt h person assumed responsibility for installation.
Bringing the ULTRIX - 3 2 system up on a processor involves the following steps :
1 . E s t a b l i s h a c r o s s - d eve
1 o p m e n t environment.
•
C language
•
•
•
Native assembler
Linker
Debugger
2 . Bu ild a boot path .
3 . Port t he kernel and a few key programs .
4 . Write special device drivers .
5 . Port the rest of the system .
The Cross Develop ment Enviro nment
When porting to a new architecture , it is neces sary to develop a set of tools that produces code for the target system . These tools constitute a cross-development system for software genera tion and often become the basis for the even tual native environment. Their construction is normally the first step in the porting process . In the M i croVAX case, the cross-deve lopment tools were not necessary, for reasons explained below.
The MicroVAX system is a subset architecture with the majority of the string manipu lation i nstructions missing.
3
MicroV AX systems can also be configured without floating point sup port in the hardware . Our c hallenge, which was a lso shared by the VMS and VA.XELN Develop ment Groups, was tO provide an execu tion environment for user programs that was com p letely compatible with larger VAX systems .
By closely examining the instructions pro duced by our C compiler, we found that, with the exception of the floating point i nstructions, not one missing string instruction was created.
Further examinations revealed that the only place where any of the missing i nstructions were used was in a handful of output formatting routines.
As an i nterim solution, the affected routi nes were rewritten to elim inate the mi ss i ng instructions.4
The Boot Path
MicroVAX systems contain the virtual memory boot (VMB) program in ROM . Normally this program loads the VMS system but has been enhanced to perform an alternate i nitial pro gram load operation, called a boot-block boot.
1 00 Digital Technical ]oun�al
No.
2 March 1986
This operation is the mechanism used to boot the ULTRIX system and is based on block num ber 0 of the boot disk being i n a special format.
Booting is a m u ltistage process.
1 . VMB first checks for an ODS- I I file struc ture s In the default case, VMB will per form a "sniffer boot , " which consists of first checking the removable media, then the fixed disks, and finally the Ethernet.
The system can also be booted from the
TKSO cartridge tape drive and a special
PROM board .
2 . I f an ODS-II file structure is not present,
VMB looks for a valid boot-block image in the first block on the disk . This block contains a table that specifies the size and location o f the secondary boot image . If the table is valid, VMB reads the secondary boot image i ntO memory and transfers control to the image . (If the table is inval i d , control is transferred back to step 1 above.)
3. The secondary boot image on ULTRIX systems is a program that locates, reads, and executes the tertiary boot program from an ULTRIX file system. The func t i o n a l i ty of the s e c o n d a ry b o o t i s severely constrained because it resides outside the file system in a fixed-size
(7 . S KB) area adjacent to the boot block.
4. The tertiary boot is capable of loading and running other programs . Unl ike the secondary boot program , it su pports interactive terminal I/0 and can prompt the user for an alternate program to load .
As a default, the tertiary boot loads the operating system kernel , called vmunix,6 from the boot disk.
S. After the steps above have been com pleted , the kernel i s i n memory and ready to run .
The two boot programs are part o f the stand a lone system , which in itself constitutes a port i ng problem that is not very different from port ing the kernel . The problems encountered are similar, although simpl ified, because the stand alone system runs with the interrupts and mem ory management d isabled . The stand -alone sys tem is not nearly as flexible as the kerne l .
Porting The Kernel
The VAX Architecture Standard ( D igital Stan dard 0 3 2) specifies the VAX instruction set, memory management, and process environ ment. However, the standard leaves many other areas open for change . These areas are typically ones that need to be supported on each new processor. For the M icroVAX system , it was necessary to address problems in the fol lowing areas :
•
•
Startup code
I/0 architecture
•
•
Console su pport
System clock
•
Missing instruction emulation
Initial Startup Code
After the kernel is loaded into memory, control is transferred tO the initial startup code. This is entered with the processor interrupt priority
" raised " to disable the interrupts, and with memory management tu rned off. The code sets up the memory management system and then
"handcrafts" the processor to run the first VAX process. The majority of this code is located in a single assembly language fil e , called locore .s.
In the case of the MicroVAX system , the instruc tion emu latOr and several changes tO the I/0 system requ ired special mapping support dur ing startup. (This su pport is discussed i n the l ast section of this paper.)
In addition to the startup code, locore .s con tains time-critical routines that use the VAX process-management i nstru ct ions . Some of them contain a easel instruction based on the processor type for processor-specific opera tions . Those routines had to be extended to include the MicroVAX processors .
1/0 A rchitecture
VAX processors do not contain I/0 instructions; instead , device and device adapters exist in various sections of the physical address space of the processor . The control and data registers for these adapters appear as memory locations and are accessed using normal instructions . A key e lement of system software for any new proces sor is support for these devices and their associ ated address spaces.
As an example, the physi cal address space of the VAX - 1 1 /780 system is pictu red in Figure l .
New Products
Digital TecbnicaljoUJ-nal
No. 2 March 1986
1 0 1
------
Porting UL TRIX Software to the Micro VAX System
PHYSICAL ADDRESS
0000 0000
FUNCTION
MEMORY
1 FFF FFFF
2000 0000
2000 2000
2000 4000
2000 6000
TRACKO
TRACK!
TRACK2
TRACK3
8KB
8KB
8KB
8KB
ADAPTER OR
NEXUS REGISTER
ADDRESS SPACE
2001 cooo
2001 EOOO
TRACK 1 4
TRACK I S
8 K B
8 K B
2 0 1 0 0000
2014 0000
20 1 8 0000
20 1 C 0000
UNIBUSO ADDRESS SPACE
UNIBUS! ADDRESS SPACE
UNIBUS2 ADDRESS SPACE
UNIBUS3 ADDRESS SPACE
1 28K RESERVED
256K EACH
Figure 1 VAX- l lj780 Physical Address
Space
Each of the UNIBUS spaces can be further broken down as shown in Figure 2 .
T h e p h ys i c a l a d d r e s s s p a c e o f t h e
MicroVAX I I system i s somewhat si mpler, as depicted in Figure 3 .
With the exception of the memory sections, the address spaces of the two processors appear to be very different. I n fact there are a su rpris i ng number of similarities, as shown in Table 1 .
The NEXUS space is where the adapter con trol and status registers reside. In the case of a
UNIBUS adapter, the registers that control the mapping from the bus to main memory are located in the NEXUS space. The equivalent
MicroVAX area, called local register space , also contains the mapping registers for the Q-bus to main memory.
These physical address spaces are eventually mapped into virtual addresses through entries in the VAX Page Tabl e . The result is pictured in
Figure 4 .
One development goal that we set for each new processor support project is to minimize the changes necessary in the operating system .
I n the case of the MicroVAX I I syste m , we exa m i ned the d i fferences i n the physical address spaces between that system and larger
VAX systems. Although the names, sizes, and positions were different, they are functionally equivalent on both the small and larger sys tems.
As a result, we "coerced" the local regis ter space i nto the NEXUS map , and the Q-bus memory and I/0 spaces were arranged to look l i ke a large UNI B US adapter. Wi th this approach we were not forced to drasti cally alter the ker nel's view of the machine, thus mini mizing changes to other portions of the kerne l .
A similar situation existed with respect to the
Q-bus map . A device installed in a UNIBUS adapter sees an 1 8-bit address for a 2 5 6KB
PHYSICAL ADDRESS
0000 0000
IFFF FFFF
2000 0000
0-BUS
1/0
2008 0000
BKB
FUNCTION
MEMORY
DEVICE REGISTERS
LOCAL
REGISTER SPA
C
E
(256KB)
UNIBUS MEMORY
(248K)
2008
FFFF
3000 0000
O·BUS MEMOAY SPACE
(4MB)
UNIBUS 1/0
8KB
DEVICE REGISTERS
303F
FFFF
Figure 2
VAX- 1 1/780 UNIBUS Space
Figure 3 Micro VA X
II
Physical Address
Space
1 02
Digital Tecb11ical Journal
No. 2 March 1986
Table 1 Comparison of Physical Add ress Spaces for the VAX- 1 1 /780 System and the MicroVAX
II
System
Physical Address Spaces
VAX Function Size
MicroVAX I I
Function Size Purpose of Function
Memory
N EX U S
U N I B U S Memory
U N I B U S 1/0
2 M B-64MB
8K each
248K
8K each
Memory
Local Register
0-bus Memory
0-bus 1/0
1 6M B
256K
4 M B
8K
Execute Programs
CPU and Bus Control
Registers
Device Memory
Device Reg isters
New Products
PHYSICAL ADDRESS
8000 0000
F U NCTION
KERNEL
NEXUS SPACE
UNIBUS
SPACE
VAX
MEMORY
II
I
I
I
/
/
I
0-BUS
MEMORY
I
I
I
BUS
M
AP
�
I
I
I
I
Figure
4
Result of Physical- to- Virtual
Mapping
Fig ure 5 Q- bus Memory Mapping address space . The adapter has a set of registers that maps this 2 5 6KB space onto the much larger VAX memory space . These registers per form the equ ivalent function that is provided by VAX Page Table entries . In effect, they "vir tual i ze" the memory that devices access .
Figure 5 depicts this mapping.
The MicroVAX II system contains a similar set of registers with the principal difference being that it has enough to map all four megabytes of main memory. Although that appears advanta geou s , it in fact posed a serious proble m . The
ULTRIX system dynamically allocates the bus mappi ng registers from a central routine . It wou ld have been easy to modify this rou tine to
" know" about the extra registers. The problem encountered here was that these al location rou -
3 1
BUFFERED DATAPATH
NUMBER
Figure
6 Coding of Allocation Routine
Wo rd
BIT
POSITION
BUS
VIRTUAL ADDRESS tines retu rn a word that is encoded as shown in
Figure 6 .
The upper part contai ns the nu mber of the buffered datapath allocated , the middle is the nu mber of registers use d , and the lower is
Digital Tech-nical ]01n-nal
No. 2 March I
'J86
1 03
the bus virtual address . The format of this
3 2 -bit word is known by all device drivers that do DMA transfers . To change the word to use all the map registers available meant that the vir tual address portion would need 2 2 bits instead of 1 8 . That wou ld have required correspond ing changes in each of the device drivers . To deter mine the severity of these problems, we did some tests to see if the 1 8-bit format would be a limiting factor. Fortu nately, we fou nd that there were always registers availabl e.
The end resu lt of the mapping and map-regis ter al location scheme was that UNIBUS device drivers could be left unchanged as long as the
Q - b u s hardware was compat i b l e w i t h the
UNIJ3US versions . We took advantage of that fact and thus were able to su pport the TSVO S ,
DHV1 1 , and RL02 disk su bsystems without any impact on the development schedu le 7
Console Port
Trad itional VAX systems have a separate proces sor that perfo rms console fun c t i o n s . This processor is used to control the main CPU and replaces the older-style front panel . Instead of having switches for ha lt or ru n , the console ru ns a program that prov i d es h a l t , r u n , examine , and initial program load capabilities .
Programs running in a VAX system can commu nicate with the console through an internal processor register. Com mands sent in this regis ter are used by the operating system to reboot and restart the machine .
The MicroVAX system is different: the con sole functions are handled by the microproces sor, the M icroVAX 7 8 0 3 2 chip, which runs a program resident in ROM . like the larger VAX systems, a register is used to communicate with the console . A code can be placed in this regis ter. When a subsequent HALT instru ction is exe cuted, execution switches to the console pro gram in ROM , which then examines the code in the register 8 In fact the register is actually a memory location in RAM that is backed up by batteries .
The U L TRIX system contains a reboot and halt routine that is accessed by a privileged sys tem cal l . That routine was modified to commu nicate with the console program.
System Clock
The ULTRlX system keeps track of the cu rrent time by coun ting cloc k i nterru pts from the
1 0 ms interval timer. The time is kept in memory as an unsigned integer; it is initialized from the time-of-year (TOY) register during system boot . The time is set by a privileged program through standard system calls and can be read by normal user programs . That set procedure i s normally done b y the system manager using the
DATE command . DATE converts the time from a format of year, month, day, hour, m i nute, and second to the integer format needed by the sys tem cal l .
User e nters : yymmddhhmmss
- set
__,
System converts to :
Integer
<-- read where yymmddhhmmss
=
Year, Month, Day,
Hour, M i nute, Second
The MicroVAX system does not have a TOY register; instead, it has a watch chip backed up by a battery. The chip contains a n umber of counters that correspond to the year, month , day, hour, minute, and second . We could have mod i fied the system cal l or added a new one to explicitly set the MicroVAX TOY clock. That wou l d have avoid e d t h e c o nv e rs i o n t o integer format, given that the user has t o enter date and time information in the format needed by the watch chip. However, it would have meant that we needed two versions of the DATE command, one for existing systems and the other for the MicroVAX system , to use the new format. To avoid that, we borrowed the conversion routines from the DATE command and used them in MicroVAX versions of the sys tem time-setting routine. The irony here is that t h e d a t e is n ow c o n v e r t e d twi c e . The i nteger format is present on either side of the system ca l l .
User enters : System reads: yymmddhhmmss integer yymmddhhmmss
Missing Instru ction Emula tion
As mentioned previously, the MicroVAX hard ware i mplements a subset of the ful l VAX instructi on set. Most stri ng i nstruct ions are missing and are e m ulated i n macrocode i nstead of implemented in hardware . The emu lation code could have been p laced in l i bra ries , where it could be linked with user-level code .
To do that, however, wou ld mean that l inked
1 04
Digital Technical journal
No .
2 March 1986
i mages from other VAX systems wou ld not run on a MicroVAX system, thus violating one of its basic objectives.
Rather than using libraries , we chose to use an emu lator designed by the VMS Development
Grou p and ported that emulator to the ULTRIX system 9 The emu lator l i nks with the kernel and is almost completely invisible to user programs .
It is supported by new traps in the hardware that help to decode each missing i nstruction .
When the kernel or a user program executes one of the missing i nstructions , a trap occurs and the emulation code takes over. That hap pens without changing mode; in other words, if an emu lation trap occurs in a user program, the emu lator is entered in user mode, not kernel mode like other traps . The resu lt is user-mode execution of code in the kernel address space .
(Unlike the VMS system, the entire ULTRIX ker nel is normal ly unreadable by user programs.)
The startup code now initializes the pages con taining the emulatOr so that they can be read and execu ted by user- level code .
As stated earlier, the end result is a combina tion of hardware and software that is almost completely compatible with systems running the fu l l VAX i nstruction set . I n fact, executable images from other VAX systems can run without re linking. The only point of i ncompatibility is that the emu lation code runs on the user stack when one of the m iss ing i nstructions is exe cuted by user code . (We have seen one cus tomer appl ication that was affected by this situ ation. The application used knowledge of its past usage of the stack to do "garbage collec tion" and was confused by the intermediate results of the emulation code . That is norma l ly not a problem; the ULTRIX- 3 2 and ULTRIX-
3 2 m kits have over 500 user-level programs .
They are compiled and linked once on a fu ll
VAX system and then ru n without modification on the MicroVAX syste m .)
Summary
In porting the ULTRIX system to the MicroVAX processor, we opted to maintain compatibi l i ty with other versions of the system , wherever possible. We choose not tO su pport hardware features if they violated internal or external interfaces . The refore , we were able ro deliver a broader range of peripheral support with a min i m u m of developme n t . The end res u l t-the
M i croVAX syste m - com b i nes ha rdware and software to provide customers, including devel opers of software device-d rivers, with a product that runs all VAX programs for a fraction of the cost of a larger VAX system.
References
1 . A detailed h istory of and supplemental information about the UNIX system can be found in the
A T& T Bell Laboratories
Te c h n ical j o u rnal, v o l . 5 7 , n o . 6
Ou lyjAugust 1 9 78) and vol . 6 3 , No. 8
(October 1 984) .
2 . Some programmers consider C to be a low-level language ; i n fact, i t has proven ro be more than adequate for program m i ng an operati ng system l ike the UNIX syste m .
3 . D.W . Dobberpu hl et a l , "The MicroVAX
7 8 0 3 2 Chip, A 32 -Bit Microprocessor,"
Digital Technical journal
(March
I
986 , this issue) : 1 2 -23
4 . This work was done long before the first hardware protOtype was developed .
5 . ODS-U is t he VMS on-disk fi le structure.
6 . The AT&T versions call this fi le "u nix ," w h i l e the Be rke le y versions call i t
"vmunix, " denoting "virtual u n i x . "
7 . However, we did have tO expend time and energy tO do additional configura tion testing.
8 . The MicroVAX 7 8 0 3 2 chip never hal ts; it is ru nning ei ther ROM console code or programs in
RA.t\1 .
9 . K . D . Morse and L.] . Kenah, "The Evolu tion of I nstruction Emu lation for the
Mi croV AX Systems , "
Digital Technical journal
(March 1 986, this issue ) : 76-8 5 .
New Products
Di_v,ital Technical ]OIII"Illll
Nv
.
.! Marcb I 'J86
1 05
Letöltés
Hirdetés