NVIDIA GPU Programming Guide

NVIDIA GPU Programming Guide
Version 2.4.0
1
Notice
ALL NVIDIA DESIGN SPECIFICATIONS, REFERENCE BOARDS, FILES, DRAWINGS,
DIAGNOSTICS, LISTS, AND OTHER DOCUMENTS (TOGETHER AND SEPARATELY,
“MATERIALS”) ARE BEING PROVIDED “AS IS.” NVIDIA MAKES NO WARRANTIES,
EXPRESSED, IMPLIED, STATUTORY, OR OTHERWISE WITH RESPECT TO THE
MATERIALS, AND EXPRESSLY DISCLAIMS ALL IMPLIED WARRANTIES OF
NONINFRINGEMENT, MERCHANTABILITY, AND FITNESS FOR A PARTICULAR
PURPOSE.
Information furnished is believed to be accurate and reliable. However, NVIDIA
Corporation assumes no responsibility for the consequences of use of such
information or for any infringement of patents or other rights of third parties that
may result from its use. No license is granted by implication or otherwise under
any patent or patent rights of NVIDIA Corporation. Specifications mentioned in this
publication are subject to change without notice. This publication supersedes and
replaces all information previously supplied. NVIDIA Corporation products are not
authorized for use as critical components in life support devices or systems without
express written approval of NVIDIA Corporation.
Trademarks
NVIDIA, the NVIDIA logo, GeForce, and NVIDIA Quadro are registered trademarks
of NVIDIA Corporation. Other company and product names may be trademarks of
the respective companies with which they are associated.
Copyright
© 2005 by NVIDIA Corporation. All rights reserved.
HISTORY OF MAJOR REVISIONS
Version
Date
Changes
2.4.0
07/08/2005
Updated cover
Added GeForce 7 Series content
2.3.0
02/08/2005
Added 2D & Video Programming chapter
Added more SLI information
2.2.1
11/23/2004
Minor formatting improvements
2.2.0
11/16/2004
Added normal map format advice
Added ps_3_0 performance advice
Added General Advice chapter
2
2.1.0
07/20/2004
Added Stereoscopic Development chapter
2.0.4
07/15/2004
Updated MRT section
2.0.3
06/25/2004
Added Multi-GPU Support chapter
NVIDIA GPU Programming Guide
Table of Contents
Chapter 1. About This Document ..................................................................9
1.1.
Introduction ..............................................................................9
1.2.
Sending Feedback.................................................................... 10
Chapter 2. How to Optimize Your Application ............................................11
2.1.
Making Accurate Measurements ................................................ 11
2.2.
Finding the Bottleneck.............................................................. 12
2.2.1.
Understanding Bottlenecks
12
2.2.2.
Basic Tests
13
2.2.3.
Using NVPerfHUD
14
2.3.
Bottleneck: CPU ....................................................................... 14
2.4.
Bottleneck: GPU....................................................................... 16
Chapter 3. General GPU Performance Tips .................................................17
3.1.
List of Tips .............................................................................. 17
3.2.
Batching.................................................................................. 19
3.2.1.
3.3.
3.3.1.
3.4.
Use Fewer Batches
19
Vertex Shader.......................................................................... 19
Use Indexed Primitive Calls
19
Shaders .................................................................................. 20
3.4.1.
Choose the Lowest Pixel Shader Version That Works
20
3.4.2.
Compile Pixel Shaders Using the ps_2_a Profile
20
3.4.3.
Choose the Lowest Data Precision That Works
21
3.4.4.
Save Computations by Using Algebra
22
3.4.5.
Don’t Pack Vector Values into Scalar Components of Multiple
Interpolants
23
3
3.4.6.
Don’t Write Overly Generic Library Functions
23
3.4.7.
Don’t Compute the Length of Normalized Vectors
23
3.4.8.
Fold Uniform Constant Expressions
24
3.4.9.
Don’t Use Uniform Parameters for Constants That Won’t Change
24
Over the Life of a Pixel Shader
3.4.10.
Balance the Vertex and Pixel Shaders
3.4.11.
Push Linearizable Calculations to the Vertex Shader If You’re Bound
25
by the Pixel Shader
3.4.12.
Use the mul() Standard Library Function
3.4.13.
Use D3DTADDRESS_CLAMP (or GL_CLAMP_TO_EDGE) Instead of
saturate() for Dependent Texture Coordinates
26
3.4.14.
Use Lower-Numbered Interpolants First
3.5.
25
25
26
Texturing ................................................................................ 26
3.5.1.
Use Mipmapping
26
3.5.2.
Use Trilinear and Anisotropic Filtering Prudently
26
3.5.3.
Replace Complex Functions with Texture Lookups
27
3.6.
Performance............................................................................ 29
3.6.1.
Double-Speed Z-Only and Stencil Rendering
29
3.6.2.
Early-Z Optimization
29
3.6.3.
Lay Down Depth First
30
3.6.4.
Allocating Memory
30
3.7.
Antialiasing.............................................................................. 31
Chapter 4. GeForce 6 & 7 Series Programming Tips ...................................33
4.1.
4
Shader Model 3.0 Support ........................................................ 33
4.1.1.
Pixel Shader 3.0
34
4.1.2.
Vertex Shader 3.0
35
4.1.3.
Dynamic Branching
35
4.1.4.
Easier Code Maintenance
36
4.1.5.
Instancing
36
4.1.6.
Summary
37
4.2.
sRGB Encoding ........................................................................ 37
4.3.
Separate Alpha Blending........................................................... 38
NVIDIA GPU Programming Guide
4.4.
Supported Texture Formats ...................................................... 39
4.5.
Floating-Point Textures............................................................. 40
4.5.1.
Limitations
40
4.6.
Multiple Render Targets (MRTs) ................................................ 40
4.7.
Vertex Texturing ...................................................................... 42
4.8.
General Performance Advice ..................................................... 42
4.9.
Normal Maps ........................................................................... 43
Chapter 5. GeForce FX Programming Tips ..................................................45
5.1.
Vertex Shaders ........................................................................ 45
5.2.
Pixel Shader Length ................................................................. 45
5.3.
DirectX-Specific Pixel Shaders ................................................... 46
5.4.
OpenGL-Specific Pixel Shaders .................................................. 46
5.5.
Using 16-Bit Floating-Point........................................................ 47
5.6.
Supported Texture Formats ...................................................... 48
5.7.
Using ps_2_x and ps_2_a in DirectX ....................................... 49
5.8.
Using Floating-Point Render Targets .......................................... 49
5.9.
Normal Maps ........................................................................... 49
5.10.
Newer Chips and Architectures.................................................. 50
5.11.
Summary ................................................................................ 50
Chapter 6. General Advice...........................................................................51
6.1.
Identifying GPUs ...................................................................... 51
6.2.
Hardware Shadow Maps ........................................................... 52
Chapter 7. 2D and Video Programming.......................................................55
7.1.
OpenGL Performance Tips for Video .......................................... 55
7.1.1.
POT with and without Mipmaps
56
7.1.2.
NP2 with Mipmaps
56
7.1.3.
NP2 without Mipmaps (Recommended)
57
7.1.4.
Texture Performance with Pixel Buffer Objects (PBOs)
57
Chapter 8. NVIDIA SLI and Multi-GPU Performance Tips...........................59
8.1.
What is SLI? ............................................................................ 59
5
8.2.
Choosing SLI Modes................................................................. 61
8.3.
Avoid CPU Bottlenecks.............................................................. 61
8.4.
Disable VSync by Default .......................................................... 62
8.5.
DirectX SLI Performance Tips.................................................... 63
8.5.1.
Limit Lag to At Least 2 Frames
8.5.2.
Update All Render-Target Textures in All Frames that Use Them 64
8.5.3.
Clear Color and Z for Render Targets and Frame Buffers
8.6.
63
64
OpenGL SLI Performance Tips................................................... 65
8.6.1.
Limit OpenGL Rendering to a Single Window
65
8.6.2.
Request PDF_SWAP_EXCHANGE Pixel Formats
65
8.6.3.
Avoid Front Buffer Rendering
65
8.6.4.
Limit pbuffer Usage
65
8.6.5.
Render Directly into Textures Instead of Using glCopyTexSubImage66
8.6.6.
Use Vertex Buffer Objects or Display Lists
66
8.6.7.
Limit Texture Working Set
67
8.6.8.
Render the Entire Frame
67
8.6.9.
Limit Data Readback
67
8.6.10.
Never Call glFinish()
67
Chapter 9. Stereoscopic Game Development..............................................69
6
9.1.
Why Care About Stereo?........................................................... 69
9.2.
How Stereo Works ................................................................... 70
9.3.
Things That Hurt Stereo ........................................................... 70
9.3.1.
Rendering at an Incorrect Depth
70
9.3.2.
Billboard Effects
71
9.3.3.
Post-Processing and Screen-Space Effects
71
9.3.4.
Using 2D Rendering in Your 3D Scene
71
9.3.5.
Sub-View Rendering
71
9.3.6.
Updating the Screen with Dirty Rectangles
72
9.3.7.
Resolving Collisions with Too Much Separation
72
9.3.8.
Changing Depth Range for Difference Objects in the Scene
72
9.3.9.
Not Providing Depth Data with Vertices
72
NVIDIA GPU Programming Guide
9.3.10.
Rendering in Windowed Mode
72
9.3.11.
Shadows
72
9.3.12.
Software Rendering
73
9.3.13.
Manually Writing to Render Targets
73
9.3.14.
Very Dark or High-Contrast Scenes
73
9.3.15.
Objects with Small Gaps between Vertices
73
9.4.
Improving the Stereo Effect ...................................................... 73
9.4.1.
Test Your Game in Stereo
73
9.4.2.
Get “Out of the Monitor” Effects
74
9.4.3.
Use High-Detail Geometry
74
9.4.4.
Provide Alternate Views
74
9.4.5.
Look Up Current Issues with Your Games
74
9.5.
Stereo APIs ............................................................................. 74
9.6.
More Information..................................................................... 75
Chapter 10. Performance Tools Overview...................................................77
10.1.
NVPerfHUD.............................................................................. 77
10.2.
NVShaderPerf .......................................................................... 78
10.3.
NVIDIA Melody ........................................................................ 78
10.4.
FX Composer ........................................................................... 79
10.5.
Developer Tools Questions and Feedback.................................. 79
7
8
Chapter 1.
About This Document
1.1.
Introduction
This guide will help you to get the highest graphics performance out of your
application, graphics API, and graphics processing unit (GPU). Understanding
the information in this guide will help you to write better graphical applications,
but keep in mind that it is never too early to send an e-mail to
[email protected] asking for help or advice.
This document is organized in the following way:
‰
Chapter 1(this chapter) gives a brief overview of the document’s contents.
‰
Chapter 2 explains how to optimize your application by finding and
addressing common bottlenecks.
‰
Chapter 3 lists tips that help you address bottlenecks once you’ve identified
them. The tips are categorized and prioritized so you can make the most
important optimizations first.
‰
Chapter 4 presents several useful programming tips for GeForce 7 Series,
GeForce 6 Series, and NV4X-based Quadro FX GPUs. These tips focus on
features, but also address performance in some cases.
‰
Chapter 5 offers several useful programming tips for NVIDIA®
GeForce™ FX and NV3X-based Quadro FX GPUs. These tips focus on
features, but also address performance in some cases.
9
How to Optimize Your Application
1.2.
‰
Chapter 6 presents general advice for NVIDIA GPUs, covering a variety of
different topics such as performance, GPU identification, and more.
‰
Chapter 7 explains NVIDIA’s Scalable Link Interface (SLI) technology,
which allows you to achieve dramatic performance increases with multiple
GPUs.
‰
Chapter 8 describes how to take advantage of our stereoscopic gaming
support. Well-written stereo games are vibrant and far more visually
immersive than their non-stereo counterparts.
‰
Chapter 9 provides an overview of NVIDIA’s performance tools.
Sending Feedback
If you have comments or suggestions for this document, please send them to
[email protected]
10
Chapter 2.
How to Optimize Your Application
This section reviews the typical steps to find and remove performance
bottlenecks in a graphics application.
2.1.
Making Accurate Measurements
Many convenient tools allow you to measure performance while providing
tested and reliable performance indicators. For example, NVPerfHUD’s yellow
line (see the NVPerfHUD documentation for more information) measures total
milliseconds (ms) per frame and displays the current frame rate.
To enable valid performance comparisons:
‰
Verify that the application runs cleanly. For example, when the
application runs with Microsoft’s DirectX Debug runtime, it should not
generate any errors or warnings.
‰
Ensure that the test environment is valid. That is, make sure you are
running release versions of the application and its DLLs, as well as the
release runtime of the latest version of DirectX.
‰
Use release versions (not debug builds) for all software.
‰
Make sure all display settings are set correctly. Typically, this means
that they are at their default values. Anisotropic filtering and antialiasing
settings particularly influence performance.
‰
Disable vertical sync. This ensures that your frame rate is not limited by
your monitor’s refresh rate.
11
How to Optimize Your Application
‰
Run on the target hardware. If you’re trying to find out if a particular
hardware configuration will perform sufficiently, make sure you’re running
on the correct CPU, GPU, and with the right amount of memory on the
system. Bottlenecks can change significantly as you move from a low-end
system to a high-end system.
2.2.
Finding the Bottleneck
2.2.1.
Understanding Bottlenecks
At this point, assume we have identified a situation that shows poor
performance. Now we need to find the performance bottleneck. The bottleneck
generally shifts depending on the content of the scene. To make things more
complicated, it often shifts over the course of a single frame. So “finding the
bottleneck” really means “Let’s find the bottleneck that limits us the most for
this scenario.” Eliminating this bottleneck achieves the largest performance boost.
Figure 1. Potential Bottlenecks
In an ideal case, there won’t be any one bottleneck—the CPU, AGP bus, and
GPU pipeline stages are all equally loaded (see Figure 1). Unfortunately, that
case is impossible to achieve in real-world applications—in practice, something
always holds back performance.
12
NVIDIA GPU Programming Guide
The bottleneck may reside on the CPU or the GPU. NVPerfHUD’s green line
(see Section 10.1 for more information about NVPerfHUD) shows how many
milliseconds the GPU is idle during a frame. If the GPU is idle for even one
millisecond per frame, it indicates that the application is at least partially CPUlimited. If the GPU is idle for a large percentage of frame time, or if it’s idle for
even one millisecond in all frames and the application does not synchronize
CPU and GPU, then the CPU is the biggest bottleneck. Improving GPU
performance simply increases GPU idle time.
2.2.2.
Basic Tests
You can perform several simple tests to identify your application’s bottleneck.
You don’t need any special tools or drivers to try these, so they are often the
easiest to start with.
‰
Eliminate all file accesses. Any hard disk access will kill your frame rate.
This condition is easy enough to detect—just take a look at your computer's
"hard disk in use" light or disk performance monitor signals using
Windows’ perfmon tool, AMD’s CodeAnalyst, (http://www.amd.com/usen/Processors/DevelopWithAMD/0,,30_2252_3604,00.html) or Intel’s
VTune (http://www.intel.com/software/products/vtune/). Keep in mind
that hard disk accesses can also be caused by memory swapping, particularly
if your application uses a lot of memory.
‰
Run identical GPUs on CPUs with different speeds. It’s helpful to find
a system BIOS that allows you to adjust (i.e., down-clock) the CPU speed,
because that lets you test with just one system. If the frame rate varies
proportionally depending on the CPU speed, your application is CPUlimited.
‰
Reduce your GPU's core clock. You can use publicly available utilities
such as Coolbits (see Chapter 6) to do this. If a slower core clock
proportionally reduces performance, then your application is limited by the
vertex shader, rasterization, or the fragment shader (that is, shader-limited).
‰
Reduce your GPU's memory clock. You can use publicly available
utilities such as Coolbits (see Chapter 6) to do this. If the slower memory
clock affects performance, your application is limited by texture or frame
buffer bandwidth (GPU bandwidth-limited).
Generally, changing CPU speed, GPU core clock, and GPU memory clock are
easy ways to quickly determine CPU bottlenecks versus GPU bottlenecks. If
underclocking the CPU by n percent reduces performance by n percent, then
the application is CPU-limited. If under-locking the GPU’s core and memory
13
How to Optimize Your Application
clocks by n percent reduces performance by n percent, then the application is
GPU-limited.
2.2.3.
Using NVPerfHUD
To verify CPU bottlenecks, run your application on a special driver, called a
Null Hardware (Null HW) driver. This driver works like a normal driver (it runs
through all the same code paths of a normal driver), except it never actually
hands any work to the GPU. A Null HW driver therefore emulates an infinitely
fast GPU. If the performance of a Null HW driver is no better than a normal driver, the
application is completely CPU-bound.
NVPerfHUD 2.0 has a special mode that suppresses all draw calls of an
application and thus emulates a Null Hardware Driver. However, omitting all
draw calls also lightens the CPU load, since the driver no longer has to process
and commit any of the state-changes made prior to each draw call.
NVPerfHUD also has a variety of other useful features that can help you to
identify performance problems.
If NVPerfHUD shows the GPU is never idle, then the application is GPUlimited. The blue line on NVPerfHUD shows how many milliseconds the driver
is waiting for the GPU, so it’s able to verify GPU-bound performance.
In addition, NVPerfHUD 3.0 has a “Frame Analysis” mode that allows you to
freeze your application and single-step though the current frame, one draw call
at a time. You can drill down to see what is happening inside the GPU at each
stage in the graphics pipeline using Advanced State Inspectors for the Index
Unit, Vertex Shader, Pixel Shader & Render Operations unit.
The NVPerfHUD User Guide contains detailed methodology for identifying and
removing bottlenecks, troubleshooting, and more. It is available at
http://developer.nvidia.com/object/nvperfhud_home.html.
2.3.
Bottleneck: CPU
If an application is CPU-bound, use profiling to pinpoint what’s consuming
CPU time. The following modules typically use significant amounts of CPU
time:
14
‰
Application (the executable as well as related DLLs)
‰
Driver (nv4disp.dll, nvoglnt.dll)
‰
DirectX Runtime (d3d9.dll)
NVIDIA GPU Programming Guide
‰
DirectX Hardware Abstraction Layer (hal32.dll)
Because the goal at this stage is to reduce the CPU overhead so that the CPU is
no longer the bottleneck, it is relatively important what consumes the most
CPU time. The usual advice applies: choose algorithmic improvements over
minor optimizations. And of course, find the biggest CPU consumers to yield
the largest performance gains.
Next, we need to drill into the application code and see if it’s possible to
remove or reduce code modules. If the application spends large amounts of
CPU in hal32.dll, d3d9.dll, or nvoglnt.dll, this may indicate API
abuse. If the driver consumes large amounts of CPU, is it possible to reduce the
number of calls made to the driver? Improving batch sizes helps reduce driver
calls. Detailed information about batching is available in the following
presentations:
http://developer.nvidia.com/docs/IO/8230/BatchBatchBatch.ppt
http://download.nvidia.com/developer/presentations/GDC_2004/Dx9Optim
ization.pdf
NVPerfHUD also helps to identify driver overhead. It can display the amount
of time spent in the driver per frame (plotted as a red line) and it graphs the
number of batches drawn per frame.
Other areas to check when performance is CPU-bound:
‰
Is the application locking resources, such as the frame buffer or
textures? Locking resources can serialize the CPU and GPU, in effect
stalling the CPU until the GPU is ready to return the lock. So the CPU is
actively waiting and not available to process the application code. Locking
therefore causes CPU overhead.
‰
Does the application use the CPU to protect the GPU? Culling small
sets of triangles creates work for the CPU and saves work on the GPU, but
the GPU is already idle! Removing these CPU-side optimizations actually
increase performance when CPU-bound.
‰
Consider offloading CPU work to the GPU. Can you reformulate your
algorithms so that they fit into the GPU’s vertex or pixel processors?
‰
Use shaders to increase batch size and decrease driver overhead. For
example, you may be able to combine two materials into a single shader and
draw the geometry as one batch, instead of drawing two batches each with
its own shader. Shader Model 3.0 can be useful in a variety of situations to
collapse multiple batches into one, and reduce both batch and draw
overhead. See Section 4.1 for more on Shader Model 3.0.
15
How to Optimize Your Application
2.4.
Bottleneck: GPU
GPUs are deeply pipelined architectures. If the GPU is the bottleneck, we need
to find out which pipeline stage is the largest bottleneck. For an overview of the
various stages of the graphics pipeline, see
http://developer.nvidia.com/docs/IO/4449/SUPP/GDC2003_PipelinePerfor
mance.ppt.
NVPerfHUD simplifies things by letting you force various GPU and driver
features on or off. For example, it can force a mipmap LOD bias to make all
textures 2 × 2. If performance improves a lot, then texture cache misses are the
bottleneck. NVPerfHUD similarly permits control over pixel shader execution
times by forcing all or part of the shaders to run in a single cycle.
If you determine that the GPU is the bottleneck for your application, use the
tips presented in Chapter 3 to improve performance.
16
Chapter 3.
General GPU Performance Tips
This chapter presents the top performance tips that will help you achieve
optimal performance on GeForce FX, GeForce 6 Series, and GeForce 7 Series
GPUs. For your convenience, the tips are organized by pipeline stage. Within
each subsection, the tips are roughly ordered by importance, so you know
where to concentrate your efforts first.
A great place to get an overview of modern GPU pipeline performance is the
Graphics Pipeline Performance chapter of the book GPU Gems: Programming
Techniques, Tips, and Tricks for Real-Time Graphics. The chapter covers bottleneck
identification as well as how to address potential performance problems in all
parts of the graphics pipeline.
Graphics Pipeline Peformance is freely available at
http://developer.nvidia.com/object/gpu_gems_samples.html.
3.1.
List of Tips
When used correctly, recent GPUs can achieve extremely high levels of
performance. This list presents an overview of available performance tips that
the subsequent sections explain in more detail.
‰
Poor Batching Causes CPU Bottleneck
‰ Use fewer batches
‰ Use texture atlases to avoid texture state changes.
http://developer.nvidia.com/object/nv_texture_tools.html
17
General GPU Performance Tips
‰ In DirectX, use the Instancing API to avoid SetMatrix and
similar instancing state changes.
‰
Vertex Shaders Cause GPU Bottleneck
‰ Use indexed primitive calls
‰ Use DirectX 9’s mesh optimization calls
[ID3DXMesh:OptimizeInplace() or
ID3DXMesh:Optimize()]
‰ Use our NVTriStrip utility if an indexed list won’t work
http://developer.nvidia.com/object/nvtristrip_library.html
‰
Pixel Shaders Cause GPU Bottleneck
‰ Choose the minimum pixel shader version that works for what
‰
‰
‰
‰
‰
‰
‰
‰
‰
18
you’re doing
‰ When developing your shader, it’s okay to use a higher
version. Make it work first, then look for opportunities to
optimize it by reducing the pixel shader version.
If you need ps_2_* functionality, use the ps_2_a profile
Choose the lowest data precision that works for what you’re doing:
‰ Prefer half to float
Use the half type for everything that you can:
‰ Varying parameters
‰ Uniform parameters
‰ Variables
‰ Constants
Balance the vertex and pixel shaders.
Push linearizable calculations to the vertex shader if you’re bound
by the pixel shader.
Don’t use uniform parameters for constants that will not change
over the life of a pixel shader.
Look for opportunities to save computations by using algebra.
Replace complex functions with texture lookups
‰ Per-pixel specular lighting
‰ Use FX Composer to bake programmatically generated
textures to files
‰ But sincos, log, exp are native instructions and do not
need to be replaced by texture lookups
Texturing Causes GPU Bottleneck
NVIDIA GPU Programming Guide
‰ Use mipmapping
‰ Use trilinear and anisotropic filtering prudently
‰ Match the level of anisotropic filtering to texture
complexity.
‰ Use our Photoshop plug-in to vary the anisotropic filtering
level and see what it looks like.
http://developer.nvidia.com/object/nv_texture_tools.html
Follow this simple rule of thumb: If the texture is noisy,
turn anisotropic filtering on.
‰
Rasterization Causes GPU bottleneck
‰ Double-speed z-only and stencil rendering
‰ Early-z (Z-cull) optimizations
‰
Antialiasing
‰ How to take advantage of antialiasing
3.2.
Batching
3.2.1.
Use Fewer Batches
“Batching” refers to grouping geometry together so many triangles can be
drawn with one API call, instead of using (in the worse case) one API call per
triangle. There is driver overhead whenever you make an API call, and the best
way to amortize this overhead is to call the API as little as possible. In other
words, reduce the total number of draw calls by drawing several thousand
triangles at once. Using a smaller number of larger batches is a great way to
improve performance. As GPUs become ever more powerful, effective
batching becomes ever more important in order to achieve optimal rendering
rates.
3.3.
Vertex Shader
3.3.1.
Use Indexed Primitive Calls
Using indexed primitive calls allows the GPU to take advantage of its posttransform-and-lighting vertex cache. If it sees a vertex it’s already transformed,
it doesn’t transform it a second time—it simply uses a cached result.
19
General GPU Performance Tips
In DirectX, you can use the ID3DXMesh class’s OptimizeInPlace() or
Optimize() functions to optimize meshes and make them more friendly
towards the vertex cache.
You can also use our own NVTriStrip utility to create optimized cache-friendly
meshes. NVTriStrip is a standalone program that is available at
http://developer.nvidia.com/object/nvtristrip_library.html.
3.4.
Shaders
High-level shading languages provide a powerful and flexible mechanism that
makes writing shaders easy. Unfortunately, this means that writing slow shaders
is easier than ever. If you’re not careful, you can end up with a spontaneous
explosion of slow shaders that brings your application to a halt. The following
tips will help you avoid writing inefficient shaders for simple effects. In
addition, you’ll learn how to take full advantage of the GPU’s computational
power. Used correctly, the high-end GeForce FX GPUs can deliver more than
20 operations per clock cycle! And the latest GeForce 6 and 7 Series GPUs can
deliver many times more performance.
3.4.1.
Choose the Lowest Pixel Shader Version That Works
Choose the lowest pixel shader version that will get the job done. For example,
if you’re doing a simple texture fetch and a blend operation on a texture that’s
just 8 bits per component, there’s no need to use a ps_2_0 or higher shader.
3.4.2.
Compile Pixel Shaders Using the ps_2_a Profile
Microsoft’s HLSL compiler (fxc.exe) adds chip-specific optimizations based
on the profile that you’re compiling for. If you’re using a GeForce FX GPU and
your shaders require ps_2_0 or higher, you should use the ps_2_a profile,
which is a superset of ps_2_0 functionality that directly corresponds to the
GeForce FX family. Compiling to the ps_2_a profile will probably give you
better performance than compiling to the generic ps_2_0 profile. Please note
that the ps_2_a profile was only available starting with the July 2003 HLSL
release.
In general, you should use the latest version of fxc (with DirectX 9.0c or
newer), since Microsoft will add smarter compilation and fix bugs with each
20
NVIDIA GPU Programming Guide
release. For GeForce 6 and 7 Series GPUs, simply compiling with the
appropriate profile and latest compiler is sufficient.
3.4.3.
Choose the Lowest Data Precision That Works
Another factor that affects both performance and quality is the precision used
for operations and registers. The GeForce FX, GeForce 6 Series, and GeForce
7 Series GPUs support 32-bit and 16-bit floating point formats (called float
and half, respectively), and a 12-bit fixed-point format (called fixed). The
float data type is very IEEE-like, with an s23e8 format. The half is also
IEEE-like, in an s10e5 format. The 12-bit fixed type covers a range from [2,2) and is not available in the ps_2_0 and higher profiles. The fixed type is
available with the ps_1_0 to ps_1_4 profiles in DirectX, and with either the
NV_fragment_program extension or Cg in OpenGL.
The performance of these various types varies with precision:
‰
The fixed type is fastest and should be used for low-precision
calculations, such as color computation.
‰
If you need floating-point precision, the half type delivers higher
performance than the float type. Prudent use of the half type can triple
frame rates, with more than 99% of the rendered pixels within one leastsignificant bit (LSB) of a fully 32-bit result in most applications!
‰
If you need the highest possible accuracy, use the float type.
You can use the /Gpp flag (available in the July 2003 HLSL update) to force
everything in your shaders to half precision. After you get your shaders
working and follow the tips in this section, enable this flag to see its effect on
performance and quality. If no errors appear, leave this flag enabled. Otherwise,
manually demote to half precision when it is beneficial (/Gpp provides an
upper performance bound that you can work toward).
When you use the half or fixed types, make sure you use them for varying
parameters, uniform parameters, variables, and constants. If you’re using
DirectX’s assembly language with the ps_2_0 profile, use the _pp modifier to
reduce the precision of your calculations.
If you’re using the OpenGL ARB_fragment_program language, use the
ARB_precision_hint_fastest option if you want to minimize execution
time, with possibly reduced precision, or use the NV_fragment_program
option if you want to control precision on a per-instruction basis (see
http://www.nvidia.com/dev_content/nvopenglspecs/GL_NV_fragment_prog
ram_option.txt).
21
General GPU Performance Tips
Many color-based operations can be performed with the fixed or half data
types without any loss of precision (for example, a tex2D*diffuseColor
operation).
On GeForce FX hardware in OpenGL, you can speed up shaders consisting of
mostly floating-point operations by doing operations (like dot products of
normalized vectors) in fixed-point precision.
For instance, the result of any normalize can be half-precision, as can colors.
Positions can be half-precision as well, but they may need to be scaled in the
vertex shader to make the relevant values near zero.
For instance, moving values to local tangent space, and then scaling positions
down can eliminate banding artifacts seen when very large positions are
converted to half precision.
3.4.4.
Save Computations by Using Algebra
Once you’ve got your shader working, look at your computations and figure out
if you can collapse them by using mathematical properties. This is especially
true for library functions shared across multiple shaders. For example:
‰
Generic sphere map projection is often expressed in terms of
p = sqrt(Rx2 + Ry2 + (Rz + 1)2)
This expands to :
p = sqrt(Rx2 + Ry2 + Rz2 + 2Rz + 1)
If you know the reflection vector is normalized (see Sections 3.4.8 and
3.4.6), the sum of the first three terms is guaranteed to be 1.0. This
expression can then be refactored as:
p = sqrt(2 * (Rz + 1)) = 1.414*sqrt(Rz + 1)
‰
Fold the multiplication by 1.414 into another constant (see Section 3.4.8),
saving a dot product.
‰ dot(normalize(N), normalize(L))
can be computed far more
efficiently.
‰ It’s usually computed as (N/|N|) dot (L/|L|), which requires
two expensive reciprocal square root (rsq) computations.
‰ Doing a little algebra gives us:
‰ (N/|N|) dot (L/|L|)
‰ = (N dot L) / (|N| * |L|)
22
NVIDIA GPU Programming Guide
‰ = (N dot L) / (sqrt( (N dot N) * (L dot L) )
‰ = (N dot L) * rsq( (N dot N) * (L dot L) )
‰ which requires only one expensive rsq operation.
3.4.5.
Don’t Pack Vector Values into Scalar Components of Multiple
Interpolants
Packing too much information into a calculation can make it harder for the
compiler to optimize your code efficiently. For example, if you are passing
down a tangent matrix, do not include the view vector in the 3 q components.
This mistake is illustrated below:
// Bad practice
tangent = float4(tangentVec, viewVec.x)
binormal = float4(binormalVec, viewVec.y)
normal = float4(normalVec, viewVec.z)
Instead, place the view vector in a fourth interpolant.
3.4.6.
Don’t Write Overly Generic Library Functions
Functions that are shared across multiple shaders are frequently written very
generically. For example, reflection is often computed as:
float3 reflect(float3 I, float3 N) {
return (2.0*dot(I,N)/dot(N,N))*N – I;
}
Written this way, the reflection vector can be computed independent of the
length of the normal or incident vectors. However, shader authors frequently
want at least the normal vector normalized in order to perform lighting
calculations. If this is the case, then a dot product, a reciprocal, and a scalar
multiply can be removed from reflect(). Optimizations like these can
dramatically improve performance.
3.4.7.
Don’t Compute the Length of Normalized Vectors
A common (and expensive) example of an overly-generic library function is one
that computes the lengths of the input vectors locally. However, the vectors
have often been normalized prior to calling the function. Compilers don’t detect
this, which means substantial per-pixel arithmetic is performed to compute 1.0.
23
General GPU Performance Tips
If your library functions must work correctly independent of the vector’s
lengths, consider making length a scalar parameter to the functions. That way,
the shaders that normalize vectors before calling the function can pass down a
constant value of 1.0 (providing all the benefits of not computing the length),
and those that don’t normalize vectors can compute the length.
3.4.8.
Fold Uniform Constant Expressions
Many developers compute expressions involving dynamic constants in their
pixel shaders. If more than one uniform constant (or a uniform and an in-lined
constant) is used in an expression, there is often a way to fold the constants
together and improve performance. For example:
half4 main(float2 diffuse : TEXCOORD0,
uniform sampler2D diffuseTex,
uniform half4 g_OverbrightColor) {
return tex2D(diffuseTex, diffuse) * g_OverbrightColor * 3.0;
}
g_OverbrightColor can be premultiplied by 3.0 on the CPU, saving a per-
pixel multiplication on potentially millions of pixels each frame.
You may need to distribute or factor expressions in order to fold as many
constant expressions as possible. In addition, you can use HLSL preshaders to
perform precomputation on the CPU before a shader runs.
Another common example is computing materialColor * lightColor
at each vertex. Because this expression has the same value for all vertices in a
given batch, it should be calculated on the CPU.
You should also compute matrix inverses and transposes on the CPU instead of
on the GPU, because they only need to be calculated once instead of per-vertex
or per-fragment. The /Zpr (pack row-major) and /Zpc (pack column-major)
compiler options can help store matrices the way you want.
3.4.9.
Don’t Use Uniform Parameters for Constants That Won’t Change
Over the Life of a Pixel Shader
Developers sometimes use uniform parameters to pass in commonly used
constants like 0, 1, and 255. This practice should be avoided. It makes it harder
for compilers to distinguish between constants and shader parameters, reducing
performance.
24
NVIDIA GPU Programming Guide
3.4.10. Balance the Vertex and Pixel Shaders
Achieving high performance is all about removing bottlenecks—which really
means that you have to balance every piece of the pipeline: the CPU, the AGP
bus, and the stages of the graphics pipeline. The decision to use a vertex shader
or a pixel shader depends on a few factors:
‰
How tessellated are your objects? You may want to lighten the load on
the vertex shader if you have millions of vertices in each frame. This is
especially true if you’re using a multipass algorithm.
‰
What resolution are you targeting? If you expect your application to be
run at higher resolutions, the pixel shader is more likely to become the
bottleneck. So, you may want to push more computations to the vertex
shader.
‰
How long are your pixel shaders? If you’re doing complex shading, the
pixel shader will probably be your bottleneck. If your pixel shaders compile
to more than 20 cycles (on average) and occupy more than half the screen,
your application will likely be pixel shader-bound on GeForce FX
hardware. So, look for opportunities to move calculations to the vertex
shader. (See Section 3.4.11 for examples.) You can use our NVShaderPerf
tool to find out how many cycles your shaders are using. Also, note that
newer hardware such as GeForce 6 and 7 Series GPUs will allow more
complex pixel shaders before becoming pixel shader-bound.
3.4.11. Push Linearizable Calculations to the Vertex Shader If You’re
Bound by the Pixel Shader
The rasterizer takes per-vertex values and interpolates them per-fragment while
accounting for perspective correction. Take advantage of the hardware that
already takes care of this for you by moving linear calculations to the vertex
shader. You may be able to perform the calculation for fewer vertices and still
receive an interpolated result in the pixel shader.
For example, you can move from world space to light space for attenuation. Or,
if you’re doing bump mapping, you can make the move into tangent space pervertex, unless you’re doing per-pixel reflection into a cube map.
3.4.12. Use the mul() Standard Library Function
Instead of performing matrix multiplication manually, use the mul() Standard
Library function. This will avoid some row-major/column-major issues that
may appear when applications pass down matrices in interpolants.
25
General GPU Performance Tips
3.4.13. Use D3DTADDRESS_CLAMP (or GL_CLAMP_TO_EDGE) Instead
of saturate() for Dependent Texture Coordinates
Using saturate()can cost extra on some GPUs. If the clamped result is
used as a texture coordinate, it is preferable to use the texture hardware’s ability
to clamp texture coordinate to the [0..1] range, rather than doing this in the
shader.
3.4.14. Use Lower-Numbered Interpolants First
You will see higher performance if you use lower-numbered texture coordinate
sets (TEXCOORD sets) first. Start by using TEXCOORD0, and move upwards from
there to TEXCOORD1, TEXCOORD2, and so on.
3.5.
Texturing
3.5.1.
Use Mipmapping
To prevent minified textures from causing “sparkling” artifacts, always use
mipmapping in your applications. You’ll achieve better image quality, improved
texture cache behavior, and higher performance. You get all this for just 33%
more memory usage, which is a great trade-off. 3D textures, in particular, can
benefit greatly from mipmapping—we’ve seen performance increases of 30% to
40% when mipmapping was enabled.
When creating mipmaps, don’t simply use a box filter to generate smaller and
smaller mipmaps. Instead, use a Gaussian or Mitchell filter to take more
samples—this will produce a higher quality result. But spending a little more
time in the preprocess to create mipmaps, you can make your application look
better continuously at runtime. Our Photoshop plug-in (part of the NVIDIA
Texture Tools suite) can quickly create high-quality mipmaps for you. The suite
is available at http://developer.nvidia.com/object/nv_texture_tools.html.
3.5.2.
Use Trilinear and Anisotropic Filtering Prudently
Trilinear and anisotropic filtering both help to improve image quality, but they
each bring a performance penalty. Try to use trilinear and anisotropic filtering
only where they’re needed. In general, you’ll want to use them on textures that
have a lot of high-contrast detail. For anisotropic filtering, you may also want to
consider the orientation of the texture. If you know a texture will be oblique to
26
NVIDIA GPU Programming Guide
the viewer (for example, a floor texture), increase the level of anisotropic
filtering for that texture. For multitextured surfaces, you should have an
appropriate level of filtering for each of the different layers.
Our Adobe Photoshop plug-in is helpful for determining the level of
anisotropic filtering to use. This tool allows you to try different filtering levels
and see the visual effects. It is available at
http://developer.nvidia.com/object/nv_texture_tools.html. Your artists may
want to use this tool to help them decide which textures require anisotropic or
trilinear filtering.
3.5.3.
Replace Complex Functions with Texture Lookups
Textures are a great way to encode complex functions—think of them as
multidimensional arrays that you can index on-the-fly. The GeForce FX family
can access textures efficiently—often at the same cost as an arithmetic
operation. You can use our FX Composer tool to prototype this kind of
optimization. FX Composer is available at
http://developer.nvidia.com/FXComposer.
Any time you can encode a complex sequence of arithmetic operations in a
texture, you can improve performance. Keep in mind that some complex
functions, such as log and exp, are micro-instructions in ps_2_0 and higher
profiles, and therefore don’t need to be encoded in textures for optimal
performance.
3.5.3.1.
Per-Pixel Lighting
Using a 2D Texture
One common situation where a texture can be useful is in per-pixel lighting.
You can use a 2D texture that you index with (N dot L) on one axis and (N
dot H) on the other axis. At each (u, v) location, the texture would encode:
max(N dot L,0) + Ks*pow((N dot L>0) ? max(N dot H,0) : 0), n)
This is the standard Blinn lighting model, including clamping for the diffuse and
specular terms.
Using a 1D ARGB Texture
A useful trick is to use a 1D ARGB texture, indexed by (N dot H). The
texture encodes (N dot H) to various exponents in each channel. For
example, it may encode:
( (N dot H)4, (N dot H)8, (N dot H)12, (N dot H)16 )
27
General GPU Performance Tips
Then, each material is assigned a four-component weighting constant that
blends these values, giving a monochrome specular value for shading. The
beauty of this approach is that it works on GeForce 4-class hardware and is
flexible enough to enable a variety of appearances.
Using a 3D Texture
You can also add the specular exponentiation to the mix by using a 3D texture.
The first two axes use the 2D texture technique described in the previous
section, and the third axis encodes the specular exponent (shininess).
Remember, however, that cache performance may suffer if the texture is too
large. You may want to encode only the most frequently used exponents.
3.5.3.2.
Normalizing Vectors
If you’re writing a ps_1_* shader, use normalization cube maps to normalize
vectors quickly. For higher quality, you can use two 16-bit signed cube maps:
one for x and y, and the other for z.
Another optimization is based on the fact that, in practice, the vectors V that
are normalized are often of norm close to 1 because they’re interpolated or
filtered normals. This means that you can approximate 1 /||V|| by the first
terms of the Taylor expansion of 1 / sqrt(x) at x = 1:
1 / sqrt(x) ~ 1 + ½ (1 - x)
such that:
V / ||V|| = V / sqrt(||V||^2) = V + ½ V (1 - ||V||^2)
This formula can be written in two assembly instructions:
dp3_sat r1, r0, r0
mad_d2 r1, r0, 1-r1, r0_x2
with r0 containing V and the final value of r1 containing V / ||V||.
The code above is only valid for ps_1_4 because of the _x2 register modifier.
For lower pixel shader versions, the following formula can be used instead:
dp3_sat r1, r0_bx2, r0_bx2
mad r1, r0_bias, 1-r1, r0_bx2
This is assuming that r0 contains ½ (V + 1), which is rarely a constraint as V
often needs to be passed on range-compressed from [ -1, 1 ] to [ 0, 1 ] to the
pixel shader.
28
NVIDIA GPU Programming Guide
GeForce 6 and 7 Series GPUs have a special half-precision normalize unit that
can normalize an fp16 vector for free during a shader cycle. Take advantage of
this feature, simply perform a normalization on an fp16 quantity and the
compiler will generate a nrmh instruction.
For more on normalization, please see our Normalization Heuristics and Bump
Map Compression whitepapers, available at the following URLs:
http://developer.nvidia.com/object/normalization_heuristics.html
http://developer.nvidia.com/object/bump_map_compression.html
3.5.3.3.
The sincos() Function
Despite the preceding advice, the GeForce FX family and later GPUs support
some complex mathematical functions natively in hardware. One such function
that is convenient is the sincos function, which allows you to simultaneously
calculate the sine and cosine of a value.
3.6.
Performance
3.6.1.
Double-Speed Z-Only and Stencil Rendering
The GeForce FX, GeForce 6 Series, and GeForce 7 Series GPUs render at
double speed when rendering only depth or stencil values. To enable this special
rendering mode, you must follow the following rules:
3.6.2.
‰
Color writes are disabled
‰
The active depth-stencil surface is not multisampled
‰
Texkill has not been applied to any fragments
‰
Depth replace (oDepth, texm3x2depth, texdepth) has not been
applied to any fragments
‰
Alpha test is disabled
‰
No color key is used in any of the active textures
‰
No user clip planes are enabled
Early-Z Optimization
Early-z optimization (sometimes called “z-cull”) improves performance by
avoiding the rendering of occluded surfaces. If the occluded surfaces have
29
General GPU Performance Tips
expensive shaders applied to them, z-cull can save a large amount of
computation time. To take advantage of z-cull, follow these guidelines:
‰
Don’t create triangles with holes in them (that is, avoid alpha test or texkill)
‰
Don’t modify depth (that is, allow the GPU to use the interpolated depth
value)
Violating these rules can invalidate the data the GPU uses for early
optimization, and can disable z-cull until the depth buffer is cleared again.
3.6.3.
Lay Down Depth First
The best way to take advantage of the two aforementioned performance
features is to “lay down depth first.” By this, we mean that you should use
double-speed depth rendering to draw your scene (without shading) as a first
pass. This then establishes the closest surfaces to the viewer. Now you can
render the scene again, but with full shading. Z-cull will automatically cull out
fragments that aren’t visible, meaning that you save on shading computations.
Laying down depth first requires its own render pass, but can be a performance
win if many occluded surfaces have expensive shading applied to them. Doublespeed rendering is less efficient as triangles get small. And, small triangles can
reduce z-cull efficiency.
Another related technique is Deferred Shading, which you can find in NVSDK
7.1 and later.
3.6.4.
Allocating Memory
In order to minimize the chance of your application thrashing video memory,
the best way to allocate shaders and render targets is:
1. Allocate render targets first
‰ Sort the order of allocation by pitch (width * bpp).
‰ Sort the different pitch groups based on frequency of use. The
surfaces that are rendered to most frequently should be allocated
first.
2. Create vertex and pixel shaders
3. Load remaining textures
30
NVIDIA GPU Programming Guide
3.7.
Antialiasing
GeForce FX, GeForce 6 Series, and GeForce 7 Series GPUs all have powerful
antialiasing engines. They perform best with antialiasing enabled, so we
recommend that you enable your applications for antialiasing.
If you need to use techniques that don’t work with antialiasing, contact us—
we’re happy to discuss the problem with you and to help you find solutions.
One issue that is now solved with DirectX 9.0b or later is using antialiasing with
post-processing effects. The StretchRect() call can copy the back buffer to
an off-screen texture in concert with multisampling.
For instance, if 4x multisampling is enabled, on a 100 × 100 back buffer, the
driver actually internally creates a 200 × 200 back buffer and depth buffer in
order to perform the antialiasing. If the application creates a 100 × 100 offscreen texture, it can StretchRect() the entire back buffer to the off-screen
surface, and the GPU will filter down the antialiased buffer into the off-screen
buffer.
Then glows and other post-processing effects can be performed on the 100 ×
100 texture, and then applied back to the main back buffer.
This resolution mismatch between the real back buffer size (200 × 200) and the
application’s view of it (100 × 100) is the reason why you can’t attach a
multisampled z buffer to a non-multisampled render target.
31
Chapter 4.
GeForce 6 & 7 Series Programming
Tips
This chapter presents several useful tips that help you fully use the capabilities
of GeForce 6 & 7 Series as well as NV4X-based Quadro FX GPUs. These are
mostly feature oriented, though some may affect performance as well.
4.1.
Shader Model 3.0 Support
Microsoft DirectX 9.0 introduced several new standards for advanced vertex
and pixel shader technology, version 2.0 and version 3.0. Shader Model 2.0
hardware has been available since late 2002, and the vast majority of GPUs
shipped today support Shader Model 2.0 or better. Shader Model 2.0 includes
technologies useful for advanced lighting and animation techniques, but has
limited shader program length, and complexity, which limits the fidelity of the
effects that can be achieved.
As developers push against the limits inherent in Pixel Shader 2.0 and Vertex
Shader 2.0, they have started to adopt the newer, more advanced Shader Model
3.0. This shader model has advances in several areas, in both pixel and vertex
shader processing.
33
GeForce 6 Series Programming Tips
4.1.1.
Pixel Shader 3.0
The following is a feature summary outlining the key differentiators between
Pixel Shader 2.0 and 3.0.
34
Pixel Shader
Feature
Shader 2.0
Shader 3.0
Description
Shader length
96
65535+
Allows more complex shading,
lighting, and procedural
materials
Dynamic
branching
No
Yes
Saves performance by skipping
complex shading on irrelevant
pixels
Shader
antialiasing
Not
supported
Built-in
derivative
instructions
Developers can calculate the
screen space derivatives of any
function, allowing them to
adjust shading frequencies or
over-sampling to eliminate
artifacts
Back-face
register
No
Yes
Allows two-sided lighting in a
single pass
Interpolated
color format
8-bit integer
minimum
32-bit floating
point
minimum
Higher range and precision
color allows high-dynamic range
lighting at the vertex level
Multiple
render
targets
Optional
4 required
Allows advanced lighting
algorithms to save filtering and
vertex work – thus more lights
for minimal cost
Fog and
specular
8-bit fixed
function
minimum
Custom fp16fp32 shader
program
Shader Model 3.0 gives
developers full and precise
control over specular and fog
computations, previously fixedfunction
Texture
coordinate
count
8
10
More per-pixel inputs allows
more realistic rendering,
especially for skin
NVIDIA GPU Programming Guide
4.1.2.
Vertex Shader 3.0
Here is a similar listing of key features developers enjoy when moving from
Vertex Shader Model 2.0 to 3.0.
4.1.3.
Vertex
shader
feature
Shader 2.0
Shader 3.0
Description
Shader
length
256
Instructions
65535 instructions
More instructions allow more
detailed character lighting
and animation
Dynamic
branching
No
Yes
Saves performance by
skipping animation and
calculations on irrelevant
vertices
Vertex
texture
No
Any number of
lookups from up to
4 textures
Allows displacement mapping,
particle effects
Instancing
support
No
Required
Allows many varied objects to
be drawn with only a single
command
Dynamic Branching
One major feature of both Shader 3.0 models (vertex and pixel) is Dynamic
Branching. Put simply, this allows a shader author to create true loops and
conditionals in their shader programs. For instance, one could write a shader
that looped through a certain number of vertex lights, determine which ones
might influence a particular vertex, and then pass down the index of each
relevant light to the pixel shader. The pixel shader could then use this ‘light
index’ to determine which light parameters to apply. The pixel shader would
then loop over the active lights, then use dynamic branching to exit the shader
early once all lights are processed.
Most light types only apply to the front side of an object—the side facing the
light. Therefore, you can use both vertex and pixel branching to skip
processing for lights that the shader detects as facing away from the light (using
the new “back-face” register). This can save significant processing time, and
speed up the shader. Similar speedups can be used to skip processing of
character bone animation as well as many similar algorithms.
35
GeForce 6 Series Programming Tips
4.1.4.
Easier Code Maintenance
As game engines become more and more complex, they often create many
different versions of each shader in order to fit them all in to the Pixel Shader
2.0 program length limitations. This adds to code maintenance, shader compile
time, and level load time, as well as taking up valuable system memory at
runtime. Shader Model 3.0 eliminates this issue, through its comprehensive
looping and branching, allowing the engine to write a single vertex and single
pixel shader containing appropriate static and dynamic branching in order to
select the correct execution path at runtime, thus greatly simplifying the shader
combinatorial explosion issue.
4.1.5.
Instancing
Another key feature of Shader Model 3.0 is the support for the Microsoft
DirectX® Instancing API. Currently, games face limits on the number of
unique objects they can display in the scene, not because of graphics
horsepower, but often because of the CPU-side overhead of either storing or
submitting many slightly different variations of the same object. For instance, a
forest is made up of trees that are often similar to each other, but each would be
in a different position, have differing height, leaf color, and so on. In order to
add the desired variation, developers have to choose between storing many
separate copies of the tree, each slightly different, or making expensive render
state changes in order to rotate, scale, color and place each tree.
Instancing allows the programmer to store a single tree, and then several other
vertex data streams to specify the per-instance color, height, branch size and so
on. For instance, a single 1,000-vertex tree model would contain the vertex
positions and normals, and a 200-element vertex streams would contain
positions, colors, and heights. Instancing allows the programmer to submit a
single draw call, which renders each of the 200 trees, using the same data for the
basic tree shape, but then vary it through the per-instance streams.
Our instancing code sample is available at
http://download.nvidia.com/developer/SDK/Individual_Samples/samples.html
36
NVIDIA GPU Programming Guide
4.1.6.
Summary
In summary, DirectX 9.0 Shader Model 3.0 is a significant step forward in terms
of ease of use, performance, and shader complexity. Dynamic branching brings
speed-ups to many algorithms which contain early-out opportunities, while also
simplifying shader code paths in graphics engines and tools. Lastly, instancing
allows extreme complexity for very low CPU and memory overhead.
4.2.
GeForce 7 Series Features
With 8 vertex pipelines and 24 pixel pipelines, the GeForce 7800 GTX features
a superscalar architecture and the NVIDIA CineFX 4.0 engine. The CineFX 4.0
engine includes a redesigned vertex shader unit that shortens the time required
to set up complex geometry processing. This results in higher vertex
throughput.
In addition, a new pixel shader unit delivers up to 2x the floating point shading
power of previous generation products due to additional computational
resources in the shader core.
An improved texture unit incorporates new hardware to speed up anisotropic
and floating-point filtering operations. Also included are new features such as
transparency antialiasing, improved multisampling, advanced video processing,
and more.
Learn more about the GeForce 7 Series at
http://www.nvidia.com/page/geforce_7800.html.
4.3.
Transparency Antialiasing
Transparency antialiasing is a new feature available for GeForce 7 Series GPUs.
When alpha testing is used to create transparent regions in textures,
transparency antialiasing will antialias the resulting edges, resulting in higher
image quality.
With GeForce 7800 GTX, NVIDIA has exposed a render state that allows
individual primitives to be multisampled. Multisampling is accomplished by
using the computed alpha as a measure of how many samples the pixel will take.
A texel with an alpha value of one will contribute to all samples in the
destination pixel, while texels with successively lower alpha values will
37
GeForce 6 Series Programming Tips
contribute to fewer samples. Transparency multisampling is especially effective
for textures where soft edges are desirable (as is the case with most vegetation).
When using transparency multisampling, your art assets must have a gradual
falloff towards their edges in the alpha channel. The alpha channel should be
created or scaled in the shader such the computed alpha goes from the alpha
reference value to 0.0f at the boundaries. This ensures that the coverage values
will fall off nicely, producing smooth antialiased edges.
Learn more about transparency antialiasing at
http://developer.nvidia.com/object/transparency_aa.html.
4.4.
sRGB Encoding
sRGB encoding is a format that uses gamma conversion to provide more
precision near zero, mimicking the human visual system. sRGB also works with
DXT compression formats, allowing applications to benefit from both
increased color fidelity and reduced storage size.
On GeForce 6 & 7 Series GPUs, you may want to use a floating-point format in
some cases, as this offers several advantages:
‰
Higher precision across the entire range.
‰
Much larger, linear, dynamic range
‰
Linear blending for frame buffers
For textures, sRGB is vastly preferable to floating-point in most cases, because
of its smaller memory footprint and bandwidth requirements.
4.5.
Separate Alpha Blending
The GeForce 6 & 7 Series GPUs allow you to specify a separate blending
function for color and alpha. This gives you more flexibility with respect to
what values are stored in the alpha channel of your textures, for example. One
use for this would be to modulate the color channels while preserving alpha.
38
NVIDIA GPU Programming Guide
4.6.
Supported Texture Formats
The following table shows the texture formats supported by GeForce 6 & 7
Series GPUs.
Integer Formats
R8G8B8
A8R8G8B8
X8R8G8B8
R5G6B5
X1R5G5B5
A1R5G5B5
A4R4G4B4
R3G3B2
A8
A8R3G3B2
X4R4G4B4
A2B10G10R10
A8B8G8R8
X8B8G8R8
G16R16
A2R10G10B10
A16B16G16R16
A8P8
P8
L8
L16
A8L8
A4L4
2D
N
Y
Y
Y
Y
Y
Y
N
Y
N
N
N
N
N
Y
N
N
N
N
Y
Y
Y
N
Cube
N
Y
Y
Y
Y
Y
Y
N
Y
N
N
N
N
N
Y
N
N
N
N
Y
Y
Y
N
3D
N
Y
Y
Y
Y
Y
Y
N
Y
N
N
N
N
N
Y
N
N
N
N
Y
Y
Y
N
MIP
N
Y
Y
Y
Y
Y
Y
N
Y
N
N
N
N
N
Y
N
N
N
N
Y
Y
Y
N
Filter
N
Y
Y
Y
Y
Y
Y
N
Y
N
N
N
N
N
Y
N
N
N
N
Y
Y
Y
N
sRGB
N
Y
Y
Y
Y
Y
Y
N
N/A
N
N
N
N
N
N
N
N
N
N
N
N
N
N
Render
N
Y
Y
Y
Y
N
N
N
N
N
N
N
N
N
N
N
N
N
N
N
N
N
N
Blend
N
Y
Y
Y
Y
N
N
N
N
N
N
N
N
N
N
N
N
N
N
N
N
N
N
Vertex
N
N
N
N
N
N
N
N
N
N
N
N
N
N
N
N
N
N
N
N
N
N
N
Float Formats
R16F
G16R16F
A16B16G16R16F
R32F
G32R32F
A32B32G32R32F
2D
N
Y
Y
Y
N
Y
Cube
N
Y
Y
Y
N
Y
3D
N
Y
Y
Y
N
Y
MIP
N
Y
Y
Y
N
Y
Filter
N
Y
Y
N
N
N
sRGB
N/A
N/A
N/A
N/A
N/A
N/A
Render
N
Y
Y
Y
N
Y
Blend
N
N
Y
N
N
N
Vertex
N
N
N
Y
N
Y
Shadow Map
D24X8
D24S8
D16
2D
Y
Y
Y
Cube
N
N
N
3D
N
N
N
MIP
Y
Y
Y
Filter
Y
Y
Y
sRGB
N/A
N/A
N/A
Render
Y
Y
Y
Blend
N/A
N/A
N/A
Vertex
N
N
N
39
GeForce 6 Series Programming Tips
4.7.
Floating-Point Textures
GeForce 6 & 7 Series GPUs offer improved support for floating-point textures.
The following table shows the various features that are supported for both 16bit per component (fp16) and 32-bit per component (fp32) floating-point
textures.
4.7.1.
Texture
Component
Type
Nearest
Filtering
Bilinear
and
Trilinear
Filtering
Anisotropic
Filtering
NonMipmap
3D
Cube power-ofSupport Textures Maps
2
Textures
16-bit
Yes
Yes
Yes
Yes
Yes
Yes
Yes
32-bit
Yes
No
No
Yes
Yes
Yes
Yes
Limitations
Please note that we do not support the R16F format – use G16R16F instead. In
addition, you can only blend to an A16B16G16R16F surface, not a G16R16F or
R32F surface. However, filtering is supported for G16R16F textures.
4.8.
Multiple Render Targets (MRTs)
GeForce 6 & 7 Series GPUs support MRTs, which allow a pixel shader to write
out data to up to four different targets. MRTs are useful whenever the pixel
shader computes more than four float values and needs to store these
intermediate results in a texture.
Examples uses of MRTs include particle physics simultaneously computing
positions and velocities, and similar GPGPU algorithms. Deferred shading is
another technique that computes and stores multiple float4 values
simultaneously: it computes all material properties, such as for example, surface
normal, diffuse and specular material properties and stores these in separate
textures. These properties become used when lighting the scene with multiple
lights in subsequent passes.
The DirectX caps bit NumSimultaneousRTs indicates how many render
targets the graphics device can render to simultaneously. For GeForce 6 & 7
Series GPUs that caps bit is four. To enable MRTs use the
40
NVIDIA GPU Programming Guide
SetRenderTarget(index, pRenderTarget) API call. This call binds
the passed in render-target texture to the given render index. A pixel shader
then outputs to those bound render-target textures using the oC0, oC1, oC2,
and oC3 output registers. Remember to reset the render targets for indices one
through three to NULL to turn MRT rendering off.
MRTs restrict other GPU features. Most important, hardware-accelerated
antialiasing is inapplicable to MRT render targets. Furthermore, all render
targets must have the same width, height, and bit depth. For the GeForce 6 &
7 Series, that means one can freely mix-and-match within the 32-bit formats
(i.e., A8R8G8B8, X8R8G8B8, G16R16F, and R32F), or use up to four rendertargets with 64bits each (i.e., use the A16R16G16B16F format), or up to four
render-targets with 128 bits each (i.e., use the A32R32G32B32F format).
Furthermore, the post pixel shader blend operations alpha-blending, alphatesting, fogging, and dithering are only available for MRT if the
D3DMISCCAPS_MRTPOSTPIXELSHADERBLENDING cap bit is set and querying
a render-format for USAGE_QUERY_POSTPIXELSHADERBLENDING returns
yes.
GeForce 6 & 7 Series chips support this capability for all MRT formats, except
R32F, G16R16F, and A32R32G32B32F. Thus, note in particular that GeForce
6 & 7 Series GPUs (aside from the GeForce 6200) do support post-blend
operations on A16R16G16B16F floating point render-targets. The DirectX
specification further modifies the behavior of these MRT post-blend
operations: MRT rendering ignores the dithering state, respects the fog state
only for render-target zero (render-targets one through three act as if fog is
disabled), and alpha-testing uses only the value of oC0.a to determine whether
or not to discard all four render-target pixels.
Finally, using MRTs has performance implications. MRTs have a large
associated frame-buffer bandwidth cost, especially when using the wider bitdepths. For example, rendering to four A32R32G32B32F surfaces consumes 16
times the frame-buffer bandwidth of rendering to a single A8R8G8B8! In
addition, the GeForce 6 & 7 Series has a performance sweet spot when using
three or fewer render-targets simultaneously.
The following general performance advice thus applies: Use MRTs only as
needed, i.e., when MRT saves multiple passes. Minimize the number of rendertargets and their bit-depths, for example by tightly packing your data and not
wasting the alpha channels. Make sure to allocate MRT render-targets early (see
Section 3.6.4. Allocating Memory).
Also, split MRT outputs into groups of 3 or fewer, when doing so does not
increase the total number of passes. For example, if your application renders an
ambient pass, followed by a pass which outputs to 4 MRTs, consider outputting
41
GeForce 6 Series Programming Tips
one of the targets during the ambient pass, and then outputting to just 3 MRTs.
Doing so is particularly beneficial if one of the targets can be stored at a lower
precision than the others, and is easily computed independently of the other
targets (e.g., a material diffuse texture map). You can learn more about deferred
shading in version 7.1 of our SDK or by downloading our Deferred Shading
demo clip at ftp://download.nvidia.com/developer/Movies/NV40-LowResClips/Deferred_Shading.avi.
4.9.
Vertex Texturing
The GeForce 6 & 7 Series GPUs support vertex texturing, but vertex textures
should not be treated as constant RAM. Vertex textures generate latency for
fetching data, unlike true constant reads. Therefore, the best way to use vertex
textures is to do a texture fetch and follow it with many arithmetic operations to
hide the latency before using the result of the texture fetch.
Vertex Textures are not a substitute for large arrays of constants. They
instead are designed for sparse per vertex data, such that there are only a small
number of vertex texture fetches per vertex.
Learn more about vertex textures by reading our Using Vertex Textures
whitepaper, which is available at:
http://developer.nvidia.com/object/using_vertex_textures.html.
4.10. General Performance Advice
The GeForce 6 & 7 Series architecture contains numerous improvements that
make it more efficient and versatile. Here are some tips to help you take
advantage of its capabilities:
42
‰
Use write masks and swizzles. The GeForce 6 & 7 Series shader
architecture is able to schedule portions of 4-component vectors on
different units (through co-issue and dual-issue), which improves shader
utilization. By using write masks and swizzles, you can help the compiler to
identify these types of schedule opportunities.
‰
Use partial precision whenever possible. There are two reasons to use
partial precision with GeForce 6 & 7 Series GPUs. First, the GeForce 6 & 7
Series have a special free fp16 normalize unit in the shader, which allows
16-bit floating-point normalization to happen very efficiently in parallel
NVIDIA GPU Programming Guide
with other computations. To take advantage of this, simply use partial
precision whenever appropriate in your programs. The second reason is
that partial precision helps to reduce register pressure, potentially resulting
in higher performance.
‰
Use dynamic branching when the branches will be fairly coherent. As
mentioned in Section 4.1.3, dynamic branching can make code faster and
easier to implement. But in order for it to work optimally, branches should
be fairly coherent (for example, over regions of roughly 30 x 30 pixels).
4.11. Normal Maps
If normal map storage is an issue for your application, normal map compression
for unit-length tangent-space normals can be achieved on GeForce 6 & 7 GPUs
by utilizing a technique known as hemisphere remapping to eliminate the need
to store one of the components:
N.z = sqrt( 1 – N.x*N.x – N.y*N.y );
This compiles to 3-5 pixel shader instructions, so this technique may or may
yield a performance improvement, depending on whether these instructions can
be co-issued with other (previously existing) shader instructions, and whether or
not fetching textures is a bottleneck.
If you decide to pursue hemisphere remapping, the preferred texture formats on
GeForce 6 & 7 Series GPUs are D3DFMT_V8U8 in DirectX, and
GL_LUMINANCE8_ALPHA8 in OpenGL. These formats are 16 bits/pixel,
which provide 2:1 lossless compression.
Note that hemisphere remapping does reduce some flexibility, since only
positive, unit-length normals are generated. Techniques which rely on negative
values (such as object-space normal mapping) or non-unit normals (such as the
antialiasing technique described in
http://developer.nvidia.com/object/mipmapping_normal_maps.html) will not
be possible.
To learn more about normal maps, please see our Bump Map Compression
whitepaper at
http://developer.nvidia.com/object/bump_map_compression.html.
To create high quality normal maps that make a low-poly model look like a
high-poly model, use NVIDIA Melody. Simply load your low poly working
model, then load your high-poly reference model, click the "Generate Normal
Map" button and watch Melody go to town. Melody is available at
http://developer.nvidia.com/object/melody_home.html.
43
Chapter 5.
GeForce FX Programming Tips
This chapter presents several useful tips that help you fully use the capabilities
of the GeForce FX family. These are mostly feature oriented, though some may
affect performance as well.
5.1.
Vertex Shaders
The powerful GeForce FX vertex engine can achieve over 200 million triangles
per second on the GeForce FX 5900. It supports full dynamic branching, which
allows a shader author to branch, based on the number and type of lights,
bones, and so forth.
Each active branch thread will slow the overall execution of the program, so
branch only to save significant vertex calculations or to increase primitive batch
sizes through the API.
5.2.
Pixel Shader Length
The GeForce FX can natively handle 512 pixel instructions per pass in
Direct3D and 1,024 in OpenGL. In DirectX, compile to the ps_2_a or
ps_2_x profiles. In OpenGL, you can use GLSL, Cg (by compiling to the
arbfp1 or fp30 profiles), or use the ARB_fragment_program extension
directly.
45
GeForce FX Programming Tips
Quadro FX cards can handle 2,048 pixel instructions per pass.
5.3.
DirectX-Specific Pixel Shaders
The latest DirectX 9 ps_2_0 and higher shading models require that math and
temporaries be computed at 24-bit or higher precision by default (the GeForce
FX family uses the 32-bit float type for this case). Applications can specify a
_pp modifier on the assembly to achieve 16-bit floating-point precision.
When you use HLSL or Cg, this task is very easy—declare your variables as
float for 32-bit precision, and half for 16-bit precision.
We recommend writing your shaders the most convenient way first, and then
optimizing for register use and half-precision as needed.
Fixed-point blending is mainly used in the fixed-function texture blending
pipeline and in the texture arithmetic sections of shader models ps_1_0 –
ps_1_4.
In DirectX, you cannot request fixed-point precision through ps_2_0. If you
can fit a program into ps_1_1 – ps_1_4, it will often run faster because of the
increased usage of the fixed-point shading hardware.
5.4.
OpenGL-Specific Pixel Shaders
The ARB_fragment_program extension requires 24-bit floating-point
precision at a minimum, by default. You can place various flags at the top of the
ARB_fragment_program source code to control precision:
‰ NV_fragment_program.
Allows the half (16-bit floating-point) and
fixed (12-bit fixed-point) formats to be used explicitly for maximum
control and performance.
‰ ARB_precision_hint_fastest.
The Unified Compiler will determine
the appropriate precision for each shader operation at run-time, increasing
overall performance with minimal visual artifacts.
‰ ARB_precision_hint_nicest.
float precision.
46
Forces the entire program to run in
NVIDIA GPU Programming Guide
5.5.
Using 16-Bit Floating-Point
Many developers haven’t worked with half before, and view float as a
format that “just works,” with very few concerns about range and precision.
half has 10 mantissa bits and five bits of exponent. float has 23 bits of
mantissa and 8 bits of exponent. Each mantissa bit can be viewed as a tick mark
on a ruler. This dictates the maximum precision of the format. In half, the
precision between each mantissa bit is .1%. The exponent value can be viewed
as the length of the ruler. As the ruler gets longer with higher exponents, each
tick on the ruler represents a greater unit of length. This is how floating-point
formats automatically trade precision for range.
With half, you can only exactly represent the integers from –2048 to 2048,
with no fractional bits left over. If you perform view- or world-space
calculations in the pixel shader, you may run out of precision. For instance, if
your character is at position 4096 and the light is at position 4097, both
characters are represented by the same 16-bit floating point number. When you
subtract these values, you get zero. Then, squaring and normalizing the result
yields INF. The workaround is simple, elegant, and even faster than the original
shader: Move matrix and vector subtraction operations into the vertex shader.
Problems with lighting calculations when you first use a GeForce FX GPU can
often be traced to misuse of the half format. This is not surprising because the
half type is not a common format for games, although it is used extensively in
films for colors.
Vertex shaders are required at a minimum to support float, so they can easily
handle large world and view spaces. Also, it’s easy to see why linear calculations
belong in the vertex shader to begin with—why recalculate something at each
pixel when it can be calculated only per vertex and iterated for free by the
GPU?
Therefore, our recommendation is to move vertices into light space or tangent
space in the vertex shader, and pass the resulting position down to the pixel
shader. One nice approach is to subtract light positions from vertex positions in
the vertex shader. Then uniformly scale the vector by a constant to make the
value closer to zero, and normalize the vector in the fragment shader. (Note
that the uniform scale doesn’t affect the normalized result.)
Often the goal is to perform lighting in tangent space anyway. This is especially
useful because it lets you center the coordinate system on the vertex. Working
near zero gives you the sign bit to work with as well, so you have an extra
mantissa bit to work with.
47
GeForce FX Programming Tips
In general, perform constant calculations on the CPU, linear calculations in the
vertex shader, and nonlinear calculations in the pixel shader.
5.6.
Supported Texture Formats
Integer Formats
R8G8B8
A8R8G8B8
X8R8G8B8
R5G6B5
X1R5G5B5
A1R5G5B5
A4R4G4B4
R3G3B2
A8
A8R3G3B2
X4R4G4B4
A2B10G10R10
A8B8G8R8
X8B8G8R8
G16R16
A2R10G10B10
A16B16G16R16
A8P8
P8
L8
L16
A8L8
A4L4
2D
N
Y
Y
Y
Y
Y
Y
N
N
N
N
N
N
N
Y
N
N
N
N
Y
Y
Y
N
Cube
N
Y
Y
Y
Y
Y
Y
N
N
N
N
N
N
N
Y
N
N
N
N
Y
Y
Y
N
3D
N
Y
Y
Y
Y
Y
Y
N
N
N
N
N
N
N
Y
N
N
N
N
Y
Y
Y
N
MIP
N
Y
Y
Y
Y
Y
Y
N
N
N
N
N
N
N
Y
N
N
N
N
Y
Y
Y
N
Filter
N
Y
Y
Y
Y
Y
Y
N
N
N
N
N
N
N
Y
N
N
N
N
Y
Y
Y
N
sRGB
N
Y
Y
Y
Y
Y
Y
N
N/A
N
N
N
N
N
N
N
N
N
N
N
N
N
N
Render
N
Y
Y
Y
Y
N
N
N
N
N
N
N
N
N
N
N
N
N
N
N
N
N
N
Blend
N
Y
Y
Y
Y
N
N
N
N
N
N
N
N
N
N
N
N
N
N
N
N
N
N
Vertex
N
N
N
N
N
N
N
N
N
N
N
N
N
N
N
N
N
N
N
N
N
N
N
Float Formats
R16F
G16R16F
A16B16G16R16F*
R32F
G32R32F
A32B32G32R32F*
2D
N
Y
Y
Y
N
Y
Cube
N
Y
N
Y
N
N
3D
N
Y
N
Y
N
N
MIP
N
Y
N
Y
N
N
Filter
N
N
N
N
N
N
sRGB
N/A
N/A
N/A
N/A
N/A
N/A
Render
N
Y
Y
Y
N
Y
Blend
N
N
N
N
N
N
Vertex
N
N
N
N
N
N
Shadow Map
2D Cube 3D MIP Filter sRGB
Render
Blend
D24X8
Y
N
N
Y
Y
N/A
Y
N/A
D24S8
Y
N
N
Y
Y
N/A
Y
N/A
D16
Y
N
N
Y
Y
N/A
Y
N/A
* Exposed in DX9.0c without wrapping or mipmapping capability
Vertex
N
N
N
48
NVIDIA GPU Programming Guide
5.7.
Using ps_2_x and ps_2_a in
DirectX
GeForce FX supports several features beyond ps_2_0 functionality, including
derivative calculations (through DDX and DDY), longer shaders, and
predication support. You can access this functionality from a high-level shading
language in two ways. One is to compile using HLSL or Cg with the ps_2_x
profile, which uses the above functionality but does not check against a
particular set of capability bits. The better approach is to use the ps_2_a
profile, which is a profile that exactly matches the GeForce FX family’s shading
capabilities and produces more optimized code.
5.8.
Using Floating-Point Render
Targets
The GeForce FX family supports four-component floating-point render targets
in 64-bit and 128-bit, and one- and two-component 32-bit floating point render
targets and mipmapped textures.
Under OpenGL, floating-point textures are exposed through the
NV_float_buffer extension, and multiple low-precision components can be
packed into a single larger-precision component using the pack/unpack
instructions in NV_fragment_program. Applications can emulate
NEAREST_MIPMAP_NEAREST floating-point cubemaps and volume textures
using the pack/unpack instructions on ordinary RGBA8 textures (support is
limited to a single fp32 or 2 fp16 components).
On GeForce FX hardware, floating-point render targets don’t support blending
and floating-point textures larger than 32-bits per texel don’t support
mipmapping or filtering.
5.9.
Normal Maps
GeForce FX, GeForce4, and GeForce3 GPUs have dedicated hardware to
normal map hemisphere remapping from a 2-component normal map. For
these GPUs, you should use the CxV8U8 format, since on these parts it may be
faster than to use this format rather than deriving Z in the shader.
49
GeForce FX Programming Tips
For more information about normal maps, please see our Bump Map
compression whitepaper at
http://developer.nvidia.com/object/bump_map_compression.html.
To create high quality normal maps that make a low-poly model look like a
high-poly model, use NVIDIA Melody. Simply load your low poly working
model, then load your high-poly reference model, click the "Generate Normal
Map" button and watch Melody go to town. Melody is available at
http://developer.nvidia.com/object/melody_home.html.
5.10. Newer Chips and Architectures
The GeForce FX architecture is now available in a wide range of models and
prices. Each product in the family has the same vertex and pixel shading
features. These features let developers easily scale their games targeting
GeForce FX and beyond—for example, by handling scalability through only
geometry LOD or screen resolution. The only differences are with respect to
performance: that is, the number of vertex and pixel processors.
GeForce 6 & 7 Series GPUs have much faster float performance, but
continue to support half because of its higher performance. They are also
more orthogonal with respect to the float and half types, floating-point
render targets, and floating-point textures compared to their fixed-point
counterparts.
5.11. Summary
The GeForce FX, GeForce 6 Series, and GeForce 7 Series architectures have
the most flexible shader capabilities in the industry—from long shader
programs to true derivative calculations. However, on GeForce FX hardware,
pure floating-point shaders do not run as fast as a combination of fixed- and
floating-point shaders.
For many shaders, the best way to achieve maximum performance on the
GeForce FX architecture may be to use a mixture of ps_1_* and ps_2_*
shaders. For instance, for per-pixel lighting it may be faster to do the diffuse
lighting term in a ps_1_1 shader, and the specular term in another pass using a
ps_1_4 or ps_2_0 shader.
50
Chapter 6.
General Advice
This chapter covers general advice about programming GPUs that can be
leveraged across multiple GPU families.
6.1.
Identifying GPUs
In the past, developers often queried a GPU’s device ID (through Windows) to
find out what GPU they were running on. The device IDs have historically been
monotonically increasing. However, with the GeForce 6 & 7 Series GPUs, this
is no longer the case. Therefore, we recommend that you rely on caps bits (in
DirectX) or the extensions string (in OpenGL) to establish the features of the
GPU you’re running on. If you’re using OpenGL’s renderer string, don’t forget
that NV40-based chips do not all have an “FX” moniker in their name (they are
named “GeForce 6xxx” or “Quadro FX x400”). Similarly, G70-based chips are
named “GeForce 7xxx”.
Device IDs are often used by developers to try to reduce support calls. If you
mishandle Device IDs, you will instead create support calls. Often, when we
create a new GPU, many applications will not recognize it and fail to run.
One key idea that cannot be stressed enough is that not recognizing a Device
ID does not give you any information. Do not take any drastic action just
because you do not recognize a Device ID.
The only reasonable use of Device ID is to take action when you recognize the
ID, and you know there is a special capability or issue you wish to address.
51
GeForce FX Programming Tips
Some games are failing to run on GeForce 6 & 7 Series GPUs because they misidentify the GPU as a TNT-class GPU, or don’t recognize the Device ID. This
behavior creates a support nightmare, as the NV4X and G70 generation of
chips is the most capable ever, and yet some games won’t run due to poor
coding practices.
Device IDs are also not a substitute for caps and extension strings. Caps have
and do change over time, due to various reasons. Mostly, caps will be turned on
over time, but caps also get turned off, due to specs being tightened up and
clarified, or simply the difficulty or cost of maintaining certain driver or
hardware capabilities.
Render target and texture formats also have been turned off from time to time,
so be sure to check for support.
If you are having problems with Device IDs, please contact our Developer
Relations group at [email protected]
The current list of Device IDs for all NVIDIA GPUs is here:
http://developer.nvidia.com/object/device_ids.html.
6.2.
Hardware Shadow Maps
NVIDIA hardware from the GeForce 3 GPU and up supports hardware
shadow mapping in OpenGL and DirectX. “Hardware shadow mapping”
means that we have dedicated special transistors specifically for performing the
shadow map depth comparison and percentage-closer filtering operations. We
recommend that you take advantage of this feature, as it produces higher quality
filtered shadow map edges very efficiently. Because dedicated transistors exist
for hardware shadow mapping, you will lose performance and quality if you try
to emulate our shadow mapping algorithm with ps_2_0 of higher.
If you’re using shadow maps in a game engine, you may want to take a look at:
52
‰
http://developer.nvidia.com/object/hwshadowmap_paper.html
‰
http://developer.nvidia.com/object/cedec_shadowmap.html
‰
http://developer.nvidia.com/object/d3dshadowmap.html
‰
http://developer.nvidia.com/object/Shadow_Map.html
‰
Perspective Shadow Maps from SIGGRAPH 2002
NVIDIA GPU Programming Guide
‰
Perspective Shadow Maps: Care and Feeding in GPU Gems: Programming Techniques,
Tips, and Tricks for Real-Time Graphics
(http://developer.nvidia.com/GPUGems)
Simon Kozlov’s “Perspective Shadow Maps: Care and Feeding” chapter in GPU
Gems explains some improvements that make perspective shadow maps usable
in practice. We have taken the concepts in Kozlov’s chapter and implemented
them in an engine of our own as proof of concept, and we’ve found that they
work well in real-world situations. This example is available in version 7.0 and
higher of our SDK. It’s available at
http://download.nvidia.com/developer/SDK/Individual_Samples/samples.html.
In DirectX, you can create a hardware shadow map in the following way:
1) Create a texture with usage D3DUSAGE_DEPTHSTENCIL
2) The format should be D3DFMT_D16, D3DFMT_D24X8 ( or
D3DFMT_D24S8, but you can’t access the stencil bits in your shader )
3) Use GetSurfaceLevel(0) to get the IDirectDrawSurface9
Interface
4) Set the Surface pointer as the Z buffer in SetDepthStencilSurface()
5) DirectX requires that you set a color render target as well, but you can
disable color writes by setting D3DRS_COLORWRITEENABLE to zero.
6) Render your shadow-casting geometry into the shadow map z buffer
7) Save off the view-projection matrix used in this step.
8) Switch render targets and z buffer back to your main scene buffers
9) Bind the shadow map texture to a sampler, and set the texture coordinates
to be:
V’ = Bias(0.5/TexWidth, 0.5/TexHeight, 0) *
Bias(0.5, 0.5, 0) *
Scale(0.5,0.5,1) *
ViewProjsaved * World * Object * V
The matrices can be concatenated on the CPU, and the concatenated
transformation can be applied in either a vertex shader or using the fixedfunction pipeline’s texture matrices.
10) If using the fixed-function pipeline or ps_1.0-1.3, set the projection flags to
be D3DTTFF_COUNT4 | D3DTTFF_PROJECTED.
11) If using pixel shaders 1.4 or higher, perform a projected texture fetch from
the shadow map sampler.
53
GeForce FX Programming Tips
12) The hardware will use the shadow map texture coordinate’s projected x and
y coordinates to look up into the texture.
13) It will compare the shadow map’s depth value to the texture coordinate’s
projected z value. If the texture coordinate depth is greater than the
shadow map depth, the result returned for the fetch will be 0 (in shadow);
otherwise, the result will be 1.
14) If you turn on D3DFILTER_LINEAR for the shadow map sampler, the
hardware will perform 4 depth comparisons, and bilinearly filter the results
for the same cost as one sample—this just makes things look better.
15) Use this value to modulate with your lighting
Early NVIDIA drivers (version 45.23 and earlier) implicitly assumed that
shadow maps were to be projected. This behavior changed with NVIDIA
drivers 52.16 and later—programmers now need to explicitly set the appropriate
texture stage flags. In particular, to use shadow-maps in the ps.2.0 shader model
one has to explicitly issue a projective texture look-up (for example,
tex2Dproj(ShadowMapSampler, IN.TexCoord0).rgb). Emulating the
same command by doing the w-divide by hand, followed by a non-projective
texture look-up does not work! For example, tex2D(ShadowMapSampler,
IN.TexCoord0/IN.TexCoord0.w) does not work.
Similarly, when using the ps1.1-1.3 shader models, drivers version 52.16 and
later now require that the projective flag is explicitly set for the texture-stage
sampling the shadow-map (for example, SetTextureStageState(0,
D3DTSS_TEXTURETRANSFORMFLAGS, D3DTTFF_PROJECTED).
NOTE: In ForceWare 61.71 and later, any texture instruction (tex2D,
tex2Dlod, etc.) will work correctly with shadow maps.
Our SDK contains simple examples of setting up hardware shadow mapping in
both DirectX and OpenGL. They are available at:
http://download.nvidia.com/developer/SDK/Individual_Samples/samples.html.
54
Chapter 7.
2D and Video Programming
In addition to delivering phenomenal 3D performance, GPUs are of course able
to shoulder 2D and video programming tasks. This chapter focuses on
performance advice for that area of computer graphics.
7.1.
OpenGL Performance Tips for
Video
When used properly, Quadro FX, GeForce 6 Series, and GeForce 7 Series
GPUs can achieve extremely high levels of performance for video applications.
This section explains some of the guidelines video developers should follow to
get the most out of the GPU.
Video developers writing a playback application are concerned with uploading
textures as quickly as possible to the video card. Video typically has a video
frame size with a resolution that is not a power of 2, some typical video
resolutions are:
Standard Definition Video (SD)
High Definition Video (HD)
720 × 480i
720 × 576i
1280 × 720p
1920 × 1080i
There are three methods that can be used for texturing video images:
55
GeForce FX Programming Tips
1. GL_TEXTURE_2D
POT (power of two) texture coordinates range from [0..uscale] x [0..vscale].
2. GL_TEXTURE_2D
NP2 (non-power of two) texture coordinates range from [0..1] x [0..1].
3. GL_TEXTURE_RECTANGLE_NV
NP2 (non-power of two) texture coordinates range from
[0..width] x [0..height].
7.1.1.
POT with and without Mipmaps
A POT texture with the source as a NP2 texture size requires the driver to pad
extra data for each texture upload. With mipmapping enabled, the driver needs
to create all the subsequent mipmap levels. Both of these impose a significant
performance penalty. Texture coordinates must range from [0..uscale] x [0..vscale].
This [uscale, vscale] factor shown below handles all various image widths and
heights.
⎡ NP 2 width
NP 2 height ⎤
⎤
⎡
,
⎥
⎢u scale , v scale ⎥ = ⎢ POT
POTheight ⎦⎥
⎢⎣
⎦
⎣
width
This method to texture video supports all mipmap and non-mipmap texture
filtering as well as all the texture wrap and border modes.
To specify a POT texture size, use this parameter for target.
glTexSubImage2D (GL_TEXTURE_2D, … )
7.1.2.
NP2 with Mipmaps
A NP2 texture with mipmapping enabled is useful for compositing Standard
Definition video with computer generated graphics. This technique has the
same drawbacks as POT textures, as the driver will need to generate all mipmap
levels and perform texture padding when textures are being uploaded.
Before using this option, the application first needs to query for the OpenGL
extension string ARB_texture_non_power_of_two. If this capability is
supported by the driver, this method may be used. Texture coordinates must
range from [0..1] x [0..1].
This method to texture video supports all mipmap texture filtering and texture
wrap and border modes.
56
NVIDIA GPU Programming Guide
To specify a NP2 texture size, use this parameter for target:
glTexSubImage2D (GL_TEXTURE_2D, … )
7.1.3.
NP2 without Mipmaps (Recommended)
For video playback at HD resolutions, the goal is to upload each texture to the
video card as quickly and efficiently as possible. This method eliminates the
need for texture padding and the mipmap creation. NP2 textures have the
benefit of using less video memory than POT textures and reduce bandwidth
requirements significantly. Texture coordinates are non-normalized and range
from [0..width] x [0..height]. There are some restrictions: NP2 textures may not
use mipmap filtering and do not support the optional 1-texel border. For
texture wrap and border modes, only the following are supported:
‰ GL_CLAMP
‰ GL_CLAMP_TO_EDGE
‰ GL_CLAMP_TO_BORDER_ARB.
All non-mipmap texture filtering is supported with one exception. For the
GeForce FX family and the GeForce 6200 GPU, only GL_NEAREST filtering is
supported for 16-bit floating point textures. For GeForce 6 & 7 Series as well as
NV4XGL-based Quadro FX GPUs, filtering is supported with 16-bit floating
point textures.
To specify a NP2 texture size, use the following parameter for target:
glTexSubImage2D (GL_TEXTURE_RECTANGLE_NV, … )
7.1.4.
Texture Performance with Pixel Buffer Objects (PBOs)
PBOs enable the fast path for texture uploads/downloads in the driver. There is
an excellent example with source code included in the NVIDIA SDK 8.5 (and
newer versions) that uses ARB approved extensions for PBO’s. This shows
example demonstrates how to optimally use OpenGL Pixel Buffer Objects
(PBOs) for transferring textures to and from the GPU.
This example can be found at:
http://download.nvidia.com/developer/SDK/Individual_Samples/samples.html.
57
Chapter 8.
NVIDIA SLI and Multi-GPU
Performance Tips
This chapter presents several useful tips that allow your application to get the
maximum possible performance boost when running on a multi-GPU
configuration, such as NVIDIA’s SLI technology. For more information, please
see also
ftp://download.nvidia.com/developer/presentations/2004/GPU_Jackpot/SLI
_and_Stereo.pdf.
8.1.
What is SLI?
SLI, which stands for “Scalable Link Interface”,
allows multiple GPUs to work together to
deliver higher performance. SLI-certified
motherboards are PCI Express motherboards
with two x16 physical lanes; each one of these
slots accepts a PCI Express graphics board.
When the two graphics boards are then linked
via an external SLI bridge connector, the driver
recognizes the configuration and allows you to
enter SLI Multi-GPU mode. In SLI Multi-GPU mode the driver configures
both boards as a single device: the two GPUs look like a single logical device to
all graphics applications.
59
NVIDIA SLI and Multi-GPU Performance Tips
This single logical device runs up to 1.9 times faster than a single GPU, since
the driver splits the rendering load across the two physical GPUs. Note that
running in SLI mode does not double available video memory. For example,
plugging in two 256 MB graphics boards still only results in a device with at
most 256 MB of video memory. The reason is that the driver generally
replicates all video memory across both GPUs. That is, at any given time the
video memory of GPU 0 contains the same data as the video memory of GPU
1.
When running an application on a SLI system, it may run in one of several
modes: compatibility mode, alternate frame rendering (AFR), or split frame
rendering (SFR) mode.
Compatibility mode simply uses only a single GPU to render everything (that is,
the second GPU is idle at all times). This mode cannot show any performance
enhancement, but also ensures an application is compatible with SLI.
For AFR the driver renders all of frame n on GPU 0 and all of frame n+1 on
GPU 1. Frame n+2 renders on GPU 0 and so on. As long as each frame is
self-contained (that is, frames share little to no data) AFR is maximally efficient,
since all rendering work, such as per-vertex, rasterization, and per-pixel work
splits evenly across GPUs. If some data is shared between frames (for example,
reusing previously rendered-to textures), the data needs to transfer between the
GPUs. This data transfer constitutes communications overhead preventing a
full 2x speed-up.
For SFR the driver assigns the top portion of a frame to GPU 0 and the bottom
portion to GPU 1. The size of the top versus the bottom portion of the frame
is load balanced: if GPU 0 is underutilized in a frame, because the top portion is
less work to render than the bottom portion, the driver makes the top portion
larger in an attempt to keep both GPUs equally busy. Clipping the scene to the
top and bottom portions for, respectively, GPU 0 and 1, attempts to avoid
processing all vertices in a frame on both GPUs.
SFR mode still requires data sharing, for example, for render-to-texture
operations. Because AFR generally has less communications overhead and
better vertex-load balancing than SFR, AFR is the preferred mode. Sometimes,
however, AFR fails to apply, for example, if an application limits the maximum
number of frames buffered to less than two.
60
NVIDIA GPU Programming Guide
8.2.
Choosing SLI Modes
Applications in development default to compatibility mode in current drivers
(66.93 and later). Application developers can force AFR and SFR modes by
creating an application-specific profile in the driver, i.e., in the display driver
control panel, click on the GeForce tab, select “Performance & Quality
Settings,” click on “Add Profile,” and enter the name of your application’s
executable. Then enable “Show advanced settings” and choose “AFR” or
“SFR”.
Another option is for the application to programmatically choose an SLI mode.
All NVIDIA drivers support a control panel API that lets applications query
and set various control panel options. As part of that API the
NvCplGetDataInt() entry point allows querying for how many total GPUs
are installed and how many of those are currently in SLI mode. The
NvCplSetDataInt() entry point lets application set whether the calling
application runs in compatibility, AFR, or SFR mode.
Complete documentation of these interfaces and source code showing how to
use these interfaces are available as part of the NVSDK; the NVCpl sample in
particular exercises the SLI detection and mode-setting functionality
(http://developer.nvidia.com/object/sdk_home.html).
The following is a list of dos and don’ts to get a maximum performance boost
on an SLI system.
8.3.
Avoid CPU Bottlenecks
If your application is CPU-bound running it on a more powerful graphics
solution has little to no performance impact. To take advantage of a multi-GPU
configuration, you must avoid becoming bottlenecked by the CPU. Chapter 2 of
this Programming Guide provides guidance how to detect and avoid CPU
bottlenecks.
Similarly, if your application artificially throttles its frame rate to an arbitrary and
fixed value, then the frame rate cannot exceed that value. Please avoid that
situation.
61
NVIDIA SLI and Multi-GPU Performance Tips
8.4.
Disable VSync by Default
Enabling VSync forces the frame rate to (at best) match the monitor refresh
rate. Multi-GPU configurations are likely to achieve frame rates that are much
higher than monitor refresh rates, due to their higher graphics performance.
These frame rates are thus only attainable if VSync is turned off.
For some applications, i.e., visual simulations, VSync is an absolute requirement
to avoid tearing. In these cases SLI still demonstrates significant benefit when
the graphics workload and fill-rate requirement are high. For example, using SLI
it may become possible to run those applications in resolutions of 1920×1200
or 2048×1536 along with 4× or higher full-scene antialiasing at real-time rates.
Triple buffering is not a solution for attaining higher frame rates: Triplebuffering simply allocates an additional back buffer that a graphics adapter can
render to.
With triple buffering, a graphics adapter can render into up to three buffers
round-robin style. But if the graphics adapter consistently finishes rendering
each buffer at a rate higher than the monitor refresh rate, then the number of
back buffers is irrelevant; the monitor refresh continues to gate the overall
frame rate.
Worse, triple buffering has two significant disadvantages:
1. It consumes more video memory (in the case that screen resolution is high
and antialiasing is enabled, the amount of additional video memory
consumed is significant).
2. It increases lag as more frames are in flight between issuing a rendering
command and seeing it onscreen.
To turn off VSync in DirectX, the PresentationInterval member of the
D3DPRESENT_PARAMETERS structure has to be set to
D3DPRESENT_INTERVAL_IMMEDIATE when calling
IDirect3D9::CreateDevice().
62
NVIDIA GPU Programming Guide
8.5.
DirectX SLI Performance Tips
8.5.1.
Limit Lag to At Least 2 Frames
DirectX allows drivers to buffer up to three frames. Buffering frames is
desirable to ensure that CPU and GPU can work independent of one another
and thus achieve maximum performance. On the other hand, the more frames
are buffered the longer it takes between issuing a command and seeing its result
on-screen. This lag is generally undesirable, since humans can detect and object
to lag-times of as little as 30 ms (depending on the test scenario).
Some games thus artificially limit how many frames are buffered. For example,
locking the back-buffer forces a hard synchronization between the CPU and the
GPU. Locking the back-buffer first stalls the CPU, drains all buffers, and then
stalls the GPU. At the end of the lock all systems are idle, and the number of
buffered frames is zero.
Stalling a system in this manner has a severe performance penalty and is to be
avoided, especially on multi-GPU configurations.
A less objectionable solution to limiting the number of buffered frames is to
insert tokens into the command stream (e.g., DirectX event queries) at the end
of each frame. Checking that these events have been consumed before issuing
additional rendering commands limits the number of buffered frames and thus
lag to anywhere between one to three frames.
Multi-GPU systems are particularly sensitive to limiting the number of buffered
frames. In general, a system with n GPUs requires at the very least n frames to
be buffered to be maximally efficient.
Surprisingly, doing so does not increase lag, since a dual GPU system generally
runs twice as fast as a single GPU system. For example, buffering two frames
(that take 15 ms each to render) on a dual GPU system has the same 30 ms lag
as buffering one frame (that takes 30ms to render) on a single GPU system.
We thus recommend that applications check how many GPUs are available and,
if they must, limit the number of buffered frames to at least that number of
GPUs. The control panel API (see Section 8.2) allows querying whether a
system is in SLI mode and how many GPUs are currently in use. In particular,
the function NvCplGetDataInt(NVCPL_API_NUMBER_OF_SLI_GPUS,
&number) returns the number of SLI-enabled GPUs in the system. The
NVControlPanel_API.pdf document and the NVSDK sample NVCpl provides
details.
63
NVIDIA SLI and Multi-GPU Performance Tips
8.5.2.
Update All Render-Target Textures in All Frames that Use Them
The efficiency of a multi-GPU system is inversely proportional to how much
data the GPUs share. In the best case, the GPUs share no data, thus have no
synchronization overhead, and thus are maximally efficient.
To minimize the amount of shared data, each rendered frame should be
independent of all previous frames. In particular, when using render-to-texture
techniques, it is desirable that all render-target textures used in a frame are also
generated during that same frame. Conversely, avoid updating a render target
only every other frame, yet using the same render-target as a texture in every
frame.
If an application is explicitly skipping render-target updates to increase
rendering speed on single GPU systems, then it might be of advantage to
modify that algorithm for multi-GPU configurations. For example, detect if the
application is running with multiple GPUs and if so, update render-targets every
frame (i.e., to increase visual fidelity) or update render-targets two frames in a
row and then skip updates two frames in a row.
Alternatively, rendering to render-targets early on and only using the result late
in the frame is also beneficial for SLI systems. It avoids stalling one GPU
waiting for the results of the other GPU’s render to texture operation.
8.5.3.
Clear Color and Z for Render Targets and Frame Buffers
Clearing the color and z information of a render target prior to its use indicates
to the driver and the GPU that any existing data in the render target is
irrelevant. Conversely, not clearing that data indicates that the data may be
relevant and thus needs to be maintained, and in the case of SLI configurations
shared between GPUs.
Clearing z is generally advisable even knowing that all z-values in the render
target are going to be overwritten later (see Section 3.6.2): clearing z is close to
free and enables early z-cull optimizations. On SLI configurations, an added
benefit is avoiding synchronizing this z-information across the GPUs.
Similarly, clearing color is advisable on SLI configurations as it avoids
synchronization overhead between the GPUs. Thus, you should always clear
color, even when knowing that every pixel in the render target is going to be
overwritten later anyway.
64
NVIDIA GPU Programming Guide
8.6.
OpenGL SLI Performance Tips
8.6.1.
Limit OpenGL Rendering to a Single Window
Optimal scaling with multiple GPUs happens with a single OpenGL window.
Any child windows your application creates should not contain an OpenGL
context.
With the AFR mode you are limited to exactly one double buffered window
with page flipping. Any additional OpenGL windows will prohibit page flipping
and force the driver to fall back to a single GPU until window positioning
makes page flipping possible again.
8.6.2.
Request PDF_SWAP_EXCHANGE Pixel Formats
When creating an OpenGL context, make sure to ask for a pixel format with
the PFD_SWAP_EXCHANGE flag set instead of PFD_SWAP_COPY. The swap
exchange flag implies that the application does not rely on the back buffer
content after a SwapBuffers() is performed.
This is essential for the AFR mode because the driver cycles through GPUs per
frame so the back buffer may be stored on another GPU.
For the SFR mode PFD_SWAP_EXCHANGE is preferred over PFD_SWAP_COPY
because the swap copy mode requires additional synchronization between the
GPUs.
8.6.3.
Avoid Front Buffer Rendering
Rendering to the front buffer requires heavy synchronization when rendering to
multiple GPUs and should therefore be avoided at all costs. OpenGL overlays
are recommended instead of front buffer rendering. They are supported on
Quadro and work on both GPUs independently in an SLI configuration. Make
sure to use a R5G5B5 high-color overlay for optimal performance on NVIDIA
GPUs.
8.6.4.
Limit pbuffer Usage
Rendering to a pbuffer on a multi GPU system requires the driver to broadcast
the rendering to both GPUs because the rendered result may be used by either
GPU later on. This limits the amount of scaling available when running your
application on multiple GPUs.
65
NVIDIA SLI and Multi-GPU Performance Tips
In some cases it is possible for the driver to detect that your application only
uses the output from a pbuffer in the same frame and can limit the pbuffer
rendering to a single GPU.
8.6.5.
Render Directly into Textures Instead of Using
glCopyTexSubImage
If your application is doing any sort of render-to-texture operation the preferred
method is to use pbuffers (WGL_ARB_pbuffer) with the render-to-texture
extension (WGL_ARB_render_texture) or the upcoming framebuffer object
extension (GL_EXT_framebuffer_object) to render directly into a texture.
Using glCopyTexSubImage() to copy data from a pbuffer or the back
buffer is not recommended as the texture needs to get updated on both GPUs.
For more information, please see the following resources on our Developer
web site:
http://www.nvidia.com/dev_content/nvopenglspecs/GL_EXT_framebuffer_object.txt
http://www.nvidia.com/dev_content/nvopenglspecs/WGL_ARB_pbuffer.txt
http://www.nvidia.com/dev_content/nvopenglspecs/WGL_ARB_render_texture.txt
8.6.6.
Use Vertex Buffer Objects or Display Lists
To limit the CPU overhead when specifying geometry, your application should
use vertex arrays with the vertex buffer object extension
(GL_ARB_vertex_buffer_object) or display lists instead of immediate
mode calls.
Here are some general VBO performance tips:
66
‰
Load your VBO working set first or textures may block faster memory until
the working set stabilizes.
‰
Avoid huge batches when drawing geometry, there is a performance dropoff on NVIDIA GPUs when using more than 64k vertices per draw call.
‰
Use unsigned short for indices.
‰
Use glDrawRangeElements() instead of glDrawElements().
‰
Use the correct VBO usage hints for the type of data in the VBO
NVIDIA GPU Programming Guide
For additional VBO performance tips, please see the following resources on our
Developer web site:
http://developer.nvidia.com/object/using_VBOs.html
http://www.nvidia.com/dev_content/nvopenglspecs/GL_ARB_vertex_buffer_object.txt
8.6.7.
Limit Texture Working Set
Texture data is shared between both GPUs, so your application should limit the
amount of textures in use at a given time and not stream in new textures too
often.
For game applications it is a good idea to render some pre-warming frames to
get the textures into the right memory space to avoid low initial frame rates.
For procedural textures it’s a good idea to generate them using a shader instead
of downloading different versions of the texture every frame. This saves
bandwidth and texture memory on both GPUs.
8.6.8.
Render the Entire Frame
Avoid rendering to only a section of the frame using methods such as
glViewport() or glScissor(). This disables load balancing with the SFR
mode and hurts performance with the AFR method.
8.6.9.
Limit Data Readback
Avoid reading back the color or depth buffers using glReadPixels() and
never use glCopyPixels(). These commands cause pipeline stalls and inhibit
parallelism. However, reading the contents of the back buffer for reuse in the
same frame is ok.
8.6.10. Never Call glFinish()
The glFinish() command does not return until all pending OpenGL
commands are complete. This kills asynchronous work and reduces any
performance benefits from using multiple GPUs.
67
Chapter 9.
Stereoscopic Game Development
This section explains how NVIDIA’s stereo rendering implementation works,
and how to take full advantage of it in your applications.
9.1.
Why Care About Stereo?
Game developers often overlook one factor in their never ending quest for
greater realism in games: People see with two eyes in the real world. While
artificial stereoscopic viewing (on a screen versus in real life) is not a huge
market, many gamers enjoy the extra sense of presence obtained by playing
games with inexpensive shutter glasses along with NVIDIA’s stereo override
driver.
In addition, there are benefits to viewing your game in stereo during
development. You immediately pick up on things that look fake. Keep in mind
that motion parallax gives similar visual cues to stereo but the stereo viewer
perceives instantaneously what users see if they move around and obtain depth
information via motion parallax.
Using stereoscopic viewing while developing a game is a competitive advantage;
you see and correct visual defects before they even come out in a game. This of
course will also enhance the experience of those who play your game in stereo.
69
Stereoscopic Game Development
9.2.
How Stereo Works
The NVIDIA 3D Stereo Driver allows full-screen stereoscopic viewing of
DirectX and OpenGL based games. The 3D Stereo Driver supports red and
blue anaglyph rendering as well as page-flipped viewing suitable for shutter
glasses. With compatible hardware you’ll see the image with the perception of
depth. Please note that the Stereo Driver version must match the Display
Driver version in order to function.
When playing 3D games with the stereoscopic driver enabled, the scene will be
rendered from two viewpoints; each a little to the side of the real viewpoint, as
though they were coming from the left and right eye. This works with both
fixed-function rendering and vertex shaders.
For an accurate stereo effect, developers need consider a number of issues that
we list below.
9.3.
Things That Hurt Stereo
Here we list common things that negatively impact stereo, and provide some
ideas to work around them.
9.3.1.
Rendering at an Incorrect Depth
This problem is the number one thing you should take care of. The 3D Stereo
Driver uses the depth to create the stereo effect, so anything that is not at the
correct depth stands out like a sore thumb when viewed in Stereo.
70
‰
Place background images, sky boxes, and sky domes at the farthest possible
depth. Otherwise, the world will look like it’s in a little box in stereo.
‰
Place HUD items at their proper 3D depth. If you have name labels that
hover over objects, putting them at the 3D depth of the object gives a
better stereo effect than putting them on the near plane of the view
frustum.
‰
It’s also helpful to render the HUD as far into the scene as possible. This
trick gives you a greater perceived depth in the rest of your scene while not
causing eyestrain when looking at the HUD.
‰
Laser sights, crosshairs, and cursors do not look correct unless placed in the
3D world at the depth of the objects they are pointing to. When they aren’t,
NVIDIA GPU Programming Guide
it is almost impossible to use them since the user’s eyes are converging on
one depth, but the cursor is at another depth; users see two cursors, neither
of which point at the correct place.
‰
9.3.2.
Highlighting objects should happen at the depth of the object itself, not in
screen space.
Billboard Effects
Billboard effects look flat and bad in regular 3D; they look even worse in stereo.
In regular 3D you see the billboards as you move around, but in 3D Stereo the
problem immediately pops out at you, even in a static scene. Most billboard
effects look extremely flat in stereo, so use real geometry everywhere you can
instead, even low resolution geometry looks better.
For particle effect billboards (sparks, smoke, dust, etc), these may or may not
look ok, the best thing you can do is test your app in stereo to see what it looks
like and judge if the quality is good enough, and make sure that the billboards
have a meaningful depth in 3D.
9.3.3.
Post-Processing and Screen-Space Effects
2D screen-space effects can greatly hurt the stereo effect. Things like blurry
glow, bloom filters, image-based motion blur fall into this category. These
effects are usually created by rendering the 3D geometry to a texture, then
rendering a 2D screen aligned quad to the screen. The geometry in the texture is
no longer at the correct depth in the world as the effect should be, thus working
poorly in 3D Stereo.
You should provide the option to disable these effects for people playing in
stereo and render the geometry to the back buffer.
9.3.4.
Using 2D Rendering in Your 3D Scene
Any object rendered as 2D doesn’t really have a real 3D depth, so it is placed at
the monitor depth. This will look very flat in 3D Stereo. If you are mixing 2D
and 3D in your HUD you may have inconsistent depths leading to eye strain.
Again, render everything at the proper depth, in 3D, and test with stereo on.
9.3.5.
Sub-View Rendering
When rendering a sub view to the screen, such as a picture in picture display,
car mirror, or small map in the upper corner, you must set the viewport to
71
Stereoscopic Game Development
cover the area before rendering. This trick prevents strange stereo effects
bleeding outside of the intended section of the screen.
9.3.6.
Updating the Screen with Dirty Rectangles
If you are determining only the parts of the screen that have changed and not
updating the rest of the screen this can cause odd looking rendering in 3D
stereo. Just render all visible objects each frame.
9.3.7.
Resolving Collisions with Too Much Separation
If you are resolving collisions by pushing objects away from each other, make
sure you don’t push them away too far. It looks bad in normal 3D when moving
around but stands out right away in stereo, making things appear to hover
above the ground.
9.3.8.
Changing Depth Range for Difference Objects in the Scene
Splitting the scene into multiple depth ranges can cause distortions in the stereo
effect making some objects look shorter or elongated. All objects should be
rendered in a consistent depth range for the best stereo effect.
9.3.9.
Not Providing Depth Data with Vertices
When sending vertices for rendering to D3D for software transform and
lighting, include the RHW depth information for stereo to function properly.
9.3.10. Rendering in Windowed Mode
NVIDIA’s 3D Stereo only works when your application is in full-screen
exclusive mode. If you don’t support fullscreen mode, then game players can’t
take advantage of 3D Stereo.
9.3.11. Shadows
Rendering stencil shadows using a fullscreen shadow color quad will not work
properly in stereo. However, re-rendering shadowed objects in the scene at their
proper depth in shadow color will function correctly in stereo. Shadow maps
function fine, and projection shadows function as long as you are projecting to
the proper depth for the shadow.
72
NVIDIA GPU Programming Guide
9.3.12. Software Rendering
NVIDIA 3D Stereo drivers support DirectX and OpenGL automatically. If you
use any other APIs to render 3D geometry it will not be rendered in stereo.
9.3.13. Manually Writing to Render Targets
Don’t lock render targets and do direct writes; doing so bypasses the stereo
driver.
9.3.14. Very Dark or High-Contrast Scenes
Very dark scenes can become even darker when using 3D Stereo shutter glasses.
Providing a brightness or gamma adjustment will help this problem. Using very
bright objects on very bright and very dark objects causes ghosting, which hurts
stereo. Testing your game in stereo quickly shows whether or not this is a
problem.
9.3.15. Objects with Small Gaps between Vertices
Small gaps in meshes can become much more obvious when rendered in stereo.
Make sure your meshes are tight and test in stereo to be sure this isn’t
happening.
9.4.
Improving the Stereo Effect
Here are some ideas you can use to create a more immersive experience using
stereo.
9.4.1.
Test Your Game in Stereo
The best thing you can do to get a great 3D Stereo experience is to test your
game while running in stereo. Most problems are obvious right away and simple
to fix. You can get low cost stereo kits from IODisplay (http://www.iglasses.com). The NVIDIA 3D Stereo Driver also supports Red and Blue
Anaglyph Stereo Mode, so if you have a pair of paper 3D glasses lying around
it’s even easier to test.
73
Stereoscopic Game Development
9.4.2.
Get “Out of the Monitor” Effects
You can design things in your game that are very close to the near plane but do
not intersect the edges of the view frustum. This design lets you get the great
“popping out of the monitor” stereo effect. Hovering orbs, space ships, flying
characters, and so on are a few possibilities. These all really look fantastic in
stereo.
9.4.3.
Use High-Detail Geometry
Use more polygons for objects to get better realism in 3D. This strategy is of
course always true, but even more so for stereo. Use polygons everywhere–for
buildings, plants, trees, characters… anywhere possible!
9.4.4.
Provide Alternate Views
Give the users a choice of viewpoint in your game, or at least control over the
camera. First person, third person, top down, etc, some may look better than
others in stereo.
9.4.5.
Look Up Current Issues with Your Games
Once the 3D Stereo Driver is installed, you can view its control panel and look
under “Stereo Game Configuration”. On this page you can see if your game is
listed and what issues we have already found with it. Your NVIDIA Developer
Relations contact can also help out here.
9.5.
Stereo APIs
We are currently developing two separate APIs to better support stereo:
74
‰
StereoBLT API – Display Pre-Rendered Stereo Images in 3D
Allows display of pre-created left/right images while the stereoscopic
display is active.
‰
IStereoAPI – Real-time Control Over Stereoscopic Rendering
Lets you query and control the convergence, stereo separation and other
details of the stereo driver. This API consists of a header and library that
you can use to control these settings in real time in your game.
Modifications take effect immediately and can be changed every frame.
NVIDIA GPU Programming Guide
‰
9.6.
OpenGL Quad-Buffered Stereo. This is available on the NVIDIA
Quadro GPU family, requires no special stereo driver, and works in
windowed mode too.
More Information
To learn more, contact your NVIDIA Developer Relations representative, or email [email protected] to request more info or the prerelease Stereo
APIs.
You can also find updated Stereo Information online at
http://developer.nvidia.com/object/3D_Stereoscopic_Dev.html
75
Chapter 10.
Performance Tools Overview
This section describes several of our tools that will help you identify and remedy
performance bottlenecks.
10.1. NVPerfHUD
NVPerfHUD displays four
informative graphs overlaid
on top of any DirectX
application. These graphs
show important statistics
about your application,
which helps you to identify
potential bottlenecks. The
graphs display sample data
in a heart-rate monitor
format. By scrolling from
right to left, you can see the
values of the last 256
frames.
You can get NVPerfHUD on the NVIDIA Developer Web Site at:
http://developer.nvidia.com/object/nvperfhud_home.html.
77
GPU Codename and Product Name List
10.2. NVShaderPerf
The NVShaderPerf command line utility uses the same
technology as the Shader Perf panel in FX Composer to
report shader performance metrics. It supports DirectX
and OpenGL shaders written in HLSL, GLSL, Cg,
!!FP1.0, !!ARBfp1.0, ps_1_x, and ps_2_x. You
can get performance reports for your shaders on the
entire family of GeForce 6 & 7 Series and GeForce FX
GPUs, including cycle count, register usage and a GPU utilization rating.
You can get NVShaderPerf at:
http://developer.nvidia.com/object/nvshaderperf_home.html.
10.3. NVIDIA Melody
To create high quality normal maps that make
a low-poly model look like a high-poly model,
use NVIDIA Melody. Simply load your low
poly working model, then load your high-poly
reference model, click the "Generate Normal
Map" button and watch Melody go to town.
Melody is available at
http://developer.nvidia.com/object/melo
dy_home.html.
78
NVIDIA GPU Programming Guide
10.4. FX Composer
FX Composer empowers
developers to create high
performance shaders in an
integrated development
environment with unique realtime preview and
optimization features. FX
Composer was designed with
the goal of making shader
development and
optimization easier for
programmers while providing
an intuitive GUI for artists
customizing shaders for a particular scene.
FX Composer allows you to tune your shader performance with advanced
analysis and optimization:
‰
Enables performance tuning workflow for vertex and pixel shaders
‰
Simulates performance for the entire family of GeForce 6 & 7 Series and
GeForce FX GPUs
‰
Capture of pre-calculated functions to texture look-up table
‰
Provides empirical performance metrics such as GPU cycle count, register
usage, utilization rating, and FPS.
‰
Optimization hints notify you of performance bottlenecks
You can download the latest version of FX Composer from:
http://developer.nvidia.com/fxcomposer.
10.5. Developer Tools
Questions and Feedback
We would like to receive your feedback on our tools. Please send your
comments or concerns to [email protected]
79
Was this manual useful for you? yes no
Thank you for your participation!

* Your assessment is very important for improving the work of artificial intelligence, which forms the content of this project

Download PDF

advertisement