AMD DirectX Optimizations

1
AMD DirectX Optimizations CodeAnalyst

• Howard Dortch

2
AMDs Optimizations of DirectX for AMD Processors

• Fixed Function Pipeline (since DX5)The
  optimizations for DX5-DX7 for this fixed function
  pipeline mainly take advantage of 3DNow! SIMD
  over the reference C implementation which
  compiles to x87.
• Vertex Shader Software Rendering Pipeline (since
  DX8)
• D3DX Library (about 60 of the more important
  functions)

3
MSSDK Vertex Shader run without optimizations
4
MSSDK Vertex shader run with optimizations
5
Introduction to CodeAnalyst

• CodeAnalyst helps programmers find slow code and
  gives hints as to how to make it run faster on
  AMD processors. It does this using a Timer Based
  Sampler to show what part of the code is getting
  hit the most. Then it allows the user to take a
  closer look at an area of interest by simulating
  the code.
• TSIM - The Trace Simulator included in
  CodeAnalyst was designed to allow the user to
  simulate code on all AMD processors regardless of
  the machine used for testing.

6
What Needs Optimization?

• Servers - The online gaming industry is growing
  fast. Game servers have to do a lot of
  calculations.
• Games - Video games are becoming more complex.
  Game designers have to write software for high
  end and low-end graphics cards. No matter how
  powerful, the graphics cards, or the API,
  designers will push the limits.
• Streaming video - Moving large quantities of data
  through a computer is processor and memory
  bandwidth intensive.
• Math intensive programs - Any software that moves
  data and performs math functions on the data is a
  candidate for optimization.

7
CodeAnalyst Demo

• MSVC build in release mode with debug info.

8
CodeAnalyst Demo

• MSVC build in release mode with debug info
  (cont).

9
Setup CodeAnalyst project

• Starting a new project is easy, just follow the
  path to the exe file in the release directory and
  click on it. You can then select any special
  dlls or pass parameters.

10
Timer Based Profiling is easy just click the
button!
11
Using the Trace Simulator

• Set start point

12
Using the Trace Simulator

• set end point

13
Using the Trace Simulator

• And GO!

14
Using the Trace Simulator

• Getting results

15
Code Sample 1

• /
  
• Xform
• THIS FUNCTION SHOWS AN UNOPTIMIZED FUNCTION
• 
  /
• void Xform(float r, float V, Mat M)
• for (int i0ilt4i)
• ri0
• for (int j0jlt4j)
• riMjiVj

16
Code Sample 2

• /
  
• XformU
• THIS FUNCTION IS OPTIMIZED BY UNROLLING THE LOOP
• 
  /
• void XformU(float r, float V, Mat M)
• r0 M00 V0 M10 V1 M20
  V2 M30 V3
• r1 M01 V0 M11 V1 M21
  V2 M31 V3
• r2 M02 V0 M12 V1 M22
  V2 M32 V3
• r3 M03 V0 M13 V1 M23
  V2 M33 V3

17
Final Slide

• The AMD Code Optimization Guide has lots of good
  information
• www.amd.com/products/cpg/athlon/techdocs/
• AMD Developers Connection
• www.amd.com/devconn
• More info on CodeAnalyst can be found at
• www.amd.com/devconn/3dsdk/canalyst.html
• Visit the AMD Suite
• Email howard.dortch_at_amd.com