Kenneth%20Hurley%20Sr.%20Software%20Engineer

1
Kenneth HurleySr. Software Engineer

• khurley_at_nvidia.com

2
What are the problems we are seeing when 3D
engines are written?

• Misuse of Vertex Buffers
• Concurrency Limitations
• Frame Rate Limiters
• Non-Optimized surface usage
• Cache misses
• Data Ordering

3
Misuse of Vertex Buffers

• Bad Things can happen unless you know the right
  way to use a vertex Buffer
• Dynamic vertex buffer vs. static vertex buffers
• When creating the vertex buffer, use
  D3DVBCABS_WRITEONLY
• Use D3DLOCK_DISCARDCONTENTS
• Use D3DLOCK_NOOVERWRITE
• Vertex buffer ordering
• Use ordered vertex buffers because of cache
  coherency

4
Using Vertex Buffers Correctly
5
Example vertex buffer flow

• CreateVB(WRITEONLY, 1000-12000)
• A I 0
• B Space in VB for M vertices?
• Yes Lock(NOOVERWRITE)
• No GOTO C
• Fill in M vertices at index I
• Unlock() DIPVB(I) I M GOTO B
• C Lock(DISCARDCONTENTS) GOTO A

6
Concurrency

• Why do I need it?
• Concurrency helps parallelism between the CPU and
  the GPU.
• OK, How do I achieve it?
• Use NVPAT to see if Spin Lock is happening.
• Spin Locks are when the driver has to stall
  waiting for the hardware to finish with an object
• These objects can be vertex buffers or texture
  surfaces

7
Concurrency (cont.)

• Use the vertex buffer and texture surface flags
  so the driver can give you another buffer while
  the hardware is using the other one.

8
Frame Rate Limiters

• Can cause concurrency issues
• Better ways to achieve constant frame rates
• Makes effective triangle rate much lower, because
  driver has to do some work with vertex data.

9
Frame Rate Limiter Problem

• Serialization of code loop

Rescheduled for concurrency
10
Non Optimized Surface Usage

• Locking a texture before the GPU is finished with
  it causes concurrency problems by stalling the
  CPU inside the driver.
• Typical examples include locking the backbuffer
  to do 2D operations on it
• The best solution for this is to use 2 screen
  aligned triangles (quad) instead and put them
  directly in the 3D pipeline

11
Cache Misses

• Big slowdowns can occur here
• CPU cache misses can occur because of ordering of
  vertex data. Check these carefully with VTune.
• GPU has a vertex cache also. Geforce has a 16
  entry cache, but optimal cache use is 10, because
  6 triangles can be in flight at any given time.
• GPU vertex cache statistics will be added to
  NVPAT.

12
Vertex Ordering

• Best performance is to also order vertex data and
  vertex indices in sequential order. This helps
  both the CPU and the GPU
• Out of order vertices makes the CPU hit the cache
  more often
• It does the same thing to the GPU

13
How do we solve these problems?

• VTune
• GPT
• NVPAT

14
VTune 4.5

• Will help your application optimize for CPU
• Works well in conjunction with NVPAT
• I personally use the Time-Based Sampling Wizard
• VTune is excellent for application specific
  analysis
• It doesnt show where in the driver time is
  spent, unless you have symbols for the driver.
  You almost certainly dont have driver symbols.

15
VTune 4.5

• Flare Application

16
GPT 3.5

• Excellent tool to help you achieve maximum
  performance.
• Works on both D3D and OpenGL
• Helps with application ?? API slowdowns
• Works well in conjunction with VTune and NVPAT.
  GPT is excellent for application to
  Direct3D/OpenGL analysis.
• It still cant tell you what is occurring inside
  the driver that may be slowing your application
  down

17
GPT 3.5 (cont)

• Quad view for visual analysis modes

View of alien world in Half-Life
18
NVPAT 1.07

• Analyze interaction with driver
• Works on NVIDIA hardware only
• Windows 98/Windows 2000 capable
• Hotkey capable
• Online help via F1 function key
• Logging
• Frame Rate Display
• Natural Extension to VTune and GPT

19
NVPAT 1.07

• Demo Flare VS NewFlare
• NVPAT Available free at http//www.nvidia.com/Mark
  eting/Developer/SwDevStaticPages.nsf/pages/StatsDr
  iver
• You must be a registered NVIDIA developer

20
VTune DLL SDK

• Soon, all these performance tools should be
  integrated into VTune using the DLL SDK
• NVPAT will be integrated into the VTune DLL SDK
• VTune DLL SDK is available from Intel and gives
  you the ability to integrate performance tools
  into VTune.
• http//developer.intel.com/vtune/analyzer/vtperfd
  ll
• Common User Interface/API means less to learn for
  developers

21
Action Items

• Profile often and early in the process
• Use the tools available to you
• Some are free, the rest are reasonable
• Architect engine with concurrency in mind
• Ask for enhancements from your tool vendor

22
Questions?

• Comments/Suggestions?
• Enhancement requests for NVPAT can be sent to
  statdriver_at_nvidia.com