NVIDIA OpenGL Update Simon Green

NVIDIA OpenGL Update Simon Green

Overview SLI How it works OpenGL Programming Tips SLI Futures New extensions NVX instanced arrays – OpenGL instancing! EXT timer query Copyright NVIDIA Corporation 2004

What is SLI? Allows scaling graphics performance by combining multiple GPUs in a single system Works best with NVIDIA nForce motherboards Improves rendering performance up to 2x with two GPUs Copyright NVIDIA Corporation 2004

SLI-Ready PCs affordable for Everyone nForce4 SLI Motherboard Dual GeForce 6800 GTs nForce4 SLI Motherboard GeForce 6600 LE Copyright NVIDIA Corporation 2004

Quad SLI 4 GPUs is better than 2! 2 cards, each with 2 GPUs Copyright NVIDIA Corporation 2004

SLI Notebooks Copyright NVIDIA Corporation 2004

SLI Game Performance Scaling GeForce 7800 GTX GeForce 7800 GTX SLI Project Snowblind Brothers in Arms Thief: Deadly Shadows Code Creatures Chronicles of Riddick Splinter Cell Chaos Theory Ground Control 2 X2 Doom 3 3Dmark05 3DMark03 Painkiller: Battle Out of Hell Tests run on AMD FX 55 with 1GB memory Colin McRea 5 0.8 Copyright NVIDIA Corporation 2004 1600x1200, 4X/8X 1 1.2 1.4 1.6 1.8 2

How SLI Works Plug multiple GPUs into motherboard Have to be same model currently NVIDIA driver reports as one logical device Video memory does NOT double Video scan out happens from one board Bridge connector transmits digital video between boards Copyright NVIDIA Corporation 2004

SLI and Game Development Developing a game now takes 2 years or more CPU performance doubles (or less) GPU performance quadruples CPU / GPU balance shifts Worse: CPU-hungry modules are developed later AI, physics, full game play SLI allows you to preview future GPU performance now Copyright NVIDIA Corporation 2004

SLI Rendering Modes Compatibility mode Only uses one GPU No SLI benefits Alternate frame rendering (AFR) Split frame rendering (SFR) SLI AA SLI Stereo? Copyright NVIDIA Corporation 2004

AFR GPUs work on alternate frames in parallel GPU 0: GPU 1: 1 3 2 4 Scan-out toggles which framebuffer to read from Copyright NVIDIA Corporation 2004

AFR Advantages Advantages All work is parallelized Scales geometry and pixel fill performance Preferred SLI mode Disadvantages Requires pushing data to other GPU if frame is not selfcontained For example, if application updates a render-to-texture target only every other frame Copyright NVIDIA Corporation 2004

SFR GPUs work on the same frame For two GPUs GPU 0 renders top region GPU 1 renders bottom region GPU 0 GPU 1 1 2 Scan-out combines framebuffer data Copyright NVIDIA Corporation 2004 3

SFR Advantages Driver load-balances by changing region size Based on time each GPU took to render Driver clips geometry to regions Avoids both GPUs processing all vertices But not perfect Still requires sharing data between GPUs E.g., render to texture Copyright NVIDIA Corporation 2004

SFR Compared to AFR SFR works even when few frames are buffered Or when AFR otherwise fails In general, SFR has more communications overhead Applications with heavy vertex load benefit less from SFR Copyright NVIDIA Corporation 2004

Overview: Things Interfering with SLI CPU-bound applications Or vertical-sync enabled Applications that limit the number of frames buffered Communications overhead Copyright NVIDIA Corporation 2004

CPU-Bound Applications SLI cannot help Reduce CPU work Move CPU work onto the GPU See http://www.gpgpu.org Don’t deliberately throttle frame-rate Copyright NVIDIA Corporation 2004

V-Sync Enabling vertical-sync limits frame rate to multiples of the monitor refresh rate Copyright NVIDIA Corporation 2004

Limiting Number of Frames Buffered Some apps allow at most one frame buffered To reduce lag Via occlusion queries Don’t read back-buffer - this causes CPU stall Breaks AFR SLI SLI is faster anyway e.g. 2 GPU SLI systems 2.0x less lag Copyright NVIDIA Corporation 2004

Why Reading the Back Buffer Is Bad Back buffer read: wait for GPU to finish rendering CPU GPU Frame n Copyright NVIDIA Corporation 2004 CPU GPU Frame n 1

OpenGL SLI Tips Limit OpenGL rendering to a single window child windows shouldn’t have OpenGL contexts Request pixel format with PDF SWAP EXCHANGE tells driver that app doesn’t need the back buffer contents after SwapBuffers() Avoid rendering to FRONT buffer use overlays instead on Quadro GPUs Copyright NVIDIA Corporation 2004

Offscreen Rendering and Textures Limit P-buffer usage Often requires broadcasting rendering to both GPUs Use render-to-texture rather than glCopyTexSubImage glCopyTexSubImage requires texture to be copied to both GPUs Use FBO or P-buffers instead Limit texture working set Textures have to be stored on both GPUs Don’t download new textures unnecessarily Copyright NVIDIA Corporation 2004

Geometry Use Vertex Buffer Objects or display lists to render geometry Don’t use immediate mode Reduces CPU overhead Render the entire frame Don’t use use glViewport or glScissor Disables load balancing in SFR mode, and hurts performance in AFR mode Copyright NVIDIA Corporation 2004

More OpenGL SLI Tips Limit read-backs e.g. glReadPixel, glCopyPixels causes pipeline to stall Never call glFinish() doesn’t return until all rendering is finished prevents parallelism Avoid glGetError() in release code Causes sync point Copyright NVIDIA Corporation 2004

How Do I Detect SLI Systems? NVCpl API: NVIDIA-specific API supported by all NV drivers Function support for: Detecting that NVCpl API is available Bus mode (PCI/AGP/PCI-E) and rate (1x-8x) Video RAM size SLI Copyright NVIDIA Corporation 2004

NVCpl API SLI Detection SDK sample and full documentation available HINSTANCE hLib ::LoadLibrary("NVCPL.dll"); NvCplGetDataIntType NvCplGetDataInt; NvCplGetDataInt (NvCplGetDataIntType)::GetProcAddress(hLib, "NvCplGetDataInt"); long numSLIGPUs 0L; NvCplGetDataInt(NVCPL API NUMBER OF SLI GPUS, &numSLIGPUs); Copyright NVIDIA Corporation 2004

Forcing SLI Support In Your Game Use NVCpl NvCplSetDataInt() sets AFR, SFR, Compatibility mode See SDK sample Modify or create a profile: http://nzone.com/object/nzone sli appprofile.html End-users can create profile as well Copyright NVIDIA Corporation 2004

SLI Performance Tools NVPerfKit has support for SLI Provides performance counters for Total SLI peer-to-peer bytes Total SLI peer-to-peer transactions Above originating from Vertex/index buffers: bytes and transactions Textures: bytes and transactions Render targets: bytes and transactions Copyright NVIDIA Corporation 2004

What is Instancing? Rendering multiple instances of a given geometry Some attributes can vary across instances Transformation matrix Color Examples Trees in a forest Characters in a crowd Boulders in a avalanche Screws in an assembly Copyright NVIDIA Corporation 2004

Instancing Methods in OpenGL Send transform as vertex program constants Relatively slow Can also pack several transforms into constant memory and index in vertex program Send transform using immediate mode texture coordinates (“pseudo instancing”) Usually much faster (glTexCoord calls are inlined) Requires custom vertex program Can use glArrayElement to set current texture coordinates from a vertex array (not efficient on NV hardware) NVX instanced arrays Single draw call Fastest Copyright NVIDIA Corporation 2004

NVX instanced arrays Allows rendering multiple instances of an object with a single draw call Similar to Direct3D instancing functionality OpenGL draw call cost is lower than Direct3D, but still gives a significant performance benefit Combined with render-to-vertex array, can be used for controlling object transformations on the GPU Performance is dependent on CPU speed, GPU speed, number of objects and number of vertices per object Will improve on next generation GPU hardware Copyright NVIDIA Corporation 2004

OpenGL Instancing Performance Verts/object Constants (fps) Texcoords (fps) Instancing (fps) 8 205 323 560 24 200 266 440 60 183 190 246 120 138 135 155 220 72 77 77 8192 objects, Quadro FX 4500, P4 3.4 GHz Copyright NVIDIA Corporation 2004

NVX instanced arrays Allows rendering an array of primitives multiple times, while stepping specified vertex attribute arrays only once per N objects Only supports generic attribute arrays No immediate mode Warning – experimental extension -API may change! Typically 3 attribute arrays are used to store a 3x4 transformation matrix Attribute divisor is set to 1 for these arrays Custom vertex program transforms geometry from object to world space based on input attributes Copyright NVIDIA Corporation 2004

NVX instanced arrays API void VertexAttribDivisorNVX(uint attrib, uint divisor); Specifies rate at which to advance attribute per object 0 disabled Attribute 0 (position) cannot be changed Future – fractional divisor to allow geometry amplification? void DrawArraysInstancedNVX(enum mode, int start, sizei count, sizei primCount); void DrawElementsInstancedNVX(enum mode, sizei count, enum type, const void *indices, sizei primCount); Renders primCount instances of specified geometric primitives, using attribute divisors Copyright NVIDIA Corporation 2004

DrawArraysInstancedNVX Pseudocode for (instance 0; instance primCount; instance ) { Begin(mode); for (vertex 0; vertex count; vertex ) { for (attrib 1; attrib MAX ATTRIB; attrib ) { if (ArrayAttribEnabled[attrib]) { if (InstanceDivisors[attrib] 0) { offset instance / InstanceDivisors[attrib]; } else { offset start vertex; } offset * CookedAttribStride[attrib]; VertexAttribvFunc[attrib]( VertexAttribPointers[attrib] offset); } } if (ArrayAttribEnabled[0]) { offset start vertex; offset * CookedAttribStride[0]; VertexAttribvFunc[0]( VertexAttribPointers[0] offset); } } End(); } Copyright NVIDIA Corporation 2004

Standard Rendering Loop // load vertex arrays and transform data for(int i 0; i nobjects; i ) { // send transformation as texture coordinates glMultiTexCoord4fv(GL TEXTURE0, &transform data[0][i*4]); glMultiTexCoord4fv(GL TEXTURE1, &transform data[1][i*4]); glMultiTexCoord4fv(GL TEXTURE2, &transform data[2][i*4]); // draw instance glDrawElements(GL TRIANGLES, nindices, GL UNSIGNED SHORT, indices); } Copyright NVIDIA Corporation 2004

Using Instancing // set vertex array pointers // enable transform attribute arrays and set divisors glEnableVertexAttribArrayARB(8); // texcoord0 glVertexAttribDivisorNVX(8, 1); glEnableVertexAttribArrayARB(9); // texcoord1 glVertexAttribDivisorNVX(9, 1); glEnableVertexAttribArrayARB(10); // texcoord2 glVertexAttribDivisorNVX(10, 1); // draw all instances at once glDrawElementsInstancedNVX(GL TRIANGLES, nindices, GL UNSIGNED SHORT, indices, nobjects); glDisableVertexAttribArrayARB(8); glVertexAttribDivisorNVX(8, 0); glDisableVertexAttribArrayARB(9); glVertexAttribDivisorNVX(9, 0); glDisableVertexAttribArrayARB(10); glVertexAttribDivisorNVX(10, 0); Copyright NVIDIA Corporation 2004

HavokFX Instancing Results Readback (fps) Instancing (fps) Instancing / Readback 4096 bricks 240 280 1.17 8000 bricks 130 150 1.15 27000 bricks 40 46 1.15 5000 boulders 173 223 1.29 10000 boulders 90 114 1.27 30000 boulders 31 41 1.32 Copyright NVIDIA Corporation 2004

GPU Timing Timing is important for performance tuning How can you improve something if you can’t measure it accurately? Problem with timing the GPU is that it is asynchronous and has a deep pipeline There’s no way to know if a particular command has completed before reading the timer Usual solution is to insert glFinish() commands Guarantees that all rendering commands have completed, but stalls pipeline and changes performance! Copyright NVIDIA Corporation 2004

EXT timer query Provides a method for timing a sequence of OpenGL commands, without stalling the pipeline Based on the query object mechanism introduced by the occlusion query extension glBeginQuery() Timer starts when all prior commands have completed glEndQuery() Timer stops when all prior commands have completed Measures total time elapsed (driver hardware) Measured in nanoseconds (10-9 seconds) 32 bit counter can represent about 4 seconds maximum Introduces GLuint64 type to allow 64 bit counters Copyright NVIDIA Corporation 2004

Code Example GLint queries[N]; glGenQueries(N, queries); // generate query objects for(int i 0; i N; i ) { glBeginQuery(GL TIME ELAPSED EXT, queries[i]); // Start query // Draw object i glEndQuery(GL TIME ELAPSED EXT); // End query } // Wait for all results to become available // (should really only wait for previous frame’s results) int available 0; while (!available) { glGetQueryObjectiv(GL QUERY RESULT AVAILABLE, queries[N-1], &available); } // See how much time the rendering of object i took in nanoseconds GLuint64EXT timeElapsed; for (i 0; i N; i ) { glGetQueryObjectui64vEXT(queries[i], GL QUERY RESULT, &timeElapsed); // do something with result } Copyright NVIDIA Corporation 2004

Questions? GPU Programming Guide: http://developer.nvidia.com/object/gpu programmin g guide.html http://developer.nvidia.com Thanks: Matthias Wloka, Jason Allen, Michael Gold Copyright NVIDIA Corporation 2004