� Notes

Outline

On the Future of the GPU in Computer Graphics: Where are we going? Prof. John C. Hart Dept. of Computer Science University of Illinois, Urbana-Champaign

Faster than Moore�s Law

Local Illumination GPU designed for efficient local illumination computations to make video games more interesting Bump mapping, BRDF Vertex shader processes per-vertex attributes (normal, texcoords, color) Displacement mapping, skinning Rasterization interpolates vertex attributes across pixels Projective, perspective correct Pixel shader computes colors from interpolated values and texture lookups texture shading, perspective texturing

Modern GPU Org.

Global Illumination Simulates all paths of light from source to eye Techniques Ray tracing Radiosity Photon mapping Effects Shadows Reflection Refraction Color bleeding Caustics Subsurface scattering Expands role of GPU from game engine to DCC platform

Subsurface Scattering Simulates scattering of light within a translucent material Needed for all non-metallic surfaces SKIN, milk, bread, stone Real-time GPU Implementations one based on recent pixel shader enhancements (floating point) one based on recent vertex shader enhancements (# of constants)

Pixel Shader Subsurface Scattering

Vertex Shader Subsurface Scattering

GPU Global Problems

Early GPU Ray Tracing Implemented on ATI Radeon 8500 DX 8.1 Pixel Shaders 1.4 16-bit fixed point Limited precision resulted in errors 114M ray-triangle ints/s Much faster than best single CPU time (20-40M on 800MHz P3, Wald et al. EGRW01) Expect gap to widen further

Modern GPU Ray Tracing Nate Carr, Jesse Hall, John Hart. The Ray Engine. Proc. Graphics Hardware 2002. Uses GPU to intersect rays with triangles Turns �Geometry Engine� into a �Ray Engine� Uses CPU to organize which rays and triangles to send to GPU

Results Ray Engine GH02 200K rays/s for highly coherent, small (2.5K) scenes 115K rays/s for large (34K) complicated scenes Wald et al. EGRW01 P3 SSE: 200K � 1.5M rays/s CPU-SSE tightly coupled Purcell et al. S02 300K rays/s large (35K) up to 4M rays/s small (35)

Readback Drivers currently use PCI readback hardware and drivers, even for AGP cards! Only get 160MB/s (should get 1GB/s) An older IBM card got 800 MB/s readback PCI Express promises fast bidirectional GPU access But only if hardware manufacturers use both directions Would help if benchmarks include readback rates