Anselmo Lastra, Steve Molnar, Marc Olano, Yulan Wang. Marc Olano, Yulan Wang University of North Carolina at Chapel Hill

Transcription

1 Real-Time Programmable Shading Real-Time Programmable Shading ( ( Anselmo Lastra, Steve Molnar, Anselmo Lastra, Steve Molnar, Marc Olano, Yulan Wang Marc Olano, Yulan Wang University of North Carolina at Chapel Hill University of North Carolina at Chapel Hill Introduction Organization of programmable shading shading architecture Programmable what? Surface shader Simple functions Run at each pixel/sample Compute surface shading, lighting, displacement maps, atmospheric effects,... Whitted 82, Cook 84, Perlin 85, Hanrahan 90 Rhoades 92 For one sample Inputs: Intrinsic color, normal, texture coordinates, width, length, bumpiness, swirliness,... Outputs: Color, (opacity) 1

2 Resource requirements Programmability Programmability Programmable processors at sample level High level language (i.e. RenderMan) al power Hanrahan 90, Upstill 90 Resource requirements Memory Programmability al power Table memory Local memory Shader Local Variables carpet 24 marble 26 stippled 33 stone 23 Resource requirements Parallelism Programmability al power Parallelism Deferred shading Uniform/varying Fixed point/floating point Pixel-Planes 5 (Fuchs 89) 2 50 Graphics processors (i860) 1 20 Renderers (custom) 16k Pixel processors / renderer (custom) Reality Engine (Akeley 93) 8 12 Geometry engines (i860xp) 5 20 Fragment generators (custom) Image engines (custom) 2

3 Resource requirements Deferred shading Programmability al power Parallelism Deferred shading Uniform/varying Fixed point/floating point Keep shading parameters at each pixel Shade after visibility is determined Pros: Doesn t shade hidden pixels Shading independant of geometric complexity! Better utilization on SIMD Cons: Can t affect visibility (No transparency, no displacement maps!) Resource requirements Uniform/varying Programmability al power Parallelism Deferred shading Uniform/varying Fixed point/floating point Uniform = constant across pixels/samples Wood grain, marble vein frequency,... Varying = different in each pixel/sample Normal, texture coordinates,... Don t compute uniform values at every pixel compute once and broadcast Resource requirements Fixed point/floating point Programmability al power Parallelism Deferred shading Uniform/varying Fixed point/floating point Pixel processors Many processors Simple instruction set Floating point acceleration is unlikely Fixed point When required precision is known More efficient in time and memory 3

4 node Node description Timings System Geometry network HP PA-RISC HP PA-RISC 128 x 64 SIMD array Table Memory Optional video card Composition network Memory Node description Timings System 16MB table memory 256 bytes local memory 128 bytes local memory/communication Timings (µs) Timings (µs) float fixed 4 byte 4 byte 2 byte µs * / sqrt sqrt / * + 2 byte fixed 4 byte fixed 4 byte float 4

5 system Node description Timings System Show video Shading functions Shading functions Pins Crown, label, scuffs, dirt, Phong Alley Wood, reflection map Ball Phong Light Shadow map 5

6 Time: 7 ms - shadow map Use of multiple processors 33 ms Time: 7 ms - reflection map Use of multiple processors 33 ms Time: 15.7 ms - final image Use of multiple processors 33 ms 6

7 Shading: 2µs - crown Shading: 15µs - label Shading: 39µs - scuffs & dirt 7

8 Shading: 15µs - wood Shading: 28 µs - light/shadows Shading: 12µs - Phong (pins) 8

9 Shading: 27µs - reflection Shading: 12µs - Phong (ball) Time for table lookups About 23ns per pixel Worst case Bowling pin (4 lookups) in all pixels» Label image» Scuff bump map» Dirt image» Shadow map Total 760µs per region 9

10 10

11 11

12 12

13 Future work Acknowledgements RenderMan-like shading compiler Allow programming other places Convince commercial vendors that they can and should do programmable shading too Lawrence Kesteloot, Fredrik Fatemi UNC Group Tony Apodaca, Pixar Hewlett-Packard ARPA ISTO Order No. A410 NSF Grant No. MIP