Real-time REYES

1
Real-time REYES

• Jeremy Sugerman

2
This Talk

• Still exploring future GPUs, Direct3D / GL
• This time in a software context
• Offline REYES, as RenderMan, is king of the
  rendering world
• What would real-time REYES look like?
• Where are current GPUs / Direct3D and REYES the
  same? Similar? Different?

3
Background

• The Reyes Image Rendering Architecture
• Pixar Rob Cook, Loren Carpenter, Ed Catmull
• SIGGRAPH 1987, First film 1985, Oscar 1993
• An architecture optimized for fast high-quality
  rendering of complex animated scenes.
• Fast feature-length film in a year
• High-Quality virtually indistinguishable from
  live action
• Complex as visually rich as real scenes.

4
REYES Features

• Primitives are smooth surfaces
• Automatic, adaptive refinement
• Programmable shading including displacement
• On surface-space grids
• Very high resolution geometry for visibility
• After shading
• Stochastic sampling in time, space, lens
• Motion-blur, anti-aliasing, depth of field
• Order-independent (depth sorted) blending
• Wide pixel reconstruction filtering

5
Basic REYES Pipeline
Parallel scalability is added with screen-space
bucketing computed during the Bounding and
Splitting. Deferring Dicing also allows each
bucket to be depth-sorted and occlusion culled.
6
REYES on a GPU?

• If you squint at it right
• Grids of micropolygons ? quads
• Displace and shade ? shade vertices()
• Bust bound, sample ? rasterize()
• Blend filter ? blend filter()
• Technically, OpenGL even includes implicit
  surface (NURBS) evaluators, but there is no
  serious support in (current) GPUs.

7
Opaque GPU-REYES

• Tesselate to grids in advance (CPU, cache, )
• Render grids as GL_QUADS
• Displace, shade grids with a vertex shader
• Cull, sample, and produce visible points with the
  rasterizer (MSAA / FSAA)
• Reconstruction during ROP / FB
• Flat fragment shading with micropolygons!
• Current GPU lacking in tesselation, shading
  richness, stochastic sampling, blending

8
GPU Mismatches

• High level surfaces versus triangles
• Surface vertex shading versus fragment shading
• Including a constrained shading language
• Stochastic sampling versus MSAA rasterization
• Depth order versus primitive order blending
• Note REYES only tries to make pictures.
• Bonus Opaque REYES is totally unordered.

9
Straightforward GPU Evolution

• High level surfaces are already coming
• Without ordering, performance is easy
• Shading surface points versus pixels is fakeable
• Less flexible GLSL/HLSL is fine
• Unified shader cores help a lot
• Adjust rast for small quad workloads
• Stochastic sampling versus MSAA rasterization
• Allow rast to time-sample linear motion

10
Sticking Point Blending

• Primitive order blending generates poor images
• Depth order risks unbounded storage needs
• Samples must be buffered until all closer samples
  are blended.
• Closest sample may not happen until the last grid
  of the last primitive.
• Unbounded storage is anathema in GPU environments
  and real-time systems in general.

11
Options for Depth-Order Blending

• Classic REYES buckets grids into screen-space
  partitions.
• Reduces, but still does not bound, footprint
• Assume GPU REYES has fixed on-chip buffers
• Three options when a bucket fills
• Rendering fails (awful)
• Spill to memory (slow? when memory fills?)
• Shrink bucket, discard samples newly outside
  (recomputation hard? too compute wasteful?)

12
GPU REYES Summary

• Unordered
• Retained mode input (at least for blending)
• Simplified shading language
• Separate opaque and blended modes
• Constrained reconstruction filters
• Bound and split on CPU? Simplified adaptivity?
• Mandatory displacement shader separation?
• Constraints on stochastic sampling?
• Blending policy Primitive order? Spill? Shrink?

13
Acknowledgements

• Kayvon has provided explanation and discussion
  beyond a point where I can separate any ideas
  unique to myself.
• Kurt and Pat provided excellent historical
  context and insight.
• Gordon Stoll and Bill Mark repeated the word
  REYES and insisted that its shading /
  anti-aliasing properties were valuable until I
  dug into it in self-defense.