1 Ray Tracing with Existing Graphics Systems Jeremy Sugerman, FLASHG 31 January 2006
1
Ray Tracing with Existing Graphics Systems
Jeremy Sugerman,
FLASHG 31 January 2006
2
Why Consider Tracing Rays?
Some techniques are hard to mimic with rasterization.
Shadows (arbitrary view visibility) Reflection, Refraction effects Global, Indirect Illumination
3
Why Now?
Raytracing is getting fast enough to use interactively.
New architectures seem potentially suited or not too far from being suited.
4
“Interactive”, Really?
Intel MLTRA (SIGGRAPH 2005)20-36 fps (20-36 MRays/sec) on one 3.2 GHzWithout MLTRA, still 7-12 fps / MRays/sec
My toy SSE raytracerAbout 2 MRays/sec (primary rays)
‘Quick Hack’ Cell system at SIGGRAPHClaims 20-25 fps / MRays/sec (unreviewed)
5
What about Rasterization?
60+ fps over 1024x768 on $99 GPUs Ray tracing would need 50+ MRays/sec
casting primary rays alone. GPUs are getting faster at a phenomenal
rate.
Ray tracing is DOOMED!
6
It’s Not About Rasterization
The vast majority of render time is actually shading in most 3D apps.
True for offline rendering too. Whether fragments are rasterized or ray
traced, shading is the same…
To the extent the fragments are the same
7
And Who Traces Primary Rays?
We want ray tracing for shadows, reflection/refraction, and indirect illumination.
Those are all applications of secondary rays. The rasterizer already produces (x, y, z),
normals, etc. from primary hits. Might as well rasterize them if you have a fast
existing mechanism. It doesn’t matter.
8
Synthesis (Thesis + Antithesis)
Bolt a ray tracer onto a conventional rasterization based system.
Add a bit of ray tracer friendliness to a GPU. Or, wire together a Cell and a GPU (don’t tell
Sony). Window systems (2D), text are rasterization
tasks fundamentally (plus optional compositing)
9
“Ray Tracer Friendliness”?
Ray tracing strongly favours threaded architectures over SIMD.
Packet tracing leverages bandwidth at the cost of very simple horizontal communication.
Being able to use queues / write buffers seems critical.
10
How Would It Work?
Operate on ‘hits’ Analogous to fragments, plus weights
Each hit feeds any of three (independent) shading choices Gather rays Shadowing and local lighting contribution Secondary rays
Each produces a colour and a weight which are accumulated into a final pixel colour
11
Shadowing And Local Lighting
Effectively runs: Interpolate shading data (BRDF, Normal, etc.) Generate as many shadow rays as are desired Foreach shadow ray {
If (shadow ray hits light) {Compute local light contribution from the light
}
Fits in the same per-fragment storage Shadow computations are indepedent
12
Gather and Secondary Rays
Gather rays just perform a weighted lookup in a data structure (e.g. photon map)
Secondary rays are generated based on the surface shading information (BRDF, Normal, …)Hits are fed back into the same pipelineOnce generated, independent of parent
13
Kinda Like a GPU (If you squint)
GPU shading just does the weighted local lighting calculation.
But a fragment program can generate and trace shadow rays to mask local lighting
And a fragment program can generate rays for final gathers.
All this formulation does is expose parallelism and offer natural places to optimize hardware
14
I Saw You Palm That Card!
Secondary rays are a bit harder to cram into a fragment program. No variable output buffers No recursion
Once generated, secondary rays are completely independent. Only need an unordered write buffer No state, so recursion becomes iteration
So start with only shadow and gather rays Exact same system supports them, though
15
What About Coherence?
Rasterization lie: Rasterization systems exploit coherence vastly better than ray tracing.
Really means shading coherence Coherence between fragments in texture lookups Can bind a single material shader and rasterize only
relevant geometry Can perform (expensive!) shading math at a different
resolution than visibility math.
16
Ray Tracing Is Coherent Too
Packet tracing is useful because of visibility coherence.
The same coherence is also somewhat relevant for local shading and lighting.
Even, render only objects with the same material in each pass and z-cull.Just as coherent as rasterization
17
Ray Tracing Is Coherent Too
Packet tracing literature demonstrates ray coherence helps visibility most, but also shadow / secondary rays and shading.
Rasterization systems save overdraw by rendering a depth-only pass before material shading with z-culling.
The same technique works with rays!
18
Global Effects Are Incoherent
Nearly by definition, indirect and inter-object shading effects are incoherent. In rasterization or ray-traced systems!
Packet tracing literature indicates secondary rays retain some coherence.
Geometry level of detail can regain coherence.
19
Coherence: Bottom Line
Without secondary rays, ray tracing shares the coherence characteristics of rasterization (modulo implementation).
With secondary rays, ray tracing offers natural mechanisms for effects that are only awkwardly kludged via rasterization.
Let the developer / director pick the tradeoff.
20
What Needs To Be Done?
Persuade Pat to get a PS3 dev kit? Get a good ray tracer running on Cell and/or a
GPU. Simulate various extensions
Cram code in fragment programs Readback to the CPU for now Readback to a Cell Talk to the Smart Memories folks? Talk to GPU vendors (unlikely, they’re busy folks)
21
What’s Behind the Curtain?
K-d tree building and updates But, Rasterization systems do this today
K-d tree storage and bandwidth needs Hardware changes for traversing k-d trees or
intersecting triangles Unclear if it’s even necessary initially
Level of detail with secondary rays Decoupling visibility and shading resolutions
22
Contributions and Inspirations
Tim, Pat Gordon Stoll Bill Mark Phil Slusallek and the Saarland crew Tim Purcell, John Nickolls (NVIDIA)
23
Rasterization Lies
Rasterization has coherence Ray tracing lacks. The rasterization argument goes: rasterize all
the geometry of a given material and there’s excellent coherence of shading samples.
Without secondary rays, ray tracing is the same. With secondary rays, you get better pictures. The programmer gets to choose.