![Proven Tech – Out in the Wild
●Tiled Deferred [Andersson09]
●Frostbite
●UE4
●Ryse
●Forward+ [Harada et al 12]
●DiRT & GRID Series
●The Order: 1886
●Ryse](https://image.slidesharecdn.com/advancements-in-tiled-rendering-150323053848-conversion-gate01/85/Advancements-in-tiled-rendering-3-320.jpg)
![Tiled Rendering 101
1
2
3
[1] [1,2,3] [2,3]](https://image.slidesharecdn.com/advancements-in-tiled-rendering-150323053848-conversion-gate01/85/Advancements-in-tiled-rendering-4-320.jpg)
![Depth Bounds
● Determine min and max
bounds of the depth buffer
on a per tile basis
● Atomic Min Max [Andersson09]](https://image.slidesharecdn.com/advancements-in-tiled-rendering-150323053848-conversion-gate01/85/Advancements-in-tiled-rendering-10-320.jpg)
![Parallel Reduction
●Atomics are useful but not efficient
●Compute-friendly algorithm
●Great material already available:
●“Optimizing Parallel Reduction in CUDA” [Harris07]
●“Compute Shader Optimizations for AMD GPUs: Parallel Reduction” [Engel14]](https://image.slidesharecdn.com/advancements-in-tiled-rendering-150323053848-conversion-gate01/85/Advancements-in-tiled-rendering-12-320.jpg)
![59 86 95 53 97 18 28 46
57 16 25 43
depth[tid] = min(depth[tid],depth[tid+8])
25 13
depth[tid] = min(depth[tid],depth[tid+4])
13
depth[tid] = min(depth[tid],depth[tid+2])
1
depth[tid] = min(depth[tid],depth[tid+1])](https://image.slidesharecdn.com/advancements-in-tiled-rendering-150323053848-conversion-gate01/85/Advancements-in-tiled-rendering-13-320.jpg)
![Arvo Intersection Test [Arvo90]](https://image.slidesharecdn.com/advancements-in-tiled-rendering-150323053848-conversion-gate01/85/Advancements-in-tiled-rendering-22-320.jpg)
![2.5D Culling [Harada et al 12]
Scene Geometry
Geometry Mask
1 1 1 1
1 1 1
Light Mask](https://image.slidesharecdn.com/advancements-in-tiled-rendering-150323053848-conversion-gate01/85/Advancements-in-tiled-rendering-29-320.jpg)
![Clustered Rendering [Olsson et al12]
●Production proven in Forza Horizon 2
●Additional benefits on top of 2D
culling:
●No mandatory Z prepass
●Just works™ for transparencies and
volumetric effects
●Can a further reduction in lights per
pixel improve performance?](https://image.slidesharecdn.com/advancements-in-tiled-rendering-150323053848-conversion-gate01/85/Advancements-in-tiled-rendering-46-320.jpg)
![Compacted List
●Option 1:
●Do all culling on CPU [Olsson et al12] [Persson13][Dufresne14]
●But some of the lights may be spawned by the GPU
●My CPU is a precious resource!
● Option 2:
●Cull on GPU
●Keep track of how many lights per slice in TGSM
●Write table of offsets in light list header
●Only need maxLights x “safety factor” per tile](https://image.slidesharecdn.com/advancements-in-tiled-rendering-150323053848-conversion-gate01/85/Advancements-in-tiled-rendering-52-320.jpg)
![References
●[Andersson09] Johan Andersson, “Parallel Graphics in Frostbite – Current & Future”, Beyond
Programmable Shading, SIGGRAPH 2009
●[Harada et al12] Takahiro Harada, Jay McKee, Jason C Yang, “Forward+: Bringing Deferred
Lighting to the Next Level”, Eurographics 2012
●[Harris07] Mark Harris, “Optimizing Parallel Reduction in CUDA”, NVIDIA 2007
●[Engel14] Wolfgang Engel, “Compute Shader Optimizations for AMD GPUs: Parallel Reduction”,
Confetti 2014
●[Harada12] Takahiro Harada, “A 2.5D Culling for Forward+”, Technical Briefs, SIGGRAPH Asia
2012
●[Arvo90] Jim Arvo, “A simple method for box-sphere intersection testing”, Graphics Gems 1990
●[Dufresne14] Marc Fauconneau Dufresne, “Forward Clustered Shading”, Intel 2014
●[Persson13] Emil Persson, “Practical Clustered Shading”, Avalanche 2013
●[Olsson et al12] Ola Olsson, Markus Billeter, Ulf Assarsson, “Clustered Deferred and Forward
Shading”, HPG 2012
●[Schulz14] Nicolas Schulz, “Moving to the Next Generation – The Rendering Technology of
Ryse”, GDC 2014](https://image.slidesharecdn.com/advancements-in-tiled-rendering-150323053848-conversion-gate01/85/Advancements-in-tiled-rendering-59-320.jpg)
Advancements in-tiled-rendering
- 1. Advancements in Tiled-Based Compute Rendering Gareth Thomas Developer Technology Engineer, AMD
- 2. Agenda ●Current Tech ●Culling Improvements ●Clustered Rendering ●Summary
- 3. Proven Tech – Out in the Wild ●Tiled Deferred [Andersson09] ●Frostbite ●UE4 ●Ryse ●Forward+ [Harada et al 12] ●DiRT & GRID Series ●The Order: 1886 ●Ryse
- 4. Tiled Rendering 101 1 2 3 [1] [1,2,3] [2,3]
- 5. Tiled Rendering 101 ● Divide screen into tiles ● Fit asymmetric frustum around each tile Tile0 Tile1 Tile3Tile2
- 6. Tiled Rendering 101 ● Use z buffer from depth pre-pass as input ● Find min and max depth per tile ● Use this frustum for intersection testing
- 8. Index1 •1 Tiled Rendering 101 •Position •Radius Light0 •Position •Radius Light1 •Position •Radius Light2 •Position •Radius Light3 •Position •Radius Light4 … •Position •Radius Light10 •4Index2 •Lights=2Index0 Index3 •Empty Index4 •Empty … 1 4
- 9. Targets for Improvement ●Z Prepass (on Forward+) ●Depth bounds ●Light Culling ●Color Pass
- 10. Depth Bounds ● Determine min and max bounds of the depth buffer on a per tile basis ● Atomic Min Max [Andersson09]
- 11. // read one depth sample per thread // reinterpret as uint // atomic min & max // reinterpret back to float
- 12. Parallel Reduction ●Atomics are useful but not efficient ●Compute-friendly algorithm ●Great material already available: ●“Optimizing Parallel Reduction in CUDA” [Harris07] ●“Compute Shader Optimizations for AMD GPUs: Parallel Reduction” [Engel14]
- 13. 59 86 95 53 97 18 28 46 57 16 25 43 depth[tid] = min(depth[tid],depth[tid+8]) 25 13 depth[tid] = min(depth[tid],depth[tid+4]) 13 depth[tid] = min(depth[tid],depth[tid+2]) 1 depth[tid] = min(depth[tid],depth[tid+1])
- 14. Implementation details ●First pass reads 4 depth samples ●Needs to be separate pass ●Write bounds to UAV ●Maybe useful for other things too
- 16. Parallel Reduction - Performance Atomic Min/Max Parallel Reduction AMD R9 290X 1.8ms 1.60ms NVIDIA GTX 980 1.8ms 1.54ms ● Combined cost of depth bounds and light culling of 2048 lights at 3840x2160 ● Parallel reduction pass takes ~0.35ms ● Faster than Atomic Min/Max on the GPUs tested
- 17. Light Culling: The Intersection Test
- 20. AABB around Frustum Frustum planes
- 21. AABB around long frustum AABB around short frustum
- 22. Arvo Intersection Test [Arvo90]
- 26. Culling Spot Lights ●Don’t put bounding sphere around spot light origin ●Tightly bound spot light inside sphere at P with radius r spot position P r θ r d
- 28. Depth Discontinuities False Positives Scene Geometry
- 29. 2.5D Culling [Harada et al 12] Scene Geometry Geometry Mask 1 1 1 1 1 1 1 Light Mask
- 30. HalfZ Scene Geometry HalfZ MinZ MaxZ HalfZ low bits HalfZ high bits numLights near side numLights far side light indices… 3 4 lo hi 16 bit light index buffer size: maxLightsPerTile x 2 + 4
- 31. Modified HalfZ HalfZ MinZ MaxZ MinZ2 MaxZ2 ●Calculate Min & Max Z as normal ●Calculate HalfZ ●Second set of Min and Max values using HalfZ and max & min respectively ●Test against near bounds and far bounds ●Write to either one list ●Or write to two lists cf. HalfZ ●Doubles the work in the depth bounds pass ●Worst case converges on HalfZ
- 32. Sponza Atrium + 1 million sub pixel triangles
- 34. MinMax depth bounds, Frustum culling
- 35. MinMax depth bounds, AABB culling
- 36. MinMax depth bounds, Hybrid culling (AABB + Frustum sides)
- 37. Modified HalfZ depth bounds, AABB culling
- 38. Unreal Engine 4, Infiltrator Demo
- 39. Modified HalfZ in one light list MinMax Depth Bounds
- 43. What happens if we cull 32x32 tiles? Still using 16x16 thread groups
- 45. Culling Conclusion ●Modified HalfZ with AABBs generally works best ●Even though generating MinZ2 and MaxZ2 adds a little cost ●Even though culling each light against two AABBs instead of one ●32x32 tiles saves a good chunk of time in the culling stage ●…at the cost of color pass efficiency when pushing larger number of lights
- 46. Clustered Rendering [Olsson et al12] ●Production proven in Forza Horizon 2 ●Additional benefits on top of 2D culling: ●No mandatory Z prepass ●Just works™ for transparencies and volumetric effects ●Can a further reduction in lights per pixel improve performance?
- 47. Clustered Rendering 101 ● Divide screen into tiles ● Fit asymmetric frustum around each tile Tile0 Tile1 Tile3Tile2 ● Divide down Z axis into n slices or clusters
- 48. Clustered Rendering ●Divide up Z axis exponentially ●Start at some sensible near slice ●Cap at some sensible value
- 49. Provision for far lights ● Fade them out ● Drop back to glares ● Prebake
- 50. Light Culling ●View space AABBs worked best on 2D grid ●Bad when running say 16 slices ●View space frustum planes are better ●Calculate per tile planes ●Then test each slice near and far ●Optionally, then test AABBs
- 51. VRAM Usage ●16x16 pixel 2D grid requires numTilesX x numTilesY x maxLights ●1080p: 120x68x512xuint16 = 8MB ●4k: 240x135x512xuint16 = 32MB ●List for each light type (points & spots): 64MB ●So 32 slices: 1GB for point lights only ●Either use coarser grid ●Or use a compacted list
- 52. Compacted List ●Option 1: ●Do all culling on CPU [Olsson et al12] [Persson13][Dufresne14] ●But some of the lights may be spawned by the GPU ●My CPU is a precious resource! ● Option 2: ●Cull on GPU ●Keep track of how many lights per slice in TGSM ●Write table of offsets in light list header ●Only need maxLights x “safety factor” per tile
- 53. Coarse Grid ●Example: ●4k resolution ●64x64 pixel tiles with 64 slices ●maxLights = 512 ●60 x 34 tiles x 64 slices x 512 x uint16 = 128MB
- 57. Z Prepass ●Very scene dependant ●Often considered too expensive ●DirectX12 can help draw submission cost ●Should already have a super optimized depth only path for shadows! ● Position only streams ● Index buffer to batch materials together ●A partial prepass can really help lighten the geometry load
- 58. Conclusions ●Parallel Reduction - faster than atomic min/max ●AABB-Sphere test in conjunction with Modified HalfZ is a good choice ●Clustered shading ●Potentially a big saving on the tile culling ●Less overhead for low light numbers ●Offers other benefits over 2D tiling ●Aggressive culling is very worthwhile ●The best optimisation for your expensive color scene
- 59. References ●[Andersson09] Johan Andersson, “Parallel Graphics in Frostbite – Current & Future”, Beyond Programmable Shading, SIGGRAPH 2009 ●[Harada et al12] Takahiro Harada, Jay McKee, Jason C Yang, “Forward+: Bringing Deferred Lighting to the Next Level”, Eurographics 2012 ●[Harris07] Mark Harris, “Optimizing Parallel Reduction in CUDA”, NVIDIA 2007 ●[Engel14] Wolfgang Engel, “Compute Shader Optimizations for AMD GPUs: Parallel Reduction”, Confetti 2014 ●[Harada12] Takahiro Harada, “A 2.5D Culling for Forward+”, Technical Briefs, SIGGRAPH Asia 2012 ●[Arvo90] Jim Arvo, “A simple method for box-sphere intersection testing”, Graphics Gems 1990 ●[Dufresne14] Marc Fauconneau Dufresne, “Forward Clustered Shading”, Intel 2014 ●[Persson13] Emil Persson, “Practical Clustered Shading”, Avalanche 2013 ●[Olsson et al12] Ola Olsson, Markus Billeter, Ulf Assarsson, “Clustered Deferred and Forward Shading”, HPG 2012 ●[Schulz14] Nicolas Schulz, “Moving to the Next Generation – The Rendering Technology of Ryse”, GDC 2014
- 60. Thanks ●Jason Stewart, AMD ●Epic Rendering Team ●Emil Persson, Avalanche Studios
Editor's Notes
- #15: Max Z might also be useful for transparent light list or tiled-based particle rendering
- #19: Simple test, prone to false positives
- #21: Alternative approach – use AABBs. Still prone to false positives.
- #22: AABB case is much better when depth bounds are small, but bad with large depth discontinuities. AABB suggested by Brian Karis from Epic
- #23: Martin Mittring from Epic initially implemented this method
- #30: Doesn’t require any changes to colour pixel shaders. Just trims light lists. Not perfect – consider purple light.
- #31: HalfZ requires a code change at in the colour pixel shaders to determine which light list to read
- #32: UE4 uses just one list, so like 2.5D, requires no extra work in colour pass. Probably a good idea if the number of lights per tile is low.
- #36: Better results for small depth ranges. Long frusta generate large AABBs
- #40: UE4 Infiltrator Demo
- #49: Add diagrams
- #53: Show diagram of layout. Mention that some lights might overlap slices, hence the safety factor Mention that TGSM needs to be kept under control or waves in flight will be reduced.
- #61: Plug Jason’s GPU Pro article