Optimized Soft Shadow Volume Algorithm with Real-Time Performance

1. An Optimized Soft Shadow Volume Algorithm with Real-Time Performance Ulf Assarsson 1 , Michael Dougherty 2 , Michael Mounier 2 , and Tomas Akenine-Mller 1 1 Department of Computer Engineering Chalmers University of Technology 2 Xbox Advanced Technology Group, Microsoft

2. Soft Shadow Volume Algorithm Soft Shadow Volume Algorithm Soft Shadow Volume Algorithm Soft Shadow Volume Algorithm Ulf Assarsson and Tomas Akenine-Mller . A Geometry-Based Soft Shadow Volume A lgorithm Using Graphics Hardware SIGGRAPH 2003 , Tuesday, June 29, ~4.30 pm Area light sources Simple volumetric light sources Textures and short video textures as lights

3. This Paper This Paper This Paper This Paper Presents new optimizations Real-Time performance for: area light sources spherical light sources gray scale textures

4. Hard vs. soft shadows point source umbra area source umbra penumbra Two different light source types:

5. A Real-Time Soft Shadow Volume Algorithm

6. A wedge for each silhouette edge A wedge for each silhouette edge

7. A wedge for each silhouette edge A wedge for each silhouette edge

8. A wedge for each silhouette edge A wedge for each silhouette edge

9. A wedge for each silhouette edge A wedge for each silhouette edge

10. A wedge for each silhouette edge A wedge for each silhouette edge

11. Rasterizing the wedges Rasterizing the wedges

12. Rasterizing the wedges Rasterizing the wedges

13. Rasterizing the wedges Rasterizing the wedges

14. Rasterizing the wedges Rasterizing the wedges

15. Rasterizing the wedges Rasterizing the wedges

16. Optimizations Culling Only run pixel shader for points inside the wedge Tighter wedges around the penumbra 1.2 2 times speedup Optimized Pixel Shader Previously ~250 instructions, now ~60 instructions.

17. Artifacts and a Solution Artifacts and a Solution Artifacts and a Solution Artifacts and a Solution Two major approximations: Assume constant silhouette I.e., one single silhouette as seen from all locations of the light area. Ignore overlap between occluders Two major approximations: Assume constant silhouette I.e., one single silhouette as seen from all locations of the light area. Ignore overlap between occluders

18. Worst Case Example One light source 1024 point lights One light source 1024 point lights

19. For better quality, use 2x2 area lights (2x2 does not cost 4 times as much!) One light source 2x2 light sources 1024 point lights One light source 2x2 light sources 1024 point lights

20. For better quality, use 2x2 area lights (2x2 does not cost 4 times as much!) One light source 2x2 light sources 1024 point lights One light source 2x2 light sources 1024 point lights One light source 1024 point lights One light source 1024 point lights

21. For better quality, use 2x2 area lights (2x2 does not cost 4 times as much!) One light source 2x2 light sources 1024 point lights One light source 2x2 light sources 1024 point lights One light source 2x2 light sources 3x3 light sources 1024 point lights One light source 2x2 light sources 3x3 light sources 1024 point lights

22. Optimized Shaders We wrote two optimizated shaders, one for a sperical light source and one for a planar light source Both shaders follow approximatly the same steps 1. Transform the silhouette edge to projected space 2. Clip the edge and do a perspective divide 3. Calculate the coverage using the two resulting 2D points 4. Split the coverage and add it to the visibility mask

23. Transform the Silouhette Edge to Projection Space Projection space is defined as the point to be shaded at the origin and the light center located on the z axis at z = 1

24. Clip Edge and Do a Perspective Divide Clipping is done in homogenous space The square light shader clips edge to planes in homogenous space. The spherical light shader solves the equation of an intersection with a cone and line in homogenous space. The perspective divide is simply x/z and y/z.

25. Calculate Coverage Using the Resulting 2D Points The angles between A and the x axis ( 0 ) and B and the x axis ( 1 ) are calculated by looking up the x and y of the vectors in a cube map that implements atan2. 0 and 1 are used as coordinates to lookup in a 2D texture that returns the area defined by the intersection of the area swept by the rays from the light center through the projected points and the square or circle area. the cross product of A and B is subtracted from the area overestimate to give the coverage (shaded in gray).

27. Split the Coverage and Add it to the Visibility Mask A 1D texture is used to split the coverage value across four 8-bit channels. The spilt is required to allow for overflow when incrementing the visibility mask. Positive and negative contributions to the mask are stored separately. The split values are recombined with a dot product that scales each channel and adds them together The split and recombine steps could be skipped on hardware that supports blending on greater than 8-bit per channel render targets.

29. Quality Quality Quality Quality

30. Reference image Our algorithm Reference image Our algorithm Quality Quality Quality Quality

31. Reference image Our algorithm Reference image Our algorithm Quality Quality Quality Quality

32. Small light source Small light source Large light source Large light source Quality Quality Quality Quality

33. 512 point lights Our algorithm 512 point lights Our algorithm Quality Quality Quality Quality

34. Real-Time Demo coded by: Michael Dougherty and Michael Mounier Xbox Advanced Technology Group, Microsoft coded by: Michael Dougherty and Michael Mounier Xbox Advanced Technology Group, Microsoft

35. Demo Notes We do a pass to write out the world space xyz. The depth z and screen space x and y could be transformed by the inverse of the view*projection matrix to save bandwidth but we were limited on instruction count. There was no profiling done on the sample. It may be bound by the massive number of render state changes required to render each wedge half. On the ATI 9700 The default clip planes did not provide enough accuracy on clipped texture coordinate values. This caused the sampled world space xyz to be off. User clip planes fixed the problem. The center plane of each wedge needed to be shared (save vertices and cull order) by both wedge halves in order to avoid artifacts.

36. Thanks for listening Questions? Questions? Questions? Questions? More info here: www.ce.chalmers.se/staff/tomasm/soft/ More info here: www.ce.chalmers.se/staff/tomasm/soft/