![Previous Work
• [Wang et al. 2007]
̵ 16-bpp HDR texture format
̵ Utilization of current generation GPUs
• [Roimela et al. 2006, 2008]
̵ 8-bpp HDR texture format
̵ Simple hardware decoder
• [Munkberg et al. 2006, 2008]
̵ 8-bpp HDR texture format
̵ Near lossless visual quality
5](https://image.slidesharecdn.com/yanlu-dhtchdrtexturecompression-100331003512-phpapp01/85/Yanlu-Dhtc-Hdr-Texture-Compression-5-320.jpg)
![Chrominance Local Dynamic Range Reduction
• Adaptive color transform significantly reduces the
chrominance dynamic range
Chrominance value distribution in [0, 1]
12](https://image.slidesharecdn.com/yanlu-dhtchdrtexturecompression-100331003512-phpapp01/85/Yanlu-Dhtc-Hdr-Texture-Compression-12-320.jpg)
![Results
• Log[RGB] RMSE
Munkberg Roimela
Textures DHTC
2007 2008
BigFogMap 0.06 0.06 0.07
Cathedral 0.17 0.17 0.35
Memorial 0.14 0.13 0.31
Room 0.09 0.08 0.15
Desk 0.17 0.22 1.26
Tubes 0.32 0.28 0.81
Average 0.16 0.16 0.49
27](https://image.slidesharecdn.com/yanlu-dhtchdrtexturecompression-100331003512-phpapp01/85/Yanlu-Dhtc-Hdr-Texture-Compression-27-320.jpg)
Yanlu Dhtc Hdr Texture Compression
- 1. DHTC: An Effective DXTC-based HDR Texture Compression Scheme Wen Sun1,2 Yan Lu1 Feng Wu1 Shipeng Li1 yanlu@microsoft.com 1 Microsoft Research Asia 2 University of Science and Technology of China
- 2. Outline • Background and related work ̵ High dynamic range (HDR) texture and its compression • DHTC: DXTC-based HDR texture compression ̵ Extension of LDR texture format (i.e. DXTC) to HDR texture compression ̵ Unified 8-bpp format for LDR textures, HDR textures and alpha maps • Results and summary 2
- 3. HDR Textures Low exposure Medium exposure High exposure • The real world is high dynamic range • A dynamic range of 10000:1 is common • HDR rendering is gaining popularity in practice 3
- 4. HDR Texture Compression • HDR textures are huge in size ̵ Currently used FP32/FP16 formats: 96/48 bits per pixel, 4x/2x size of raw LDR RGB textures ̵ Consume too much memory and bandwidth • Current status ̵ No HDR texture compression standard in industry ̵ No graphics card supports rendering from block-wise compressed HDR textures 4
- 5. Previous Work • [Wang et al. 2007] ̵ 16-bpp HDR texture format ̵ Utilization of current generation GPUs • [Roimela et al. 2006, 2008] ̵ 8-bpp HDR texture format ̵ Simple hardware decoder • [Munkberg et al. 2006, 2008] ̵ 8-bpp HDR texture format ̵ Near lossless visual quality 5
- 6. Our Insights • HDR texture compression scheme can be built upon existing LDR texture compression scheme ̵ Lead to a unified compression framework ̵ Reuse existing hardware • Joint color-channel compression can result in better visual quality ̵ Joint-channel bit allocation ̵ Utilization of cross-channel correlations • It is a plus to support LDR textures and alpha maps in a single HDR texture format 6
- 7. Our Solution • DXTC-based HDR texture compression framework ̵ Utilize joint color-channel compression to provide advanced bit allocation ̵ Utilize the existing DXTC hardware to reduce the adoption cost in industry • 8 bpp compressed DHTC texture format ̵ Near lossless visual quality for HDR textures ̵ Support 1 bit alpha channel for HDR textures ̵ Backward compatible to LDR RGBA textures 7
- 8. Framework Extract joint Avoid error Utilize local channel accumulation property correlation Original Adaptive Local HDR Joint-Channel Compressed HDR Texture Color Trans. Reduction Compression HDR Texture Point Translation Reshape texel distribution 8
- 9. Adaptive Color Transform • Traditional color transform Forward Color Inverse Color Transform Transform Y wr R wg G wb B R U Y / wr Error controllable channels U wr R Explicit G V Y / wg Y channels B (Y wr R wg G ) / wb Error accumulative wg G channel V Y wB Implicit W b Y channel Absolute errors 9
- 10. Adaptive Color Transform • Our solution ̵ Adaptively select the implicit channel to minimize the impact of error accumulation Keep the block dominant color Luminance and channel from chrominance channels being explicitly encoded Adaptive color transform mode 10
- 11. Luminance Local Dynamic Range Reduction • Extract the block upper and lower bounds Extract the block upper & lower bounds (Global Base Luma – L0, L1) Mapping Global HDR to Local LDR (Floating Point -> integer) 11
- 12. Chrominance Local Dynamic Range Reduction • Adaptive color transform significantly reduces the chrominance dynamic range Chrominance value distribution in [0, 1] 12
- 13. Chrominance Local Dynamic Range Reduction • Proposed adaptive log/linear encoding Use a 1-bit flag in each texture block to adaptively select the encoding mode Linear encoding is better Log encoding is better 13
- 14. Joint Channel Compression & Point Translation • Basic idea – Joint channel linear fitting ̵ Simple and hardware friendly Y ̵ But rely on texel distribution V • Point translation U ̵ Translate texels along Y axis to reshape the distribution 14
- 15. Point Translation • We use a constant modifier table to provide translation vector for each texel Look Up Translation Vectors One HDR texture block Modifier table 15
- 16. Point Translation • We use a constant modifier table to provide translation vector for each texel Look Up Translation Vectors One HDR texture block Modifier table 16
- 17. DHTC Format • Bits layout of a 4x4 block 64 bits 64 bits Extension Block DXT1 Block Local dynamic range reduction L0 L1 (5+5) bits 4 bits T_idx Ch_mode 2 bits Adaptive color XXX XXX XXX XXX transform M_idx XXX XXX XXX XXX 16x3 bits XXX XXX XXX XXX XXX XXX XXX XXX Point translation Base Y0 X0 Z0 2x(5+6+5) Color Y1 X1 Z1 bits Joint channel XX XX XX XX compression XX XX XX XX 16x2 C_idx XX XX XX XX bits XX XX XX XX 17
- 18. Decoding Logic for HDR Textures L0 L1 Ch_mode T_idx M_idx C_idx X0 Y0 Z0 X1 Y1 Z1 Modifier Table DXT Decoder MUX Yint Uint Vint Log Decoder Y U V Inverse Color Transform Rh Gh Bh 18
- 19. Decoding Logic for LDR Textures L0 L1 Ch_mode T_idx M_idx C_idx X0 Y0 Z0 X1 Y1 Z1 Modifier Table DXT Decoder MUX MUX 0 255 x y (which texel) MUX A R G B 19
- 20. Results • Original HDR texture at different exposures 20
- 21. Results • DHTC compressed at 8 bpp 21
- 22. Results • Original HDR texture at different exposures 22
- 23. Results • DHTC compressed at 8 bpp 23
- 24. Results • Visual comparison with the state-of-the-art Munkberg et al Roimela et al Original DHTC 24 2007 2008
- 25. Results • Visual comparison with the state-of-the-art Munkberg et al Roimela et al Original DHTC 25 2007 2008
- 26. Results • mPSNR (dB) Munkberg Roimela Textures DHTC 2007 2008 BigFogMap 51.0 51.9 50.4 Cathedral 39.7 40.0 34.3 Memorial 46.8 46.5 41.7 Room 48.1 48.6 44.0 Desk 41.5 40.3 28.4 Tubes 35.7 35.7 27.0 Average 43.8 43.8 37.6 26
- 27. Results • Log[RGB] RMSE Munkberg Roimela Textures DHTC 2007 2008 BigFogMap 0.06 0.06 0.07 Cathedral 0.17 0.17 0.35 Memorial 0.14 0.13 0.31 Room 0.09 0.08 0.15 Desk 0.17 0.22 1.26 Tubes 0.32 0.28 0.81 Average 0.16 0.16 0.49 27
- 28. Results • HDR-VDP above 75% error (%) Munkberg Roimela Textures DHTC 2007 2008 BigFogMap 0.00 0.00 0.00 Cathedral 0.10 0.02 0.03 Memorial 0.01 0.00 0.00 Room 0.01 0.01 0.00 Desk 0.03 0.01 0.00 Tubes 0.87 1.25 1.20 Average 0.17 0.22 0.21 28
- 29. Results • Rendered scenes Original DHTC, 50dB 29
- 30. Results • Rendered scenes Original DHTC, 50dB 30
- 31. Results • Rendered scenes Original DHTC, 53dB 31
- 32. Results • 1-bit HDR alpha blending DHTC compressed HDR texture with 1-bit alpha channel 32
- 33. Results • 1-bit HDR alpha blending Alpha blended scenes 33
- 34. Results • Alpha coding for LDR textures ̵ We test 12 alpha maps from textures used in an Xbox game 34
- 35. Results • Alpha coding for LDR textures PSNR(dB) DHTC DXT5 AlphaMap1 49.2 48.6 Our method provides AlphaMap2 53.6 54. 8 better results in most AlphaMap3 54.4 57.1 cases than DXT5 alpha AlphaMap4 56.4 61.2 coding AlphaMap5 46.3 45.3 AlphaMap6 47.5 45.2 AlphaMap7 40.2 39.0 AlphaMap8 47.1 46.1 AlphaMap9 47.3 44.2 AlphaMap10 44.9 42.9 AlphaMap11 45.0 43.8 AlphaMap12 41.9 40.2 35 Average 47.8 47.4
- 36. Results • Alpha coding for LDR textures ̵ Less block artifact in our method DHTC DXT5 36
- 37. Summary • DXTC-based HDR texture compression ̵ Compress HDR textures into 8 bpp with the best quality so far ̵ Utilize the existing DXTC decoding hardware to minimize the adoption cost ̵ Provide a unified solution to compress HDR textures, LDR textures and alpha maps 37
- 38. Acknowledgements • Kimmo Roimela of Nokia Research Center and Jacob Munkberg of Lund University for providing their testing results for comparison. • Xin Tong, Liyi Wei, Baining Guo, John Tardif, Matt Bronder, Andrew Flavell of Microsoft for the valuable comments and suggestions. • John Owens of Univ. of California, Davis, for the helps in improving the paper quality.