Files
UnrealEngineUWP/Engine/Source/Developer/TextureCompressor
danny couture 90ed5f144f Optimize NVTT texture compression using AVX2
- Load the proper delay-loaded dll depending on instruction support of the current platform.
  - Replaced reciprocal approximation with proper division.
    - This improves performance as the division is faster than a reciprocal with refinement on recent CPUs.
    - This improves quality as the division provides higher precision.
    - It fixes a nasty non-determinism in the DDC because reciprocal yield different results on different CPUs.
  - Configured so that AVX2 and SSE2 results are exactly the same (see config.h).
  - More details can be found in config.h

  - 80.773s -> 47.547s for single-threaded 8k texture compression
  - 18m34s -> 14m50s for recompressing all textures of Reverb P_World during level loading.

#rb Francis.Hurteau, Geoff.Evans

[CL 14362917 by danny couture in ue5-main branch]
2020-09-22 05:48:20 -04:00
..