Repository navigation

cuda: add psnr_cuda, ssim_cuda and ciede_cuda feature extractors - #1563

Closed

BardieJoensen wants to merge 14 commits into

Netflix:masterfrom

BardieJoensen:cuda-psnr-ssim

Commits on Sep 4, 2026
- cuda: don't flush CUDA feature extractors twice with a thread pool
  BardieJoensen
  committed
  ae338ad
- cuda: add psnr_cuda feature extractor
  BardieJoensen
  committed
  046700a
- cuda: add ssim_cuda feature extractor (float_ssim port)
  BardieJoensen
  committed
  04d9ecd
- cuda: add CPU/CUDA parity test for psnr_cuda and ssim_cuda
  BardieJoensen
  committed
  378b8ce
- cuda/psnr: one atomic per block instead of per warp
  BardieJoensen
  committed
  de0efdb
- cuda: address review findings on psnr_cuda/ssim_cuda
  BardieJoensen
  committed
  bfcdf24
- cuda: run psnr_cuda/ssim_cuda kernels on their own streams
  BardieJoensen
  committed
  5d6bae3
- cuda/psnr: vectorized grid-stride loads
  BardieJoensen
  committed
  d59fff3
- cuda/ssim: fuse float conversion into the decimation kernel
  BardieJoensen
  committed
  d48f7d8
- cuda: make the extractor work streams legacy-blocking
  BardieJoensen
  committed
  271cd32
- cuda: attribute the new files' copyright headers
  BardieJoensen
  committed
  84f38d8
- cuda: add ciede_cuda feature extractor
  BardieJoensen
  committed
  5224989
- ciede: fix 4:2:2 chroma upsampling
  BardieJoensen
  committed
  347437b
- cuda: strengthen PSNR and SSIM parity coverage
  BardieJoensen
  committed
  3f64bed