mirror of
https://github.com/ddidderr/wuffs.git
synced 2026-11-17 16:32:11 +01:00
This only affects truncated (malformed) JPEGs. It moves Wuffs' std/jpeg decoder closer to libjpeg-turbo, as seen by the "hash of the decoded output" for test/data/peacock.default.truncated.jpeg now matching test/data/peacock.default.truncated.decoded-by-djpeg.png instead of test/data/peacock.default.truncated.decoded-by-wuffs.png The decoding of test/data/peacock.progressive.truncated.jpeg is not yet affected, requiring further investigation, but a future commit could similarly make that match djpeg. The two libraries don't use exactly the same algorithm because, for performance, Wuffs' decoder reads from the bitstream 8 bytes at a time where libjpeg-turbo reads only 1 byte at a time. While this change moves Wuffs' output closer to libjpeg-turbo (to repeat, only for truncated JPEGs; well-formed JPEGs are unaffected), the benchmarks below show a negative performance impact, and visual inspection of test/data/peacock.*.truncated.decoded-by-*.png shows that matching libjpeg-turbo *exactly* on truncated JPEGs doesn't actually always look better. On balance, it doesn't seem worth it to keep this change (and spend more time on that further investigation). This commit will soon be followed by a rollback, but it is committed anyway so that we can refer to these numbers in the git log. name old speed new speed delta wuffs_jpeg_decode_19k_8bpp/clang11 378MB/s ± 0% 364MB/s ± 0% -3.53% (p=0.008 n=5+5) wuffs_jpeg_decode_30k_24bpp_progressive/clang11 325MB/s ± 0% 340MB/s ± 0% +4.62% (p=0.016 n=5+4) wuffs_jpeg_decode_30k_24bpp_sequential/clang11 1.02GB/s ± 0% 0.99GB/s ± 0% -2.64% (p=0.008 n=5+5) wuffs_jpeg_decode_77k_24bpp/clang11 842MB/s ± 0% 815MB/s ± 0% -3.16% (p=0.008 n=5+5) wuffs_jpeg_decode_552k_24bpp_420/clang11 916MB/s ± 0% 887MB/s ± 0% -3.08% (p=0.008 n=5+5) wuffs_jpeg_decode_552k_24bpp_444/clang11 1.10GB/s ± 0% 1.07GB/s ± 2% -3.39% (p=0.008 n=5+5) wuffs_jpeg_decode_4002k_24bpp/clang11 919MB/s ± 0% 892MB/s ± 0% -2.95% (p=0.008 n=5+5) wuffs_jpeg_decode_19k_8bpp/gcc10 360MB/s ± 0% 348MB/s ± 0% -3.34% (p=0.008 n=5+5) wuffs_jpeg_decode_30k_24bpp_progressive/gcc10 290MB/s ± 0% 283MB/s ± 0% -2.46% (p=0.008 n=5+5) wuffs_jpeg_decode_30k_24bpp_sequential/gcc10 925MB/s ± 0% 909MB/s ± 0% -1.70% (p=0.008 n=5+5) wuffs_jpeg_decode_77k_24bpp/gcc10 758MB/s ± 0% 732MB/s ± 0% -3.43% (p=0.008 n=5+5) wuffs_jpeg_decode_552k_24bpp_420/gcc10 829MB/s ± 0% 799MB/s ± 0% -3.54% (p=0.008 n=5+5) wuffs_jpeg_decode_552k_24bpp_444/gcc10 981MB/s ± 0% 964MB/s ± 0% -1.72% (p=0.008 n=5+5) wuffs_jpeg_decode_4002k_24bpp/gcc10 835MB/s ± 0% 811MB/s ± 0% -2.89% (p=0.008 n=5+5)