There is an issue with the constants used in YUV to YUV range conversion,
where the upper bound is not respected when converting to mpeg range.
With this patchset, the constants are calculated at runtime, depending on
the bit depth. This approach also allows us to more easily understand how
the constants are derived.
These are the speedups for the entire patchset:
x86_64:
chrRangeFromJpeg8_1920_c: 5827.4 5845.2 ( 1.00x)
chrRangeFromJpeg8_1920_sse2: 1945.6 1955.2 ( 1.00x)
chrRangeFromJpeg8_1920_avx2: 992.0 988.9 ( 1.00x)
chrRangeFromJpeg16_1920_c: 5793.2 5809.1 ( 1.00x)
chrRangeToJpeg8_1920_c: 11726.2 9462.2 ( 1.24x)
chrRangeToJpeg8_1920_sse2: 1965.5 1949.9 ( 1.01x)
chrRangeToJpeg8_1920_avx2: 984.2 988.5 ( 1.00x)
chrRangeToJpeg16_1920_c: 10610.8 9261.5 ( 1.15x)
lumRangeFromJpeg8_1920_c: 4165.7 4191.4 ( 0.99x)
lumRangeFromJpeg8_1920_sse2: 1032.0 1040.5 ( 0.99x)
lumRangeFromJpeg8_1920_avx2: 575.2 520.5 ( 1.11x)
lumRangeFromJpeg16_1920_c: 4530.0 4143.4 ( 1.09x)
lumRangeToJpeg8_1920_c: 6044.8 5720.5 ( 1.06x)
lumRangeToJpeg8_1920_sse2: 1034.2 1046.0 ( 0.99x)
lumRangeToJpeg8_1920_avx2: 513.5 540.5 ( 0.95x)
lumRangeToJpeg16_1920_c: 5343.6 5139.5 ( 1.04x)
aarch64 A55:
chrRangeFromJpeg8_1920_c: 28839.3 28834.8 ( 1.00x)
chrRangeFromJpeg8_1920_neon: 5312.2 5313.1 ( 1.00x)
chrRangeFromJpeg16_1920_c: 28843.8 28840.6 ( 1.00x)
chrRangeToJpeg8_1920_c: 44196.1 23072.5 ( 1.92x)
chrRangeToJpeg8_1920_neon: 6035.9 5550.8 ( 1.09x)
chrRangeToJpeg16_1920_c: 36526.7 23075.1 ( 1.58x)
lumRangeFromJpeg8_1920_c: 15384.3 15386.7 ( 1.00x)
lumRangeFromJpeg8_1920_neon: 3148.6 3145.8 ( 1.00x)
lumRangeFromJpeg16_1920_c: 15390.1 15383.8 ( 1.00x)
lumRangeToJpeg8_1920_c: 23066.7 19223.6 ( 1.20x)
lumRangeToJpeg8_1920_neon: 3868.8 3624.9 ( 1.07x)
lumRangeToJpeg16_1920_c: 19224.6 19225.5 ( 1.00x)
aarch64 A76:
chrRangeFromJpeg8_1920_c: 6316.2 6318.5 ( 1.00x)
chrRangeFromJpeg8_1920_neon: 2263.5 2304.2 ( 0.98x)
chrRangeFromJpeg16_1920_c: 6321.9 6323.5 ( 1.00x)
chrRangeToJpeg8_1920_c: 11389.3 9170.0 ( 1.24x)
chrRangeToJpeg8_1920_neon: 2644.2 2793.8 ( 0.95x)
chrRangeToJpeg16_1920_c: 9514.4 9195.6 ( 1.03x)
lumRangeFromJpeg8_1920_c: 4376.0 4425.5 ( 0.99x)
lumRangeFromJpeg8_1920_neon: 1110.8 1105.0 ( 1.01x)
lumRangeFromJpeg16_1920_c: 4437.9 4436.8 ( 1.00x)
lumRangeToJpeg8_1920_c: 6667.0 6017.2 ( 1.11x)
lumRangeToJpeg8_1920_neon: 1327.5 1328.0 ( 1.00x)
lumRangeToJpeg16_1920_c: 6062.5 6017.2 ( 1.01x)
NOTE: simd optimizations for x86 and aarch64 have been updated, but riscv
and loongarch are still missing (and therefore disabled).
NOTE2: the same issue still exists in rgb2yuv conversions, which is not
addressed in this patchset.
Changes from v1:
- Saturate the output value instead of limiting the input with amax;
- Add more comprehensive benchmarks to commit messages;
- Add comments when disabling code with "#if 0";
Ramiro Polla (16):
swscale/range_convert: call arch-specific init functions from main
init function
swscale/range_convert: drop redundant conditionals from arch-specific
init functions
swscale/range_convert: indent after previous commit
checkasm: use FF_ARRAY_ELEMS instead of hardcoding size of arrays
checkasm/sw_range_convert: use YUV pixel formats instead of YUVJ
checkasm/sw_range_convert: reduce number of input sizes tested
checkasm/sw_range_convert: only run benchmarks on largest input width
checkasm/sw_range_convert: test all supported bit depths
checkasm/sw_range_convert: indent after previous couple of commits
swscale/range_convert: saturate output instead of limiting input
swscale/aarch64/range_convert: saturate output instead of limiting
input
swscale/range_convert: fix mpeg ranges in yuv range conversion for
non-8-bit pixel formats
swscale/x86/range_convert: update sse2 and avx2 range_convert
functions to new API
swscale/x86: add sse2, sse4, and avx2 {lum,chr}ConvertRange16
swscale/aarch64/range_convert: update neon range_convert functions to
new API
swscale/aarch64: add neon {lum,chr}ConvertRange16
libswscale/aarch64/range_convert_neon.S | 152 ++++++++++----
libswscale/aarch64/swscale.c | 41 +++-
libswscale/hscale.c | 6 +-
libswscale/loongarch/swscale_init_loongarch.c | 38 ++--
libswscale/riscv/swscale.c | 15 +-
libswscale/swscale.c | 122 ++++++++++--
libswscale/swscale_internal.h | 11 +-
libswscale/utils.c | 10 +-
libswscale/x86/range_convert.asm | 161 ++++++++++-----
libswscale/x86/swscale.c | 56 ++++--
tests/checkasm/sw_gbrp.c | 15 +-
tests/checkasm/sw_range_convert.c | 186 +++++++++++++-----
tests/checkasm/sw_scale.c | 11 +-
.../fate/filter-alphaextract_alphamerge_rgb | 100 +++++-----
tests/ref/fate/filter-pixdesc-gray10be | 2 +-
tests/ref/fate/filter-pixdesc-gray10le | 2 +-
tests/ref/fate/filter-pixdesc-gray12be | 2 +-
tests/ref/fate/filter-pixdesc-gray12le | 2 +-
tests/ref/fate/filter-pixdesc-gray14be | 2 +-
tests/ref/fate/filter-pixdesc-gray14le | 2 +-
tests/ref/fate/filter-pixdesc-gray16be | 2 +-
tests/ref/fate/filter-pixdesc-gray16le | 2 +-
tests/ref/fate/filter-pixdesc-gray9be | 2 +-
tests/ref/fate/filter-pixdesc-gray9le | 2 +-
tests/ref/fate/filter-pixdesc-ya16be | 2 +-
tests/ref/fate/filter-pixdesc-ya16le | 2 +-
tests/ref/fate/filter-pixdesc-yuvj411p | 2 +-
tests/ref/fate/filter-pixdesc-yuvj420p | 2 +-
tests/ref/fate/filter-pixdesc-yuvj422p | 2 +-
tests/ref/fate/filter-pixdesc-yuvj440p | 2 +-
tests/ref/fate/filter-pixdesc-yuvj444p | 2 +-
tests/ref/fate/filter-pixfmts-copy | 34 ++--
tests/ref/fate/filter-pixfmts-crop | 34 ++--
tests/ref/fate/filter-pixfmts-field | 34 ++--
tests/ref/fate/filter-pixfmts-fieldorder | 30 +--
tests/ref/fate/filter-pixfmts-hflip | 34 ++--
tests/ref/fate/filter-pixfmts-il | 34 ++--
tests/ref/fate/filter-pixfmts-lut | 18 +-
tests/ref/fate/filter-pixfmts-null | 34 ++--
tests/ref/fate/filter-pixfmts-pad | 22 +--
tests/ref/fate/filter-pixfmts-pullup | 10 +-
tests/ref/fate/filter-pixfmts-rotate | 4 +-
tests/ref/fate/filter-pixfmts-scale | 34 ++--
tests/ref/fate/filter-pixfmts-swapuv | 10 +-
.../ref/fate/filter-pixfmts-tinterlace_cvlpf | 8 +-
.../ref/fate/filter-pixfmts-tinterlace_merge | 8 +-
tests/ref/fate/filter-pixfmts-tinterlace_pad | 8 +-
tests/ref/fate/filter-pixfmts-tinterlace_vlpf | 8 +-
tests/ref/fate/filter-pixfmts-transpose | 28 +--
tests/ref/fate/filter-pixfmts-vflip | 34 ++--
tests/ref/fate/fitsenc-gray | 2 +-
tests/ref/fate/fitsenc-gray16be | 10 +-
tests/ref/fate/gifenc-gray | 186 +++++++++---------
tests/ref/fate/idroq-video-encode | 2 +-
tests/ref/fate/jpg-icc | 8 +-
tests/ref/fate/sws-yuv-colorspace | 2 +-
tests/ref/fate/sws-yuv-range | 2 +-
tests/ref/fate/vvc-conformance-SCALING_A_1 | 128 ++++++------
tests/ref/lavf/gray16be.fits | 4 +-
tests/ref/lavf/gray16be.pam | 4 +-
tests/ref/lavf/gray16be.png | 6 +-
tests/ref/lavf/jpg | 6 +-
tests/ref/lavf/smjpeg | 6 +-
tests/ref/pixfmt/yuvj420p | 2 +-
tests/ref/pixfmt/yuvj422p | 2 +-
tests/ref/pixfmt/yuvj440p | 2 +-
tests/ref/pixfmt/yuvj444p | 2 +-
tests/ref/seek/lavf-jpg | 8 +-
tests/ref/seek/vsynth_lena-mjpeg | 40 ++--
tests/ref/seek/vsynth_lena-roqvideo | 2 +-
tests/ref/vsynth/vsynth1-amv | 8 +-
tests/ref/vsynth/vsynth1-mjpeg | 6 +-
tests/ref/vsynth/vsynth1-mjpeg-422 | 6 +-
tests/ref/vsynth/vsynth1-mjpeg-444 | 6 +-
tests/ref/vsynth/vsynth1-mjpeg-huffman | 6 +-
tests/ref/vsynth/vsynth1-mjpeg-trell | 8 +-
tests/ref/vsynth/vsynth1-mjpeg-trell-huffman | 8 +-
tests/ref/vsynth/vsynth1-roqvideo | 8 +-
tests/ref/vsynth/vsynth2-amv | 6 +-
tests/ref/vsynth/vsynth2-mjpeg | 6 +-
tests/ref/vsynth/vsynth2-mjpeg-422 | 6 +-
tests/ref/vsynth/vsynth2-mjpeg-444 | 6 +-
tests/ref/vsynth/vsynth2-mjpeg-huffman | 6 +-
tests/ref/vsynth/vsynth2-mjpeg-trell | 8 +-
tests/ref/vsynth/vsynth2-mjpeg-trell-huffman | 8 +-
tests/ref/vsynth/vsynth2-roqvideo | 8 +-
tests/ref/vsynth/vsynth3-amv | 8 +-
tests/ref/vsynth/vsynth3-mjpeg | 8 +-
tests/ref/vsynth/vsynth3-mjpeg-422 | 8 +-
tests/ref/vsynth/vsynth3-mjpeg-444 | 6 +-
tests/ref/vsynth/vsynth3-mjpeg-huffman | 8 +-
tests/ref/vsynth/vsynth3-mjpeg-trell | 6 +-
tests/ref/vsynth/vsynth3-mjpeg-trell-huffman | 6 +-
tests/ref/vsynth/vsynth_lena-amv | 6 +-
tests/ref/vsynth/vsynth_lena-mjpeg | 8 +-
tests/ref/vsynth/vsynth_lena-mjpeg-422 | 6 +-
tests/ref/vsynth/vsynth_lena-mjpeg-444 | 6 +-
tests/ref/vsynth/vsynth_lena-mjpeg-huffman | 8 +-
tests/ref/vsynth/vsynth_lena-mjpeg-trell | 8 +-
.../vsynth/vsynth_lena-mjpeg-trell-huffman | 8 +-
tests/ref/vsynth/vsynth_lena-roqvideo | 8 +-
101 files changed, 1193 insertions(+), 833 deletions(-)
--
2.30.2
_______________________________________________
ffmpeg-devel mailing list
ffmpeg-devel@ffmpeg.org
https://ffmpeg.org/mailman/listinfo/ffmpeg-devel
To unsubscribe, visit link above, or email
ffmpeg-devel-requ...@ffmpeg.org with subject "unsubscribe".
