Nigel Tao
|
5afaa8911c
|
pixconv: guard swizzle_ycc__etc__avx2 with #ifdef
|
2023-07-30 18:52:16 +10:00 |
|
Nigel Tao
|
2687d1daaf
|
pixconv: fix -Werror=pedantic/overflow errors
|
2023-07-23 22:11:19 +10:00 |
|
Nigel Tao
|
e2b9566fc5
|
base: restrict x86_avx2 to 64-bit (not 32-bit) x86
32-bit C/C++ compilers don't always compile the AVX intrinsics. Per the
60773 bug below,
> clang++ -m32 etc etc fuzz/c/std/jpeg_fuzzer.c
> undefined reference to `_mm_cvtsi128_si64'
> undefined reference to `_mm256_extract_epi64'
See also the "#ifdef __x86_64__" guards in
https://github.com/gcc-mirror/gcc/blob/4a3895903c29ed85da6fcb886f31ff23d4c6e935/gcc/config/i386/emmintrin.h#L228-L231
https://github.com/llvm/llvm-project/blob/08fab9ebecf72a682279d75489dc2460121cbeed/clang/lib/Headers/emmintrin.h#L3502-L3515
Note that -m32 is not the same as -mx32
> $ touch empty.c
> $ gcc -dM -E empty.c | grep -E '__i386__|__x86_64__'
> #define __x86_64__ 1
> $ gcc -m32 -dM -E empty.c | grep -E '__i386__|__x86_64__'
> #define __i386__ 1
> $ gcc -mx32 -dM -E empty.c | grep -E '__i386__|__x86_64__'
> #define __x86_64__ 1
Updates https://bugs.chromium.org/p/oss-fuzz/issues/detail?id=60773
|
2023-07-23 21:20:35 +10:00 |
|
Nigel Tao
|
cd769a0cdf
|
pixconv: add upsample_inv_h2v2_triangle_x86_avx2
name old speed new speed delta
wuffs_jpeg_decode_19k_8bpp/clang11 370MB/s ± 0% 376MB/s ± 0% +1.84% (p=0.008 n=5+5)
wuffs_jpeg_decode_30k_24bpp_progressive/clang11 295MB/s ± 0% 311MB/s ± 0% +5.56% (p=0.008 n=5+5)
wuffs_jpeg_decode_30k_24bpp_sequential/clang11 726MB/s ± 0% 1021MB/s ± 0% +40.62% (p=0.008 n=5+5)
wuffs_jpeg_decode_77k_24bpp/clang11 622MB/s ± 0% 842MB/s ± 0% +35.49% (p=0.008 n=5+5)
wuffs_jpeg_decode_552k_24bpp_420/clang11 659MB/s ± 0% 917MB/s ± 0% +39.15% (p=0.008 n=5+5)
wuffs_jpeg_decode_552k_24bpp_444/clang11 649MB/s ± 0% 649MB/s ± 0% -0.09% (p=0.016 n=5+5)
wuffs_jpeg_decode_4002k_24bpp/clang11 675MB/s ± 0% 921MB/s ± 0% +36.49% (p=0.008 n=5+5)
wuffs_jpeg_decode_19k_8bpp/gcc10 351MB/s ± 0% 326MB/s ± 0% -7.21% (p=0.008 n=5+5)
wuffs_jpeg_decode_30k_24bpp_progressive/gcc10 334MB/s ± 0% 292MB/s ± 0% -12.67% (p=0.008 n=5+5)
wuffs_jpeg_decode_30k_24bpp_sequential/gcc10 924MB/s ± 0% 920MB/s ± 0% -0.41% (p=0.008 n=5+5)
wuffs_jpeg_decode_77k_24bpp/gcc10 740MB/s ± 0% 742MB/s ± 0% +0.27% (p=0.008 n=5+5)
wuffs_jpeg_decode_552k_24bpp_420/gcc10 812MB/s ± 0% 792MB/s ± 0% -2.38% (p=0.008 n=5+5)
wuffs_jpeg_decode_552k_24bpp_444/gcc10 647MB/s ± 0% 627MB/s ± 0% -3.16% (p=0.016 n=5+4)
wuffs_jpeg_decode_4002k_24bpp/gcc10 817MB/s ± 0% 792MB/s ± 0% -3.01% (p=0.016 n=4+5)
|
2023-07-23 12:03:07 +10:00 |
|
Nigel Tao
|
35c1f76f86
|
pixconv: add swizzle_ycc__convert_rgbx_x86_avx2
Some of the benchmark numbers move around even though they don't use the
new etc__convert_rgbx_x86_avx2 code paths. It's possibly just a code
alignment coincidence per https://news.ycombinator.com/item?id=23633583
name old speed new speed delta
wuffs_jpeg_decode_19k_8bpp/clang11 372MB/s ± 0% 370MB/s ± 0% -0.64% (p=0.008 n=5+5)
wuffs_jpeg_decode_30k_24bpp_progressive/clang11 278MB/s ± 0% 295MB/s ± 0% +5.89% (p=0.008 n=5+5)
wuffs_jpeg_decode_30k_24bpp_sequential/clang11 710MB/s ± 0% 726MB/s ± 0% +2.30% (p=0.008 n=5+5)
wuffs_jpeg_decode_77k_24bpp/clang11 603MB/s ± 0% 622MB/s ± 0% +3.12% (p=0.008 n=5+5)
wuffs_jpeg_decode_552k_24bpp_420/clang11 635MB/s ± 0% 659MB/s ± 0% +3.82% (p=0.008 n=5+5)
wuffs_jpeg_decode_552k_24bpp_444/clang11 650MB/s ± 0% 649MB/s ± 0% -0.17% (p=0.008 n=5+5)
wuffs_jpeg_decode_4002k_24bpp/clang11 629MB/s ± 0% 675MB/s ± 0% +7.20% (p=0.008 n=5+5)
wuffs_jpeg_decode_19k_8bpp/gcc10 364MB/s ± 0% 351MB/s ± 0% -3.62% (p=0.008 n=5+5)
wuffs_jpeg_decode_30k_24bpp_progressive/gcc10 334MB/s ± 0% 334MB/s ± 0% -0.06% (p=0.032 n=5+5)
wuffs_jpeg_decode_30k_24bpp_sequential/gcc10 944MB/s ± 0% 924MB/s ± 0% -2.19% (p=0.008 n=5+5)
wuffs_jpeg_decode_77k_24bpp/gcc10 769MB/s ± 0% 740MB/s ± 0% -3.75% (p=0.008 n=5+5)
wuffs_jpeg_decode_552k_24bpp_420/gcc10 842MB/s ± 0% 812MB/s ± 0% -3.60% (p=0.008 n=5+5)
wuffs_jpeg_decode_552k_24bpp_444/gcc10 661MB/s ± 0% 647MB/s ± 0% -2.13% (p=0.008 n=5+5)
wuffs_jpeg_decode_4002k_24bpp/gcc10 849MB/s ± 0% 817MB/s ± 0% -3.77% (p=0.016 n=5+4)
|
2023-07-22 13:11:12 +10:00 |
|
Nigel Tao
|
4facef4e01
|
pixconv: add swizzle_ycc__convert_bgrx_x86_avx2
name old speed new speed delta
wuffs_jpeg_decode_19k_8bpp/clang11 365MB/s ± 1% 372MB/s ± 0% +1.77% (p=0.008 n=5+5)
wuffs_jpeg_decode_30k_24bpp_progressive/clang11 241MB/s ± 0% 278MB/s ± 0% +15.49% (p=0.008 n=5+5)
wuffs_jpeg_decode_30k_24bpp_sequential/clang11 477MB/s ± 0% 710MB/s ± 0% +48.91% (p=0.008 n=5+5)
wuffs_jpeg_decode_77k_24bpp/clang11 429MB/s ± 0% 603MB/s ± 0% +40.56% (p=0.008 n=5+5)
wuffs_jpeg_decode_552k_24bpp_420/clang11 446MB/s ± 0% 635MB/s ± 0% +42.18% (p=0.008 n=5+5)
wuffs_jpeg_decode_552k_24bpp_444/clang11 644MB/s ± 0% 650MB/s ± 0% +1.06% (p=0.016 n=4+5)
wuffs_jpeg_decode_4002k_24bpp/clang11 458MB/s ± 0% 629MB/s ± 0% +37.46% (p=0.008 n=5+5)
wuffs_jpeg_decode_19k_8bpp/gcc10 352MB/s ± 0% 364MB/s ± 0% +3.62% (p=0.008 n=5+5)
wuffs_jpeg_decode_30k_24bpp_progressive/gcc10 295MB/s ± 0% 334MB/s ± 0% +13.19% (p=0.008 n=5+5)
wuffs_jpeg_decode_30k_24bpp_sequential/gcc10 669MB/s ± 0% 944MB/s ± 0% +41.07% (p=0.008 n=5+5)
wuffs_jpeg_decode_77k_24bpp/gcc10 567MB/s ± 0% 769MB/s ± 0% +35.59% (p=0.008 n=5+5)
wuffs_jpeg_decode_552k_24bpp_420/gcc10 605MB/s ± 0% 842MB/s ± 0% +39.07% (p=0.008 n=5+5)
wuffs_jpeg_decode_552k_24bpp_444/gcc10 647MB/s ± 0% 661MB/s ± 0% +2.19% (p=0.008 n=5+5)
wuffs_jpeg_decode_4002k_24bpp/gcc10 613MB/s ± 0% 849MB/s ± 0% +38.61% (p=0.008 n=5+5)
|
2023-07-22 12:51:51 +10:00 |
|