tree 9d37b5aae9ee42a155897f332ae5fba74fd8d6ce
parent ec32eb3eb78bdc90fb09b6889b668e54c9735d21
author Nigel Tao <nigeltao@golang.org> 1624331672 +1000
committer Nigel Tao <nigeltao@golang.org> 1624332445 +1000

Have "x86_avx2" require PCLMUL, POPCNT and SSE4.2

The std/crc32 benchmark numbers get worse. The benches still run AVX2
code, but it runs slower for an unknown reason. The std/gzip and std/png
benchmark numbers are mixed.

Still, it's cleaner for Wuffs' "x86_avx2" to be a superset of
"x86_sse42" (and apart from PCLMUL, they're now similar to Clang 12 and
GCC 11's "x86-64-v3" and "x86-64-v2" micro-architecture feature levels).
Maybe future compiler versions can claw back the std/crc32 performance.

On a mid-range x86_64 laptop (2016, Skylake):

name                                                               old speed      new speed      delta

wuffs_crc32_ieee_10k/clang9                                        8.07GB/s ± 0%  7.61GB/s ± 1%   -5.72%  (p=0.000 n=8+8)
wuffs_crc32_ieee_100k/clang9                                       11.0GB/s ± 1%   9.7GB/s ± 3%  -11.45%  (p=0.000 n=7+8)

wuffs_crc32_ieee_10k/gcc10                                         9.77GB/s ± 1%  7.72GB/s ± 1%  -21.00%  (p=0.000 n=10+9)
wuffs_crc32_ieee_100k/gcc10                                        12.7GB/s ± 2%  11.9GB/s ± 2%   -5.84%  (p=0.000 n=9+9)

wuffs_gzip_decode_10k/clang9                                        217MB/s ± 2%   221MB/s ± 0%   +1.80%  (p=0.000 n=10+8)
wuffs_gzip_decode_100k/clang9                                       280MB/s ± 1%   283MB/s ± 0%   +1.38%  (p=0.000 n=9+8)

wuffs_gzip_decode_10k/gcc10                                         215MB/s ± 1%   220MB/s ± 1%   +2.52%  (p=0.000 n=10+9)
wuffs_gzip_decode_100k/gcc10                                        272MB/s ± 0%   274MB/s ± 1%   +0.74%  (p=0.000 n=9+10)

wuffs_png_decode_image_19k_8bpp/clang9                              136MB/s ± 0%   142MB/s ± 0%   +3.95%  (p=0.000 n=8+8)
wuffs_png_decode_image_40k_24bpp/clang9                             171MB/s ± 1%   172MB/s ± 0%   +0.84%  (p=0.000 n=10+8)
wuffs_png_decode_image_77k_8bpp/clang9                              492MB/s ± 3%   501MB/s ± 2%   +1.88%  (p=0.009 n=10+10)
wuffs_png_decode_image_552k_32bpp_ignore_checksum/clang9            456MB/s ± 1%   461MB/s ± 1%   +1.04%  (p=0.000 n=9+9)
wuffs_png_decode_image_552k_32bpp_verify_checksum/clang9            439MB/s ± 1%   440MB/s ± 3%     ~     (p=0.173 n=8+10)
wuffs_png_decode_image_4002k_24bpp/clang9                           172MB/s ± 1%   174MB/s ± 1%   +1.29%  (p=0.005 n=10+10)

wuffs_png_decode_image_19k_8bpp/gcc10                               152MB/s ± 0%   149MB/s ± 1%   -1.92%  (p=0.000 n=9+9)
wuffs_png_decode_image_40k_24bpp/gcc10                              184MB/s ± 0%   186MB/s ± 1%   +1.37%  (p=0.000 n=8+9)
wuffs_png_decode_image_77k_8bpp/gcc10                               550MB/s ± 1%   542MB/s ± 2%   -1.46%  (p=0.000 n=10+10)
wuffs_png_decode_image_552k_32bpp_ignore_checksum/gcc10             473MB/s ± 1%   478MB/s ± 1%   +1.23%  (p=0.000 n=9+10)
wuffs_png_decode_image_552k_32bpp_verify_checksum/gcc10             453MB/s ± 1%   458MB/s ± 1%   +0.91%  (p=0.000 n=9+10)
wuffs_png_decode_image_4002k_24bpp/gcc10                            185MB/s ± 2%   188MB/s ± 0%   +1.82%  (p=0.000 n=9+8)