tree 49102729db91c942352fc67f764ef1848d522739
parent 85452d2f6200e5e76592ffc4269c45d5ad9d13f3
author greatroar <61184462+greatroar@users.noreply.github.com> 1705311579 +0100
committer GitHub <noreply@github.com> 1705311579 +0100
gpgsig -----BEGIN PGP SIGNATURE-----
 
 wsBcBAABCAAQBQJlpP1bCRBK7hj4Ov3rIwAA2NMIAI0bxws1T3pNa8dwHNWC6oIR
 osWGapMPbQkeUtxWQRjfrsOdfIZQkRRwq7MPzyaTFLAi2Nhn62tMdnnkQ7dYZoNN
 YNkCF4qWzMIXFXEV79wvxqTGOH9XQU3Y/qxPiX35KBEACvHMBJzKgZiBzNoAA1Do
 IQMiguLCn1IS6MgsJjAREBvAptIModQl2cvSFJK2hW5Db9YFNmiXtHL6ohrBDx27
 puGIRh0zBieNrY4sOYM53BQl54MzhPOqU5MhL2w/qKT0t5qPSs1YHLkFWlfXhOzF
 pRg+6+n5jBlHu41NsRPkFclzdrVoGzKfT1wj9eXWY/Ve9LHb19eRXZiBW4DhJkU=
 =pUXq
 -----END PGP SIGNATURE-----
 

zstd: Tweak noasm FSE decoder (#910)

Benchmark results below. On a different machine, I saw somewhat bigger
speedups when benchmarking a version that has the noasm version of FSE,
but asm seqdec. The paper-100k benchmarks have some extreme variance in
the first 12 runs, so I did 12 more to get a more reliable average. The
apparent regression on that benchmark may be a fluke.

The asm FSE decoder still wins.

goos: linux
goarch: amd64
pkg: github.com/klauspost/compress/zstd
cpu: Intel(R) Core(TM) i7-3770K CPU @ 3.50GHz
                                                     │     old      │                 new                 │
                                                     │     B/s      │     B/s       vs base               │
Decoder_DecoderSmall/kppkn.gtb.zst/buffered-8          252.3Mi ± 0%   251.9Mi ± 0%  -0.12% (p=0.027 n=12)
Decoder_DecoderSmall/kppkn.gtb.zst/unbuffered-8        300.2Mi ± 1%   299.5Mi ± 1%       ~ (p=0.977 n=12)
Decoder_DecoderSmall/geo.protodata.zst/buffered-8      702.1Mi ± 0%   701.4Mi ± 0%       ~ (p=0.551 n=12)
Decoder_DecoderSmall/geo.protodata.zst/unbuffered-8    703.1Mi ± 1%   695.9Mi ± 1%       ~ (p=0.178 n=12)
Decoder_DecoderSmall/plrabn12.txt.zst/buffered-8       200.3Mi ± 0%   200.2Mi ± 0%       ~ (p=0.417 n=12)
Decoder_DecoderSmall/plrabn12.txt.zst/unbuffered-8     294.3Mi ± 2%   294.2Mi ± 2%       ~ (p=0.755 n=12)
Decoder_DecoderSmall/lcet10.txt.zst/buffered-8         237.7Mi ± 0%   237.7Mi ± 0%       ~ (p=0.561 n=12)
Decoder_DecoderSmall/lcet10.txt.zst/unbuffered-8       315.7Mi ± 1%   313.0Mi ± 2%       ~ (p=0.713 n=12)
Decoder_DecoderSmall/asyoulik.txt.zst/buffered-8       205.7Mi ± 3%   212.8Mi ± 0%  +3.43% (p=0.002 n=12)
Decoder_DecoderSmall/asyoulik.txt.zst/unbuffered-8     267.1Mi ± 2%   273.4Mi ± 4%       ~ (p=0.078 n=12)
Decoder_DecoderSmall/alice29.txt.zst/buffered-8        194.8Mi ± 0%   194.8Mi ± 0%       ~ (p=0.876 n=12)
Decoder_DecoderSmall/alice29.txt.zst/unbuffered-8      219.8Mi ± 2%   221.5Mi ± 1%       ~ (p=0.410 n=12)
Decoder_DecoderSmall/html_x_4.zst/buffered-8           1.512Gi ± 1%   1.550Gi ± 0%  +2.49% (p=0.000 n=12)
Decoder_DecoderSmall/html_x_4.zst/unbuffered-8         1.149Gi ± 2%   1.196Gi ± 2%  +4.09% (p=0.000 n=12)
Decoder_DecoderSmall/paper-100k.pdf.zst/buffered-8     2.504Gi ± 1%   2.513Gi ± 0%       ~ (p=0.136 n=24)
Decoder_DecoderSmall/paper-100k.pdf.zst/unbuffered-8   1.041Gi ± 3%   1.023Gi ± 0%  -1.72% (p=0.002 n=24)
Decoder_DecoderSmall/fireworks.jpeg.zst/buffered-8     6.315Gi ± 0%   6.349Gi ± 0%  +0.53% (p=0.000 n=12)
Decoder_DecoderSmall/fireworks.jpeg.zst/unbuffered-8   3.178Gi ± 3%   3.234Gi ± 2%       ~ (p=0.713 n=12)
Decoder_DecoderSmall/urls.10K.zst/buffered-8           350.9Mi ± 0%   351.0Mi ± 0%       ~ (p=1.000 n=12)
Decoder_DecoderSmall/urls.10K.zst/unbuffered-8         525.8Mi ± 1%   524.9Mi ± 3%       ~ (p=0.671 n=12)
Decoder_DecoderSmall/html.zst/buffered-8               562.0Mi ± 0%   561.2Mi ± 0%       ~ (p=0.054 n=12)
Decoder_DecoderSmall/html.zst/unbuffered-8             537.4Mi ± 2%   543.5Mi ± 2%       ~ (p=0.178 n=12)
Decoder_DecoderSmall/comp-data.bin.zst/buffered-8      283.4Mi ± 0%   283.8Mi ± 0%       ~ (p=0.164 n=12)
Decoder_DecoderSmall/comp-data.bin.zst/unbuffered-8    118.0Mi ± 1%   120.3Mi ± 3%       ~ (p=0.128 n=12)
geomean                                                505.1Mi        508.1Mi       +0.60%