bytes: speed up Compare() on amd64

Use AVX2 if available.
Results (haswell), below:

name                           old time/op    new time/op     delta
BytesCompare1-6                  11.4ns ± 0%     11.4ns ± 0%     ~     (all samples are equal)
BytesCompare2-6                  11.4ns ± 0%     11.4ns ± 0%     ~     (all samples are equal)
BytesCompare4-6                  11.4ns ± 0%     11.4ns ± 0%     ~     (all samples are equal)
BytesCompare8-6                  9.29ns ± 2%     8.76ns ± 0%   -5.72%        (p=0.000 n=16+17)
BytesCompare16-6                 9.29ns ± 2%     9.20ns ± 0%   -1.02%        (p=0.000 n=20+16)
BytesCompare32-6                 11.4ns ± 1%     11.4ns ± 0%     ~           (p=0.191 n=20+20)
BytesCompare64-6                 14.4ns ± 0%     13.1ns ± 0%   -8.68%        (p=0.000 n=20+20)
BytesCompare128-6                20.2ns ± 0%     18.5ns ± 0%   -8.27%        (p=0.000 n=16+20)
BytesCompare256-6                29.3ns ± 0%     24.5ns ± 0%  -16.38%        (p=0.000 n=16+16)
BytesCompare512-6                46.8ns ± 0%     37.1ns ± 0%  -20.78%        (p=0.000 n=18+16)
BytesCompare1024-6               82.9ns ± 0%     62.3ns ± 0%  -24.86%        (p=0.000 n=20+14)
BytesCompare2048-6                155ns ± 0%      112ns ± 0%  -27.74%        (p=0.000 n=20+20)
CompareBytesEqual-6              10.1ns ± 1%     10.0ns ± 1%     ~           (p=0.527 n=20+20)
CompareBytesToNil-6              10.0ns ± 2%      9.4ns ± 0%   -6.57%        (p=0.000 n=20+17)
CompareBytesEmpty-6              8.76ns ± 0%     8.76ns ± 0%     ~     (all samples are equal)
CompareBytesIdentical-6          8.76ns ± 0%     8.76ns ± 0%     ~     (all samples are equal)
CompareBytesSameLength-6         10.6ns ± 1%     10.6ns ± 1%     ~           (p=0.240 n=20+20)
CompareBytesDifferentLength-6    10.6ns ± 0%     10.6ns ± 1%     ~           (p=1.000 n=20+20)
CompareBytesBigUnaligned-6        132±s ± 1%      105±s ± 1%  -20.61%        (p=0.000 n=20+18)
CompareBytesBig-6                 125±s ± 1%      105±s ± 1%  -16.31%        (p=0.000 n=20+20)
CompareBytesBigIdentical-6       8.13ns ± 0%     8.13ns ± 0%     ~     (all samples are equal)

name                           old speed      new speed       delta
CompareBytesBigUnaligned-6     7.94GB/s ± 1%  10.01GB/s ± 1%  +25.96%        (p=0.000 n=20+18)
CompareBytesBig-6              8.38GB/s ± 1%  10.01GB/s ± 1%  +19.48%        (p=0.000 n=20+20)
CompareBytesBigIdentical-6      129TB/s ± 0%    129TB/s ± 0%   +0.01%        (p=0.003 n=17+19)

Change-Id: I820f31bab4582dd4204b146bb077c0d2f24cd8f5
Reviewed-on: https://go-review.googlesource.com/16434
Run-TryBot: Ilya Tocar <ilya.tocar@intel.com>
Reviewed-by: Klaus Post <klauspost@gmail.com>
TryBot-Result: Gobot Gobot <gobot@golang.org>
Reviewed-by: Keith Randall <khr@golang.org>
This commit is contained in:
Ilya Tocar 2015-10-28 23:20:26 +03:00 committed by Keith Randall
parent cf73357e37
commit 0e23ca41d9
5 changed files with 126 additions and 8 deletions

View file

@ -627,6 +627,8 @@ var (
cpuid_ecx uint32
cpuid_edx uint32
lfenceBeforeRdtsc bool
support_avx bool
support_avx2 bool
goarm uint8 // set by cmd/link on arm systems
)