go /
image /
93fad3647fcda43e12f413a6dd7568896dc06ba9 vector: remove a shift from the inner loop.
PSHUFB already lets us pick certain bytes out of an XMM register.
There's no need to shift by 8 bits (1 byte) beforehand.
The code generator is simpler. There's also a small win on the
benchmarks, especially for FixedAccumulateOpOverSIMD.
name old time/op new time/op delta
FixedAccumulateOpOverSIMD16-8 183ns ± 0% 174ns ± 0% -4.92% (p=0.000 n=8+8)
FixedAccumulateOpSrcSIMD16-8 87.0ns ± 1% 86.3ns ± 0% -0.77% (p=0.000 n=10+9)
FixedAccumulateMaskSIMD16-8 80.4ns ± 1% 81.2ns ± 1% +1.01% (p=0.000 n=10+10)
FloatingAccumulateOpOverSIMD16-8 250ns ± 1% 244ns ± 0% -2.39% (p=0.000 n=10+8)
FloatingAccumulateOpSrcSIMD16-8 176ns ± 1% 176ns ± 0% ~ (p=0.142 n=10+8)
FloatingAccumulateMaskSIMD16-8 167ns ± 0% 167ns ± 0% ~ (p=0.137 n=8+10)
FixedAccumulateOpOverSIMD64-8 2.73µs ± 1% 2.58µs ± 0% -5.36% (p=0.000 n=10+7)
FixedAccumulateOpSrcSIMD64-8 1.18µs ± 1% 1.17µs ± 0% -0.33% (p=0.003 n=10+9)
FixedAccumulateMaskSIMD64-8 1.09µs ± 0% 1.09µs ± 0% -0.17% (p=0.047 n=9+9)
FloatingAccumulateOpOverSIMD64-8 3.67µs ± 0% 3.61µs ± 1% -1.47% (p=0.000 n=7+10)
FloatingAccumulateOpSrcSIMD64-8 2.60µs ± 0% 2.61µs ± 0% +0.19% (p=0.003 n=8+8)
FloatingAccumulateMaskSIMD64-8 2.47µs ± 0% 2.46µs ± 0% ~ (p=0.162 n=10+9)
GlyphAlpha16Over-8 2.99µs ± 0% 2.98µs ± 1% -0.50% (p=0.021 n=9+10)
GlyphAlpha16Src-8 2.89µs ± 1% 2.89µs ± 0% ~ (p=0.381 n=10+10)
GlyphAlpha32Over-8 4.53µs ± 0% 4.50µs ± 0% -0.83% (p=0.000 n=10+10)
GlyphAlpha32Src-8 4.14µs ± 0% 4.13µs ± 0% -0.21% (p=0.026 n=9+10)
GlyphAlpha64Over-8 8.97µs ± 1% 8.80µs ± 0% -1.85% (p=0.000 n=10+9)
GlyphAlpha64Src-8 7.42µs ± 1% 7.39µs ± 0% -0.45% (p=0.011 n=10+10)
GlyphAlpha128Over-8 21.8µs ± 0% 21.2µs ± 0% -2.91% (p=0.000 n=9+9)
GlyphAlpha128Src-8 15.6µs ± 0% 15.6µs ± 0% ~ (p=0.982 n=10+7)
GlyphAlpha256Over-8 66.3µs ± 1% 63.7µs ± 0% -3.84% (p=0.000 n=10+9)
GlyphAlpha256Src-8 41.2µs ± 1% 41.2µs ± 1% ~ (p=1.000 n=10+10)
GlyphRGBA16Over-8 4.75µs ± 0% 4.75µs ± 1% ~ (p=0.735 n=9+10)
GlyphRGBA16Src-8 4.20µs ± 0% 4.20µs ± 0% ~ (p=0.503 n=8+8)
GlyphRGBA32Over-8 11.4µs ± 0% 11.4µs ± 0% ~ (p=0.119 n=9+9)
GlyphRGBA32Src-8 9.34µs ± 1% 9.32µs ± 0% ~ (p=0.062 n=9+8)
GlyphRGBA64Over-8 36.0µs ± 0% 36.1µs ± 0% ~ (p=0.209 n=8+9)
GlyphRGBA64Src-8 27.9µs ± 1% 27.8µs ± 0% ~ (p=0.796 n=10+10)
GlyphRGBA128Over-8 131µs ± 0% 131µs ± 0% ~ (p=0.931 n=9+9)
GlyphRGBA128Src-8 97.9µs ± 0% 97.7µs ± 1% ~ (p=0.053 n=9+10)
GlyphRGBA256Over-8 503µs ± 0% 503µs ± 1% ~ (p=0.274 n=8+10)
GlyphRGBA256Src-8 370µs ± 0% 369µs ± 0% ~ (p=0.497 n=9+10)
Change-Id: I56651e70b258792b83ea2a74904756243c88bef4
Reviewed-on: https://go-review.googlesource.com/31537
Reviewed-by: David Crawshaw <crawshaw@golang.org>
3 files changed