golang-image/vector/acc_amd64.s

1029 lines
23 KiB
ArmAsm
Raw Normal View History

// generated by go run gen.go; DO NOT EDIT
// +build !appengine
// +build gc
// +build !noasm
#include "textflag.h"
// fl is short for floating point math. fx is short for fixed point math.
DATA flAlmost65536<>+0x00(SB)/8, $0x477fffff477fffff
DATA flAlmost65536<>+0x08(SB)/8, $0x477fffff477fffff
DATA flOne<>+0x00(SB)/8, $0x3f8000003f800000
DATA flOne<>+0x08(SB)/8, $0x3f8000003f800000
DATA flSignMask<>+0x00(SB)/8, $0x7fffffff7fffffff
DATA flSignMask<>+0x08(SB)/8, $0x7fffffff7fffffff
// scatterAndMulBy0x101 is a PSHUFB mask that brings the low four bytes of an
// XMM register to the low byte of that register's four uint32 values. It
// duplicates those bytes, effectively multiplying each uint32 by 0x101.
//
// It transforms a little-endian 16-byte XMM value from
// ijkl????????????
// to
// ii00jj00kk00ll00
DATA scatterAndMulBy0x101<>+0x00(SB)/8, $0x8080010180800000
DATA scatterAndMulBy0x101<>+0x08(SB)/8, $0x8080030380800202
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>
2016-10-20 06:15:53 +02:00
// gather is a PSHUFB mask that brings the second-lowest byte of the XMM
// register's four uint32 values to the low four bytes of that register.
//
// It transforms a little-endian 16-byte XMM value from
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>
2016-10-20 06:15:53 +02:00
// ?i???j???k???l??
// to
// ijkl000000000000
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>
2016-10-20 06:15:53 +02:00
DATA gather<>+0x00(SB)/8, $0x808080800d090501
DATA gather<>+0x08(SB)/8, $0x8080808080808080
DATA fxAlmost65536<>+0x00(SB)/8, $0x0000ffff0000ffff
DATA fxAlmost65536<>+0x08(SB)/8, $0x0000ffff0000ffff
DATA inverseFFFF<>+0x00(SB)/8, $0x8000800180008001
DATA inverseFFFF<>+0x08(SB)/8, $0x8000800180008001
GLOBL flAlmost65536<>(SB), (NOPTR+RODATA), $16
GLOBL flOne<>(SB), (NOPTR+RODATA), $16
GLOBL flSignMask<>(SB), (NOPTR+RODATA), $16
GLOBL scatterAndMulBy0x101<>(SB), (NOPTR+RODATA), $16
GLOBL gather<>(SB), (NOPTR+RODATA), $16
GLOBL fxAlmost65536<>(SB), (NOPTR+RODATA), $16
GLOBL inverseFFFF<>(SB), (NOPTR+RODATA), $16
// func haveSSE4_1() bool
TEXT ·haveSSE4_1(SB), NOSPLIT, $0
MOVQ $1, AX
CPUID
SHRQ $19, CX
ANDQ $1, CX
MOVB CX, ret+0(FP)
RET
// ----------------------------------------------------------------------------
// func fixedAccumulateOpOverSIMD(dst []uint8, src []uint32)
//
// XMM registers. Variable names are per
// https://github.com/google/font-rs/blob/master/src/accumulate.c
//
// xmm0 scratch
// xmm1 x
// xmm2 y, z
// xmm3 -
// xmm4 -
// xmm5 fxAlmost65536
// xmm6 gather
// xmm7 offset
// xmm8 scatterAndMulBy0x101
// xmm9 fxAlmost65536
// xmm10 inverseFFFF
TEXT ·fixedAccumulateOpOverSIMD(SB), NOSPLIT, $0-48
vector: add SIMD versions of xxxAccumulateMask. name old time/op new time/op delta GlyphAlphaLoose16Over-8 3.96µs ± 0% 3.64µs ± 1% -8.08% (p=0.000 n=8+10) GlyphAlphaLoose16Src-8 3.64µs ± 0% 3.35µs ± 0% -7.88% (p=0.000 n=8+9) GlyphAlphaLoose32Over-8 8.45µs ± 0% 6.74µs ± 0% -20.22% (p=0.000 n=8+9) GlyphAlphaLoose32Src-8 7.24µs ± 0% 5.54µs ± 1% -23.48% (p=0.000 n=8+10) GlyphAlphaLoose64Over-8 22.2µs ± 0% 17.4µs ± 0% -21.67% (p=0.000 n=9+9) GlyphAlphaLoose64Src-8 17.6µs ± 1% 12.2µs ± 1% -30.32% (p=0.000 n=10+10) GlyphAlphaLoose128Over-8 67.9µs ± 0% 53.3µs ± 1% -21.53% (p=0.000 n=10+10) GlyphAlphaLoose128Src-8 48.2µs ± 0% 32.6µs ± 2% -32.41% (p=0.000 n=9+10) GlyphAlphaLoose256Over-8 242µs ± 1% 187µs ± 1% -22.96% (p=0.000 n=9+9) GlyphAlphaLoose256Src-8 163µs ± 0% 105µs ± 1% -35.83% (p=0.000 n=9+9) GlyphRGBA16Over-8 5.25µs ± 1% 4.95µs ± 0% -5.78% (p=0.000 n=9+9) GlyphRGBA16Src-8 4.72µs ± 0% 4.43µs ± 1% -6.22% (p=0.000 n=9+10) GlyphRGBA32Over-8 13.5µs ± 0% 11.9µs ± 1% -12.19% (p=0.000 n=9+10) GlyphRGBA32Src-8 11.5µs ± 1% 9.8µs ± 0% -14.72% (p=0.000 n=9+9) GlyphRGBA64Over-8 42.0µs ± 2% 36.9µs ± 1% -12.19% (p=0.000 n=10+10) GlyphRGBA64Src-8 34.1µs ± 1% 28.5µs ± 0% -16.25% (p=0.000 n=9+7) GlyphRGBA128Over-8 149µs ± 2% 133µs ± 1% -10.24% (p=0.000 n=10+9) GlyphRGBA128Src-8 115µs ± 1% 99µs ± 1% -13.57% (p=0.000 n=9+10) GlyphRGBA256Over-8 566µs ± 0% 511µs ± 1% -9.85% (p=0.000 n=9+10) GlyphRGBA256Src-8 435µs ± 0% 372µs ± 0% -14.64% (p=0.000 n=9+8) GlyphNRGBA16Over-8 26.9µs ± 3% 26.0µs ± 3% -3.55% (p=0.000 n=10+9) GlyphNRGBA16Src-8 18.8µs ± 2% 18.4µs ± 2% -2.21% (p=0.000 n=9+10) GlyphNRGBA32Over-8 99.1µs ± 2% 95.9µs ± 3% -3.23% (p=0.000 n=10+10) GlyphNRGBA32Src-8 65.6µs ± 3% 62.8µs ± 2% -4.36% (p=0.000 n=10+10) GlyphNRGBA64Over-8 376µs ± 4% 370µs ± 2% ~ (p=0.063 n=10+10) GlyphNRGBA64Src-8 238µs ± 3% 233µs ± 1% -2.21% (p=0.000 n=9+10) GlyphNRGBA128Over-8 1.52ms ± 2% 1.48ms ± 0% -2.11% (p=0.000 n=10+8) GlyphNRGBA128Src-8 951µs ± 3% 935µs ± 1% -1.69% (p=0.013 n=10+9) GlyphNRGBA256Over-8 6.00ms ± 1% 5.87ms ± 3% -2.12% (p=0.002 n=10+10) GlyphNRGBA256Src-8 3.94ms ± 2% 3.80ms ± 2% -3.64% (p=0.000 n=10+10) A comparison of the non-SIMD and SIMD versions: name time/op FixedAccumulateMask16-8 237ns ± 0% FixedAccumulateMaskSIMD16-8 80.0ns ± 1% FloatingAccumulateMask16-8 413ns ± 2% FloatingAccumulateMaskSIMD16-8 166ns ± 0% FixedAccumulateMask64-8 3.42µs ± 0% FixedAccumulateMaskSIMD64-8 1.09µs ± 0% FloatingAccumulateMask64-8 6.92µs ± 0% FloatingAccumulateMaskSIMD64-8 2.47µs ± 1% Change-Id: Ib6980e5975ed2842ff2a372f76dd5f2e95c5526c Reviewed-on: https://go-review.googlesource.com/30898 Reviewed-by: David Crawshaw <crawshaw@golang.org>
2016-10-13 05:50:52 +02:00
MOVQ dst_base+0(FP), DI
MOVQ dst_len+8(FP), BX
MOVQ src_base+24(FP), SI
MOVQ src_len+32(FP), R10
// Sanity check that len(dst) >= len(src).
CMPQ BX, R10
JLT fxAccOpOverEnd
// R10 = len(src) &^ 3
// R11 = len(src)
MOVQ R10, R11
ANDQ $-4, R10
// fxAlmost65536 := XMM(0x0000ffff repeated four times) // Maximum of an uint16.
MOVOU fxAlmost65536<>(SB), X5
vector: add SIMD versions of xxxAccumulateMask. name old time/op new time/op delta GlyphAlphaLoose16Over-8 3.96µs ± 0% 3.64µs ± 1% -8.08% (p=0.000 n=8+10) GlyphAlphaLoose16Src-8 3.64µs ± 0% 3.35µs ± 0% -7.88% (p=0.000 n=8+9) GlyphAlphaLoose32Over-8 8.45µs ± 0% 6.74µs ± 0% -20.22% (p=0.000 n=8+9) GlyphAlphaLoose32Src-8 7.24µs ± 0% 5.54µs ± 1% -23.48% (p=0.000 n=8+10) GlyphAlphaLoose64Over-8 22.2µs ± 0% 17.4µs ± 0% -21.67% (p=0.000 n=9+9) GlyphAlphaLoose64Src-8 17.6µs ± 1% 12.2µs ± 1% -30.32% (p=0.000 n=10+10) GlyphAlphaLoose128Over-8 67.9µs ± 0% 53.3µs ± 1% -21.53% (p=0.000 n=10+10) GlyphAlphaLoose128Src-8 48.2µs ± 0% 32.6µs ± 2% -32.41% (p=0.000 n=9+10) GlyphAlphaLoose256Over-8 242µs ± 1% 187µs ± 1% -22.96% (p=0.000 n=9+9) GlyphAlphaLoose256Src-8 163µs ± 0% 105µs ± 1% -35.83% (p=0.000 n=9+9) GlyphRGBA16Over-8 5.25µs ± 1% 4.95µs ± 0% -5.78% (p=0.000 n=9+9) GlyphRGBA16Src-8 4.72µs ± 0% 4.43µs ± 1% -6.22% (p=0.000 n=9+10) GlyphRGBA32Over-8 13.5µs ± 0% 11.9µs ± 1% -12.19% (p=0.000 n=9+10) GlyphRGBA32Src-8 11.5µs ± 1% 9.8µs ± 0% -14.72% (p=0.000 n=9+9) GlyphRGBA64Over-8 42.0µs ± 2% 36.9µs ± 1% -12.19% (p=0.000 n=10+10) GlyphRGBA64Src-8 34.1µs ± 1% 28.5µs ± 0% -16.25% (p=0.000 n=9+7) GlyphRGBA128Over-8 149µs ± 2% 133µs ± 1% -10.24% (p=0.000 n=10+9) GlyphRGBA128Src-8 115µs ± 1% 99µs ± 1% -13.57% (p=0.000 n=9+10) GlyphRGBA256Over-8 566µs ± 0% 511µs ± 1% -9.85% (p=0.000 n=9+10) GlyphRGBA256Src-8 435µs ± 0% 372µs ± 0% -14.64% (p=0.000 n=9+8) GlyphNRGBA16Over-8 26.9µs ± 3% 26.0µs ± 3% -3.55% (p=0.000 n=10+9) GlyphNRGBA16Src-8 18.8µs ± 2% 18.4µs ± 2% -2.21% (p=0.000 n=9+10) GlyphNRGBA32Over-8 99.1µs ± 2% 95.9µs ± 3% -3.23% (p=0.000 n=10+10) GlyphNRGBA32Src-8 65.6µs ± 3% 62.8µs ± 2% -4.36% (p=0.000 n=10+10) GlyphNRGBA64Over-8 376µs ± 4% 370µs ± 2% ~ (p=0.063 n=10+10) GlyphNRGBA64Src-8 238µs ± 3% 233µs ± 1% -2.21% (p=0.000 n=9+10) GlyphNRGBA128Over-8 1.52ms ± 2% 1.48ms ± 0% -2.11% (p=0.000 n=10+8) GlyphNRGBA128Src-8 951µs ± 3% 935µs ± 1% -1.69% (p=0.013 n=10+9) GlyphNRGBA256Over-8 6.00ms ± 1% 5.87ms ± 3% -2.12% (p=0.002 n=10+10) GlyphNRGBA256Src-8 3.94ms ± 2% 3.80ms ± 2% -3.64% (p=0.000 n=10+10) A comparison of the non-SIMD and SIMD versions: name time/op FixedAccumulateMask16-8 237ns ± 0% FixedAccumulateMaskSIMD16-8 80.0ns ± 1% FloatingAccumulateMask16-8 413ns ± 2% FloatingAccumulateMaskSIMD16-8 166ns ± 0% FixedAccumulateMask64-8 3.42µs ± 0% FixedAccumulateMaskSIMD64-8 1.09µs ± 0% FloatingAccumulateMask64-8 6.92µs ± 0% FloatingAccumulateMaskSIMD64-8 2.47µs ± 1% Change-Id: Ib6980e5975ed2842ff2a372f76dd5f2e95c5526c Reviewed-on: https://go-review.googlesource.com/30898 Reviewed-by: David Crawshaw <crawshaw@golang.org>
2016-10-13 05:50:52 +02:00
// gather := XMM(see above) // PSHUFB shuffle mask.
// scatterAndMulBy0x101 := XMM(see above) // PSHUFB shuffle mask.
// fxAlmost65536 := XMM(0x0000ffff repeated four times) // 0xffff.
// inverseFFFF := XMM(0x80008001 repeated four times) // Magic constant for dividing by 0xffff.
vector: add SIMD versions of xxxAccumulateMask. name old time/op new time/op delta GlyphAlphaLoose16Over-8 3.96µs ± 0% 3.64µs ± 1% -8.08% (p=0.000 n=8+10) GlyphAlphaLoose16Src-8 3.64µs ± 0% 3.35µs ± 0% -7.88% (p=0.000 n=8+9) GlyphAlphaLoose32Over-8 8.45µs ± 0% 6.74µs ± 0% -20.22% (p=0.000 n=8+9) GlyphAlphaLoose32Src-8 7.24µs ± 0% 5.54µs ± 1% -23.48% (p=0.000 n=8+10) GlyphAlphaLoose64Over-8 22.2µs ± 0% 17.4µs ± 0% -21.67% (p=0.000 n=9+9) GlyphAlphaLoose64Src-8 17.6µs ± 1% 12.2µs ± 1% -30.32% (p=0.000 n=10+10) GlyphAlphaLoose128Over-8 67.9µs ± 0% 53.3µs ± 1% -21.53% (p=0.000 n=10+10) GlyphAlphaLoose128Src-8 48.2µs ± 0% 32.6µs ± 2% -32.41% (p=0.000 n=9+10) GlyphAlphaLoose256Over-8 242µs ± 1% 187µs ± 1% -22.96% (p=0.000 n=9+9) GlyphAlphaLoose256Src-8 163µs ± 0% 105µs ± 1% -35.83% (p=0.000 n=9+9) GlyphRGBA16Over-8 5.25µs ± 1% 4.95µs ± 0% -5.78% (p=0.000 n=9+9) GlyphRGBA16Src-8 4.72µs ± 0% 4.43µs ± 1% -6.22% (p=0.000 n=9+10) GlyphRGBA32Over-8 13.5µs ± 0% 11.9µs ± 1% -12.19% (p=0.000 n=9+10) GlyphRGBA32Src-8 11.5µs ± 1% 9.8µs ± 0% -14.72% (p=0.000 n=9+9) GlyphRGBA64Over-8 42.0µs ± 2% 36.9µs ± 1% -12.19% (p=0.000 n=10+10) GlyphRGBA64Src-8 34.1µs ± 1% 28.5µs ± 0% -16.25% (p=0.000 n=9+7) GlyphRGBA128Over-8 149µs ± 2% 133µs ± 1% -10.24% (p=0.000 n=10+9) GlyphRGBA128Src-8 115µs ± 1% 99µs ± 1% -13.57% (p=0.000 n=9+10) GlyphRGBA256Over-8 566µs ± 0% 511µs ± 1% -9.85% (p=0.000 n=9+10) GlyphRGBA256Src-8 435µs ± 0% 372µs ± 0% -14.64% (p=0.000 n=9+8) GlyphNRGBA16Over-8 26.9µs ± 3% 26.0µs ± 3% -3.55% (p=0.000 n=10+9) GlyphNRGBA16Src-8 18.8µs ± 2% 18.4µs ± 2% -2.21% (p=0.000 n=9+10) GlyphNRGBA32Over-8 99.1µs ± 2% 95.9µs ± 3% -3.23% (p=0.000 n=10+10) GlyphNRGBA32Src-8 65.6µs ± 3% 62.8µs ± 2% -4.36% (p=0.000 n=10+10) GlyphNRGBA64Over-8 376µs ± 4% 370µs ± 2% ~ (p=0.063 n=10+10) GlyphNRGBA64Src-8 238µs ± 3% 233µs ± 1% -2.21% (p=0.000 n=9+10) GlyphNRGBA128Over-8 1.52ms ± 2% 1.48ms ± 0% -2.11% (p=0.000 n=10+8) GlyphNRGBA128Src-8 951µs ± 3% 935µs ± 1% -1.69% (p=0.013 n=10+9) GlyphNRGBA256Over-8 6.00ms ± 1% 5.87ms ± 3% -2.12% (p=0.002 n=10+10) GlyphNRGBA256Src-8 3.94ms ± 2% 3.80ms ± 2% -3.64% (p=0.000 n=10+10) A comparison of the non-SIMD and SIMD versions: name time/op FixedAccumulateMask16-8 237ns ± 0% FixedAccumulateMaskSIMD16-8 80.0ns ± 1% FloatingAccumulateMask16-8 413ns ± 2% FloatingAccumulateMaskSIMD16-8 166ns ± 0% FixedAccumulateMask64-8 3.42µs ± 0% FixedAccumulateMaskSIMD64-8 1.09µs ± 0% FloatingAccumulateMask64-8 6.92µs ± 0% FloatingAccumulateMaskSIMD64-8 2.47µs ± 1% Change-Id: Ib6980e5975ed2842ff2a372f76dd5f2e95c5526c Reviewed-on: https://go-review.googlesource.com/30898 Reviewed-by: David Crawshaw <crawshaw@golang.org>
2016-10-13 05:50:52 +02:00
MOVOU gather<>(SB), X6
MOVOU scatterAndMulBy0x101<>(SB), X8
MOVOU fxAlmost65536<>(SB), X9
MOVOU inverseFFFF<>(SB), X10
// offset := XMM(0x00000000 repeated four times) // Cumulative sum.
XORPS X7, X7
// i := 0
MOVQ $0, R9
fxAccOpOverLoop4:
// for i < (len(src) &^ 3)
CMPQ R9, R10
JAE fxAccOpOverLoop1
// x = XMM(s0, s1, s2, s3)
//
// Where s0 is src[i+0], s1 is src[i+1], etc.
MOVOU (SI), X1
// scratch = XMM(0, s0, s1, s2)
// x += scratch // yields x == XMM(s0, s0+s1, s1+s2, s2+s3)
MOVOU X1, X0
PSLLO $4, X0
PADDD X0, X1
// scratch = XMM(0, 0, 0, 0)
// scratch = XMM(scratch@0, scratch@0, x@0, x@1) // yields scratch == XMM(0, 0, s0, s0+s1)
// x += scratch // yields x == XMM(s0, s0+s1, s0+s1+s2, s0+s1+s2+s3)
XORPS X0, X0
SHUFPS $0x40, X1, X0
PADDD X0, X1
// x += offset
PADDD X7, X1
// y = abs(x)
vector: change ϕ from 10 to 9. This slight loss in quality allows us to use int32 math exclusively throughout raster_fixed.go, instead of occasionally dropping into int64 math. The change in ϕ doesn't affect the benchmarks noticably, but staying in int32 does. The net effect: name old time/op new time/op delta GlyphAlpha16Over-8 3.36µs ± 0% 2.99µs ± 0% -10.89% (p=0.000 n=10+9) GlyphAlpha16Src-8 3.26µs ± 0% 2.89µs ± 1% -11.34% (p=0.000 n=9+10) GlyphAlpha32Over-8 5.20µs ± 0% 4.53µs ± 0% -12.76% (p=0.000 n=8+10) GlyphAlpha32Src-8 4.81µs ± 1% 4.14µs ± 0% -13.91% (p=0.000 n=9+9) GlyphAlpha64Over-8 10.2µs ± 0% 9.0µs ± 1% -11.99% (p=0.000 n=9+10) GlyphAlpha64Src-8 8.62µs ± 0% 7.42µs ± 1% -13.89% (p=0.000 n=9+10) GlyphAlpha128Over-8 24.1µs ± 0% 21.8µs ± 0% -9.32% (p=0.000 n=9+9) GlyphAlpha128Src-8 17.9µs ± 0% 15.6µs ± 0% -12.68% (p=0.000 n=9+10) GlyphAlpha256Over-8 70.1µs ± 0% 66.3µs ± 1% -5.44% (p=0.000 n=10+10) GlyphAlpha256Src-8 45.2µs ± 1% 41.2µs ± 1% -8.92% (p=0.000 n=10+10) GlyphRGBA16Over-8 5.12µs ± 0% 4.75µs ± 0% -7.15% (p=0.000 n=10+9) GlyphRGBA16Src-8 4.57µs ± 1% 4.20µs ± 0% -8.18% (p=0.000 n=9+8) GlyphRGBA32Over-8 12.1µs ± 0% 11.4µs ± 0% -5.50% (p=0.000 n=10+9) GlyphRGBA32Src-8 10.0µs ± 0% 9.3µs ± 1% -6.80% (p=0.000 n=10+9) GlyphRGBA64Over-8 37.2µs ± 0% 36.0µs ± 0% -3.17% (p=0.000 n=9+8) GlyphRGBA64Src-8 29.0µs ± 1% 27.9µs ± 1% -4.05% (p=0.000 n=9+10) GlyphRGBA128Over-8 134µs ± 1% 131µs ± 0% -1.85% (p=0.000 n=9+9) GlyphRGBA128Src-8 100µs ± 1% 98µs ± 0% -2.27% (p=0.000 n=10+9) GlyphRGBA256Over-8 506µs ± 0% 503µs ± 0% -0.56% (p=0.000 n=10+8) GlyphRGBA256Src-8 373µs ± 0% 370µs ± 0% -1.01% (p=0.000 n=10+9) Change-Id: Ie02afac6fd6fa95f090bf3fe0a5c744799ea9dc5 Reviewed-on: https://go-review.googlesource.com/31532 Reviewed-by: David Crawshaw <crawshaw@golang.org>
2016-10-20 01:34:23 +02:00
// y >>= 2 // Shift by 2*ϕ - 16.
// y = min(y, fxAlmost65536)
PABSD X1, X2
PSRLL $2, X2
PMINUD X5, X2
// z = convertToInt32(y)
// No-op.
// Blend over the dst's prior value. SIMD for i in 0..3:
//
// dstA := uint32(dst[i]) * 0x101
// maskA := z@i
// outA := dstA*(0xffff-maskA)/0xffff + maskA
// dst[i] = uint8(outA >> 8)
//
// First, set X0 to dstA*(0xfff-maskA).
MOVL (DI), X0
PSHUFB X8, X0
MOVOU X9, X11
PSUBL X2, X11
PMULLD X11, X0
// We implement uint32 division by 0xffff as multiplication by a magic
// constant (0x800080001) and then a shift by a magic constant (47).
// See TestDivideByFFFF for a justification.
//
// That multiplication widens from uint32 to uint64, so we have to
// duplicate and shift our four uint32s from one XMM register (X0) to
// two XMM registers (X0 and X11).
//
// Move the second and fourth uint32s in X0 to be the first and third
// uint32s in X11.
MOVOU X0, X11
PSRLQ $32, X11
// Multiply by magic, shift by magic.
PMULULQ X10, X0
PMULULQ X10, X11
PSRLQ $47, X0
PSRLQ $47, X11
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>
2016-10-20 06:15:53 +02:00
// Merge the two registers back to one, X11, and add maskA.
PSLLQ $32, X11
XORPS X0, X11
PADDD X11, X2
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>
2016-10-20 06:15:53 +02:00
// As per opSrcStore4, shuffle and copy the 4 second-lowest bytes.
PSHUFB X6, X2
MOVL X2, (DI)
// offset = XMM(x@3, x@3, x@3, x@3)
MOVOU X1, X7
SHUFPS $0xff, X1, X7
// i += 4
// dst = dst[4:]
// src = src[4:]
ADDQ $4, R9
ADDQ $4, DI
ADDQ $16, SI
JMP fxAccOpOverLoop4
fxAccOpOverLoop1:
// for i < len(src)
CMPQ R9, R11
vector: add SIMD versions of xxxAccumulateMask. name old time/op new time/op delta GlyphAlphaLoose16Over-8 3.96µs ± 0% 3.64µs ± 1% -8.08% (p=0.000 n=8+10) GlyphAlphaLoose16Src-8 3.64µs ± 0% 3.35µs ± 0% -7.88% (p=0.000 n=8+9) GlyphAlphaLoose32Over-8 8.45µs ± 0% 6.74µs ± 0% -20.22% (p=0.000 n=8+9) GlyphAlphaLoose32Src-8 7.24µs ± 0% 5.54µs ± 1% -23.48% (p=0.000 n=8+10) GlyphAlphaLoose64Over-8 22.2µs ± 0% 17.4µs ± 0% -21.67% (p=0.000 n=9+9) GlyphAlphaLoose64Src-8 17.6µs ± 1% 12.2µs ± 1% -30.32% (p=0.000 n=10+10) GlyphAlphaLoose128Over-8 67.9µs ± 0% 53.3µs ± 1% -21.53% (p=0.000 n=10+10) GlyphAlphaLoose128Src-8 48.2µs ± 0% 32.6µs ± 2% -32.41% (p=0.000 n=9+10) GlyphAlphaLoose256Over-8 242µs ± 1% 187µs ± 1% -22.96% (p=0.000 n=9+9) GlyphAlphaLoose256Src-8 163µs ± 0% 105µs ± 1% -35.83% (p=0.000 n=9+9) GlyphRGBA16Over-8 5.25µs ± 1% 4.95µs ± 0% -5.78% (p=0.000 n=9+9) GlyphRGBA16Src-8 4.72µs ± 0% 4.43µs ± 1% -6.22% (p=0.000 n=9+10) GlyphRGBA32Over-8 13.5µs ± 0% 11.9µs ± 1% -12.19% (p=0.000 n=9+10) GlyphRGBA32Src-8 11.5µs ± 1% 9.8µs ± 0% -14.72% (p=0.000 n=9+9) GlyphRGBA64Over-8 42.0µs ± 2% 36.9µs ± 1% -12.19% (p=0.000 n=10+10) GlyphRGBA64Src-8 34.1µs ± 1% 28.5µs ± 0% -16.25% (p=0.000 n=9+7) GlyphRGBA128Over-8 149µs ± 2% 133µs ± 1% -10.24% (p=0.000 n=10+9) GlyphRGBA128Src-8 115µs ± 1% 99µs ± 1% -13.57% (p=0.000 n=9+10) GlyphRGBA256Over-8 566µs ± 0% 511µs ± 1% -9.85% (p=0.000 n=9+10) GlyphRGBA256Src-8 435µs ± 0% 372µs ± 0% -14.64% (p=0.000 n=9+8) GlyphNRGBA16Over-8 26.9µs ± 3% 26.0µs ± 3% -3.55% (p=0.000 n=10+9) GlyphNRGBA16Src-8 18.8µs ± 2% 18.4µs ± 2% -2.21% (p=0.000 n=9+10) GlyphNRGBA32Over-8 99.1µs ± 2% 95.9µs ± 3% -3.23% (p=0.000 n=10+10) GlyphNRGBA32Src-8 65.6µs ± 3% 62.8µs ± 2% -4.36% (p=0.000 n=10+10) GlyphNRGBA64Over-8 376µs ± 4% 370µs ± 2% ~ (p=0.063 n=10+10) GlyphNRGBA64Src-8 238µs ± 3% 233µs ± 1% -2.21% (p=0.000 n=9+10) GlyphNRGBA128Over-8 1.52ms ± 2% 1.48ms ± 0% -2.11% (p=0.000 n=10+8) GlyphNRGBA128Src-8 951µs ± 3% 935µs ± 1% -1.69% (p=0.013 n=10+9) GlyphNRGBA256Over-8 6.00ms ± 1% 5.87ms ± 3% -2.12% (p=0.002 n=10+10) GlyphNRGBA256Src-8 3.94ms ± 2% 3.80ms ± 2% -3.64% (p=0.000 n=10+10) A comparison of the non-SIMD and SIMD versions: name time/op FixedAccumulateMask16-8 237ns ± 0% FixedAccumulateMaskSIMD16-8 80.0ns ± 1% FloatingAccumulateMask16-8 413ns ± 2% FloatingAccumulateMaskSIMD16-8 166ns ± 0% FixedAccumulateMask64-8 3.42µs ± 0% FixedAccumulateMaskSIMD64-8 1.09µs ± 0% FloatingAccumulateMask64-8 6.92µs ± 0% FloatingAccumulateMaskSIMD64-8 2.47µs ± 1% Change-Id: Ib6980e5975ed2842ff2a372f76dd5f2e95c5526c Reviewed-on: https://go-review.googlesource.com/30898 Reviewed-by: David Crawshaw <crawshaw@golang.org>
2016-10-13 05:50:52 +02:00
JAE fxAccOpOverEnd
// x = src[i] + offset
MOVL (SI), X1
PADDD X7, X1
// y = abs(x)
vector: change ϕ from 10 to 9. This slight loss in quality allows us to use int32 math exclusively throughout raster_fixed.go, instead of occasionally dropping into int64 math. The change in ϕ doesn't affect the benchmarks noticably, but staying in int32 does. The net effect: name old time/op new time/op delta GlyphAlpha16Over-8 3.36µs ± 0% 2.99µs ± 0% -10.89% (p=0.000 n=10+9) GlyphAlpha16Src-8 3.26µs ± 0% 2.89µs ± 1% -11.34% (p=0.000 n=9+10) GlyphAlpha32Over-8 5.20µs ± 0% 4.53µs ± 0% -12.76% (p=0.000 n=8+10) GlyphAlpha32Src-8 4.81µs ± 1% 4.14µs ± 0% -13.91% (p=0.000 n=9+9) GlyphAlpha64Over-8 10.2µs ± 0% 9.0µs ± 1% -11.99% (p=0.000 n=9+10) GlyphAlpha64Src-8 8.62µs ± 0% 7.42µs ± 1% -13.89% (p=0.000 n=9+10) GlyphAlpha128Over-8 24.1µs ± 0% 21.8µs ± 0% -9.32% (p=0.000 n=9+9) GlyphAlpha128Src-8 17.9µs ± 0% 15.6µs ± 0% -12.68% (p=0.000 n=9+10) GlyphAlpha256Over-8 70.1µs ± 0% 66.3µs ± 1% -5.44% (p=0.000 n=10+10) GlyphAlpha256Src-8 45.2µs ± 1% 41.2µs ± 1% -8.92% (p=0.000 n=10+10) GlyphRGBA16Over-8 5.12µs ± 0% 4.75µs ± 0% -7.15% (p=0.000 n=10+9) GlyphRGBA16Src-8 4.57µs ± 1% 4.20µs ± 0% -8.18% (p=0.000 n=9+8) GlyphRGBA32Over-8 12.1µs ± 0% 11.4µs ± 0% -5.50% (p=0.000 n=10+9) GlyphRGBA32Src-8 10.0µs ± 0% 9.3µs ± 1% -6.80% (p=0.000 n=10+9) GlyphRGBA64Over-8 37.2µs ± 0% 36.0µs ± 0% -3.17% (p=0.000 n=9+8) GlyphRGBA64Src-8 29.0µs ± 1% 27.9µs ± 1% -4.05% (p=0.000 n=9+10) GlyphRGBA128Over-8 134µs ± 1% 131µs ± 0% -1.85% (p=0.000 n=9+9) GlyphRGBA128Src-8 100µs ± 1% 98µs ± 0% -2.27% (p=0.000 n=10+9) GlyphRGBA256Over-8 506µs ± 0% 503µs ± 0% -0.56% (p=0.000 n=10+8) GlyphRGBA256Src-8 373µs ± 0% 370µs ± 0% -1.01% (p=0.000 n=10+9) Change-Id: Ie02afac6fd6fa95f090bf3fe0a5c744799ea9dc5 Reviewed-on: https://go-review.googlesource.com/31532 Reviewed-by: David Crawshaw <crawshaw@golang.org>
2016-10-20 01:34:23 +02:00
// y >>= 2 // Shift by 2*ϕ - 16.
// y = min(y, fxAlmost65536)
PABSD X1, X2
PSRLL $2, X2
PMINUD X5, X2
// z = convertToInt32(y)
// No-op.
// Blend over the dst's prior value.
//
// dstA := uint32(dst[0]) * 0x101
// maskA := z
// outA := dstA*(0xffff-maskA)/0xffff + maskA
// dst[0] = uint8(outA >> 8)
MOVBLZX (DI), R12
IMULL $0x101, R12
MOVL X2, R13
MOVL $0xffff, AX
SUBL R13, AX
MULL R12 // MULL's implicit arg is AX, and the result is stored in DX:AX.
MOVL $0x80008001, BX // Divide by 0xffff is to first multiply by a magic constant...
MULL BX // MULL's implicit arg is AX, and the result is stored in DX:AX.
SHRL $15, DX // ...and then shift by another magic constant (47 - 32 = 15).
ADDL DX, R13
SHRL $8, R13
MOVB R13, (DI)
// offset = x
MOVOU X1, X7
// i += 1
// dst = dst[1:]
// src = src[1:]
ADDQ $1, R9
ADDQ $1, DI
ADDQ $4, SI
JMP fxAccOpOverLoop1
fxAccOpOverEnd:
RET
// ----------------------------------------------------------------------------
// func fixedAccumulateOpSrcSIMD(dst []uint8, src []uint32)
//
// XMM registers. Variable names are per
// https://github.com/google/font-rs/blob/master/src/accumulate.c
//
// xmm0 scratch
// xmm1 x
// xmm2 y, z
// xmm3 -
// xmm4 -
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>
2016-10-20 06:15:53 +02:00
// xmm5 fxAlmost65536
// xmm6 gather
// xmm7 offset
// xmm8 -
// xmm9 -
// xmm10 -
TEXT ·fixedAccumulateOpSrcSIMD(SB), NOSPLIT, $0-48
vector: add SIMD versions of xxxAccumulateMask. name old time/op new time/op delta GlyphAlphaLoose16Over-8 3.96µs ± 0% 3.64µs ± 1% -8.08% (p=0.000 n=8+10) GlyphAlphaLoose16Src-8 3.64µs ± 0% 3.35µs ± 0% -7.88% (p=0.000 n=8+9) GlyphAlphaLoose32Over-8 8.45µs ± 0% 6.74µs ± 0% -20.22% (p=0.000 n=8+9) GlyphAlphaLoose32Src-8 7.24µs ± 0% 5.54µs ± 1% -23.48% (p=0.000 n=8+10) GlyphAlphaLoose64Over-8 22.2µs ± 0% 17.4µs ± 0% -21.67% (p=0.000 n=9+9) GlyphAlphaLoose64Src-8 17.6µs ± 1% 12.2µs ± 1% -30.32% (p=0.000 n=10+10) GlyphAlphaLoose128Over-8 67.9µs ± 0% 53.3µs ± 1% -21.53% (p=0.000 n=10+10) GlyphAlphaLoose128Src-8 48.2µs ± 0% 32.6µs ± 2% -32.41% (p=0.000 n=9+10) GlyphAlphaLoose256Over-8 242µs ± 1% 187µs ± 1% -22.96% (p=0.000 n=9+9) GlyphAlphaLoose256Src-8 163µs ± 0% 105µs ± 1% -35.83% (p=0.000 n=9+9) GlyphRGBA16Over-8 5.25µs ± 1% 4.95µs ± 0% -5.78% (p=0.000 n=9+9) GlyphRGBA16Src-8 4.72µs ± 0% 4.43µs ± 1% -6.22% (p=0.000 n=9+10) GlyphRGBA32Over-8 13.5µs ± 0% 11.9µs ± 1% -12.19% (p=0.000 n=9+10) GlyphRGBA32Src-8 11.5µs ± 1% 9.8µs ± 0% -14.72% (p=0.000 n=9+9) GlyphRGBA64Over-8 42.0µs ± 2% 36.9µs ± 1% -12.19% (p=0.000 n=10+10) GlyphRGBA64Src-8 34.1µs ± 1% 28.5µs ± 0% -16.25% (p=0.000 n=9+7) GlyphRGBA128Over-8 149µs ± 2% 133µs ± 1% -10.24% (p=0.000 n=10+9) GlyphRGBA128Src-8 115µs ± 1% 99µs ± 1% -13.57% (p=0.000 n=9+10) GlyphRGBA256Over-8 566µs ± 0% 511µs ± 1% -9.85% (p=0.000 n=9+10) GlyphRGBA256Src-8 435µs ± 0% 372µs ± 0% -14.64% (p=0.000 n=9+8) GlyphNRGBA16Over-8 26.9µs ± 3% 26.0µs ± 3% -3.55% (p=0.000 n=10+9) GlyphNRGBA16Src-8 18.8µs ± 2% 18.4µs ± 2% -2.21% (p=0.000 n=9+10) GlyphNRGBA32Over-8 99.1µs ± 2% 95.9µs ± 3% -3.23% (p=0.000 n=10+10) GlyphNRGBA32Src-8 65.6µs ± 3% 62.8µs ± 2% -4.36% (p=0.000 n=10+10) GlyphNRGBA64Over-8 376µs ± 4% 370µs ± 2% ~ (p=0.063 n=10+10) GlyphNRGBA64Src-8 238µs ± 3% 233µs ± 1% -2.21% (p=0.000 n=9+10) GlyphNRGBA128Over-8 1.52ms ± 2% 1.48ms ± 0% -2.11% (p=0.000 n=10+8) GlyphNRGBA128Src-8 951µs ± 3% 935µs ± 1% -1.69% (p=0.013 n=10+9) GlyphNRGBA256Over-8 6.00ms ± 1% 5.87ms ± 3% -2.12% (p=0.002 n=10+10) GlyphNRGBA256Src-8 3.94ms ± 2% 3.80ms ± 2% -3.64% (p=0.000 n=10+10) A comparison of the non-SIMD and SIMD versions: name time/op FixedAccumulateMask16-8 237ns ± 0% FixedAccumulateMaskSIMD16-8 80.0ns ± 1% FloatingAccumulateMask16-8 413ns ± 2% FloatingAccumulateMaskSIMD16-8 166ns ± 0% FixedAccumulateMask64-8 3.42µs ± 0% FixedAccumulateMaskSIMD64-8 1.09µs ± 0% FloatingAccumulateMask64-8 6.92µs ± 0% FloatingAccumulateMaskSIMD64-8 2.47µs ± 1% Change-Id: Ib6980e5975ed2842ff2a372f76dd5f2e95c5526c Reviewed-on: https://go-review.googlesource.com/30898 Reviewed-by: David Crawshaw <crawshaw@golang.org>
2016-10-13 05:50:52 +02:00
MOVQ dst_base+0(FP), DI
MOVQ dst_len+8(FP), BX
MOVQ src_base+24(FP), SI
MOVQ src_len+32(FP), R10
// Sanity check that len(dst) >= len(src).
CMPQ BX, R10
JLT fxAccOpSrcEnd
// R10 = len(src) &^ 3
// R11 = len(src)
MOVQ R10, R11
ANDQ $-4, R10
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>
2016-10-20 06:15:53 +02:00
// fxAlmost65536 := XMM(0x0000ffff repeated four times) // Maximum of an uint16.
MOVOU fxAlmost65536<>(SB), X5
vector: add SIMD versions of xxxAccumulateMask. name old time/op new time/op delta GlyphAlphaLoose16Over-8 3.96µs ± 0% 3.64µs ± 1% -8.08% (p=0.000 n=8+10) GlyphAlphaLoose16Src-8 3.64µs ± 0% 3.35µs ± 0% -7.88% (p=0.000 n=8+9) GlyphAlphaLoose32Over-8 8.45µs ± 0% 6.74µs ± 0% -20.22% (p=0.000 n=8+9) GlyphAlphaLoose32Src-8 7.24µs ± 0% 5.54µs ± 1% -23.48% (p=0.000 n=8+10) GlyphAlphaLoose64Over-8 22.2µs ± 0% 17.4µs ± 0% -21.67% (p=0.000 n=9+9) GlyphAlphaLoose64Src-8 17.6µs ± 1% 12.2µs ± 1% -30.32% (p=0.000 n=10+10) GlyphAlphaLoose128Over-8 67.9µs ± 0% 53.3µs ± 1% -21.53% (p=0.000 n=10+10) GlyphAlphaLoose128Src-8 48.2µs ± 0% 32.6µs ± 2% -32.41% (p=0.000 n=9+10) GlyphAlphaLoose256Over-8 242µs ± 1% 187µs ± 1% -22.96% (p=0.000 n=9+9) GlyphAlphaLoose256Src-8 163µs ± 0% 105µs ± 1% -35.83% (p=0.000 n=9+9) GlyphRGBA16Over-8 5.25µs ± 1% 4.95µs ± 0% -5.78% (p=0.000 n=9+9) GlyphRGBA16Src-8 4.72µs ± 0% 4.43µs ± 1% -6.22% (p=0.000 n=9+10) GlyphRGBA32Over-8 13.5µs ± 0% 11.9µs ± 1% -12.19% (p=0.000 n=9+10) GlyphRGBA32Src-8 11.5µs ± 1% 9.8µs ± 0% -14.72% (p=0.000 n=9+9) GlyphRGBA64Over-8 42.0µs ± 2% 36.9µs ± 1% -12.19% (p=0.000 n=10+10) GlyphRGBA64Src-8 34.1µs ± 1% 28.5µs ± 0% -16.25% (p=0.000 n=9+7) GlyphRGBA128Over-8 149µs ± 2% 133µs ± 1% -10.24% (p=0.000 n=10+9) GlyphRGBA128Src-8 115µs ± 1% 99µs ± 1% -13.57% (p=0.000 n=9+10) GlyphRGBA256Over-8 566µs ± 0% 511µs ± 1% -9.85% (p=0.000 n=9+10) GlyphRGBA256Src-8 435µs ± 0% 372µs ± 0% -14.64% (p=0.000 n=9+8) GlyphNRGBA16Over-8 26.9µs ± 3% 26.0µs ± 3% -3.55% (p=0.000 n=10+9) GlyphNRGBA16Src-8 18.8µs ± 2% 18.4µs ± 2% -2.21% (p=0.000 n=9+10) GlyphNRGBA32Over-8 99.1µs ± 2% 95.9µs ± 3% -3.23% (p=0.000 n=10+10) GlyphNRGBA32Src-8 65.6µs ± 3% 62.8µs ± 2% -4.36% (p=0.000 n=10+10) GlyphNRGBA64Over-8 376µs ± 4% 370µs ± 2% ~ (p=0.063 n=10+10) GlyphNRGBA64Src-8 238µs ± 3% 233µs ± 1% -2.21% (p=0.000 n=9+10) GlyphNRGBA128Over-8 1.52ms ± 2% 1.48ms ± 0% -2.11% (p=0.000 n=10+8) GlyphNRGBA128Src-8 951µs ± 3% 935µs ± 1% -1.69% (p=0.013 n=10+9) GlyphNRGBA256Over-8 6.00ms ± 1% 5.87ms ± 3% -2.12% (p=0.002 n=10+10) GlyphNRGBA256Src-8 3.94ms ± 2% 3.80ms ± 2% -3.64% (p=0.000 n=10+10) A comparison of the non-SIMD and SIMD versions: name time/op FixedAccumulateMask16-8 237ns ± 0% FixedAccumulateMaskSIMD16-8 80.0ns ± 1% FloatingAccumulateMask16-8 413ns ± 2% FloatingAccumulateMaskSIMD16-8 166ns ± 0% FixedAccumulateMask64-8 3.42µs ± 0% FixedAccumulateMaskSIMD64-8 1.09µs ± 0% FloatingAccumulateMask64-8 6.92µs ± 0% FloatingAccumulateMaskSIMD64-8 2.47µs ± 1% Change-Id: Ib6980e5975ed2842ff2a372f76dd5f2e95c5526c Reviewed-on: https://go-review.googlesource.com/30898 Reviewed-by: David Crawshaw <crawshaw@golang.org>
2016-10-13 05:50:52 +02:00
// gather := XMM(see above) // PSHUFB shuffle mask.
MOVOU gather<>(SB), X6
vector: add SIMD versions of xxxAccumulateMask. name old time/op new time/op delta GlyphAlphaLoose16Over-8 3.96µs ± 0% 3.64µs ± 1% -8.08% (p=0.000 n=8+10) GlyphAlphaLoose16Src-8 3.64µs ± 0% 3.35µs ± 0% -7.88% (p=0.000 n=8+9) GlyphAlphaLoose32Over-8 8.45µs ± 0% 6.74µs ± 0% -20.22% (p=0.000 n=8+9) GlyphAlphaLoose32Src-8 7.24µs ± 0% 5.54µs ± 1% -23.48% (p=0.000 n=8+10) GlyphAlphaLoose64Over-8 22.2µs ± 0% 17.4µs ± 0% -21.67% (p=0.000 n=9+9) GlyphAlphaLoose64Src-8 17.6µs ± 1% 12.2µs ± 1% -30.32% (p=0.000 n=10+10) GlyphAlphaLoose128Over-8 67.9µs ± 0% 53.3µs ± 1% -21.53% (p=0.000 n=10+10) GlyphAlphaLoose128Src-8 48.2µs ± 0% 32.6µs ± 2% -32.41% (p=0.000 n=9+10) GlyphAlphaLoose256Over-8 242µs ± 1% 187µs ± 1% -22.96% (p=0.000 n=9+9) GlyphAlphaLoose256Src-8 163µs ± 0% 105µs ± 1% -35.83% (p=0.000 n=9+9) GlyphRGBA16Over-8 5.25µs ± 1% 4.95µs ± 0% -5.78% (p=0.000 n=9+9) GlyphRGBA16Src-8 4.72µs ± 0% 4.43µs ± 1% -6.22% (p=0.000 n=9+10) GlyphRGBA32Over-8 13.5µs ± 0% 11.9µs ± 1% -12.19% (p=0.000 n=9+10) GlyphRGBA32Src-8 11.5µs ± 1% 9.8µs ± 0% -14.72% (p=0.000 n=9+9) GlyphRGBA64Over-8 42.0µs ± 2% 36.9µs ± 1% -12.19% (p=0.000 n=10+10) GlyphRGBA64Src-8 34.1µs ± 1% 28.5µs ± 0% -16.25% (p=0.000 n=9+7) GlyphRGBA128Over-8 149µs ± 2% 133µs ± 1% -10.24% (p=0.000 n=10+9) GlyphRGBA128Src-8 115µs ± 1% 99µs ± 1% -13.57% (p=0.000 n=9+10) GlyphRGBA256Over-8 566µs ± 0% 511µs ± 1% -9.85% (p=0.000 n=9+10) GlyphRGBA256Src-8 435µs ± 0% 372µs ± 0% -14.64% (p=0.000 n=9+8) GlyphNRGBA16Over-8 26.9µs ± 3% 26.0µs ± 3% -3.55% (p=0.000 n=10+9) GlyphNRGBA16Src-8 18.8µs ± 2% 18.4µs ± 2% -2.21% (p=0.000 n=9+10) GlyphNRGBA32Over-8 99.1µs ± 2% 95.9µs ± 3% -3.23% (p=0.000 n=10+10) GlyphNRGBA32Src-8 65.6µs ± 3% 62.8µs ± 2% -4.36% (p=0.000 n=10+10) GlyphNRGBA64Over-8 376µs ± 4% 370µs ± 2% ~ (p=0.063 n=10+10) GlyphNRGBA64Src-8 238µs ± 3% 233µs ± 1% -2.21% (p=0.000 n=9+10) GlyphNRGBA128Over-8 1.52ms ± 2% 1.48ms ± 0% -2.11% (p=0.000 n=10+8) GlyphNRGBA128Src-8 951µs ± 3% 935µs ± 1% -1.69% (p=0.013 n=10+9) GlyphNRGBA256Over-8 6.00ms ± 1% 5.87ms ± 3% -2.12% (p=0.002 n=10+10) GlyphNRGBA256Src-8 3.94ms ± 2% 3.80ms ± 2% -3.64% (p=0.000 n=10+10) A comparison of the non-SIMD and SIMD versions: name time/op FixedAccumulateMask16-8 237ns ± 0% FixedAccumulateMaskSIMD16-8 80.0ns ± 1% FloatingAccumulateMask16-8 413ns ± 2% FloatingAccumulateMaskSIMD16-8 166ns ± 0% FixedAccumulateMask64-8 3.42µs ± 0% FixedAccumulateMaskSIMD64-8 1.09µs ± 0% FloatingAccumulateMask64-8 6.92µs ± 0% FloatingAccumulateMaskSIMD64-8 2.47µs ± 1% Change-Id: Ib6980e5975ed2842ff2a372f76dd5f2e95c5526c Reviewed-on: https://go-review.googlesource.com/30898 Reviewed-by: David Crawshaw <crawshaw@golang.org>
2016-10-13 05:50:52 +02:00
// offset := XMM(0x00000000 repeated four times) // Cumulative sum.
XORPS X7, X7
// i := 0
MOVQ $0, R9
fxAccOpSrcLoop4:
// for i < (len(src) &^ 3)
CMPQ R9, R10
JAE fxAccOpSrcLoop1
// x = XMM(s0, s1, s2, s3)
//
// Where s0 is src[i+0], s1 is src[i+1], etc.
MOVOU (SI), X1
// scratch = XMM(0, s0, s1, s2)
// x += scratch // yields x == XMM(s0, s0+s1, s1+s2, s2+s3)
MOVOU X1, X0
PSLLO $4, X0
PADDD X0, X1
// scratch = XMM(0, 0, 0, 0)
// scratch = XMM(scratch@0, scratch@0, x@0, x@1) // yields scratch == XMM(0, 0, s0, s0+s1)
// x += scratch // yields x == XMM(s0, s0+s1, s0+s1+s2, s0+s1+s2+s3)
XORPS X0, X0
SHUFPS $0x40, X1, X0
PADDD X0, X1
// x += offset
PADDD X7, X1
// y = abs(x)
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>
2016-10-20 06:15:53 +02:00
// y >>= 2 // Shift by 2*ϕ - 16.
// y = min(y, fxAlmost65536)
PABSD X1, X2
PSRLL $2, X2
PMINUD X5, X2
// z = convertToInt32(y)
// No-op.
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>
2016-10-20 06:15:53 +02:00
// z = shuffleTheSecondLowestBytesOfEach4ByteElement(z)
// copy(dst[:4], low4BytesOf(z))
PSHUFB X6, X2
MOVL X2, (DI)
// offset = XMM(x@3, x@3, x@3, x@3)
MOVOU X1, X7
SHUFPS $0xff, X1, X7
// i += 4
// dst = dst[4:]
// src = src[4:]
ADDQ $4, R9
ADDQ $4, DI
ADDQ $16, SI
JMP fxAccOpSrcLoop4
fxAccOpSrcLoop1:
// for i < len(src)
CMPQ R9, R11
vector: add SIMD versions of xxxAccumulateMask. name old time/op new time/op delta GlyphAlphaLoose16Over-8 3.96µs ± 0% 3.64µs ± 1% -8.08% (p=0.000 n=8+10) GlyphAlphaLoose16Src-8 3.64µs ± 0% 3.35µs ± 0% -7.88% (p=0.000 n=8+9) GlyphAlphaLoose32Over-8 8.45µs ± 0% 6.74µs ± 0% -20.22% (p=0.000 n=8+9) GlyphAlphaLoose32Src-8 7.24µs ± 0% 5.54µs ± 1% -23.48% (p=0.000 n=8+10) GlyphAlphaLoose64Over-8 22.2µs ± 0% 17.4µs ± 0% -21.67% (p=0.000 n=9+9) GlyphAlphaLoose64Src-8 17.6µs ± 1% 12.2µs ± 1% -30.32% (p=0.000 n=10+10) GlyphAlphaLoose128Over-8 67.9µs ± 0% 53.3µs ± 1% -21.53% (p=0.000 n=10+10) GlyphAlphaLoose128Src-8 48.2µs ± 0% 32.6µs ± 2% -32.41% (p=0.000 n=9+10) GlyphAlphaLoose256Over-8 242µs ± 1% 187µs ± 1% -22.96% (p=0.000 n=9+9) GlyphAlphaLoose256Src-8 163µs ± 0% 105µs ± 1% -35.83% (p=0.000 n=9+9) GlyphRGBA16Over-8 5.25µs ± 1% 4.95µs ± 0% -5.78% (p=0.000 n=9+9) GlyphRGBA16Src-8 4.72µs ± 0% 4.43µs ± 1% -6.22% (p=0.000 n=9+10) GlyphRGBA32Over-8 13.5µs ± 0% 11.9µs ± 1% -12.19% (p=0.000 n=9+10) GlyphRGBA32Src-8 11.5µs ± 1% 9.8µs ± 0% -14.72% (p=0.000 n=9+9) GlyphRGBA64Over-8 42.0µs ± 2% 36.9µs ± 1% -12.19% (p=0.000 n=10+10) GlyphRGBA64Src-8 34.1µs ± 1% 28.5µs ± 0% -16.25% (p=0.000 n=9+7) GlyphRGBA128Over-8 149µs ± 2% 133µs ± 1% -10.24% (p=0.000 n=10+9) GlyphRGBA128Src-8 115µs ± 1% 99µs ± 1% -13.57% (p=0.000 n=9+10) GlyphRGBA256Over-8 566µs ± 0% 511µs ± 1% -9.85% (p=0.000 n=9+10) GlyphRGBA256Src-8 435µs ± 0% 372µs ± 0% -14.64% (p=0.000 n=9+8) GlyphNRGBA16Over-8 26.9µs ± 3% 26.0µs ± 3% -3.55% (p=0.000 n=10+9) GlyphNRGBA16Src-8 18.8µs ± 2% 18.4µs ± 2% -2.21% (p=0.000 n=9+10) GlyphNRGBA32Over-8 99.1µs ± 2% 95.9µs ± 3% -3.23% (p=0.000 n=10+10) GlyphNRGBA32Src-8 65.6µs ± 3% 62.8µs ± 2% -4.36% (p=0.000 n=10+10) GlyphNRGBA64Over-8 376µs ± 4% 370µs ± 2% ~ (p=0.063 n=10+10) GlyphNRGBA64Src-8 238µs ± 3% 233µs ± 1% -2.21% (p=0.000 n=9+10) GlyphNRGBA128Over-8 1.52ms ± 2% 1.48ms ± 0% -2.11% (p=0.000 n=10+8) GlyphNRGBA128Src-8 951µs ± 3% 935µs ± 1% -1.69% (p=0.013 n=10+9) GlyphNRGBA256Over-8 6.00ms ± 1% 5.87ms ± 3% -2.12% (p=0.002 n=10+10) GlyphNRGBA256Src-8 3.94ms ± 2% 3.80ms ± 2% -3.64% (p=0.000 n=10+10) A comparison of the non-SIMD and SIMD versions: name time/op FixedAccumulateMask16-8 237ns ± 0% FixedAccumulateMaskSIMD16-8 80.0ns ± 1% FloatingAccumulateMask16-8 413ns ± 2% FloatingAccumulateMaskSIMD16-8 166ns ± 0% FixedAccumulateMask64-8 3.42µs ± 0% FixedAccumulateMaskSIMD64-8 1.09µs ± 0% FloatingAccumulateMask64-8 6.92µs ± 0% FloatingAccumulateMaskSIMD64-8 2.47µs ± 1% Change-Id: Ib6980e5975ed2842ff2a372f76dd5f2e95c5526c Reviewed-on: https://go-review.googlesource.com/30898 Reviewed-by: David Crawshaw <crawshaw@golang.org>
2016-10-13 05:50:52 +02:00
JAE fxAccOpSrcEnd
// x = src[i] + offset
MOVL (SI), X1
PADDD X7, X1
// y = abs(x)
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>
2016-10-20 06:15:53 +02:00
// y >>= 2 // Shift by 2*ϕ - 16.
// y = min(y, fxAlmost65536)
PABSD X1, X2
PSRLL $2, X2
PMINUD X5, X2
// z = convertToInt32(y)
// No-op.
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>
2016-10-20 06:15:53 +02:00
// dst[0] = uint8(z>>8)
MOVL X2, BX
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>
2016-10-20 06:15:53 +02:00
SHRL $8, BX
MOVB BX, (DI)
// offset = x
MOVOU X1, X7
// i += 1
// dst = dst[1:]
// src = src[1:]
ADDQ $1, R9
ADDQ $1, DI
ADDQ $4, SI
JMP fxAccOpSrcLoop1
vector: add SIMD versions of xxxAccumulateMask. name old time/op new time/op delta GlyphAlphaLoose16Over-8 3.96µs ± 0% 3.64µs ± 1% -8.08% (p=0.000 n=8+10) GlyphAlphaLoose16Src-8 3.64µs ± 0% 3.35µs ± 0% -7.88% (p=0.000 n=8+9) GlyphAlphaLoose32Over-8 8.45µs ± 0% 6.74µs ± 0% -20.22% (p=0.000 n=8+9) GlyphAlphaLoose32Src-8 7.24µs ± 0% 5.54µs ± 1% -23.48% (p=0.000 n=8+10) GlyphAlphaLoose64Over-8 22.2µs ± 0% 17.4µs ± 0% -21.67% (p=0.000 n=9+9) GlyphAlphaLoose64Src-8 17.6µs ± 1% 12.2µs ± 1% -30.32% (p=0.000 n=10+10) GlyphAlphaLoose128Over-8 67.9µs ± 0% 53.3µs ± 1% -21.53% (p=0.000 n=10+10) GlyphAlphaLoose128Src-8 48.2µs ± 0% 32.6µs ± 2% -32.41% (p=0.000 n=9+10) GlyphAlphaLoose256Over-8 242µs ± 1% 187µs ± 1% -22.96% (p=0.000 n=9+9) GlyphAlphaLoose256Src-8 163µs ± 0% 105µs ± 1% -35.83% (p=0.000 n=9+9) GlyphRGBA16Over-8 5.25µs ± 1% 4.95µs ± 0% -5.78% (p=0.000 n=9+9) GlyphRGBA16Src-8 4.72µs ± 0% 4.43µs ± 1% -6.22% (p=0.000 n=9+10) GlyphRGBA32Over-8 13.5µs ± 0% 11.9µs ± 1% -12.19% (p=0.000 n=9+10) GlyphRGBA32Src-8 11.5µs ± 1% 9.8µs ± 0% -14.72% (p=0.000 n=9+9) GlyphRGBA64Over-8 42.0µs ± 2% 36.9µs ± 1% -12.19% (p=0.000 n=10+10) GlyphRGBA64Src-8 34.1µs ± 1% 28.5µs ± 0% -16.25% (p=0.000 n=9+7) GlyphRGBA128Over-8 149µs ± 2% 133µs ± 1% -10.24% (p=0.000 n=10+9) GlyphRGBA128Src-8 115µs ± 1% 99µs ± 1% -13.57% (p=0.000 n=9+10) GlyphRGBA256Over-8 566µs ± 0% 511µs ± 1% -9.85% (p=0.000 n=9+10) GlyphRGBA256Src-8 435µs ± 0% 372µs ± 0% -14.64% (p=0.000 n=9+8) GlyphNRGBA16Over-8 26.9µs ± 3% 26.0µs ± 3% -3.55% (p=0.000 n=10+9) GlyphNRGBA16Src-8 18.8µs ± 2% 18.4µs ± 2% -2.21% (p=0.000 n=9+10) GlyphNRGBA32Over-8 99.1µs ± 2% 95.9µs ± 3% -3.23% (p=0.000 n=10+10) GlyphNRGBA32Src-8 65.6µs ± 3% 62.8µs ± 2% -4.36% (p=0.000 n=10+10) GlyphNRGBA64Over-8 376µs ± 4% 370µs ± 2% ~ (p=0.063 n=10+10) GlyphNRGBA64Src-8 238µs ± 3% 233µs ± 1% -2.21% (p=0.000 n=9+10) GlyphNRGBA128Over-8 1.52ms ± 2% 1.48ms ± 0% -2.11% (p=0.000 n=10+8) GlyphNRGBA128Src-8 951µs ± 3% 935µs ± 1% -1.69% (p=0.013 n=10+9) GlyphNRGBA256Over-8 6.00ms ± 1% 5.87ms ± 3% -2.12% (p=0.002 n=10+10) GlyphNRGBA256Src-8 3.94ms ± 2% 3.80ms ± 2% -3.64% (p=0.000 n=10+10) A comparison of the non-SIMD and SIMD versions: name time/op FixedAccumulateMask16-8 237ns ± 0% FixedAccumulateMaskSIMD16-8 80.0ns ± 1% FloatingAccumulateMask16-8 413ns ± 2% FloatingAccumulateMaskSIMD16-8 166ns ± 0% FixedAccumulateMask64-8 3.42µs ± 0% FixedAccumulateMaskSIMD64-8 1.09µs ± 0% FloatingAccumulateMask64-8 6.92µs ± 0% FloatingAccumulateMaskSIMD64-8 2.47µs ± 1% Change-Id: Ib6980e5975ed2842ff2a372f76dd5f2e95c5526c Reviewed-on: https://go-review.googlesource.com/30898 Reviewed-by: David Crawshaw <crawshaw@golang.org>
2016-10-13 05:50:52 +02:00
fxAccOpSrcEnd:
RET
// ----------------------------------------------------------------------------
// func fixedAccumulateMaskSIMD(buf []uint32)
//
// XMM registers. Variable names are per
// https://github.com/google/font-rs/blob/master/src/accumulate.c
//
// xmm0 scratch
// xmm1 x
// xmm2 y, z
// xmm3 -
// xmm4 -
// xmm5 fxAlmost65536
// xmm6 -
// xmm7 offset
// xmm8 -
// xmm9 -
// xmm10 -
TEXT ·fixedAccumulateMaskSIMD(SB), NOSPLIT, $0-24
MOVQ buf_base+0(FP), DI
MOVQ buf_len+8(FP), BX
MOVQ buf_base+0(FP), SI
MOVQ buf_len+8(FP), R10
// R10 = len(src) &^ 3
// R11 = len(src)
MOVQ R10, R11
ANDQ $-4, R10
// fxAlmost65536 := XMM(0x0000ffff repeated four times) // Maximum of an uint16.
MOVOU fxAlmost65536<>(SB), X5
// offset := XMM(0x00000000 repeated four times) // Cumulative sum.
XORPS X7, X7
// i := 0
MOVQ $0, R9
fxAccMaskLoop4:
// for i < (len(src) &^ 3)
CMPQ R9, R10
JAE fxAccMaskLoop1
// x = XMM(s0, s1, s2, s3)
//
// Where s0 is src[i+0], s1 is src[i+1], etc.
MOVOU (SI), X1
// scratch = XMM(0, s0, s1, s2)
// x += scratch // yields x == XMM(s0, s0+s1, s1+s2, s2+s3)
MOVOU X1, X0
PSLLO $4, X0
PADDD X0, X1
// scratch = XMM(0, 0, 0, 0)
// scratch = XMM(scratch@0, scratch@0, x@0, x@1) // yields scratch == XMM(0, 0, s0, s0+s1)
// x += scratch // yields x == XMM(s0, s0+s1, s0+s1+s2, s0+s1+s2+s3)
XORPS X0, X0
SHUFPS $0x40, X1, X0
PADDD X0, X1
// x += offset
PADDD X7, X1
// y = abs(x)
vector: change ϕ from 10 to 9. This slight loss in quality allows us to use int32 math exclusively throughout raster_fixed.go, instead of occasionally dropping into int64 math. The change in ϕ doesn't affect the benchmarks noticably, but staying in int32 does. The net effect: name old time/op new time/op delta GlyphAlpha16Over-8 3.36µs ± 0% 2.99µs ± 0% -10.89% (p=0.000 n=10+9) GlyphAlpha16Src-8 3.26µs ± 0% 2.89µs ± 1% -11.34% (p=0.000 n=9+10) GlyphAlpha32Over-8 5.20µs ± 0% 4.53µs ± 0% -12.76% (p=0.000 n=8+10) GlyphAlpha32Src-8 4.81µs ± 1% 4.14µs ± 0% -13.91% (p=0.000 n=9+9) GlyphAlpha64Over-8 10.2µs ± 0% 9.0µs ± 1% -11.99% (p=0.000 n=9+10) GlyphAlpha64Src-8 8.62µs ± 0% 7.42µs ± 1% -13.89% (p=0.000 n=9+10) GlyphAlpha128Over-8 24.1µs ± 0% 21.8µs ± 0% -9.32% (p=0.000 n=9+9) GlyphAlpha128Src-8 17.9µs ± 0% 15.6µs ± 0% -12.68% (p=0.000 n=9+10) GlyphAlpha256Over-8 70.1µs ± 0% 66.3µs ± 1% -5.44% (p=0.000 n=10+10) GlyphAlpha256Src-8 45.2µs ± 1% 41.2µs ± 1% -8.92% (p=0.000 n=10+10) GlyphRGBA16Over-8 5.12µs ± 0% 4.75µs ± 0% -7.15% (p=0.000 n=10+9) GlyphRGBA16Src-8 4.57µs ± 1% 4.20µs ± 0% -8.18% (p=0.000 n=9+8) GlyphRGBA32Over-8 12.1µs ± 0% 11.4µs ± 0% -5.50% (p=0.000 n=10+9) GlyphRGBA32Src-8 10.0µs ± 0% 9.3µs ± 1% -6.80% (p=0.000 n=10+9) GlyphRGBA64Over-8 37.2µs ± 0% 36.0µs ± 0% -3.17% (p=0.000 n=9+8) GlyphRGBA64Src-8 29.0µs ± 1% 27.9µs ± 1% -4.05% (p=0.000 n=9+10) GlyphRGBA128Over-8 134µs ± 1% 131µs ± 0% -1.85% (p=0.000 n=9+9) GlyphRGBA128Src-8 100µs ± 1% 98µs ± 0% -2.27% (p=0.000 n=10+9) GlyphRGBA256Over-8 506µs ± 0% 503µs ± 0% -0.56% (p=0.000 n=10+8) GlyphRGBA256Src-8 373µs ± 0% 370µs ± 0% -1.01% (p=0.000 n=10+9) Change-Id: Ie02afac6fd6fa95f090bf3fe0a5c744799ea9dc5 Reviewed-on: https://go-review.googlesource.com/31532 Reviewed-by: David Crawshaw <crawshaw@golang.org>
2016-10-20 01:34:23 +02:00
// y >>= 2 // Shift by 2*ϕ - 16.
vector: add SIMD versions of xxxAccumulateMask. name old time/op new time/op delta GlyphAlphaLoose16Over-8 3.96µs ± 0% 3.64µs ± 1% -8.08% (p=0.000 n=8+10) GlyphAlphaLoose16Src-8 3.64µs ± 0% 3.35µs ± 0% -7.88% (p=0.000 n=8+9) GlyphAlphaLoose32Over-8 8.45µs ± 0% 6.74µs ± 0% -20.22% (p=0.000 n=8+9) GlyphAlphaLoose32Src-8 7.24µs ± 0% 5.54µs ± 1% -23.48% (p=0.000 n=8+10) GlyphAlphaLoose64Over-8 22.2µs ± 0% 17.4µs ± 0% -21.67% (p=0.000 n=9+9) GlyphAlphaLoose64Src-8 17.6µs ± 1% 12.2µs ± 1% -30.32% (p=0.000 n=10+10) GlyphAlphaLoose128Over-8 67.9µs ± 0% 53.3µs ± 1% -21.53% (p=0.000 n=10+10) GlyphAlphaLoose128Src-8 48.2µs ± 0% 32.6µs ± 2% -32.41% (p=0.000 n=9+10) GlyphAlphaLoose256Over-8 242µs ± 1% 187µs ± 1% -22.96% (p=0.000 n=9+9) GlyphAlphaLoose256Src-8 163µs ± 0% 105µs ± 1% -35.83% (p=0.000 n=9+9) GlyphRGBA16Over-8 5.25µs ± 1% 4.95µs ± 0% -5.78% (p=0.000 n=9+9) GlyphRGBA16Src-8 4.72µs ± 0% 4.43µs ± 1% -6.22% (p=0.000 n=9+10) GlyphRGBA32Over-8 13.5µs ± 0% 11.9µs ± 1% -12.19% (p=0.000 n=9+10) GlyphRGBA32Src-8 11.5µs ± 1% 9.8µs ± 0% -14.72% (p=0.000 n=9+9) GlyphRGBA64Over-8 42.0µs ± 2% 36.9µs ± 1% -12.19% (p=0.000 n=10+10) GlyphRGBA64Src-8 34.1µs ± 1% 28.5µs ± 0% -16.25% (p=0.000 n=9+7) GlyphRGBA128Over-8 149µs ± 2% 133µs ± 1% -10.24% (p=0.000 n=10+9) GlyphRGBA128Src-8 115µs ± 1% 99µs ± 1% -13.57% (p=0.000 n=9+10) GlyphRGBA256Over-8 566µs ± 0% 511µs ± 1% -9.85% (p=0.000 n=9+10) GlyphRGBA256Src-8 435µs ± 0% 372µs ± 0% -14.64% (p=0.000 n=9+8) GlyphNRGBA16Over-8 26.9µs ± 3% 26.0µs ± 3% -3.55% (p=0.000 n=10+9) GlyphNRGBA16Src-8 18.8µs ± 2% 18.4µs ± 2% -2.21% (p=0.000 n=9+10) GlyphNRGBA32Over-8 99.1µs ± 2% 95.9µs ± 3% -3.23% (p=0.000 n=10+10) GlyphNRGBA32Src-8 65.6µs ± 3% 62.8µs ± 2% -4.36% (p=0.000 n=10+10) GlyphNRGBA64Over-8 376µs ± 4% 370µs ± 2% ~ (p=0.063 n=10+10) GlyphNRGBA64Src-8 238µs ± 3% 233µs ± 1% -2.21% (p=0.000 n=9+10) GlyphNRGBA128Over-8 1.52ms ± 2% 1.48ms ± 0% -2.11% (p=0.000 n=10+8) GlyphNRGBA128Src-8 951µs ± 3% 935µs ± 1% -1.69% (p=0.013 n=10+9) GlyphNRGBA256Over-8 6.00ms ± 1% 5.87ms ± 3% -2.12% (p=0.002 n=10+10) GlyphNRGBA256Src-8 3.94ms ± 2% 3.80ms ± 2% -3.64% (p=0.000 n=10+10) A comparison of the non-SIMD and SIMD versions: name time/op FixedAccumulateMask16-8 237ns ± 0% FixedAccumulateMaskSIMD16-8 80.0ns ± 1% FloatingAccumulateMask16-8 413ns ± 2% FloatingAccumulateMaskSIMD16-8 166ns ± 0% FixedAccumulateMask64-8 3.42µs ± 0% FixedAccumulateMaskSIMD64-8 1.09µs ± 0% FloatingAccumulateMask64-8 6.92µs ± 0% FloatingAccumulateMaskSIMD64-8 2.47µs ± 1% Change-Id: Ib6980e5975ed2842ff2a372f76dd5f2e95c5526c Reviewed-on: https://go-review.googlesource.com/30898 Reviewed-by: David Crawshaw <crawshaw@golang.org>
2016-10-13 05:50:52 +02:00
// y = min(y, fxAlmost65536)
PABSD X1, X2
PSRLL $2, X2
PMINUD X5, X2
vector: add SIMD versions of xxxAccumulateMask. name old time/op new time/op delta GlyphAlphaLoose16Over-8 3.96µs ± 0% 3.64µs ± 1% -8.08% (p=0.000 n=8+10) GlyphAlphaLoose16Src-8 3.64µs ± 0% 3.35µs ± 0% -7.88% (p=0.000 n=8+9) GlyphAlphaLoose32Over-8 8.45µs ± 0% 6.74µs ± 0% -20.22% (p=0.000 n=8+9) GlyphAlphaLoose32Src-8 7.24µs ± 0% 5.54µs ± 1% -23.48% (p=0.000 n=8+10) GlyphAlphaLoose64Over-8 22.2µs ± 0% 17.4µs ± 0% -21.67% (p=0.000 n=9+9) GlyphAlphaLoose64Src-8 17.6µs ± 1% 12.2µs ± 1% -30.32% (p=0.000 n=10+10) GlyphAlphaLoose128Over-8 67.9µs ± 0% 53.3µs ± 1% -21.53% (p=0.000 n=10+10) GlyphAlphaLoose128Src-8 48.2µs ± 0% 32.6µs ± 2% -32.41% (p=0.000 n=9+10) GlyphAlphaLoose256Over-8 242µs ± 1% 187µs ± 1% -22.96% (p=0.000 n=9+9) GlyphAlphaLoose256Src-8 163µs ± 0% 105µs ± 1% -35.83% (p=0.000 n=9+9) GlyphRGBA16Over-8 5.25µs ± 1% 4.95µs ± 0% -5.78% (p=0.000 n=9+9) GlyphRGBA16Src-8 4.72µs ± 0% 4.43µs ± 1% -6.22% (p=0.000 n=9+10) GlyphRGBA32Over-8 13.5µs ± 0% 11.9µs ± 1% -12.19% (p=0.000 n=9+10) GlyphRGBA32Src-8 11.5µs ± 1% 9.8µs ± 0% -14.72% (p=0.000 n=9+9) GlyphRGBA64Over-8 42.0µs ± 2% 36.9µs ± 1% -12.19% (p=0.000 n=10+10) GlyphRGBA64Src-8 34.1µs ± 1% 28.5µs ± 0% -16.25% (p=0.000 n=9+7) GlyphRGBA128Over-8 149µs ± 2% 133µs ± 1% -10.24% (p=0.000 n=10+9) GlyphRGBA128Src-8 115µs ± 1% 99µs ± 1% -13.57% (p=0.000 n=9+10) GlyphRGBA256Over-8 566µs ± 0% 511µs ± 1% -9.85% (p=0.000 n=9+10) GlyphRGBA256Src-8 435µs ± 0% 372µs ± 0% -14.64% (p=0.000 n=9+8) GlyphNRGBA16Over-8 26.9µs ± 3% 26.0µs ± 3% -3.55% (p=0.000 n=10+9) GlyphNRGBA16Src-8 18.8µs ± 2% 18.4µs ± 2% -2.21% (p=0.000 n=9+10) GlyphNRGBA32Over-8 99.1µs ± 2% 95.9µs ± 3% -3.23% (p=0.000 n=10+10) GlyphNRGBA32Src-8 65.6µs ± 3% 62.8µs ± 2% -4.36% (p=0.000 n=10+10) GlyphNRGBA64Over-8 376µs ± 4% 370µs ± 2% ~ (p=0.063 n=10+10) GlyphNRGBA64Src-8 238µs ± 3% 233µs ± 1% -2.21% (p=0.000 n=9+10) GlyphNRGBA128Over-8 1.52ms ± 2% 1.48ms ± 0% -2.11% (p=0.000 n=10+8) GlyphNRGBA128Src-8 951µs ± 3% 935µs ± 1% -1.69% (p=0.013 n=10+9) GlyphNRGBA256Over-8 6.00ms ± 1% 5.87ms ± 3% -2.12% (p=0.002 n=10+10) GlyphNRGBA256Src-8 3.94ms ± 2% 3.80ms ± 2% -3.64% (p=0.000 n=10+10) A comparison of the non-SIMD and SIMD versions: name time/op FixedAccumulateMask16-8 237ns ± 0% FixedAccumulateMaskSIMD16-8 80.0ns ± 1% FloatingAccumulateMask16-8 413ns ± 2% FloatingAccumulateMaskSIMD16-8 166ns ± 0% FixedAccumulateMask64-8 3.42µs ± 0% FixedAccumulateMaskSIMD64-8 1.09µs ± 0% FloatingAccumulateMask64-8 6.92µs ± 0% FloatingAccumulateMaskSIMD64-8 2.47µs ± 1% Change-Id: Ib6980e5975ed2842ff2a372f76dd5f2e95c5526c Reviewed-on: https://go-review.googlesource.com/30898 Reviewed-by: David Crawshaw <crawshaw@golang.org>
2016-10-13 05:50:52 +02:00
// z = convertToInt32(y)
// No-op.
vector: add SIMD versions of xxxAccumulateMask. name old time/op new time/op delta GlyphAlphaLoose16Over-8 3.96µs ± 0% 3.64µs ± 1% -8.08% (p=0.000 n=8+10) GlyphAlphaLoose16Src-8 3.64µs ± 0% 3.35µs ± 0% -7.88% (p=0.000 n=8+9) GlyphAlphaLoose32Over-8 8.45µs ± 0% 6.74µs ± 0% -20.22% (p=0.000 n=8+9) GlyphAlphaLoose32Src-8 7.24µs ± 0% 5.54µs ± 1% -23.48% (p=0.000 n=8+10) GlyphAlphaLoose64Over-8 22.2µs ± 0% 17.4µs ± 0% -21.67% (p=0.000 n=9+9) GlyphAlphaLoose64Src-8 17.6µs ± 1% 12.2µs ± 1% -30.32% (p=0.000 n=10+10) GlyphAlphaLoose128Over-8 67.9µs ± 0% 53.3µs ± 1% -21.53% (p=0.000 n=10+10) GlyphAlphaLoose128Src-8 48.2µs ± 0% 32.6µs ± 2% -32.41% (p=0.000 n=9+10) GlyphAlphaLoose256Over-8 242µs ± 1% 187µs ± 1% -22.96% (p=0.000 n=9+9) GlyphAlphaLoose256Src-8 163µs ± 0% 105µs ± 1% -35.83% (p=0.000 n=9+9) GlyphRGBA16Over-8 5.25µs ± 1% 4.95µs ± 0% -5.78% (p=0.000 n=9+9) GlyphRGBA16Src-8 4.72µs ± 0% 4.43µs ± 1% -6.22% (p=0.000 n=9+10) GlyphRGBA32Over-8 13.5µs ± 0% 11.9µs ± 1% -12.19% (p=0.000 n=9+10) GlyphRGBA32Src-8 11.5µs ± 1% 9.8µs ± 0% -14.72% (p=0.000 n=9+9) GlyphRGBA64Over-8 42.0µs ± 2% 36.9µs ± 1% -12.19% (p=0.000 n=10+10) GlyphRGBA64Src-8 34.1µs ± 1% 28.5µs ± 0% -16.25% (p=0.000 n=9+7) GlyphRGBA128Over-8 149µs ± 2% 133µs ± 1% -10.24% (p=0.000 n=10+9) GlyphRGBA128Src-8 115µs ± 1% 99µs ± 1% -13.57% (p=0.000 n=9+10) GlyphRGBA256Over-8 566µs ± 0% 511µs ± 1% -9.85% (p=0.000 n=9+10) GlyphRGBA256Src-8 435µs ± 0% 372µs ± 0% -14.64% (p=0.000 n=9+8) GlyphNRGBA16Over-8 26.9µs ± 3% 26.0µs ± 3% -3.55% (p=0.000 n=10+9) GlyphNRGBA16Src-8 18.8µs ± 2% 18.4µs ± 2% -2.21% (p=0.000 n=9+10) GlyphNRGBA32Over-8 99.1µs ± 2% 95.9µs ± 3% -3.23% (p=0.000 n=10+10) GlyphNRGBA32Src-8 65.6µs ± 3% 62.8µs ± 2% -4.36% (p=0.000 n=10+10) GlyphNRGBA64Over-8 376µs ± 4% 370µs ± 2% ~ (p=0.063 n=10+10) GlyphNRGBA64Src-8 238µs ± 3% 233µs ± 1% -2.21% (p=0.000 n=9+10) GlyphNRGBA128Over-8 1.52ms ± 2% 1.48ms ± 0% -2.11% (p=0.000 n=10+8) GlyphNRGBA128Src-8 951µs ± 3% 935µs ± 1% -1.69% (p=0.013 n=10+9) GlyphNRGBA256Over-8 6.00ms ± 1% 5.87ms ± 3% -2.12% (p=0.002 n=10+10) GlyphNRGBA256Src-8 3.94ms ± 2% 3.80ms ± 2% -3.64% (p=0.000 n=10+10) A comparison of the non-SIMD and SIMD versions: name time/op FixedAccumulateMask16-8 237ns ± 0% FixedAccumulateMaskSIMD16-8 80.0ns ± 1% FloatingAccumulateMask16-8 413ns ± 2% FloatingAccumulateMaskSIMD16-8 166ns ± 0% FixedAccumulateMask64-8 3.42µs ± 0% FixedAccumulateMaskSIMD64-8 1.09µs ± 0% FloatingAccumulateMask64-8 6.92µs ± 0% FloatingAccumulateMaskSIMD64-8 2.47µs ± 1% Change-Id: Ib6980e5975ed2842ff2a372f76dd5f2e95c5526c Reviewed-on: https://go-review.googlesource.com/30898 Reviewed-by: David Crawshaw <crawshaw@golang.org>
2016-10-13 05:50:52 +02:00
// copy(dst[:4], z)
MOVOU X2, (DI)
// offset = XMM(x@3, x@3, x@3, x@3)
MOVOU X1, X7
SHUFPS $0xff, X1, X7
// i += 4
// dst = dst[4:]
// src = src[4:]
ADDQ $4, R9
ADDQ $16, DI
ADDQ $16, SI
JMP fxAccMaskLoop4
fxAccMaskLoop1:
// for i < len(src)
CMPQ R9, R11
JAE fxAccMaskEnd
// x = src[i] + offset
MOVL (SI), X1
PADDD X7, X1
// y = abs(x)
vector: change ϕ from 10 to 9. This slight loss in quality allows us to use int32 math exclusively throughout raster_fixed.go, instead of occasionally dropping into int64 math. The change in ϕ doesn't affect the benchmarks noticably, but staying in int32 does. The net effect: name old time/op new time/op delta GlyphAlpha16Over-8 3.36µs ± 0% 2.99µs ± 0% -10.89% (p=0.000 n=10+9) GlyphAlpha16Src-8 3.26µs ± 0% 2.89µs ± 1% -11.34% (p=0.000 n=9+10) GlyphAlpha32Over-8 5.20µs ± 0% 4.53µs ± 0% -12.76% (p=0.000 n=8+10) GlyphAlpha32Src-8 4.81µs ± 1% 4.14µs ± 0% -13.91% (p=0.000 n=9+9) GlyphAlpha64Over-8 10.2µs ± 0% 9.0µs ± 1% -11.99% (p=0.000 n=9+10) GlyphAlpha64Src-8 8.62µs ± 0% 7.42µs ± 1% -13.89% (p=0.000 n=9+10) GlyphAlpha128Over-8 24.1µs ± 0% 21.8µs ± 0% -9.32% (p=0.000 n=9+9) GlyphAlpha128Src-8 17.9µs ± 0% 15.6µs ± 0% -12.68% (p=0.000 n=9+10) GlyphAlpha256Over-8 70.1µs ± 0% 66.3µs ± 1% -5.44% (p=0.000 n=10+10) GlyphAlpha256Src-8 45.2µs ± 1% 41.2µs ± 1% -8.92% (p=0.000 n=10+10) GlyphRGBA16Over-8 5.12µs ± 0% 4.75µs ± 0% -7.15% (p=0.000 n=10+9) GlyphRGBA16Src-8 4.57µs ± 1% 4.20µs ± 0% -8.18% (p=0.000 n=9+8) GlyphRGBA32Over-8 12.1µs ± 0% 11.4µs ± 0% -5.50% (p=0.000 n=10+9) GlyphRGBA32Src-8 10.0µs ± 0% 9.3µs ± 1% -6.80% (p=0.000 n=10+9) GlyphRGBA64Over-8 37.2µs ± 0% 36.0µs ± 0% -3.17% (p=0.000 n=9+8) GlyphRGBA64Src-8 29.0µs ± 1% 27.9µs ± 1% -4.05% (p=0.000 n=9+10) GlyphRGBA128Over-8 134µs ± 1% 131µs ± 0% -1.85% (p=0.000 n=9+9) GlyphRGBA128Src-8 100µs ± 1% 98µs ± 0% -2.27% (p=0.000 n=10+9) GlyphRGBA256Over-8 506µs ± 0% 503µs ± 0% -0.56% (p=0.000 n=10+8) GlyphRGBA256Src-8 373µs ± 0% 370µs ± 0% -1.01% (p=0.000 n=10+9) Change-Id: Ie02afac6fd6fa95f090bf3fe0a5c744799ea9dc5 Reviewed-on: https://go-review.googlesource.com/31532 Reviewed-by: David Crawshaw <crawshaw@golang.org>
2016-10-20 01:34:23 +02:00
// y >>= 2 // Shift by 2*ϕ - 16.
vector: add SIMD versions of xxxAccumulateMask. name old time/op new time/op delta GlyphAlphaLoose16Over-8 3.96µs ± 0% 3.64µs ± 1% -8.08% (p=0.000 n=8+10) GlyphAlphaLoose16Src-8 3.64µs ± 0% 3.35µs ± 0% -7.88% (p=0.000 n=8+9) GlyphAlphaLoose32Over-8 8.45µs ± 0% 6.74µs ± 0% -20.22% (p=0.000 n=8+9) GlyphAlphaLoose32Src-8 7.24µs ± 0% 5.54µs ± 1% -23.48% (p=0.000 n=8+10) GlyphAlphaLoose64Over-8 22.2µs ± 0% 17.4µs ± 0% -21.67% (p=0.000 n=9+9) GlyphAlphaLoose64Src-8 17.6µs ± 1% 12.2µs ± 1% -30.32% (p=0.000 n=10+10) GlyphAlphaLoose128Over-8 67.9µs ± 0% 53.3µs ± 1% -21.53% (p=0.000 n=10+10) GlyphAlphaLoose128Src-8 48.2µs ± 0% 32.6µs ± 2% -32.41% (p=0.000 n=9+10) GlyphAlphaLoose256Over-8 242µs ± 1% 187µs ± 1% -22.96% (p=0.000 n=9+9) GlyphAlphaLoose256Src-8 163µs ± 0% 105µs ± 1% -35.83% (p=0.000 n=9+9) GlyphRGBA16Over-8 5.25µs ± 1% 4.95µs ± 0% -5.78% (p=0.000 n=9+9) GlyphRGBA16Src-8 4.72µs ± 0% 4.43µs ± 1% -6.22% (p=0.000 n=9+10) GlyphRGBA32Over-8 13.5µs ± 0% 11.9µs ± 1% -12.19% (p=0.000 n=9+10) GlyphRGBA32Src-8 11.5µs ± 1% 9.8µs ± 0% -14.72% (p=0.000 n=9+9) GlyphRGBA64Over-8 42.0µs ± 2% 36.9µs ± 1% -12.19% (p=0.000 n=10+10) GlyphRGBA64Src-8 34.1µs ± 1% 28.5µs ± 0% -16.25% (p=0.000 n=9+7) GlyphRGBA128Over-8 149µs ± 2% 133µs ± 1% -10.24% (p=0.000 n=10+9) GlyphRGBA128Src-8 115µs ± 1% 99µs ± 1% -13.57% (p=0.000 n=9+10) GlyphRGBA256Over-8 566µs ± 0% 511µs ± 1% -9.85% (p=0.000 n=9+10) GlyphRGBA256Src-8 435µs ± 0% 372µs ± 0% -14.64% (p=0.000 n=9+8) GlyphNRGBA16Over-8 26.9µs ± 3% 26.0µs ± 3% -3.55% (p=0.000 n=10+9) GlyphNRGBA16Src-8 18.8µs ± 2% 18.4µs ± 2% -2.21% (p=0.000 n=9+10) GlyphNRGBA32Over-8 99.1µs ± 2% 95.9µs ± 3% -3.23% (p=0.000 n=10+10) GlyphNRGBA32Src-8 65.6µs ± 3% 62.8µs ± 2% -4.36% (p=0.000 n=10+10) GlyphNRGBA64Over-8 376µs ± 4% 370µs ± 2% ~ (p=0.063 n=10+10) GlyphNRGBA64Src-8 238µs ± 3% 233µs ± 1% -2.21% (p=0.000 n=9+10) GlyphNRGBA128Over-8 1.52ms ± 2% 1.48ms ± 0% -2.11% (p=0.000 n=10+8) GlyphNRGBA128Src-8 951µs ± 3% 935µs ± 1% -1.69% (p=0.013 n=10+9) GlyphNRGBA256Over-8 6.00ms ± 1% 5.87ms ± 3% -2.12% (p=0.002 n=10+10) GlyphNRGBA256Src-8 3.94ms ± 2% 3.80ms ± 2% -3.64% (p=0.000 n=10+10) A comparison of the non-SIMD and SIMD versions: name time/op FixedAccumulateMask16-8 237ns ± 0% FixedAccumulateMaskSIMD16-8 80.0ns ± 1% FloatingAccumulateMask16-8 413ns ± 2% FloatingAccumulateMaskSIMD16-8 166ns ± 0% FixedAccumulateMask64-8 3.42µs ± 0% FixedAccumulateMaskSIMD64-8 1.09µs ± 0% FloatingAccumulateMask64-8 6.92µs ± 0% FloatingAccumulateMaskSIMD64-8 2.47µs ± 1% Change-Id: Ib6980e5975ed2842ff2a372f76dd5f2e95c5526c Reviewed-on: https://go-review.googlesource.com/30898 Reviewed-by: David Crawshaw <crawshaw@golang.org>
2016-10-13 05:50:52 +02:00
// y = min(y, fxAlmost65536)
PABSD X1, X2
PSRLL $2, X2
PMINUD X5, X2
vector: add SIMD versions of xxxAccumulateMask. name old time/op new time/op delta GlyphAlphaLoose16Over-8 3.96µs ± 0% 3.64µs ± 1% -8.08% (p=0.000 n=8+10) GlyphAlphaLoose16Src-8 3.64µs ± 0% 3.35µs ± 0% -7.88% (p=0.000 n=8+9) GlyphAlphaLoose32Over-8 8.45µs ± 0% 6.74µs ± 0% -20.22% (p=0.000 n=8+9) GlyphAlphaLoose32Src-8 7.24µs ± 0% 5.54µs ± 1% -23.48% (p=0.000 n=8+10) GlyphAlphaLoose64Over-8 22.2µs ± 0% 17.4µs ± 0% -21.67% (p=0.000 n=9+9) GlyphAlphaLoose64Src-8 17.6µs ± 1% 12.2µs ± 1% -30.32% (p=0.000 n=10+10) GlyphAlphaLoose128Over-8 67.9µs ± 0% 53.3µs ± 1% -21.53% (p=0.000 n=10+10) GlyphAlphaLoose128Src-8 48.2µs ± 0% 32.6µs ± 2% -32.41% (p=0.000 n=9+10) GlyphAlphaLoose256Over-8 242µs ± 1% 187µs ± 1% -22.96% (p=0.000 n=9+9) GlyphAlphaLoose256Src-8 163µs ± 0% 105µs ± 1% -35.83% (p=0.000 n=9+9) GlyphRGBA16Over-8 5.25µs ± 1% 4.95µs ± 0% -5.78% (p=0.000 n=9+9) GlyphRGBA16Src-8 4.72µs ± 0% 4.43µs ± 1% -6.22% (p=0.000 n=9+10) GlyphRGBA32Over-8 13.5µs ± 0% 11.9µs ± 1% -12.19% (p=0.000 n=9+10) GlyphRGBA32Src-8 11.5µs ± 1% 9.8µs ± 0% -14.72% (p=0.000 n=9+9) GlyphRGBA64Over-8 42.0µs ± 2% 36.9µs ± 1% -12.19% (p=0.000 n=10+10) GlyphRGBA64Src-8 34.1µs ± 1% 28.5µs ± 0% -16.25% (p=0.000 n=9+7) GlyphRGBA128Over-8 149µs ± 2% 133µs ± 1% -10.24% (p=0.000 n=10+9) GlyphRGBA128Src-8 115µs ± 1% 99µs ± 1% -13.57% (p=0.000 n=9+10) GlyphRGBA256Over-8 566µs ± 0% 511µs ± 1% -9.85% (p=0.000 n=9+10) GlyphRGBA256Src-8 435µs ± 0% 372µs ± 0% -14.64% (p=0.000 n=9+8) GlyphNRGBA16Over-8 26.9µs ± 3% 26.0µs ± 3% -3.55% (p=0.000 n=10+9) GlyphNRGBA16Src-8 18.8µs ± 2% 18.4µs ± 2% -2.21% (p=0.000 n=9+10) GlyphNRGBA32Over-8 99.1µs ± 2% 95.9µs ± 3% -3.23% (p=0.000 n=10+10) GlyphNRGBA32Src-8 65.6µs ± 3% 62.8µs ± 2% -4.36% (p=0.000 n=10+10) GlyphNRGBA64Over-8 376µs ± 4% 370µs ± 2% ~ (p=0.063 n=10+10) GlyphNRGBA64Src-8 238µs ± 3% 233µs ± 1% -2.21% (p=0.000 n=9+10) GlyphNRGBA128Over-8 1.52ms ± 2% 1.48ms ± 0% -2.11% (p=0.000 n=10+8) GlyphNRGBA128Src-8 951µs ± 3% 935µs ± 1% -1.69% (p=0.013 n=10+9) GlyphNRGBA256Over-8 6.00ms ± 1% 5.87ms ± 3% -2.12% (p=0.002 n=10+10) GlyphNRGBA256Src-8 3.94ms ± 2% 3.80ms ± 2% -3.64% (p=0.000 n=10+10) A comparison of the non-SIMD and SIMD versions: name time/op FixedAccumulateMask16-8 237ns ± 0% FixedAccumulateMaskSIMD16-8 80.0ns ± 1% FloatingAccumulateMask16-8 413ns ± 2% FloatingAccumulateMaskSIMD16-8 166ns ± 0% FixedAccumulateMask64-8 3.42µs ± 0% FixedAccumulateMaskSIMD64-8 1.09µs ± 0% FloatingAccumulateMask64-8 6.92µs ± 0% FloatingAccumulateMaskSIMD64-8 2.47µs ± 1% Change-Id: Ib6980e5975ed2842ff2a372f76dd5f2e95c5526c Reviewed-on: https://go-review.googlesource.com/30898 Reviewed-by: David Crawshaw <crawshaw@golang.org>
2016-10-13 05:50:52 +02:00
// z = convertToInt32(y)
// No-op.
// dst[0] = uint32(z)
MOVL X2, (DI)
// offset = x
MOVOU X1, X7
// i += 1
// dst = dst[1:]
// src = src[1:]
ADDQ $1, R9
ADDQ $4, DI
ADDQ $4, SI
JMP fxAccMaskLoop1
fxAccMaskEnd:
RET
// ----------------------------------------------------------------------------
// func floatingAccumulateOpOverSIMD(dst []uint8, src []float32)
//
// XMM registers. Variable names are per
// https://github.com/google/font-rs/blob/master/src/accumulate.c
//
// xmm0 scratch
// xmm1 x
// xmm2 y, z
// xmm3 flSignMask
// xmm4 flOne
// xmm5 flAlmost65536
// xmm6 gather
// xmm7 offset
// xmm8 scatterAndMulBy0x101
// xmm9 fxAlmost65536
// xmm10 inverseFFFF
TEXT ·floatingAccumulateOpOverSIMD(SB), NOSPLIT, $8-48
vector: add SIMD versions of xxxAccumulateMask. name old time/op new time/op delta GlyphAlphaLoose16Over-8 3.96µs ± 0% 3.64µs ± 1% -8.08% (p=0.000 n=8+10) GlyphAlphaLoose16Src-8 3.64µs ± 0% 3.35µs ± 0% -7.88% (p=0.000 n=8+9) GlyphAlphaLoose32Over-8 8.45µs ± 0% 6.74µs ± 0% -20.22% (p=0.000 n=8+9) GlyphAlphaLoose32Src-8 7.24µs ± 0% 5.54µs ± 1% -23.48% (p=0.000 n=8+10) GlyphAlphaLoose64Over-8 22.2µs ± 0% 17.4µs ± 0% -21.67% (p=0.000 n=9+9) GlyphAlphaLoose64Src-8 17.6µs ± 1% 12.2µs ± 1% -30.32% (p=0.000 n=10+10) GlyphAlphaLoose128Over-8 67.9µs ± 0% 53.3µs ± 1% -21.53% (p=0.000 n=10+10) GlyphAlphaLoose128Src-8 48.2µs ± 0% 32.6µs ± 2% -32.41% (p=0.000 n=9+10) GlyphAlphaLoose256Over-8 242µs ± 1% 187µs ± 1% -22.96% (p=0.000 n=9+9) GlyphAlphaLoose256Src-8 163µs ± 0% 105µs ± 1% -35.83% (p=0.000 n=9+9) GlyphRGBA16Over-8 5.25µs ± 1% 4.95µs ± 0% -5.78% (p=0.000 n=9+9) GlyphRGBA16Src-8 4.72µs ± 0% 4.43µs ± 1% -6.22% (p=0.000 n=9+10) GlyphRGBA32Over-8 13.5µs ± 0% 11.9µs ± 1% -12.19% (p=0.000 n=9+10) GlyphRGBA32Src-8 11.5µs ± 1% 9.8µs ± 0% -14.72% (p=0.000 n=9+9) GlyphRGBA64Over-8 42.0µs ± 2% 36.9µs ± 1% -12.19% (p=0.000 n=10+10) GlyphRGBA64Src-8 34.1µs ± 1% 28.5µs ± 0% -16.25% (p=0.000 n=9+7) GlyphRGBA128Over-8 149µs ± 2% 133µs ± 1% -10.24% (p=0.000 n=10+9) GlyphRGBA128Src-8 115µs ± 1% 99µs ± 1% -13.57% (p=0.000 n=9+10) GlyphRGBA256Over-8 566µs ± 0% 511µs ± 1% -9.85% (p=0.000 n=9+10) GlyphRGBA256Src-8 435µs ± 0% 372µs ± 0% -14.64% (p=0.000 n=9+8) GlyphNRGBA16Over-8 26.9µs ± 3% 26.0µs ± 3% -3.55% (p=0.000 n=10+9) GlyphNRGBA16Src-8 18.8µs ± 2% 18.4µs ± 2% -2.21% (p=0.000 n=9+10) GlyphNRGBA32Over-8 99.1µs ± 2% 95.9µs ± 3% -3.23% (p=0.000 n=10+10) GlyphNRGBA32Src-8 65.6µs ± 3% 62.8µs ± 2% -4.36% (p=0.000 n=10+10) GlyphNRGBA64Over-8 376µs ± 4% 370µs ± 2% ~ (p=0.063 n=10+10) GlyphNRGBA64Src-8 238µs ± 3% 233µs ± 1% -2.21% (p=0.000 n=9+10) GlyphNRGBA128Over-8 1.52ms ± 2% 1.48ms ± 0% -2.11% (p=0.000 n=10+8) GlyphNRGBA128Src-8 951µs ± 3% 935µs ± 1% -1.69% (p=0.013 n=10+9) GlyphNRGBA256Over-8 6.00ms ± 1% 5.87ms ± 3% -2.12% (p=0.002 n=10+10) GlyphNRGBA256Src-8 3.94ms ± 2% 3.80ms ± 2% -3.64% (p=0.000 n=10+10) A comparison of the non-SIMD and SIMD versions: name time/op FixedAccumulateMask16-8 237ns ± 0% FixedAccumulateMaskSIMD16-8 80.0ns ± 1% FloatingAccumulateMask16-8 413ns ± 2% FloatingAccumulateMaskSIMD16-8 166ns ± 0% FixedAccumulateMask64-8 3.42µs ± 0% FixedAccumulateMaskSIMD64-8 1.09µs ± 0% FloatingAccumulateMask64-8 6.92µs ± 0% FloatingAccumulateMaskSIMD64-8 2.47µs ± 1% Change-Id: Ib6980e5975ed2842ff2a372f76dd5f2e95c5526c Reviewed-on: https://go-review.googlesource.com/30898 Reviewed-by: David Crawshaw <crawshaw@golang.org>
2016-10-13 05:50:52 +02:00
MOVQ dst_base+0(FP), DI
MOVQ dst_len+8(FP), BX
MOVQ src_base+24(FP), SI
MOVQ src_len+32(FP), R10
// Sanity check that len(dst) >= len(src).
CMPQ BX, R10
JLT flAccOpOverEnd
// R10 = len(src) &^ 3
// R11 = len(src)
MOVQ R10, R11
ANDQ $-4, R10
vector: add SIMD versions of xxxAccumulateMask. name old time/op new time/op delta GlyphAlphaLoose16Over-8 3.96µs ± 0% 3.64µs ± 1% -8.08% (p=0.000 n=8+10) GlyphAlphaLoose16Src-8 3.64µs ± 0% 3.35µs ± 0% -7.88% (p=0.000 n=8+9) GlyphAlphaLoose32Over-8 8.45µs ± 0% 6.74µs ± 0% -20.22% (p=0.000 n=8+9) GlyphAlphaLoose32Src-8 7.24µs ± 0% 5.54µs ± 1% -23.48% (p=0.000 n=8+10) GlyphAlphaLoose64Over-8 22.2µs ± 0% 17.4µs ± 0% -21.67% (p=0.000 n=9+9) GlyphAlphaLoose64Src-8 17.6µs ± 1% 12.2µs ± 1% -30.32% (p=0.000 n=10+10) GlyphAlphaLoose128Over-8 67.9µs ± 0% 53.3µs ± 1% -21.53% (p=0.000 n=10+10) GlyphAlphaLoose128Src-8 48.2µs ± 0% 32.6µs ± 2% -32.41% (p=0.000 n=9+10) GlyphAlphaLoose256Over-8 242µs ± 1% 187µs ± 1% -22.96% (p=0.000 n=9+9) GlyphAlphaLoose256Src-8 163µs ± 0% 105µs ± 1% -35.83% (p=0.000 n=9+9) GlyphRGBA16Over-8 5.25µs ± 1% 4.95µs ± 0% -5.78% (p=0.000 n=9+9) GlyphRGBA16Src-8 4.72µs ± 0% 4.43µs ± 1% -6.22% (p=0.000 n=9+10) GlyphRGBA32Over-8 13.5µs ± 0% 11.9µs ± 1% -12.19% (p=0.000 n=9+10) GlyphRGBA32Src-8 11.5µs ± 1% 9.8µs ± 0% -14.72% (p=0.000 n=9+9) GlyphRGBA64Over-8 42.0µs ± 2% 36.9µs ± 1% -12.19% (p=0.000 n=10+10) GlyphRGBA64Src-8 34.1µs ± 1% 28.5µs ± 0% -16.25% (p=0.000 n=9+7) GlyphRGBA128Over-8 149µs ± 2% 133µs ± 1% -10.24% (p=0.000 n=10+9) GlyphRGBA128Src-8 115µs ± 1% 99µs ± 1% -13.57% (p=0.000 n=9+10) GlyphRGBA256Over-8 566µs ± 0% 511µs ± 1% -9.85% (p=0.000 n=9+10) GlyphRGBA256Src-8 435µs ± 0% 372µs ± 0% -14.64% (p=0.000 n=9+8) GlyphNRGBA16Over-8 26.9µs ± 3% 26.0µs ± 3% -3.55% (p=0.000 n=10+9) GlyphNRGBA16Src-8 18.8µs ± 2% 18.4µs ± 2% -2.21% (p=0.000 n=9+10) GlyphNRGBA32Over-8 99.1µs ± 2% 95.9µs ± 3% -3.23% (p=0.000 n=10+10) GlyphNRGBA32Src-8 65.6µs ± 3% 62.8µs ± 2% -4.36% (p=0.000 n=10+10) GlyphNRGBA64Over-8 376µs ± 4% 370µs ± 2% ~ (p=0.063 n=10+10) GlyphNRGBA64Src-8 238µs ± 3% 233µs ± 1% -2.21% (p=0.000 n=9+10) GlyphNRGBA128Over-8 1.52ms ± 2% 1.48ms ± 0% -2.11% (p=0.000 n=10+8) GlyphNRGBA128Src-8 951µs ± 3% 935µs ± 1% -1.69% (p=0.013 n=10+9) GlyphNRGBA256Over-8 6.00ms ± 1% 5.87ms ± 3% -2.12% (p=0.002 n=10+10) GlyphNRGBA256Src-8 3.94ms ± 2% 3.80ms ± 2% -3.64% (p=0.000 n=10+10) A comparison of the non-SIMD and SIMD versions: name time/op FixedAccumulateMask16-8 237ns ± 0% FixedAccumulateMaskSIMD16-8 80.0ns ± 1% FloatingAccumulateMask16-8 413ns ± 2% FloatingAccumulateMaskSIMD16-8 166ns ± 0% FixedAccumulateMask64-8 3.42µs ± 0% FixedAccumulateMaskSIMD64-8 1.09µs ± 0% FloatingAccumulateMask64-8 6.92µs ± 0% FloatingAccumulateMaskSIMD64-8 2.47µs ± 1% Change-Id: Ib6980e5975ed2842ff2a372f76dd5f2e95c5526c Reviewed-on: https://go-review.googlesource.com/30898 Reviewed-by: David Crawshaw <crawshaw@golang.org>
2016-10-13 05:50:52 +02:00
// Prepare to set MXCSR bits 13 and 14, so that the CVTPS2PL below is
// "Round To Zero".
STMXCSR mxcsrOrig-8(SP)
MOVL mxcsrOrig-8(SP), AX
ORL $0x6000, AX
MOVL AX, mxcsrNew-4(SP)
// flSignMask := XMM(0x7fffffff repeated four times) // All but the sign bit of a float32.
// flOne := XMM(0x3f800000 repeated four times) // 1 as a float32.
// flAlmost65536 := XMM(0x477fffff repeated four times) // 255.99998 * 256 as a float32.
MOVOU flSignMask<>(SB), X3
MOVOU flOne<>(SB), X4
MOVOU flAlmost65536<>(SB), X5
vector: add SIMD versions of xxxAccumulateMask. name old time/op new time/op delta GlyphAlphaLoose16Over-8 3.96µs ± 0% 3.64µs ± 1% -8.08% (p=0.000 n=8+10) GlyphAlphaLoose16Src-8 3.64µs ± 0% 3.35µs ± 0% -7.88% (p=0.000 n=8+9) GlyphAlphaLoose32Over-8 8.45µs ± 0% 6.74µs ± 0% -20.22% (p=0.000 n=8+9) GlyphAlphaLoose32Src-8 7.24µs ± 0% 5.54µs ± 1% -23.48% (p=0.000 n=8+10) GlyphAlphaLoose64Over-8 22.2µs ± 0% 17.4µs ± 0% -21.67% (p=0.000 n=9+9) GlyphAlphaLoose64Src-8 17.6µs ± 1% 12.2µs ± 1% -30.32% (p=0.000 n=10+10) GlyphAlphaLoose128Over-8 67.9µs ± 0% 53.3µs ± 1% -21.53% (p=0.000 n=10+10) GlyphAlphaLoose128Src-8 48.2µs ± 0% 32.6µs ± 2% -32.41% (p=0.000 n=9+10) GlyphAlphaLoose256Over-8 242µs ± 1% 187µs ± 1% -22.96% (p=0.000 n=9+9) GlyphAlphaLoose256Src-8 163µs ± 0% 105µs ± 1% -35.83% (p=0.000 n=9+9) GlyphRGBA16Over-8 5.25µs ± 1% 4.95µs ± 0% -5.78% (p=0.000 n=9+9) GlyphRGBA16Src-8 4.72µs ± 0% 4.43µs ± 1% -6.22% (p=0.000 n=9+10) GlyphRGBA32Over-8 13.5µs ± 0% 11.9µs ± 1% -12.19% (p=0.000 n=9+10) GlyphRGBA32Src-8 11.5µs ± 1% 9.8µs ± 0% -14.72% (p=0.000 n=9+9) GlyphRGBA64Over-8 42.0µs ± 2% 36.9µs ± 1% -12.19% (p=0.000 n=10+10) GlyphRGBA64Src-8 34.1µs ± 1% 28.5µs ± 0% -16.25% (p=0.000 n=9+7) GlyphRGBA128Over-8 149µs ± 2% 133µs ± 1% -10.24% (p=0.000 n=10+9) GlyphRGBA128Src-8 115µs ± 1% 99µs ± 1% -13.57% (p=0.000 n=9+10) GlyphRGBA256Over-8 566µs ± 0% 511µs ± 1% -9.85% (p=0.000 n=9+10) GlyphRGBA256Src-8 435µs ± 0% 372µs ± 0% -14.64% (p=0.000 n=9+8) GlyphNRGBA16Over-8 26.9µs ± 3% 26.0µs ± 3% -3.55% (p=0.000 n=10+9) GlyphNRGBA16Src-8 18.8µs ± 2% 18.4µs ± 2% -2.21% (p=0.000 n=9+10) GlyphNRGBA32Over-8 99.1µs ± 2% 95.9µs ± 3% -3.23% (p=0.000 n=10+10) GlyphNRGBA32Src-8 65.6µs ± 3% 62.8µs ± 2% -4.36% (p=0.000 n=10+10) GlyphNRGBA64Over-8 376µs ± 4% 370µs ± 2% ~ (p=0.063 n=10+10) GlyphNRGBA64Src-8 238µs ± 3% 233µs ± 1% -2.21% (p=0.000 n=9+10) GlyphNRGBA128Over-8 1.52ms ± 2% 1.48ms ± 0% -2.11% (p=0.000 n=10+8) GlyphNRGBA128Src-8 951µs ± 3% 935µs ± 1% -1.69% (p=0.013 n=10+9) GlyphNRGBA256Over-8 6.00ms ± 1% 5.87ms ± 3% -2.12% (p=0.002 n=10+10) GlyphNRGBA256Src-8 3.94ms ± 2% 3.80ms ± 2% -3.64% (p=0.000 n=10+10) A comparison of the non-SIMD and SIMD versions: name time/op FixedAccumulateMask16-8 237ns ± 0% FixedAccumulateMaskSIMD16-8 80.0ns ± 1% FloatingAccumulateMask16-8 413ns ± 2% FloatingAccumulateMaskSIMD16-8 166ns ± 0% FixedAccumulateMask64-8 3.42µs ± 0% FixedAccumulateMaskSIMD64-8 1.09µs ± 0% FloatingAccumulateMask64-8 6.92µs ± 0% FloatingAccumulateMaskSIMD64-8 2.47µs ± 1% Change-Id: Ib6980e5975ed2842ff2a372f76dd5f2e95c5526c Reviewed-on: https://go-review.googlesource.com/30898 Reviewed-by: David Crawshaw <crawshaw@golang.org>
2016-10-13 05:50:52 +02:00
// gather := XMM(see above) // PSHUFB shuffle mask.
// scatterAndMulBy0x101 := XMM(see above) // PSHUFB shuffle mask.
// fxAlmost65536 := XMM(0x0000ffff repeated four times) // 0xffff.
// inverseFFFF := XMM(0x80008001 repeated four times) // Magic constant for dividing by 0xffff.
vector: add SIMD versions of xxxAccumulateMask. name old time/op new time/op delta GlyphAlphaLoose16Over-8 3.96µs ± 0% 3.64µs ± 1% -8.08% (p=0.000 n=8+10) GlyphAlphaLoose16Src-8 3.64µs ± 0% 3.35µs ± 0% -7.88% (p=0.000 n=8+9) GlyphAlphaLoose32Over-8 8.45µs ± 0% 6.74µs ± 0% -20.22% (p=0.000 n=8+9) GlyphAlphaLoose32Src-8 7.24µs ± 0% 5.54µs ± 1% -23.48% (p=0.000 n=8+10) GlyphAlphaLoose64Over-8 22.2µs ± 0% 17.4µs ± 0% -21.67% (p=0.000 n=9+9) GlyphAlphaLoose64Src-8 17.6µs ± 1% 12.2µs ± 1% -30.32% (p=0.000 n=10+10) GlyphAlphaLoose128Over-8 67.9µs ± 0% 53.3µs ± 1% -21.53% (p=0.000 n=10+10) GlyphAlphaLoose128Src-8 48.2µs ± 0% 32.6µs ± 2% -32.41% (p=0.000 n=9+10) GlyphAlphaLoose256Over-8 242µs ± 1% 187µs ± 1% -22.96% (p=0.000 n=9+9) GlyphAlphaLoose256Src-8 163µs ± 0% 105µs ± 1% -35.83% (p=0.000 n=9+9) GlyphRGBA16Over-8 5.25µs ± 1% 4.95µs ± 0% -5.78% (p=0.000 n=9+9) GlyphRGBA16Src-8 4.72µs ± 0% 4.43µs ± 1% -6.22% (p=0.000 n=9+10) GlyphRGBA32Over-8 13.5µs ± 0% 11.9µs ± 1% -12.19% (p=0.000 n=9+10) GlyphRGBA32Src-8 11.5µs ± 1% 9.8µs ± 0% -14.72% (p=0.000 n=9+9) GlyphRGBA64Over-8 42.0µs ± 2% 36.9µs ± 1% -12.19% (p=0.000 n=10+10) GlyphRGBA64Src-8 34.1µs ± 1% 28.5µs ± 0% -16.25% (p=0.000 n=9+7) GlyphRGBA128Over-8 149µs ± 2% 133µs ± 1% -10.24% (p=0.000 n=10+9) GlyphRGBA128Src-8 115µs ± 1% 99µs ± 1% -13.57% (p=0.000 n=9+10) GlyphRGBA256Over-8 566µs ± 0% 511µs ± 1% -9.85% (p=0.000 n=9+10) GlyphRGBA256Src-8 435µs ± 0% 372µs ± 0% -14.64% (p=0.000 n=9+8) GlyphNRGBA16Over-8 26.9µs ± 3% 26.0µs ± 3% -3.55% (p=0.000 n=10+9) GlyphNRGBA16Src-8 18.8µs ± 2% 18.4µs ± 2% -2.21% (p=0.000 n=9+10) GlyphNRGBA32Over-8 99.1µs ± 2% 95.9µs ± 3% -3.23% (p=0.000 n=10+10) GlyphNRGBA32Src-8 65.6µs ± 3% 62.8µs ± 2% -4.36% (p=0.000 n=10+10) GlyphNRGBA64Over-8 376µs ± 4% 370µs ± 2% ~ (p=0.063 n=10+10) GlyphNRGBA64Src-8 238µs ± 3% 233µs ± 1% -2.21% (p=0.000 n=9+10) GlyphNRGBA128Over-8 1.52ms ± 2% 1.48ms ± 0% -2.11% (p=0.000 n=10+8) GlyphNRGBA128Src-8 951µs ± 3% 935µs ± 1% -1.69% (p=0.013 n=10+9) GlyphNRGBA256Over-8 6.00ms ± 1% 5.87ms ± 3% -2.12% (p=0.002 n=10+10) GlyphNRGBA256Src-8 3.94ms ± 2% 3.80ms ± 2% -3.64% (p=0.000 n=10+10) A comparison of the non-SIMD and SIMD versions: name time/op FixedAccumulateMask16-8 237ns ± 0% FixedAccumulateMaskSIMD16-8 80.0ns ± 1% FloatingAccumulateMask16-8 413ns ± 2% FloatingAccumulateMaskSIMD16-8 166ns ± 0% FixedAccumulateMask64-8 3.42µs ± 0% FixedAccumulateMaskSIMD64-8 1.09µs ± 0% FloatingAccumulateMask64-8 6.92µs ± 0% FloatingAccumulateMaskSIMD64-8 2.47µs ± 1% Change-Id: Ib6980e5975ed2842ff2a372f76dd5f2e95c5526c Reviewed-on: https://go-review.googlesource.com/30898 Reviewed-by: David Crawshaw <crawshaw@golang.org>
2016-10-13 05:50:52 +02:00
MOVOU gather<>(SB), X6
MOVOU scatterAndMulBy0x101<>(SB), X8
MOVOU fxAlmost65536<>(SB), X9
MOVOU inverseFFFF<>(SB), X10
// offset := XMM(0x00000000 repeated four times) // Cumulative sum.
XORPS X7, X7
// i := 0
MOVQ $0, R9
flAccOpOverLoop4:
// for i < (len(src) &^ 3)
CMPQ R9, R10
JAE flAccOpOverLoop1
// x = XMM(s0, s1, s2, s3)
//
// Where s0 is src[i+0], s1 is src[i+1], etc.
MOVOU (SI), X1
// scratch = XMM(0, s0, s1, s2)
// x += scratch // yields x == XMM(s0, s0+s1, s1+s2, s2+s3)
MOVOU X1, X0
PSLLO $4, X0
ADDPS X0, X1
// scratch = XMM(0, 0, 0, 0)
// scratch = XMM(scratch@0, scratch@0, x@0, x@1) // yields scratch == XMM(0, 0, s0, s0+s1)
// x += scratch // yields x == XMM(s0, s0+s1, s0+s1+s2, s0+s1+s2+s3)
XORPS X0, X0
SHUFPS $0x40, X1, X0
ADDPS X0, X1
// x += offset
ADDPS X7, X1
// y = x & flSignMask
// y = min(y, flOne)
// y = mul(y, flAlmost65536)
MOVOU X3, X2
ANDPS X1, X2
MINPS X4, X2
MULPS X5, X2
// z = convertToInt32(y)
vector: add SIMD versions of xxxAccumulateMask. name old time/op new time/op delta GlyphAlphaLoose16Over-8 3.96µs ± 0% 3.64µs ± 1% -8.08% (p=0.000 n=8+10) GlyphAlphaLoose16Src-8 3.64µs ± 0% 3.35µs ± 0% -7.88% (p=0.000 n=8+9) GlyphAlphaLoose32Over-8 8.45µs ± 0% 6.74µs ± 0% -20.22% (p=0.000 n=8+9) GlyphAlphaLoose32Src-8 7.24µs ± 0% 5.54µs ± 1% -23.48% (p=0.000 n=8+10) GlyphAlphaLoose64Over-8 22.2µs ± 0% 17.4µs ± 0% -21.67% (p=0.000 n=9+9) GlyphAlphaLoose64Src-8 17.6µs ± 1% 12.2µs ± 1% -30.32% (p=0.000 n=10+10) GlyphAlphaLoose128Over-8 67.9µs ± 0% 53.3µs ± 1% -21.53% (p=0.000 n=10+10) GlyphAlphaLoose128Src-8 48.2µs ± 0% 32.6µs ± 2% -32.41% (p=0.000 n=9+10) GlyphAlphaLoose256Over-8 242µs ± 1% 187µs ± 1% -22.96% (p=0.000 n=9+9) GlyphAlphaLoose256Src-8 163µs ± 0% 105µs ± 1% -35.83% (p=0.000 n=9+9) GlyphRGBA16Over-8 5.25µs ± 1% 4.95µs ± 0% -5.78% (p=0.000 n=9+9) GlyphRGBA16Src-8 4.72µs ± 0% 4.43µs ± 1% -6.22% (p=0.000 n=9+10) GlyphRGBA32Over-8 13.5µs ± 0% 11.9µs ± 1% -12.19% (p=0.000 n=9+10) GlyphRGBA32Src-8 11.5µs ± 1% 9.8µs ± 0% -14.72% (p=0.000 n=9+9) GlyphRGBA64Over-8 42.0µs ± 2% 36.9µs ± 1% -12.19% (p=0.000 n=10+10) GlyphRGBA64Src-8 34.1µs ± 1% 28.5µs ± 0% -16.25% (p=0.000 n=9+7) GlyphRGBA128Over-8 149µs ± 2% 133µs ± 1% -10.24% (p=0.000 n=10+9) GlyphRGBA128Src-8 115µs ± 1% 99µs ± 1% -13.57% (p=0.000 n=9+10) GlyphRGBA256Over-8 566µs ± 0% 511µs ± 1% -9.85% (p=0.000 n=9+10) GlyphRGBA256Src-8 435µs ± 0% 372µs ± 0% -14.64% (p=0.000 n=9+8) GlyphNRGBA16Over-8 26.9µs ± 3% 26.0µs ± 3% -3.55% (p=0.000 n=10+9) GlyphNRGBA16Src-8 18.8µs ± 2% 18.4µs ± 2% -2.21% (p=0.000 n=9+10) GlyphNRGBA32Over-8 99.1µs ± 2% 95.9µs ± 3% -3.23% (p=0.000 n=10+10) GlyphNRGBA32Src-8 65.6µs ± 3% 62.8µs ± 2% -4.36% (p=0.000 n=10+10) GlyphNRGBA64Over-8 376µs ± 4% 370µs ± 2% ~ (p=0.063 n=10+10) GlyphNRGBA64Src-8 238µs ± 3% 233µs ± 1% -2.21% (p=0.000 n=9+10) GlyphNRGBA128Over-8 1.52ms ± 2% 1.48ms ± 0% -2.11% (p=0.000 n=10+8) GlyphNRGBA128Src-8 951µs ± 3% 935µs ± 1% -1.69% (p=0.013 n=10+9) GlyphNRGBA256Over-8 6.00ms ± 1% 5.87ms ± 3% -2.12% (p=0.002 n=10+10) GlyphNRGBA256Src-8 3.94ms ± 2% 3.80ms ± 2% -3.64% (p=0.000 n=10+10) A comparison of the non-SIMD and SIMD versions: name time/op FixedAccumulateMask16-8 237ns ± 0% FixedAccumulateMaskSIMD16-8 80.0ns ± 1% FloatingAccumulateMask16-8 413ns ± 2% FloatingAccumulateMaskSIMD16-8 166ns ± 0% FixedAccumulateMask64-8 3.42µs ± 0% FixedAccumulateMaskSIMD64-8 1.09µs ± 0% FloatingAccumulateMask64-8 6.92µs ± 0% FloatingAccumulateMaskSIMD64-8 2.47µs ± 1% Change-Id: Ib6980e5975ed2842ff2a372f76dd5f2e95c5526c Reviewed-on: https://go-review.googlesource.com/30898 Reviewed-by: David Crawshaw <crawshaw@golang.org>
2016-10-13 05:50:52 +02:00
LDMXCSR mxcsrNew-4(SP)
CVTPS2PL X2, X2
vector: add SIMD versions of xxxAccumulateMask. name old time/op new time/op delta GlyphAlphaLoose16Over-8 3.96µs ± 0% 3.64µs ± 1% -8.08% (p=0.000 n=8+10) GlyphAlphaLoose16Src-8 3.64µs ± 0% 3.35µs ± 0% -7.88% (p=0.000 n=8+9) GlyphAlphaLoose32Over-8 8.45µs ± 0% 6.74µs ± 0% -20.22% (p=0.000 n=8+9) GlyphAlphaLoose32Src-8 7.24µs ± 0% 5.54µs ± 1% -23.48% (p=0.000 n=8+10) GlyphAlphaLoose64Over-8 22.2µs ± 0% 17.4µs ± 0% -21.67% (p=0.000 n=9+9) GlyphAlphaLoose64Src-8 17.6µs ± 1% 12.2µs ± 1% -30.32% (p=0.000 n=10+10) GlyphAlphaLoose128Over-8 67.9µs ± 0% 53.3µs ± 1% -21.53% (p=0.000 n=10+10) GlyphAlphaLoose128Src-8 48.2µs ± 0% 32.6µs ± 2% -32.41% (p=0.000 n=9+10) GlyphAlphaLoose256Over-8 242µs ± 1% 187µs ± 1% -22.96% (p=0.000 n=9+9) GlyphAlphaLoose256Src-8 163µs ± 0% 105µs ± 1% -35.83% (p=0.000 n=9+9) GlyphRGBA16Over-8 5.25µs ± 1% 4.95µs ± 0% -5.78% (p=0.000 n=9+9) GlyphRGBA16Src-8 4.72µs ± 0% 4.43µs ± 1% -6.22% (p=0.000 n=9+10) GlyphRGBA32Over-8 13.5µs ± 0% 11.9µs ± 1% -12.19% (p=0.000 n=9+10) GlyphRGBA32Src-8 11.5µs ± 1% 9.8µs ± 0% -14.72% (p=0.000 n=9+9) GlyphRGBA64Over-8 42.0µs ± 2% 36.9µs ± 1% -12.19% (p=0.000 n=10+10) GlyphRGBA64Src-8 34.1µs ± 1% 28.5µs ± 0% -16.25% (p=0.000 n=9+7) GlyphRGBA128Over-8 149µs ± 2% 133µs ± 1% -10.24% (p=0.000 n=10+9) GlyphRGBA128Src-8 115µs ± 1% 99µs ± 1% -13.57% (p=0.000 n=9+10) GlyphRGBA256Over-8 566µs ± 0% 511µs ± 1% -9.85% (p=0.000 n=9+10) GlyphRGBA256Src-8 435µs ± 0% 372µs ± 0% -14.64% (p=0.000 n=9+8) GlyphNRGBA16Over-8 26.9µs ± 3% 26.0µs ± 3% -3.55% (p=0.000 n=10+9) GlyphNRGBA16Src-8 18.8µs ± 2% 18.4µs ± 2% -2.21% (p=0.000 n=9+10) GlyphNRGBA32Over-8 99.1µs ± 2% 95.9µs ± 3% -3.23% (p=0.000 n=10+10) GlyphNRGBA32Src-8 65.6µs ± 3% 62.8µs ± 2% -4.36% (p=0.000 n=10+10) GlyphNRGBA64Over-8 376µs ± 4% 370µs ± 2% ~ (p=0.063 n=10+10) GlyphNRGBA64Src-8 238µs ± 3% 233µs ± 1% -2.21% (p=0.000 n=9+10) GlyphNRGBA128Over-8 1.52ms ± 2% 1.48ms ± 0% -2.11% (p=0.000 n=10+8) GlyphNRGBA128Src-8 951µs ± 3% 935µs ± 1% -1.69% (p=0.013 n=10+9) GlyphNRGBA256Over-8 6.00ms ± 1% 5.87ms ± 3% -2.12% (p=0.002 n=10+10) GlyphNRGBA256Src-8 3.94ms ± 2% 3.80ms ± 2% -3.64% (p=0.000 n=10+10) A comparison of the non-SIMD and SIMD versions: name time/op FixedAccumulateMask16-8 237ns ± 0% FixedAccumulateMaskSIMD16-8 80.0ns ± 1% FloatingAccumulateMask16-8 413ns ± 2% FloatingAccumulateMaskSIMD16-8 166ns ± 0% FixedAccumulateMask64-8 3.42µs ± 0% FixedAccumulateMaskSIMD64-8 1.09µs ± 0% FloatingAccumulateMask64-8 6.92µs ± 0% FloatingAccumulateMaskSIMD64-8 2.47µs ± 1% Change-Id: Ib6980e5975ed2842ff2a372f76dd5f2e95c5526c Reviewed-on: https://go-review.googlesource.com/30898 Reviewed-by: David Crawshaw <crawshaw@golang.org>
2016-10-13 05:50:52 +02:00
LDMXCSR mxcsrOrig-8(SP)
// Blend over the dst's prior value. SIMD for i in 0..3:
//
// dstA := uint32(dst[i]) * 0x101
// maskA := z@i
// outA := dstA*(0xffff-maskA)/0xffff + maskA
// dst[i] = uint8(outA >> 8)
//
// First, set X0 to dstA*(0xfff-maskA).
MOVL (DI), X0
PSHUFB X8, X0
MOVOU X9, X11
PSUBL X2, X11
PMULLD X11, X0
// We implement uint32 division by 0xffff as multiplication by a magic
// constant (0x800080001) and then a shift by a magic constant (47).
// See TestDivideByFFFF for a justification.
//
// That multiplication widens from uint32 to uint64, so we have to
// duplicate and shift our four uint32s from one XMM register (X0) to
// two XMM registers (X0 and X11).
//
// Move the second and fourth uint32s in X0 to be the first and third
// uint32s in X11.
MOVOU X0, X11
PSRLQ $32, X11
// Multiply by magic, shift by magic.
PMULULQ X10, X0
PMULULQ X10, X11
PSRLQ $47, X0
PSRLQ $47, X11
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>
2016-10-20 06:15:53 +02:00
// Merge the two registers back to one, X11, and add maskA.
PSLLQ $32, X11
XORPS X0, X11
PADDD X11, X2
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>
2016-10-20 06:15:53 +02:00
// As per opSrcStore4, shuffle and copy the 4 second-lowest bytes.
PSHUFB X6, X2
MOVL X2, (DI)
// offset = XMM(x@3, x@3, x@3, x@3)
MOVOU X1, X7
SHUFPS $0xff, X1, X7
// i += 4
// dst = dst[4:]
// src = src[4:]
ADDQ $4, R9
ADDQ $4, DI
ADDQ $16, SI
JMP flAccOpOverLoop4
flAccOpOverLoop1:
// for i < len(src)
CMPQ R9, R11
vector: add SIMD versions of xxxAccumulateMask. name old time/op new time/op delta GlyphAlphaLoose16Over-8 3.96µs ± 0% 3.64µs ± 1% -8.08% (p=0.000 n=8+10) GlyphAlphaLoose16Src-8 3.64µs ± 0% 3.35µs ± 0% -7.88% (p=0.000 n=8+9) GlyphAlphaLoose32Over-8 8.45µs ± 0% 6.74µs ± 0% -20.22% (p=0.000 n=8+9) GlyphAlphaLoose32Src-8 7.24µs ± 0% 5.54µs ± 1% -23.48% (p=0.000 n=8+10) GlyphAlphaLoose64Over-8 22.2µs ± 0% 17.4µs ± 0% -21.67% (p=0.000 n=9+9) GlyphAlphaLoose64Src-8 17.6µs ± 1% 12.2µs ± 1% -30.32% (p=0.000 n=10+10) GlyphAlphaLoose128Over-8 67.9µs ± 0% 53.3µs ± 1% -21.53% (p=0.000 n=10+10) GlyphAlphaLoose128Src-8 48.2µs ± 0% 32.6µs ± 2% -32.41% (p=0.000 n=9+10) GlyphAlphaLoose256Over-8 242µs ± 1% 187µs ± 1% -22.96% (p=0.000 n=9+9) GlyphAlphaLoose256Src-8 163µs ± 0% 105µs ± 1% -35.83% (p=0.000 n=9+9) GlyphRGBA16Over-8 5.25µs ± 1% 4.95µs ± 0% -5.78% (p=0.000 n=9+9) GlyphRGBA16Src-8 4.72µs ± 0% 4.43µs ± 1% -6.22% (p=0.000 n=9+10) GlyphRGBA32Over-8 13.5µs ± 0% 11.9µs ± 1% -12.19% (p=0.000 n=9+10) GlyphRGBA32Src-8 11.5µs ± 1% 9.8µs ± 0% -14.72% (p=0.000 n=9+9) GlyphRGBA64Over-8 42.0µs ± 2% 36.9µs ± 1% -12.19% (p=0.000 n=10+10) GlyphRGBA64Src-8 34.1µs ± 1% 28.5µs ± 0% -16.25% (p=0.000 n=9+7) GlyphRGBA128Over-8 149µs ± 2% 133µs ± 1% -10.24% (p=0.000 n=10+9) GlyphRGBA128Src-8 115µs ± 1% 99µs ± 1% -13.57% (p=0.000 n=9+10) GlyphRGBA256Over-8 566µs ± 0% 511µs ± 1% -9.85% (p=0.000 n=9+10) GlyphRGBA256Src-8 435µs ± 0% 372µs ± 0% -14.64% (p=0.000 n=9+8) GlyphNRGBA16Over-8 26.9µs ± 3% 26.0µs ± 3% -3.55% (p=0.000 n=10+9) GlyphNRGBA16Src-8 18.8µs ± 2% 18.4µs ± 2% -2.21% (p=0.000 n=9+10) GlyphNRGBA32Over-8 99.1µs ± 2% 95.9µs ± 3% -3.23% (p=0.000 n=10+10) GlyphNRGBA32Src-8 65.6µs ± 3% 62.8µs ± 2% -4.36% (p=0.000 n=10+10) GlyphNRGBA64Over-8 376µs ± 4% 370µs ± 2% ~ (p=0.063 n=10+10) GlyphNRGBA64Src-8 238µs ± 3% 233µs ± 1% -2.21% (p=0.000 n=9+10) GlyphNRGBA128Over-8 1.52ms ± 2% 1.48ms ± 0% -2.11% (p=0.000 n=10+8) GlyphNRGBA128Src-8 951µs ± 3% 935µs ± 1% -1.69% (p=0.013 n=10+9) GlyphNRGBA256Over-8 6.00ms ± 1% 5.87ms ± 3% -2.12% (p=0.002 n=10+10) GlyphNRGBA256Src-8 3.94ms ± 2% 3.80ms ± 2% -3.64% (p=0.000 n=10+10) A comparison of the non-SIMD and SIMD versions: name time/op FixedAccumulateMask16-8 237ns ± 0% FixedAccumulateMaskSIMD16-8 80.0ns ± 1% FloatingAccumulateMask16-8 413ns ± 2% FloatingAccumulateMaskSIMD16-8 166ns ± 0% FixedAccumulateMask64-8 3.42µs ± 0% FixedAccumulateMaskSIMD64-8 1.09µs ± 0% FloatingAccumulateMask64-8 6.92µs ± 0% FloatingAccumulateMaskSIMD64-8 2.47µs ± 1% Change-Id: Ib6980e5975ed2842ff2a372f76dd5f2e95c5526c Reviewed-on: https://go-review.googlesource.com/30898 Reviewed-by: David Crawshaw <crawshaw@golang.org>
2016-10-13 05:50:52 +02:00
JAE flAccOpOverEnd
// x = src[i] + offset
MOVL (SI), X1
ADDPS X7, X1
// y = x & flSignMask
// y = min(y, flOne)
// y = mul(y, flAlmost65536)
MOVOU X3, X2
ANDPS X1, X2
MINPS X4, X2
MULPS X5, X2
// z = convertToInt32(y)
vector: add SIMD versions of xxxAccumulateMask. name old time/op new time/op delta GlyphAlphaLoose16Over-8 3.96µs ± 0% 3.64µs ± 1% -8.08% (p=0.000 n=8+10) GlyphAlphaLoose16Src-8 3.64µs ± 0% 3.35µs ± 0% -7.88% (p=0.000 n=8+9) GlyphAlphaLoose32Over-8 8.45µs ± 0% 6.74µs ± 0% -20.22% (p=0.000 n=8+9) GlyphAlphaLoose32Src-8 7.24µs ± 0% 5.54µs ± 1% -23.48% (p=0.000 n=8+10) GlyphAlphaLoose64Over-8 22.2µs ± 0% 17.4µs ± 0% -21.67% (p=0.000 n=9+9) GlyphAlphaLoose64Src-8 17.6µs ± 1% 12.2µs ± 1% -30.32% (p=0.000 n=10+10) GlyphAlphaLoose128Over-8 67.9µs ± 0% 53.3µs ± 1% -21.53% (p=0.000 n=10+10) GlyphAlphaLoose128Src-8 48.2µs ± 0% 32.6µs ± 2% -32.41% (p=0.000 n=9+10) GlyphAlphaLoose256Over-8 242µs ± 1% 187µs ± 1% -22.96% (p=0.000 n=9+9) GlyphAlphaLoose256Src-8 163µs ± 0% 105µs ± 1% -35.83% (p=0.000 n=9+9) GlyphRGBA16Over-8 5.25µs ± 1% 4.95µs ± 0% -5.78% (p=0.000 n=9+9) GlyphRGBA16Src-8 4.72µs ± 0% 4.43µs ± 1% -6.22% (p=0.000 n=9+10) GlyphRGBA32Over-8 13.5µs ± 0% 11.9µs ± 1% -12.19% (p=0.000 n=9+10) GlyphRGBA32Src-8 11.5µs ± 1% 9.8µs ± 0% -14.72% (p=0.000 n=9+9) GlyphRGBA64Over-8 42.0µs ± 2% 36.9µs ± 1% -12.19% (p=0.000 n=10+10) GlyphRGBA64Src-8 34.1µs ± 1% 28.5µs ± 0% -16.25% (p=0.000 n=9+7) GlyphRGBA128Over-8 149µs ± 2% 133µs ± 1% -10.24% (p=0.000 n=10+9) GlyphRGBA128Src-8 115µs ± 1% 99µs ± 1% -13.57% (p=0.000 n=9+10) GlyphRGBA256Over-8 566µs ± 0% 511µs ± 1% -9.85% (p=0.000 n=9+10) GlyphRGBA256Src-8 435µs ± 0% 372µs ± 0% -14.64% (p=0.000 n=9+8) GlyphNRGBA16Over-8 26.9µs ± 3% 26.0µs ± 3% -3.55% (p=0.000 n=10+9) GlyphNRGBA16Src-8 18.8µs ± 2% 18.4µs ± 2% -2.21% (p=0.000 n=9+10) GlyphNRGBA32Over-8 99.1µs ± 2% 95.9µs ± 3% -3.23% (p=0.000 n=10+10) GlyphNRGBA32Src-8 65.6µs ± 3% 62.8µs ± 2% -4.36% (p=0.000 n=10+10) GlyphNRGBA64Over-8 376µs ± 4% 370µs ± 2% ~ (p=0.063 n=10+10) GlyphNRGBA64Src-8 238µs ± 3% 233µs ± 1% -2.21% (p=0.000 n=9+10) GlyphNRGBA128Over-8 1.52ms ± 2% 1.48ms ± 0% -2.11% (p=0.000 n=10+8) GlyphNRGBA128Src-8 951µs ± 3% 935µs ± 1% -1.69% (p=0.013 n=10+9) GlyphNRGBA256Over-8 6.00ms ± 1% 5.87ms ± 3% -2.12% (p=0.002 n=10+10) GlyphNRGBA256Src-8 3.94ms ± 2% 3.80ms ± 2% -3.64% (p=0.000 n=10+10) A comparison of the non-SIMD and SIMD versions: name time/op FixedAccumulateMask16-8 237ns ± 0% FixedAccumulateMaskSIMD16-8 80.0ns ± 1% FloatingAccumulateMask16-8 413ns ± 2% FloatingAccumulateMaskSIMD16-8 166ns ± 0% FixedAccumulateMask64-8 3.42µs ± 0% FixedAccumulateMaskSIMD64-8 1.09µs ± 0% FloatingAccumulateMask64-8 6.92µs ± 0% FloatingAccumulateMaskSIMD64-8 2.47µs ± 1% Change-Id: Ib6980e5975ed2842ff2a372f76dd5f2e95c5526c Reviewed-on: https://go-review.googlesource.com/30898 Reviewed-by: David Crawshaw <crawshaw@golang.org>
2016-10-13 05:50:52 +02:00
LDMXCSR mxcsrNew-4(SP)
CVTPS2PL X2, X2
vector: add SIMD versions of xxxAccumulateMask. name old time/op new time/op delta GlyphAlphaLoose16Over-8 3.96µs ± 0% 3.64µs ± 1% -8.08% (p=0.000 n=8+10) GlyphAlphaLoose16Src-8 3.64µs ± 0% 3.35µs ± 0% -7.88% (p=0.000 n=8+9) GlyphAlphaLoose32Over-8 8.45µs ± 0% 6.74µs ± 0% -20.22% (p=0.000 n=8+9) GlyphAlphaLoose32Src-8 7.24µs ± 0% 5.54µs ± 1% -23.48% (p=0.000 n=8+10) GlyphAlphaLoose64Over-8 22.2µs ± 0% 17.4µs ± 0% -21.67% (p=0.000 n=9+9) GlyphAlphaLoose64Src-8 17.6µs ± 1% 12.2µs ± 1% -30.32% (p=0.000 n=10+10) GlyphAlphaLoose128Over-8 67.9µs ± 0% 53.3µs ± 1% -21.53% (p=0.000 n=10+10) GlyphAlphaLoose128Src-8 48.2µs ± 0% 32.6µs ± 2% -32.41% (p=0.000 n=9+10) GlyphAlphaLoose256Over-8 242µs ± 1% 187µs ± 1% -22.96% (p=0.000 n=9+9) GlyphAlphaLoose256Src-8 163µs ± 0% 105µs ± 1% -35.83% (p=0.000 n=9+9) GlyphRGBA16Over-8 5.25µs ± 1% 4.95µs ± 0% -5.78% (p=0.000 n=9+9) GlyphRGBA16Src-8 4.72µs ± 0% 4.43µs ± 1% -6.22% (p=0.000 n=9+10) GlyphRGBA32Over-8 13.5µs ± 0% 11.9µs ± 1% -12.19% (p=0.000 n=9+10) GlyphRGBA32Src-8 11.5µs ± 1% 9.8µs ± 0% -14.72% (p=0.000 n=9+9) GlyphRGBA64Over-8 42.0µs ± 2% 36.9µs ± 1% -12.19% (p=0.000 n=10+10) GlyphRGBA64Src-8 34.1µs ± 1% 28.5µs ± 0% -16.25% (p=0.000 n=9+7) GlyphRGBA128Over-8 149µs ± 2% 133µs ± 1% -10.24% (p=0.000 n=10+9) GlyphRGBA128Src-8 115µs ± 1% 99µs ± 1% -13.57% (p=0.000 n=9+10) GlyphRGBA256Over-8 566µs ± 0% 511µs ± 1% -9.85% (p=0.000 n=9+10) GlyphRGBA256Src-8 435µs ± 0% 372µs ± 0% -14.64% (p=0.000 n=9+8) GlyphNRGBA16Over-8 26.9µs ± 3% 26.0µs ± 3% -3.55% (p=0.000 n=10+9) GlyphNRGBA16Src-8 18.8µs ± 2% 18.4µs ± 2% -2.21% (p=0.000 n=9+10) GlyphNRGBA32Over-8 99.1µs ± 2% 95.9µs ± 3% -3.23% (p=0.000 n=10+10) GlyphNRGBA32Src-8 65.6µs ± 3% 62.8µs ± 2% -4.36% (p=0.000 n=10+10) GlyphNRGBA64Over-8 376µs ± 4% 370µs ± 2% ~ (p=0.063 n=10+10) GlyphNRGBA64Src-8 238µs ± 3% 233µs ± 1% -2.21% (p=0.000 n=9+10) GlyphNRGBA128Over-8 1.52ms ± 2% 1.48ms ± 0% -2.11% (p=0.000 n=10+8) GlyphNRGBA128Src-8 951µs ± 3% 935µs ± 1% -1.69% (p=0.013 n=10+9) GlyphNRGBA256Over-8 6.00ms ± 1% 5.87ms ± 3% -2.12% (p=0.002 n=10+10) GlyphNRGBA256Src-8 3.94ms ± 2% 3.80ms ± 2% -3.64% (p=0.000 n=10+10) A comparison of the non-SIMD and SIMD versions: name time/op FixedAccumulateMask16-8 237ns ± 0% FixedAccumulateMaskSIMD16-8 80.0ns ± 1% FloatingAccumulateMask16-8 413ns ± 2% FloatingAccumulateMaskSIMD16-8 166ns ± 0% FixedAccumulateMask64-8 3.42µs ± 0% FixedAccumulateMaskSIMD64-8 1.09µs ± 0% FloatingAccumulateMask64-8 6.92µs ± 0% FloatingAccumulateMaskSIMD64-8 2.47µs ± 1% Change-Id: Ib6980e5975ed2842ff2a372f76dd5f2e95c5526c Reviewed-on: https://go-review.googlesource.com/30898 Reviewed-by: David Crawshaw <crawshaw@golang.org>
2016-10-13 05:50:52 +02:00
LDMXCSR mxcsrOrig-8(SP)
// Blend over the dst's prior value.
//
// dstA := uint32(dst[0]) * 0x101
// maskA := z
// outA := dstA*(0xffff-maskA)/0xffff + maskA
// dst[0] = uint8(outA >> 8)
MOVBLZX (DI), R12
IMULL $0x101, R12
MOVL X2, R13
MOVL $0xffff, AX
SUBL R13, AX
MULL R12 // MULL's implicit arg is AX, and the result is stored in DX:AX.
MOVL $0x80008001, BX // Divide by 0xffff is to first multiply by a magic constant...
MULL BX // MULL's implicit arg is AX, and the result is stored in DX:AX.
SHRL $15, DX // ...and then shift by another magic constant (47 - 32 = 15).
ADDL DX, R13
SHRL $8, R13
MOVB R13, (DI)
// offset = x
MOVOU X1, X7
// i += 1
// dst = dst[1:]
// src = src[1:]
ADDQ $1, R9
ADDQ $1, DI
ADDQ $4, SI
JMP flAccOpOverLoop1
flAccOpOverEnd:
RET
// ----------------------------------------------------------------------------
// func floatingAccumulateOpSrcSIMD(dst []uint8, src []float32)
//
// XMM registers. Variable names are per
// https://github.com/google/font-rs/blob/master/src/accumulate.c
//
// xmm0 scratch
// xmm1 x
// xmm2 y, z
// xmm3 flSignMask
// xmm4 flOne
// xmm5 flAlmost65536
// xmm6 gather
// xmm7 offset
// xmm8 -
// xmm9 -
// xmm10 -
TEXT ·floatingAccumulateOpSrcSIMD(SB), NOSPLIT, $8-48
vector: add SIMD versions of xxxAccumulateMask. name old time/op new time/op delta GlyphAlphaLoose16Over-8 3.96µs ± 0% 3.64µs ± 1% -8.08% (p=0.000 n=8+10) GlyphAlphaLoose16Src-8 3.64µs ± 0% 3.35µs ± 0% -7.88% (p=0.000 n=8+9) GlyphAlphaLoose32Over-8 8.45µs ± 0% 6.74µs ± 0% -20.22% (p=0.000 n=8+9) GlyphAlphaLoose32Src-8 7.24µs ± 0% 5.54µs ± 1% -23.48% (p=0.000 n=8+10) GlyphAlphaLoose64Over-8 22.2µs ± 0% 17.4µs ± 0% -21.67% (p=0.000 n=9+9) GlyphAlphaLoose64Src-8 17.6µs ± 1% 12.2µs ± 1% -30.32% (p=0.000 n=10+10) GlyphAlphaLoose128Over-8 67.9µs ± 0% 53.3µs ± 1% -21.53% (p=0.000 n=10+10) GlyphAlphaLoose128Src-8 48.2µs ± 0% 32.6µs ± 2% -32.41% (p=0.000 n=9+10) GlyphAlphaLoose256Over-8 242µs ± 1% 187µs ± 1% -22.96% (p=0.000 n=9+9) GlyphAlphaLoose256Src-8 163µs ± 0% 105µs ± 1% -35.83% (p=0.000 n=9+9) GlyphRGBA16Over-8 5.25µs ± 1% 4.95µs ± 0% -5.78% (p=0.000 n=9+9) GlyphRGBA16Src-8 4.72µs ± 0% 4.43µs ± 1% -6.22% (p=0.000 n=9+10) GlyphRGBA32Over-8 13.5µs ± 0% 11.9µs ± 1% -12.19% (p=0.000 n=9+10) GlyphRGBA32Src-8 11.5µs ± 1% 9.8µs ± 0% -14.72% (p=0.000 n=9+9) GlyphRGBA64Over-8 42.0µs ± 2% 36.9µs ± 1% -12.19% (p=0.000 n=10+10) GlyphRGBA64Src-8 34.1µs ± 1% 28.5µs ± 0% -16.25% (p=0.000 n=9+7) GlyphRGBA128Over-8 149µs ± 2% 133µs ± 1% -10.24% (p=0.000 n=10+9) GlyphRGBA128Src-8 115µs ± 1% 99µs ± 1% -13.57% (p=0.000 n=9+10) GlyphRGBA256Over-8 566µs ± 0% 511µs ± 1% -9.85% (p=0.000 n=9+10) GlyphRGBA256Src-8 435µs ± 0% 372µs ± 0% -14.64% (p=0.000 n=9+8) GlyphNRGBA16Over-8 26.9µs ± 3% 26.0µs ± 3% -3.55% (p=0.000 n=10+9) GlyphNRGBA16Src-8 18.8µs ± 2% 18.4µs ± 2% -2.21% (p=0.000 n=9+10) GlyphNRGBA32Over-8 99.1µs ± 2% 95.9µs ± 3% -3.23% (p=0.000 n=10+10) GlyphNRGBA32Src-8 65.6µs ± 3% 62.8µs ± 2% -4.36% (p=0.000 n=10+10) GlyphNRGBA64Over-8 376µs ± 4% 370µs ± 2% ~ (p=0.063 n=10+10) GlyphNRGBA64Src-8 238µs ± 3% 233µs ± 1% -2.21% (p=0.000 n=9+10) GlyphNRGBA128Over-8 1.52ms ± 2% 1.48ms ± 0% -2.11% (p=0.000 n=10+8) GlyphNRGBA128Src-8 951µs ± 3% 935µs ± 1% -1.69% (p=0.013 n=10+9) GlyphNRGBA256Over-8 6.00ms ± 1% 5.87ms ± 3% -2.12% (p=0.002 n=10+10) GlyphNRGBA256Src-8 3.94ms ± 2% 3.80ms ± 2% -3.64% (p=0.000 n=10+10) A comparison of the non-SIMD and SIMD versions: name time/op FixedAccumulateMask16-8 237ns ± 0% FixedAccumulateMaskSIMD16-8 80.0ns ± 1% FloatingAccumulateMask16-8 413ns ± 2% FloatingAccumulateMaskSIMD16-8 166ns ± 0% FixedAccumulateMask64-8 3.42µs ± 0% FixedAccumulateMaskSIMD64-8 1.09µs ± 0% FloatingAccumulateMask64-8 6.92µs ± 0% FloatingAccumulateMaskSIMD64-8 2.47µs ± 1% Change-Id: Ib6980e5975ed2842ff2a372f76dd5f2e95c5526c Reviewed-on: https://go-review.googlesource.com/30898 Reviewed-by: David Crawshaw <crawshaw@golang.org>
2016-10-13 05:50:52 +02:00
MOVQ dst_base+0(FP), DI
MOVQ dst_len+8(FP), BX
MOVQ src_base+24(FP), SI
MOVQ src_len+32(FP), R10
// Sanity check that len(dst) >= len(src).
CMPQ BX, R10
JLT flAccOpSrcEnd
// R10 = len(src) &^ 3
// R11 = len(src)
MOVQ R10, R11
ANDQ $-4, R10
vector: add SIMD versions of xxxAccumulateMask. name old time/op new time/op delta GlyphAlphaLoose16Over-8 3.96µs ± 0% 3.64µs ± 1% -8.08% (p=0.000 n=8+10) GlyphAlphaLoose16Src-8 3.64µs ± 0% 3.35µs ± 0% -7.88% (p=0.000 n=8+9) GlyphAlphaLoose32Over-8 8.45µs ± 0% 6.74µs ± 0% -20.22% (p=0.000 n=8+9) GlyphAlphaLoose32Src-8 7.24µs ± 0% 5.54µs ± 1% -23.48% (p=0.000 n=8+10) GlyphAlphaLoose64Over-8 22.2µs ± 0% 17.4µs ± 0% -21.67% (p=0.000 n=9+9) GlyphAlphaLoose64Src-8 17.6µs ± 1% 12.2µs ± 1% -30.32% (p=0.000 n=10+10) GlyphAlphaLoose128Over-8 67.9µs ± 0% 53.3µs ± 1% -21.53% (p=0.000 n=10+10) GlyphAlphaLoose128Src-8 48.2µs ± 0% 32.6µs ± 2% -32.41% (p=0.000 n=9+10) GlyphAlphaLoose256Over-8 242µs ± 1% 187µs ± 1% -22.96% (p=0.000 n=9+9) GlyphAlphaLoose256Src-8 163µs ± 0% 105µs ± 1% -35.83% (p=0.000 n=9+9) GlyphRGBA16Over-8 5.25µs ± 1% 4.95µs ± 0% -5.78% (p=0.000 n=9+9) GlyphRGBA16Src-8 4.72µs ± 0% 4.43µs ± 1% -6.22% (p=0.000 n=9+10) GlyphRGBA32Over-8 13.5µs ± 0% 11.9µs ± 1% -12.19% (p=0.000 n=9+10) GlyphRGBA32Src-8 11.5µs ± 1% 9.8µs ± 0% -14.72% (p=0.000 n=9+9) GlyphRGBA64Over-8 42.0µs ± 2% 36.9µs ± 1% -12.19% (p=0.000 n=10+10) GlyphRGBA64Src-8 34.1µs ± 1% 28.5µs ± 0% -16.25% (p=0.000 n=9+7) GlyphRGBA128Over-8 149µs ± 2% 133µs ± 1% -10.24% (p=0.000 n=10+9) GlyphRGBA128Src-8 115µs ± 1% 99µs ± 1% -13.57% (p=0.000 n=9+10) GlyphRGBA256Over-8 566µs ± 0% 511µs ± 1% -9.85% (p=0.000 n=9+10) GlyphRGBA256Src-8 435µs ± 0% 372µs ± 0% -14.64% (p=0.000 n=9+8) GlyphNRGBA16Over-8 26.9µs ± 3% 26.0µs ± 3% -3.55% (p=0.000 n=10+9) GlyphNRGBA16Src-8 18.8µs ± 2% 18.4µs ± 2% -2.21% (p=0.000 n=9+10) GlyphNRGBA32Over-8 99.1µs ± 2% 95.9µs ± 3% -3.23% (p=0.000 n=10+10) GlyphNRGBA32Src-8 65.6µs ± 3% 62.8µs ± 2% -4.36% (p=0.000 n=10+10) GlyphNRGBA64Over-8 376µs ± 4% 370µs ± 2% ~ (p=0.063 n=10+10) GlyphNRGBA64Src-8 238µs ± 3% 233µs ± 1% -2.21% (p=0.000 n=9+10) GlyphNRGBA128Over-8 1.52ms ± 2% 1.48ms ± 0% -2.11% (p=0.000 n=10+8) GlyphNRGBA128Src-8 951µs ± 3% 935µs ± 1% -1.69% (p=0.013 n=10+9) GlyphNRGBA256Over-8 6.00ms ± 1% 5.87ms ± 3% -2.12% (p=0.002 n=10+10) GlyphNRGBA256Src-8 3.94ms ± 2% 3.80ms ± 2% -3.64% (p=0.000 n=10+10) A comparison of the non-SIMD and SIMD versions: name time/op FixedAccumulateMask16-8 237ns ± 0% FixedAccumulateMaskSIMD16-8 80.0ns ± 1% FloatingAccumulateMask16-8 413ns ± 2% FloatingAccumulateMaskSIMD16-8 166ns ± 0% FixedAccumulateMask64-8 3.42µs ± 0% FixedAccumulateMaskSIMD64-8 1.09µs ± 0% FloatingAccumulateMask64-8 6.92µs ± 0% FloatingAccumulateMaskSIMD64-8 2.47µs ± 1% Change-Id: Ib6980e5975ed2842ff2a372f76dd5f2e95c5526c Reviewed-on: https://go-review.googlesource.com/30898 Reviewed-by: David Crawshaw <crawshaw@golang.org>
2016-10-13 05:50:52 +02:00
// Prepare to set MXCSR bits 13 and 14, so that the CVTPS2PL below is
// "Round To Zero".
STMXCSR mxcsrOrig-8(SP)
MOVL mxcsrOrig-8(SP), AX
ORL $0x6000, AX
MOVL AX, mxcsrNew-4(SP)
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>
2016-10-20 06:15:53 +02:00
// flSignMask := XMM(0x7fffffff repeated four times) // All but the sign bit of a float32.
// flOne := XMM(0x3f800000 repeated four times) // 1 as a float32.
// flAlmost65536 := XMM(0x477fffff repeated four times) // 255.99998 * 256 as a float32.
MOVOU flSignMask<>(SB), X3
MOVOU flOne<>(SB), X4
MOVOU flAlmost65536<>(SB), X5
vector: add SIMD versions of xxxAccumulateMask. name old time/op new time/op delta GlyphAlphaLoose16Over-8 3.96µs ± 0% 3.64µs ± 1% -8.08% (p=0.000 n=8+10) GlyphAlphaLoose16Src-8 3.64µs ± 0% 3.35µs ± 0% -7.88% (p=0.000 n=8+9) GlyphAlphaLoose32Over-8 8.45µs ± 0% 6.74µs ± 0% -20.22% (p=0.000 n=8+9) GlyphAlphaLoose32Src-8 7.24µs ± 0% 5.54µs ± 1% -23.48% (p=0.000 n=8+10) GlyphAlphaLoose64Over-8 22.2µs ± 0% 17.4µs ± 0% -21.67% (p=0.000 n=9+9) GlyphAlphaLoose64Src-8 17.6µs ± 1% 12.2µs ± 1% -30.32% (p=0.000 n=10+10) GlyphAlphaLoose128Over-8 67.9µs ± 0% 53.3µs ± 1% -21.53% (p=0.000 n=10+10) GlyphAlphaLoose128Src-8 48.2µs ± 0% 32.6µs ± 2% -32.41% (p=0.000 n=9+10) GlyphAlphaLoose256Over-8 242µs ± 1% 187µs ± 1% -22.96% (p=0.000 n=9+9) GlyphAlphaLoose256Src-8 163µs ± 0% 105µs ± 1% -35.83% (p=0.000 n=9+9) GlyphRGBA16Over-8 5.25µs ± 1% 4.95µs ± 0% -5.78% (p=0.000 n=9+9) GlyphRGBA16Src-8 4.72µs ± 0% 4.43µs ± 1% -6.22% (p=0.000 n=9+10) GlyphRGBA32Over-8 13.5µs ± 0% 11.9µs ± 1% -12.19% (p=0.000 n=9+10) GlyphRGBA32Src-8 11.5µs ± 1% 9.8µs ± 0% -14.72% (p=0.000 n=9+9) GlyphRGBA64Over-8 42.0µs ± 2% 36.9µs ± 1% -12.19% (p=0.000 n=10+10) GlyphRGBA64Src-8 34.1µs ± 1% 28.5µs ± 0% -16.25% (p=0.000 n=9+7) GlyphRGBA128Over-8 149µs ± 2% 133µs ± 1% -10.24% (p=0.000 n=10+9) GlyphRGBA128Src-8 115µs ± 1% 99µs ± 1% -13.57% (p=0.000 n=9+10) GlyphRGBA256Over-8 566µs ± 0% 511µs ± 1% -9.85% (p=0.000 n=9+10) GlyphRGBA256Src-8 435µs ± 0% 372µs ± 0% -14.64% (p=0.000 n=9+8) GlyphNRGBA16Over-8 26.9µs ± 3% 26.0µs ± 3% -3.55% (p=0.000 n=10+9) GlyphNRGBA16Src-8 18.8µs ± 2% 18.4µs ± 2% -2.21% (p=0.000 n=9+10) GlyphNRGBA32Over-8 99.1µs ± 2% 95.9µs ± 3% -3.23% (p=0.000 n=10+10) GlyphNRGBA32Src-8 65.6µs ± 3% 62.8µs ± 2% -4.36% (p=0.000 n=10+10) GlyphNRGBA64Over-8 376µs ± 4% 370µs ± 2% ~ (p=0.063 n=10+10) GlyphNRGBA64Src-8 238µs ± 3% 233µs ± 1% -2.21% (p=0.000 n=9+10) GlyphNRGBA128Over-8 1.52ms ± 2% 1.48ms ± 0% -2.11% (p=0.000 n=10+8) GlyphNRGBA128Src-8 951µs ± 3% 935µs ± 1% -1.69% (p=0.013 n=10+9) GlyphNRGBA256Over-8 6.00ms ± 1% 5.87ms ± 3% -2.12% (p=0.002 n=10+10) GlyphNRGBA256Src-8 3.94ms ± 2% 3.80ms ± 2% -3.64% (p=0.000 n=10+10) A comparison of the non-SIMD and SIMD versions: name time/op FixedAccumulateMask16-8 237ns ± 0% FixedAccumulateMaskSIMD16-8 80.0ns ± 1% FloatingAccumulateMask16-8 413ns ± 2% FloatingAccumulateMaskSIMD16-8 166ns ± 0% FixedAccumulateMask64-8 3.42µs ± 0% FixedAccumulateMaskSIMD64-8 1.09µs ± 0% FloatingAccumulateMask64-8 6.92µs ± 0% FloatingAccumulateMaskSIMD64-8 2.47µs ± 1% Change-Id: Ib6980e5975ed2842ff2a372f76dd5f2e95c5526c Reviewed-on: https://go-review.googlesource.com/30898 Reviewed-by: David Crawshaw <crawshaw@golang.org>
2016-10-13 05:50:52 +02:00
// gather := XMM(see above) // PSHUFB shuffle mask.
MOVOU gather<>(SB), X6
vector: add SIMD versions of xxxAccumulateMask. name old time/op new time/op delta GlyphAlphaLoose16Over-8 3.96µs ± 0% 3.64µs ± 1% -8.08% (p=0.000 n=8+10) GlyphAlphaLoose16Src-8 3.64µs ± 0% 3.35µs ± 0% -7.88% (p=0.000 n=8+9) GlyphAlphaLoose32Over-8 8.45µs ± 0% 6.74µs ± 0% -20.22% (p=0.000 n=8+9) GlyphAlphaLoose32Src-8 7.24µs ± 0% 5.54µs ± 1% -23.48% (p=0.000 n=8+10) GlyphAlphaLoose64Over-8 22.2µs ± 0% 17.4µs ± 0% -21.67% (p=0.000 n=9+9) GlyphAlphaLoose64Src-8 17.6µs ± 1% 12.2µs ± 1% -30.32% (p=0.000 n=10+10) GlyphAlphaLoose128Over-8 67.9µs ± 0% 53.3µs ± 1% -21.53% (p=0.000 n=10+10) GlyphAlphaLoose128Src-8 48.2µs ± 0% 32.6µs ± 2% -32.41% (p=0.000 n=9+10) GlyphAlphaLoose256Over-8 242µs ± 1% 187µs ± 1% -22.96% (p=0.000 n=9+9) GlyphAlphaLoose256Src-8 163µs ± 0% 105µs ± 1% -35.83% (p=0.000 n=9+9) GlyphRGBA16Over-8 5.25µs ± 1% 4.95µs ± 0% -5.78% (p=0.000 n=9+9) GlyphRGBA16Src-8 4.72µs ± 0% 4.43µs ± 1% -6.22% (p=0.000 n=9+10) GlyphRGBA32Over-8 13.5µs ± 0% 11.9µs ± 1% -12.19% (p=0.000 n=9+10) GlyphRGBA32Src-8 11.5µs ± 1% 9.8µs ± 0% -14.72% (p=0.000 n=9+9) GlyphRGBA64Over-8 42.0µs ± 2% 36.9µs ± 1% -12.19% (p=0.000 n=10+10) GlyphRGBA64Src-8 34.1µs ± 1% 28.5µs ± 0% -16.25% (p=0.000 n=9+7) GlyphRGBA128Over-8 149µs ± 2% 133µs ± 1% -10.24% (p=0.000 n=10+9) GlyphRGBA128Src-8 115µs ± 1% 99µs ± 1% -13.57% (p=0.000 n=9+10) GlyphRGBA256Over-8 566µs ± 0% 511µs ± 1% -9.85% (p=0.000 n=9+10) GlyphRGBA256Src-8 435µs ± 0% 372µs ± 0% -14.64% (p=0.000 n=9+8) GlyphNRGBA16Over-8 26.9µs ± 3% 26.0µs ± 3% -3.55% (p=0.000 n=10+9) GlyphNRGBA16Src-8 18.8µs ± 2% 18.4µs ± 2% -2.21% (p=0.000 n=9+10) GlyphNRGBA32Over-8 99.1µs ± 2% 95.9µs ± 3% -3.23% (p=0.000 n=10+10) GlyphNRGBA32Src-8 65.6µs ± 3% 62.8µs ± 2% -4.36% (p=0.000 n=10+10) GlyphNRGBA64Over-8 376µs ± 4% 370µs ± 2% ~ (p=0.063 n=10+10) GlyphNRGBA64Src-8 238µs ± 3% 233µs ± 1% -2.21% (p=0.000 n=9+10) GlyphNRGBA128Over-8 1.52ms ± 2% 1.48ms ± 0% -2.11% (p=0.000 n=10+8) GlyphNRGBA128Src-8 951µs ± 3% 935µs ± 1% -1.69% (p=0.013 n=10+9) GlyphNRGBA256Over-8 6.00ms ± 1% 5.87ms ± 3% -2.12% (p=0.002 n=10+10) GlyphNRGBA256Src-8 3.94ms ± 2% 3.80ms ± 2% -3.64% (p=0.000 n=10+10) A comparison of the non-SIMD and SIMD versions: name time/op FixedAccumulateMask16-8 237ns ± 0% FixedAccumulateMaskSIMD16-8 80.0ns ± 1% FloatingAccumulateMask16-8 413ns ± 2% FloatingAccumulateMaskSIMD16-8 166ns ± 0% FixedAccumulateMask64-8 3.42µs ± 0% FixedAccumulateMaskSIMD64-8 1.09µs ± 0% FloatingAccumulateMask64-8 6.92µs ± 0% FloatingAccumulateMaskSIMD64-8 2.47µs ± 1% Change-Id: Ib6980e5975ed2842ff2a372f76dd5f2e95c5526c Reviewed-on: https://go-review.googlesource.com/30898 Reviewed-by: David Crawshaw <crawshaw@golang.org>
2016-10-13 05:50:52 +02:00
// offset := XMM(0x00000000 repeated four times) // Cumulative sum.
XORPS X7, X7
// i := 0
MOVQ $0, R9
flAccOpSrcLoop4:
// for i < (len(src) &^ 3)
CMPQ R9, R10
JAE flAccOpSrcLoop1
// x = XMM(s0, s1, s2, s3)
//
// Where s0 is src[i+0], s1 is src[i+1], etc.
MOVOU (SI), X1
// scratch = XMM(0, s0, s1, s2)
// x += scratch // yields x == XMM(s0, s0+s1, s1+s2, s2+s3)
MOVOU X1, X0
PSLLO $4, X0
ADDPS X0, X1
// scratch = XMM(0, 0, 0, 0)
// scratch = XMM(scratch@0, scratch@0, x@0, x@1) // yields scratch == XMM(0, 0, s0, s0+s1)
// x += scratch // yields x == XMM(s0, s0+s1, s0+s1+s2, s0+s1+s2+s3)
XORPS X0, X0
SHUFPS $0x40, X1, X0
ADDPS X0, X1
// x += offset
ADDPS X7, X1
// y = x & flSignMask
// y = min(y, flOne)
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>
2016-10-20 06:15:53 +02:00
// y = mul(y, flAlmost65536)
MOVOU X3, X2
ANDPS X1, X2
MINPS X4, X2
MULPS X5, X2
// z = convertToInt32(y)
vector: add SIMD versions of xxxAccumulateMask. name old time/op new time/op delta GlyphAlphaLoose16Over-8 3.96µs ± 0% 3.64µs ± 1% -8.08% (p=0.000 n=8+10) GlyphAlphaLoose16Src-8 3.64µs ± 0% 3.35µs ± 0% -7.88% (p=0.000 n=8+9) GlyphAlphaLoose32Over-8 8.45µs ± 0% 6.74µs ± 0% -20.22% (p=0.000 n=8+9) GlyphAlphaLoose32Src-8 7.24µs ± 0% 5.54µs ± 1% -23.48% (p=0.000 n=8+10) GlyphAlphaLoose64Over-8 22.2µs ± 0% 17.4µs ± 0% -21.67% (p=0.000 n=9+9) GlyphAlphaLoose64Src-8 17.6µs ± 1% 12.2µs ± 1% -30.32% (p=0.000 n=10+10) GlyphAlphaLoose128Over-8 67.9µs ± 0% 53.3µs ± 1% -21.53% (p=0.000 n=10+10) GlyphAlphaLoose128Src-8 48.2µs ± 0% 32.6µs ± 2% -32.41% (p=0.000 n=9+10) GlyphAlphaLoose256Over-8 242µs ± 1% 187µs ± 1% -22.96% (p=0.000 n=9+9) GlyphAlphaLoose256Src-8 163µs ± 0% 105µs ± 1% -35.83% (p=0.000 n=9+9) GlyphRGBA16Over-8 5.25µs ± 1% 4.95µs ± 0% -5.78% (p=0.000 n=9+9) GlyphRGBA16Src-8 4.72µs ± 0% 4.43µs ± 1% -6.22% (p=0.000 n=9+10) GlyphRGBA32Over-8 13.5µs ± 0% 11.9µs ± 1% -12.19% (p=0.000 n=9+10) GlyphRGBA32Src-8 11.5µs ± 1% 9.8µs ± 0% -14.72% (p=0.000 n=9+9) GlyphRGBA64Over-8 42.0µs ± 2% 36.9µs ± 1% -12.19% (p=0.000 n=10+10) GlyphRGBA64Src-8 34.1µs ± 1% 28.5µs ± 0% -16.25% (p=0.000 n=9+7) GlyphRGBA128Over-8 149µs ± 2% 133µs ± 1% -10.24% (p=0.000 n=10+9) GlyphRGBA128Src-8 115µs ± 1% 99µs ± 1% -13.57% (p=0.000 n=9+10) GlyphRGBA256Over-8 566µs ± 0% 511µs ± 1% -9.85% (p=0.000 n=9+10) GlyphRGBA256Src-8 435µs ± 0% 372µs ± 0% -14.64% (p=0.000 n=9+8) GlyphNRGBA16Over-8 26.9µs ± 3% 26.0µs ± 3% -3.55% (p=0.000 n=10+9) GlyphNRGBA16Src-8 18.8µs ± 2% 18.4µs ± 2% -2.21% (p=0.000 n=9+10) GlyphNRGBA32Over-8 99.1µs ± 2% 95.9µs ± 3% -3.23% (p=0.000 n=10+10) GlyphNRGBA32Src-8 65.6µs ± 3% 62.8µs ± 2% -4.36% (p=0.000 n=10+10) GlyphNRGBA64Over-8 376µs ± 4% 370µs ± 2% ~ (p=0.063 n=10+10) GlyphNRGBA64Src-8 238µs ± 3% 233µs ± 1% -2.21% (p=0.000 n=9+10) GlyphNRGBA128Over-8 1.52ms ± 2% 1.48ms ± 0% -2.11% (p=0.000 n=10+8) GlyphNRGBA128Src-8 951µs ± 3% 935µs ± 1% -1.69% (p=0.013 n=10+9) GlyphNRGBA256Over-8 6.00ms ± 1% 5.87ms ± 3% -2.12% (p=0.002 n=10+10) GlyphNRGBA256Src-8 3.94ms ± 2% 3.80ms ± 2% -3.64% (p=0.000 n=10+10) A comparison of the non-SIMD and SIMD versions: name time/op FixedAccumulateMask16-8 237ns ± 0% FixedAccumulateMaskSIMD16-8 80.0ns ± 1% FloatingAccumulateMask16-8 413ns ± 2% FloatingAccumulateMaskSIMD16-8 166ns ± 0% FixedAccumulateMask64-8 3.42µs ± 0% FixedAccumulateMaskSIMD64-8 1.09µs ± 0% FloatingAccumulateMask64-8 6.92µs ± 0% FloatingAccumulateMaskSIMD64-8 2.47µs ± 1% Change-Id: Ib6980e5975ed2842ff2a372f76dd5f2e95c5526c Reviewed-on: https://go-review.googlesource.com/30898 Reviewed-by: David Crawshaw <crawshaw@golang.org>
2016-10-13 05:50:52 +02:00
LDMXCSR mxcsrNew-4(SP)
CVTPS2PL X2, X2
vector: add SIMD versions of xxxAccumulateMask. name old time/op new time/op delta GlyphAlphaLoose16Over-8 3.96µs ± 0% 3.64µs ± 1% -8.08% (p=0.000 n=8+10) GlyphAlphaLoose16Src-8 3.64µs ± 0% 3.35µs ± 0% -7.88% (p=0.000 n=8+9) GlyphAlphaLoose32Over-8 8.45µs ± 0% 6.74µs ± 0% -20.22% (p=0.000 n=8+9) GlyphAlphaLoose32Src-8 7.24µs ± 0% 5.54µs ± 1% -23.48% (p=0.000 n=8+10) GlyphAlphaLoose64Over-8 22.2µs ± 0% 17.4µs ± 0% -21.67% (p=0.000 n=9+9) GlyphAlphaLoose64Src-8 17.6µs ± 1% 12.2µs ± 1% -30.32% (p=0.000 n=10+10) GlyphAlphaLoose128Over-8 67.9µs ± 0% 53.3µs ± 1% -21.53% (p=0.000 n=10+10) GlyphAlphaLoose128Src-8 48.2µs ± 0% 32.6µs ± 2% -32.41% (p=0.000 n=9+10) GlyphAlphaLoose256Over-8 242µs ± 1% 187µs ± 1% -22.96% (p=0.000 n=9+9) GlyphAlphaLoose256Src-8 163µs ± 0% 105µs ± 1% -35.83% (p=0.000 n=9+9) GlyphRGBA16Over-8 5.25µs ± 1% 4.95µs ± 0% -5.78% (p=0.000 n=9+9) GlyphRGBA16Src-8 4.72µs ± 0% 4.43µs ± 1% -6.22% (p=0.000 n=9+10) GlyphRGBA32Over-8 13.5µs ± 0% 11.9µs ± 1% -12.19% (p=0.000 n=9+10) GlyphRGBA32Src-8 11.5µs ± 1% 9.8µs ± 0% -14.72% (p=0.000 n=9+9) GlyphRGBA64Over-8 42.0µs ± 2% 36.9µs ± 1% -12.19% (p=0.000 n=10+10) GlyphRGBA64Src-8 34.1µs ± 1% 28.5µs ± 0% -16.25% (p=0.000 n=9+7) GlyphRGBA128Over-8 149µs ± 2% 133µs ± 1% -10.24% (p=0.000 n=10+9) GlyphRGBA128Src-8 115µs ± 1% 99µs ± 1% -13.57% (p=0.000 n=9+10) GlyphRGBA256Over-8 566µs ± 0% 511µs ± 1% -9.85% (p=0.000 n=9+10) GlyphRGBA256Src-8 435µs ± 0% 372µs ± 0% -14.64% (p=0.000 n=9+8) GlyphNRGBA16Over-8 26.9µs ± 3% 26.0µs ± 3% -3.55% (p=0.000 n=10+9) GlyphNRGBA16Src-8 18.8µs ± 2% 18.4µs ± 2% -2.21% (p=0.000 n=9+10) GlyphNRGBA32Over-8 99.1µs ± 2% 95.9µs ± 3% -3.23% (p=0.000 n=10+10) GlyphNRGBA32Src-8 65.6µs ± 3% 62.8µs ± 2% -4.36% (p=0.000 n=10+10) GlyphNRGBA64Over-8 376µs ± 4% 370µs ± 2% ~ (p=0.063 n=10+10) GlyphNRGBA64Src-8 238µs ± 3% 233µs ± 1% -2.21% (p=0.000 n=9+10) GlyphNRGBA128Over-8 1.52ms ± 2% 1.48ms ± 0% -2.11% (p=0.000 n=10+8) GlyphNRGBA128Src-8 951µs ± 3% 935µs ± 1% -1.69% (p=0.013 n=10+9) GlyphNRGBA256Over-8 6.00ms ± 1% 5.87ms ± 3% -2.12% (p=0.002 n=10+10) GlyphNRGBA256Src-8 3.94ms ± 2% 3.80ms ± 2% -3.64% (p=0.000 n=10+10) A comparison of the non-SIMD and SIMD versions: name time/op FixedAccumulateMask16-8 237ns ± 0% FixedAccumulateMaskSIMD16-8 80.0ns ± 1% FloatingAccumulateMask16-8 413ns ± 2% FloatingAccumulateMaskSIMD16-8 166ns ± 0% FixedAccumulateMask64-8 3.42µs ± 0% FixedAccumulateMaskSIMD64-8 1.09µs ± 0% FloatingAccumulateMask64-8 6.92µs ± 0% FloatingAccumulateMaskSIMD64-8 2.47µs ± 1% Change-Id: Ib6980e5975ed2842ff2a372f76dd5f2e95c5526c Reviewed-on: https://go-review.googlesource.com/30898 Reviewed-by: David Crawshaw <crawshaw@golang.org>
2016-10-13 05:50:52 +02:00
LDMXCSR mxcsrOrig-8(SP)
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>
2016-10-20 06:15:53 +02:00
// z = shuffleTheSecondLowestBytesOfEach4ByteElement(z)
// copy(dst[:4], low4BytesOf(z))
PSHUFB X6, X2
MOVL X2, (DI)
// offset = XMM(x@3, x@3, x@3, x@3)
MOVOU X1, X7
SHUFPS $0xff, X1, X7
// i += 4
// dst = dst[4:]
// src = src[4:]
ADDQ $4, R9
ADDQ $4, DI
ADDQ $16, SI
JMP flAccOpSrcLoop4
flAccOpSrcLoop1:
// for i < len(src)
CMPQ R9, R11
vector: add SIMD versions of xxxAccumulateMask. name old time/op new time/op delta GlyphAlphaLoose16Over-8 3.96µs ± 0% 3.64µs ± 1% -8.08% (p=0.000 n=8+10) GlyphAlphaLoose16Src-8 3.64µs ± 0% 3.35µs ± 0% -7.88% (p=0.000 n=8+9) GlyphAlphaLoose32Over-8 8.45µs ± 0% 6.74µs ± 0% -20.22% (p=0.000 n=8+9) GlyphAlphaLoose32Src-8 7.24µs ± 0% 5.54µs ± 1% -23.48% (p=0.000 n=8+10) GlyphAlphaLoose64Over-8 22.2µs ± 0% 17.4µs ± 0% -21.67% (p=0.000 n=9+9) GlyphAlphaLoose64Src-8 17.6µs ± 1% 12.2µs ± 1% -30.32% (p=0.000 n=10+10) GlyphAlphaLoose128Over-8 67.9µs ± 0% 53.3µs ± 1% -21.53% (p=0.000 n=10+10) GlyphAlphaLoose128Src-8 48.2µs ± 0% 32.6µs ± 2% -32.41% (p=0.000 n=9+10) GlyphAlphaLoose256Over-8 242µs ± 1% 187µs ± 1% -22.96% (p=0.000 n=9+9) GlyphAlphaLoose256Src-8 163µs ± 0% 105µs ± 1% -35.83% (p=0.000 n=9+9) GlyphRGBA16Over-8 5.25µs ± 1% 4.95µs ± 0% -5.78% (p=0.000 n=9+9) GlyphRGBA16Src-8 4.72µs ± 0% 4.43µs ± 1% -6.22% (p=0.000 n=9+10) GlyphRGBA32Over-8 13.5µs ± 0% 11.9µs ± 1% -12.19% (p=0.000 n=9+10) GlyphRGBA32Src-8 11.5µs ± 1% 9.8µs ± 0% -14.72% (p=0.000 n=9+9) GlyphRGBA64Over-8 42.0µs ± 2% 36.9µs ± 1% -12.19% (p=0.000 n=10+10) GlyphRGBA64Src-8 34.1µs ± 1% 28.5µs ± 0% -16.25% (p=0.000 n=9+7) GlyphRGBA128Over-8 149µs ± 2% 133µs ± 1% -10.24% (p=0.000 n=10+9) GlyphRGBA128Src-8 115µs ± 1% 99µs ± 1% -13.57% (p=0.000 n=9+10) GlyphRGBA256Over-8 566µs ± 0% 511µs ± 1% -9.85% (p=0.000 n=9+10) GlyphRGBA256Src-8 435µs ± 0% 372µs ± 0% -14.64% (p=0.000 n=9+8) GlyphNRGBA16Over-8 26.9µs ± 3% 26.0µs ± 3% -3.55% (p=0.000 n=10+9) GlyphNRGBA16Src-8 18.8µs ± 2% 18.4µs ± 2% -2.21% (p=0.000 n=9+10) GlyphNRGBA32Over-8 99.1µs ± 2% 95.9µs ± 3% -3.23% (p=0.000 n=10+10) GlyphNRGBA32Src-8 65.6µs ± 3% 62.8µs ± 2% -4.36% (p=0.000 n=10+10) GlyphNRGBA64Over-8 376µs ± 4% 370µs ± 2% ~ (p=0.063 n=10+10) GlyphNRGBA64Src-8 238µs ± 3% 233µs ± 1% -2.21% (p=0.000 n=9+10) GlyphNRGBA128Over-8 1.52ms ± 2% 1.48ms ± 0% -2.11% (p=0.000 n=10+8) GlyphNRGBA128Src-8 951µs ± 3% 935µs ± 1% -1.69% (p=0.013 n=10+9) GlyphNRGBA256Over-8 6.00ms ± 1% 5.87ms ± 3% -2.12% (p=0.002 n=10+10) GlyphNRGBA256Src-8 3.94ms ± 2% 3.80ms ± 2% -3.64% (p=0.000 n=10+10) A comparison of the non-SIMD and SIMD versions: name time/op FixedAccumulateMask16-8 237ns ± 0% FixedAccumulateMaskSIMD16-8 80.0ns ± 1% FloatingAccumulateMask16-8 413ns ± 2% FloatingAccumulateMaskSIMD16-8 166ns ± 0% FixedAccumulateMask64-8 3.42µs ± 0% FixedAccumulateMaskSIMD64-8 1.09µs ± 0% FloatingAccumulateMask64-8 6.92µs ± 0% FloatingAccumulateMaskSIMD64-8 2.47µs ± 1% Change-Id: Ib6980e5975ed2842ff2a372f76dd5f2e95c5526c Reviewed-on: https://go-review.googlesource.com/30898 Reviewed-by: David Crawshaw <crawshaw@golang.org>
2016-10-13 05:50:52 +02:00
JAE flAccOpSrcEnd
// x = src[i] + offset
MOVL (SI), X1
ADDPS X7, X1
// y = x & flSignMask
// y = min(y, flOne)
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>
2016-10-20 06:15:53 +02:00
// y = mul(y, flAlmost65536)
MOVOU X3, X2
ANDPS X1, X2
MINPS X4, X2
MULPS X5, X2
// z = convertToInt32(y)
vector: add SIMD versions of xxxAccumulateMask. name old time/op new time/op delta GlyphAlphaLoose16Over-8 3.96µs ± 0% 3.64µs ± 1% -8.08% (p=0.000 n=8+10) GlyphAlphaLoose16Src-8 3.64µs ± 0% 3.35µs ± 0% -7.88% (p=0.000 n=8+9) GlyphAlphaLoose32Over-8 8.45µs ± 0% 6.74µs ± 0% -20.22% (p=0.000 n=8+9) GlyphAlphaLoose32Src-8 7.24µs ± 0% 5.54µs ± 1% -23.48% (p=0.000 n=8+10) GlyphAlphaLoose64Over-8 22.2µs ± 0% 17.4µs ± 0% -21.67% (p=0.000 n=9+9) GlyphAlphaLoose64Src-8 17.6µs ± 1% 12.2µs ± 1% -30.32% (p=0.000 n=10+10) GlyphAlphaLoose128Over-8 67.9µs ± 0% 53.3µs ± 1% -21.53% (p=0.000 n=10+10) GlyphAlphaLoose128Src-8 48.2µs ± 0% 32.6µs ± 2% -32.41% (p=0.000 n=9+10) GlyphAlphaLoose256Over-8 242µs ± 1% 187µs ± 1% -22.96% (p=0.000 n=9+9) GlyphAlphaLoose256Src-8 163µs ± 0% 105µs ± 1% -35.83% (p=0.000 n=9+9) GlyphRGBA16Over-8 5.25µs ± 1% 4.95µs ± 0% -5.78% (p=0.000 n=9+9) GlyphRGBA16Src-8 4.72µs ± 0% 4.43µs ± 1% -6.22% (p=0.000 n=9+10) GlyphRGBA32Over-8 13.5µs ± 0% 11.9µs ± 1% -12.19% (p=0.000 n=9+10) GlyphRGBA32Src-8 11.5µs ± 1% 9.8µs ± 0% -14.72% (p=0.000 n=9+9) GlyphRGBA64Over-8 42.0µs ± 2% 36.9µs ± 1% -12.19% (p=0.000 n=10+10) GlyphRGBA64Src-8 34.1µs ± 1% 28.5µs ± 0% -16.25% (p=0.000 n=9+7) GlyphRGBA128Over-8 149µs ± 2% 133µs ± 1% -10.24% (p=0.000 n=10+9) GlyphRGBA128Src-8 115µs ± 1% 99µs ± 1% -13.57% (p=0.000 n=9+10) GlyphRGBA256Over-8 566µs ± 0% 511µs ± 1% -9.85% (p=0.000 n=9+10) GlyphRGBA256Src-8 435µs ± 0% 372µs ± 0% -14.64% (p=0.000 n=9+8) GlyphNRGBA16Over-8 26.9µs ± 3% 26.0µs ± 3% -3.55% (p=0.000 n=10+9) GlyphNRGBA16Src-8 18.8µs ± 2% 18.4µs ± 2% -2.21% (p=0.000 n=9+10) GlyphNRGBA32Over-8 99.1µs ± 2% 95.9µs ± 3% -3.23% (p=0.000 n=10+10) GlyphNRGBA32Src-8 65.6µs ± 3% 62.8µs ± 2% -4.36% (p=0.000 n=10+10) GlyphNRGBA64Over-8 376µs ± 4% 370µs ± 2% ~ (p=0.063 n=10+10) GlyphNRGBA64Src-8 238µs ± 3% 233µs ± 1% -2.21% (p=0.000 n=9+10) GlyphNRGBA128Over-8 1.52ms ± 2% 1.48ms ± 0% -2.11% (p=0.000 n=10+8) GlyphNRGBA128Src-8 951µs ± 3% 935µs ± 1% -1.69% (p=0.013 n=10+9) GlyphNRGBA256Over-8 6.00ms ± 1% 5.87ms ± 3% -2.12% (p=0.002 n=10+10) GlyphNRGBA256Src-8 3.94ms ± 2% 3.80ms ± 2% -3.64% (p=0.000 n=10+10) A comparison of the non-SIMD and SIMD versions: name time/op FixedAccumulateMask16-8 237ns ± 0% FixedAccumulateMaskSIMD16-8 80.0ns ± 1% FloatingAccumulateMask16-8 413ns ± 2% FloatingAccumulateMaskSIMD16-8 166ns ± 0% FixedAccumulateMask64-8 3.42µs ± 0% FixedAccumulateMaskSIMD64-8 1.09µs ± 0% FloatingAccumulateMask64-8 6.92µs ± 0% FloatingAccumulateMaskSIMD64-8 2.47µs ± 1% Change-Id: Ib6980e5975ed2842ff2a372f76dd5f2e95c5526c Reviewed-on: https://go-review.googlesource.com/30898 Reviewed-by: David Crawshaw <crawshaw@golang.org>
2016-10-13 05:50:52 +02:00
LDMXCSR mxcsrNew-4(SP)
CVTPS2PL X2, X2
vector: add SIMD versions of xxxAccumulateMask. name old time/op new time/op delta GlyphAlphaLoose16Over-8 3.96µs ± 0% 3.64µs ± 1% -8.08% (p=0.000 n=8+10) GlyphAlphaLoose16Src-8 3.64µs ± 0% 3.35µs ± 0% -7.88% (p=0.000 n=8+9) GlyphAlphaLoose32Over-8 8.45µs ± 0% 6.74µs ± 0% -20.22% (p=0.000 n=8+9) GlyphAlphaLoose32Src-8 7.24µs ± 0% 5.54µs ± 1% -23.48% (p=0.000 n=8+10) GlyphAlphaLoose64Over-8 22.2µs ± 0% 17.4µs ± 0% -21.67% (p=0.000 n=9+9) GlyphAlphaLoose64Src-8 17.6µs ± 1% 12.2µs ± 1% -30.32% (p=0.000 n=10+10) GlyphAlphaLoose128Over-8 67.9µs ± 0% 53.3µs ± 1% -21.53% (p=0.000 n=10+10) GlyphAlphaLoose128Src-8 48.2µs ± 0% 32.6µs ± 2% -32.41% (p=0.000 n=9+10) GlyphAlphaLoose256Over-8 242µs ± 1% 187µs ± 1% -22.96% (p=0.000 n=9+9) GlyphAlphaLoose256Src-8 163µs ± 0% 105µs ± 1% -35.83% (p=0.000 n=9+9) GlyphRGBA16Over-8 5.25µs ± 1% 4.95µs ± 0% -5.78% (p=0.000 n=9+9) GlyphRGBA16Src-8 4.72µs ± 0% 4.43µs ± 1% -6.22% (p=0.000 n=9+10) GlyphRGBA32Over-8 13.5µs ± 0% 11.9µs ± 1% -12.19% (p=0.000 n=9+10) GlyphRGBA32Src-8 11.5µs ± 1% 9.8µs ± 0% -14.72% (p=0.000 n=9+9) GlyphRGBA64Over-8 42.0µs ± 2% 36.9µs ± 1% -12.19% (p=0.000 n=10+10) GlyphRGBA64Src-8 34.1µs ± 1% 28.5µs ± 0% -16.25% (p=0.000 n=9+7) GlyphRGBA128Over-8 149µs ± 2% 133µs ± 1% -10.24% (p=0.000 n=10+9) GlyphRGBA128Src-8 115µs ± 1% 99µs ± 1% -13.57% (p=0.000 n=9+10) GlyphRGBA256Over-8 566µs ± 0% 511µs ± 1% -9.85% (p=0.000 n=9+10) GlyphRGBA256Src-8 435µs ± 0% 372µs ± 0% -14.64% (p=0.000 n=9+8) GlyphNRGBA16Over-8 26.9µs ± 3% 26.0µs ± 3% -3.55% (p=0.000 n=10+9) GlyphNRGBA16Src-8 18.8µs ± 2% 18.4µs ± 2% -2.21% (p=0.000 n=9+10) GlyphNRGBA32Over-8 99.1µs ± 2% 95.9µs ± 3% -3.23% (p=0.000 n=10+10) GlyphNRGBA32Src-8 65.6µs ± 3% 62.8µs ± 2% -4.36% (p=0.000 n=10+10) GlyphNRGBA64Over-8 376µs ± 4% 370µs ± 2% ~ (p=0.063 n=10+10) GlyphNRGBA64Src-8 238µs ± 3% 233µs ± 1% -2.21% (p=0.000 n=9+10) GlyphNRGBA128Over-8 1.52ms ± 2% 1.48ms ± 0% -2.11% (p=0.000 n=10+8) GlyphNRGBA128Src-8 951µs ± 3% 935µs ± 1% -1.69% (p=0.013 n=10+9) GlyphNRGBA256Over-8 6.00ms ± 1% 5.87ms ± 3% -2.12% (p=0.002 n=10+10) GlyphNRGBA256Src-8 3.94ms ± 2% 3.80ms ± 2% -3.64% (p=0.000 n=10+10) A comparison of the non-SIMD and SIMD versions: name time/op FixedAccumulateMask16-8 237ns ± 0% FixedAccumulateMaskSIMD16-8 80.0ns ± 1% FloatingAccumulateMask16-8 413ns ± 2% FloatingAccumulateMaskSIMD16-8 166ns ± 0% FixedAccumulateMask64-8 3.42µs ± 0% FixedAccumulateMaskSIMD64-8 1.09µs ± 0% FloatingAccumulateMask64-8 6.92µs ± 0% FloatingAccumulateMaskSIMD64-8 2.47µs ± 1% Change-Id: Ib6980e5975ed2842ff2a372f76dd5f2e95c5526c Reviewed-on: https://go-review.googlesource.com/30898 Reviewed-by: David Crawshaw <crawshaw@golang.org>
2016-10-13 05:50:52 +02:00
LDMXCSR mxcsrOrig-8(SP)
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>
2016-10-20 06:15:53 +02:00
// dst[0] = uint8(z>>8)
MOVL X2, BX
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>
2016-10-20 06:15:53 +02:00
SHRL $8, BX
MOVB BX, (DI)
// offset = x
MOVOU X1, X7
// i += 1
// dst = dst[1:]
// src = src[1:]
ADDQ $1, R9
ADDQ $1, DI
ADDQ $4, SI
JMP flAccOpSrcLoop1
flAccOpSrcEnd:
RET
vector: add SIMD versions of xxxAccumulateMask. name old time/op new time/op delta GlyphAlphaLoose16Over-8 3.96µs ± 0% 3.64µs ± 1% -8.08% (p=0.000 n=8+10) GlyphAlphaLoose16Src-8 3.64µs ± 0% 3.35µs ± 0% -7.88% (p=0.000 n=8+9) GlyphAlphaLoose32Over-8 8.45µs ± 0% 6.74µs ± 0% -20.22% (p=0.000 n=8+9) GlyphAlphaLoose32Src-8 7.24µs ± 0% 5.54µs ± 1% -23.48% (p=0.000 n=8+10) GlyphAlphaLoose64Over-8 22.2µs ± 0% 17.4µs ± 0% -21.67% (p=0.000 n=9+9) GlyphAlphaLoose64Src-8 17.6µs ± 1% 12.2µs ± 1% -30.32% (p=0.000 n=10+10) GlyphAlphaLoose128Over-8 67.9µs ± 0% 53.3µs ± 1% -21.53% (p=0.000 n=10+10) GlyphAlphaLoose128Src-8 48.2µs ± 0% 32.6µs ± 2% -32.41% (p=0.000 n=9+10) GlyphAlphaLoose256Over-8 242µs ± 1% 187µs ± 1% -22.96% (p=0.000 n=9+9) GlyphAlphaLoose256Src-8 163µs ± 0% 105µs ± 1% -35.83% (p=0.000 n=9+9) GlyphRGBA16Over-8 5.25µs ± 1% 4.95µs ± 0% -5.78% (p=0.000 n=9+9) GlyphRGBA16Src-8 4.72µs ± 0% 4.43µs ± 1% -6.22% (p=0.000 n=9+10) GlyphRGBA32Over-8 13.5µs ± 0% 11.9µs ± 1% -12.19% (p=0.000 n=9+10) GlyphRGBA32Src-8 11.5µs ± 1% 9.8µs ± 0% -14.72% (p=0.000 n=9+9) GlyphRGBA64Over-8 42.0µs ± 2% 36.9µs ± 1% -12.19% (p=0.000 n=10+10) GlyphRGBA64Src-8 34.1µs ± 1% 28.5µs ± 0% -16.25% (p=0.000 n=9+7) GlyphRGBA128Over-8 149µs ± 2% 133µs ± 1% -10.24% (p=0.000 n=10+9) GlyphRGBA128Src-8 115µs ± 1% 99µs ± 1% -13.57% (p=0.000 n=9+10) GlyphRGBA256Over-8 566µs ± 0% 511µs ± 1% -9.85% (p=0.000 n=9+10) GlyphRGBA256Src-8 435µs ± 0% 372µs ± 0% -14.64% (p=0.000 n=9+8) GlyphNRGBA16Over-8 26.9µs ± 3% 26.0µs ± 3% -3.55% (p=0.000 n=10+9) GlyphNRGBA16Src-8 18.8µs ± 2% 18.4µs ± 2% -2.21% (p=0.000 n=9+10) GlyphNRGBA32Over-8 99.1µs ± 2% 95.9µs ± 3% -3.23% (p=0.000 n=10+10) GlyphNRGBA32Src-8 65.6µs ± 3% 62.8µs ± 2% -4.36% (p=0.000 n=10+10) GlyphNRGBA64Over-8 376µs ± 4% 370µs ± 2% ~ (p=0.063 n=10+10) GlyphNRGBA64Src-8 238µs ± 3% 233µs ± 1% -2.21% (p=0.000 n=9+10) GlyphNRGBA128Over-8 1.52ms ± 2% 1.48ms ± 0% -2.11% (p=0.000 n=10+8) GlyphNRGBA128Src-8 951µs ± 3% 935µs ± 1% -1.69% (p=0.013 n=10+9) GlyphNRGBA256Over-8 6.00ms ± 1% 5.87ms ± 3% -2.12% (p=0.002 n=10+10) GlyphNRGBA256Src-8 3.94ms ± 2% 3.80ms ± 2% -3.64% (p=0.000 n=10+10) A comparison of the non-SIMD and SIMD versions: name time/op FixedAccumulateMask16-8 237ns ± 0% FixedAccumulateMaskSIMD16-8 80.0ns ± 1% FloatingAccumulateMask16-8 413ns ± 2% FloatingAccumulateMaskSIMD16-8 166ns ± 0% FixedAccumulateMask64-8 3.42µs ± 0% FixedAccumulateMaskSIMD64-8 1.09µs ± 0% FloatingAccumulateMask64-8 6.92µs ± 0% FloatingAccumulateMaskSIMD64-8 2.47µs ± 1% Change-Id: Ib6980e5975ed2842ff2a372f76dd5f2e95c5526c Reviewed-on: https://go-review.googlesource.com/30898 Reviewed-by: David Crawshaw <crawshaw@golang.org>
2016-10-13 05:50:52 +02:00
// ----------------------------------------------------------------------------
// func floatingAccumulateMaskSIMD(dst []uint32, src []float32)
//
// XMM registers. Variable names are per
// https://github.com/google/font-rs/blob/master/src/accumulate.c
//
// xmm0 scratch
// xmm1 x
// xmm2 y, z
// xmm3 flSignMask
vector: add SIMD versions of xxxAccumulateMask. name old time/op new time/op delta GlyphAlphaLoose16Over-8 3.96µs ± 0% 3.64µs ± 1% -8.08% (p=0.000 n=8+10) GlyphAlphaLoose16Src-8 3.64µs ± 0% 3.35µs ± 0% -7.88% (p=0.000 n=8+9) GlyphAlphaLoose32Over-8 8.45µs ± 0% 6.74µs ± 0% -20.22% (p=0.000 n=8+9) GlyphAlphaLoose32Src-8 7.24µs ± 0% 5.54µs ± 1% -23.48% (p=0.000 n=8+10) GlyphAlphaLoose64Over-8 22.2µs ± 0% 17.4µs ± 0% -21.67% (p=0.000 n=9+9) GlyphAlphaLoose64Src-8 17.6µs ± 1% 12.2µs ± 1% -30.32% (p=0.000 n=10+10) GlyphAlphaLoose128Over-8 67.9µs ± 0% 53.3µs ± 1% -21.53% (p=0.000 n=10+10) GlyphAlphaLoose128Src-8 48.2µs ± 0% 32.6µs ± 2% -32.41% (p=0.000 n=9+10) GlyphAlphaLoose256Over-8 242µs ± 1% 187µs ± 1% -22.96% (p=0.000 n=9+9) GlyphAlphaLoose256Src-8 163µs ± 0% 105µs ± 1% -35.83% (p=0.000 n=9+9) GlyphRGBA16Over-8 5.25µs ± 1% 4.95µs ± 0% -5.78% (p=0.000 n=9+9) GlyphRGBA16Src-8 4.72µs ± 0% 4.43µs ± 1% -6.22% (p=0.000 n=9+10) GlyphRGBA32Over-8 13.5µs ± 0% 11.9µs ± 1% -12.19% (p=0.000 n=9+10) GlyphRGBA32Src-8 11.5µs ± 1% 9.8µs ± 0% -14.72% (p=0.000 n=9+9) GlyphRGBA64Over-8 42.0µs ± 2% 36.9µs ± 1% -12.19% (p=0.000 n=10+10) GlyphRGBA64Src-8 34.1µs ± 1% 28.5µs ± 0% -16.25% (p=0.000 n=9+7) GlyphRGBA128Over-8 149µs ± 2% 133µs ± 1% -10.24% (p=0.000 n=10+9) GlyphRGBA128Src-8 115µs ± 1% 99µs ± 1% -13.57% (p=0.000 n=9+10) GlyphRGBA256Over-8 566µs ± 0% 511µs ± 1% -9.85% (p=0.000 n=9+10) GlyphRGBA256Src-8 435µs ± 0% 372µs ± 0% -14.64% (p=0.000 n=9+8) GlyphNRGBA16Over-8 26.9µs ± 3% 26.0µs ± 3% -3.55% (p=0.000 n=10+9) GlyphNRGBA16Src-8 18.8µs ± 2% 18.4µs ± 2% -2.21% (p=0.000 n=9+10) GlyphNRGBA32Over-8 99.1µs ± 2% 95.9µs ± 3% -3.23% (p=0.000 n=10+10) GlyphNRGBA32Src-8 65.6µs ± 3% 62.8µs ± 2% -4.36% (p=0.000 n=10+10) GlyphNRGBA64Over-8 376µs ± 4% 370µs ± 2% ~ (p=0.063 n=10+10) GlyphNRGBA64Src-8 238µs ± 3% 233µs ± 1% -2.21% (p=0.000 n=9+10) GlyphNRGBA128Over-8 1.52ms ± 2% 1.48ms ± 0% -2.11% (p=0.000 n=10+8) GlyphNRGBA128Src-8 951µs ± 3% 935µs ± 1% -1.69% (p=0.013 n=10+9) GlyphNRGBA256Over-8 6.00ms ± 1% 5.87ms ± 3% -2.12% (p=0.002 n=10+10) GlyphNRGBA256Src-8 3.94ms ± 2% 3.80ms ± 2% -3.64% (p=0.000 n=10+10) A comparison of the non-SIMD and SIMD versions: name time/op FixedAccumulateMask16-8 237ns ± 0% FixedAccumulateMaskSIMD16-8 80.0ns ± 1% FloatingAccumulateMask16-8 413ns ± 2% FloatingAccumulateMaskSIMD16-8 166ns ± 0% FixedAccumulateMask64-8 3.42µs ± 0% FixedAccumulateMaskSIMD64-8 1.09µs ± 0% FloatingAccumulateMask64-8 6.92µs ± 0% FloatingAccumulateMaskSIMD64-8 2.47µs ± 1% Change-Id: Ib6980e5975ed2842ff2a372f76dd5f2e95c5526c Reviewed-on: https://go-review.googlesource.com/30898 Reviewed-by: David Crawshaw <crawshaw@golang.org>
2016-10-13 05:50:52 +02:00
// xmm4 flOne
// xmm5 flAlmost65536
vector: add SIMD versions of xxxAccumulateMask. name old time/op new time/op delta GlyphAlphaLoose16Over-8 3.96µs ± 0% 3.64µs ± 1% -8.08% (p=0.000 n=8+10) GlyphAlphaLoose16Src-8 3.64µs ± 0% 3.35µs ± 0% -7.88% (p=0.000 n=8+9) GlyphAlphaLoose32Over-8 8.45µs ± 0% 6.74µs ± 0% -20.22% (p=0.000 n=8+9) GlyphAlphaLoose32Src-8 7.24µs ± 0% 5.54µs ± 1% -23.48% (p=0.000 n=8+10) GlyphAlphaLoose64Over-8 22.2µs ± 0% 17.4µs ± 0% -21.67% (p=0.000 n=9+9) GlyphAlphaLoose64Src-8 17.6µs ± 1% 12.2µs ± 1% -30.32% (p=0.000 n=10+10) GlyphAlphaLoose128Over-8 67.9µs ± 0% 53.3µs ± 1% -21.53% (p=0.000 n=10+10) GlyphAlphaLoose128Src-8 48.2µs ± 0% 32.6µs ± 2% -32.41% (p=0.000 n=9+10) GlyphAlphaLoose256Over-8 242µs ± 1% 187µs ± 1% -22.96% (p=0.000 n=9+9) GlyphAlphaLoose256Src-8 163µs ± 0% 105µs ± 1% -35.83% (p=0.000 n=9+9) GlyphRGBA16Over-8 5.25µs ± 1% 4.95µs ± 0% -5.78% (p=0.000 n=9+9) GlyphRGBA16Src-8 4.72µs ± 0% 4.43µs ± 1% -6.22% (p=0.000 n=9+10) GlyphRGBA32Over-8 13.5µs ± 0% 11.9µs ± 1% -12.19% (p=0.000 n=9+10) GlyphRGBA32Src-8 11.5µs ± 1% 9.8µs ± 0% -14.72% (p=0.000 n=9+9) GlyphRGBA64Over-8 42.0µs ± 2% 36.9µs ± 1% -12.19% (p=0.000 n=10+10) GlyphRGBA64Src-8 34.1µs ± 1% 28.5µs ± 0% -16.25% (p=0.000 n=9+7) GlyphRGBA128Over-8 149µs ± 2% 133µs ± 1% -10.24% (p=0.000 n=10+9) GlyphRGBA128Src-8 115µs ± 1% 99µs ± 1% -13.57% (p=0.000 n=9+10) GlyphRGBA256Over-8 566µs ± 0% 511µs ± 1% -9.85% (p=0.000 n=9+10) GlyphRGBA256Src-8 435µs ± 0% 372µs ± 0% -14.64% (p=0.000 n=9+8) GlyphNRGBA16Over-8 26.9µs ± 3% 26.0µs ± 3% -3.55% (p=0.000 n=10+9) GlyphNRGBA16Src-8 18.8µs ± 2% 18.4µs ± 2% -2.21% (p=0.000 n=9+10) GlyphNRGBA32Over-8 99.1µs ± 2% 95.9µs ± 3% -3.23% (p=0.000 n=10+10) GlyphNRGBA32Src-8 65.6µs ± 3% 62.8µs ± 2% -4.36% (p=0.000 n=10+10) GlyphNRGBA64Over-8 376µs ± 4% 370µs ± 2% ~ (p=0.063 n=10+10) GlyphNRGBA64Src-8 238µs ± 3% 233µs ± 1% -2.21% (p=0.000 n=9+10) GlyphNRGBA128Over-8 1.52ms ± 2% 1.48ms ± 0% -2.11% (p=0.000 n=10+8) GlyphNRGBA128Src-8 951µs ± 3% 935µs ± 1% -1.69% (p=0.013 n=10+9) GlyphNRGBA256Over-8 6.00ms ± 1% 5.87ms ± 3% -2.12% (p=0.002 n=10+10) GlyphNRGBA256Src-8 3.94ms ± 2% 3.80ms ± 2% -3.64% (p=0.000 n=10+10) A comparison of the non-SIMD and SIMD versions: name time/op FixedAccumulateMask16-8 237ns ± 0% FixedAccumulateMaskSIMD16-8 80.0ns ± 1% FloatingAccumulateMask16-8 413ns ± 2% FloatingAccumulateMaskSIMD16-8 166ns ± 0% FixedAccumulateMask64-8 3.42µs ± 0% FixedAccumulateMaskSIMD64-8 1.09µs ± 0% FloatingAccumulateMask64-8 6.92µs ± 0% FloatingAccumulateMaskSIMD64-8 2.47µs ± 1% Change-Id: Ib6980e5975ed2842ff2a372f76dd5f2e95c5526c Reviewed-on: https://go-review.googlesource.com/30898 Reviewed-by: David Crawshaw <crawshaw@golang.org>
2016-10-13 05:50:52 +02:00
// xmm6 -
// xmm7 offset
// xmm8 -
// xmm9 -
// xmm10 -
TEXT ·floatingAccumulateMaskSIMD(SB), NOSPLIT, $8-48
MOVQ dst_base+0(FP), DI
MOVQ dst_len+8(FP), BX
MOVQ src_base+24(FP), SI
MOVQ src_len+32(FP), R10
// Sanity check that len(dst) >= len(src).
CMPQ BX, R10
JLT flAccMaskEnd
// R10 = len(src) &^ 3
// R11 = len(src)
MOVQ R10, R11
ANDQ $-4, R10
// Prepare to set MXCSR bits 13 and 14, so that the CVTPS2PL below is
// "Round To Zero".
STMXCSR mxcsrOrig-8(SP)
MOVL mxcsrOrig-8(SP), AX
ORL $0x6000, AX
MOVL AX, mxcsrNew-4(SP)
// flSignMask := XMM(0x7fffffff repeated four times) // All but the sign bit of a float32.
// flOne := XMM(0x3f800000 repeated four times) // 1 as a float32.
// flAlmost65536 := XMM(0x477fffff repeated four times) // 255.99998 * 256 as a float32.
MOVOU flSignMask<>(SB), X3
vector: add SIMD versions of xxxAccumulateMask. name old time/op new time/op delta GlyphAlphaLoose16Over-8 3.96µs ± 0% 3.64µs ± 1% -8.08% (p=0.000 n=8+10) GlyphAlphaLoose16Src-8 3.64µs ± 0% 3.35µs ± 0% -7.88% (p=0.000 n=8+9) GlyphAlphaLoose32Over-8 8.45µs ± 0% 6.74µs ± 0% -20.22% (p=0.000 n=8+9) GlyphAlphaLoose32Src-8 7.24µs ± 0% 5.54µs ± 1% -23.48% (p=0.000 n=8+10) GlyphAlphaLoose64Over-8 22.2µs ± 0% 17.4µs ± 0% -21.67% (p=0.000 n=9+9) GlyphAlphaLoose64Src-8 17.6µs ± 1% 12.2µs ± 1% -30.32% (p=0.000 n=10+10) GlyphAlphaLoose128Over-8 67.9µs ± 0% 53.3µs ± 1% -21.53% (p=0.000 n=10+10) GlyphAlphaLoose128Src-8 48.2µs ± 0% 32.6µs ± 2% -32.41% (p=0.000 n=9+10) GlyphAlphaLoose256Over-8 242µs ± 1% 187µs ± 1% -22.96% (p=0.000 n=9+9) GlyphAlphaLoose256Src-8 163µs ± 0% 105µs ± 1% -35.83% (p=0.000 n=9+9) GlyphRGBA16Over-8 5.25µs ± 1% 4.95µs ± 0% -5.78% (p=0.000 n=9+9) GlyphRGBA16Src-8 4.72µs ± 0% 4.43µs ± 1% -6.22% (p=0.000 n=9+10) GlyphRGBA32Over-8 13.5µs ± 0% 11.9µs ± 1% -12.19% (p=0.000 n=9+10) GlyphRGBA32Src-8 11.5µs ± 1% 9.8µs ± 0% -14.72% (p=0.000 n=9+9) GlyphRGBA64Over-8 42.0µs ± 2% 36.9µs ± 1% -12.19% (p=0.000 n=10+10) GlyphRGBA64Src-8 34.1µs ± 1% 28.5µs ± 0% -16.25% (p=0.000 n=9+7) GlyphRGBA128Over-8 149µs ± 2% 133µs ± 1% -10.24% (p=0.000 n=10+9) GlyphRGBA128Src-8 115µs ± 1% 99µs ± 1% -13.57% (p=0.000 n=9+10) GlyphRGBA256Over-8 566µs ± 0% 511µs ± 1% -9.85% (p=0.000 n=9+10) GlyphRGBA256Src-8 435µs ± 0% 372µs ± 0% -14.64% (p=0.000 n=9+8) GlyphNRGBA16Over-8 26.9µs ± 3% 26.0µs ± 3% -3.55% (p=0.000 n=10+9) GlyphNRGBA16Src-8 18.8µs ± 2% 18.4µs ± 2% -2.21% (p=0.000 n=9+10) GlyphNRGBA32Over-8 99.1µs ± 2% 95.9µs ± 3% -3.23% (p=0.000 n=10+10) GlyphNRGBA32Src-8 65.6µs ± 3% 62.8µs ± 2% -4.36% (p=0.000 n=10+10) GlyphNRGBA64Over-8 376µs ± 4% 370µs ± 2% ~ (p=0.063 n=10+10) GlyphNRGBA64Src-8 238µs ± 3% 233µs ± 1% -2.21% (p=0.000 n=9+10) GlyphNRGBA128Over-8 1.52ms ± 2% 1.48ms ± 0% -2.11% (p=0.000 n=10+8) GlyphNRGBA128Src-8 951µs ± 3% 935µs ± 1% -1.69% (p=0.013 n=10+9) GlyphNRGBA256Over-8 6.00ms ± 1% 5.87ms ± 3% -2.12% (p=0.002 n=10+10) GlyphNRGBA256Src-8 3.94ms ± 2% 3.80ms ± 2% -3.64% (p=0.000 n=10+10) A comparison of the non-SIMD and SIMD versions: name time/op FixedAccumulateMask16-8 237ns ± 0% FixedAccumulateMaskSIMD16-8 80.0ns ± 1% FloatingAccumulateMask16-8 413ns ± 2% FloatingAccumulateMaskSIMD16-8 166ns ± 0% FixedAccumulateMask64-8 3.42µs ± 0% FixedAccumulateMaskSIMD64-8 1.09µs ± 0% FloatingAccumulateMask64-8 6.92µs ± 0% FloatingAccumulateMaskSIMD64-8 2.47µs ± 1% Change-Id: Ib6980e5975ed2842ff2a372f76dd5f2e95c5526c Reviewed-on: https://go-review.googlesource.com/30898 Reviewed-by: David Crawshaw <crawshaw@golang.org>
2016-10-13 05:50:52 +02:00
MOVOU flOne<>(SB), X4
MOVOU flAlmost65536<>(SB), X5
vector: add SIMD versions of xxxAccumulateMask. name old time/op new time/op delta GlyphAlphaLoose16Over-8 3.96µs ± 0% 3.64µs ± 1% -8.08% (p=0.000 n=8+10) GlyphAlphaLoose16Src-8 3.64µs ± 0% 3.35µs ± 0% -7.88% (p=0.000 n=8+9) GlyphAlphaLoose32Over-8 8.45µs ± 0% 6.74µs ± 0% -20.22% (p=0.000 n=8+9) GlyphAlphaLoose32Src-8 7.24µs ± 0% 5.54µs ± 1% -23.48% (p=0.000 n=8+10) GlyphAlphaLoose64Over-8 22.2µs ± 0% 17.4µs ± 0% -21.67% (p=0.000 n=9+9) GlyphAlphaLoose64Src-8 17.6µs ± 1% 12.2µs ± 1% -30.32% (p=0.000 n=10+10) GlyphAlphaLoose128Over-8 67.9µs ± 0% 53.3µs ± 1% -21.53% (p=0.000 n=10+10) GlyphAlphaLoose128Src-8 48.2µs ± 0% 32.6µs ± 2% -32.41% (p=0.000 n=9+10) GlyphAlphaLoose256Over-8 242µs ± 1% 187µs ± 1% -22.96% (p=0.000 n=9+9) GlyphAlphaLoose256Src-8 163µs ± 0% 105µs ± 1% -35.83% (p=0.000 n=9+9) GlyphRGBA16Over-8 5.25µs ± 1% 4.95µs ± 0% -5.78% (p=0.000 n=9+9) GlyphRGBA16Src-8 4.72µs ± 0% 4.43µs ± 1% -6.22% (p=0.000 n=9+10) GlyphRGBA32Over-8 13.5µs ± 0% 11.9µs ± 1% -12.19% (p=0.000 n=9+10) GlyphRGBA32Src-8 11.5µs ± 1% 9.8µs ± 0% -14.72% (p=0.000 n=9+9) GlyphRGBA64Over-8 42.0µs ± 2% 36.9µs ± 1% -12.19% (p=0.000 n=10+10) GlyphRGBA64Src-8 34.1µs ± 1% 28.5µs ± 0% -16.25% (p=0.000 n=9+7) GlyphRGBA128Over-8 149µs ± 2% 133µs ± 1% -10.24% (p=0.000 n=10+9) GlyphRGBA128Src-8 115µs ± 1% 99µs ± 1% -13.57% (p=0.000 n=9+10) GlyphRGBA256Over-8 566µs ± 0% 511µs ± 1% -9.85% (p=0.000 n=9+10) GlyphRGBA256Src-8 435µs ± 0% 372µs ± 0% -14.64% (p=0.000 n=9+8) GlyphNRGBA16Over-8 26.9µs ± 3% 26.0µs ± 3% -3.55% (p=0.000 n=10+9) GlyphNRGBA16Src-8 18.8µs ± 2% 18.4µs ± 2% -2.21% (p=0.000 n=9+10) GlyphNRGBA32Over-8 99.1µs ± 2% 95.9µs ± 3% -3.23% (p=0.000 n=10+10) GlyphNRGBA32Src-8 65.6µs ± 3% 62.8µs ± 2% -4.36% (p=0.000 n=10+10) GlyphNRGBA64Over-8 376µs ± 4% 370µs ± 2% ~ (p=0.063 n=10+10) GlyphNRGBA64Src-8 238µs ± 3% 233µs ± 1% -2.21% (p=0.000 n=9+10) GlyphNRGBA128Over-8 1.52ms ± 2% 1.48ms ± 0% -2.11% (p=0.000 n=10+8) GlyphNRGBA128Src-8 951µs ± 3% 935µs ± 1% -1.69% (p=0.013 n=10+9) GlyphNRGBA256Over-8 6.00ms ± 1% 5.87ms ± 3% -2.12% (p=0.002 n=10+10) GlyphNRGBA256Src-8 3.94ms ± 2% 3.80ms ± 2% -3.64% (p=0.000 n=10+10) A comparison of the non-SIMD and SIMD versions: name time/op FixedAccumulateMask16-8 237ns ± 0% FixedAccumulateMaskSIMD16-8 80.0ns ± 1% FloatingAccumulateMask16-8 413ns ± 2% FloatingAccumulateMaskSIMD16-8 166ns ± 0% FixedAccumulateMask64-8 3.42µs ± 0% FixedAccumulateMaskSIMD64-8 1.09µs ± 0% FloatingAccumulateMask64-8 6.92µs ± 0% FloatingAccumulateMaskSIMD64-8 2.47µs ± 1% Change-Id: Ib6980e5975ed2842ff2a372f76dd5f2e95c5526c Reviewed-on: https://go-review.googlesource.com/30898 Reviewed-by: David Crawshaw <crawshaw@golang.org>
2016-10-13 05:50:52 +02:00
// offset := XMM(0x00000000 repeated four times) // Cumulative sum.
XORPS X7, X7
// i := 0
MOVQ $0, R9
flAccMaskLoop4:
// for i < (len(src) &^ 3)
CMPQ R9, R10
JAE flAccMaskLoop1
// x = XMM(s0, s1, s2, s3)
//
// Where s0 is src[i+0], s1 is src[i+1], etc.
MOVOU (SI), X1
// scratch = XMM(0, s0, s1, s2)
// x += scratch // yields x == XMM(s0, s0+s1, s1+s2, s2+s3)
MOVOU X1, X0
PSLLO $4, X0
ADDPS X0, X1
// scratch = XMM(0, 0, 0, 0)
// scratch = XMM(scratch@0, scratch@0, x@0, x@1) // yields scratch == XMM(0, 0, s0, s0+s1)
// x += scratch // yields x == XMM(s0, s0+s1, s0+s1+s2, s0+s1+s2+s3)
XORPS X0, X0
SHUFPS $0x40, X1, X0
ADDPS X0, X1
// x += offset
ADDPS X7, X1
// y = x & flSignMask
// y = min(y, flOne)
// y = mul(y, flAlmost65536)
MOVOU X3, X2
vector: add SIMD versions of xxxAccumulateMask. name old time/op new time/op delta GlyphAlphaLoose16Over-8 3.96µs ± 0% 3.64µs ± 1% -8.08% (p=0.000 n=8+10) GlyphAlphaLoose16Src-8 3.64µs ± 0% 3.35µs ± 0% -7.88% (p=0.000 n=8+9) GlyphAlphaLoose32Over-8 8.45µs ± 0% 6.74µs ± 0% -20.22% (p=0.000 n=8+9) GlyphAlphaLoose32Src-8 7.24µs ± 0% 5.54µs ± 1% -23.48% (p=0.000 n=8+10) GlyphAlphaLoose64Over-8 22.2µs ± 0% 17.4µs ± 0% -21.67% (p=0.000 n=9+9) GlyphAlphaLoose64Src-8 17.6µs ± 1% 12.2µs ± 1% -30.32% (p=0.000 n=10+10) GlyphAlphaLoose128Over-8 67.9µs ± 0% 53.3µs ± 1% -21.53% (p=0.000 n=10+10) GlyphAlphaLoose128Src-8 48.2µs ± 0% 32.6µs ± 2% -32.41% (p=0.000 n=9+10) GlyphAlphaLoose256Over-8 242µs ± 1% 187µs ± 1% -22.96% (p=0.000 n=9+9) GlyphAlphaLoose256Src-8 163µs ± 0% 105µs ± 1% -35.83% (p=0.000 n=9+9) GlyphRGBA16Over-8 5.25µs ± 1% 4.95µs ± 0% -5.78% (p=0.000 n=9+9) GlyphRGBA16Src-8 4.72µs ± 0% 4.43µs ± 1% -6.22% (p=0.000 n=9+10) GlyphRGBA32Over-8 13.5µs ± 0% 11.9µs ± 1% -12.19% (p=0.000 n=9+10) GlyphRGBA32Src-8 11.5µs ± 1% 9.8µs ± 0% -14.72% (p=0.000 n=9+9) GlyphRGBA64Over-8 42.0µs ± 2% 36.9µs ± 1% -12.19% (p=0.000 n=10+10) GlyphRGBA64Src-8 34.1µs ± 1% 28.5µs ± 0% -16.25% (p=0.000 n=9+7) GlyphRGBA128Over-8 149µs ± 2% 133µs ± 1% -10.24% (p=0.000 n=10+9) GlyphRGBA128Src-8 115µs ± 1% 99µs ± 1% -13.57% (p=0.000 n=9+10) GlyphRGBA256Over-8 566µs ± 0% 511µs ± 1% -9.85% (p=0.000 n=9+10) GlyphRGBA256Src-8 435µs ± 0% 372µs ± 0% -14.64% (p=0.000 n=9+8) GlyphNRGBA16Over-8 26.9µs ± 3% 26.0µs ± 3% -3.55% (p=0.000 n=10+9) GlyphNRGBA16Src-8 18.8µs ± 2% 18.4µs ± 2% -2.21% (p=0.000 n=9+10) GlyphNRGBA32Over-8 99.1µs ± 2% 95.9µs ± 3% -3.23% (p=0.000 n=10+10) GlyphNRGBA32Src-8 65.6µs ± 3% 62.8µs ± 2% -4.36% (p=0.000 n=10+10) GlyphNRGBA64Over-8 376µs ± 4% 370µs ± 2% ~ (p=0.063 n=10+10) GlyphNRGBA64Src-8 238µs ± 3% 233µs ± 1% -2.21% (p=0.000 n=9+10) GlyphNRGBA128Over-8 1.52ms ± 2% 1.48ms ± 0% -2.11% (p=0.000 n=10+8) GlyphNRGBA128Src-8 951µs ± 3% 935µs ± 1% -1.69% (p=0.013 n=10+9) GlyphNRGBA256Over-8 6.00ms ± 1% 5.87ms ± 3% -2.12% (p=0.002 n=10+10) GlyphNRGBA256Src-8 3.94ms ± 2% 3.80ms ± 2% -3.64% (p=0.000 n=10+10) A comparison of the non-SIMD and SIMD versions: name time/op FixedAccumulateMask16-8 237ns ± 0% FixedAccumulateMaskSIMD16-8 80.0ns ± 1% FloatingAccumulateMask16-8 413ns ± 2% FloatingAccumulateMaskSIMD16-8 166ns ± 0% FixedAccumulateMask64-8 3.42µs ± 0% FixedAccumulateMaskSIMD64-8 1.09µs ± 0% FloatingAccumulateMask64-8 6.92µs ± 0% FloatingAccumulateMaskSIMD64-8 2.47µs ± 1% Change-Id: Ib6980e5975ed2842ff2a372f76dd5f2e95c5526c Reviewed-on: https://go-review.googlesource.com/30898 Reviewed-by: David Crawshaw <crawshaw@golang.org>
2016-10-13 05:50:52 +02:00
ANDPS X1, X2
MINPS X4, X2
MULPS X5, X2
vector: add SIMD versions of xxxAccumulateMask. name old time/op new time/op delta GlyphAlphaLoose16Over-8 3.96µs ± 0% 3.64µs ± 1% -8.08% (p=0.000 n=8+10) GlyphAlphaLoose16Src-8 3.64µs ± 0% 3.35µs ± 0% -7.88% (p=0.000 n=8+9) GlyphAlphaLoose32Over-8 8.45µs ± 0% 6.74µs ± 0% -20.22% (p=0.000 n=8+9) GlyphAlphaLoose32Src-8 7.24µs ± 0% 5.54µs ± 1% -23.48% (p=0.000 n=8+10) GlyphAlphaLoose64Over-8 22.2µs ± 0% 17.4µs ± 0% -21.67% (p=0.000 n=9+9) GlyphAlphaLoose64Src-8 17.6µs ± 1% 12.2µs ± 1% -30.32% (p=0.000 n=10+10) GlyphAlphaLoose128Over-8 67.9µs ± 0% 53.3µs ± 1% -21.53% (p=0.000 n=10+10) GlyphAlphaLoose128Src-8 48.2µs ± 0% 32.6µs ± 2% -32.41% (p=0.000 n=9+10) GlyphAlphaLoose256Over-8 242µs ± 1% 187µs ± 1% -22.96% (p=0.000 n=9+9) GlyphAlphaLoose256Src-8 163µs ± 0% 105µs ± 1% -35.83% (p=0.000 n=9+9) GlyphRGBA16Over-8 5.25µs ± 1% 4.95µs ± 0% -5.78% (p=0.000 n=9+9) GlyphRGBA16Src-8 4.72µs ± 0% 4.43µs ± 1% -6.22% (p=0.000 n=9+10) GlyphRGBA32Over-8 13.5µs ± 0% 11.9µs ± 1% -12.19% (p=0.000 n=9+10) GlyphRGBA32Src-8 11.5µs ± 1% 9.8µs ± 0% -14.72% (p=0.000 n=9+9) GlyphRGBA64Over-8 42.0µs ± 2% 36.9µs ± 1% -12.19% (p=0.000 n=10+10) GlyphRGBA64Src-8 34.1µs ± 1% 28.5µs ± 0% -16.25% (p=0.000 n=9+7) GlyphRGBA128Over-8 149µs ± 2% 133µs ± 1% -10.24% (p=0.000 n=10+9) GlyphRGBA128Src-8 115µs ± 1% 99µs ± 1% -13.57% (p=0.000 n=9+10) GlyphRGBA256Over-8 566µs ± 0% 511µs ± 1% -9.85% (p=0.000 n=9+10) GlyphRGBA256Src-8 435µs ± 0% 372µs ± 0% -14.64% (p=0.000 n=9+8) GlyphNRGBA16Over-8 26.9µs ± 3% 26.0µs ± 3% -3.55% (p=0.000 n=10+9) GlyphNRGBA16Src-8 18.8µs ± 2% 18.4µs ± 2% -2.21% (p=0.000 n=9+10) GlyphNRGBA32Over-8 99.1µs ± 2% 95.9µs ± 3% -3.23% (p=0.000 n=10+10) GlyphNRGBA32Src-8 65.6µs ± 3% 62.8µs ± 2% -4.36% (p=0.000 n=10+10) GlyphNRGBA64Over-8 376µs ± 4% 370µs ± 2% ~ (p=0.063 n=10+10) GlyphNRGBA64Src-8 238µs ± 3% 233µs ± 1% -2.21% (p=0.000 n=9+10) GlyphNRGBA128Over-8 1.52ms ± 2% 1.48ms ± 0% -2.11% (p=0.000 n=10+8) GlyphNRGBA128Src-8 951µs ± 3% 935µs ± 1% -1.69% (p=0.013 n=10+9) GlyphNRGBA256Over-8 6.00ms ± 1% 5.87ms ± 3% -2.12% (p=0.002 n=10+10) GlyphNRGBA256Src-8 3.94ms ± 2% 3.80ms ± 2% -3.64% (p=0.000 n=10+10) A comparison of the non-SIMD and SIMD versions: name time/op FixedAccumulateMask16-8 237ns ± 0% FixedAccumulateMaskSIMD16-8 80.0ns ± 1% FloatingAccumulateMask16-8 413ns ± 2% FloatingAccumulateMaskSIMD16-8 166ns ± 0% FixedAccumulateMask64-8 3.42µs ± 0% FixedAccumulateMaskSIMD64-8 1.09µs ± 0% FloatingAccumulateMask64-8 6.92µs ± 0% FloatingAccumulateMaskSIMD64-8 2.47µs ± 1% Change-Id: Ib6980e5975ed2842ff2a372f76dd5f2e95c5526c Reviewed-on: https://go-review.googlesource.com/30898 Reviewed-by: David Crawshaw <crawshaw@golang.org>
2016-10-13 05:50:52 +02:00
// z = convertToInt32(y)
LDMXCSR mxcsrNew-4(SP)
CVTPS2PL X2, X2
LDMXCSR mxcsrOrig-8(SP)
// copy(dst[:4], z)
MOVOU X2, (DI)
// offset = XMM(x@3, x@3, x@3, x@3)
MOVOU X1, X7
SHUFPS $0xff, X1, X7
// i += 4
// dst = dst[4:]
// src = src[4:]
ADDQ $4, R9
ADDQ $16, DI
ADDQ $16, SI
JMP flAccMaskLoop4
flAccMaskLoop1:
// for i < len(src)
CMPQ R9, R11
JAE flAccMaskEnd
// x = src[i] + offset
MOVL (SI), X1
ADDPS X7, X1
// y = x & flSignMask
// y = min(y, flOne)
// y = mul(y, flAlmost65536)
MOVOU X3, X2
vector: add SIMD versions of xxxAccumulateMask. name old time/op new time/op delta GlyphAlphaLoose16Over-8 3.96µs ± 0% 3.64µs ± 1% -8.08% (p=0.000 n=8+10) GlyphAlphaLoose16Src-8 3.64µs ± 0% 3.35µs ± 0% -7.88% (p=0.000 n=8+9) GlyphAlphaLoose32Over-8 8.45µs ± 0% 6.74µs ± 0% -20.22% (p=0.000 n=8+9) GlyphAlphaLoose32Src-8 7.24µs ± 0% 5.54µs ± 1% -23.48% (p=0.000 n=8+10) GlyphAlphaLoose64Over-8 22.2µs ± 0% 17.4µs ± 0% -21.67% (p=0.000 n=9+9) GlyphAlphaLoose64Src-8 17.6µs ± 1% 12.2µs ± 1% -30.32% (p=0.000 n=10+10) GlyphAlphaLoose128Over-8 67.9µs ± 0% 53.3µs ± 1% -21.53% (p=0.000 n=10+10) GlyphAlphaLoose128Src-8 48.2µs ± 0% 32.6µs ± 2% -32.41% (p=0.000 n=9+10) GlyphAlphaLoose256Over-8 242µs ± 1% 187µs ± 1% -22.96% (p=0.000 n=9+9) GlyphAlphaLoose256Src-8 163µs ± 0% 105µs ± 1% -35.83% (p=0.000 n=9+9) GlyphRGBA16Over-8 5.25µs ± 1% 4.95µs ± 0% -5.78% (p=0.000 n=9+9) GlyphRGBA16Src-8 4.72µs ± 0% 4.43µs ± 1% -6.22% (p=0.000 n=9+10) GlyphRGBA32Over-8 13.5µs ± 0% 11.9µs ± 1% -12.19% (p=0.000 n=9+10) GlyphRGBA32Src-8 11.5µs ± 1% 9.8µs ± 0% -14.72% (p=0.000 n=9+9) GlyphRGBA64Over-8 42.0µs ± 2% 36.9µs ± 1% -12.19% (p=0.000 n=10+10) GlyphRGBA64Src-8 34.1µs ± 1% 28.5µs ± 0% -16.25% (p=0.000 n=9+7) GlyphRGBA128Over-8 149µs ± 2% 133µs ± 1% -10.24% (p=0.000 n=10+9) GlyphRGBA128Src-8 115µs ± 1% 99µs ± 1% -13.57% (p=0.000 n=9+10) GlyphRGBA256Over-8 566µs ± 0% 511µs ± 1% -9.85% (p=0.000 n=9+10) GlyphRGBA256Src-8 435µs ± 0% 372µs ± 0% -14.64% (p=0.000 n=9+8) GlyphNRGBA16Over-8 26.9µs ± 3% 26.0µs ± 3% -3.55% (p=0.000 n=10+9) GlyphNRGBA16Src-8 18.8µs ± 2% 18.4µs ± 2% -2.21% (p=0.000 n=9+10) GlyphNRGBA32Over-8 99.1µs ± 2% 95.9µs ± 3% -3.23% (p=0.000 n=10+10) GlyphNRGBA32Src-8 65.6µs ± 3% 62.8µs ± 2% -4.36% (p=0.000 n=10+10) GlyphNRGBA64Over-8 376µs ± 4% 370µs ± 2% ~ (p=0.063 n=10+10) GlyphNRGBA64Src-8 238µs ± 3% 233µs ± 1% -2.21% (p=0.000 n=9+10) GlyphNRGBA128Over-8 1.52ms ± 2% 1.48ms ± 0% -2.11% (p=0.000 n=10+8) GlyphNRGBA128Src-8 951µs ± 3% 935µs ± 1% -1.69% (p=0.013 n=10+9) GlyphNRGBA256Over-8 6.00ms ± 1% 5.87ms ± 3% -2.12% (p=0.002 n=10+10) GlyphNRGBA256Src-8 3.94ms ± 2% 3.80ms ± 2% -3.64% (p=0.000 n=10+10) A comparison of the non-SIMD and SIMD versions: name time/op FixedAccumulateMask16-8 237ns ± 0% FixedAccumulateMaskSIMD16-8 80.0ns ± 1% FloatingAccumulateMask16-8 413ns ± 2% FloatingAccumulateMaskSIMD16-8 166ns ± 0% FixedAccumulateMask64-8 3.42µs ± 0% FixedAccumulateMaskSIMD64-8 1.09µs ± 0% FloatingAccumulateMask64-8 6.92µs ± 0% FloatingAccumulateMaskSIMD64-8 2.47µs ± 1% Change-Id: Ib6980e5975ed2842ff2a372f76dd5f2e95c5526c Reviewed-on: https://go-review.googlesource.com/30898 Reviewed-by: David Crawshaw <crawshaw@golang.org>
2016-10-13 05:50:52 +02:00
ANDPS X1, X2
MINPS X4, X2
MULPS X5, X2
vector: add SIMD versions of xxxAccumulateMask. name old time/op new time/op delta GlyphAlphaLoose16Over-8 3.96µs ± 0% 3.64µs ± 1% -8.08% (p=0.000 n=8+10) GlyphAlphaLoose16Src-8 3.64µs ± 0% 3.35µs ± 0% -7.88% (p=0.000 n=8+9) GlyphAlphaLoose32Over-8 8.45µs ± 0% 6.74µs ± 0% -20.22% (p=0.000 n=8+9) GlyphAlphaLoose32Src-8 7.24µs ± 0% 5.54µs ± 1% -23.48% (p=0.000 n=8+10) GlyphAlphaLoose64Over-8 22.2µs ± 0% 17.4µs ± 0% -21.67% (p=0.000 n=9+9) GlyphAlphaLoose64Src-8 17.6µs ± 1% 12.2µs ± 1% -30.32% (p=0.000 n=10+10) GlyphAlphaLoose128Over-8 67.9µs ± 0% 53.3µs ± 1% -21.53% (p=0.000 n=10+10) GlyphAlphaLoose128Src-8 48.2µs ± 0% 32.6µs ± 2% -32.41% (p=0.000 n=9+10) GlyphAlphaLoose256Over-8 242µs ± 1% 187µs ± 1% -22.96% (p=0.000 n=9+9) GlyphAlphaLoose256Src-8 163µs ± 0% 105µs ± 1% -35.83% (p=0.000 n=9+9) GlyphRGBA16Over-8 5.25µs ± 1% 4.95µs ± 0% -5.78% (p=0.000 n=9+9) GlyphRGBA16Src-8 4.72µs ± 0% 4.43µs ± 1% -6.22% (p=0.000 n=9+10) GlyphRGBA32Over-8 13.5µs ± 0% 11.9µs ± 1% -12.19% (p=0.000 n=9+10) GlyphRGBA32Src-8 11.5µs ± 1% 9.8µs ± 0% -14.72% (p=0.000 n=9+9) GlyphRGBA64Over-8 42.0µs ± 2% 36.9µs ± 1% -12.19% (p=0.000 n=10+10) GlyphRGBA64Src-8 34.1µs ± 1% 28.5µs ± 0% -16.25% (p=0.000 n=9+7) GlyphRGBA128Over-8 149µs ± 2% 133µs ± 1% -10.24% (p=0.000 n=10+9) GlyphRGBA128Src-8 115µs ± 1% 99µs ± 1% -13.57% (p=0.000 n=9+10) GlyphRGBA256Over-8 566µs ± 0% 511µs ± 1% -9.85% (p=0.000 n=9+10) GlyphRGBA256Src-8 435µs ± 0% 372µs ± 0% -14.64% (p=0.000 n=9+8) GlyphNRGBA16Over-8 26.9µs ± 3% 26.0µs ± 3% -3.55% (p=0.000 n=10+9) GlyphNRGBA16Src-8 18.8µs ± 2% 18.4µs ± 2% -2.21% (p=0.000 n=9+10) GlyphNRGBA32Over-8 99.1µs ± 2% 95.9µs ± 3% -3.23% (p=0.000 n=10+10) GlyphNRGBA32Src-8 65.6µs ± 3% 62.8µs ± 2% -4.36% (p=0.000 n=10+10) GlyphNRGBA64Over-8 376µs ± 4% 370µs ± 2% ~ (p=0.063 n=10+10) GlyphNRGBA64Src-8 238µs ± 3% 233µs ± 1% -2.21% (p=0.000 n=9+10) GlyphNRGBA128Over-8 1.52ms ± 2% 1.48ms ± 0% -2.11% (p=0.000 n=10+8) GlyphNRGBA128Src-8 951µs ± 3% 935µs ± 1% -1.69% (p=0.013 n=10+9) GlyphNRGBA256Over-8 6.00ms ± 1% 5.87ms ± 3% -2.12% (p=0.002 n=10+10) GlyphNRGBA256Src-8 3.94ms ± 2% 3.80ms ± 2% -3.64% (p=0.000 n=10+10) A comparison of the non-SIMD and SIMD versions: name time/op FixedAccumulateMask16-8 237ns ± 0% FixedAccumulateMaskSIMD16-8 80.0ns ± 1% FloatingAccumulateMask16-8 413ns ± 2% FloatingAccumulateMaskSIMD16-8 166ns ± 0% FixedAccumulateMask64-8 3.42µs ± 0% FixedAccumulateMaskSIMD64-8 1.09µs ± 0% FloatingAccumulateMask64-8 6.92µs ± 0% FloatingAccumulateMaskSIMD64-8 2.47µs ± 1% Change-Id: Ib6980e5975ed2842ff2a372f76dd5f2e95c5526c Reviewed-on: https://go-review.googlesource.com/30898 Reviewed-by: David Crawshaw <crawshaw@golang.org>
2016-10-13 05:50:52 +02:00
// z = convertToInt32(y)
LDMXCSR mxcsrNew-4(SP)
CVTPS2PL X2, X2
LDMXCSR mxcsrOrig-8(SP)
// dst[0] = uint32(z)
MOVL X2, (DI)
// offset = x
MOVOU X1, X7
// i += 1
// dst = dst[1:]
// src = src[1:]
ADDQ $1, R9
ADDQ $4, DI
ADDQ $4, SI
JMP flAccMaskLoop1
flAccMaskEnd:
RET