golang-image/vector/gen_acc_amd64.s.tmpl
Nigel Tao 8edbaf3f9e vector: add SIMD versions of xxxAccumulateOpOver.
name                             old time/op  new time/op  delta
GlyphAlpha16Over-8               3.55µs ± 0%  3.17µs ± 0%  -10.58%  (p=0.000 n=10+10)
GlyphAlpha32Over-8               6.73µs ± 1%  4.94µs ± 0%  -26.55%   (p=0.000 n=10+9)
GlyphAlpha64Over-8               16.4µs ± 0%   9.6µs ± 0%  -41.30%   (p=0.000 n=9+10)
GlyphAlpha128Over-8              47.3µs ± 0%  23.1µs ± 1%  -51.09%    (p=0.000 n=9+9)
GlyphAlpha256Over-8               159µs ± 0%    69µs ± 0%  -56.82%    (p=0.000 n=9+8)

A comparison of the non-SIMD and SIMD versions:

name                              time/op
FixedAccumulateOpOver16-8          579ns ± 0%
FixedAccumulateOpOverSIMD16-8      183ns ± 0%
FloatingAccumulateOpOver16-8       670ns ± 1%
FloatingAccumulateOpOverSIMD16-8   242ns ± 0%
FixedAccumulateOpOver64-8         9.61µs ± 0%
FixedAccumulateOpOverSIMD64-8     2.72µs ± 0%
FloatingAccumulateOpOver64-8      11.1µs ± 0%
FloatingAccumulateOpOverSIMD64-8  3.65µs ± 0%

Change-Id: I08273c40ac5445f39b77a88fb8b6b07fd3e5f84b
Reviewed-on: https://go-review.googlesource.com/30831
Reviewed-by: David Crawshaw <crawshaw@golang.org>
2016-10-13 03:04:36 +00:00

191 lines
4.6 KiB
Cheetah

// Copyright 2016 The Go Authors. All rights reserved.
// Use of this source code is governed by a BSD-style
// license that can be found in the LICENSE file.
// +build !appengine
// +build gc
// +build !noasm
#include "textflag.h"
// fl is short for floating point math. fx is short for fixed point math.
DATA flAlmost256<>+0x00(SB)/8, $0x437fffff437fffff
DATA flAlmost256<>+0x08(SB)/8, $0x437fffff437fffff
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
// gather is a PSHUFB mask that brings the low 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
// i???j???k???l???
// to
// ijkl000000000000
DATA gather<>+0x00(SB)/8, $0x808080800c080400
DATA gather<>+0x08(SB)/8, $0x8080808080808080
DATA fxAlmost256<>+0x00(SB)/8, $0x000000ff000000ff
DATA fxAlmost256<>+0x08(SB)/8, $0x000000ff000000ff
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 flAlmost256<>(SB), (NOPTR+RODATA), $16
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 fxAlmost256<>(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 {{.LongName}}SIMD(dst []uint8, src {{.SrcType}})
//
// 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 {{.XMM3}}
// xmm4 {{.XMM4}}
// xmm5 {{.XMM5}}
// xmm6 gather
// xmm7 offset
// xmm8 {{.XMM8}}
// xmm9 {{.XMM9}}
// xmm10 {{.XMM10}}
TEXT ·{{.LongName}}SIMD(SB), NOSPLIT, ${{.FrameSize}}-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 {{.ShortName}}End
// R10 = len(src) &^ 3
// R11 = len(src)
MOVQ R10, R11
ANDQ $-4, R10
{{.Setup}}
{{.LoadXMMRegs}}
// gather := XMM(see above) // PSHUFB shuffle mask.
// offset := XMM(0x00000000 repeated four times) // Cumulative sum.
MOVOU gather<>(SB), X6
XORPS X7, X7
// i := 0
MOVQ $0, R9
{{.ShortName}}Loop4:
// for i < (len(src) &^ 3)
CMPQ R9, R10
JAE {{.ShortName}}Loop1
// 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
{{.Add}} 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
{{.Add}} X0, X1
// x += offset
{{.Add}} X7, X1
{{.ClampAndScale}}
// z = convertToInt32(y)
{{.ConvertToInt32}}
{{.Store4}}
// 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 {{.ShortName}}Loop4
{{.ShortName}}Loop1:
// for i < len(src)
CMPQ R9, R11
JAE {{.ShortName}}Cleanup
// x = src[i] + offset
MOVL (SI), X1
{{.Add}} X7, X1
{{.ClampAndScale}}
// z = convertToInt32(y)
{{.ConvertToInt32}}
{{.Store1}}
// offset = x
MOVOU X1, X7
// i += 1
// dst = dst[1:]
// src = src[1:]
ADDQ $1, R9
ADDQ $1, DI
ADDQ $4, SI
JMP {{.ShortName}}Loop1
{{.ShortName}}Cleanup:
{{.Cleanup}}
{{.ShortName}}End:
RET