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>
diff --git a/vector/acc_amd64.s b/vector/acc_amd64.s
index 8d9719c..dee7ad1 100644
--- a/vector/acc_amd64.s
+++ b/vector/acc_amd64.s
@@ -8,8 +8,6 @@
// 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
@@ -28,30 +26,26 @@
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.
+// 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
-// i???j???k???l???
+// ?i???j???k???l??
// to
// ijkl000000000000
-DATA gather<>+0x00(SB)/8, $0x808080800c080400
+DATA gather<>+0x00(SB)/8, $0x808080800d090501
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
@@ -196,15 +190,12 @@
PSRLQ $47, X0
PSRLQ $47, X11
- // Merge the two registers back to one, X11.
+ // Merge the two registers back to one, X11, and add maskA.
PSLLQ $32, X11
XORPS X0, X11
-
- // Add maskA, shift from 16 bit color to 8 bit color.
PADDD X11, X2
- PSRLQ $8, X2
- // As per opSrcStore4, shuffle and copy the low 4 bytes.
+ // As per opSrcStore4, shuffle and copy the 4 second-lowest bytes.
PSHUFB X6, X2
MOVL X2, (DI)
@@ -292,7 +283,7 @@
// xmm2 y, z
// xmm3 -
// xmm4 -
-// xmm5 fxAlmost256
+// xmm5 fxAlmost65536
// xmm6 gather
// xmm7 offset
// xmm8 -
@@ -314,8 +305,8 @@
MOVQ R10, R11
ANDQ $-4, R10
- // fxAlmost256 := XMM(0x000000ff repeated four times) // Maximum of an uint8.
- MOVOU fxAlmost256<>(SB), X5
+ // fxAlmost65536 := XMM(0x0000ffff repeated four times) // Maximum of an uint16.
+ MOVOU fxAlmost65536<>(SB), X5
// gather := XMM(see above) // PSHUFB shuffle mask.
MOVOU gather<>(SB), X6
@@ -353,24 +344,24 @@
PADDD X7, X1
// y = abs(x)
- // y >>= 10 // Shift by 2*ϕ - 8.
- // y = min(y, fxAlmost256)
+ // y >>= 2 // Shift by 2*ϕ - 16.
+ // y = min(y, fxAlmost65536)
//
// pabsd %xmm1,%xmm2
- // psrld $0xa,%xmm2
+ // psrld $0x2,%xmm2
// pminud %xmm5,%xmm2
//
// Hopefully we'll get these opcode mnemonics into the assembler for Go
// 1.8. https://golang.org/issue/16007 isn't exactly the same thing, but
// it's similar.
BYTE $0x66; BYTE $0x0f; BYTE $0x38; BYTE $0x1e; BYTE $0xd1
- BYTE $0x66; BYTE $0x0f; BYTE $0x72; BYTE $0xd2; BYTE $0x0a
+ BYTE $0x66; BYTE $0x0f; BYTE $0x72; BYTE $0xd2; BYTE $0x02
BYTE $0x66; BYTE $0x0f; BYTE $0x38; BYTE $0x3b; BYTE $0xd5
// z = convertToInt32(y)
// No-op.
- // z = shuffleTheLowBytesOfEach4ByteElement(z)
+ // z = shuffleTheSecondLowestBytesOfEach4ByteElement(z)
// copy(dst[:4], low4BytesOf(z))
PSHUFB X6, X2
MOVL X2, (DI)
@@ -397,25 +388,26 @@
PADDD X7, X1
// y = abs(x)
- // y >>= 10 // Shift by 2*ϕ - 8.
- // y = min(y, fxAlmost256)
+ // y >>= 2 // Shift by 2*ϕ - 16.
+ // y = min(y, fxAlmost65536)
//
// pabsd %xmm1,%xmm2
- // psrld $0xa,%xmm2
+ // psrld $0x2,%xmm2
// pminud %xmm5,%xmm2
//
// Hopefully we'll get these opcode mnemonics into the assembler for Go
// 1.8. https://golang.org/issue/16007 isn't exactly the same thing, but
// it's similar.
BYTE $0x66; BYTE $0x0f; BYTE $0x38; BYTE $0x1e; BYTE $0xd1
- BYTE $0x66; BYTE $0x0f; BYTE $0x72; BYTE $0xd2; BYTE $0x0a
+ BYTE $0x66; BYTE $0x0f; BYTE $0x72; BYTE $0xd2; BYTE $0x02
BYTE $0x66; BYTE $0x0f; BYTE $0x38; BYTE $0x3b; BYTE $0xd5
// z = convertToInt32(y)
// No-op.
- // dst[0] = uint8(z)
+ // dst[0] = uint8(z>>8)
MOVL X2, BX
+ SHRL $8, BX
MOVB BX, (DI)
// offset = x
@@ -712,15 +704,12 @@
PSRLQ $47, X0
PSRLQ $47, X11
- // Merge the two registers back to one, X11.
+ // Merge the two registers back to one, X11, and add maskA.
PSLLQ $32, X11
XORPS X0, X11
-
- // Add maskA, shift from 16 bit color to 8 bit color.
PADDD X11, X2
- PSRLQ $8, X2
- // As per opSrcStore4, shuffle and copy the low 4 bytes.
+ // As per opSrcStore4, shuffle and copy the 4 second-lowest bytes.
PSHUFB X6, X2
MOVL X2, (DI)
@@ -801,7 +790,7 @@
// xmm0 scratch
// xmm1 x
// xmm2 y, z
-// xmm3 flAlmost256
+// xmm3 flAlmost65536
// xmm4 flOne
// xmm5 flSignMask
// xmm6 gather
@@ -832,10 +821,10 @@
ORL $0x6000, AX
MOVL AX, mxcsrNew-4(SP)
- // flAlmost256 := XMM(0x437fffff repeated four times) // 255.99998 as a float32.
- // flOne := XMM(0x3f800000 repeated four times) // 1 as a float32.
- // flSignMask := XMM(0x7fffffff repeated four times) // All but the sign bit of a float32.
- MOVOU flAlmost256<>(SB), X3
+ // flAlmost65536 := XMM(0x477fffff repeated four times) // 255.99998 * 256 as a float32.
+ // flOne := XMM(0x3f800000 repeated four times) // 1 as a float32.
+ // flSignMask := XMM(0x7fffffff repeated four times) // All but the sign bit of a float32.
+ MOVOU flAlmost65536<>(SB), X3
MOVOU flOne<>(SB), X4
MOVOU flSignMask<>(SB), X5
@@ -876,7 +865,7 @@
// y = x & flSignMask
// y = min(y, flOne)
- // y = mul(y, flAlmost256)
+ // y = mul(y, flAlmost65536)
MOVOU X5, X2
ANDPS X1, X2
MINPS X4, X2
@@ -887,7 +876,7 @@
CVTPS2PL X2, X2
LDMXCSR mxcsrOrig-8(SP)
- // z = shuffleTheLowBytesOfEach4ByteElement(z)
+ // z = shuffleTheSecondLowestBytesOfEach4ByteElement(z)
// copy(dst[:4], low4BytesOf(z))
PSHUFB X6, X2
MOVL X2, (DI)
@@ -915,7 +904,7 @@
// y = x & flSignMask
// y = min(y, flOne)
- // y = mul(y, flAlmost256)
+ // y = mul(y, flAlmost65536)
MOVOU X5, X2
ANDPS X1, X2
MINPS X4, X2
@@ -926,8 +915,9 @@
CVTPS2PL X2, X2
LDMXCSR mxcsrOrig-8(SP)
- // dst[0] = uint8(z)
+ // dst[0] = uint8(z>>8)
MOVL X2, BX
+ SHRL $8, BX
MOVB BX, (DI)
// offset = x
diff --git a/vector/gen.go b/vector/gen.go
index 92be417..cafd860 100644
--- a/vector/gen.go
+++ b/vector/gen.go
@@ -101,16 +101,16 @@
DstElemSize4: 4 * sizeOfUint8,
XMM3: fxXMM3,
XMM4: fxXMM4,
- XMM5: fxXMM5_65536,
+ XMM5: fxXMM5,
XMM6: opOverXMM6,
XMM8: opOverXMM8,
XMM9: opOverXMM9,
XMM10: opOverXMM10,
LoadArgs: twoArgLoadArgs,
Setup: fxSetup,
- LoadXMMRegs: fxLoadXMMRegs65536 + "\n" + opOverLoadXMMRegs,
+ LoadXMMRegs: fxLoadXMMRegs + "\n" + opOverLoadXMMRegs,
Add: fxAdd,
- ClampAndScale: fxClampAndScale65536,
+ ClampAndScale: fxClampAndScale,
ConvertToInt32: fxConvertToInt32,
Store4: opOverStore4,
Store1: opOverStore1,
@@ -124,16 +124,16 @@
DstElemSize4: 4 * sizeOfUint8,
XMM3: fxXMM3,
XMM4: fxXMM4,
- XMM5: fxXMM5_256,
+ XMM5: fxXMM5,
XMM6: opSrcXMM6,
XMM8: opSrcXMM8,
XMM9: opSrcXMM9,
XMM10: opSrcXMM10,
LoadArgs: twoArgLoadArgs,
Setup: fxSetup,
- LoadXMMRegs: fxLoadXMMRegs256 + "\n" + opSrcLoadXMMRegs,
+ LoadXMMRegs: fxLoadXMMRegs + "\n" + opSrcLoadXMMRegs,
Add: fxAdd,
- ClampAndScale: fxClampAndScale256,
+ ClampAndScale: fxClampAndScale,
ConvertToInt32: fxConvertToInt32,
Store4: opSrcStore4,
Store1: opSrcStore1,
@@ -147,16 +147,16 @@
DstElemSize4: 4 * sizeOfUint32,
XMM3: fxXMM3,
XMM4: fxXMM4,
- XMM5: fxXMM5_65536,
+ XMM5: fxXMM5,
XMM6: maskXMM6,
XMM8: maskXMM8,
XMM9: maskXMM9,
XMM10: maskXMM10,
LoadArgs: oneArgLoadArgs,
Setup: fxSetup,
- LoadXMMRegs: fxLoadXMMRegs65536 + "\n" + maskLoadXMMRegs,
+ LoadXMMRegs: fxLoadXMMRegs + "\n" + maskLoadXMMRegs,
Add: fxAdd,
- ClampAndScale: fxClampAndScale65536,
+ ClampAndScale: fxClampAndScale,
ConvertToInt32: fxConvertToInt32,
Store4: maskStore4,
Store1: maskStore1,
@@ -168,7 +168,7 @@
Args: "dst []uint8, src []float32",
DstElemSize1: 1 * sizeOfUint8,
DstElemSize4: 4 * sizeOfUint8,
- XMM3: flXMM3_65536,
+ XMM3: flXMM3,
XMM4: flXMM4,
XMM5: flXMM5,
XMM6: opOverXMM6,
@@ -177,9 +177,9 @@
XMM10: opOverXMM10,
LoadArgs: twoArgLoadArgs,
Setup: flSetup,
- LoadXMMRegs: flLoadXMMRegs65536 + "\n" + opOverLoadXMMRegs,
+ LoadXMMRegs: flLoadXMMRegs + "\n" + opOverLoadXMMRegs,
Add: flAdd,
- ClampAndScale: flClampAndScale65536,
+ ClampAndScale: flClampAndScale,
ConvertToInt32: flConvertToInt32,
Store4: opOverStore4,
Store1: opOverStore1,
@@ -191,7 +191,7 @@
Args: "dst []uint8, src []float32",
DstElemSize1: 1 * sizeOfUint8,
DstElemSize4: 4 * sizeOfUint8,
- XMM3: flXMM3_256,
+ XMM3: flXMM3,
XMM4: flXMM4,
XMM5: flXMM5,
XMM6: opSrcXMM6,
@@ -200,9 +200,9 @@
XMM10: opSrcXMM10,
LoadArgs: twoArgLoadArgs,
Setup: flSetup,
- LoadXMMRegs: flLoadXMMRegs256 + "\n" + opSrcLoadXMMRegs,
+ LoadXMMRegs: flLoadXMMRegs + "\n" + opSrcLoadXMMRegs,
Add: flAdd,
- ClampAndScale: flClampAndScale256,
+ ClampAndScale: flClampAndScale,
ConvertToInt32: flConvertToInt32,
Store4: opSrcStore4,
Store1: opSrcStore1,
@@ -214,7 +214,7 @@
Args: "dst []uint32, src []float32",
DstElemSize1: 1 * sizeOfUint32,
DstElemSize4: 4 * sizeOfUint32,
- XMM3: flXMM3_65536,
+ XMM3: flXMM3,
XMM4: flXMM4,
XMM5: flXMM5,
XMM6: maskXMM6,
@@ -223,9 +223,9 @@
XMM10: maskXMM10,
LoadArgs: twoArgLoadArgs,
Setup: flSetup,
- LoadXMMRegs: flLoadXMMRegs65536 + "\n" + maskLoadXMMRegs,
+ LoadXMMRegs: flLoadXMMRegs + "\n" + maskLoadXMMRegs,
Add: flAdd,
- ClampAndScale: flClampAndScale65536,
+ ClampAndScale: flClampAndScale,
ConvertToInt32: flConvertToInt32,
Store4: maskStore4,
Store1: maskStore1,
@@ -241,16 +241,14 @@
sizeOfUint8 = 1
sizeOfUint32 = 4
- fxXMM3 = `-`
- flXMM3_256 = `flAlmost256`
- flXMM3_65536 = `flAlmost65536`
+ fxXMM3 = `-`
+ flXMM3 = `flAlmost65536`
fxXMM4 = `-`
flXMM4 = `flOne`
- fxXMM5_256 = `fxAlmost256`
- fxXMM5_65536 = `fxAlmost65536`
- flXMM5 = `flSignMask`
+ fxXMM5 = `fxAlmost65536`
+ flXMM5 = `flSignMask`
oneArgLoadArgs = `
MOVQ buf_base+0(FP), DI
@@ -278,23 +276,11 @@
MOVL AX, mxcsrNew-4(SP)
`
- fxLoadXMMRegs256 = `
- // fxAlmost256 := XMM(0x000000ff repeated four times) // Maximum of an uint8.
- MOVOU fxAlmost256<>(SB), X5
- `
- fxLoadXMMRegs65536 = `
+ fxLoadXMMRegs = `
// fxAlmost65536 := XMM(0x0000ffff repeated four times) // Maximum of an uint16.
MOVOU fxAlmost65536<>(SB), X5
`
- flLoadXMMRegs256 = `
- // flAlmost256 := XMM(0x437fffff repeated four times) // 255.99998 as a float32.
- // flOne := XMM(0x3f800000 repeated four times) // 1 as a float32.
- // flSignMask := XMM(0x7fffffff repeated four times) // All but the sign bit of a float32.
- MOVOU flAlmost256<>(SB), X3
- MOVOU flOne<>(SB), X4
- MOVOU flSignMask<>(SB), X5
- `
- flLoadXMMRegs65536 = `
+ flLoadXMMRegs = `
// flAlmost65536 := XMM(0x477fffff repeated four times) // 255.99998 * 256 as a float32.
// flOne := XMM(0x3f800000 repeated four times) // 1 as a float32.
// flSignMask := XMM(0x7fffffff repeated four times) // All but the sign bit of a float32.
@@ -306,23 +292,7 @@
fxAdd = `PADDD`
flAdd = `ADDPS`
- fxClampAndScale256 = `
- // y = abs(x)
- // y >>= 10 // Shift by 2*ϕ - 8.
- // y = min(y, fxAlmost256)
- //
- // pabsd %xmm1,%xmm2
- // psrld $0xa,%xmm2
- // pminud %xmm5,%xmm2
- //
- // Hopefully we'll get these opcode mnemonics into the assembler for Go
- // 1.8. https://golang.org/issue/16007 isn't exactly the same thing, but
- // it's similar.
- BYTE $0x66; BYTE $0x0f; BYTE $0x38; BYTE $0x1e; BYTE $0xd1
- BYTE $0x66; BYTE $0x0f; BYTE $0x72; BYTE $0xd2; BYTE $0x0a
- BYTE $0x66; BYTE $0x0f; BYTE $0x38; BYTE $0x3b; BYTE $0xd5
- `
- fxClampAndScale65536 = `
+ fxClampAndScale = `
// y = abs(x)
// y >>= 2 // Shift by 2*ϕ - 16.
// y = min(y, fxAlmost65536)
@@ -338,16 +308,7 @@
BYTE $0x66; BYTE $0x0f; BYTE $0x72; BYTE $0xd2; BYTE $0x02
BYTE $0x66; BYTE $0x0f; BYTE $0x38; BYTE $0x3b; BYTE $0xd5
`
- flClampAndScale256 = `
- // y = x & flSignMask
- // y = min(y, flOne)
- // y = mul(y, flAlmost256)
- MOVOU X5, X2
- ANDPS X1, X2
- MINPS X4, X2
- MULPS X3, X2
- `
- flClampAndScale65536 = `
+ flClampAndScale = `
// y = x & flSignMask
// y = min(y, flOne)
// y = mul(y, flAlmost65536)
@@ -402,18 +363,16 @@
BYTE $0x66; BYTE $0x45; BYTE $0x0f; BYTE $0xf4; BYTE $0xda
PSRLQ $47, X0
PSRLQ $47, X11
- // Merge the two registers back to one, X11.
+ // Merge the two registers back to one, X11, and add maskA.
PSLLQ $32, X11
XORPS X0, X11
- // Add maskA, shift from 16 bit color to 8 bit color.
- PADDD X11, X2
- PSRLQ $8, X2
- // As per opSrcStore4, shuffle and copy the low 4 bytes.
+ PADDD X11, X2
+ // As per opSrcStore4, shuffle and copy the 4 second-lowest bytes.
PSHUFB X6, X2
MOVL X2, (DI)
`
opSrcStore4 = `
- // z = shuffleTheLowBytesOfEach4ByteElement(z)
+ // z = shuffleTheSecondLowestBytesOfEach4ByteElement(z)
// copy(dst[:4], low4BytesOf(z))
PSHUFB X6, X2
MOVL X2, (DI)
@@ -444,8 +403,9 @@
MOVB R13, (DI)
`
opSrcStore1 = `
- // dst[0] = uint8(z)
+ // dst[0] = uint8(z>>8)
MOVL X2, BX
+ SHRL $8, BX
MOVB BX, (DI)
`
maskStore1 = `
diff --git a/vector/gen_acc_amd64.s.tmpl b/vector/gen_acc_amd64.s.tmpl
index 615d7a0..05ce25b 100644
--- a/vector/gen_acc_amd64.s.tmpl
+++ b/vector/gen_acc_amd64.s.tmpl
@@ -10,8 +10,6 @@
// 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
@@ -30,30 +28,26 @@
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.
+// 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
-// i???j???k???l???
+// ?i???j???k???l??
// to
// ijkl000000000000
-DATA gather<>+0x00(SB)/8, $0x808080800c080400
+DATA gather<>+0x00(SB)/8, $0x808080800d090501
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