| // Copyright 2025 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. |
| |
| //go:build goexperiment.simd && amd64 |
| |
| package simd_test |
| |
| import ( |
| "simd/archsimd" |
| "testing" |
| ) |
| |
| func TestShift(t *testing.T) { |
| if !archsimd.X86.AVX2() { |
| t.Skip("requires AVX2") |
| } |
| |
| testInt32x4Binary(t, |
| func(x, y archsimd.Int32x4) archsimd.Int32x4 { return x.ShiftLeft(y.AsUint32x4()) }, |
| map2(func(x, y int32) int32 { return x << uint32(y) })) |
| testInt32x4Binary(t, |
| func(x, y archsimd.Int32x4) archsimd.Int32x4 { return x.ShiftRight(y.AsUint32x4()) }, |
| map2(func(x, y int32) int32 { return x >> uint32(y) })) |
| testUint32x4Binary(t, |
| func(x, y archsimd.Uint32x4) archsimd.Uint32x4 { return x.ShiftRight(y) }, |
| map2(func(x, y uint32) uint32 { return x >> y })) |
| } |
| |
| func TestShiftAll(t *testing.T) { |
| // Test both const and non-const shifts. |
| // Test both regular and over-shifts. |
| |
| hide := hideConst[uint64] |
| |
| // ShiftAllLeft |
| |
| testInt32x4Unary(t, |
| func(x archsimd.Int32x4) archsimd.Int32x4 { return x.ShiftAllLeft(2) }, |
| map1(func(x int32) int32 { return x << 2 })) |
| testInt32x4Unary(t, |
| func(x archsimd.Int32x4) archsimd.Int32x4 { return x.ShiftAllLeft(hide(2)) }, |
| map1(func(x int32) int32 { return x << hide(2) })) |
| |
| // Ironically, we have to hide the constant in the want function so the |
| // compiler doesn't complain about a silly shift. |
| testInt32x4Unary(t, |
| func(x archsimd.Int32x4) archsimd.Int32x4 { return x.ShiftAllLeft(0x1000) }, |
| map1(func(x int32) int32 { return x << hide(0x1000) })) |
| testInt32x4Unary(t, |
| func(x archsimd.Int32x4) archsimd.Int32x4 { return x.ShiftAllLeft(hide(0x1000)) }, |
| map1(func(x int32) int32 { return x << hide(0x1000) })) |
| |
| // Signed ShiftAllRight |
| |
| testInt32x4Unary(t, |
| func(x archsimd.Int32x4) archsimd.Int32x4 { return x.ShiftAllRight(2) }, |
| map1(func(x int32) int32 { return x >> 2 })) |
| testInt32x4Unary(t, |
| func(x archsimd.Int32x4) archsimd.Int32x4 { return x.ShiftAllRight(hide(2)) }, |
| map1(func(x int32) int32 { return x >> hide(2) })) |
| |
| testInt32x4Unary(t, |
| func(x archsimd.Int32x4) archsimd.Int32x4 { return x.ShiftAllRight(0x1000) }, |
| map1(func(x int32) int32 { return x >> hide(0x1000) })) |
| testInt32x4Unary(t, |
| func(x archsimd.Int32x4) archsimd.Int32x4 { return x.ShiftAllRight(hide(0x1000)) }, |
| map1(func(x int32) int32 { return x >> hide(0x1000) })) |
| |
| // Unsigned ShiftAllRight |
| |
| testUint32x4Unary(t, |
| func(x archsimd.Uint32x4) archsimd.Uint32x4 { return x.ShiftAllRight(2) }, |
| map1(func(x uint32) uint32 { return x >> 2 })) |
| testUint32x4Unary(t, |
| func(x archsimd.Uint32x4) archsimd.Uint32x4 { return x.ShiftAllRight(hide(2)) }, |
| map1(func(x uint32) uint32 { return x >> hide(2) })) |
| |
| testUint32x4Unary(t, |
| func(x archsimd.Uint32x4) archsimd.Uint32x4 { return x.ShiftAllRight(0x1000) }, |
| map1(func(x uint32) uint32 { return x >> hide(0x1000) })) |
| testUint32x4Unary(t, |
| func(x archsimd.Uint32x4) archsimd.Uint32x4 { return x.ShiftAllRight(hide(0x1000)) }, |
| map1(func(x uint32) uint32 { return x >> hide(0x1000) })) |
| } |
| |
| func concatInt32s(x, y int32) int64 { |
| return (int64(x) << 32) | int64(uint32(y)) |
| } |
| |
| func concatUint32s(x, y uint32) uint64 { |
| return (uint64(x) << 32) | uint64(y) |
| } |
| |
| func TestShiftAllConcat(t *testing.T) { |
| if !archsimd.X86.AVX512VBMI2() { |
| t.Skip("requires AVX512-VBMI2") |
| } |
| |
| // Note that unlike their non-Concat counterparts, these wrap the shift count. |
| |
| hide := hideConst[uint8] |
| |
| // ShiftAllLeftConcat |
| salc := func(shift uint8) func(x, y int32) int32 { |
| return func(x, y int32) int32 { |
| return int32(concatInt32s(x, y) >> (32 - shift%32)) |
| } |
| } |
| |
| testInt32x4Binary(t, |
| func(x, y archsimd.Int32x4) archsimd.Int32x4 { return x.ShiftAllLeftConcat(2, y) }, |
| map2(salc(2))) |
| testInt32x4Binary(t, |
| func(x, y archsimd.Int32x4) archsimd.Int32x4 { return x.ShiftAllLeftConcat(hide(2), y) }, |
| map2(salc(hide(2)))) |
| |
| // TODO: If we expand the shift from uint8, add larger cases (e.g., 0x1000). |
| // The uint8 conversion is only here so the build will fail if we change it. |
| |
| testInt32x4Binary(t, |
| func(x, y archsimd.Int32x4) archsimd.Int32x4 { return x.ShiftAllLeftConcat(uint8(128), y) }, |
| map2(salc(128))) |
| testInt32x4Binary(t, |
| func(x, y archsimd.Int32x4) archsimd.Int32x4 { return x.ShiftAllLeftConcat(hide(128), y) }, |
| map2(salc(hide(128)))) |
| |
| // Signed ShiftAllRightConcat |
| sarc := func(shift uint8) func(x, y int32) int32 { |
| return func(x, y int32) int32 { |
| return int32(concatInt32s(y, x) >> (shift % 32)) |
| } |
| } |
| |
| testInt32x4Binary(t, |
| func(x, y archsimd.Int32x4) archsimd.Int32x4 { return x.ShiftAllRightConcat(2, y) }, |
| map2(sarc(2))) |
| testInt32x4Binary(t, |
| func(x, y archsimd.Int32x4) archsimd.Int32x4 { return x.ShiftAllRightConcat(hide(2), y) }, |
| map2(sarc(hide(2)))) |
| |
| testInt32x4Binary(t, |
| func(x, y archsimd.Int32x4) archsimd.Int32x4 { return x.ShiftAllRightConcat(128, y) }, |
| map2(sarc(128))) |
| testInt32x4Binary(t, |
| func(x, y archsimd.Int32x4) archsimd.Int32x4 { return x.ShiftAllRightConcat(hide(128), y) }, |
| map2(sarc(hide(128)))) |
| |
| // Unsigned ShiftAllRightConcat |
| usarc := func(shift uint8) func(x, y uint32) uint32 { |
| return func(x, y uint32) uint32 { |
| return uint32(concatUint32s(y, x) >> (shift % 32)) |
| } |
| } |
| |
| testUint32x4Binary(t, |
| func(x, y archsimd.Uint32x4) archsimd.Uint32x4 { return x.ShiftAllRightConcat(2, y) }, |
| map2(usarc(2))) |
| testUint32x4Binary(t, |
| func(x, y archsimd.Uint32x4) archsimd.Uint32x4 { return x.ShiftAllRightConcat(hide(2), y) }, |
| map2(usarc(hide(2)))) |
| |
| testUint32x4Binary(t, |
| func(x, y archsimd.Uint32x4) archsimd.Uint32x4 { return x.ShiftAllRightConcat(128, y) }, |
| map2(usarc(128))) |
| testUint32x4Binary(t, |
| func(x, y archsimd.Uint32x4) archsimd.Uint32x4 { return x.ShiftAllRightConcat(hide(128), y) }, |
| map2(usarc(hide(128)))) |
| } |
| |
| func TestShiftConcat(t *testing.T) { |
| if !archsimd.X86.AVX512VBMI2() { |
| t.Skip("requires AVX512-VBMI2") |
| } |
| |
| // Note that unlike their non-Concat counterparts, these wrap the shift count. |
| |
| testInt32x4Ternary(t, |
| func(x, y, z archsimd.Int32x4) archsimd.Int32x4 { return x.ShiftLeftConcat(y, z.AsUint32x4()) }, |
| map3(func(x, y, z int32) int32 { |
| return int32(concatInt32s(x, y) >> (32 - uint32(z)%32)) |
| })) |
| |
| testInt32x4Ternary(t, |
| func(x, y, z archsimd.Int32x4) archsimd.Int32x4 { return x.ShiftRightConcat(y, z.AsUint32x4()) }, |
| map3(func(x, y, z int32) int32 { |
| return int32(concatInt32s(y, x) >> (uint32(z) % 32)) |
| })) |
| |
| testUint32x4Ternary(t, |
| func(x, y, z archsimd.Uint32x4) archsimd.Uint32x4 { return x.ShiftRightConcat(y, z) }, |
| map3(func(x, y, z uint32) uint32 { |
| return uint32(concatUint32s(y, x) >> (z % 32)) |
| })) |
| } |
| |
| func TestConcatShiftBytesRight(t *testing.T) { |
| hide := hideConst[uint8] |
| |
| csbr := func(shift uint8) func(x, y []uint8) []uint8 { |
| return func(x, y []uint8) []uint8 { |
| z := make([]uint8, len(x)) |
| for i := range z { |
| target := i + int(shift) |
| if target < 16 { |
| z[i] = y[target] |
| } else if target < 32 { |
| z[i] = x[(target - 16)] |
| } |
| } |
| return z |
| } |
| } |
| |
| t.Run("Uint8x16", func(t *testing.T) { |
| if !archsimd.X86.AVX() { |
| t.Skip("requires AVX") |
| } |
| // TODO: If we expand the shift from uint8, add larger cases (e.g., 0x1000) |
| for _, shift := range []uint8{0, 2, 16, 20, 32, 128} { |
| t.Log("shift", shift) |
| testUint8x16Binary(t, |
| func(x, y archsimd.Uint8x16) archsimd.Uint8x16 { return x.ConcatShiftBytesRight(shift, y) }, |
| csbr(shift)) |
| testUint8x16Binary(t, |
| func(x, y archsimd.Uint8x16) archsimd.Uint8x16 { return x.ConcatShiftBytesRight(hide(shift), y) }, |
| csbr(hide(shift))) |
| } |
| }) |
| |
| t.Run("Uint8x32", func(t *testing.T) { |
| if !archsimd.X86.AVX2() { |
| t.Skip("requires AVX2") |
| } |
| for _, shift := range []uint8{0, 2, 16, 20, 32, 128} { |
| t.Log("shift", shift) |
| testUint8x32Binary(t, |
| func(x, y archsimd.Uint8x32) archsimd.Uint8x32 { return x.ConcatShiftBytesRightGrouped(shift, y) }, |
| grouped2(csbr(shift))) |
| testUint8x32Binary(t, |
| func(x, y archsimd.Uint8x32) archsimd.Uint8x32 { return x.ConcatShiftBytesRightGrouped(hide(shift), y) }, |
| grouped2(csbr(hide(shift)))) |
| } |
| }) |
| |
| t.Run("Uint8x64", func(t *testing.T) { |
| if !archsimd.X86.AVX512() { |
| t.Skip("requires AVX512") |
| } |
| for _, shift := range []uint8{0, 2, 16, 20, 32, 128} { |
| t.Log("shift", shift) |
| testUint8x64Binary(t, |
| func(x, y archsimd.Uint8x64) archsimd.Uint8x64 { return x.ConcatShiftBytesRightGrouped(shift, y) }, |
| grouped2(csbr(shift))) |
| testUint8x64Binary(t, |
| func(x, y archsimd.Uint8x64) archsimd.Uint8x64 { return x.ConcatShiftBytesRightGrouped(hide(shift), y) }, |
| grouped2(csbr(hide(shift)))) |
| } |
| }) |
| } |
