unary_amd64_test.go

     1  // Copyright 2025 The Go Authors. All rights reserved.
     2  // Use of this source code is governed by a BSD-style
     3  // license that can be found in the LICENSE file.
     4  
     5  //go:build goexperiment.simd && amd64
     6  
     7  package simd_test
     8  
     9  import (
    10  	"math"
    11  	"simd/archsimd"
    12  	"testing"
    13  )
    14  
    15  func TestAbsAMD64(t *testing.T) {
    16  	testFloat32x8Unary(t, archsimd.Float32x8.Abs, map1[float32](abs))
    17  	testFloat64x4Unary(t, archsimd.Float64x4.Abs, map1[float64](abs))
    18  }
    19  
    20  func TestNegAMD64(t *testing.T) {
    21  	testFloat32x8Unary(t, archsimd.Float32x8.Neg, map1[float32](neg))
    22  	testFloat64x4Unary(t, archsimd.Float64x4.Neg, map1[float64](neg))
    23  	if archsimd.X86.AVX2() {
    24  		testInt8x32Unary(t, archsimd.Int8x32.Neg, map1[int8](neg))
    25  		testInt16x16Unary(t, archsimd.Int16x16.Neg, map1[int16](neg))
    26  		testInt32x8Unary(t, archsimd.Int32x8.Neg, map1[int32](neg))
    27  		testInt64x4Unary(t, archsimd.Int64x4.Neg, map1[int64](neg))
    28  	}
    29  	if archsimd.X86.AVX512() {
    30  		testFloat32x16Unary(t, archsimd.Float32x16.Neg, map1[float32](neg))
    31  		testFloat64x8Unary(t, archsimd.Float64x8.Neg, map1[float64](neg))
    32  		testInt8x64Unary(t, archsimd.Int8x64.Neg, map1[int8](neg))
    33  		testInt16x32Unary(t, archsimd.Int16x32.Neg, map1[int16](neg))
    34  		testInt32x16Unary(t, archsimd.Int32x16.Neg, map1[int32](neg))
    35  		testInt64x8Unary(t, archsimd.Int64x8.Neg, map1[int64](neg))
    36  	}
    37  }
    38  
    39  func TestCeilAMD64(t *testing.T) {
    40  	testFloat32x8Unary(t, archsimd.Float32x8.Ceil, ceilSlice[float32])
    41  	testFloat64x4Unary(t, archsimd.Float64x4.Ceil, ceilSlice[float64])
    42  	if archsimd.X86.AVX512() {
    43  		// testFloat32x16Unary(t, archsimd.Float32x16.Ceil, ceilSlice[float32]) // missing
    44  		// testFloat64x8Unary(t, archsimd.Float64x8.Ceil, ceilSlice[float64])   // missing
    45  	}
    46  }
    47  
    48  func TestFloorAMD64(t *testing.T) {
    49  	testFloat32x8Unary(t, archsimd.Float32x8.Floor, floorSlice[float32])
    50  	testFloat64x4Unary(t, archsimd.Float64x4.Floor, floorSlice[float64])
    51  	if archsimd.X86.AVX512() {
    52  		// testFloat32x16Unary(t, archsimd.Float32x16.Floor, floorSlice[float32]) // missing
    53  		// testFloat64x8Unary(t, archsimd.Float64x8.Floor, floorSlice[float64])   // missing
    54  	}
    55  }
    56  
    57  func TestTruncAMD64(t *testing.T) {
    58  	testFloat32x8Unary(t, archsimd.Float32x8.Trunc, truncSlice[float32])
    59  	testFloat64x4Unary(t, archsimd.Float64x4.Trunc, truncSlice[float64])
    60  	if archsimd.X86.AVX512() {
    61  		// testFloat32x16Unary(t, archsimd.Float32x16.Trunc, truncSlice[float32]) // missing
    62  		// testFloat64x8Unary(t, archsimd.Float64x8.Trunc, truncSlice[float64])   // missing
    63  	}
    64  }
    65  
    66  func TestRoundAMD64(t *testing.T) {
    67  	testFloat32x8Unary(t, archsimd.Float32x8.Round, roundSlice[float32])
    68  	testFloat64x4Unary(t, archsimd.Float64x4.Round, roundSlice[float64])
    69  	if archsimd.X86.AVX512() {
    70  		// testFloat32x16Unary(t, archsimd.Float32x16.Round, roundSlice[float32]) // missing
    71  		// testFloat64x8Unary(t, archsimd.Float64x8.Round, roundSlice[float64])   // missing
    72  	}
    73  }
    74  
    75  func TestSqrtAMD64(t *testing.T) {
    76  	testFloat32x8Unary(t, archsimd.Float32x8.Sqrt, sqrtSlice[float32])
    77  	testFloat64x4Unary(t, archsimd.Float64x4.Sqrt, sqrtSlice[float64])
    78  	if archsimd.X86.AVX512() {
    79  		testFloat32x16Unary(t, archsimd.Float32x16.Sqrt, sqrtSlice[float32])
    80  		testFloat64x8Unary(t, archsimd.Float64x8.Sqrt, sqrtSlice[float64])
    81  	}
    82  }
    83  
    84  func TestNotAMD64(t *testing.T) {
    85  	testInt16x8Unary(t, archsimd.Int16x8.Not, map1[int16](not))
    86  
    87  	if archsimd.X86.AVX2() {
    88  		testInt8x32Unary(t, archsimd.Int8x32.Not, map1[int8](not))
    89  		testInt16x16Unary(t, archsimd.Int16x16.Not, map1[int16](not))
    90  		testInt32x8Unary(t, archsimd.Int32x8.Not, map1[int32](not))
    91  	}
    92  }
    93  
    94  func TestCeilScaledResidue(t *testing.T) {
    95  	if !archsimd.X86.AVX512() {
    96  		t.Skip("Needs AVX512")
    97  	}
    98  	testFloat64x8UnaryFlaky(t,
    99  		func(x archsimd.Float64x8) archsimd.Float64x8 { return x.CeilScaledResidue(0) },
   100  		map1(ceilResidueForPrecision[float64](0)),
   101  		0.001)
   102  	testFloat64x8UnaryFlaky(t,
   103  		func(x archsimd.Float64x8) archsimd.Float64x8 { return x.CeilScaledResidue(1) },
   104  		map1(ceilResidueForPrecision[float64](1)),
   105  		0.001)
   106  	testFloat64x8Unary(t,
   107  		func(x archsimd.Float64x8) archsimd.Float64x8 { return x.Sub(x.CeilScaled(0)) },
   108  		map1[float64](func(x float64) float64 { return x - math.Ceil(x) }))
   109  }
   110  
   111  func TestConvert(t *testing.T) {
   112  	testFloat64x2ConvertToFloat32(t, archsimd.Float64x2.ConvertToFloat32, map1n[float64](toFloat32, 4))
   113  	testFloat64x4ConvertToFloat32(t, archsimd.Float64x4.ConvertToFloat32, map1[float64](toFloat32))
   114  	testFloat32x4ConvertToFloat64(t, archsimd.Float32x4.ConvertToFloat64, map1[float32](toFloat64))
   115  
   116  	testFloat32x4ConvertToInt32(t, archsimd.Float32x4.ConvertToInt32, map1[float32](floatToInt32_x86))
   117  	testFloat32x8ConvertToInt32(t, archsimd.Float32x8.ConvertToInt32, map1[float32](floatToInt32_x86))
   118  	testFloat64x2ConvertToInt32(t, archsimd.Float64x2.ConvertToInt32, map1n[float64](floatToInt32_x86, 4))
   119  	testFloat64x4ConvertToInt32(t, archsimd.Float64x4.ConvertToInt32, map1[float64](floatToInt32_x86))
   120  
   121  	testInt32x4ConvertToFloat32(t, archsimd.Int32x4.ConvertToFloat32, map1[int32](toFloat32))
   122  	testInt32x8ConvertToFloat32(t, archsimd.Int32x8.ConvertToFloat32, map1[int32](toFloat32))
   123  	testInt32x4ConvertToFloat64(t, archsimd.Int32x4.ConvertToFloat64, map1[int32](toFloat64))
   124  
   125  	if archsimd.X86.AVX512() {
   126  		testFloat32x8ConvertToFloat64(t, archsimd.Float32x8.ConvertToFloat64, map1[float32](toFloat64))
   127  		testFloat64x8ConvertToFloat32(t, archsimd.Float64x8.ConvertToFloat32, map1[float64](toFloat32))
   128  
   129  		testFloat32x16ConvertToInt32(t, archsimd.Float32x16.ConvertToInt32, map1[float32](floatToInt32_x86))
   130  		testFloat64x8ConvertToInt32(t, archsimd.Float64x8.ConvertToInt32, map1[float64](floatToInt32_x86))
   131  		testFloat32x4ConvertToInt64(t, archsimd.Float32x4.ConvertToInt64, map1[float32](floatToInt64_x86))
   132  		testFloat32x8ConvertToInt64(t, archsimd.Float32x8.ConvertToInt64, map1[float32](floatToInt64_x86))
   133  		testFloat64x2ConvertToInt64(t, archsimd.Float64x2.ConvertToInt64, map1[float64](floatToInt64_x86))
   134  		testFloat64x4ConvertToInt64(t, archsimd.Float64x4.ConvertToInt64, map1[float64](floatToInt64_x86))
   135  		testFloat64x8ConvertToInt64(t, archsimd.Float64x8.ConvertToInt64, map1[float64](floatToInt64_x86))
   136  
   137  		testFloat32x4ConvertToUint32(t, archsimd.Float32x4.ConvertToUint32, map1[float32](floatToUint32_x86))
   138  		testFloat32x8ConvertToUint32(t, archsimd.Float32x8.ConvertToUint32, map1[float32](floatToUint32_x86))
   139  		testFloat32x16ConvertToUint32(t, archsimd.Float32x16.ConvertToUint32, map1[float32](floatToUint32_x86))
   140  		testFloat64x2ConvertToUint32(t, archsimd.Float64x2.ConvertToUint32, map1n[float64](floatToUint32_x86, 4))
   141  		testFloat64x4ConvertToUint32(t, archsimd.Float64x4.ConvertToUint32, map1[float64](floatToUint32_x86))
   142  		testFloat64x8ConvertToUint32(t, archsimd.Float64x8.ConvertToUint32, map1[float64](floatToUint32_x86))
   143  		testFloat32x4ConvertToUint64(t, archsimd.Float32x4.ConvertToUint64, map1[float32](floatToUint64_x86))
   144  		testFloat32x8ConvertToUint64(t, archsimd.Float32x8.ConvertToUint64, map1[float32](floatToUint64_x86))
   145  		testFloat64x2ConvertToUint64(t, archsimd.Float64x2.ConvertToUint64, map1[float64](floatToUint64_x86))
   146  		testFloat64x4ConvertToUint64(t, archsimd.Float64x4.ConvertToUint64, map1[float64](floatToUint64_x86))
   147  		testFloat64x8ConvertToUint64(t, archsimd.Float64x8.ConvertToUint64, map1[float64](floatToUint64_x86))
   148  
   149  		testInt32x16ConvertToFloat32(t, archsimd.Int32x16.ConvertToFloat32, map1[int32](toFloat32))
   150  		testInt64x2ConvertToFloat32(t, archsimd.Int64x2.ConvertToFloat32, map1n[int64](toFloat32, 4))
   151  		testInt64x4ConvertToFloat32(t, archsimd.Int64x4.ConvertToFloat32, map1[int64](toFloat32))
   152  		testInt64x8ConvertToFloat32(t, archsimd.Int64x8.ConvertToFloat32, map1[int64](toFloat32))
   153  		testInt64x2ConvertToFloat64(t, archsimd.Int64x2.ConvertToFloat64, map1[int64](toFloat64))
   154  		testInt64x4ConvertToFloat64(t, archsimd.Int64x4.ConvertToFloat64, map1[int64](toFloat64))
   155  		testInt64x8ConvertToFloat64(t, archsimd.Int64x8.ConvertToFloat64, map1[int64](toFloat64))
   156  
   157  		testUint32x4ConvertToFloat32(t, archsimd.Uint32x4.ConvertToFloat32, map1[uint32](toFloat32))
   158  		testUint32x8ConvertToFloat32(t, archsimd.Uint32x8.ConvertToFloat32, map1[uint32](toFloat32))
   159  		testUint32x16ConvertToFloat32(t, archsimd.Uint32x16.ConvertToFloat32, map1[uint32](toFloat32))
   160  		testUint64x2ConvertToFloat32(t, archsimd.Uint64x2.ConvertToFloat32, map1n[uint64](toFloat32, 4))
   161  		testUint64x4ConvertToFloat32(t, archsimd.Uint64x4.ConvertToFloat32, map1[uint64](toFloat32))
   162  		testUint64x8ConvertToFloat32(t, archsimd.Uint64x8.ConvertToFloat32, map1[uint64](toFloat32))
   163  		testUint32x4ConvertToFloat64(t, archsimd.Uint32x4.ConvertToFloat64, map1[uint32](toFloat64))
   164  		testUint32x8ConvertToFloat64(t, archsimd.Uint32x8.ConvertToFloat64, map1[uint32](toFloat64))
   165  		testUint64x2ConvertToFloat64(t, archsimd.Uint64x2.ConvertToFloat64, map1[uint64](toFloat64))
   166  		testUint64x4ConvertToFloat64(t, archsimd.Uint64x4.ConvertToFloat64, map1[uint64](toFloat64))
   167  		testUint64x8ConvertToFloat64(t, archsimd.Uint64x8.ConvertToFloat64, map1[uint64](toFloat64))
   168  	}
   169  }
   170  
   171  func TestExtend(t *testing.T) {
   172  	if archsimd.X86.AVX2() {
   173  		testInt8x16ConvertToInt16(t, archsimd.Int8x16.ExtendToInt16, map1[int8](toInt16))
   174  		testInt16x8ConvertToInt32(t, archsimd.Int16x8.ExtendToInt32, map1[int16](toInt32))
   175  		testInt32x4ConvertToInt64(t, archsimd.Int32x4.ExtendToInt64, map1[int32](toInt64))
   176  		testUint8x16ConvertToUint16(t, archsimd.Uint8x16.ExtendToUint16, map1[uint8](toUint16))
   177  		testUint16x8ConvertToUint32(t, archsimd.Uint16x8.ExtendToUint32, map1[uint16](toUint32))
   178  		testUint32x4ConvertToUint64(t, archsimd.Uint32x4.ExtendToUint64, map1[uint32](toUint64))
   179  	}
   180  
   181  	if archsimd.X86.AVX512() {
   182  		testInt8x32ConvertToInt16(t, archsimd.Int8x32.ExtendToInt16, map1[int8](toInt16))
   183  		testInt8x16ConvertToInt32(t, archsimd.Int8x16.ExtendToInt32, map1[int8](toInt32))
   184  		testInt16x16ConvertToInt32(t, archsimd.Int16x16.ExtendToInt32, map1[int16](toInt32))
   185  		testInt16x8ConvertToInt64(t, archsimd.Int16x8.ExtendToInt64, map1[int16](toInt64))
   186  		testInt32x8ConvertToInt64(t, archsimd.Int32x8.ExtendToInt64, map1[int32](toInt64))
   187  		testUint8x32ConvertToUint16(t, archsimd.Uint8x32.ExtendToUint16, map1[uint8](toUint16))
   188  		testUint8x16ConvertToUint32(t, archsimd.Uint8x16.ExtendToUint32, map1[uint8](toUint32))
   189  		testUint16x16ConvertToUint32(t, archsimd.Uint16x16.ExtendToUint32, map1[uint16](toUint32))
   190  		testUint16x8ConvertToUint64(t, archsimd.Uint16x8.ExtendToUint64, map1[uint16](toUint64))
   191  		testUint32x8ConvertToUint64(t, archsimd.Uint32x8.ExtendToUint64, map1[uint32](toUint64))
   192  	}
   193  }
   194  
   195  func TestExtendLo(t *testing.T) {
   196  	testInt8x16ConvertLoToInt64x2(t, archsimd.Int8x16.ExtendLo2ToInt64, map1n[int8](toInt64, 2))
   197  	testInt16x8ConvertLoToInt64x2(t, archsimd.Int16x8.ExtendLo2ToInt64, map1n[int16](toInt64, 2))
   198  	testInt32x4ConvertLoToInt64x2(t, archsimd.Int32x4.ExtendLo2ToInt64, map1n[int32](toInt64, 2))
   199  	testUint8x16ConvertLoToUint64x2(t, archsimd.Uint8x16.ExtendLo2ToUint64, map1n[uint8](toUint64, 2))
   200  	testUint16x8ConvertLoToUint64x2(t, archsimd.Uint16x8.ExtendLo2ToUint64, map1n[uint16](toUint64, 2))
   201  	testUint32x4ConvertLoToUint64x2(t, archsimd.Uint32x4.ExtendLo2ToUint64, map1n[uint32](toUint64, 2))
   202  	testInt8x16ConvertLoToInt32x4(t, archsimd.Int8x16.ExtendLo4ToInt32, map1n[int8](toInt32, 4))
   203  	testInt16x8ConvertLoToInt32x4(t, archsimd.Int16x8.ExtendLo4ToInt32, map1n[int16](toInt32, 4))
   204  	testUint8x16ConvertLoToUint32x4(t, archsimd.Uint8x16.ExtendLo4ToUint32, map1n[uint8](toUint32, 4))
   205  	testUint16x8ConvertLoToUint32x4(t, archsimd.Uint16x8.ExtendLo4ToUint32, map1n[uint16](toUint32, 4))
   206  	testInt8x16ConvertLoToInt16x8(t, archsimd.Int8x16.ExtendLo8ToInt16, map1n[int8](toInt16, 8))
   207  	testUint8x16ConvertLoToUint16x8(t, archsimd.Uint8x16.ExtendLo8ToUint16, map1n[uint8](toUint16, 8))
   208  
   209  	if archsimd.X86.AVX2() {
   210  		testInt8x16ConvertLoToInt64x4(t, archsimd.Int8x16.ExtendLo4ToInt64, map1n[int8](toInt64, 4))
   211  		testInt16x8ConvertLoToInt64x4(t, archsimd.Int16x8.ExtendLo4ToInt64, map1n[int16](toInt64, 4))
   212  		testUint8x16ConvertLoToUint64x4(t, archsimd.Uint8x16.ExtendLo4ToUint64, map1n[uint8](toUint64, 4))
   213  		testUint16x8ConvertLoToUint64x4(t, archsimd.Uint16x8.ExtendLo4ToUint64, map1n[uint16](toUint64, 4))
   214  		testInt8x16ConvertLoToInt32x8(t, archsimd.Int8x16.ExtendLo8ToInt32, map1n[int8](toInt32, 8))
   215  		testUint8x16ConvertLoToUint32x8(t, archsimd.Uint8x16.ExtendLo8ToUint32, map1n[uint8](toUint32, 8))
   216  	}
   217  
   218  	if archsimd.X86.AVX512() {
   219  		testInt8x16ConvertToInt64(t, archsimd.Int8x16.ExtendLo8ToInt64, map1n[int8](toInt64, 8))
   220  		testUint8x16ConvertToUint64(t, archsimd.Uint8x16.ExtendLo8ToUint64, map1n[uint8](toUint64, 8))
   221  	}
   222  }
   223  
   224  func TestTruncate(t *testing.T) {
   225  	if archsimd.X86.AVX512() {
   226  		testInt16x8ConvertToInt8(t, archsimd.Int16x8.TruncToInt8, map1n[int16](toInt8, 16))
   227  		testInt16x16ConvertToInt8(t, archsimd.Int16x16.TruncToInt8, map1[int16](toInt8))
   228  		testInt16x32ConvertToInt8(t, archsimd.Int16x32.TruncToInt8, map1[int16](toInt8))
   229  		testInt32x4ConvertToInt8(t, archsimd.Int32x4.TruncToInt8, map1n[int32](toInt8, 16))
   230  		testInt32x8ConvertToInt8(t, archsimd.Int32x8.TruncToInt8, map1n[int32](toInt8, 16))
   231  		testInt32x16ConvertToInt8(t, archsimd.Int32x16.TruncToInt8, map1[int32](toInt8))
   232  		testInt64x2ConvertToInt8(t, archsimd.Int64x2.TruncToInt8, map1n[int64](toInt8, 16))
   233  		testInt64x4ConvertToInt8(t, archsimd.Int64x4.TruncToInt8, map1n[int64](toInt8, 16))
   234  		testInt64x8ConvertToInt8(t, archsimd.Int64x8.TruncToInt8, map1n[int64](toInt8, 16))
   235  		testInt32x4ConvertToInt16(t, archsimd.Int32x4.TruncToInt16, map1n[int32](toInt16, 8))
   236  		testInt32x8ConvertToInt16(t, archsimd.Int32x8.TruncToInt16, map1[int32](toInt16))
   237  		testInt32x16ConvertToInt16(t, archsimd.Int32x16.TruncToInt16, map1[int32](toInt16))
   238  		testInt64x2ConvertToInt16(t, archsimd.Int64x2.TruncToInt16, map1n[int64](toInt16, 8))
   239  		testInt64x4ConvertToInt16(t, archsimd.Int64x4.TruncToInt16, map1n[int64](toInt16, 8))
   240  		testInt64x8ConvertToInt16(t, archsimd.Int64x8.TruncToInt16, map1[int64](toInt16))
   241  		testInt64x2ConvertToInt32(t, archsimd.Int64x2.TruncToInt32, map1n[int64](toInt32, 4))
   242  		testInt64x4ConvertToInt32(t, archsimd.Int64x4.TruncToInt32, map1[int64](toInt32))
   243  		testInt64x8ConvertToInt32(t, archsimd.Int64x8.TruncToInt32, map1[int64](toInt32))
   244  
   245  		testUint16x8ConvertToUint8(t, archsimd.Uint16x8.TruncToUint8, map1n[uint16](toUint8, 16))
   246  		testUint16x16ConvertToUint8(t, archsimd.Uint16x16.TruncToUint8, map1[uint16](toUint8))
   247  		testUint16x32ConvertToUint8(t, archsimd.Uint16x32.TruncToUint8, map1[uint16](toUint8))
   248  		testUint32x4ConvertToUint8(t, archsimd.Uint32x4.TruncToUint8, map1n[uint32](toUint8, 16))
   249  		testUint32x8ConvertToUint8(t, archsimd.Uint32x8.TruncToUint8, map1n[uint32](toUint8, 16))
   250  		testUint32x16ConvertToUint8(t, archsimd.Uint32x16.TruncToUint8, map1[uint32](toUint8))
   251  		testUint64x2ConvertToUint8(t, archsimd.Uint64x2.TruncToUint8, map1n[uint64](toUint8, 16))
   252  		testUint64x4ConvertToUint8(t, archsimd.Uint64x4.TruncToUint8, map1n[uint64](toUint8, 16))
   253  		testUint64x8ConvertToUint8(t, archsimd.Uint64x8.TruncToUint8, map1n[uint64](toUint8, 16))
   254  		testUint32x4ConvertToUint16(t, archsimd.Uint32x4.TruncToUint16, map1n[uint32](toUint16, 8))
   255  		testUint32x8ConvertToUint16(t, archsimd.Uint32x8.TruncToUint16, map1[uint32](toUint16))
   256  		testUint32x16ConvertToUint16(t, archsimd.Uint32x16.TruncToUint16, map1[uint32](toUint16))
   257  		testUint64x2ConvertToUint16(t, archsimd.Uint64x2.TruncToUint16, map1n[uint64](toUint16, 8))
   258  		testUint64x4ConvertToUint16(t, archsimd.Uint64x4.TruncToUint16, map1n[uint64](toUint16, 8))
   259  		testUint64x8ConvertToUint16(t, archsimd.Uint64x8.TruncToUint16, map1[uint64](toUint16))
   260  		testUint64x2ConvertToUint32(t, archsimd.Uint64x2.TruncToUint32, map1n[uint64](toUint32, 4))
   261  		testUint64x4ConvertToUint32(t, archsimd.Uint64x4.TruncToUint32, map1[uint64](toUint32))
   262  		testUint64x8ConvertToUint32(t, archsimd.Uint64x8.TruncToUint32, map1[uint64](toUint32))
   263  	}
   264  }
   265  
   266  func TestSaturate(t *testing.T) {
   267  	if archsimd.X86.AVX512() {
   268  		testInt16x8ConvertToInt8(t, archsimd.Int16x8.SaturateToInt8, map1n[int16](satToInt8, 16))
   269  		testInt16x16ConvertToInt8(t, archsimd.Int16x16.SaturateToInt8, map1[int16](satToInt8))
   270  		testInt16x32ConvertToInt8(t, archsimd.Int16x32.SaturateToInt8, map1[int16](satToInt8))
   271  		testInt32x4ConvertToInt8(t, archsimd.Int32x4.SaturateToInt8, map1n[int32](satToInt8, 16))
   272  		testInt32x8ConvertToInt8(t, archsimd.Int32x8.SaturateToInt8, map1n[int32](satToInt8, 16))
   273  		testInt32x16ConvertToInt8(t, archsimd.Int32x16.SaturateToInt8, map1[int32](satToInt8))
   274  		testInt64x2ConvertToInt8(t, archsimd.Int64x2.SaturateToInt8, map1n[int64](satToInt8, 16))
   275  		testInt64x4ConvertToInt8(t, archsimd.Int64x4.SaturateToInt8, map1n[int64](satToInt8, 16))
   276  		testInt64x8ConvertToInt8(t, archsimd.Int64x8.SaturateToInt8, map1n[int64](satToInt8, 16))
   277  		testInt32x4ConvertToInt16(t, archsimd.Int32x4.SaturateToInt16, map1n[int32](satToInt16, 8))
   278  		testInt32x8ConvertToInt16(t, archsimd.Int32x8.SaturateToInt16, map1[int32](satToInt16))
   279  		testInt32x16ConvertToInt16(t, archsimd.Int32x16.SaturateToInt16, map1[int32](satToInt16))
   280  		testInt64x2ConvertToInt16(t, archsimd.Int64x2.SaturateToInt16, map1n[int64](satToInt16, 8))
   281  		testInt64x4ConvertToInt16(t, archsimd.Int64x4.SaturateToInt16, map1n[int64](satToInt16, 8))
   282  		testInt64x8ConvertToInt16(t, archsimd.Int64x8.SaturateToInt16, map1[int64](satToInt16))
   283  		testInt64x2ConvertToInt32(t, archsimd.Int64x2.SaturateToInt32, map1n[int64](satToInt32, 4))
   284  		testInt64x4ConvertToInt32(t, archsimd.Int64x4.SaturateToInt32, map1[int64](satToInt32))
   285  		testInt64x8ConvertToInt32(t, archsimd.Int64x8.SaturateToInt32, map1[int64](satToInt32))
   286  
   287  		testUint16x8ConvertToUint8(t, archsimd.Uint16x8.SaturateToUint8, map1n[uint16](satToUint8, 16))
   288  		testUint16x16ConvertToUint8(t, archsimd.Uint16x16.SaturateToUint8, map1[uint16](satToUint8))
   289  		testUint16x32ConvertToUint8(t, archsimd.Uint16x32.SaturateToUint8, map1[uint16](satToUint8))
   290  		testUint32x4ConvertToUint8(t, archsimd.Uint32x4.SaturateToUint8, map1n[uint32](satToUint8, 16))
   291  		testUint32x8ConvertToUint8(t, archsimd.Uint32x8.SaturateToUint8, map1n[uint32](satToUint8, 16))
   292  		testUint32x16ConvertToUint8(t, archsimd.Uint32x16.SaturateToUint8, map1[uint32](satToUint8))
   293  		testUint64x2ConvertToUint8(t, archsimd.Uint64x2.SaturateToUint8, map1n[uint64](satToUint8, 16))
   294  		testUint64x4ConvertToUint8(t, archsimd.Uint64x4.SaturateToUint8, map1n[uint64](satToUint8, 16))
   295  		testUint64x8ConvertToUint8(t, archsimd.Uint64x8.SaturateToUint8, map1n[uint64](satToUint8, 16))
   296  		testUint32x4ConvertToUint16(t, archsimd.Uint32x4.SaturateToUint16, map1n[uint32](satToUint16, 8))
   297  		testUint32x8ConvertToUint16(t, archsimd.Uint32x8.SaturateToUint16, map1[uint32](satToUint16))
   298  		testUint32x16ConvertToUint16(t, archsimd.Uint32x16.SaturateToUint16, map1[uint32](satToUint16))
   299  		testUint64x2ConvertToUint16(t, archsimd.Uint64x2.SaturateToUint16, map1n[uint64](satToUint16, 8))
   300  		testUint64x4ConvertToUint16(t, archsimd.Uint64x4.SaturateToUint16, map1n[uint64](satToUint16, 8))
   301  		testUint64x8ConvertToUint16(t, archsimd.Uint64x8.SaturateToUint16, map1[uint64](satToUint16))
   302  		testUint64x2ConvertToUint32(t, archsimd.Uint64x2.SaturateToUint32, map1n[uint64](satToUint32, 4))
   303  		testUint64x4ConvertToUint32(t, archsimd.Uint64x4.SaturateToUint32, map1[uint64](satToUint32))
   304  		testUint64x8ConvertToUint32(t, archsimd.Uint64x8.SaturateToUint32, map1[uint64](satToUint32))
   305  	}
   306  }
   307  
   308  func TestOnesCountAMD64(t *testing.T) {
   309  	if archsimd.X86.AVX2() {
   310  		testInt8x32Unary(t, archsimd.Int8x32.OnesCount, map1[int8](onesCount))
   311  	}
   312  	if archsimd.X86.AVX512BITALG() {
   313  		// 128-bit
   314  		testInt16x8Unary(t, archsimd.Int16x8.OnesCount, map1[int16](onesCount))
   315  		testUint16x8Unary(t, archsimd.Uint16x8.OnesCount, map1[uint16](onesCount))
   316  
   317  		// 256-bit
   318  		testUint8x32Unary(t, archsimd.Uint8x32.OnesCount, map1[uint8](onesCount))
   319  		testInt16x16Unary(t, archsimd.Int16x16.OnesCount, map1[int16](onesCount))
   320  		testUint16x16Unary(t, archsimd.Uint16x16.OnesCount, map1[uint16](onesCount))
   321  
   322  		// 512-bit
   323  		testInt8x64Unary(t, archsimd.Int8x64.OnesCount, map1[int8](onesCount))
   324  		testUint8x64Unary(t, archsimd.Uint8x64.OnesCount, map1[uint8](onesCount))
   325  		testInt16x32Unary(t, archsimd.Int16x32.OnesCount, map1[int16](onesCount))
   326  		testUint16x32Unary(t, archsimd.Uint16x32.OnesCount, map1[uint16](onesCount))
   327  	}
   328  
   329  	if archsimd.X86.AVX512VPOPCNTDQ() {
   330  		// 128-bit
   331  		testInt32x4Unary(t, archsimd.Int32x4.OnesCount, map1[int32](onesCount))
   332  		testUint32x4Unary(t, archsimd.Uint32x4.OnesCount, map1[uint32](onesCount))
   333  		testInt64x2Unary(t, archsimd.Int64x2.OnesCount, map1[int64](onesCount))
   334  		testUint64x2Unary(t, archsimd.Uint64x2.OnesCount, map1[uint64](onesCount))
   335  
   336  		// 256-bit
   337  		testInt32x8Unary(t, archsimd.Int32x8.OnesCount, map1[int32](onesCount))
   338  		testUint32x8Unary(t, archsimd.Uint32x8.OnesCount, map1[uint32](onesCount))
   339  		testInt64x4Unary(t, archsimd.Int64x4.OnesCount, map1[int64](onesCount))
   340  		testUint64x4Unary(t, archsimd.Uint64x4.OnesCount, map1[uint64](onesCount))
   341  
   342  		// 512-bit
   343  		testInt32x16Unary(t, archsimd.Int32x16.OnesCount, map1[int32](onesCount))
   344  		testUint32x16Unary(t, archsimd.Uint32x16.OnesCount, map1[uint32](onesCount))
   345  		testInt64x8Unary(t, archsimd.Int64x8.OnesCount, map1[int64](onesCount))
   346  		testUint64x8Unary(t, archsimd.Uint64x8.OnesCount, map1[uint64](onesCount))
   347  	}
   348  }
   349
View as plain text