minmax_amd64_test.go

     1  // Copyright 2026 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  	"simd/archsimd"
    11  	"testing"
    12  )
    13  
    14  func TestFloatMinMax32AMD64(t *testing.T) {
    15  	for _, tc := range []struct {
    16  		name  string
    17  		lanes int
    18  		run   func([]float32, []float32, [4][]float32, bool, bool)
    19  	}{
    20  		{"Float32x8", 8, floatMinMax32x8},
    21  		{"Float32x16", 16, floatMinMax32x16},
    22  	} {
    23  		t.Run(tc.name, func(t *testing.T) {
    24  			if tc.lanes == 16 && !archsimd.X86.AVX512() {
    25  				t.Skip("requires AVX512")
    26  			}
    27  			testFloatMinMax32(t, tc.lanes, tc.run)
    28  		})
    29  	}
    30  }
    31  
    32  func TestFloatMinMax64AMD64(t *testing.T) {
    33  	for _, tc := range []struct {
    34  		name  string
    35  		lanes int
    36  		run   func([]float64, []float64, [4][]float64, bool, bool)
    37  	}{
    38  		{"Float64x4", 4, floatMinMax64x4},
    39  		{"Float64x8", 8, floatMinMax64x8},
    40  	} {
    41  		t.Run(tc.name, func(t *testing.T) {
    42  			if tc.lanes == 8 && !archsimd.X86.AVX512() {
    43  				t.Skip("requires AVX512")
    44  			}
    45  			testFloatMinMax64(t, tc.lanes, tc.run)
    46  		})
    47  	}
    48  }
    49  
    50  func floatMinMaxPlatformConfig() floatMinMaxConfig {
    51  	return floatMinMaxConfig{allowMasked: archsimd.X86.AVX512()}
    52  }
    53  
    54  func floatMinMaxMask32x4() archsimd.Mask32x4 {
    55  	return archsimd.Mask32x4FromBits(0x5)
    56  }
    57  
    58  func floatMinMaxMask64x2() archsimd.Mask64x2 {
    59  	return archsimd.Mask64x2FromBits(0x1)
    60  }
    61  
    62  //go:noinline
    63  func floatMinMax32x8(a, b []float32, out [4][]float32, reverse, masked bool) {
    64  	x, y := archsimd.LoadFloat32x8(a), archsimd.LoadFloat32x8(b)
    65  	var minXY, minYX, maxXY, maxYX archsimd.Float32x8
    66  	if masked {
    67  		mask := archsimd.Mask32x8FromBits(0x55)
    68  		if reverse {
    69  			minYX, minXY = y.Min(x).Masked(mask), x.Min(y).Masked(mask)
    70  			maxYX, maxXY = y.Max(x).Masked(mask), x.Max(y).Masked(mask)
    71  		} else {
    72  			minXY, minYX = x.Min(y).Masked(mask), y.Min(x).Masked(mask)
    73  			maxXY, maxYX = x.Max(y).Masked(mask), y.Max(x).Masked(mask)
    74  		}
    75  	} else if reverse {
    76  		minYX, minXY = y.Min(x), x.Min(y)
    77  		maxYX, maxXY = y.Max(x), x.Max(y)
    78  	} else {
    79  		minXY, minYX = x.Min(y), y.Min(x)
    80  		maxXY, maxYX = x.Max(y), y.Max(x)
    81  	}
    82  	minXY.Store(out[0])
    83  	minYX.Store(out[1])
    84  	maxXY.Store(out[2])
    85  	maxYX.Store(out[3])
    86  }
    87  
    88  //go:noinline
    89  func floatMinMax32x16(a, b []float32, out [4][]float32, reverse, masked bool) {
    90  	x, y := archsimd.LoadFloat32x16(a), archsimd.LoadFloat32x16(b)
    91  	var minXY, minYX, maxXY, maxYX archsimd.Float32x16
    92  	if masked {
    93  		mask := archsimd.Mask32x16FromBits(0x5555)
    94  		if reverse {
    95  			minYX, minXY = y.Min(x).Masked(mask), x.Min(y).Masked(mask)
    96  			maxYX, maxXY = y.Max(x).Masked(mask), x.Max(y).Masked(mask)
    97  		} else {
    98  			minXY, minYX = x.Min(y).Masked(mask), y.Min(x).Masked(mask)
    99  			maxXY, maxYX = x.Max(y).Masked(mask), y.Max(x).Masked(mask)
   100  		}
   101  	} else if reverse {
   102  		minYX, minXY = y.Min(x), x.Min(y)
   103  		maxYX, maxXY = y.Max(x), x.Max(y)
   104  	} else {
   105  		minXY, minYX = x.Min(y), y.Min(x)
   106  		maxXY, maxYX = x.Max(y), y.Max(x)
   107  	}
   108  	minXY.Store(out[0])
   109  	minYX.Store(out[1])
   110  	maxXY.Store(out[2])
   111  	maxYX.Store(out[3])
   112  }
   113  
   114  //go:noinline
   115  func floatMinMax64x4(a, b []float64, out [4][]float64, reverse, masked bool) {
   116  	x, y := archsimd.LoadFloat64x4(a), archsimd.LoadFloat64x4(b)
   117  	var minXY, minYX, maxXY, maxYX archsimd.Float64x4
   118  	if masked {
   119  		mask := archsimd.Mask64x4FromBits(0x5)
   120  		if reverse {
   121  			minYX, minXY = y.Min(x).Masked(mask), x.Min(y).Masked(mask)
   122  			maxYX, maxXY = y.Max(x).Masked(mask), x.Max(y).Masked(mask)
   123  		} else {
   124  			minXY, minYX = x.Min(y).Masked(mask), y.Min(x).Masked(mask)
   125  			maxXY, maxYX = x.Max(y).Masked(mask), y.Max(x).Masked(mask)
   126  		}
   127  	} else if reverse {
   128  		minYX, minXY = y.Min(x), x.Min(y)
   129  		maxYX, maxXY = y.Max(x), x.Max(y)
   130  	} else {
   131  		minXY, minYX = x.Min(y), y.Min(x)
   132  		maxXY, maxYX = x.Max(y), y.Max(x)
   133  	}
   134  	minXY.Store(out[0])
   135  	minYX.Store(out[1])
   136  	maxXY.Store(out[2])
   137  	maxYX.Store(out[3])
   138  }
   139  
   140  //go:noinline
   141  func floatMinMax64x8(a, b []float64, out [4][]float64, reverse, masked bool) {
   142  	x, y := archsimd.LoadFloat64x8(a), archsimd.LoadFloat64x8(b)
   143  	var minXY, minYX, maxXY, maxYX archsimd.Float64x8
   144  	if masked {
   145  		mask := archsimd.Mask64x8FromBits(0x55)
   146  		if reverse {
   147  			minYX, minXY = y.Min(x).Masked(mask), x.Min(y).Masked(mask)
   148  			maxYX, maxXY = y.Max(x).Masked(mask), x.Max(y).Masked(mask)
   149  		} else {
   150  			minXY, minYX = x.Min(y).Masked(mask), y.Min(x).Masked(mask)
   151  			maxXY, maxYX = x.Max(y).Masked(mask), y.Max(x).Masked(mask)
   152  		}
   153  	} else if reverse {
   154  		minYX, minXY = y.Min(x), x.Min(y)
   155  		maxYX, maxXY = y.Max(x), x.Max(y)
   156  	} else {
   157  		minXY, minYX = x.Min(y), y.Min(x)
   158  		maxXY, maxYX = x.Max(y), y.Max(x)
   159  	}
   160  	minXY.Store(out[0])
   161  	minYX.Store(out[1])
   162  	maxXY.Store(out[2])
   163  	maxYX.Store(out[3])
   164  }
   165
View as plain text