[dev.simd] simd: add partial slice load/store for 32/64-bits on AVX2

These all use int-vector-masked loads and stores.
Partial set of tests (for all NxK shapes, thought not all types).

Change-Id: I8f493aaa9228647e08ea5badb06dcfe716d6925d
Reviewed-on: https://go-review.googlesource.com/c/go/+/689336
LUCI-TryBot-Result: Go LUCI <golang-scoped@luci-project-accounts.iam.gserviceaccount.com>
Reviewed-by: Junyang Shao <shaojunyang@google.com>

diff --git a/src/simd/slicepart_amd64.go b/src/simd/slicepart_amd64.go
index 920cdb8ccd9771caadf94e7c6e2743f72edb6bec..00025775be114b695d5c269743477b4cf954bc9b 100644 (file)
--- a/src/simd/slicepart_amd64.go
+++ b/src/simd/slicepart_amd64.go
@@ -407,8 +407,23 @@ func paInt32x4(s []int32) *[4]int32 {
        return (*[4]int32)(unsafe.Pointer(&s[0]))
 }
 
+func paInt32x8(s []int32) *[8]int32 {
+       return (*[8]int32)(unsafe.Pointer(&s[0]))
+}
+
+func paInt64x2(s []int64) *[2]int64 {
+       return (*[2]int64)(unsafe.Pointer(&s[0]))
+}
+
+func paInt64x4(s []int64) *[4]int64 {
+       return (*[4]int64)(unsafe.Pointer(&s[0]))
+}
+
 /* 32 and 64-bit slice-part loads for AVX2 (128 and 256 bit) */
 
+// LoadInt32x4SlicePart loads a Int32x4 from the slice s.
+// If s has fewer than 4 elements, the remaining elements of the vector are filled with zeroes.
+// If s has 4 or more elements, the function is equivalent to LoadInt32x4Slice.
 func LoadInt32x4SlicePart(s []int32) Int32x4 {
        l := len(s)
        if l >= 4 {
@@ -423,6 +438,9 @@ func LoadInt32x4SlicePart(s []int32) Int32x4 {
        return LoadMaskedInt32x4(paInt32x4(s), LoadInt32x4Slice(mask).AsMask32x4())
 }
 
+// StoreSlicePart stores the elements of x into the slice s.
+// It stores as many elements as will fit in s.
+// If s has 4 or more elements, the method is equivalent to x.StoreSlice.
 func (x Int32x4) StoreSlicePart(s []int32) {
        l := len(s)
        if l >= 4 {
@@ -437,22 +455,290 @@ func (x Int32x4) StoreSlicePart(s []int32) {
        x.StoreMasked(paInt32x4(s), LoadInt32x4Slice(mask).AsMask32x4())
 }
 
-// func LoadInt32x8SlicePart(s []int32) Int32x8 {
-// }
+// LoadInt32x8SlicePart loads a Int32x8 from the slice s.
+// If s has fewer than 8 elements, the remaining elements of the vector are filled with zeroes.
+// If s has 8 or more elements, the function is equivalent to LoadInt32x8Slice.
+func LoadInt32x8SlicePart(s []int32) Int32x8 {
+       l := len(s)
+       if l >= 8 {
+               return LoadInt32x8Slice(s)
+       }
+       if l == 0 {
+               var x Int32x8
+               return x
+       }
+       p := int32atP64(&vecMask64[0])
+       mask := unsafe.Slice(p, 32)[16-l:]
+       return LoadMaskedInt32x8(paInt32x8(s), LoadInt32x8Slice(mask).AsMask32x8())
+}
 
-// func LoadInt64x2SlicePart(s []int64) Int64x2 {
-// }
+// LoadInt64x2SlicePart loads a Int64x2 from the slice s.
+// If s has fewer than 2 elements, the remaining elements of the vector are filled with zeroes.
+// If s has 2 or more elements, the function is equivalent to LoadInt64x2Slice.
+func LoadInt64x2SlicePart(s []int64) Int64x2 {
+       l := len(s)
+       if l >= 2 {
+               return LoadInt64x2Slice(s)
+       }
+       if l == 0 {
+               var x Int64x2
+               return x
+       }
 
-// func LoadInt64x4SlicePart(s []int64) Int64x4 {
-// }
+       mask := vecMask64[8-l:]
+       return LoadMaskedInt64x2(paInt64x2(s), LoadInt64x2Slice(mask).AsMask64x2())
+}
 
-// func (x Int32x8) StoreSlicePart(s []int32) {
-// }
+// LoadInt64x4SlicePart loads a Int64x4 from the slice s.
+// If s has fewer than 4 elements, the remaining elements of the vector are filled with zeroes.
+// If s has 4 or more elements, the function is equivalent to LoadInt64x4Slice.
+func LoadInt64x4SlicePart(s []int64) Int64x4 {
+       l := len(s)
+       if l >= 4 {
+               return LoadInt64x4Slice(s)
+       }
+       if l == 0 {
+               var x Int64x4
+               return x
+       }
 
-// func (x Int64x4) StoreSlicePart(s []int64) {
-// }
+       mask := vecMask64[8-l:]
+       return LoadMaskedInt64x4(paInt64x4(s), LoadInt64x4Slice(mask).AsMask64x4())
+}
 
-// func (x Int64x8) StoreSlicePart(s []int64) {
-// }
+// StoreSlicePart stores the elements of x into the slice s.
+// It stores as many elements as will fit in s.
+// If s has 8 or more elements, the method is equivalent to x.StoreSlice.
+func (x Int32x8) StoreSlicePart(s []int32) {
+       l := len(s)
+       if l >= 8 {
+               x.StoreSlice(s)
+               return
+       }
+       if l == 0 {
+               return
+       }
+       p := int32atP64(&vecMask64[0])
+       mask := unsafe.Slice(p, 32)[16-l:]
+       x.StoreMasked(paInt32x8(s), LoadInt32x8Slice(mask).AsMask32x8())
+}
+
+// StoreSlicePart stores the elements of x into the slice s.
+// It stores as many elements as will fit in s.
+// If s has 2 or more elements, the method is equivalent to x.StoreSlice.
+func (x Int64x2) StoreSlicePart(s []int64) {
+       l := len(s)
+       if l >= 2 {
+               x.StoreSlice(s)
+               return
+       }
+       if l == 0 {
+               return
+       }
+       mask := vecMask64[8-l:]
+       x.StoreMasked(paInt64x2(s), LoadInt64x2Slice(mask).AsMask64x2())
+}
+
+// StoreSlicePart stores the elements of x into the slice s.
+// It stores as many elements as will fit in s.
+// If s has 4 or more elements, the method is equivalent to x.StoreSlice.
+func (x Int64x4) StoreSlicePart(s []int64) {
+       l := len(s)
+       if l >= 4 {
+               x.StoreSlice(s)
+               return
+       }
+       if l == 0 {
+               return
+       }
+       mask := vecMask64[8-l:]
+       x.StoreMasked(paInt64x4(s), LoadInt64x4Slice(mask).AsMask64x4())
+}
 
 // Handle float32, float64, uint32, and uint64 with ugly casts.
+
+// LoadUint32x4SlicePart loads a Uint32x4 from the slice s.
+// If s has fewer than 4 elements, the remaining elements of the vector are filled with zeroes.
+// If s has 4 or more elements, the function is equivalent to LoadUint32x4Slice.
+func LoadUint32x4SlicePart(s []uint32) Uint32x4 {
+       if len(s) == 0 {
+               var zero Uint32x4
+               return zero
+       }
+       t := unsafe.Slice((*int32)(unsafe.Pointer(&s[0])), len(s))
+       return LoadInt32x4SlicePart(t).AsUint32x4()
+}
+
+// StoreSlicePart stores the elements of x into the slice s.
+// It stores as many elements as will fit in s.
+// If s has 4 or more elements, the method is equivalent to x.StoreSlice.
+func (x Uint32x4) StoreSlicePart(s []uint32) {
+       if len(s) == 0 {
+               return
+       }
+       t := unsafe.Slice((*int32)(unsafe.Pointer(&s[0])), len(s))
+       x.AsInt32x4().StoreSlicePart(t)
+}
+
+// LoadUint32x8SlicePart loads a Uint32x8 from the slice s.
+// If s has fewer than 8 elements, the remaining elements of the vector are filled with zeroes.
+// If s has 8 or more elements, the function is equivalent to LoadUint32x8Slice.
+func LoadUint32x8SlicePart(s []uint32) Uint32x8 {
+       if len(s) == 0 {
+               var zero Uint32x8
+               return zero
+       }
+       t := unsafe.Slice((*int32)(unsafe.Pointer(&s[0])), len(s))
+       return LoadInt32x8SlicePart(t).AsUint32x8()
+}
+
+// StoreSlicePart stores the elements of x into the slice s.
+// It stores as many elements as will fit in s.
+// If s has 8 or more elements, the method is equivalent to x.StoreSlice.
+func (x Uint32x8) StoreSlicePart(s []uint32) {
+       if len(s) == 0 {
+               return
+       }
+       t := unsafe.Slice((*int32)(unsafe.Pointer(&s[0])), len(s))
+       x.AsInt32x8().StoreSlicePart(t)
+}
+
+// LoadUint64x2SlicePart loads a Uint64x2 from the slice s.
+// If s has fewer than 2 elements, the remaining elements of the vector are filled with zeroes.
+// If s has 2 or more elements, the function is equivalent to LoadUint64x2Slice.
+func LoadUint64x2SlicePart(s []uint64) Uint64x2 {
+       if len(s) == 0 {
+               var zero Uint64x2
+               return zero
+       }
+       t := unsafe.Slice((*int64)(unsafe.Pointer(&s[0])), len(s))
+       return LoadInt64x2SlicePart(t).AsUint64x2()
+}
+
+// StoreSlicePart stores the elements of x into the slice s.
+// It stores as many elements as will fit in s.
+// If s has 2 or more elements, the method is equivalent to x.StoreSlice.
+func (x Uint64x2) StoreSlicePart(s []uint64) {
+       if len(s) == 0 {
+               return
+       }
+       t := unsafe.Slice((*int64)(unsafe.Pointer(&s[0])), len(s))
+       x.AsInt64x2().StoreSlicePart(t)
+}
+
+// LoadUint64x4SlicePart loads a Uint64x4 from the slice s.
+// If s has fewer than 4 elements, the remaining elements of the vector are filled with zeroes.
+// If s has 4 or more elements, the function is equivalent to LoadUint64x4Slice.
+func LoadUint64x4SlicePart(s []uint64) Uint64x4 {
+       if len(s) == 0 {
+               var zero Uint64x4
+               return zero
+       }
+       t := unsafe.Slice((*int64)(unsafe.Pointer(&s[0])), len(s))
+       return LoadInt64x4SlicePart(t).AsUint64x4()
+}
+
+// StoreSlicePart stores the elements of x into the slice s.
+// It stores as many elements as will fit in s.
+// If s has 4 or more elements, the method is equivalent to x.StoreSlice.
+func (x Uint64x4) StoreSlicePart(s []uint64) {
+       if len(s) == 0 {
+               return
+       }
+       t := unsafe.Slice((*int64)(unsafe.Pointer(&s[0])), len(s))
+       x.AsInt64x4().StoreSlicePart(t)
+}
+
+// Float32xK and Float64xK
+
+// LoadFloat32x4SlicePart loads a Float32x4 from the slice s.
+// If s has fewer than 4 elements, the remaining elements of the vector are filled with zeroes.
+// If s has 4 or more elements, the function is equivalent to LoadFloat32x4Slice.
+func LoadFloat32x4SlicePart(s []float32) Float32x4 {
+       if len(s) == 0 {
+               var zero Float32x4
+               return zero
+       }
+       t := unsafe.Slice((*int32)(unsafe.Pointer(&s[0])), len(s))
+       return LoadInt32x4SlicePart(t).AsFloat32x4()
+}
+
+// StoreSlicePart stores the elements of x into the slice s.
+// It stores as many elements as will fit in s.
+// If s has 4 or more elements, the method is equivalent to x.StoreSlice.
+func (x Float32x4) StoreSlicePart(s []float32) {
+       if len(s) == 0 {
+               return
+       }
+       t := unsafe.Slice((*int32)(unsafe.Pointer(&s[0])), len(s))
+       x.AsInt32x4().StoreSlicePart(t)
+}
+
+// LoadFloat32x8SlicePart loads a Float32x8 from the slice s.
+// If s has fewer than 8 elements, the remaining elements of the vector are filled with zeroes.
+// If s has 8 or more elements, the function is equivalent to LoadFloat32x8Slice.
+func LoadFloat32x8SlicePart(s []float32) Float32x8 {
+       if len(s) == 0 {
+               var zero Float32x8
+               return zero
+       }
+       t := unsafe.Slice((*int32)(unsafe.Pointer(&s[0])), len(s))
+       return LoadInt32x8SlicePart(t).AsFloat32x8()
+}
+
+// StoreSlicePart stores the elements of x into the slice s.
+// It stores as many elements as will fit in s.
+// If s has 8 or more elements, the method is equivalent to x.StoreSlice.
+func (x Float32x8) StoreSlicePart(s []float32) {
+       if len(s) == 0 {
+               return
+       }
+       t := unsafe.Slice((*int32)(unsafe.Pointer(&s[0])), len(s))
+       x.AsInt32x8().StoreSlicePart(t)
+}
+
+// LoadFloat64x2SlicePart loads a Float64x2 from the slice s.
+// If s has fewer than 2 elements, the remaining elements of the vector are filled with zeroes.
+// If s has 2 or more elements, the function is equivalent to LoadFloat64x2Slice.
+func LoadFloat64x2SlicePart(s []float64) Float64x2 {
+       if len(s) == 0 {
+               var zero Float64x2
+               return zero
+       }
+       t := unsafe.Slice((*int64)(unsafe.Pointer(&s[0])), len(s))
+       return LoadInt64x2SlicePart(t).AsFloat64x2()
+}
+
+// StoreSlicePart stores the elements of x into the slice s.
+// It stores as many elements as will fit in s.
+// If s has 2 or more elements, the method is equivalent to x.StoreSlice.
+func (x Float64x2) StoreSlicePart(s []float64) {
+       if len(s) == 0 {
+               return
+       }
+       t := unsafe.Slice((*int64)(unsafe.Pointer(&s[0])), len(s))
+       x.AsInt64x2().StoreSlicePart(t)
+}
+
+// LoadFloat64x4SlicePart loads a Float64x4 from the slice s.
+// If s has fewer than 4 elements, the remaining elements of the vector are filled with zeroes.
+// If s has 4 or more elements, the function is equivalent to LoadFloat64x4Slice.
+func LoadFloat64x4SlicePart(s []float64) Float64x4 {
+       if len(s) == 0 {
+               var zero Float64x4
+               return zero
+       }
+       t := unsafe.Slice((*int64)(unsafe.Pointer(&s[0])), len(s))
+       return LoadInt64x4SlicePart(t).AsFloat64x4()
+}
+
+// StoreSlicePart stores the elements of x into the slice s.
+// It stores as many elements as will fit in s.
+// If s has 4 or more elements, the method is equivalent to x.StoreSlice.
+func (x Float64x4) StoreSlicePart(s []float64) {
+       if len(s) == 0 {
+               return
+       }
+       t := unsafe.Slice((*int64)(unsafe.Pointer(&s[0])), len(s))
+       x.AsInt64x4().StoreSlicePart(t)
+}

diff --git a/src/simd/slicepart_test.go b/src/simd/slicepart_test.go
index cd282be7b1c2afb23fd5819496b44e18e12bce3d..cfdb7581d9e9e7530e83d758e643b13ce3b3051c 100644 (file)
--- a/src/simd/slicepart_test.go
+++ b/src/simd/slicepart_test.go
@@ -179,6 +179,7 @@ func TestSlicesPartStoreUint8x32(t *testing.T) {
 }
 
 func TestSlicePartInt32(t *testing.T) {
+       // 32x4
        L := 4
        c := []int32{1, 2, 3, 4, 5, -1, -1, -1, -1}
        a := c[:L+1]
@@ -217,3 +218,126 @@ func TestSlicePartInt32(t *testing.T) {
                }
        }
 }
+
+func TestSlicePartUint64(t *testing.T) {
+       // 64x4
+       L := 4
+       c := []uint64{1, 2, 3, 4, 5, 86, 86, 86, 86}
+       a := c[:L+1]
+       for i := range a {
+               // Test the load first
+               // e is a partial slice.
+               e := a[i:]
+               v := simd.LoadUint64x4SlicePart(e)
+               // d contains what a ought to contain
+               d := make([]uint64, L)
+               for j := 0; j < len(e) && j < len(d); j++ {
+                       d[j] = e[j]
+               }
+
+               b := make([]uint64, L)
+               v.StoreSlice(b)
+               // test the load
+               checkSlices(t, d, b)
+
+               // Test the store
+               f := make([]uint64, L+1)
+               for i := range f {
+                       f[i] = 99
+               }
+
+               v.StoreSlicePart(f[:len(e)])
+               if len(e) < len(b) {
+                       checkSlices(t, f, b[:len(e)])
+               } else {
+                       checkSlices(t, f, b)
+               }
+               for i := len(e); i < len(f); i++ {
+                       if f[i] != 99 {
+                               t.Errorf("StoreSlicePart altered f[%d], expected 99, saw %d", i, f[i])
+                       }
+               }
+       }
+}
+
+func TestSlicePartFloat64(t *testing.T) {
+       // 64x2
+       L := 2
+       c := []float64{1, 2, 3, 86, 86, 86, 86}
+       a := c[:L+1]
+       for i := range a {
+               // Test the load first
+               // e is a partial slice.
+               e := a[i:]
+               v := simd.LoadFloat64x2SlicePart(e)
+               // d contains what a ought to contain
+               d := make([]float64, L)
+               for j := 0; j < len(e) && j < len(d); j++ {
+                       d[j] = e[j]
+               }
+
+               b := make([]float64, L)
+               v.StoreSlice(b)
+               // test the load
+               checkSlices(t, d, b)
+
+               // Test the store
+               f := make([]float64, L+1)
+               for i := range f {
+                       f[i] = 99
+               }
+
+               v.StoreSlicePart(f[:len(e)])
+               if len(e) < len(b) {
+                       checkSlices(t, f, b[:len(e)])
+               } else {
+                       checkSlices(t, f, b)
+               }
+               for i := len(e); i < len(f); i++ {
+                       if f[i] != 99 {
+                               t.Errorf("StoreSlicePart altered f[%d], expected 99, saw %v", i, f[i])
+                       }
+               }
+       }
+}
+
+func TestSlicePartFloat32(t *testing.T) {
+       // 32x8
+       L := 8
+       c := []float32{1, 2, 3, 4, 5, 6, 7, 8, 86, 86, 86, 86}
+       a := c[:L+1]
+       for i := range a {
+               // Test the load first
+               // e is a partial slice.
+               e := a[i:]
+               v := simd.LoadFloat32x8SlicePart(e)
+               // d contains what a ought to contain
+               d := make([]float32, L)
+               for j := 0; j < len(e) && j < len(d); j++ {
+                       d[j] = e[j]
+               }
+
+               b := make([]float32, L)
+               v.StoreSlice(b)
+               // test the load
+               checkSlices(t, d, b)
+
+               // Test the store
+               f := make([]float32, L+1)
+               for i := range f {
+                       f[i] = 99
+               }
+
+               v.StoreSlicePart(f[:len(e)])
+               if len(e) < len(b) {
+                       checkSlices(t, f, b[:len(e)])
+               } else {
+                       checkSlices(t, f, b)
+               }
+               for i := len(e); i < len(f); i++ {
+                       if f[i] != 99 {
+                               t.Errorf("StoreSlicePart altered f[%d], expected 99, saw %v", i, f[i])
+                       }
+               }
+       }
+}