Skip to content

perf(exp/simd): optimize Sum funcs with horizontal reduction #849

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Draft
d-enk wants to merge 1 commit into
base: master
Choose a base branch
from
Draft

perf(exp/simd): optimize Sum funcs with horizontal reduction #849

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
1 commit
Select commit
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
35 changes: 35 additions & 0 deletions exp/simd/helpers.go
View file
Open in desktop
Original file line number Diff line number Diff line change
@@ -0,0 +1,35 @@
//go:build go1.26 && goexperiment.simd && amd64

package simd

type number interface {
~float32 | ~float64 |
~int | ~int8 | ~int16 | ~int32 | ~int64 |
~uint | ~uint8 | ~uint16 | ~uint32 | ~uint64 | ~uintptr

@samber samber Mar 31, 2026

Copy link
Copy Markdown
Owner

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

uintptr ? can you explain why ?

@d-enk d-enk Mar 31, 2026

Copy link
Copy Markdown
Contributor Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Wow, just mistake)

}

// sum16 sums a vector of 16 elements.
func sum16[T number](vec *[16]T) T {
return 0 +
vec[0] + vec[1] + vec[2] + vec[3] +
vec[4] + vec[5] + vec[6] + vec[7] +
vec[8] + vec[9] + vec[10] + vec[11] +
vec[12] + vec[13] + vec[14] + vec[15]
}

// sum8 sums a vector of 8 elements.
func sum8[T number](vec *[8]T) T {
return 0 +
vec[0] + vec[1] + vec[2] + vec[3] +
vec[4] + vec[5] + vec[6] + vec[7]
}

// sum4 sums a vector of 4 elements.
func sum4[T number](vec *[4]T) T {
return vec[0] + vec[1] + vec[2] + vec[3]
}

// sum2 sums a vector of 2 elements.
func sum2[T number](vec *[2]T) T {
return vec[0] + vec[1]
}
210 changes: 70 additions & 140 deletions exp/simd/math_avx.go
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -23,26 +23,19 @@ func SumInt8x16[T ~int8](collection []T) T {
const lanes = simdLanes16

base := unsafeSliceInt8(collection, length)
var acc archsimd.Int8x16

i := uint(0)
for ; i+lanes <= length; i += lanes {
v := archsimd.LoadInt8x16Slice(base[i : i+lanes])
acc = acc.Add(v)
i := length % lanes

acc := archsimd.LoadInt8x16SlicePart(base[:i])

for ; i < length; i += lanes {
acc = acc.Add(archsimd.LoadInt8x16Slice(base[i : i+lanes]))
}

var buf [lanes]int8
acc.Store(&buf)
var sum T
for k := uint(0); k < lanes; k++ {
sum += T(buf[k])
}

for ; i < length; i++ {
sum += collection[i]
}

return sum
return T(sum16(&buf))
}

// SumInt16x8 sums a slice of int16 using AVX SIMD (Int16x8, 8 lanes).
Expand All @@ -57,26 +50,19 @@ func SumInt16x8[T ~int16](collection []T) T {
const lanes = simdLanes8

base := unsafeSliceInt16(collection, length)
var acc archsimd.Int16x8

i := uint(0)
for ; i+lanes <= length; i += lanes {
v := archsimd.LoadInt16x8Slice(base[i : i+lanes])
acc = acc.Add(v)
i := length % lanes

acc := archsimd.LoadInt16x8SlicePart(base[:i])

for ; i < length; i += lanes {
acc = acc.Add(archsimd.LoadInt16x8Slice(base[i : i+lanes]))
}

var buf [lanes]int16
acc.Store(&buf)
var sum T
for k := uint(0); k < lanes; k++ {
sum += T(buf[k])
}

for ; i < length; i++ {
sum += collection[i]
}

return sum
return T(sum8(&buf))
}

// SumInt32x4 sums a slice of int32 using AVX SIMD (Int32x4, 4 lanes).
Expand All @@ -91,26 +77,19 @@ func SumInt32x4[T ~int32](collection []T) T {
const lanes = simdLanes4

base := unsafeSliceInt32(collection, length)
var acc archsimd.Int32x4

i := uint(0)
for ; i+lanes <= length; i += lanes {
v := archsimd.LoadInt32x4Slice(base[i : i+lanes])
acc = acc.Add(v)
i := length % lanes

acc := archsimd.LoadInt32x4SlicePart(base[:i])

for ; i < length; i += lanes {
acc = acc.Add(archsimd.LoadInt32x4Slice(base[i : i+lanes]))
}

var buf [lanes]int32
acc.Store(&buf)
var sum T
for k := uint(0); k < lanes; k++ {
sum += T(buf[k])
}

for ; i < length; i++ {
sum += collection[i]
}

return sum
return T(sum4(&buf))
}

// SumInt64x2 sums a slice of int64 using AVX SIMD (Int64x2, 2 lanes).
Expand All @@ -125,26 +104,19 @@ func SumInt64x2[T ~int64](collection []T) T {
const lanes = simdLanes2

base := unsafeSliceInt64(collection, length)
var acc archsimd.Int64x2

i := uint(0)
for ; i+lanes <= length; i += lanes {
v := archsimd.LoadInt64x2Slice(base[i : i+lanes])
acc = acc.Add(v)
i := length % lanes

acc := archsimd.LoadInt64x2SlicePart(base[:i])

for ; i < length; i += lanes {
acc = acc.Add(archsimd.LoadInt64x2Slice(base[i : i+lanes]))
}

var buf [lanes]int64
acc.Store(&buf)
var sum T
for k := uint(0); k < lanes; k++ {
sum += T(buf[k])
}

for ; i < length; i++ {
sum += collection[i]
}

return sum
return T(sum2(&buf))
}

// SumUint8x16 sums a slice of uint8 using AVX SIMD (Uint8x16, 16 lanes).
Expand All @@ -159,26 +131,19 @@ func SumUint8x16[T ~uint8](collection []T) T {
const lanes = simdLanes16

base := unsafeSliceUint8(collection, length)
var acc archsimd.Uint8x16

i := uint(0)
for ; i+lanes <= length; i += lanes {
v := archsimd.LoadUint8x16Slice(base[i : i+lanes])
acc = acc.Add(v)
i := length % lanes

acc := archsimd.LoadUint8x16SlicePart(base[:i])

for ; i < length; i += lanes {
acc = acc.Add(archsimd.LoadUint8x16Slice(base[i : i+lanes]))
}

var buf [lanes]uint8
acc.Store(&buf)
var sum T
for k := uint(0); k < lanes; k++ {
sum += T(buf[k])
}

for ; i < length; i++ {
sum += collection[i]
}

return sum
return T(sum16(&buf))
}

// SumUint16x8 sums a slice of uint16 using AVX SIMD (Uint16x8, 8 lanes).
Expand All @@ -193,26 +158,19 @@ func SumUint16x8[T ~uint16](collection []T) T {
const lanes = simdLanes8

base := unsafeSliceUint16(collection, length)
var acc archsimd.Uint16x8

i := uint(0)
for ; i+lanes <= length; i += lanes {
v := archsimd.LoadUint16x8Slice(base[i : i+lanes])
acc = acc.Add(v)
i := length % lanes

acc := archsimd.LoadUint16x8SlicePart(base[:i])

for ; i < length; i += lanes {
acc = acc.Add(archsimd.LoadUint16x8Slice(base[i : i+lanes]))
}

var buf [lanes]uint16
acc.Store(&buf)
var sum T
for k := uint(0); k < lanes; k++ {
sum += T(buf[k])
}

for ; i < length; i++ {
sum += collection[i]
}

return sum
return T(sum8(&buf))
}

// SumUint32x4 sums a slice of uint32 using AVX SIMD (Uint32x4, 4 lanes).
Expand All @@ -227,26 +185,19 @@ func SumUint32x4[T ~uint32](collection []T) T {
const lanes = simdLanes4

base := unsafeSliceUint32(collection, length)
var acc archsimd.Uint32x4

i := uint(0)
for ; i+lanes <= length; i += lanes {
v := archsimd.LoadUint32x4Slice(base[i : i+lanes])
acc = acc.Add(v)
i := length % lanes

acc := archsimd.LoadUint32x4SlicePart(base[:i])

for ; i < length; i += lanes {
acc = acc.Add(archsimd.LoadUint32x4Slice(base[i : i+lanes]))
}

var buf [lanes]uint32
acc.Store(&buf)
var sum T
for k := uint(0); k < lanes; k++ {
sum += T(buf[k])
}

for ; i < length; i++ {
sum += collection[i]
}

return sum
return T(sum4(&buf))
}

// SumUint64x2 sums a slice of uint64 using AVX SIMD (Uint64x2, 2 lanes).
Expand All @@ -261,26 +212,19 @@ func SumUint64x2[T ~uint64](collection []T) T {
const lanes = simdLanes2

base := unsafeSliceUint64(collection, length)
var acc archsimd.Uint64x2

i := uint(0)
for ; i+lanes <= length; i += lanes {
v := archsimd.LoadUint64x2Slice(base[i : i+lanes])
acc = acc.Add(v)
i := length % lanes

acc := archsimd.LoadUint64x2SlicePart(base[:i])

for ; i < length; i += lanes {
acc = acc.Add(archsimd.LoadUint64x2Slice(base[i : i+lanes]))
}

var buf [lanes]uint64
acc.Store(&buf)
var sum T
for k := uint(0); k < lanes; k++ {
sum += T(buf[k])
}

for ; i < length; i++ {
sum += collection[i]
}

return sum
return T(sum2(&buf))
}

// SumFloat32x4 sums a slice of float32 using AVX SIMD (Float32x4, 4 lanes).
Expand All @@ -294,26 +238,19 @@ func SumFloat32x4[T ~float32](collection []T) T {
const lanes = simdLanes4

base := unsafeSliceFloat32(collection, length)
var acc archsimd.Float32x4

i := uint(0)
for ; i+lanes <= length; i += lanes {
v := archsimd.LoadFloat32x4Slice(base[i : i+lanes])
acc = acc.Add(v)
i := length % lanes

acc := archsimd.LoadFloat32x4SlicePart(base[:i])

for ; i < length; i += lanes {
acc = acc.Add(archsimd.LoadFloat32x4Slice(base[i : i+lanes]))
}

var buf [lanes]float32
acc.Store(&buf)
var sum T
for k := uint(0); k < lanes; k++ {
sum += T(buf[k])
}

for ; i < length; i++ {
sum += collection[i]
}

return sum
return T(sum4(&buf))
}

// SumFloat64x2 sums a slice of float64 using AVX SIMD (Float64x2, 2 lanes).
Expand All @@ -327,26 +264,19 @@ func SumFloat64x2[T ~float64](collection []T) T {
const lanes = simdLanes2

base := unsafeSliceFloat64(collection, length)
var acc archsimd.Float64x2

i := uint(0)
for ; i+lanes <= length; i += lanes {
v := archsimd.LoadFloat64x2Slice(base[i : i+lanes])
acc = acc.Add(v)
i := length % lanes

acc := archsimd.LoadFloat64x2SlicePart(base[:i])

for ; i < length; i += lanes {
acc = acc.Add(archsimd.LoadFloat64x2Slice(base[i : i+lanes]))
}

var buf [lanes]float64
acc.Store(&buf)
var sum T
for k := uint(0); k < lanes; k++ {
sum += T(buf[k])
}

for ; i < length; i++ {
sum += collection[i]
}

return sum
return T(sum2(&buf))
}

// MeanInt8x16 calculates the mean of a slice of int8 using AVX SIMD
Expand Down
Loading
Loading