Simple experimentation and benchmarking of Go 1.26's SIMD support.

README.md

Go 1.26 SIMD Benchmarks

goos: linux
goarch: amd64
pkg: test
cpu: AMD Ryzen 9 3900X 12-Core Processor            
       │ nosimd.txt  │              simd.txt               │             tensor.txt              │
       │   sec/op    │   sec/op     vs base                │   sec/op     vs base                │
Add-24   316.2n ± 1%   434.1n ± 1%  +37.32% (p=0.000 n=10)   184.9n ± 1%  -41.52% (p=0.000 n=10)

       │ nosimd.txt │            simd.txt            │           tensor.txt           │
       │    B/op    │    B/op     vs base            │    B/op     vs base            │
Add-24   0.000 ± 0%   0.000 ± 0%  ~ (p=1.000 n=10) ¹   0.000 ± 0%  ~ (p=1.000 n=10) ¹
¹ all samples are equal

       │ nosimd.txt │            simd.txt            │           tensor.txt           │
       │ allocs/op  │ allocs/op   vs base            │ allocs/op   vs base            │
Add-24   0.000 ± 0%   0.000 ± 0%  ~ (p=1.000 n=10) ¹   0.000 ± 0%  ~ (p=1.000 n=10) ¹
¹ all samples are equal

go.mod

module test

go 1.26.0

nosimd.txt

goos: linux
goarch: amd64
pkg: test
cpu: AMD Ryzen 9 3900X 12-Core Processor            
BenchmarkAdd-24    	 3779433	       315.9 ns/op	       0 B/op	       0 allocs/op
BenchmarkAdd-24    	 3815323	       315.3 ns/op	       0 B/op	       0 allocs/op
BenchmarkAdd-24    	 3771746	       319.1 ns/op	       0 B/op	       0 allocs/op
BenchmarkAdd-24    	 3776943	       316.6 ns/op	       0 B/op	       0 allocs/op
BenchmarkAdd-24    	 3822039	       315.8 ns/op	       0 B/op	       0 allocs/op
BenchmarkAdd-24    	 3776506	       317.8 ns/op	       0 B/op	       0 allocs/op
BenchmarkAdd-24    	 3823285	       314.4 ns/op	       0 B/op	       0 allocs/op
BenchmarkAdd-24    	 3817754	       315.3 ns/op	       0 B/op	       0 allocs/op
BenchmarkAdd-24    	 3766824	       316.4 ns/op	       0 B/op	       0 allocs/op
BenchmarkAdd-24    	 3765676	       318.0 ns/op	       0 B/op	       0 allocs/op
PASS
ok  	test	12.005s

simd.txt

goos: linux
goarch: amd64
pkg: test
cpu: AMD Ryzen 9 3900X 12-Core Processor            
BenchmarkAdd-24    	 2749945	       435.4 ns/op	       0 B/op	       0 allocs/op
BenchmarkAdd-24    	 2748529	       438.4 ns/op	       0 B/op	       0 allocs/op
BenchmarkAdd-24    	 2774643	       432.3 ns/op	       0 B/op	       0 allocs/op
BenchmarkAdd-24    	 2756433	       433.8 ns/op	       0 B/op	       0 allocs/op
BenchmarkAdd-24    	 2765024	       432.6 ns/op	       0 B/op	       0 allocs/op
BenchmarkAdd-24    	 2747886	       436.3 ns/op	       0 B/op	       0 allocs/op
BenchmarkAdd-24    	 2764462	       433.3 ns/op	       0 B/op	       0 allocs/op
BenchmarkAdd-24    	 2778027	       434.5 ns/op	       0 B/op	       0 allocs/op
BenchmarkAdd-24    	 2758598	       436.8 ns/op	       0 B/op	       0 allocs/op
BenchmarkAdd-24    	 2771936	       433.3 ns/op	       0 B/op	       0 allocs/op
PASS
ok  	test	12.011s

tensor.txt

goos: linux
goarch: amd64
pkg: test
cpu: AMD Ryzen 9 3900X 12-Core Processor            
BenchmarkAdd-24    	 6450811	       185.0 ns/op	       0 B/op	       0 allocs/op
BenchmarkAdd-24    	 6462780	       185.2 ns/op	       0 B/op	       0 allocs/op
BenchmarkAdd-24    	 6475417	       185.8 ns/op	       0 B/op	       0 allocs/op
BenchmarkAdd-24    	 6490584	       184.3 ns/op	       0 B/op	       0 allocs/op
BenchmarkAdd-24    	 6493693	       185.0 ns/op	       0 B/op	       0 allocs/op
BenchmarkAdd-24    	 6478975	       185.1 ns/op	       0 B/op	       0 allocs/op
BenchmarkAdd-24    	 6501334	       184.8 ns/op	       0 B/op	       0 allocs/op
BenchmarkAdd-24    	 6528841	       183.9 ns/op	       0 B/op	       0 allocs/op
BenchmarkAdd-24    	 6509636	       183.9 ns/op	       0 B/op	       0 allocs/op
BenchmarkAdd-24    	 6541306	       183.8 ns/op	       0 B/op	       0 allocs/op
PASS
ok  	test	12.000s

test_test.go

package test_test

import (
	"math/rand/v2"
	"simd/archsimd"
	"testing"
)

//func BenchmarkAdd(b *testing.B) {
//	v1 := generateRandom()
//	v2 := generateRandom()
//
//	b.ResetTimer()
//	for b.Loop() {
//		add(v1, v2)
//	}
//}

//func BenchmarkAdd(b *testing.B) {
//	v1 := generateRandom()
//	v2 := generateRandom()
//
//	b.ResetTimer()
//	for b.Loop() {
//		addSIMD(v1, v2)
//	}
//}

func BenchmarkAdd(b *testing.B) {
	v1 := toTensor(generateRandom())
	v2 := toTensor(generateRandom())

	b.ResetTimer()
	for b.Loop() {
		addTensor(v1, v2)
	}
}

func generateRandom() []uint64 {
	v := make([]uint64, 1000)
	for i := range v {
		v[i] = rand.Uint64()
	}
	return v
}

func toTensor(v []uint64) []archsimd.Uint64x4 {
	s := make([]archsimd.Uint64x4, 0, (len(v)+4-1)/4)
	for i := 0; i < len(v); i += 4 {
		s = append(s, archsimd.LoadUint64x4SlicePart(v[i:]))
	}
	return s
}

func addTensor(v1, v2 []archsimd.Uint64x4) []archsimd.Uint64x4 {
	if len(v1) != len(v2) {
		panic("len(v1) != len(v2)")
	}

	for i, s1 := range v1 {
		v1[i] = s1.Add(v2[i])
	}

	return v1
}

func addSIMD(v1, v2 []uint64) []uint64 {
	if len(v1) != len(v2) {
		panic("len(v1) != len(v2)")
	}

	for i := 0; i < len(v1); i += 4 {
		s1 := archsimd.LoadUint64x4SlicePart(v1[i:])
		s2 := archsimd.LoadUint64x4SlicePart(v2[i:])
		r := s1.Add(s2)
		r.StoreSlicePart(v1[i:])
	}

	return v1
}

func add(v1, v2 []uint64) []uint64 {
	if len(v1) != len(v2) {
		panic("len(v1) != len(v2)")
	}

	for i := range v1 {
		v1[i] += v2[i]
	}

	return v1
}