ARM64 benchmark

bench.log

│BenchmarkAmove-4        20000000               102 ns/op        1881.62 MB/s
│BenchmarkBmove-4        50000000                34.5 ns/op      5563.29 MB/s

simd.go

package simd

//go:onescape
func amove(dst, src []byte, count uint64)

//go:onescape
func bmove(dst, src []byte, count uint64)

simd_arm64.s

#include "textflag.h"

TEXT ·amove(SB),NOSPLIT,$56
	 MOVD    dst+0(FP), R0
	 MOVD    src+24(FP), R1
	 MOVD    count+48(FP), R2
	PRFM (R1), PLDL1KEEP
loop:
	 VLD1.P 64(R1), [V0.B16, V1.B16, V2.B16, V3.B16]
	 VLD1.P 64(R1), [V4.B16, V5.B16, V6.B16, V7.B16]
	 VLD1.P 64(R1), [V8.B16, V9.B16, V10.B16, V11.B16]
	 VLD1.P 64(R1), [V12.B16, V13.B16, V14.B16, V15.B16]
	 VST1.P [V0.B16, V1.B16, V2.B16, V3.B16], 64(R0)
	 VST1.P [V4.B16, V5.B16, V6.B16, V7.B16], 64(R0)
	 VST1.P [V8.B16, V9.B16, V10.B16, V11.B16], 64(R0)
	 VST1.P [V12.B16, V13.B16, V14.B16, V15.B16], 64(R0)
	 SUBS $256, R2, R2
	 BGT loop
	 RET

TEXT ·bmove(SB),NOSPLIT,$56
	 MOVD    dst+0(FP), R0
	 MOVD    src+24(FP), R1
	 MOVD    count+48(FP), R2
	PRFM (R1), PLDL1KEEP
bloop:
	LDP (R1), (R4, R5)
	LDP 16(R1), (R6, R7)
	LDP 32(R1), (R8, R9)
	LDP 48(R1), (R10, R11)
	LDP 64(R1), (R12, R13)
	LDP 80(R1), (R14, R15)
	LDP 96(R1), (R16, R17)
	LDP 112(R1), (R18, R19)
	LDP 128(R1), (R20, R21)
	LDP 144(R1), (R22, R23)
	LDP 160(R1), (R24, R25)
	LDP.W 176(R1), (R26, R27)

	STP (R4, R5)   ,(R0)
	STP (R6, R7)   ,16(R0)  
	STP (R8, R9)   ,32(R0)  
	STP (R10, R11) ,48(R0)  
	STP (R12, R13) ,64(R0)  
	STP (R14, R15) ,80(R0)  
	STP (R16, R17) ,96(R0)  
	STP (R18, R19) ,112(R0) 
	STP (R20, R21) ,128(R0) 
	STP (R22, R23) ,144(R0) 
	STP (R24, R25) ,160(R0) 
	STP.W (R26, R27) ,176(R0) 

	 SUBS $192, R2, R2
	 BGT bloop
	 RET

simd_test.go

package simd

import (
	"crypto/rand"
	"testing"
)

const bulk = 3

func BenchmarkAmove(b *testing.B) {
	dst := make([]byte, 64*bulk)
	src := make([]byte, 64*bulk)
	b.SetBytes(64 * bulk)
	b.ResetTimer()
	for i := 0; i < b.N; i++ {
		amove(dst, src, 64*bulk)
	}
}
func BenchmarkBmove(b *testing.B) {
	dst := make([]byte, 64*bulk)
	src := make([]byte, 64*bulk)
	b.SetBytes(64 * bulk)
	b.ResetTimer()
	for i := 0; i < b.N; i++ {
		bmove(dst, src, 64*bulk)
	}
}

func TestAmove(t *testing.T) {
	dst := make([]byte, 64*bulk)
	chk := make([]byte, 64*bulk)
	src := make([]byte, 64*bulk)
	rand.Read(src)
	copy(chk, src)
	amove(dst, src, 64*bulk)

	for i, c := range chk {
		if dst[i] != c {
			t.Errorf("dst[%d]%x != src[%d]%x", i, dst[i], i, c)
			break
		}
	}
}
func TestBmove(t *testing.T) {
	dst := make([]byte, 64*bulk)
	chk := make([]byte, 64*bulk)
	src := make([]byte, 64*bulk)
	rand.Read(src)
	copy(chk, src)
	bmove(dst, src, 64*bulk)

	for i, c := range chk {
		if dst[i] != c {
			t.Errorf("dst[%d]%x != src[%d]%x", i, dst[i], i, c)
			break
		}
	}
}