xorsimd

xorsimd provides high-throughput XOR operations for byte slices. On amd64 it selects AVX-512/AVX2/SSE2 implementations, and falls back to portable Go code on other architectures.

Install

go get github.com/templexxx/xorsimd

Quick Start

package main

import "github.com/templexxx/xorsimd"

func main() {
	a := []byte{1, 2, 3, 4}
	b := []byte{9, 8, 7, 6}
	dst := make([]byte, 4)

	// dst = a ^ b
	n := xorsimd.Bytes(dst, a, b)
	_ = n

	// dst = src[0] ^ src[1] ^ src[2] ...
	src := [][]byte{a, b, []byte{1, 1, 1, 1}}
	n = xorsimd.Encode(dst, src)
	_ = n
}

API Notes

• Encode(dst, src) XORs all rows in src into dst and returns the number of bytes processed.
• Encode requires len(src) >= 1.
• Bytes(dst, a, b) is a convenience wrapper around Encode.
• Bytes8, Bytes16, Bytes8Align, and Bytes16Align are fixed-size helpers.
• BytesA processes len(a) bytes; BytesB processes len(b) bytes.

All APIs may reuse overlapping slices. For correctness, callers must ensure the input slices are long enough for the selected function variant.

Performance

Performance mainly depends on:

• CPU SIMD instruction support.
• Number of input source rows.
• Slice size.

Benchmark formula:

I/O = (src_num + 1) * vector_size / cost

Test environment:

• AWS c5d.xlarge
• Intel Xeon Platinum 8124M @ 3.00GHz
• Single physical core

Src Num	Vector Size	AVX512 I/O (MB/s)	AVX2 I/O (MB/s)	SSE2 I/O (MB/s)
5	4KB	270403.73	142825.25	74443.91
5	1MB	26948.34	26887.37	26950.65
5	8MB	17881.32	17212.56	16402.97
10	4KB	190445.30	102953.59	53244.04
10	1MB	26424.44	26618.65	26094.39
10	8MB	15471.31	14866.72	13565.80

Run Tests and Benchmarks

go test ./...
go test -bench=. -benchmem