bench.go

package main

import (
	"bytes"
	"crypto/aes"
	"fmt"
	"runtime"
	"sync"
	"time"
)

func singleThreaded(rounds int, key, origPayload [32]byte) []byte {
	payload := make([]byte, 32)
	copy(payload, origPayload[:])

	cipher, err := aes.NewCipher(key[:])
	if err != nil {
		panic(err.Error())
	}

	for i := 0; i < rounds; i++ {
		cipher.Encrypt(payload[0:16], payload[0:16])
		cipher.Encrypt(payload[16:32], payload[16:32])
	}

	return payload
}

func dualThreaded(rounds int, key, origPayload [32]byte) []byte {
	payload := make([]byte, 32)
	copy(payload, origPayload[:])

	cipher, err := aes.NewCipher(key[:])
	if err != nil {
		panic(err.Error())
	}

	wg := sync.WaitGroup{}

	wg.Add(1)
	go func() {
		for i := 0; i < rounds; i++ {
			cipher.Encrypt(payload[0:16], payload[0:16])
		}
		wg.Done()
	}()

	wg.Add(1)
	go func() {
		for i := 0; i < rounds; i++ {
			cipher.Encrypt(payload[16:32], payload[16:32])
		}
		wg.Done()
	}()

	wg.Wait()

	return payload
}

func dualThreadedAvoidFalseSharing(rounds int, key, origPayload [32]byte) []byte {
	payload := make([]byte, 80)
	copy(payload[0:16], origPayload[0:16])
	copy(payload[64:80], origPayload[16:32])

	cipher, err := aes.NewCipher(key[:])
	if err != nil {
		panic(err.Error())
	}

	wg := sync.WaitGroup{}

	wg.Add(1)
	go func() {
		for i := 0; i < rounds; i++ {
			cipher.Encrypt(payload[0:16], payload[0:16])
		}
		wg.Done()
	}()

	wg.Add(1)
	go func() {
		for i := 0; i < rounds; i++ {
			cipher.Encrypt(payload[64:80], payload[64:80])
		}
		wg.Done()
	}()

	wg.Wait()

	result := make([]byte, 32)
	copy(result[0:16], payload[0:16])
	copy(result[16:32], payload[64:80])

	return result
}

func main() {
	runtime.GOMAXPROCS(4)

	key := [32]byte{0xe, 0x0, 0x0, 0x46, 0x0, 0x0, 0x5, 0x1, 0x0, 0x5e, 0x0, 0x4f, 0x58, 0x0, 0x42, 0x45, 0x8, 0x0, 0x0, 0x0, 0x7, 0x0, 0x0, 0x38, 0x6, 0x49, 0x1, 0x1b, 0x44, 0x0, 0x0, 0x4b}
	rounds := 20_000_000
	payload := [32]byte{0x0, 0x13, 0x39, 0x1e, 0x10, 0x8, 0x0, 0x0, 0x0, 0x24, 0x6, 0x0, 0x0, 0x1, 0x60, 0x1f, 0x37, 0x3, 0x0, 0x0, 0x0, 0x8, 0x41, 0x45, 0x2, 0x0, 0x0, 0x3f, 0x1, 0xc, 0x4b, 0x3e}
	expectedResult := [32]byte{0x94, 0x6c, 0x51, 0x9b, 0x92, 0x5, 0x20, 0xbd, 0x3e, 0x16, 0x46, 0x94, 0xf0, 0xa8, 0xd4, 0x97, 0xa, 0xf5, 0xcc, 0xbe, 0xd5, 0x9c, 0xef, 0xf0, 0x19, 0x8d, 0x61, 0xcb, 0xe0, 0x1, 0x90, 0x49}

	type bench struct {
		name string
		code func(int, [32]byte, [32]byte) []byte
	}

	benches := []bench{
		{"Single threaded", singleThreaded},
		{"Dual threaded", dualThreaded},
		{"Dual threaded, avoiding false sharing", dualThreadedAvoidFalseSharing},
	}

	for _, b := range benches {
		start := time.Now()
		gotResult := b.code(rounds, key, payload)
		elapsed := time.Since(start)

		if !bytes.Equal(gotResult, expectedResult[:]) {
			fmt.Printf("%s: Incorrect result\n", b.name)
		} else {
			fmt.Printf("%s: %v\n", b.name, elapsed)
		}
	}
}

results.txt

Single threaded: 964.790093ms
Dual threaded: 1.469041773s
Dual threaded, avoiding false sharing: 666.44416ms