加法SSE2的实现与C版本的时间对比

#include<iostream>
#include<emmintrin.h>
#include<time.h>
#include<Windows.h>
using namespace std;


void interAddSimd(const unsigned char* p1, const unsigned char* p2, unsigned char* result, int num)
{
	__m128i m1 = _mm_loadu_si128((__m128i*)p1);
	__m128i m2 = _mm_loadu_si128((__m128i*)p2);
	__m128i m3 = _mm_adds_epi8(m1, m2);

	_mm_storeu_si128((__m128i*)result, m3);
}

void interAdd(const unsigned char* p1, const unsigned char* p2, unsigned char* result, int num)
{
	for (int i = 0; i < num; i++)
	{
		result[i] = p1[i] + p2[i];
	}
}


void main()
{
	LARGE_INTEGER timeStart;
	LARGE_INTEGER timeEnd;
	LARGE_INTEGER frequency;
	double        quadpart;

	QueryPerformanceFrequency(&frequency);
	quadpart = (double)frequency.QuadPart;
	

	unsigned char p1[16] = { 1, 1, 2, 2, 3, 3, 4, 4, 5, 5, 6, 6, 7, 7, 8, 8 };
	unsigned char p2[16] = { 2, 2, 3, 3, 4, 4, 5, 5, 6, 6, 7, 7, 8, 8, 9, 9 };
	unsigned char p3c[16] = { 0 };
	unsigned char p3simd[16] = { 0 };

	int cycNum = 1<<25;
	int timeAdd_c, timeAdd_simd = 0;
	// test for c
	QueryPerformanceCounter(&timeStart);
	for (int i = 0; i < cycNum; i++)
	{
		interAdd(p1, p2, p3c, 16);
	}
	QueryPerformanceCounter(&timeEnd);
	timeAdd_c = 1000 * (timeEnd.QuadPart - timeStart.QuadPart) / quadpart; // ms

	// test for sse2
	QueryPerformanceCounter(&timeStart);
	for (int i = 0; i < cycNum; i++)
	{
		interAddSimd(p1, p2, p3simd, 16);
	}
	QueryPerformanceCounter(&timeEnd);
	timeAdd_simd = 1000 * (timeEnd.QuadPart - timeStart.QuadPart) / quadpart; // ms

	for (int i = 0; i < 16; i++)
	{
		cout << (int)p3c[i] <<"--"<<(int)p3simd[i] << endl;
	}
	cout << "c time is:" << timeAdd_c << endl;
	cout << "simd time is:" << timeAdd_simd << endl;
	system("pause");
}

结果如下：

循环2^25次，release 模式下， C代码时间是144ms, SIMD 时间是28ms, simd理论上一次对16个数进行运算，应该是C版本的1/16, 但实际时间是其0.194倍，约1/5的时间，这可能是因为for循环本身的比较和加法运算的原因， 还有程序主体相对循环比较简单，当循环2^30次时， C代码时间是3743ms, SIMD时间是925ms,是C代码时间的0.25倍。