单线程、SSE、AVX运行效率对比

测试硬件：CPU-i5-4590
命令行：/arch:AVX
优化项：/O2

main.cpp

#include <iostream>
#include <vector>
#include "method.h"
#include <random>
#include <time.h>

using std::default_random_engine;
using std::uniform_real_distribution;

int main(int argc, char* argv[])
{
	//乘法累加运算
	{
		int size = 33;
		float *input1 = (float *)malloc(sizeof(float) * size);
		float *input2 = (float *)malloc(sizeof(float) * size);

		default_random_engine e;
		uniform_real_distribution<float> u(0, 1); //随机数分布对象 
		for (int i = 0; i < size; i++)
		{
			input1[i] = u(e);
			input2[i] = u(e);
		}

		int cntLoop = 10000000;

		clock_t start_t = clock();
		float org = 0.0;
		for (int i = 0; i < cntLoop; i++)
			org = MathMulAdd(input1, input2, size);
		printf("org = %f\t", org);
		printf("cost time: %d(ms)\n", clock() - start_t);

		start_t = clock();
		float sse = 0.0;
		for (int i = 0; i < cntLoop; i++)
			sse = SSEMulAdd(input1, input2, size);
		printf("sse = %f\t", sse);
		printf("cost time: %d(ms)\n", clock() - start_t);

		start_t = clock();
		float sse_ = 0.0;
		for (int i = 0; i < cntLoop; i++)
			sse_ = SSEFmAdd(input1, input2, size);
		printf("sse_= %f\t", sse_);
		printf("cost time: %d(ms)\n", clock() - start_t);

		start_t = clock();
		float avx = 0.0;
		for (int i = 0; i < cntLoop; i++)
			avx = AVXMulAdd(input1, input2, size);
		printf("avx = %f\t", avx);
		printf("cost time: %d(ms)\n", clock() - start_t);

		start_t = clock();
		float avx_ = 0.0;
		for (int i = 0; i < cntLoop; i++)
			avx_ = AVXFmAdd(input1, input2, size);
		printf("avx_= %f\t", avx_);
		printf("cost time: %d(ms)\n", clock() - start_t);

		free(input1);
		free(input2);
	}
	//结果：
	//org = 11.216135 cost time : 174(ms)
	//sse = 11.216136 cost time : 102(ms)
	//sse_ = 11.216136 cost time : 119(ms)
	//avx = 11.216136 cost time : 63(ms)
	//avx_ = 11.216136 cost time : 61(ms)


	//加法运算
	//{
	//	int size = 27;
	//	float *input = (float *)malloc(sizeof(float) * size);
	//	for (int i = 0; i < size; i++)
	//		input[i] = 0.0025;

	//	int cntLoop = 300000000;
	//	clock_t start_t = clock();
	//	float org = 0.0;
	//	for (int i = 0; i < cntLoop; i++)
	//		org = MathSum(input, size);
	//	printf("org = %f\t", org);
	//	printf("cost time: %d\n", clock() - start_t);

	//	start_t = clock();
	//	float sse = 0.0;
	//	for (int i = 0; i < cntLoop; i++)
	//		sse = SSESum(input, size);
	//	printf("sse = %f\t", sse);
	//	printf("cost time: %d\n", clock() - start_t);

	//	start_t = clock();
	//	float avx = 0.0;
	//	for (int i = 0; i < cntLoop; i++)
	//		avx = AVXSum(input, size);
	//	printf("avx = %f\t", avx);
	//	printf("cost time: %d\n", clock() - start_t);

	//	free(input);
	//}
	//结果：
	//org = 0.067500  cost time : 3062
	//sse = 0.067500  cost time : 2283
	//avx = 0.067500  cost time : 1829


	//最大值/最小值运算
	//{
	//	int size = 58;
	//	float *input = (float *)malloc(sizeof(float) * size);

	//	default_random_engine e;
	//	uniform_real_distribution<float> u(0, 3); //随机数分布对象 
	//	for (int i = 0; i < size; i++)
	//	{
	//		input[i] = u(e);
	//		printf("%f ", input[i]);
	//		if ((i + 1) % 8 == 0)
	//			printf("\n");
	//	}
	//	printf("\n");

	//	int cntLoop = 100000000;
	//	clock_t start_t = clock();
	//	float org;
	//	for (int i = 0; i < cntLoop; i++)
	//		org = MathMax(input, size);
	//	printf("org = %f\t", org);
	//	printf("cost time: %d(ms)\n", clock() - start_t);

	//	start_t = clock();
	//	float sse;
	//	for (int i = 0; i < cntLoop; i++)
	//		sse = SSEMax(input, size);
	//	printf("sse = %f\t", sse);
	//	printf("cost time: %d(ms)\n", clock() - start_t);

	//	start_t = clock();
	//	float avx;
	//	for (int i = 0; i < cntLoop; i++)
	//		avx = AVXMax(input, size);
	//	printf("avx = %f\t", avx);
	//	printf("cost time: %d(ms)\n", clock() - start_t);

	//	free(input);
	//}
	//结果：
	//org = 2.989384  cost time : 9491(ms)
	//sse = 2.989384  cost time : 1261(ms)
	//avx = 2.989384  cost time : 1413(ms)


	return 0;
}

method.h

#pragma once

#include <intrin.h>
#include <stdio.h>

float MathMulAdd(const float *input1, const float *input2, int size);
float SSEMulAdd(const float *input1, const float *input2, int size);
float SSEFmAdd(const float *input1, const float *input2, int size);
float AVXMulAdd(const float *input1, const float *input2, int size);
float AVXFmAdd(const float *input1, const float *input2, int size);


float MathSum(const float *input, int size);
float SSESum(const float *input, int size);
float AVXSum(const float *input, int size);


float MathMax(const float *input, int size);
float SSEMax(const float *input, int size);
float AVXMax(const float *input, int size);

method.cpp

#include "method.h"


float MathMulAdd(const float *input1, const float *input2, int size)
{
	float output = 0.0;
	for (int i = 0; i < size; i++)
	{
		output += input1[i] * input2[i];
	}
	return output;
}

float SSEMulAdd(const float *input1, const float *input2, int size)
{
	if (input1 == nullptr || input2 == nullptr)
	{
		printf("input data is null\n");
		return -1;
	}
	int nBlockWidth = 4;
	int cntBlock = size / nBlockWidth;
	int cntRem = size % nBlockWidth;

	float output = 0;
	__m128 loadData1, loadData2;
	__m128 mulData = _mm_setzero_ps();
	__m128 sumData = _mm_setzero_ps();
	const float *p1 = input1;
	const float *p2 = input2;
	for (int i = 0; i < cntBlock; i++)
	{
		loadData1 = _mm_load_ps(p1);
		loadData2 = _mm_load_ps(p2);
		mulData = _mm_mul_ps(loadData1, loadData2);
		sumData = _mm_add_ps(sumData, mulData);
		p1 += nBlockWidth;
		p2 += nBlockWidth;
	}
	sumData = _mm_hadd_ps(sumData, sumData); // p[0] + p[1] + p[4] + p[5] + ...
	sumData = _mm_hadd_ps(sumData, sumData); // p[2] + p[3] + p[6] + p[7] + ...
	output += sumData.m128_f32[(0)];         // 前4组

	for (int i = 0; i < cntRem; i++)
	{
		output += p1[i] * p2[i];
	}

	return output;
}

float SSEFmAdd(const float *input1, const float *input2, int size)
{
	if (input1 == nullptr || input2 == nullptr)
	{
		printf("input data is null\n");
		return -1;
	}
	int nBlockWidth = 4;
	int cntBlock = size / nBlockWidth;
	int cntRem = size % nBlockWidth;

	float output = 0;
	__m128 loadData1, loadData2;
	__m128 sumData = _mm_setzero_ps();
	const float *p1 = input1;
	const float *p2 = input2;
	for (int i = 0; i < cntBlock; i++)
	{
		loadData1 = _mm_load_ps(p1);
		loadData2 = _mm_load_ps(p2);
		sumData = _mm_fmadd_ps(loadData1, loadData2, sumData);
		p1 += nBlockWidth;
		p2 += nBlockWidth;
	}
	sumData = _mm_hadd_ps(sumData, sumData); // p[0] + p[1] + p[4] + p[5] + ...
	sumData = _mm_hadd_ps(sumData, sumData); // p[2] + p[3] + p[6] + p[7] + ...
	output += sumData.m128_f32[(0)];         // 前4组

	for (int i = 0; i < cntRem; i++)
	{
		output += p1[i] * p2[i];
	}

	return output;
}

float AVXMulAdd(const float *input1, const float *input2, int size)
{
	if (input1 == nullptr || input2 == nullptr)
	{
		printf("input data is null\n");
		return -1;
	}
	int nBlockWidth = 8;
	int cntBlock = size / nBlockWidth;
	int cntRem = size % nBlockWidth;

	float output = 0;
	__m256 loadData1, loadData2;
	__m256 mulData = _mm256_setzero_ps();
	__m256 sumData = _mm256_setzero_ps();
	const float *p1 = input1;
	const float *p2 = input2;
	for (int i = 0; i < cntBlock; i++)
	{
		loadData1 = _mm256_load_ps(p1);
		loadData2 = _mm256_load_ps(p2);
		mulData = _mm256_mul_ps(loadData1, loadData2);
		sumData = _mm256_add_ps(sumData, mulData);
		p1 += nBlockWidth;
		p2 += nBlockWidth;
	}
	sumData = _mm256_hadd_ps(sumData, sumData); // p[0] + p[1] + p[4] + p[5] + p[8] + p[9] + p[12] + p[13] + ... 
	sumData = _mm256_hadd_ps(sumData, sumData); // p[2] + p[3] + p[6] + p[7] + p[10] + p[11] + p[14] + p[15] + ... 
	output += sumData.m256_f32[(0)];            // 前4组
	output += sumData.m256_f32[(4)];            // 后4组

	for (int i = 0; i < cntRem; i++)
	{
		output += p1[i] * p2[i];
	}

	return output;
}

float AVXFmAdd(const float *input1, const float *input2, int size)
{
	if (input1 == nullptr || input2 == nullptr)
	{
		printf("input data is null\n");
		return -1;
	}
	int nBlockWidth = 8;
	int cntBlock = size / nBlockWidth;
	int cntRem = size % nBlockWidth;

	float output = 0;
	__m256 loadData1, loadData2;
	__m256 sumData = _mm256_setzero_ps();
	const float *p1 = input1;
	const float *p2 = input2;
	for (int i = 0; i < cntBlock; i++)
	{
		loadData1 = _mm256_load_ps(p1);
		loadData2 = _mm256_load_ps(p2);
		sumData = _mm256_fmadd_ps(loadData1, loadData2, sumData);
		p1 += nBlockWidth;
		p2 += nBlockWidth;
	}
	sumData = _mm256_hadd_ps(sumData, sumData); // p[0] + p[1] + p[4] + p[5] + p[8] + p[9] + p[12] + p[13] + ... 
	sumData = _mm256_hadd_ps(sumData, sumData); // p[2] + p[3] + p[6] + p[7] + p[10] + p[11] + p[14] + p[15] + ... 
	output += sumData.m256_f32[(0)];            // 前4组
	output += sumData.m256_f32[(4)];            // 后4组

	for (int i = 0; i < cntRem; i++)
	{
		output += p1[i] * p2[i];
	}

	return output;
}



float MathSum(const float *input, int size)
{
	float output = 0.0;
	for (int i = 0; i < size; i++)
	{
		output += input[i];
	}
	return output;
}

float SSESum(const float *input, int size)
{
	if (input == nullptr)
	{
		printf("input data is null\n");
		return -1;
	}
	int nBlockWidth = 4;
	int cntBlock = size / nBlockWidth;
	int cntRem = size % nBlockWidth;

	float output = 0;
	__m128 loadData;
	__m128 sumData = _mm_setzero_ps();
	const float *p = input;
	for (int i = 0; i < cntBlock; i++)
	{
		loadData = _mm_load_ps(p);
		sumData = _mm_add_ps(sumData, loadData);
		p += nBlockWidth;
	}
	sumData = _mm_hadd_ps(sumData, sumData); // p[0] + p[1] + p[4] + p[5] + ...
	sumData = _mm_hadd_ps(sumData, sumData); // p[2] + p[3] + p[6] + p[7] + ...
	output += sumData.m128_f32[(0)];         // 前4组

	for (int i = 0; i < cntRem; i++)
	{
		output += p[i];
	}

	return output;
}

float AVXSum(const float *input, int size)
{
	if (input == nullptr)
	{
		printf("input data is null\n");
		return -1;
	}
	int nBlockWidth = 8;
	int cntBlock = size / nBlockWidth;
	int cntRem = size % nBlockWidth;

	float output = 0;
	__m256 loadData;
	__m256 sumData = _mm256_setzero_ps();
	const float *p = input;
	for (int i = 0; i < cntBlock; i++)
	{
		loadData = _mm256_load_ps(p);
		sumData = _mm256_add_ps(sumData, loadData);
		p += nBlockWidth;
	}
	sumData = _mm256_hadd_ps(sumData, sumData); // p[0] + p[1] + p[4] + p[5] + p[8] + p[9] + p[12] + p[13] + ... 
	sumData = _mm256_hadd_ps(sumData, sumData); // p[2] + p[3] + p[6] + p[7] + p[10] + p[11] + p[14] + p[15] + ... 
	output += sumData.m256_f32[(0)];            // 前4组
	output += sumData.m256_f32[(4)];            // 后4组

	for (int i = 0; i < cntRem; i++)
	{
		output += p[i];
	}

	return output;
}


float MathMax(const float *input, int size)
{
	float maxVal = input[0];
	for (int i = 1; i < size; i++)
	{
		maxVal = maxVal > input[i] ? maxVal : input[i];
	}

	return maxVal;
}

float SSEMax(const float *input, int size)
{
	if (input == nullptr)
	{
		printf("input data is null\n");
		return -1;
	}
	int nBlockWidth = 4;
	int cntBlock = size / nBlockWidth;
	int cntRem = size % nBlockWidth;

	__declspec(align(16)) float output[4];
	__m128 loadData;
	const float *p = input;

	__m128 maxVal = _mm_load_ps(p);
	p += nBlockWidth;

	for (int i = 1; i < cntBlock; i++)
	{
		loadData = _mm_load_ps(p);
		maxVal = _mm_max_ps(maxVal, loadData);

		p += nBlockWidth;
	}
	_mm_store_ps(output, maxVal);

	float maxVal_ = output[0];
	for (int i = 1; i < 4; i++)
	{
		maxVal_ = maxVal_ > output[i] ? maxVal_ : output[i];
	}
	for (int i = 0; i < cntRem; i++)
	{
		maxVal_ = maxVal_ > p[i] ? maxVal_ : p[i];
	}

	return maxVal_;
}

float AVXMax(const float *input, int size)
{
	if (input == nullptr)
	{
		printf("input data is null\n");
		return -1;
	}
	int nBlockWidth = 8;
	int cntBlock = size / nBlockWidth;
	int cntRem = size % nBlockWidth;

	__declspec(align(32)) float output[8];
	__m256 loadData;
	const float *p = input;

	__m256 maxVal = _mm256_load_ps(p);
	p += nBlockWidth;

	for (int i = 1; i < cntBlock; i++)
	{
		loadData = _mm256_load_ps(p);
		maxVal = _mm256_max_ps(maxVal, loadData);

		p += nBlockWidth;
	}
	_mm256_store_ps(output, maxVal);

	float maxVal_ = output[0];
	for (int i = 1; i < 8; i++)
	{
		maxVal_ = maxVal_ > output[i] ? maxVal_ : output[i];
	}
	for (int i = 0; i < cntRem; i++)
	{
		maxVal_ = maxVal_ > p[i] ? maxVal_ : p[i];
	}

	return maxVal_;
}