單線程、SSE、AVX運行效率對比——最大值/最小值運算

前言

單列數組下，AVX比SSE稍慢！

math_function.h

#pragma once

#include <immintrin.h>
#include <stdio.h>

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

math_function.cpp

#include "math_function.h"

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_;
}

main.cpp

#include "math_function.h"
#include <random>
#include <time.h>

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

int main(int argc, char* argv[])
{
	int size = 58;
	float *input = (float *)malloc(sizeof(float) * size);

	default_random_engine e;
	uniform_real_distribution<float> u(0, 3); //隨機數分佈對象 
	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);

	getchar();
	free(input);

	return 0;
}

運行結果

測試硬件：CPU-i7-9700K

預處理器：_WINDOWS

命令行：/arch:AVX 

優化項：/O2

 

任何問題請加唯一QQ2258205918（名稱samylee）！