我们目前正在使用C++编写一些对于大量矩阵和向量进行操作的性能关键代码。根据我们的研究,std::array
和标准C数组之间应该没有太大的性能差异(参见这个问题或这个问题)。然而,在测试过程中,我们发现使用C数组比std::array
可以获得巨大的性能提升。以下是我们的演示代码:
#include <iostream>
#include <array>
#include <sys/time.h>
#define ROWS 784
#define COLS 100
#define RUNS 50
using std::array;
void DotPComplex(array<double, ROWS> &result, array<double, ROWS> &vec1, array<double, ROWS> &vec2){
for(int i = 0; i < ROWS; i++){
result[i] = vec1[i] * vec2[i];
}
}
void DotPSimple(double result[ROWS], double vec1[ROWS], double vec2[ROWS]){
for(int i = 0; i < ROWS; i++){
result[i] = vec1[i] * vec2[i];
}
}
void MatMultComplex(array<double, ROWS> &result, array<array<double, COLS>, ROWS> &mat, array<double, ROWS> &vec){
for (int i = 0; i < COLS; ++i) {
for (int j = 0; j < ROWS; ++j) {
result[i] += mat[i][j] * vec[j];
}
}
}
void MatMultSimple(double result[ROWS], double mat[ROWS][COLS], double vec[ROWS]){
for (int i = 0; i < COLS; ++i) {
for (int j = 0; j < ROWS; ++j) {
result[i] += mat[i][j] * vec[j];
}
}
}
double getTime(){
struct timeval currentTime;
gettimeofday(¤tTime, NULL);
double tmp = (double)currentTime.tv_sec * 1000.0 + (double)currentTime.tv_usec/1000.0;
return tmp;
}
array<double, ROWS> inputVectorComplex = {{ 0 }};
array<double, ROWS> resultVectorComplex = {{ 0 }};
double inputVectorSimple[ROWS] = { 0 };
double resultVectorSimple[ROWS] = { 0 };
array<array<double, COLS>, ROWS> inputMatrixComplex = {{0}};
double inputMatrixSimple[ROWS][COLS] = { 0 };
int main(){
double start;
std::cout << "DotP test with C array: " << std::endl;
start = getTime();
for(int i = 0; i < RUNS; i++){
DotPSimple(resultVectorSimple, inputVectorSimple, inputVectorSimple);
}
std::cout << "Duration: " << getTime() - start << std::endl;
std::cout << "DotP test with C++ array: " << std::endl;
start = getTime();
for(int i = 0; i < RUNS; i++){
DotPComplex(resultVectorComplex, inputVectorComplex, inputVectorComplex);
}
std::cout << "Duration: " << getTime() - start << std::endl;
std::cout << "MatMult test with C array : " << std::endl;
start = getTime();
for(int i = 0; i < RUNS; i++){
MatMultSimple(resultVectorSimple, inputMatrixSimple, inputVectorSimple);
}
std::cout << "Duration: " << getTime() - start << std::endl;
std::cout << "MatMult test with C++ array: " << std::endl;
start = getTime();
for(int i = 0; i < RUNS; i++){
MatMultComplex(resultVectorComplex, inputMatrixComplex, inputVectorComplex);
}
std::cout << "Duration: " << getTime() - start << std::endl;
}
编译命令: icpc demo.cpp -std=c++11 -O0
编译结果如下:
DotP test with C array:
Duration: 0.289795 ms
DotP test with C++ array:
Duration: 1.98413 ms
MatMult test with C array :
Duration: 28.3459 ms
MatMult test with C++ array:
Duration: 175.15 ms
通过-O3
标志:
DotP test with C array:
Duration: 0.0280762 ms
DotP test with C++ array:
Duration: 0.0288086 ms
MatMult test with C array :
Duration: 1.78296 ms
MatMult test with C++ array:
Duration: 4.90991 ms
C数组实现在没有编译器优化的情况下速度更快。为什么? 使用编译器优化后,点积同样快。但是对于矩阵乘法,使用C数组仍然可以显著提速。 是否有一种方法可以在使用std :: array时实现相等的性能?
更新:
所用编译器:icpc 17.0.0
使用gcc 4.8.5
,我们的代码运行速度比使用任何优化级别的英特尔编译器都要慢得多。因此,我们主要关注英特尔编译器的行为。
如Jonas建议,我们调整了RUNS 50,000
,并获得以下结果(英特尔编译器):
使用-O0
标志:
DotP test with C array:
Duration: 201.764 ms
DotP test with C++ array:
Duration: 1020.67 ms
MatMult test with C array :
Duration: 15069.2 ms
MatMult test with C++ array:
Duration: 123826 ms
使用 -O3
标志:
DotP test with C array:
Duration: 16.583 ms
DotP test with C++ array:
Duration: 15.635 ms
MatMult test with C array :
Duration: 980.582 ms
MatMult test with C++ array:
Duration: 2344.46 ms