有没有办法在特征平方范数中使下面的代码更快

Question

我在以下有 Eigen C++代码并进行 squredNorm 计算 10000 万次。

无论如何可以使其更强大/更快.

#include <Eigen/Core>
#include <tbb/parallel_for.h>
#include "tbb/tbb.h"
#include <mutex>
#include <opencv2/opencv.hpp>
int main(){ 
int numberOFdata = 10000008;
Eigen::MatrixXf feat = Eigen::MatrixXf::Random(numberOFdata,512);
Eigen::MatrixXf b_cmp= Eigen::MatrixXf::Random(1,512);
int count_feature = feat.rows();
std::vector<int> found_number ;
std::mutex mutex1;
for (int loop = 0 ; loop<16 ; loop++){
double start_1 = static_cast<double>(cv::getTickCount());                                                                  
tbb::affinity_partitioner ap;
tbb::parallel_for( tbb::blocked_range<int>(0,count_feature),
[&](tbb::blocked_range<int> r  ) 
{
for (int i=r.begin(); i<r.end(); ++i)
{
auto distance = ( feat.row(i)- b_cmp ).squaredNorm();                                   
if (distance < 0.5) {                                                                   
mutex1.lock();            
found_number.push_back(i);
mutex1.unlock();

}
}
},ap);
double timefin = ((double)cv::getTickCount() - start_1) / cv::getTickFrequency();
std::cout  << count_feature   << " TOTAL : "  << timefin << std::endl;
}
}

编译标志：

-Xpreprocessor   -std=c++11  -fopenmp -pthread -O3  -mavx2 -march=native -funroll-loops -fpermissive

特征版本 3.3.7 TBB OpenCV和Eigen链接。

您可以删除 opencv 并使用不同的已用时间计算。

谢谢

Answer 1

如果您以与访问feat相同的顺序存储，您应该快约 4 倍(即，在您的情况下为Eigen::RowMajor)。

删除所有与本征无关的内容的最小示例：

int numberOFdata = 10000008;
Eigen::Matrix<float,Eigen::Dynamic, Eigen::Dynamic, Eigen::RowMajor> feat = Eigen::MatrixXf::Random(numberOFdata, 512);
Eigen::RowVectorXf b_cmp = Eigen::MatrixXf::Random(1, 512);
int count_feature = feat.rows();
std::vector<int> found_number;
for (int loop = 0; loop < 16; loop++) {
auto start = std::chrono::steady_clock::now();
{
for (int i = 0; i < feat.rows(); ++i) {
float distance = (feat.row(i) - b_cmp).squaredNorm();
if (distance < 0.5f) {
found_number.push_back(i);
}
}
};
auto end = std::chrono::steady_clock::now();
std::chrono::duration<double> diff = end-start;
std::cout  << count_feature   << " TOTAL : "  <<
diff.count() << std::endl;
}

Godbolt-Demo(由于内存限制而减小了feat的维度)：https://godbolt.org/z/b6r5K4Yxv