矩阵乘法的特征码运行比使用 std::vector 的循环乘法慢

Question

我正在学习 C++ 以及机器学习，所以我决定使用 Eigen 库进行矩阵乘法。我正在训练一个感知器来识别 MNIST 数据库中的数字。对于训练阶段，我将训练周期（或纪元）的数量设置为 T = 100。

'training matrix' 是一个 10000 x 785 矩阵。每行的第零个元素包含 'label' 标识输入数据（该行的其余 784 个元素）映射到的数字。

还有一个 784 x 1 'weights' 向量，其中包含 784 个特征中每个特征的权重。权重向量将与每个输入向量（训练矩阵的一行，不包括第零个元素）相乘，并在每次迭代时更新，这对于 10000 个输入中的每一个都会发生 T 次。

我编写了以下程序（它抓住了我正在做的事情的本质），我在其中比较了将矩阵的行与权重向量相乘的“普通”方法（使用 std::vector 和循环) 到我认为我可以用 Eigen 方法做的最好的事情。它实际上不是矩阵与向量的乘法，我实际上是对训练矩阵的行进行切片并将其与权重向量相乘。

std::vector 方法的训练周期持续时间为 160.662 毫秒，而 Eigen 方法通常超过 10,000 毫秒。

我使用以下命令编译程序：

clang++ -Wall -Wextra -pedantic -O3 -march=native -Xpreprocessor -fopenmp permute.cc -o perm -std=c++17

我使用的是“中期”2012 MacBook Pro 运行 macOS Catalina 并配备 2.5 GHz 双核 i5。

#include <iostream>
#include <algorithm>
#include <random>
#include <Eigen/Dense>
#include <ctime>
#include <chrono>
using namespace Eigen;

int main() {
    Matrix<uint8_t, Dynamic, Dynamic> m = Matrix<uint8_t, Dynamic, Dynamic>::Random(10000, 785);
    Matrix<double, 784, 1> weights_m = Matrix<double, 784, 1>::Random(784, 1);
    Matrix<uint8_t, 10000, 1> y_m, t_m;

    std::minstd_rand rng;
    rng.seed(time(NULL));
    std::uniform_int_distribution<> dist(0,1); //random integers between 0 and 1
    for (int i = 0; i < y_m.rows(); i++) {
        y_m(i) = dist(rng);
        t_m(i) = dist(rng);
    }

    int T = 100;
    int err;
    double eta;
    eta = 0.25; //learning rate
    Matrix<double, 1, 1> sum_wx_m;

    auto start1 = std::chrono::steady_clock::now(); //start of Eigen Matrix loop

    for (int iter = 0; iter < T; iter++) {
        for (int i = 0; i < m.rows(); i++) {
            sum_wx_m = m.block(i, 1, 1, 784).cast<double>() * weights_m;
        
            //some code to update y_m(i) based on the value of sum_wx_m which I left out
        
            err = y_m(i) - t_m(i);
            if (fabs(err) > 0) { //update the weights_m matrix if there's a difference between target and predicted
                weights_m = weights_m - eta * err * m.block(i, 1, 1, 784).transpose().cast<double>();
            } 
        }
    }

    auto end1 = std::chrono::steady_clock::now();
    auto diff1 = end1 - start1;
    std::cout << "Eigen matrix time is "<<std::chrono::duration <double, std::milli> (diff1).count() << " ms" << std::endl;

    //checking how std::vector form performs;

    std::vector<std::vector<uint8_t>> v(10000);
    std::vector<double> weights_v(784);
    std::vector<uint8_t> y_v(10000), t_v(10000);

    for (unsigned long i = 0; i < v.size(); i++) {
        for (int j = 0; j < m.cols(); j++) {
            v[i].push_back(m(i, j));
        }
    }

    for (unsigned long i = 0; i < weights_v.size(); i++) {
        weights_v[i] = weights_m(i);
    }

    for (unsigned long i = 0; i < y_v.size(); i++) {
        y_v[i] = dist(rng);
        t_v[i] = dist(rng);
    }

    double sum_wx_v;

    auto start2 = std::chrono::steady_clock::now(); //start of vector loop

    for (int iter = 0; iter < T; iter++) {
        for(unsigned long j = 0; j < v.size(); j++) {
            sum_wx_v = 0.0;
            for (unsigned long k = 1; k < v[0].size() ; k++) {
                sum_wx_v += weights_v[k - 1] * v[j][k];
            }
        
            //some code to update y_v[i] based on the value of sum_wx_v which I left out
        
            err = y_v[j] - t_v[j];
            if (fabs(err) > 0) {//update the weights_v matrix if there's a difference between target and predicted
                for (unsigned long k = 1; k < v[0].size(); k++) {
                    weights_v[k - 1] -= eta * err * v[j][k];
                }
            }
        }
    }

    auto end2 = std::chrono::steady_clock::now();
    auto diff2 = end2 - start2;
    std::cout << "std::vector time is "<<std::chrono::duration <double, std::milli> (diff2).count() << " ms" << std::endl;
}

我应该做哪些改变才能变得更好运行次？

Answer 1

可能不是最好的解决方案，但您可以尝试：

由于 Eigen 的默认数据顺序是 Column-Major 你可以让你的训练矩阵为 785x10000 这样每个训练 label/data 对在内存中都是连续的（也改变行 sum_wx_m 计算）。
使用fixed-size版本的块操作，即，您可以将m.block(i, 1, 1, 784)替换为m.block<1,784>(i, 1)（如果您已经切换了训练矩阵布局，则按相反的顺序排列，或者您可以简单地映射训练矩阵的数据部分并使用 .col()参考[见下面的例子])

这是您根据这些想法修改的代码：

#include <iostream>
#include <algorithm>
#include <random>
#include <Eigen/Dense>
#include <ctime>
#include <chrono>
using namespace Eigen;

int main() {
    Matrix<uint8_t, Dynamic, Dynamic> m = Matrix<uint8_t, Dynamic, Dynamic>::Random(785, 10000);
    Map<Matrix<uint8_t, Dynamic, Dynamic>> m_data(m.data() + 785, 784, 10000);

    Matrix<double, 784, 1> weights_m = Matrix<double, 784, 1>::Random(784, 1);
    Matrix<uint8_t, 10000, 1> y_m, t_m;

    std::minstd_rand rng;
    rng.seed(time(NULL));
    std::uniform_int_distribution<> dist(0,1); //random integers between 0 and 1
    for (int i = 0; i < y_m.rows(); i++) {
        y_m(i) = dist(rng);
        t_m(i) = dist(rng);
    }

    int T = 100;
    int err;
    double eta;
    eta = 0.25; //learning rate
     Matrix<double, 1, 1> sum_wx_m;

    auto start1 = std::chrono::steady_clock::now(); //start of Eigen Matrix loop

    for (int iter = 0; iter < T; iter++) {
        for (int i = 0; i < m.cols(); i++) {
            sum_wx_m = weights_m.transpose() * m_data.col(i).cast<double>();
        
            //some code to update y_m(i) based on the value of sum_wx_m which I left out
        
            err = y_m(i) - t_m(i);
            if (fabs(err) > 0) { //update the weights_m matrix if there's a difference between target and predicted
                weights_m = weights_m - eta * err * m_data.col(i).cast<double>();
            } 
        }
    }

    auto end1 = std::chrono::steady_clock::now();
    auto diff1 = end1 - start1;
    std::cout << "Eigen matrix time is "<<std::chrono::duration <double, std::milli> (diff1).count() << " ms" << std::endl;

    //checking how std::vector form performs;

    std::vector<std::vector<uint8_t>> v(10000);
    std::vector<double> weights_v(784);
    std::vector<uint8_t> y_v(10000), t_v(10000);

    for (unsigned long i = 0; i < v.size(); i++) {
        for (int j = 0; j < m.rows(); j++) {
            v[i].push_back(m(j, i));
        }
    }

    for (unsigned long i = 0; i < weights_v.size(); i++) {
        weights_v[i] = weights_m(i);
    }

    for (unsigned long i = 0; i < y_v.size(); i++) {
        y_v[i] = dist(rng);
        t_v[i] = dist(rng);
    }

    double sum_wx_v;

    auto start2 = std::chrono::steady_clock::now(); //start of vector loop

    for (int iter = 0; iter < T; iter++) {
        for(unsigned long j = 0; j < v.size(); j++) {
            sum_wx_v = 0.0;
            for (unsigned long k = 1; k < v[0].size() ; k++) {
                sum_wx_v += weights_v[k - 1] * v[j][k];
            }
        
            //some code to update y_v[i] based on the value of sum_wx_v which I left out
        
            err = y_v[j] - t_v[j];
            if (fabs(err) > 0) {//update the weights_v matrix if there's a difference between target and predicted
                for (unsigned long k = 1; k < v[0].size(); k++) {
                    weights_v[k - 1] -= eta * err * v[j][k];
                }
            }
        }
    }

    auto end2 = std::chrono::steady_clock::now();
    auto diff2 = end2 - start2;
    std::cout << "std::vector time is "<<std::chrono::duration <double, std::milli> (diff2).count() << " ms" << std::endl;
}

我已经在我的 Ubuntu 带有 i7-9700K 的桌面上编译了这段代码：

g++ -Wall -Wextra -O3 -std=c++17
====================================
Eigen matrix time is 110.523 ms
std::vector time is 117.826 ms


g++ -Wall -Wextra -O3 -march=native -std=c++17
=============================================
Eigen matrix time is 66.3044 ms
std::vector time is 71.2296 ms

Answer 2

经过与用户 J. Schultke 和 puhu 的讨论，我对我的代码进行了以下更改：

我已将所有 m.block(i, 1, 1, 784) 调用更改为 m.block<1, 784 >(i, 1)，这将本征矩阵循环所需的时间减少了 third。（首先由 J. Schultke 建议）
我已声明我的 m 矩阵以 RowMajor 顺序存储。这是因为默认情况下，本征矩阵以 ColMajor (column-major) 顺序存储。这将导致连续存储一行中的每个条目。所以现在 m.block() 调用，我用来引用 m 矩阵中一行的 slice，会简单一次获取整个内存块，将“本征矩阵”时间减少到“std::vector”时间以下。 (由puhu推荐)

现在的平均运行时间是

cpp:Pro$ ./perm
Eigen matrix time is 134.76 ms
std::vector time is 155.574 ms

修改后的代码为：

#include <iostream>
#include <algorithm>
#include <random>
#include <Eigen/Dense>
#include <chrono>
#include <ctime>
using namespace Eigen;
int main() {
    Matrix<uint8_t, Dynamic, Dynamic, RowMajor> m = Matrix<uint8_t, Dynamic, Dynamic, RowMajor>::Random(10000, 785);
    Matrix<double, 784, 1> weights_m = Matrix<double, 784, 1>::Random(784, 1);
    Matrix<uint8_t, 10000, 1> y_m, t_m;
    std::minstd_rand rng;
    rng.seed(time(NULL));
    std::uniform_int_distribution<> dist(0,1); //random integers between 0 and 1
    for (int i = 0; i < y_m.rows(); i++) {
        y_m(i) = dist(rng);
        t_m(i) = dist(rng);
    }

    int T = 100;
    int err;
    double eta;
    eta = 0.25; //learning rate
    Matrix<double, 1, 1> sum_wx_m;

    auto start1 = std::chrono::steady_clock::now(); //start of Eigen Matrix loop

    for (int iter = 0; iter < T; iter++) {
        for (int i = 0; i < m.rows(); i++) {
            auto b = m.block<1, 784>(i, 1).cast<double>();
            sum_wx_m = b * weights_m;
    
            //some code to update y_m(i) based on the value of sum_wx_m which I left out
    
            err = y_m(i) - t_m(i);
            if (fabs(err) > 0) { //update the weights_m matrix if there's a difference between target and predicted
                weights_m = weights_m - eta * err * b.transpose();
            } 
        }
    }

    auto end1 = std::chrono::steady_clock::now();
    auto diff1 = end1 - start1;
    std::cout << "Eigen matrix time is "<<std::chrono::duration <double, std::milli> (diff1).count() << " ms" << std::endl;

    //checking how std::vector form performs;

    std::vector<std::vector<uint8_t>> v(10000);
    std::vector<double> weights_v(784);
    std::vector<uint8_t> y_v(10000), t_v(10000);

    for (unsigned long i = 0; i < v.size(); i++) {
        for (int j = 0; j < m.cols(); j++) {
            v[i].push_back(m(i, j));
        }
    }

    for (unsigned long i = 0; i < weights_v.size(); i++) {
        weights_v[i] = weights_m(i);
    }

    for (unsigned long i = 0; i < y_v.size(); i++) {
        y_v[i] = dist(rng);
        t_v[i] = dist(rng);
    } 

    double sum_wx_v;

    auto start2 = std::chrono::steady_clock::now(); //start of vector loop

    for (int iter = 0; iter < T; iter++) {
        for(unsigned long j = 0; j < v.size(); j++) {
            sum_wx_v = 0.0;
            for (unsigned long k = 1; k < v[0].size() ; k++) {
                sum_wx_v += weights_v[k - 1] * v[j][k];
            }
    
            //some code to update y_v[i] based on the value of sum_wx_v which I left out
    
            err = y_v[j] - t_v[j];
            if (fabs(err) > 0) {//update the weights_v matrix if there's a difference between target and predicted
                for (unsigned long k = 1; k < v[0].size(); k++) {
                    weights_v[k - 1] -= eta * err * v[j][k];
                }
            }
        }
    }

    auto end2 = std::chrono::steady_clock::now();
    auto diff2 = end2 - start2;
    std::cout << "std::vector time is "<<std::chrono::duration <double, std::milli> (diff2).count() << " ms" << std::endl;
}

矩阵乘法的特征码运行比使用 std::vector 的循环乘法慢

Eigen code for matrix multiplication running slower than looped multiplication using std::vector

c++

matrix

eigen

矩阵乘法的特征码 运行 比使用 std::vector 的循环乘法慢

Eigen code for matrix multiplication running slower than looped multiplication using std::vector

c++

matrix

eigen

矩阵乘法的特征码运行比使用 std::vector 的循环乘法慢