矩阵乘法的特征码比使用 std::vector 的循环乘法运行速度慢

Question

I am learning C++ as well as machine learning, so I decided to use the Eigen library for matrix multiplication. I was training a perceptron to recognise a digit from the MNIST database. For the training phase I set the number of training cycles (or epochs) to T = 100.

The 'training matrix' is a 10000 x 785 matrix. The zeroth element of each row contains the 'label' identifying the digit to which the input data (the remaining 784 elements of the row) maps to.

There is also a 784 x 1 'weights' vector which contains the weights for each of the 784 features. The weights vector would be multiplied with each input vector (a row of the training matrix excluding the zeroth element) and would be updated every iteration, and this would happen T times for each of the 10000 inputs.

I wrote the following program (which captures the essence of what I am doing), where I compared the "vanilla" approach of multiplying the rows of a matrix with the weight vector (using std::vector and loops) to what I felt was the best I could do with an Eigen approach. It is not really a multiplication of a matrix with a vector, I am actually slicing the row of the training matrix and multiplying that with the weight vector.

The time duration for the training period for the std::vector approach was 160.662 ms and for the Eigen method was usually over 10,000 ms.

I compile the program using the following command:

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

I am using a "mid" 2012 MacBook Pro running macOS Catalina and having 2.5 GHz dual core 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;
}

What changes should I make to get better running times?

Answer 1

Might not be the best solution but you can try:

• Since default data order of Eigen is Column-Major you can let your training matrix be 785x10000 such that each training label/data pair will be contiguous in memory (also change the line where sum_wx_m is computed).
• Use fixed-size version of block operations, ie, you can replace m.block(i, 1, 1, 784) with m.block<1,784>(i, 1) (in reverse order if you have already switched your training matrix layout or you can simply map the data part of your training matrix and use.col() reference [see the below example])

Here is your code modified based on these ideas:

#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;
}

I have compiled this code in my Ubuntu Desktop with i7-9700K:

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

After discussions with users J. Schultke and puhu, I have made the following changes in my code:

1. I have changed all the m.block(i, 1, 1, 784) calls to m.block<1, 784>(i, 1) , this reduces the time required for the Eigen matrix loop by a third . (first suggested by J. Schultke)
2. I have declared my m matrix as being stored in RowMajor order. This is because by default the Eigen matrices are stored in ColMajor (column-major) order. This would cause each entry in a row to be stored contiguously. So now the m.block() calls, which I use to refer to a slice of a row in the m matrix, would simply fetch the whole chunk of memory at once, reducing the "Eigen matrix" time to below the "std::vector" time. (suggested by puhu)

The average runtimes now are

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

and the modified code is:

#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;
}