Intel MKL dgesvd does not seem to return correct singular vectors

Angelo_L_1 — Wed, 02 Mar 2016 08:21:49 GMT

I'm using Intel MKL dgesvd in the code below. When comparing the results with Matlab, the singular values are ok, while the singular vectors are not.

The C++ code I'm using

#include <stdio.h>
#include <stdlib.h>
#include <sstream>
#include <string.h>
#include <iostream>
#include <fstream>
#include <time.h>
#include <mkl.h>
#include "mkl_lapacke.h"

#include "TimingCPU.h"
#include "InputOutput.h"

using namespace std;

/**********************************/
/* RANDOM MATRIX ENTRY GENERATION */
/**********************************/
double generateRandomMatrixEntry(int N) { return 2000. * ((double)rand() / (double)(RAND_MAX - 0.2) * (1. / (double)N)) + 100.*((double)rand() / (double)(RAND_MAX - 0.2) * (1. / (double)N)) + 24.; }

/********/
/* MAIN */
/********/
int main(int argc, char **argv) {

        int Nrows, Ncols, K, numExecutions;

    char *jobu = "S", *jobvt = "S";     // --- "N" means not to calculate the corresponding singular vectors

    TimingCPU timer;

    if (argc != 5) { printf("usage: svdMKLDouble Nrows Ncols K numExecutions\n"); exit(0); }
        { std::stringstream ss(argv[1]); ss >> Nrows; }
        { std::stringstream ss(argv[2]); ss >> Ncols; }
        { std::stringstream ss(argv[3]); ss >> K; }    
        { std::stringstream ss(argv[4]); ss >> numExecutions; }

        int lda = Nrows, ldu = Nrows, ldvt = Ncols;

        double TotalTime = 0., AverageTime = 0.;

    double *A       = (double*)mkl_malloc(Nrows * Ncols * K * sizeof(double), 64);  // --- Input matrices
    double *A_QUERY = (double*)mkl_malloc(Nrows * Ncols *     sizeof(double), 64);  // --- Input matrix for MKL query
    double *U       = (double*)mkl_malloc(Nrows * Nrows * K * sizeof(double), 64);  // --- Matrices of LEFT singular vectors
    double *S       = (double*)mkl_malloc(        Ncols * K * sizeof(double), 64);  // --- Singular values
    double *V       = (double*)mkl_malloc(Ncols * Ncols * K * sizeof(double), 64);  // --- Matrices of RIGHT singular vectors

    srand(time(NULL));
    //srand(0);
    int N = 5;

    /*************************************************/
    /* QUERY AND ALLOCATION OF THE OPTIMAL WORKSPACE */
    /*************************************************/
        for (int h = 0; h < Nrows * Ncols; h++) A_QUERY = generateRandomMatrixEntry(N);

    int info, lwork = -1;          
    double wkopt;
        dgesvd(jobu, jobvt, &Nrows, &Ncols, A_QUERY, &lda, S, U, &ldu, V, &ldvt, &wkopt, &lwork, &info);
        lwork = (int)wkopt;
        double *work = (double *)malloc(lwork * K * sizeof(double));                

    /**************/
    /* DUMMY CALL */
    /**************/
    for (int i = 0; i < K; i++)
            dgesvd(jobu, jobvt, &Nrows, &Ncols, A + (i * Nrows * Ncols), &lda, S + (i * Ncols), U + (i * Nrows * Nrows), &ldu, V + (i * Ncols * Ncols), &ldvt, work + i * lwork, &lwork, &info);

    /*********************/
    /* TIME MEASUREMENTS */
    /*********************/
    for (int k = 0; k < numExecutions; k++) {

        //srand(k);

            for (int h = 0; h < Nrows * Ncols * K; h++) { A = generateRandomMatrixEntry(N); }

            timer.StartCounter();
        #pragma omp parallel for
        for (int i = 0; i < K; i++) {
            dgesvd(jobu, jobvt, &Nrows, &Ncols, A + (i * Nrows * Ncols), &lda, S + (i * Ncols), U + (i * Nrows * Nrows), &ldu, V + (i * Ncols * Ncols), &ldvt, work + i * lwork, &lwork, &info);
            // --- Convergence check            
            if (info > 0) { printf("Convergence failure.\n"); exit(1); }
        }
        #pragma omp barrier   
            TotalTime += timer.GetCounter(); // --- TotalTime in ms

    }

    AverageTime = TotalTime / numExecutions;
    std::cout << std::scientific << Nrows << " " << Ncols << " " << K << " " << numExecutions << " " << AverageTime << " " << std::endl;

    /*****************************/
    /* SAVINGS FOR RESULT CHECKS */
    /*****************************/
    // --- Saving the last matrix
    saveCPUrealtxt<double>(A + ((K - 1) * Nrows * Ncols), "Source_Matrix.txt", Nrows * Ncols);  
    // --- Saving the SVD
    saveCPUrealtxt<double>(U + ((K - 1) * Nrows * Nrows), "U.txt", Nrows * Nrows);  
    saveCPUrealtxt<double>(V + ((K - 1) * Ncols * Ncols), "V.txt", Ncols * Ncols);  
    saveCPUrealtxt<double>(S + ((K - 1) * Ncols),         "S.txt", Ncols        );  

    return 0;
}

The compilation command

icpc -lpthread -lmkl_sequential -lmkl_core -lmkl_intel_lp64 -lrt -openmp svdMKLDouble.cpp TimingCPU.cpp InputOutput.cpp -O3 -o svdMKLDouble

Code usage as...

./svdMKLDouble 8 8 1 1

The Matlab code I'm using as reference

clear all
close all
clc

load Source_Matrix.txt
load U.txt
load V.txt
load S.txt

Source_Matrix = reshape(Source_Matrix, 8, 8);
U             = reshape(U, 8, 8);
V             = reshape(V, 8, 8);

[UU, SS, VV]  = svd(Source_Matrix);

100 * sqrt(sum(abs(diag(SS) - S).^2) / sum(abs(diag(SS)).^2))

100 * sqrt(sum(sum((abs(Source_Matrix - UU * SS * VV.').^2))) / sum(sum((abs(Source_Matrix).^2))))

100 * sqrt(sum(sum((abs(Source_Matrix - U  * diag(S)  * V).^2))) / sum(sum((abs(Source_Matrix).^2))))

The comparison results

ans =

    0.0512        ---> Singular values correctly computed


ans =

   5.1387e-13     ---> Matlab SVD works correctly


ans =

  132.3127        ---> Intel MKL singular values do not seem to be correct

The question: What I'm doing wrong?

topic Intel MKL dgesvd does not seem to return correct singular vectors in Intel® oneAPI Math Kernel Library

Intel MKL dgesvd does not seem to return correct singular vectors