About multiplying a matrix with a vector: Why cblas_?gemm is faster than cblas_?gemv ?

[mintaka]

I'm trying to perform the multiplication between a matrix (`N x p`) and a vector (`p x 1`), where `N = 1000000` and `p = 600`. I find out that `cblas_?gemm` is around 2~3 times faster than `cblas_?gemv`. Shouldn't `cblas_?gemv` always be faster than `cblas_?gemm`?

Thanks in advance.

Below is my code:

#include <opencv2/core/core.hpp>
#include <iostream>
#include <cblas.h>
#include <cstdlib>

using namespace std;
using namespace cv;

int main(int argc, char* argv[]) {
    clock_t start_time;
    RNG &rng = theRNG();
    float alpha = 1.f, gamma = 0;

    float time_openblas;
    int m = 1000000, p = 600, n = 1;

    Mat_<float> A(m, p, CV_32F), B(p, n, CV_32F);
    Mat C = Mat(m, n, CV_32F, 0.0);
    rng.fill(A, RNG::NORMAL, 1, 100);
    rng.fill(B, RNG::NORMAL, 2, 50);
    float *src1, *src2, *src3;
    src1 = (float *) A.data;
    src2 = (float *) B.data;
    src3 = (float *) C.data;

    start_time = clock();
    cblas_sgemm(CblasRowMajor, CblasNoTrans, CblasNoTrans, m, n, p,
                alpha, src1, p, src2, n, beta, src3, n);
    float duration = (clock() - start_time) / (CLOCKS_PER_SEC / 1000.0);
    cout << "cblas_sgemm: " << duration << " ms" << endl;

    start_time = clock();
    cblas_sgemv(CblasRowMajor, CblasNoTrans, m, p, alpha, src1, p, src2, 1, beta, src3, 1);
    duration = (clock() - start_time) / (CLOCKS_PER_SEC / 1000.0);
    cout << "cblas_sgemv: " << duration << " ms" << endl;

    return 0;
}

[Zhang Xianyi]

What's your CPU and OS?

Xianyi

--
You received this message because you are subscribed to the Google Groups "OpenBLAS-users" group.
To unsubscribe from this group and stop receiving emails from it, send an email to openblas-user...@googlegroups.com.
To post to this group, send email to openbla...@googlegroups.com.
For more options, visit https://groups.google.com/d/optout.

[mintaka]

Forgot to mention that :p

Ubuntu 14.04 64-bit with Intel® Core™ i7-4770 CPU @ 3.40GHz × 8

[Zhang Xianyi]

I can reproduce this on my haswell machine. Looks like a performance bug for sgemv.

[mintaka]

Hi Xianyi, thanks for confirming that. Any plan to fix it in the near future?