Hi,

I've done some benchmark on sparse matrix - dense vector multiplication with Eigen2 and Gmm++. Performance of this operation is very important to me because I have to do this lots of times in my project.

It turns out that Eigen2 is twice slower than Gmm++. Of course, there is no reason for Eigen2 to supposedly be faster than Gmm++, but I am just wondering here if I got this done correctly in my code, or if there's some possible improvements with Eigen2 for this matrix-vector operation.

TEST CODE:

#include <iostream>
#include <Eigen/Eigen>
#include <gmm/gmm.h>

const int N=30000;

int main() {

std::clock_t start;
double diff;

Eigen::DynamicSparseMatrix<double> m1(N,N);
Eigen::VectorXd v1(N), vv1(N);
gmm::col_matrix< gmm::rsvector<double> > m2(N,N);
std::vector<double> v2(N), vv2(N);

for (int i=0;i<N;i++) {
for (int j=0;j<N;j++) {
double r=drand48();
if (r<0.05) {
m1.coeffRef(i,j)=r;
m2(i,j)=r;
}
}
}

gmm::fill_random(v2);
for (int i=0;i<N;i++) v1[i]=v2[i];

start=std::clock();
for (int i=0;i<100;i++) vv1=m1*v1;
diff=(std::clock()-start)/(double)CLOCKS_PER_SEC;
std::cout << "EIGEN2: " << diff << std::endl;

start=std::clock();
for (int i=0;i<100;i++) gmm::mult(m2,v2,vv2);
diff=(std::clock()-start)/(double)CLOCKS_PER_SEC;
std::cout << "GMM++: " << diff << std::endl;

}

RESULT:

EIGEN2: 44.06
GMM++: 19.54

The code was compiled with g++ 4.1.2 and -O2 -msse2 optimization switches.

Regards,

DH

Have you tried using SparseMatrix instead of DynamicSparseMatrix? Don't know if it makes a difference, but I'd guess so.

hi,

indeed, I confirm the issue, and I'm pretty sure that wasn't the case when I implemented this product because I always bench against gmm (e.g., see what is in the bench/ folder). I'll see what's going wrong...

ok, the problem was that GCC generated poor code for some unknown reasons... the workaround was to split the function into two smaller one. Now Eigen is as fast as gmm for that product

Thanks for quick response and fix! Now I can see that eigen2-devel performs as fast as gmm++ does on that product.