Hi all,

I'm trying to optimize a Matrix * Vector multiplication. What I have originally looks something like:

Code: Select all

Eigen::SparseMatrix<float, Eigen::ColMajor> originalMatrix;
Eigen::SparseVector<float> vec[3];
Eigen::SparseMatrix<float> result[3];
for( int i = 0; i < 3; i++ )
{
  result[i] = originalMatrix * vec[i];
}


What I'm trying to do is something like:

Code: Select all

Eigen::SparseMatrix<float, Eigen::ColMajor> originalMatrix;
Eigen::SparseVector<float3> vec;
Eigen::SparseMatrix<float3> result;

result = originalMatrix * vec;


where float3 is a struct containing 3 floats, and with all the relevant operators overridden. My reasoning is that it should use SSE / AVX instructions when multiplying 3 floats by 1 float. If the code already uses SSE instructions, I'd like to see the performance difference when using this approach.

I'm currently getting errors about conj_helper not being defined (in gebp_kernel::operator() in Core\products\GeneralBlockPanelKernel.h). I'd really appreciate it if anyone could explain the high-level architecture in Core\arch\SSE\Complex.h and Core\util\BlasUtil.h (and whatever else needs to be modified).

Off topic - I have a small optimization for SparseMatrix::setFromTriplets() if anyone is interested (no custom instructions, just code re-arranging).

I'm confused because your code is showing sparse matrices while the reported error concerns dense matrix product...