MyMat0<T,ROWMAJOR> matMulOpt(MyMat0<T,ROWMAJOR> const & m1,MyMat0<T,ROWMAJOR> const & m2)
{
MyMat0<T,ROWMAJOR> res(m1.nrow(),m2.ncol(),0.);
std::vector<T> tmp(m2.nrow());
for(size_type j=0; j<m2.ncol();++j)
{
// Store the jth column of m2
for (size_type l=0; l<m2.nrow(); ++l)
{
tmp[l]=m2(l,j);
}
for(size_type i=0; i<m1.nrow();++i)
{
size_type row_starts= m1.getIndex(i,0);
res(i,j) = std::inner_product(tmp.begin(),tmp.end(),m1.cbegin()+row_starts,T(0));
/*
for(size_type k=0; k<m2.nrow();++k)
{
res(i,j) += m1(i,k)*tmp[k];
}
*/
}
}
return res;
}
MyMat0<T,ROWMAJOR> matMulOpt(MyMat0<T,ROWMAJOR> const & m1,MyMat0<T,COLUMNMAJOR> const & m2)
{
MyMat0<T,ROWMAJOR> res(m1.nrow(),m2.ncol(),0.);
for(size_type i=0; i<m1.nrow();++i)
{
for(size_type j=0; j<m2.ncol();++j)
{
size_type column_start = m2.getIndex(0,j);
size_type row_start = m1.getIndex(i,0);
size_type row_end = m1.getIndex(i+1,0);
res(i,j) = std::inner_product(
m1.cbegin()+row_start,
m1.cbegin()+row_end,
m2.cbegin()+column_start,
T(0));
}
}
/*
for(size_type k=0; k<m2.nrow();++k)
res(i,j) += m1(i,k)*m2(k,j);
*/
return res;
}
