Re: Matrix multiplication performance

Hi Palik and Oswin,

I attached a simple implementation of a matrix-matrix product (gemm.hpp) which has less
then 250 lines of code.  I also attached a simple test programm (matprod.cc).  The test
Programm has a simple wrapper for the implementation in gemm.hpp.  For a real applications
you should use traits like for the BLAS bindings.  You can compile it with

g++ -O3 -Wall -std=c++11 -I $BOOST_PATH matprod.cc

where BOOST_PATH contains your boost implementation.  Let me know if this works
for you.


Cheers,

Michael










-----------------------------------------------------------------------------------
Dr. Michael Lehn
University of Ulm, Institute for Numerical Mathematics
Helmholtzstr. 20
D-89069 Ulm, Germany
Phone: (+49) 731 50-23534, Fax: (+49) 731 50-23548
-----------------------------------------------------------------------------------

On 18 Jan 2016, at 09:25, Oswin Krause <[email protected]> wrote:

> Hi Palik,
> 
> this is a known problem. In your case you should already get better performance when using axpy_prod instead of prod. There are currently moves towards a ublas 2.0 which should make this a non-problem in the future.
> 
> 
> On 2016-01-17 21:23, palik imre wrote:
>> Hi all,
>> It seems that the matrix multiplication in ublas ends up with the
>> trivial algorithm.  On my machine, even the following function
>> outperforms it for square matrices bigger than 173*173 (by a huge
>> margin for matrices bigger than 190*190), while not performing
>> considerably worse for smaller matrices:
>> matrix<double>
>> matmul_byrow(const matrix<double> &lhs, const matrix<double> &rhs)
>> {
>>  assert(lhs.size2() == rhs.size1());
>>  matrix<double> rv(lhs.size1(), rhs.size2());
>>  matrix<double> r = trans(rhs);
>>  for (unsigned c = 0; c < rhs.size2(); c++)
>>    {
>>      matrix_column<matrix<double> > out(rv, c);
>>      matrix_row<matrix<double> > in(r, c);
>>      out = prod(lhs, in);
>>    }
>>  return rv;
>> }
>> Is there anybody working on improving the matrix multiplication performance?
>> If not, then I can try to find some spare cycles ...
>> Cheers,
>> Imre Palik
>> _______________________________________________
>> ublas mailing list
>> [email protected]
>> http://lists.boost.org/mailman/listinfo.cgi/ublas
>> Sent to: [email protected]
> _______________________________________________
> ublas mailing list
> [email protected]
> http://lists.boost.org/mailman/listinfo.cgi/ublas
> Sent to: [email protected]