lmonade project proposal

[Brice Boyer]

Hi,

this could be a late additional project for linbox/fflas-ffpack.

Best regards,

Brice.

== GPU acceleration for dense/sparse matrix multiplication on finite fields ==

Fast exact dense linear algebra on finite fields is the core of the C++ library [[http://linalg.org/projects/fflas-ffpack|FFLAS-FFPACK]] [1]. More importantly algorithms therein rely on the efficiency of matrix/matrix multiplication and matrix/vector multiplication. The numerical (sparse) BLAS are the building blocks underlying these algorithms.

Recently, a lot of effort has been put into re-factoring the dense code and introducing sparse matrix formats and operations. On the one hand, there is now a clean and efficient implementation for both sequential and shared memory matmul routines. On the other hand, using GPU acceleration  (OpenCL) for computations over \(F_p\) was introduced in [[http://linalg.org|LinBox]] [3]. The goal here is to make use of fast numerical GPU BLAS libraries ([[http://docs.nvidia.com/cuda/cublas/|cuBLAS]], [[http://docs.nvidia.com/cuda/cusparse/|cuSPARSE]]). Also, it would be nice to import/implement openCL fall back routines in FFLAS-FFPACK. 

A first project would consist in using the library for the dense/sparse matrix multipliation operations and write a \(fmod\) operation for the GPU in cuda/opencl. A little more challenging first project would consist in moving the matrix multiplication OpenCL code from LinBox to FFLAS-FFPACK. 

Depending on how this project goes and the goals of the student, it would also be interesting to add an offloading to the GPU mechanism to the existing multi-threaded code.

'''skills/prerequisites''': C/C++, basic linear algebra routines, Cuda or OpenCL

'''mentors:''' Brice Boyer, B. David Saunders

'''References''':
* [1] "Dense Linear Algebra over Finite Fields: the FFLAS and FFPACK packages." J.-G. Dumas, P. Giorgi and C. Pernet. [[http://hal.ccsd.cnrs.fr/ccsd-00018223|pdf]

 * [2] "Exact Sparse Matrix-Vector Multiplication on GPU's and Multicore Architectures" Brice Boyer,  Jean-Guillaume Dumas and Pascal Giorgi [[https://hal.archives-ouvertes.fr/hal-00475185|pdf]]

* [3] "Dealing with performance/portability and performance/accuracy trade-offs in heterogeneous computing systems: a case study with matrix multiplication modulo primes" Matthew Wezowicz, B. David Saunder and Michela Taufer [[http://gcl.cis.udel.edu/publications/conferences/2012_SPIE.pdf|pdf]]

'''contact:''' ffpack...@googlegroups.com

[Burcin Erocal]

On Sun, 15 Mar 2015 23:23:42 +0000
Brice Boyer <boyer...@gmail.com> wrote:

> this could be a late additional project for linbox/fflas-ffpack.

Thanks! I added it to the wiki page:

http://wiki.lmona.de/get_involved/projects

I hope you find good candidates.

Cheers,
Burcin