README

This is the extenstions plugin for PSBLAS, version 1.3.1.

Version 1.3.1 supports versions of CUDA up to 10, or from 11.3.
In version 11, support for HYB has been dropped.

This package contains: 
1. Extended matrix formats: ELLPACK, Hacked ELLPACK, DIAgonals, Hacked
   DIAgonals. Note: DIA and HDIA have limited support.      
2. A GPU plugin: gpu-enabled versions of the above, with the CUDA code
   from https://github.com/davidebarbieri/spgpu,  plus interfaces to
   CSR and HYB formats available in the NVIDIA CuSPARSE lib.
   Note: DIAG and HDIAG have limited support. 
3. RSB: an interface to http://sourceforge.net/projects/librsb


The architectural ideas of the GPU plugin are explained in 

 V. Cardellini, S. Filippone and D. Rouson
 Design Patterns for Scientific Computations on
 Sparse Matrices
 Scientific Computing, 22(2014), pp. 1--19.
 
The architecture of the Fortran 2003 sparse BLAS is described in 

S. Filippone, A. Buttari:
Object-Oriented Techniques for Sparse Matrix Computations in Fortran
2003, 
ACM Trans. on Math. Software, vol. 38, No. 4, 2012.



PREREQUISITES

To build this code you need to have PSBLAS 3.3.0 or later, together
with its prerequisites. 

To make use of the NVIDIA GPU you'll need:
1. An installation of the CUDA toolkit (version 4.1 or later); 
2. The SPGPU code from http://spgpu.googlecode.com



INSTALLING

./configure --prefix=/path/to/install \
            --with-psblas=/path/to/PSBLAS/install \
	    --with-cuda=/CUDA/install \
	    --with-spgpu=/SPGPU/install \
	    --with-librsb=/LIBRSB/install

make;
make install

Note: we have only tested with GNU Fortran compiler. 
Note: CUDA nvcc typically lags behind the latest  versions of GCC/GNU
      Fortran; currently nvcc supports GCC 4.8 so this is the preferred choice.
      Mixing SPGPU CUDA code  compiled with an older version and the rest with
      e.g. 4.9 has  worked fine so far: YMMV.  


TODO

Improve MPI support.


WHAT IS NOT HERE

Good preconditioners for the GPU. Performance of  triangular system
solves on the GPU is very bad: we enable it in CSRG and HYBG, we do
not even bother to implement it in ELG and HLG.  
So if you use the GPU, you are limited to no preconditioning, or
diagonal scaling. We are working on an independent plugin for mld2p4 
(www.mld2p4.it) that will deliver better alternatives based on
approximate inverses.  


Report bugs to:
 https://github.com/sfilippone/psblas3-ext/issues

Contributors
 Salvatore Filippone     
 Alessandro Fanfarillo