Open discussion for MAGMA library (Matrix Algebra on GPU and Multicore Architectures)
Post Reply
Posts: 12
Joined: Wed Jun 27, 2012 3:22 am


Post by arom » Fri Aug 10, 2012 2:32 am

Hi Mark,

1. could you please add Fortran interfaces for
  • magmaf_dgetri
  • magmaf_dstedc
  • magmaf_zhegvx
  • magmaf_zheevx
2. could you please fix zhegvx.cpp and comment/remove Mymagma_ztrmm routines

3. could you please add magma_dstedc function (see attachment)
my small updates to MAGMA-1.2.1
(18.91 KiB) Downloaded 1255 times

Posts: 916
Joined: Fri Jan 06, 2012 2:13 pm

Re: wishlist

Post by mgates3 » Mon Aug 20, 2012 1:29 pm

These seem reasonable, so hopefully we can get to them for the next release.

Posts: 12
Joined: Wed Jun 27, 2012 3:22 am

Re: wishlist

Post by arom » Wed Sep 05, 2012 1:25 am

Hi Mark!

I need zstedc function, which is absent in Magma.

Is it correct implementation:

Code: Select all

#include "common_magma.h"

extern "C" magma_int_t
magma_zstedc(char range, magma_int_t n, double* d, double* e, cuDoubleComplex* z, magma_int_t ldz,
             cuDoubleComplex* work, magma_int_t lwork, double *rwork, magma_int_t lrwork,
             magma_int_t* iwork, magma_int_t liwork, magma_int_t* info)

        double *dwork;

        if (MAGMA_SUCCESS != magma_dmalloc( &dwork, 3*n*(n/2 + 1) )) {
            *info = -15;
            return MAGMA_ERR_DEVICE_ALLOC;

        char range_t = ' ';
        if (range == 'I') range_t = 'A';
        magma_zstedx(range_t, n, 0., 0., 0, 0, d, e, z, ldz,
                     rwork, lrwork, iwork, liwork, dwork, info);

        magma_free( dwork );
Please, add it to the next release.

BTW, when will the next release be announced?


Posts: 18
Joined: Tue Jan 25, 2011 8:20 pm

Re: wishlist

Post by brom » Wed Sep 05, 2012 8:12 am

Is magma_zstedx even accelerated? It looks like a complete pass through to the host LAPACK.

Stan Tomov
Posts: 281
Joined: Fri Aug 21, 2009 10:39 pm

Re: wishlist

Post by Stan Tomov » Wed Sep 05, 2012 11:02 am

This routine becomes more compute intensive when eigenvectors are needed. In that case most of the flops are in gemm and this is what is GPU accelerated.

Post Reply