Commit a49ebd61 authored by D'azevedo, Ed's avatar D'azevedo, Ed
Browse files

add extra checks

parent 404eeace

src/dmrg_vbatch.c

+119 −4
Original line number Diff line number Diff line
@@ -17,8 +17,14 @@ 


#include "dmrg_magma.h"



#define MAXGPUS 8

   static magma_queue_t queue_array[MAXGPUS];

   static magma_device_t device_array[MAXGPUS];

   static int ngpu = 0;



   static magma_queue_t queue = 0;

   static int device = 0;



#endif
@@ -28,11 +34,35 @@ void dmrg_init() 
     is_initialized = 1;



#ifdef USE_MAGMA

     const int idebug = 1;



     device = 0;

     magma_init();



     magma_getdevices( device_array, MAXGPUS, &ngpu );

     assert( ngpu >= 1 );



      

     if (idebug >= 1) {

         printf("dmrg_init: ngpu = %d \n",ngpu);

     };





     int idev = 0;

     for (idev = 0; idev < ngpu; idev++) {

       device = device_array[idev];



       magma_setdevice( device );

       magma_queue_create( device, &queue );

       assert( queue != 0 );

       queue_array[idev] = queue;

     };



     idev = 0;

     queue = queue_array[idev];

     device = device_array[idev];





#endif

     };
@@ -355,7 +385,7 @@ double elapsed_time = 0; 
   FpType beta = beta_vbatch[0];





     

  if (ngpu == 1) { 



   magmablas_Xgemm_vbatched( transA, transB,

           m_vbatch, n_vbatch, k_vbatch,
@@ -365,9 +395,94 @@ double elapsed_time = 0; 
           beta,

           c_vbatch, ldc_vbatch,

           batch_size, queue );

  }

  else {

    /*

     * --------------------------------------------

     * simple partitioning of work to multiple GPUs

     * --------------------------------------------

     */

    int inc = (batch_size + (ngpu-1))/ngpu;

    int idev = 0;

    for(idev = 0; idev < ngpu; idev++) {

       int istart = idev * inc;

       int iend = istart+inc-1;

       if (iend >= (batch_size-1)) {

           iend = batch_size-1;

           };

       int isize =  (iend - istart + 1);



       device = device_array[idev];

       queue = queue_array[idev];

       magma_setdevice( device );





       int *pm_vbatch = &(m_vbatch[istart]);

       int *pn_vbatch = &(n_vbatch[istart]);

       int *pk_vbatch = &(k_vbatch[istart]);



       FpType **pa_vbatch = &(a_vbatch[istart]);

       FpType **pb_vbatch = &(b_vbatch[istart]);

       FpType **pc_vbatch = &(c_vbatch[istart]);





       int *plda_vbatch = &(lda_vbatch[istart]);

       int *pldb_vbatch = &(ldb_vbatch[istart]);

       int *pldc_vbatch = &(ldc_vbatch[istart]);



       /*

        * ------------

        * extra checks

        * ------------

        */

       int i = 0;

       for(i=0; i < isize; i++) {

          int mm = pm_vbatch[i];

          int nn = pn_vbatch[i];

          int kk = pk_vbatch[i];



          int lda = plda_vbatch[i];

          int ldb = pldb_vbatch[i];

          int ldc = pldc_vbatch[i];



          FpType *Amat = a_vbatch[i];

          FpType *Bmat = b_vbatch[i];

          FpType *Cmat = c_vbatch[i];





          assert( mm >= 1 );

          assert( nn >= 1 );

          assert( kk >= 1 );



          assert( lda >= 1 );

          assert( ldb >= 1 );

          assert( ldc >= 1 );



          assert( Amat != 0 );

          assert( Bmat != 0 );

          assert( Cmat != 0 );

          };



       int pbatch_size = isize;



       if (pbatch_size >= 1) {

         magmablas_Xgemm_vbatched( transA, transB,

           pm_vbatch, pn_vbatch, pk_vbatch,

           alpha,

           (FpType const * const *) pa_vbatch, plda_vbatch, 

           (FpType const * const *) pb_vbatch, pldb_vbatch,

           beta,

           pc_vbatch, pldc_vbatch,

           pbatch_size, queue );

          };



       }; /* end for idev */



       idev = 0;

       device = device_array[idev];

       queue = queue_array[idev];



     }; /* end if (ngpu > 1) */

  };

#else

  {