Commit 66f63590 authored by D'azevedo, Ed's avatar D'azevedo, Ed
Browse files

Merge branch 'master' of https://code.ornl.gov/e6d/dmrgppPluginSc

parents 3d4fcde4 50ec5d19

src/Makefile

+1 −0
Original line number Diff line number Diff line
@@ -12,6 +12,7 @@ OBJS=\ 
	setup_sparse_batch.o \

        unsetup_sparse_batch.o \

        apply_Htarget_sparse.o \

        estimate_work.o \

        BatchedGemm.o



libdmrgppPluginSc.a: $(OBJS)

src/apply_Htarget_sparse.c

+8 −2
Original line number Diff line number Diff line
@@ -103,7 +103,10 @@ void apply_Htarget_sparse( 
  if (need_allocate_X) {

    X_ = (FpType *) dmrg_malloc( sizeof(FpType) * xy_size_dim );

    assert( X_ != NULL );

    memcpy( &(X_[0]), &(Xin_[0]), sizeof(FpType) * xy_size );

    void *dest = (void *) &(X_[0]);

    void *src = (void *) &(Xin_[0]);

    size_t count = sizeof(FpType) * xy_size;

    dmrg_memcpy( dest, src, count );

    };



  if (need_allocate_Y) {
@@ -501,7 +504,10 @@ void apply_Htarget_sparse( 
    dmrg_free( X_ ); X_ = NULL;

    };

 if (need_allocate_Y) {

   memcpy( &(Yout_[0]), &(Y_[0]), sizeof(FpType) * xy_size );

   void *dest = &(Yout_[0]);

   void *src = &(Y_[0]);

   size_t count = sizeof(FpType) * xy_size;

   dmrg_memcpy( dest, src, count );

   dmrg_free( Y_ ); Y_ = NULL;

   };

#endif

src/dmrg_malloc.c

+12 −0
Original line number Diff line number Diff line
@@ -20,6 +20,18 @@ int dmrg_is_managed( const void *ptr ) 
  return( is_managed );

}

  

void dmrg_memcpy(void *dest, const void *src, size_t count)

{

#ifdef USE_MAGMA

  cudaError_t istat = cudaMemcpy( dest, src, count, cudaMemcpyDefault );

  if (istat != cudaSuccess) {

    fprintf(stderr,"dmrg_memcpy: %s\n", cudaGetErrorString(istat));

    };

  assert( istat == cudaSuccess );

#else

  memcpy( dest, src, count );

#endif

}



void *dmrg_malloc(const size_t alloc_size )

{

src/dmrg_vbatch.h

+2 −0
Original line number Diff line number Diff line
@@ -45,6 +45,8 @@ extern "C" { 
extern

void dmrg_init();



extern

void dmrg_memcpy(void *dest, const void *src, size_t n);



extern

int dmrg_is_managed( const void *ptr );

src/estimate_work.c

0 → 100644
+76 −0
Original line number Diff line number Diff line 
#include "test_vbatch.h"



void estimate_work( int npatches, 

                    int left_patch_size_[],

                    int right_patch_size_[],

                    int nC_[], 

                    double *ptotal_gflops, 

                    double *pgmemA, 

                    double *pgmemB, 

                    double *pgmemBX )

#define nC(ipatch,jpatch) nC_[ ((ipatch)-1) + ((jpatch)-1)*npatches ]

#define left_patch_size(ipatch) left_patch_size_[(ipatch)-1]

#define right_patch_size(ipatch) right_patch_size_[(ipatch)-1]

{

/*

 -------------------

 estimate total work

 -------------------

 */

 assert( ptotal_gflops != NULL );

 assert( pgmemA != NULL );

 assert( pgmemB != NULL );

 assert( pgmemBX != NULL );

  

 double gmemA = 0.0;

 double gmemB = 0.0;

 double gmemBX = 0.0;



 double total_flops = 0.0;

 {

 int ipatch = 0;

 int jpatch = 0;





 for(jpatch=1; jpatch <= npatches; jpatch++) {

 for(ipatch=1; ipatch <= npatches; ipatch++) {

    int nop = nC(ipatch,jpatch);

    if (nop <= 0) continue;



    double flops_total = 0.0;

    double flops_method1 = 0.0;

    double flops_method2 = 0.0;



    /*

     --------------------------------------

     Note: evaluate (B * X ) * transpose(A)

     --------------------------------------

     */

     

    

    int nrowA = left_patch_size(ipatch);

    int ncolA = left_patch_size(jpatch);

    int nrowB = right_patch_size(ipatch);

    int ncolB = right_patch_size(jpatch);

    int ncolX  = ncolA;



    gmemA += nop * nrowA * ncolA;

    gmemB += nop * nrowB * ncolB;

    gmemBX += nop * nrowB * ncolX;



    cal_kron_flops( nrowA, nrowB, ncolA, ncolB, 

            &flops_total, &flops_method1,   &flops_method2);



    total_flops += flops_method1*nop;

    };

    };

  };

  

 double total_gflops = total_flops/(1000.0*1000.0*1000.0);

 *ptotal_gflops = total_gflops;



 *pgmemA = gmemA;

 *pgmemB = gmemB;

 *pgmemBX = gmemBX;



}