Commit aecfdb5d authored by D'azevedo, Ed's avatar D'azevedo, Ed
Browse files

add dmrg_lacpy

parent c49ae3ef

src/dmrg_malloc.c

+47 −0
Original line number Diff line number Diff line
@@ -76,6 +76,53 @@ void *ptr = NULL; 
  return (ptr);

}





void dmrg_lacpy(  const char uplo,

                  const int m, const int n, 

                  const FpType *src, const int ld_src, 

                        FpType *dest, const int ld_dest )

{

#ifdef USE_MAGMA

        const int is_upper = (uplo == 'U') || (uplo == 'u');

        const int is_lower = (uplo == 'L') || (uplo == 'l');

        const int is_full = (!is_upper) && (!is_lower);



        int is_block_copy = is_full && (m == ld_src) && (m == ld_dest);

        if (is_block_copy) {

                const size_t nbytes = sizeof(FpType) * m * n;

                dmrg_memcpy( dest, src, nbytes );

        }

        else {

          const int min_mn = (m <= n) ? m : n;

          const int ncol = is_full ? n : min_mn;

          int jcol = 0; 



          for(jcol=0; jcol < ncol; jcol++) {

                const int irow = jcol;

                int istart = is_upper ? 0 :

                             is_lower ? irow :

                                         0;

                int iend = is_upper ? irow :

                           is_lower ? m-1  :

                                      m-1;

                int count =   iend - istart + 1;

                if (count >= 1) {



                    const FpType *psrc  =  src + jcol * ld_src  + istart;

                          FpType *pdest = dest + jcol * ld_dest + istart; 



                    const size_t nbytes = count * sizeof(FpType);

                    // dmrg_memcpy( pdest, psrc, nbytes  );

                    memcpy( pdest, psrc, nbytes );

                    };

               };

          };



#else

        Xlacpy_( &uplo, &m, &n, src, &ld_src, dest, &ld_dest );

#endif

}



void dmrg_prefetch_to_device( void *unified_memory_ptr, size_t nbytes )

{

        assert( unified_memory_ptr != 0);

src/dmrg_vbatch.h

+6 −0
Original line number Diff line number Diff line
@@ -54,6 +54,12 @@ int dmrg_is_managed( const void *ptr ); 
extern

void *dmrg_malloc( const size_t alloc_size );



extern

void dmrg_lacpy(  const char uplo,

                  const int m, const int n, 

                  const FpType *src, const int ld_src, 

                        FpType *dest, const int ld_dest );



extern

void dmrg_prefetch_to_device( void *unified_memory_ptr, size_t nbytes );

src/setup_sparse_batch.c

+9 −6
Original line number Diff line number Diff line
@@ -408,9 +408,11 @@ void setup_sparse_batch( 
        int isok = (1 <= m) && (1 <= n) && (m <= ld1) && (m <= ld2);

        assert( isok );



        Xlacpy_(  &uplo, &m, &n, 

                       Asrc, &ld1, 

                       Adest, &ld2);



        // --------------------------------------------------

        // Xlacpy_(  &uplo, &m, &n, Asrc, &ld1, Adest, &ld2);

        // --------------------------------------------------

        dmrg_lacpy( uplo, m,n,  Asrc,ld1, Adest, ld2 );

        

        Adest += (ld2 * n);

        };
@@ -425,9 +427,10 @@ void setup_sparse_batch( 
        int isok = (1 <= m) && (1 <= n) && (m <= ld1) && (m <= ld2);

        assert(isok);



        Xlacpy_(  &uplo, &m, &n, 

                       Bsrc, &ld1, 

                       Bdest, &ld2);

        // --------------------------------------------------

        // Xlacpy_(  &uplo, &m, &n, Bsrc, &ld1, Bdest, &ld2);

        // --------------------------------------------------

        dmrg_lacpy( uplo, m,n, Bsrc, ld1, Bdest, ld2 );



        Bdest += (ld2 * n);