|
libflame
12600
|
Functions | |
| FLA_Error | FLA_Gemm_external (FLA_Trans transa, FLA_Trans transb, FLA_Obj alpha, FLA_Obj A, FLA_Obj B, FLA_Obj beta, FLA_Obj C) |
| FLA_Error FLA_Gemm_external | ( | FLA_Trans | transa, |
| FLA_Trans | transb, | ||
| FLA_Obj | alpha, | ||
| FLA_Obj | A, | ||
| FLA_Obj | B, | ||
| FLA_Obj | beta, | ||
| FLA_Obj | C | ||
| ) |
References bl1_cgemm(), bl1_dgemm(), bl1_sgemm(), bl1_zgemm(), FLA_Check_error_level(), FLA_Gemm_check(), FLA_Obj_col_stride(), FLA_Obj_datatype(), FLA_Obj_has_zero_dim(), FLA_Obj_length(), FLA_Obj_row_stride(), FLA_Obj_width(), FLA_Param_map_flame_to_blis_trans(), and FLA_Scal_external().
Referenced by FLA_Bidiag_UT_u_blf_var4(), FLA_Bidiag_UT_u_blk_var4(), FLA_Bidiag_UT_u_blk_var5(), FLA_Gemm(), FLA_Gemm_cc_task(), FLA_Gemm_ch_task(), FLA_Gemm_cn_task(), FLA_Gemm_ct_task(), FLA_Gemm_hc_task(), FLA_Gemm_hh_task(), FLA_Gemm_hn_task(), FLA_Gemm_ht_task(), FLA_Gemm_nc_task(), FLA_Gemm_nh_task(), FLA_Gemm_nn_task(), FLA_Gemm_nt_task(), FLA_Gemm_task(), FLA_Gemm_tc_task(), FLA_Gemm_th_task(), FLA_Gemm_tn_task(), FLA_Gemm_tt_task(), FLA_Hess_UT_blf_var2(), FLA_Hess_UT_blf_var3(), FLA_Hess_UT_blf_var4(), FLA_Hess_UT_blk_var1(), FLA_Hess_UT_blk_var2(), FLA_Hess_UT_blk_var3(), FLA_Hess_UT_blk_var4(), FLA_Hess_UT_blk_var5(), FLA_LQ_UT_blk_var2(), FLA_QR_UT_blk_var2(), FLA_QR_UT_piv_blk_var2(), FLA_SA_FS_blk(), and FLA_SA_LU_blk().
{
FLA_Datatype datatype;
int k_AB;
int m_A, n_A;
int m_C, n_C;
int rs_A, cs_A;
int rs_B, cs_B;
int rs_C, cs_C;
trans1_t blis_transa;
trans1_t blis_transb;
if ( FLA_Check_error_level() == FLA_FULL_ERROR_CHECKING )
FLA_Gemm_check( transa, transb, alpha, A, B, beta, C );
if ( FLA_Obj_has_zero_dim( C ) ) return FLA_SUCCESS;
if ( FLA_Obj_has_zero_dim( A ) || FLA_Obj_has_zero_dim( B ) )
{
FLA_Scal_external( beta, C );
return FLA_SUCCESS;
}
datatype = FLA_Obj_datatype( A );
m_A = FLA_Obj_length( A );
n_A = FLA_Obj_width( A );
rs_A = FLA_Obj_row_stride( A );
cs_A = FLA_Obj_col_stride( A );
rs_B = FLA_Obj_row_stride( B );
cs_B = FLA_Obj_col_stride( B );
m_C = FLA_Obj_length( C );
n_C = FLA_Obj_width( C );
rs_C = FLA_Obj_row_stride( C );
cs_C = FLA_Obj_col_stride( C );
if ( transa == FLA_NO_TRANSPOSE || transa == FLA_CONJ_NO_TRANSPOSE )
k_AB = n_A;
else
k_AB = m_A;
FLA_Param_map_flame_to_blis_trans( transa, &blis_transa );
FLA_Param_map_flame_to_blis_trans( transb, &blis_transb );
switch( datatype ){
case FLA_FLOAT:
{
float *buff_A = ( float * ) FLA_FLOAT_PTR( A );
float *buff_B = ( float * ) FLA_FLOAT_PTR( B );
float *buff_C = ( float * ) FLA_FLOAT_PTR( C );
float *buff_alpha = ( float * ) FLA_FLOAT_PTR( alpha );
float *buff_beta = ( float * ) FLA_FLOAT_PTR( beta );
bl1_sgemm( blis_transa,
blis_transb,
m_C,
k_AB,
n_C,
buff_alpha,
buff_A, rs_A, cs_A,
buff_B, rs_B, cs_B,
buff_beta,
buff_C, rs_C, cs_C );
break;
}
case FLA_DOUBLE:
{
double *buff_A = ( double * ) FLA_DOUBLE_PTR( A );
double *buff_B = ( double * ) FLA_DOUBLE_PTR( B );
double *buff_C = ( double * ) FLA_DOUBLE_PTR( C );
double *buff_alpha = ( double * ) FLA_DOUBLE_PTR( alpha );
double *buff_beta = ( double * ) FLA_DOUBLE_PTR( beta );
bl1_dgemm( blis_transa,
blis_transb,
m_C,
k_AB,
n_C,
buff_alpha,
buff_A, rs_A, cs_A,
buff_B, rs_B, cs_B,
buff_beta,
buff_C, rs_C, cs_C );
break;
}
case FLA_COMPLEX:
{
scomplex *buff_A = ( scomplex * ) FLA_COMPLEX_PTR( A );
scomplex *buff_B = ( scomplex * ) FLA_COMPLEX_PTR( B );
scomplex *buff_C = ( scomplex * ) FLA_COMPLEX_PTR( C );
scomplex *buff_alpha = ( scomplex * ) FLA_COMPLEX_PTR( alpha );
scomplex *buff_beta = ( scomplex * ) FLA_COMPLEX_PTR( beta );
bl1_cgemm( blis_transa,
blis_transb,
m_C,
k_AB,
n_C,
buff_alpha,
buff_A, rs_A, cs_A,
buff_B, rs_B, cs_B,
buff_beta,
buff_C, rs_C, cs_C );
break;
}
case FLA_DOUBLE_COMPLEX:
{
dcomplex *buff_A = ( dcomplex * ) FLA_DOUBLE_COMPLEX_PTR( A );
dcomplex *buff_B = ( dcomplex * ) FLA_DOUBLE_COMPLEX_PTR( B );
dcomplex *buff_C = ( dcomplex * ) FLA_DOUBLE_COMPLEX_PTR( C );
dcomplex *buff_alpha = ( dcomplex * ) FLA_DOUBLE_COMPLEX_PTR( alpha );
dcomplex *buff_beta = ( dcomplex * ) FLA_DOUBLE_COMPLEX_PTR( beta );
bl1_zgemm( blis_transa,
blis_transb,
m_C,
k_AB,
n_C,
buff_alpha,
buff_A, rs_A, cs_A,
buff_B, rs_B, cs_B,
buff_beta,
buff_C, rs_C, cs_C );
break;
}
}
return FLA_SUCCESS;
}
1.7.6.1