void bli_symm_front ( side_t side, obj_t* alpha, obj_t* a, obj_t* b, obj_t* beta, obj_t* c, cntx_t* cntx, cntl_t* cntl ) { obj_t a_local; obj_t b_local; obj_t c_local; // Check parameters. if ( bli_error_checking_is_enabled() ) bli_symm_check( side, alpha, a, b, beta, c, cntx ); // If alpha is zero, scale by beta and return. if ( bli_obj_equals( alpha, &BLIS_ZERO ) ) { bli_scalm( beta, c ); return; } // Reinitialize the memory allocator to accommodate the blocksizes // in the current context. bli_memsys_reinit( cntx ); // Alias A, B, and C in case we need to apply transformations. bli_obj_alias_to( *a, a_local ); bli_obj_alias_to( *b, b_local ); bli_obj_alias_to( *c, c_local ); // An optimization: If C is stored by rows and the micro-kernel prefers // contiguous columns, or if C is stored by columns and the micro-kernel // prefers contiguous rows, transpose the entire operation to allow the // micro-kernel to access elements of C in its preferred manner. if ( bli_cntx_l3_ukr_dislikes_storage_of( &c_local, BLIS_GEMM_UKR, cntx ) ) { bli_toggle_side( side ); bli_obj_induce_trans( b_local ); bli_obj_induce_trans( c_local ); } // Swap A and B if multiplying A from the right so that "B" contains // the symmetric matrix. if ( bli_is_right( side ) ) { bli_obj_swap( a_local, b_local ); } // Set the operation family id in the context. bli_cntx_set_family( BLIS_GEMM, cntx ); // Record the threading for each level within the context. bli_cntx_set_thrloop_from_env( BLIS_SYMM, BLIS_LEFT, cntx ); // Invoke the internal back-end. bli_l3_thread_decorator ( bli_gemm_int, alpha, &a_local, &b_local, beta, &c_local, cntx, cntl ); }
void bli_symm_front( side_t side, obj_t* alpha, obj_t* a, obj_t* b, obj_t* beta, obj_t* c, gemm_t* cntl ) { obj_t a_local; obj_t b_local; obj_t c_local; // Check parameters. if ( bli_error_checking_is_enabled() ) bli_symm_check( side, alpha, a, b, beta, c ); // If alpha is zero, scale by beta and return. if ( bli_obj_equals( alpha, &BLIS_ZERO ) ) { bli_scalm( beta, c ); return; } // Alias A, B, and C in case we need to apply transformations. bli_obj_alias_to( *a, a_local ); bli_obj_alias_to( *b, b_local ); bli_obj_alias_to( *c, c_local ); // An optimization: If C is stored by rows and the micro-kernel prefers // contiguous columns, or if C is stored by columns and the micro-kernel // prefers contiguous rows, transpose the entire operation to allow the // micro-kernel to access elements of C in its preferred manner. if ( ( bli_obj_is_row_stored( c_local ) && bli_func_prefers_contig_cols( bli_obj_datatype( c_local ), cntl_gemm_ukrs( cntl ) ) ) || ( bli_obj_is_col_stored( c_local ) && bli_func_prefers_contig_rows( bli_obj_datatype( c_local ), cntl_gemm_ukrs( cntl ) ) ) ) { bli_toggle_side( side ); bli_obj_induce_trans( b_local ); bli_obj_induce_trans( c_local ); } // Swap A and B if multiplying A from the right so that "B" contains // the symmetric matrix. if ( bli_is_right( side ) ) { bli_obj_swap( a_local, b_local ); } gemm_thrinfo_t** infos = bli_create_gemm_thrinfo_paths(); dim_t n_threads = thread_num_threads( infos[0] ); // Invoke the internal back-end. bli_level3_thread_decorator( n_threads, (level3_int_t) bli_gemm_int, alpha, &a_local, &b_local, beta, &c_local, (void*) cntl, (void**) infos ); bli_gemm_thrinfo_free_paths( infos, n_threads ); }