void bli_gemm_int( obj_t* alpha, obj_t* a, obj_t* b, obj_t* beta, obj_t* c, gemm_t* cntl, gemm_thrinfo_t* thread ) { obj_t a_local; obj_t b_local; obj_t c_local; varnum_t n; impl_t i; FUNCPTR_T f; // Check parameters. if ( bli_error_checking_is_enabled() ) bli_gemm_int_check( alpha, a, b, beta, c, cntl ); // If C has a zero dimension, return early. if ( bli_obj_has_zero_dim( *c ) ) return; // If A or B has a zero dimension, scale C by beta and return early. if ( bli_obj_has_zero_dim( *a ) || bli_obj_has_zero_dim( *b ) ) { if( thread_am_ochief( thread ) ) bli_scalm( beta, c ); thread_obarrier( thread ); return; } // If A or B is marked as being filled with zeros, scale C by beta and // return early. if ( bli_obj_is_zeros( *a ) || bli_obj_is_zeros( *b ) ) { if( thread_am_ochief( thread ) ) bli_scalm( beta, c ); thread_obarrier( thread ); return; } // Alias A and B in case we need to update attached scalars. bli_obj_alias_to( *a, a_local ); bli_obj_alias_to( *b, b_local ); // Alias C in case we need to induce a transposition. bli_obj_alias_to( *c, c_local ); // If we are about to call a leaf-level implementation, and matrix C // still needs a transposition, then we must induce one by swapping the // strides and dimensions. Note that this transposition would normally // be handled explicitly in the packing of C, but if C is not being // packed, this is our last chance to handle the transposition. if ( cntl_is_leaf( cntl ) && bli_obj_has_trans( *c ) ) { //if( thread_am_ochief( thread ) ) { bli_obj_induce_trans( c_local ); bli_obj_set_onlytrans( BLIS_NO_TRANSPOSE, c_local ); // } } // If alpha is non-unit, typecast and apply it to the scalar attached // to B. if ( !bli_obj_equals( alpha, &BLIS_ONE ) ) { bli_obj_scalar_apply_scalar( alpha, &b_local ); } // If beta is non-unit, typecast and apply it to the scalar attached // to C. if ( !bli_obj_equals( beta, &BLIS_ONE ) ) { bli_obj_scalar_apply_scalar( beta, &c_local ); } // Extract the variant number and implementation type. n = cntl_var_num( cntl ); i = cntl_impl_type( cntl ); // Index into the variant array to extract the correct function pointer. f = vars[n][i]; // Invoke the variant. f( &a_local, &b_local, &c_local, cntl, thread ); }
void bli_trmm_int( obj_t* alpha, obj_t* a, obj_t* b, obj_t* beta, obj_t* c, trmm_t* cntl ) { obj_t a_local; obj_t b_local; obj_t c_local; bool_t side, uplo; varnum_t n; impl_t i; FUNCPTR_T f; // Check parameters. if ( bli_error_checking_is_enabled() ) bli_trmm_int_check( alpha, a, b, beta, c, cntl ); // If C has a zero dimension, return early. if ( bli_obj_has_zero_dim( *c ) ) return; // If A or B has a zero dimension, scale C by beta and return early. if ( bli_obj_has_zero_dim( *a ) || bli_obj_has_zero_dim( *b ) ) { bli_scalm( beta, c ); return; } // Alias A and B in case we need to update attached scalars. bli_obj_alias_to( *a, a_local ); bli_obj_alias_to( *b, b_local ); // Alias C in case we need to induce a transposition. bli_obj_alias_to( *c, c_local ); // If we are about to call a leaf-level implementation, and matrix C // still needs a transposition, then we must induce one by swapping the // strides and dimensions. Note that this transposition would normally // be handled explicitly in the packing of C, but if C is not being // packed, this is our last chance to handle the transposition. if ( cntl_is_leaf( cntl ) && bli_obj_has_trans( *c ) ) { bli_obj_induce_trans( c_local ); bli_obj_set_onlytrans( BLIS_NO_TRANSPOSE, c_local ); } // If alpha is non-unit, typecast and apply it to the scalar attached // to B. if ( !bli_obj_equals( alpha, &BLIS_ONE ) ) { bli_obj_scalar_apply_scalar( alpha, &b_local ); } // If beta is non-unit, typecast and apply it to the scalar attached // to C. if ( !bli_obj_equals( beta, &BLIS_ONE ) ) { bli_obj_scalar_apply_scalar( beta, &c_local ); } // Set two bools: one based on the implied side parameter (the structure // of the root object) and one based on the uplo field of the triangular // matrix's root object (whether that is matrix A or matrix B). if ( bli_obj_root_is_triangular( *a ) ) { side = 0; if ( bli_obj_root_is_lower( *a ) ) uplo = 0; else uplo = 1; } else // if ( bli_obj_root_is_triangular( *b ) ) { side = 1; // Set a bool based on the uplo field of A's root object. if ( bli_obj_root_is_lower( *b ) ) uplo = 0; else uplo = 1; } // Extract the variant number and implementation type. n = cntl_var_num( cntl ); i = cntl_impl_type( cntl ); // Index into the variant array to extract the correct function pointer. f = vars[side][uplo][n][i]; // Invoke the variant. f( &a_local, &b_local, &c_local, cntl ); }
void bli_ger_int( conj_t conjx, conj_t conjy, obj_t* alpha, obj_t* x, obj_t* y, obj_t* a, ger_t* cntl ) { varnum_t n; impl_t i; FUNCPTR_T f; obj_t alpha_local; obj_t x_local; obj_t y_local; obj_t a_local; // Check parameters. if ( bli_error_checking_is_enabled() ) bli_ger_int_check( alpha, x, y, a, cntl ); // If A has a zero dimension, return early. if ( bli_obj_has_zero_dim( *a ) ) return; // If x or y has a zero dimension, return early. if ( bli_obj_has_zero_dim( *x ) || bli_obj_has_zero_dim( *y ) ) return; // Alias the objects, applying conjx and conjy to x and y, respectively. bli_obj_alias_with_conj( conjx, *x, x_local ); bli_obj_alias_with_conj( conjy, *y, y_local ); bli_obj_alias_to( *a, a_local ); // If matrix A is marked for conjugation, we interpret this as a request // to apply a conjugation to the other operands. if ( bli_obj_has_conj( a_local ) ) { bli_obj_toggle_conj( a_local ); bli_obj_toggle_conj( x_local ); bli_obj_toggle_conj( y_local ); bli_obj_scalar_init_detached_copy_of( bli_obj_datatype( *alpha ), BLIS_CONJUGATE, alpha, &alpha_local ); } else { bli_obj_alias_to( *alpha, alpha_local ); } // If we are about the call a leaf-level implementation, and matrix A // still needs a transposition, then we must induce one by swapping the // strides and dimensions. if ( cntl_is_leaf( cntl ) && bli_obj_has_trans( a_local ) ) { bli_obj_induce_trans( a_local ); bli_obj_set_onlytrans( BLIS_NO_TRANSPOSE, a_local ); } // Extract the variant number and implementation type. n = cntl_var_num( cntl ); i = cntl_impl_type( cntl ); // Index into the variant array to extract the correct function pointer. f = vars[n][i]; // Invoke the variant. f( &alpha_local, &x_local, &y_local, &a_local, cntl ); }