void libblis_test_symv_check ( test_params_t* params, obj_t* alpha, obj_t* a, obj_t* x, obj_t* beta, obj_t* y, obj_t* y_orig, double* resid ) { num_t dt = bli_obj_dt( y ); num_t dt_real = bli_obj_dt_proj_to_real( y ); dim_t m = bli_obj_vector_dim( y ); obj_t v; obj_t norm; double junk; // // Pre-conditions: // - a is randomized and symmetric. // - x is randomized. // - y_orig is randomized. // Note: // - alpha and beta should have non-zero imaginary components in the // complex cases in order to more fully exercise the implementation. // // Under these conditions, we assume that the implementation for // // y := beta * y_orig + alpha * conja(A) * conjx(x) // // is functioning correctly if // // normf( y - v ) // // is negligible, where // // v = beta * y_orig + alpha * conja(A_dense) * x // bli_obj_scalar_init_detached( dt_real, &norm ); bli_obj_create( dt, m, 1, 0, 0, &v ); bli_copyv( y_orig, &v ); bli_mksymm( a ); bli_obj_set_struc( BLIS_GENERAL, a ); bli_obj_set_uplo( BLIS_DENSE, a ); bli_gemv( alpha, a, x, beta, &v ); bli_subv( &v, y ); bli_normfv( y, &norm ); bli_getsc( &norm, resid, &junk ); bli_obj_free( &v ); }
void libblis_test_symv_experiment ( test_params_t* params, test_op_t* op, iface_t iface, num_t datatype, char* pc_str, char* sc_str, unsigned int p_cur, double* perf, double* resid ) { unsigned int n_repeats = params->n_repeats; unsigned int i; double time_min = DBL_MAX; double time; dim_t m; uplo_t uploa; conj_t conja; conj_t conjx; obj_t alpha, a, x, beta, y; obj_t y_save; // Map the dimension specifier to an actual dimension. m = libblis_test_get_dim_from_prob_size( op->dim_spec[0], p_cur ); // Map parameter characters to BLIS constants. bli_param_map_char_to_blis_uplo( pc_str[0], &uploa ); bli_param_map_char_to_blis_conj( pc_str[1], &conja ); bli_param_map_char_to_blis_conj( pc_str[2], &conjx ); // Create test scalars. bli_obj_scalar_init_detached( datatype, &alpha ); bli_obj_scalar_init_detached( datatype, &beta ); // Create test operands (vectors and/or matrices). libblis_test_mobj_create( params, datatype, BLIS_NO_TRANSPOSE, sc_str[0], m, m, &a ); libblis_test_vobj_create( params, datatype, sc_str[1], m, &x ); libblis_test_vobj_create( params, datatype, sc_str[2], m, &y ); libblis_test_vobj_create( params, datatype, sc_str[2], m, &y_save ); // Set alpha and beta. if ( bli_obj_is_real( &y ) ) { bli_setsc( 1.0, 0.0, &alpha ); bli_setsc( -1.0, 0.0, &beta ); } else { bli_setsc( 0.5, 0.5, &alpha ); bli_setsc( -0.5, 0.5, &beta ); } // Set the structure and uplo properties of A. bli_obj_set_struc( BLIS_SYMMETRIC, &a ); bli_obj_set_uplo( uploa, &a ); // Randomize A, make it densely symmetric, and zero the unstored triangle // to ensure the implementation reads only from the stored region. libblis_test_mobj_randomize( params, TRUE, &a ); bli_mksymm( &a ); bli_mktrim( &a ); // Randomize x and y, and save y. libblis_test_vobj_randomize( params, TRUE, &x ); libblis_test_vobj_randomize( params, TRUE, &y ); bli_copyv( &y, &y_save ); // Apply the remaining parameters. bli_obj_set_conj( conja, &a ); bli_obj_set_conj( conjx, &x ); // Repeat the experiment n_repeats times and record results. for ( i = 0; i < n_repeats; ++i ) { bli_copym( &y_save, &y ); time = bli_clock(); libblis_test_symv_impl( iface, &alpha, &a, &x, &beta, &y ); time_min = bli_clock_min_diff( time_min, time ); } // Estimate the performance of the best experiment repeat. *perf = ( 1.0 * m * m ) / time_min / FLOPS_PER_UNIT_PERF; if ( bli_obj_is_complex( &y ) ) *perf *= 4.0; // Perform checks. libblis_test_symv_check( params, &alpha, &a, &x, &beta, &y, &y_save, resid ); // Zero out performance and residual if output vector is empty. libblis_test_check_empty_problem( &y, perf, resid ); // Free the test objects. bli_obj_free( &a ); bli_obj_free( &x ); bli_obj_free( &y ); bli_obj_free( &y_save ); }
void libblis_test_syr2k_experiment( test_params_t* params, test_op_t* op, iface_t iface, num_t datatype, char* pc_str, char* sc_str, unsigned int p_cur, double* perf, double* resid ) { unsigned int n_repeats = params->n_repeats; unsigned int i; double time_min = 1e9; double time; dim_t m, k; uplo_t uploc; trans_t transa, transb; obj_t kappa; obj_t alpha, a, b, beta, c; obj_t c_save; // Map the dimension specifier to actual dimensions. m = libblis_test_get_dim_from_prob_size( op->dim_spec[0], p_cur ); k = libblis_test_get_dim_from_prob_size( op->dim_spec[1], p_cur ); // Map parameter characters to BLIS constants. bli_param_map_char_to_blis_uplo( pc_str[0], &uploc ); bli_param_map_char_to_blis_trans( pc_str[1], &transa ); bli_param_map_char_to_blis_trans( pc_str[2], &transb ); // Create test scalars. bli_obj_scalar_init_detached( datatype, &kappa ); bli_obj_scalar_init_detached( datatype, &alpha ); bli_obj_scalar_init_detached( datatype, &beta ); // Create test operands (vectors and/or matrices). libblis_test_mobj_create( params, datatype, transa, sc_str[0], m, k, &a ); libblis_test_mobj_create( params, datatype, transb, sc_str[1], m, k, &b ); libblis_test_mobj_create( params, datatype, BLIS_NO_TRANSPOSE, sc_str[2], m, m, &c ); libblis_test_mobj_create( params, datatype, BLIS_NO_TRANSPOSE, sc_str[2], m, m, &c_save ); // Set alpha and beta. if ( bli_obj_is_real( c ) ) { bli_setsc( 0.8, 0.0, &alpha ); bli_setsc( -1.0, 0.0, &beta ); } else { // For syr2k, both alpha and beta may be complex since, unlike her2k, // C is symmetric in both the real and complex cases. bli_setsc( 0.8, 0.5, &alpha ); bli_setsc( -1.0, 0.5, &beta ); } // Randomize A and B. bli_randm( &a ); bli_randm( &b ); // Set the structure and uplo properties of C. bli_obj_set_struc( BLIS_SYMMETRIC, c ); bli_obj_set_uplo( uploc, c ); // Randomize A, make it densely symmetric, and zero the unstored triangle // to ensure the implementation is reads only from the stored region. bli_randm( &c ); bli_mksymm( &c ); bli_mktrim( &c ); // Save C and set its structure and uplo properties. bli_obj_set_struc( BLIS_SYMMETRIC, c_save ); bli_obj_set_uplo( uploc, c_save ); bli_copym( &c, &c_save ); // Normalize by k. bli_setsc( 1.0/( double )k, 0.0, &kappa ); bli_scalm( &kappa, &a ); bli_scalm( &kappa, &b ); // Apply the remaining parameters. bli_obj_set_conjtrans( transa, a ); bli_obj_set_conjtrans( transb, b ); // Repeat the experiment n_repeats times and record results. for ( i = 0; i < n_repeats; ++i ) { bli_copym( &c_save, &c ); time = bli_clock(); libblis_test_syr2k_impl( iface, &alpha, &a, &b, &beta, &c ); time_min = bli_clock_min_diff( time_min, time ); } // Estimate the performance of the best experiment repeat. *perf = ( 2.0 * m * m * k ) / time_min / FLOPS_PER_UNIT_PERF; if ( bli_obj_is_complex( c ) ) *perf *= 4.0; // Perform checks. libblis_test_syr2k_check( &alpha, &a, &b, &beta, &c, &c_save, resid ); // Zero out performance and residual if output matrix is empty. libblis_test_check_empty_problem( &c, perf, resid ); // Free the test objects. bli_obj_free( &a ); bli_obj_free( &b ); bli_obj_free( &c ); bli_obj_free( &c_save ); }
void libblis_test_syr2_check( obj_t* alpha, obj_t* x, obj_t* y, obj_t* a, obj_t* a_orig, double* resid ) { num_t dt = bli_obj_datatype( *a ); num_t dt_real = bli_obj_datatype_proj_to_real( *a ); dim_t m_a = bli_obj_length( *a ); obj_t xt, yt; obj_t t, v, w1, w2; obj_t tau, rho, norm; double junk; // // Pre-conditions: // - x is randomized. // - y is randomized. // - a is randomized and symmetric. // Note: // - alpha should have a non-zero imaginary component in the // complex cases in order to more fully exercise the implementation. // // Under these conditions, we assume that the implementation for // // A := A_orig + alpha * conjx(x) * conjy(y)^T + alpha * conjy(y) * conjx(x)^T // // is functioning correctly if // // normf( v - w ) // // is negligible, where // // v = A * t // w = ( A_orig + alpha * conjx(x) * conjy(y)^T + alpha * conjy(y) * conjx(x)^T ) * t // = A_orig * t + alpha * conjx(x) * conjy(y)^T * t + alpha * conjy(y) * conjx(x)^T * t // = A_orig * t + alpha * conjx(x) * conjy(y)^T * t + alpha * conjy(y) * rho // = A_orig * t + alpha * conjx(x) * conjy(y)^T * t + w1 // = A_orig * t + alpha * conjx(x) * rho + w1 // = A_orig * t + w2 + w1 // bli_mksymm( a ); bli_mksymm( a_orig ); bli_obj_set_struc( BLIS_GENERAL, *a ); bli_obj_set_struc( BLIS_GENERAL, *a_orig ); bli_obj_set_uplo( BLIS_DENSE, *a ); bli_obj_set_uplo( BLIS_DENSE, *a_orig ); bli_obj_scalar_init_detached( dt, &tau ); bli_obj_scalar_init_detached( dt, &rho ); bli_obj_scalar_init_detached( dt_real, &norm ); bli_obj_create( dt, m_a, 1, 0, 0, &t ); bli_obj_create( dt, m_a, 1, 0, 0, &v ); bli_obj_create( dt, m_a, 1, 0, 0, &w1 ); bli_obj_create( dt, m_a, 1, 0, 0, &w2 ); bli_obj_alias_to( *x, xt ); bli_obj_alias_to( *y, yt ); bli_setsc( 1.0/( double )m_a, -1.0/( double )m_a, &tau ); bli_setv( &tau, &t ); bli_gemv( &BLIS_ONE, a, &t, &BLIS_ZERO, &v ); bli_dotv( &xt, &t, &rho ); bli_mulsc( alpha, &rho ); bli_scal2v( &rho, y, &w1 ); bli_dotv( &yt, &t, &rho ); bli_mulsc( alpha, &rho ); bli_scal2v( &rho, x, &w2 ); bli_addv( &w2, &w1 ); bli_gemv( &BLIS_ONE, a_orig, &t, &BLIS_ONE, &w1 ); bli_subv( &w1, &v ); bli_normfv( &v, &norm ); bli_getsc( &norm, resid, &junk ); bli_obj_free( &t ); bli_obj_free( &v ); bli_obj_free( &w1 ); bli_obj_free( &w2 ); }
void libblis_test_syr_check ( test_params_t* params, obj_t* alpha, obj_t* x, obj_t* a, obj_t* a_orig, double* resid ) { num_t dt = bli_obj_dt( a ); num_t dt_real = bli_obj_dt_proj_to_real( a ); dim_t m_a = bli_obj_length( a ); obj_t xt, t, v, w; obj_t rho, norm; double junk; // // Pre-conditions: // - x is randomized. // - a is randomized and symmetric. // Note: // - alpha should have a non-zero imaginary component in the // complex cases in order to more fully exercise the implementation. // // Under these conditions, we assume that the implementation for // // A := A_orig + alpha * conjx(x) * conjx(x)^T // // is functioning correctly if // // normf( v - w ) // // is negligible, where // // v = A * t // w = ( A_orig + alpha * conjx(x) * conjx(x)^T ) * t // = A_orig * t + alpha * conjx(x) * conjx(x)^T * t // = A_orig * t + alpha * conjx(x) * rho // = A_orig * t + w // bli_mksymm( a ); bli_mksymm( a_orig ); bli_obj_set_struc( BLIS_GENERAL, a ); bli_obj_set_struc( BLIS_GENERAL, a_orig ); bli_obj_set_uplo( BLIS_DENSE, a ); bli_obj_set_uplo( BLIS_DENSE, a_orig ); bli_obj_scalar_init_detached( dt, &rho ); bli_obj_scalar_init_detached( dt_real, &norm ); bli_obj_create( dt, m_a, 1, 0, 0, &t ); bli_obj_create( dt, m_a, 1, 0, 0, &v ); bli_obj_create( dt, m_a, 1, 0, 0, &w ); bli_obj_alias_to( x, &xt ); libblis_test_vobj_randomize( params, TRUE, &t ); bli_gemv( &BLIS_ONE, a, &t, &BLIS_ZERO, &v ); bli_dotv( &xt, &t, &rho ); bli_mulsc( alpha, &rho ); bli_scal2v( &rho, x, &w ); bli_gemv( &BLIS_ONE, a_orig, &t, &BLIS_ONE, &w ); bli_subv( &w, &v ); bli_normfv( &v, &norm ); bli_getsc( &norm, resid, &junk ); bli_obj_free( &t ); bli_obj_free( &v ); bli_obj_free( &w ); }