void libblis_test_trmv_experiment( test_params_t* params, test_op_t* op, mt_impl_t impl, 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; uplo_t uploa; trans_t transa; diag_t diaga; obj_t kappa; obj_t alpha, a, x; obj_t x_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_trans( pc_str[1], &transa ); bli_param_map_char_to_blis_diag( pc_str[2], &diaga ); // Create test scalars. bli_obj_init_scalar( datatype, &alpha ); bli_obj_init_scalar( datatype, &kappa ); // 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[1], m, &x_save ); // Set alpha. if ( bli_obj_is_real( x ) ) bli_setsc( -1.0, 0.0, &alpha ); else bli_setsc( 0.0, -1.0, &alpha ); // Set the structure and uplo properties of A. bli_obj_set_struc( BLIS_TRIANGULAR, a ); bli_obj_set_uplo( uploa, a ); // Randomize A, make it densely triangular. bli_randm( &a ); bli_mktrim( &a ); // Randomize x and save. bli_randv( &x ); bli_copyv( &x, &x_save ); // Normalize vectors by m. bli_setsc( 1.0/( double )m, 0.0, &kappa ); bli_scalv( &kappa, &x ); bli_scalv( &kappa, &x_save ); // Apply the remaining parameters. bli_obj_set_conjtrans( transa, a ); bli_obj_set_diag( diaga, a ); // Repeat the experiment n_repeats times and record results. for ( i = 0; i < n_repeats; ++i ) { bli_copym( &x_save, &x ); time = bli_clock(); libblis_test_trmv_impl( impl, &alpha, &a, &x ); 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( x ) ) *perf *= 4.0; // Perform checks. libblis_test_trmv_check( &alpha, &a, &x, &x_save, resid ); // Zero out performance and residual if output vector is empty. libblis_test_check_empty_problem( &x, perf, resid ); // Free the test objects. bli_obj_free( &a ); bli_obj_free( &x ); bli_obj_free( &x_save ); }
void libblis_test_trsm_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, n; dim_t mn_side; side_t side; uplo_t uploa; trans_t transa; diag_t diaga; obj_t alpha, a, b; obj_t b_save; // Map the dimension specifier to actual dimensions. m = libblis_test_get_dim_from_prob_size( op->dim_spec[0], p_cur ); n = 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_side( pc_str[0], &side ); bli_param_map_char_to_blis_uplo( pc_str[1], &uploa ); bli_param_map_char_to_blis_trans( pc_str[2], &transa ); bli_param_map_char_to_blis_diag( pc_str[3], &diaga ); // Create test scalars. bli_obj_scalar_init_detached( datatype, &alpha ); // Create test operands (vectors and/or matrices). bli_set_dim_with_side( side, m, n, mn_side ); libblis_test_mobj_create( params, datatype, transa, sc_str[0], mn_side, mn_side, &a ); libblis_test_mobj_create( params, datatype, BLIS_NO_TRANSPOSE, sc_str[1], m, n, &b ); libblis_test_mobj_create( params, datatype, BLIS_NO_TRANSPOSE, sc_str[1], m, n, &b_save ); // Set alpha. if ( bli_obj_is_real( b ) ) { bli_setsc( 2.0, 0.0, &alpha ); } else { bli_setsc( 2.0, 0.0, &alpha ); } // Set the structure and uplo properties of A. bli_obj_set_struc( BLIS_TRIANGULAR, a ); bli_obj_set_uplo( uploa, a ); // Randomize A, load the diagonal, make it densely triangular. libblis_test_mobj_randomize( params, TRUE, &a ); libblis_test_mobj_load_diag( params, &a ); bli_mktrim( &a ); // Randomize B and save B. libblis_test_mobj_randomize( params, TRUE, &b ); bli_copym( &b, &b_save ); // Apply the remaining parameters. bli_obj_set_conjtrans( transa, a ); bli_obj_set_diag( diaga, a ); // Repeat the experiment n_repeats times and record results. for ( i = 0; i < n_repeats; ++i ) { bli_copym( &b_save, &b ); time = bli_clock(); libblis_test_trsm_impl( iface, side, &alpha, &a, &b ); time_min = bli_clock_min_diff( time_min, time ); } // Estimate the performance of the best experiment repeat. *perf = ( 1.0 * mn_side * m * n ) / time_min / FLOPS_PER_UNIT_PERF; if ( bli_obj_is_complex( b ) ) *perf *= 4.0; // Perform checks. libblis_test_trsm_check( params, side, &alpha, &a, &b, &b_save, resid ); // Zero out performance and residual if output matrix is empty. libblis_test_check_empty_problem( &b, perf, resid ); // Free the test objects. bli_obj_free( &a ); bli_obj_free( &b ); bli_obj_free( &b_save ); }
int main( int argc, char** argv ) { obj_t a, x; obj_t x_save; obj_t alpha; dim_t m; dim_t p; dim_t p_begin, p_end, p_inc; int m_input; num_t dt_a, dt_x; num_t dt_alpha; int r, n_repeats; uplo_t uplo; double dtime; double dtime_save; double gflops; //bli_init(); n_repeats = 3; #ifndef PRINT p_begin = 40; p_end = 2000; p_inc = 40; m_input = -1; #else p_begin = 16; p_end = 16; p_inc = 1; m_input = 15; n_input = 15; #endif dt_alpha = dt_a = dt_x = BLIS_DOUBLE; uplo = BLIS_LOWER; // Begin with initializing the last entry to zero so that // matlab allocates space for the entire array once up-front. for ( p = p_begin; p + p_inc <= p_end; p += p_inc ) ; #ifdef BLIS printf( "data_trsv_blis" ); #else printf( "data_trv_%s", BLAS ); #endif printf( "( %2lu, 1:2 ) = [ %4lu %7.2f ];\n", ( unsigned long )(p - p_begin + 1)/p_inc + 1, ( unsigned long )0, 0.0 ); for ( p = p_begin; p <= p_end; p += p_inc ) { if ( m_input < 0 ) m = p * ( dim_t )abs(m_input); else m = ( dim_t ) m_input; bli_obj_create( dt_alpha, 1, 1, 0, 0, &alpha ); bli_obj_create( dt_a, m, m, 0, 0, &a ); bli_obj_create( dt_x, m, 1, 0, 0, &x ); bli_obj_create( dt_x, m, 1, 0, 0, &x_save ); bli_randm( &a ); bli_randm( &x ); bli_obj_set_struc( BLIS_TRIANGULAR, &a ); bli_obj_set_uplo( uplo, &a ); bli_obj_set_onlytrans( BLIS_NO_TRANSPOSE, &a ); bli_obj_set_diag( BLIS_NONUNIT_DIAG, &a ); // Randomize A and zero the unstored triangle to ensure the // implementation reads only from the stored region. bli_randm( &a ); bli_mktrim( &a ); // Load the diagonal of A to make it more likely to be invertible. bli_shiftd( &BLIS_TWO, &a ); bli_setsc( (1.0/1.0), 0.0, &alpha ); bli_copym( &x, &x_save ); dtime_save = DBL_MAX; for ( r = 0; r < n_repeats; ++r ) { bli_copym( &x_save, &x ); dtime = bli_clock(); #ifdef PRINT bli_printm( "a", &a, "%4.1f", "" ); bli_printm( "x", &x, "%4.1f", "" ); #endif #ifdef BLIS bli_trsv( &BLIS_ONE, &a, &x ); #else f77_char uploa = 'L'; f77_char transa = 'N'; f77_char diaga = 'N'; f77_int mm = bli_obj_length( &a ); f77_int lda = bli_obj_col_stride( &a ); f77_int incx = bli_obj_vector_inc( &x ); double* ap = bli_obj_buffer( &a ); double* xp = bli_obj_buffer( &x ); dtrsv_( &uploa, &transa, &diaga, &mm, ap, &lda, xp, &incx ); #endif #ifdef PRINT bli_printm( "x after", &x, "%4.1f", "" ); exit(1); #endif dtime_save = bli_clock_min_diff( dtime_save, dtime ); } gflops = ( 1.0 * m * m ) / ( dtime_save * 1.0e9 ); #ifdef BLIS printf( "data_trsv_blis" ); #else printf( "data_trsv_%s", BLAS ); #endif printf( "( %2lu, 1:2 ) = [ %4lu %7.2f ];\n", ( unsigned long )(p - p_begin + 1)/p_inc + 1, ( unsigned long )m, gflops ); bli_obj_free( &alpha ); bli_obj_free( &a ); bli_obj_free( &x ); bli_obj_free( &x_save ); } //bli_finalize(); return 0; }
int main( int argc, char** argv ) { obj_t a, c; obj_t c_save; obj_t alpha; dim_t m, n; dim_t p; dim_t p_begin, p_max, p_inc; int m_input, n_input; ind_t ind; num_t dt; char dt_ch; int r, n_repeats; side_t side; uplo_t uploa; trans_t transa; diag_t diaga; f77_char f77_side; f77_char f77_uploa; f77_char f77_transa; f77_char f77_diaga; double dtime; double dtime_save; double gflops; //bli_init(); //bli_error_checking_level_set( BLIS_NO_ERROR_CHECKING ); n_repeats = 3; dt = DT; ind = IND; p_begin = P_BEGIN; p_max = P_MAX; p_inc = P_INC; m_input = -1; n_input = -1; // Supress compiler warnings about unused variable 'ind'. ( void )ind; #if 0 cntx_t* cntx; ind_t ind_mod = ind; // A hack to use 3m1 as 1mpb (with 1m as 1mbp). if ( ind == BLIS_3M1 ) ind_mod = BLIS_1M; // Initialize a context for the current induced method and datatype. cntx = bli_gks_query_ind_cntx( ind_mod, dt ); // Set k to the kc blocksize for the current datatype. k_input = bli_cntx_get_blksz_def_dt( dt, BLIS_KC, cntx ); #elif 1 //k_input = 256; #endif // Choose the char corresponding to the requested datatype. if ( bli_is_float( dt ) ) dt_ch = 's'; else if ( bli_is_double( dt ) ) dt_ch = 'd'; else if ( bli_is_scomplex( dt ) ) dt_ch = 'c'; else dt_ch = 'z'; #if 0 side = BLIS_LEFT; #else side = BLIS_RIGHT; #endif #if 0 uploa = BLIS_LOWER; #else uploa = BLIS_UPPER; #endif transa = BLIS_NO_TRANSPOSE; diaga = BLIS_NONUNIT_DIAG; bli_param_map_blis_to_netlib_side( side, &f77_side ); bli_param_map_blis_to_netlib_uplo( uploa, &f77_uploa ); bli_param_map_blis_to_netlib_trans( transa, &f77_transa ); bli_param_map_blis_to_netlib_diag( diaga, &f77_diaga ); // Begin with initializing the last entry to zero so that // matlab allocates space for the entire array once up-front. for ( p = p_begin; p + p_inc <= p_max; p += p_inc ) ; printf( "data_%s_%ctrsm_%s", THR_STR, dt_ch, STR ); printf( "( %2lu, 1:3 ) = [ %4lu %4lu %7.2f ];\n", ( unsigned long )(p - p_begin + 1)/p_inc + 1, ( unsigned long )0, ( unsigned long )0, 0.0 ); for ( p = p_begin; p <= p_max; p += p_inc ) { if ( m_input < 0 ) m = p / ( dim_t )abs(m_input); else m = ( dim_t ) m_input; if ( n_input < 0 ) n = p / ( dim_t )abs(n_input); else n = ( dim_t ) n_input; bli_obj_create( dt, 1, 1, 0, 0, &alpha ); if ( bli_is_left( side ) ) bli_obj_create( dt, m, m, 0, 0, &a ); else bli_obj_create( dt, n, n, 0, 0, &a ); bli_obj_create( dt, m, n, 0, 0, &c ); //bli_obj_create( dt, m, n, n, 1, &c ); bli_obj_create( dt, m, n, 0, 0, &c_save ); bli_randm( &a ); bli_randm( &c ); bli_obj_set_struc( BLIS_TRIANGULAR, &a ); bli_obj_set_uplo( uploa, &a ); bli_obj_set_conjtrans( transa, &a ); bli_obj_set_diag( diaga, &a ); bli_randm( &a ); bli_mktrim( &a ); // Load the diagonal of A to make it more likely to be invertible. bli_shiftd( &BLIS_TWO, &a ); bli_setsc( (2.0/1.0), 0.0, &alpha ); bli_copym( &c, &c_save ); #if 0 //def BLIS bli_ind_disable_all_dt( dt ); bli_ind_enable_dt( ind, dt ); #endif dtime_save = DBL_MAX; for ( r = 0; r < n_repeats; ++r ) { bli_copym( &c_save, &c ); dtime = bli_clock(); #ifdef PRINT bli_printm( "a", &a, "%4.1f", "" ); bli_printm( "c", &c, "%4.1f", "" ); #endif #ifdef BLIS bli_trsm( side, &alpha, &a, &c ); #else if ( bli_is_float( dt ) ) { f77_int mm = bli_obj_length( &c ); f77_int kk = bli_obj_width( &c ); f77_int lda = bli_obj_col_stride( &a ); f77_int ldc = bli_obj_col_stride( &c ); float* alphap = ( float* )bli_obj_buffer( &alpha ); float* ap = ( float* )bli_obj_buffer( &a ); float* cp = ( float* )bli_obj_buffer( &c ); strsm_( &f77_side, &f77_uploa, &f77_transa, &f77_diaga, &mm, &kk, alphap, ap, &lda, cp, &ldc ); } else if ( bli_is_double( dt ) ) { f77_int mm = bli_obj_length( &c ); f77_int kk = bli_obj_width( &c ); f77_int lda = bli_obj_col_stride( &a ); f77_int ldc = bli_obj_col_stride( &c ); double* alphap = ( double* )bli_obj_buffer( &alpha ); double* ap = ( double* )bli_obj_buffer( &a ); double* cp = ( double* )bli_obj_buffer( &c ); dtrsm_( &f77_side, &f77_uploa, &f77_transa, &f77_diaga, &mm, &kk, alphap, ap, &lda, cp, &ldc ); } else if ( bli_is_scomplex( dt ) ) { f77_int mm = bli_obj_length( &c ); f77_int kk = bli_obj_width( &c ); f77_int lda = bli_obj_col_stride( &a ); f77_int ldc = bli_obj_col_stride( &c ); #ifdef EIGEN float* alphap = ( float* )bli_obj_buffer( &alpha ); float* ap = ( float* )bli_obj_buffer( &a ); float* cp = ( float* )bli_obj_buffer( &c ); #else scomplex* alphap = ( scomplex* )bli_obj_buffer( &alpha ); scomplex* ap = ( scomplex* )bli_obj_buffer( &a ); scomplex* cp = ( scomplex* )bli_obj_buffer( &c ); #endif ctrsm_( &f77_side, &f77_uploa, &f77_transa, &f77_diaga, &mm, &kk, alphap, ap, &lda, cp, &ldc ); } else if ( bli_is_dcomplex( dt ) ) { f77_int mm = bli_obj_length( &c ); f77_int kk = bli_obj_width( &c ); f77_int lda = bli_obj_col_stride( &a ); f77_int ldc = bli_obj_col_stride( &c ); #ifdef EIGEN double* alphap = ( double* )bli_obj_buffer( &alpha ); double* ap = ( double* )bli_obj_buffer( &a ); double* cp = ( double* )bli_obj_buffer( &c ); #else dcomplex* alphap = ( dcomplex* )bli_obj_buffer( &alpha ); dcomplex* ap = ( dcomplex* )bli_obj_buffer( &a ); dcomplex* cp = ( dcomplex* )bli_obj_buffer( &c ); #endif ztrsm_( &f77_side, &f77_uploa, &f77_transa, &f77_diaga, &mm, &kk, alphap, ap, &lda, cp, &ldc ); } #endif #ifdef PRINT bli_printm( "c after", &c, "%4.1f", "" ); exit(1); #endif dtime_save = bli_clock_min_diff( dtime_save, dtime ); } if ( bli_is_left( side ) ) gflops = ( 1.0 * m * m * n ) / ( dtime_save * 1.0e9 ); else gflops = ( 1.0 * m * n * n ) / ( dtime_save * 1.0e9 ); if ( bli_is_complex( dt ) ) gflops *= 4.0; printf( "data_%s_%ctrsm_%s", THR_STR, dt_ch, STR ); printf( "( %2lu, 1:3 ) = [ %4lu %4lu %7.2f ];\n", ( unsigned long )(p - p_begin + 1)/p_inc + 1, ( unsigned long )m, ( unsigned long )n, gflops ); bli_obj_free( &alpha ); bli_obj_free( &a ); bli_obj_free( &c ); bli_obj_free( &c_save ); } //bli_finalize(); return 0; }
int main( int argc, char** argv ) { obj_t a, c; obj_t c_save; obj_t alpha; dim_t m, n; dim_t p; dim_t p_begin, p_end, p_inc; int m_input, n_input; num_t dt; int r, n_repeats; side_t side; uplo_t uploa; trans_t transa; diag_t diaga; f77_char f77_side; f77_char f77_uploa; f77_char f77_transa; f77_char f77_diaga; double dtime; double dtime_save; double gflops; //bli_init(); //bli_error_checking_level_set( BLIS_NO_ERROR_CHECKING ); n_repeats = 3; #ifndef PRINT p_begin = 200; p_end = 2000; p_inc = 200; m_input = -1; n_input = -1; #else p_begin = 16; p_end = 16; p_inc = 1; m_input = 4; n_input = 4; #endif #if 1 //dt = BLIS_FLOAT; dt = BLIS_DOUBLE; #else //dt = BLIS_SCOMPLEX; dt = BLIS_DCOMPLEX; #endif side = BLIS_LEFT; //side = BLIS_RIGHT; uploa = BLIS_LOWER; //uploa = BLIS_UPPER; transa = BLIS_NO_TRANSPOSE; diaga = BLIS_NONUNIT_DIAG; bli_param_map_blis_to_netlib_side( side, &f77_side ); bli_param_map_blis_to_netlib_uplo( uploa, &f77_uploa ); bli_param_map_blis_to_netlib_trans( transa, &f77_transa ); bli_param_map_blis_to_netlib_diag( diaga, &f77_diaga ); for ( p = p_begin; p <= p_end; p += p_inc ) { if ( m_input < 0 ) m = p * ( dim_t )abs(m_input); else m = ( dim_t ) m_input; if ( n_input < 0 ) n = p * ( dim_t )abs(n_input); else n = ( dim_t ) n_input; bli_obj_create( dt, 1, 1, 0, 0, &alpha ); if ( bli_is_left( side ) ) bli_obj_create( dt, m, m, 0, 0, &a ); else bli_obj_create( dt, n, n, 0, 0, &a ); bli_obj_create( dt, m, n, 0, 0, &c ); bli_obj_create( dt, m, n, 0, 0, &c_save ); bli_randm( &a ); bli_randm( &c ); bli_obj_set_struc( BLIS_TRIANGULAR, &a ); bli_obj_set_uplo( uploa, &a ); bli_obj_set_conjtrans( transa, &a ); bli_obj_set_diag( diaga, &a ); // Randomize A, make it densely Hermitian, and zero the unstored // triangle to ensure the implementation reads only from the stored // region. bli_randm( &a ); bli_mkherm( &a ); bli_mktrim( &a ); bli_setsc( (2.0/1.0), 1.0, &alpha ); bli_copym( &c, &c_save ); dtime_save = DBL_MAX; for ( r = 0; r < n_repeats; ++r ) { bli_copym( &c_save, &c ); dtime = bli_clock(); #ifdef PRINT bli_invertd( &a ); bli_printm( "a", &a, "%4.1f", "" ); bli_invertd( &a ); bli_printm( "c", &c, "%4.1f", "" ); #endif #ifdef BLIS bli_trsm( side, &alpha, &a, &c ); #else if ( bli_is_float( dt ) ) { f77_int mm = bli_obj_length( &c ); f77_int nn = bli_obj_width( &c ); f77_int lda = bli_obj_col_stride( &a ); f77_int ldc = bli_obj_col_stride( &c ); float* alphap = bli_obj_buffer( &alpha ); float* ap = bli_obj_buffer( &a ); float* cp = bli_obj_buffer( &c ); strsm_( &f77_side, &f77_uploa, &f77_transa, &f77_diaga, &mm, &nn, alphap, ap, &lda, cp, &ldc ); } else if ( bli_is_double( dt ) ) { f77_int mm = bli_obj_length( &c ); f77_int nn = bli_obj_width( &c ); f77_int lda = bli_obj_col_stride( &a ); f77_int ldc = bli_obj_col_stride( &c ); double* alphap = bli_obj_buffer( &alpha ); double* ap = bli_obj_buffer( &a ); double* cp = bli_obj_buffer( &c ); dtrsm_( &f77_side, &f77_uploa, &f77_transa, &f77_diaga, &mm, &nn, alphap, ap, &lda, cp, &ldc ); } else if ( bli_is_scomplex( dt ) ) { f77_int mm = bli_obj_length( &c ); f77_int nn = bli_obj_width( &c ); f77_int lda = bli_obj_col_stride( &a ); f77_int ldc = bli_obj_col_stride( &c ); scomplex* alphap = bli_obj_buffer( &alpha ); scomplex* ap = bli_obj_buffer( &a ); scomplex* cp = bli_obj_buffer( &c ); ctrsm_( &f77_side, &f77_uploa, &f77_transa, &f77_diaga, &mm, &nn, alphap, ap, &lda, cp, &ldc ); } else if ( bli_is_dcomplex( dt ) ) { f77_int mm = bli_obj_length( &c ); f77_int nn = bli_obj_width( &c ); f77_int lda = bli_obj_col_stride( &a ); f77_int ldc = bli_obj_col_stride( &c ); dcomplex* alphap = bli_obj_buffer( &alpha ); dcomplex* ap = bli_obj_buffer( &a ); dcomplex* cp = bli_obj_buffer( &c ); ztrsm_( &f77_side, &f77_uploa, &f77_transa, &f77_diaga, &mm, &nn, alphap, ap, &lda, cp, &ldc ); } #endif #ifdef PRINT bli_printm( "c after", &c, "%9.5f", "" ); exit(1); #endif dtime_save = bli_clock_min_diff( dtime_save, dtime ); } if ( bli_is_left( side ) ) gflops = ( 1.0 * m * m * n ) / ( dtime_save * 1.0e9 ); else gflops = ( 1.0 * m * n * n ) / ( dtime_save * 1.0e9 ); if ( bli_is_complex( dt ) ) gflops *= 4.0; #ifdef BLIS printf( "data_trsm_blis" ); #else printf( "data_trsm_%s", BLAS ); #endif printf( "( %2lu, 1:4 ) = [ %4lu %4lu %10.3e %6.3f ];\n", ( unsigned long )(p - p_begin + 1)/p_inc + 1, ( unsigned long )m, ( unsigned long )n, dtime_save, gflops ); bli_obj_free( &alpha ); bli_obj_free( &a ); bli_obj_free( &c ); bli_obj_free( &c_save ); } //bli_finalize(); return 0; }