void libblis_test_trmv_check( obj_t* alpha,
obj_t* a,
obj_t* x,
obj_t* x_orig,
double* resid )
{
num_t dt = bli_obj_datatype( *x );
num_t dt_real = bli_obj_datatype_proj_to_real( *x );
dim_t m = bli_obj_vector_dim( *x );
uplo_t uploa = bli_obj_uplo( *a );
trans_t transa = bli_obj_conjtrans_status( *a );
obj_t a_local, y;
obj_t norm;
double junk;
//
// Pre-conditions:
// - a is randomized and triangular.
// - x is randomized.
// 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
//
// x := alpha * transa(A) * x_orig
//
// is functioning correctly if
//
// fnorm( y - x )
//
// is negligible, where
//
// y = alpha * conja(A_dense) * x_orig
//
bli_obj_init_scalar( dt_real, &norm );
bli_obj_create( dt, m, 1, 0, 0, &y );
bli_obj_create( dt, m, m, 0, 0, &a_local );
bli_obj_set_struc( BLIS_TRIANGULAR, a_local );
bli_obj_set_uplo( uploa, a_local );
bli_obj_toggle_uplo_if_trans( transa, a_local );
bli_copym( a, &a_local );
bli_mktrim( &a_local );
bli_obj_set_struc( BLIS_GENERAL, a_local );
bli_obj_set_uplo( BLIS_DENSE, a_local );
bli_gemv( alpha, &a_local, x_orig, &BLIS_ZERO, &y );
bli_subv( x, &y );
bli_fnormv( &y, &norm );
bli_getsc( &norm, resid, &junk );
bli_obj_free( &y );
bli_obj_free( &a_local );
}
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 );
}
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 );
}
int main( int argc, char** argv )
{
obj_t a, b, c;
obj_t c_save;
obj_t alpha, beta;
dim_t m, n;
dim_t p;
dim_t p_begin, p_end, p_inc;
int m_input, n_input;
num_t dt_a, dt_b, dt_c;
num_t dt_alpha, dt_beta;
int r, n_repeats;
side_t side;
uplo_t uplo;
double dtime;
double dtime_save;
double gflops;
bli_init();
n_repeats = 3;
if( argc < 7 )
{
printf("Usage:\n");
printf("test_foo.x m n k p_begin p_inc p_end:\n");
exit;
}
int world_size, world_rank, provided;
MPI_Init_thread( NULL, NULL, MPI_THREAD_FUNNELED, &provided );
MPI_Comm_size( MPI_COMM_WORLD, &world_size );
MPI_Comm_rank( MPI_COMM_WORLD, &world_rank );
m_input = strtol( argv[1], NULL, 10 );
n_input = strtol( argv[2], NULL, 10 );
p_begin = strtol( argv[4], NULL, 10 );
p_inc = strtol( argv[5], NULL, 10 );
p_end = strtol( argv[6], NULL, 10 );
#if 1
dt_a = BLIS_DOUBLE;
dt_b = BLIS_DOUBLE;
dt_c = BLIS_DOUBLE;
dt_alpha = BLIS_DOUBLE;
dt_beta = BLIS_DOUBLE;
#else
dt_a = dt_b = dt_c = dt_alpha = dt_beta = BLIS_DCOMPLEX;
#endif
side = BLIS_LEFT;
//side = BLIS_RIGHT;
uplo = BLIS_LOWER;
//uplo = BLIS_UPPER;
for ( p = p_begin + world_rank * p_inc; p <= p_end; p += p_inc * world_size )
{
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_alpha, 1, 1, 0, 0, &alpha );
bli_obj_create( dt_beta, 1, 1, 0, 0, &beta );
if ( bli_is_left( side ) )
bli_obj_create( dt_a, m, m, 0, 0, &a );
else
bli_obj_create( dt_a, n, n, 0, 0, &a );
bli_obj_create( dt_b, m, n, 0, 0, &b );
bli_obj_create( dt_c, m, n, 0, 0, &c );
bli_obj_create( dt_c, m, n, 0, 0, &c_save );
bli_randm( &a );
bli_randm( &b );
bli_randm( &c );
bli_obj_set_struc( BLIS_HERMITIAN, &a );
bli_obj_set_uplo( uplo, &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_obj_toggle_uplo( &a );
bli_obj_inc_diag_offset( 1, &a );
bli_setm( &BLIS_ZERO, &a );
bli_obj_inc_diag_offset( -1, &a );
bli_obj_toggle_uplo( &a );
bli_obj_set_diag( BLIS_NONUNIT_DIAG, &a );
bli_scalm( &BLIS_TWO, &a );
bli_scalm( &BLIS_TWO, &a );
*/
bli_setsc( (2.0/1.0), 1.0, &alpha );
bli_setsc( (1.0/1.0), 0.0, &beta );
bli_copym( &c, &c_save );
dtime_save = 1.0e9;
for ( r = 0; r < n_repeats; ++r )
{
bli_copym( &c_save, &c );
dtime = bli_clock();
#ifdef PRINT
/*
obj_t ar, ai;
bli_obj_alias_to( &a, &ar );
bli_obj_alias_to( &a, &ai );
bli_obj_set_dt( BLIS_DOUBLE, &ar ); ar.rs *= 2; ar.cs *= 2;
bli_obj_set_dt( BLIS_DOUBLE, &ai ); ai.rs *= 2; ai.cs *= 2; ai.buffer = ( double* )ai.buffer + 1;
bli_printm( "ar", &ar, "%4.1f", "" );
bli_printm( "ai", &ai, "%4.1f", "" );
*/
bli_printm( "a", &a, "%4.1f", "" );
bli_printm( "b", &b, "%4.1f", "" );
bli_printm( "c", &c, "%4.1f", "" );
#endif
#ifdef BLIS
//bli_error_checking_level_set( BLIS_NO_ERROR_CHECKING );
bli_hemm( side,
//bli_hemm4m( side,
&alpha,
&a,
&b,
&beta,
&c );
#else
f77_char side = 'L';
f77_char uplo = 'L';
f77_int mm = bli_obj_length( &c );
f77_int nn = bli_obj_width( &c );
f77_int lda = bli_obj_col_stride( &a );
f77_int ldb = bli_obj_col_stride( &b );
f77_int ldc = bli_obj_col_stride( &c );
double* alphap = bli_obj_buffer( &alpha );
double* ap = bli_obj_buffer( &a );
double* bp = bli_obj_buffer( &b );
double* betap = bli_obj_buffer( &beta );
double* cp = bli_obj_buffer( &c );
dsymm_( &side,
&uplo,
&mm,
&nn,
alphap,
ap, &lda,
bp, &ldb,
betap,
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 = ( 2.0 * m * m * n ) / ( dtime_save * 1.0e9 );
else
gflops = ( 2.0 * m * n * n ) / ( dtime_save * 1.0e9 );
if ( bli_is_complex( dt_a ) ) gflops *= 4.0;
#ifdef BLIS
printf( "data_hemm_blis" );
#else
printf( "data_hemm_%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( &beta );
bli_obj_free( &a );
bli_obj_free( &b );
bli_obj_free( &c );
bli_obj_free( &c_save );
}
bli_finalize();
return 0;
}
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_her2k_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, k;
uplo_t uploc;
trans_t transa, transb;
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, &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 her2k, alpha may be complex, but beta must be real-valued
// (in order to preserve the Hermitian structure of C).
bli_setsc( 0.8, 0.5, &alpha );
bli_setsc( -1.0, 0.0, &beta );
}
// Randomize A and B.
libblis_test_mobj_randomize( params, TRUE, &a );
libblis_test_mobj_randomize( params, TRUE, &b );
// Set the structure and uplo properties of C.
bli_obj_set_struc( BLIS_HERMITIAN, c );
bli_obj_set_uplo( uploc, c );
// Randomize A, make it densely Hermitian, and zero the unstored triangle
// to ensure the implementation is reads only from the stored region.
libblis_test_mobj_randomize( params, TRUE, &c );
bli_mkherm( &c );
bli_mktrim( &c );
// Save C and set its structure and uplo properties.
bli_obj_set_struc( BLIS_HERMITIAN, c_save );
bli_obj_set_uplo( uploc, c_save );
bli_copym( &c, &c_save );
// 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_her2k_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_her2k_check( params, &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 );
}
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;
}
void libblis_test_her_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;
uplo_t uploa;
conj_t conjx;
obj_t alpha, x, a;
obj_t a_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], &conjx );
// Create test scalars.
bli_obj_scalar_init_detached( datatype, &alpha );
// Create test operands (vectors and/or matrices).
libblis_test_vobj_create( params, datatype,
sc_str[0], m, &x );
libblis_test_mobj_create( params, datatype, BLIS_NO_TRANSPOSE,
sc_str[1], m, m, &a );
libblis_test_mobj_create( params, datatype, BLIS_NO_TRANSPOSE,
sc_str[1], m, m, &a_save );
// Set alpha.
//bli_copysc( &BLIS_MINUS_ONE, &alpha );
bli_setsc( -1.0, 0.0, &alpha );
// Randomize x.
bli_randv( &x );
// Set the structure and uplo properties of A.
bli_obj_set_struc( BLIS_HERMITIAN, a );
bli_obj_set_uplo( uploa, a );
// Randomize A, make it densely Hermitian, and zero the unstored triangle
// to ensure the implementation is reads only from the stored region.
bli_randm( &a );
bli_mkherm( &a );
bli_mktrim( &a );
// Save A and set its structure and uplo properties.
bli_obj_set_struc( BLIS_HERMITIAN, a_save );
bli_obj_set_uplo( uploa, a_save );
bli_copym( &a, &a_save );
// Apply the remaining parameters.
bli_obj_set_conj( conjx, x );
// Repeat the experiment n_repeats times and record results.
for ( i = 0; i < n_repeats; ++i )
{
bli_copym( &a_save, &a );
time = bli_clock();
libblis_test_her_impl( iface, &alpha, &x, &a );
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( a ) ) *perf *= 4.0;
// Perform checks.
libblis_test_her_check( &alpha, &x, &a, &a_save, resid );
// Zero out performance and residual if output matrix is empty.
libblis_test_check_empty_problem( &a, perf, resid );
// Free the test objects.
bli_obj_free( &x );
bli_obj_free( &a );
bli_obj_free( &a_save );
}
int main( int argc, char** argv )
{
obj_t a, b, c;
obj_t c_save;
obj_t alpha, beta;
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;
f77_char f77_side;
f77_char f77_uploa;
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;
bli_param_map_blis_to_netlib_side( side, &f77_side );
bli_param_map_blis_to_netlib_uplo( uploa, &f77_uploa );
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 );
bli_obj_create( dt, 1, 1, 0, 0, &beta );
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, &b );
bli_obj_create( dt, m, n, 0, 0, &c );
bli_obj_create( dt, m, n, 0, 0, &c_save );
bli_randm( &a );
bli_randm( &b );
bli_randm( &c );
bli_obj_set_struc( BLIS_HERMITIAN, a );
bli_obj_set_uplo( uploa, 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_obj_toggle_uplo( a );
bli_obj_inc_diag_off( 1, a );
bli_setm( &BLIS_ZERO, &a );
bli_obj_inc_diag_off( -1, a );
bli_obj_toggle_uplo( a );
bli_obj_set_diag( BLIS_NONUNIT_DIAG, a );
bli_scalm( &BLIS_TWO, &a );
bli_scalm( &BLIS_TWO, &a );
*/
bli_setsc( (2.0/1.0), 1.0, &alpha );
bli_setsc( -(1.0/1.0), 0.0, &beta );
bli_copym( &c, &c_save );
dtime_save = 1.0e9;
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( "b", &b, "%4.1f", "" );
bli_printm( "c", &c, "%4.1f", "" );
#endif
#ifdef BLIS
bli_hemm( side,
&alpha,
&a,
&b,
&beta,
&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 ldb = bli_obj_col_stride( b );
f77_int ldc = bli_obj_col_stride( c );
float* alphap = bli_obj_buffer( alpha );
float* ap = bli_obj_buffer( a );
float* bp = bli_obj_buffer( b );
float* betap = bli_obj_buffer( beta );
float* cp = bli_obj_buffer( c );
ssymm_( &f77_side,
&f77_uploa,
&mm,
&nn,
alphap,
ap, &lda,
bp, &ldb,
betap,
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 ldb = bli_obj_col_stride( b );
f77_int ldc = bli_obj_col_stride( c );
double* alphap = bli_obj_buffer( alpha );
double* ap = bli_obj_buffer( a );
double* bp = bli_obj_buffer( b );
double* betap = bli_obj_buffer( beta );
double* cp = bli_obj_buffer( c );
dsymm_( &f77_side,
&f77_uploa,
&mm,
&nn,
alphap,
ap, &lda,
bp, &ldb,
betap,
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 ldb = bli_obj_col_stride( b );
f77_int ldc = bli_obj_col_stride( c );
scomplex* alphap = bli_obj_buffer( alpha );
scomplex* ap = bli_obj_buffer( a );
scomplex* bp = bli_obj_buffer( b );
scomplex* betap = bli_obj_buffer( beta );
scomplex* cp = bli_obj_buffer( c );
chemm_( &f77_side,
&f77_uploa,
&mm,
&nn,
alphap,
ap, &lda,
bp, &ldb,
betap,
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 ldb = bli_obj_col_stride( b );
f77_int ldc = bli_obj_col_stride( c );
dcomplex* alphap = bli_obj_buffer( alpha );
dcomplex* ap = bli_obj_buffer( a );
dcomplex* bp = bli_obj_buffer( b );
dcomplex* betap = bli_obj_buffer( beta );
dcomplex* cp = bli_obj_buffer( c );
zhemm_( &f77_side,
&f77_uploa,
&mm,
&nn,
alphap,
ap, &lda,
bp, &ldb,
betap,
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 = ( 2.0 * m * m * n ) / ( dtime_save * 1.0e9 );
else
gflops = ( 2.0 * m * n * n ) / ( dtime_save * 1.0e9 );
if ( bli_is_complex( dt ) ) gflops *= 4.0;
#ifdef BLIS
printf( "data_hemm_blis" );
#else
printf( "data_hemm_%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( &beta );
bli_obj_free( &a );
bli_obj_free( &b );
bli_obj_free( &c );
bli_obj_free( &c_save );
}
bli_finalize();
return 0;
}
void libblis_test_syrk_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, k;
uplo_t uploc;
trans_t transa;
obj_t kappa;
obj_t alpha, a, 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 );
// 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, BLIS_NO_TRANSPOSE,
sc_str[1], m, m, &c );
libblis_test_mobj_create( params, datatype, BLIS_NO_TRANSPOSE,
sc_str[1], m, m, &c_save );
// Set alpha and beta.
if ( bli_obj_is_real( c ) )
{
bli_setsc( 1.2, 0.0, &alpha );
bli_setsc( -1.0, 0.0, &beta );
}
else
{
// For syrk, both alpha and beta may be complex since, unlike herk,
// C is symmetric in both the real and complex cases.
bli_setsc( 1.2, 0.5, &alpha );
bli_setsc( -1.0, 0.5, &beta );
}
// Randomize A.
bli_randm( &a );
// 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 );
// Apply the remaining parameters.
bli_obj_set_conjtrans( transa, a );
// 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_syrk_impl( impl, &alpha, &a, &beta, &c );
time_min = bli_clock_min_diff( time_min, time );
}
// Estimate the performance of the best experiment repeat.
*perf = ( 1.0 * m * m * k ) / time_min / FLOPS_PER_UNIT_PERF;
if ( bli_obj_is_complex( c ) ) *perf *= 4.0;
// Perform checks.
libblis_test_syrk_check( &alpha, &a, &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( &c );
bli_obj_free( &c_save );
}
void libblis_test_hemm_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;
conj_t conja;
trans_t 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 );
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_conj( pc_str[2], &conja );
bli_param_map_char_to_blis_trans( pc_str[3], &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).
bli_set_dim_with_side( side, m, n, mn_side );
libblis_test_mobj_create( params, datatype, BLIS_NO_TRANSPOSE,
sc_str[0], mn_side, mn_side, &a );
libblis_test_mobj_create( params, datatype, transb,
sc_str[1], m, n, &b );
libblis_test_mobj_create( params, datatype, BLIS_NO_TRANSPOSE,
sc_str[2], m, n, &c );
libblis_test_mobj_create( params, datatype, BLIS_NO_TRANSPOSE,
sc_str[2], m, n, &c_save );
// Set alpha and beta.
if ( bli_obj_is_real( c ) )
{
bli_setsc( 1.2, 0.0, &alpha );
bli_setsc( -1.0, 0.0, &beta );
}
else
{
bli_setsc( 1.2, 0.8, &alpha );
bli_setsc( -1.0, 1.0, &beta );
}
// Set the structure and uplo properties of A.
bli_obj_set_struc( BLIS_HERMITIAN, a );
bli_obj_set_uplo( uploa, 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 );
// Randomize B and C, and save C.
bli_randm( &b );
bli_randm( &c );
bli_copym( &c, &c_save );
// Normalize by m.
bli_setsc( 1.0/( double )m, 0.0, &kappa );
bli_scalm( &kappa, &b );
// Apply the remaining parameters.
bli_obj_set_conj( conja, 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_hemm_impl( iface, side, &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 * mn_side * m * n ) / time_min / FLOPS_PER_UNIT_PERF;
if ( bli_obj_is_complex( c ) ) *perf *= 4.0;
// Perform checks.
libblis_test_hemm_check( side, &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 );
}
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;
}
