void libblis_test_addv_check( obj_t* alpha,
obj_t* beta,
obj_t* x,
obj_t* y,
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 );
conj_t conjx = bli_obj_conj_status( *x );
obj_t aplusb;
obj_t alpha_conj;
obj_t norm_r, m_r, temp_r;
double junk;
//
// Pre-conditions:
// - x is set to alpha.
// - y_orig is set to beta.
// 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 := y_orig + conjx(x)
//
// is functioning correctly if
//
// fnormv(y) - sqrt( absqsc( beta + conjx(alpha) ) * m )
//
// is negligible.
//
bli_obj_scalar_init_detached( dt, &aplusb );
bli_obj_scalar_init_detached( dt_real, &temp_r );
bli_obj_scalar_init_detached( dt_real, &norm_r );
bli_obj_scalar_init_detached( dt_real, &m_r );
bli_obj_scalar_init_detached_copy_of( dt, conjx, alpha, &alpha_conj );
bli_fnormv( y, &norm_r );
bli_copysc( beta, &aplusb );
bli_addsc( &alpha_conj, &aplusb );
bli_setsc( ( double )m, 0.0, &m_r );
bli_absqsc( &aplusb, &temp_r );
bli_mulsc( &m_r, &temp_r );
bli_sqrtsc( &temp_r, &temp_r );
bli_subsc( &temp_r, &norm_r );
bli_getsc( &norm_r, resid, &junk );
}
void libblis_test_scalv_check
(
test_params_t* params,
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 norm_y_r;
obj_t nbeta;
obj_t y2;
double junk;
//
// Pre-conditions:
// - y_orig is randomized.
// Note:
// - beta 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
//
// y := conjbeta(beta) * y_orig
//
// is functioning correctly if
//
// normf( y + -conjbeta(beta) * y_orig )
//
// is negligible.
//
bli_obj_create( dt, m, 1, 0, 0, &y2 );
bli_copyv( y_orig, &y2 );
bli_obj_scalar_init_detached( dt, &nbeta );
bli_obj_scalar_init_detached( dt_real, &norm_y_r );
bli_copysc( beta, &nbeta );
bli_mulsc( &BLIS_MINUS_ONE, &nbeta );
bli_scalv( &nbeta, &y2 );
bli_addv( &y2, y );
bli_normfv( y, &norm_y_r );
bli_getsc( &norm_y_r, resid, &junk );
bli_obj_free( &y2 );
}
void libblis_test_normfm_check
(
test_params_t* params,
obj_t* beta,
obj_t* x,
obj_t* norm,
double* resid
)
{
num_t dt_real = bli_obj_datatype_proj_to_real( *x );
dim_t m = bli_obj_length( *x );
dim_t n = bli_obj_width( *x );
obj_t m_r, n_r, temp_r;
double junk;
//
// Pre-conditions:
// - x is set to beta.
// Note:
// - beta 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
//
// norm := normf( x )
//
// is functioning correctly if
//
// norm = sqrt( absqsc( beta ) * m * n )
//
// where m and n are the dimensions of x.
//
bli_obj_scalar_init_detached( dt_real, &temp_r );
bli_obj_scalar_init_detached( dt_real, &m_r );
bli_obj_scalar_init_detached( dt_real, &n_r );
bli_setsc( ( double )m, 0.0, &m_r );
bli_setsc( ( double )n, 0.0, &n_r );
bli_absqsc( beta, &temp_r );
bli_mulsc( &m_r, &temp_r );
bli_mulsc( &n_r, &temp_r );
bli_sqrtsc( &temp_r, &temp_r );
bli_subsc( &temp_r, norm );
bli_getsc( norm, resid, &junk );
}
void libblis_test_xpbym_check
(
test_params_t* params,
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_length( y );
dim_t n = bli_obj_width( y );
obj_t x_temp, y_temp;
obj_t norm;
double junk;
//
// Pre-conditions:
// - x is randomized.
// - y_orig 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
//
// y := beta * y_orig + conjx(x)
//
// is functioning correctly if
//
// normf( y - ( beta * y_orig + conjx(x) ) )
//
// is negligible.
//
bli_obj_scalar_init_detached( dt_real, &norm );
bli_obj_create( dt, m, n, 0, 0, &x_temp );
bli_obj_create( dt, m, n, 0, 0, &y_temp );
bli_copym( x, &x_temp );
bli_copym( y_orig, &y_temp );
bli_scalm( beta, &y_temp );
bli_addm( &x_temp, &y_temp );
bli_subm( &y_temp, y );
bli_normfm( y, &norm );
bli_getsc( &norm, resid, &junk );
bli_obj_free( &x_temp );
bli_obj_free( &y_temp );
}
void libblis_test_hemv_check( 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_datatype( *y );
num_t dt_real = bli_obj_datatype_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 Hermitian.
// - 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_mkherm( 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_axpbyv_check( obj_t* alpha,
obj_t* x,
obj_t* beta,
obj_t* y,
obj_t* y_orig,
double* resid )
{
num_t dt = bli_obj_datatype( *y );
num_t dt_real = bli_obj_datatype_proj_to_real( *y );
dim_t m = bli_obj_vector_dim( *y );
obj_t x_temp, y_temp;
obj_t norm;
double junk;
//
// Pre-conditions:
// - x is randomized.
// - y_orig 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
//
// y := beta * y_orig + alpha * conjx(x)
//
// is functioning correctly if
//
// normf( y - ( beta * y_orig + alpha * conjx(x) ) )
//
// is negligible.
//
bli_obj_scalar_init_detached( dt_real, &norm );
bli_obj_create( dt, m, 1, 0, 0, &x_temp );
bli_obj_create( dt, m, 1, 0, 0, &y_temp );
bli_copyv( x, &x_temp );
bli_copyv( y_orig, &y_temp );
bli_scalv( alpha, &x_temp );
bli_scalv( beta, &y_temp );
bli_addv( &x_temp, &y_temp );
bli_subv( &y_temp, y );
bli_normfv( y, &norm );
bli_getsc( &norm, resid, &junk );
bli_obj_free( &x_temp );
bli_obj_free( &y_temp );
}
void bli_obj_scalar_detach( obj_t* a,
obj_t* alpha )
{
num_t dt_a = bli_obj_datatype( *a );
// Initialize alpha to be a bufferless internal scalar of the same
// datatype as A.
bli_obj_scalar_init_detached( dt_a, alpha );
// Copy the internal scalar in A to alpha.
bli_obj_copy_internal_scalar( *a, *alpha );
}
void libblis_test_amaxv_check
(
test_params_t* params,
obj_t* x,
obj_t* index,
double* resid
)
{
obj_t index_test;
obj_t chi_i;
obj_t chi_i_test;
dim_t i;
dim_t i_test;
double i_d, junk;
double i_d_test;
//
// Pre-conditions:
// - x is randomized.
//
// Under these conditions, we assume that the implementation for
//
// index := amaxv( x )
//
// is functioning correctly if
//
// x[ index ] = max( x )
//
// where max() is implemented via the bli_?amaxv_test() function.
//
// The following two calls have already been made by the caller. That
// is, the index object has already been created and the library's
// amaxv implementation has already been tested.
//bli_obj_scalar_init_detached( BLIS_INT, &index );
//bli_amaxv( x, &index );
bli_getsc( index, &i_d, &junk );
i = i_d;
bli_acquire_vi( i, x, &chi_i );
bli_obj_scalar_init_detached( BLIS_INT, &index_test );
bli_amaxv_test( x, &index_test );
bli_getsc( &index_test, &i_d_test, &junk );
i_test = i_d_test;
bli_acquire_vi( i_test, x, &chi_i_test );
// Verify that the values referenced by index and index_test are equal.
if ( bli_obj_equals( &chi_i, &chi_i_test ) ) *resid = 0.0;
else *resid = 1.0;
}
void bli_obj_scalar_init_detached_copy_of( num_t dt,
conj_t conj,
obj_t* alpha,
obj_t* beta )
{
obj_t alpha_local;
// Make a local copy of alpha so we can apply the conj parameter.
bli_obj_alias_to( *alpha, alpha_local );
bli_obj_apply_conj( conj, alpha_local );
// Initialize beta without a buffer and then attach its internal buffer.
bli_obj_scalar_init_detached( dt, beta );
// Copy the scalar value in a to object b, conjugating and/or
// typecasting if needed.
bli_copysc( &alpha_local, beta );
}
void libblis_test_randv_check
(
test_params_t* params,
obj_t* x,
double* resid
)
{
num_t dt_real = bli_obj_datatype_proj_to_real( *x );
dim_t m_x = bli_obj_vector_dim( *x );
obj_t sum;
*resid = 0.0;
//
// The two most likely ways that randv would fail is if all elements
// were zero, or if all elements were greater than or equal to one.
// We check both of these conditions by computing the sum of the
// absolute values of the elements of x.
//
bli_obj_scalar_init_detached( dt_real, &sum );
bli_norm1v( x, &sum );
if ( bli_is_float( dt_real ) )
{
float* sum_x = bli_obj_buffer_at_off( sum );
if ( *sum_x == *bli_d0 ) *resid = 1.0;
else if ( *sum_x >= 2.0 * m_x ) *resid = 2.0;
}
else // if ( bli_is_double( dt_real ) )
{
double* sum_x = bli_obj_buffer_at_off( sum );
if ( *sum_x == *bli_d0 ) *resid = 1.0;
else if ( *sum_x >= 2.0 * m_x ) *resid = 2.0;
}
}
void libblis_test_copym_check
(
test_params_t* params,
obj_t* x,
obj_t* y,
double* resid
)
{
num_t dt_real = bli_obj_dt_proj_to_real( x );
obj_t norm_y_r;
double junk;
//
// Pre-conditions:
// - x is randomized.
//
// Under these conditions, we assume that the implementation for
//
// y := conjx(x)
//
// is functioning correctly if
//
// normfm( y - conjx(x) )
//
// is negligible.
//
bli_obj_scalar_init_detached( dt_real, &norm_y_r );
bli_subm( x, y );
bli_normfm( y, &norm_y_r );
bli_getsc( &norm_y_r, resid, &junk );
}
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 );
}
void libblis_test_scalv_experiment
(
test_params_t* params,
test_op_t* op,
iface_t iface,
char* dc_str,
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;
num_t datatype;
dim_t m;
conj_t conjbeta;
obj_t beta, y;
obj_t y_save;
// Use the datatype of the first char in the datatype combination string.
bli_param_map_char_to_blis_dt( dc_str[0], &datatype );
// 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_conj( pc_str[0], &conjbeta );
// Create test scalars.
bli_obj_scalar_init_detached( datatype, &beta );
// Create test operands (vectors and/or matrices).
libblis_test_vobj_create( params, datatype, sc_str[0], m, &y );
libblis_test_vobj_create( params, datatype, sc_str[0], m, &y_save );
// Set beta.
if ( bli_obj_is_real( &y ) )
bli_setsc( -2.0, 0.0, &beta );
else
bli_setsc( 0.0, -2.0, &beta );
// Randomize and save y.
libblis_test_vobj_randomize( params, FALSE, &y );
bli_copyv( &y, &y_save );
// Apply the parameters.
bli_obj_set_conj( conjbeta, &beta );
// Repeat the experiment n_repeats times and record results.
for ( i = 0; i < n_repeats; ++i )
{
bli_copyv( &y_save, &y );
time = bli_clock();
libblis_test_scalv_impl( iface, &beta, &y );
time_min = bli_clock_min_diff( time_min, time );
}
// Estimate the performance of the best experiment repeat.
*perf = ( 1.0 * m ) / time_min / FLOPS_PER_UNIT_PERF;
if ( bli_obj_is_complex( &y ) ) *perf *= 6.0;
// Perform checks.
libblis_test_scalv_check( params, &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( &y );
bli_obj_free( &y_save );
}
void libblis_test_addv_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 )
{
double time_min = 1e9;
double time;
dim_t m;
conj_t conjx;
obj_t alpha, beta;
obj_t x, y;
// 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_conj( pc_str[0], &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_vobj_create( params, datatype, sc_str[0], m, &x );
libblis_test_vobj_create( params, datatype, sc_str[1], m, &y );
// Initialize alpha and beta.
bli_setsc( -1.0, -1.0, &alpha );
bli_setsc( 3.0, 3.0, &beta );
// Set x and y to alpha and beta, respectively.
bli_setv( &alpha, &x );
bli_setv( &beta, &y );
// Apply the parameters.
bli_obj_set_conj( conjx, x );
// Disable repeats since bli_copyv() is not yet tested.
//for ( i = 0; i < n_repeats; ++i )
{
time = bli_clock();
libblis_test_addv_impl( impl, &x, &y );
time_min = bli_clock_min_diff( time_min, time );
}
// Estimate the performance of the best experiment repeat.
*perf = ( 2.0 * m ) / time_min / FLOPS_PER_UNIT_PERF;
if ( bli_obj_is_complex( x ) ) *perf *= 2.0;
// Perform checks.
libblis_test_addv_check( &alpha, &beta, &x, &y, 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( &x );
bli_obj_free( &y );
}
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_dotxaxpyf_check
(
test_params_t* params,
obj_t* alpha,
obj_t* at,
obj_t* a,
obj_t* w,
obj_t* x,
obj_t* beta,
obj_t* y,
obj_t* z,
obj_t* y_orig,
obj_t* z_orig,
double* resid
)
{
num_t dt = bli_obj_datatype( *y );
num_t dt_real = bli_obj_datatype_proj_to_real( *y );
dim_t m = bli_obj_vector_dim( *z );
dim_t b_n = bli_obj_vector_dim( *y );
dim_t i;
obj_t a1, chi1, psi1, v, q;
obj_t alpha_chi1;
obj_t norm;
double resid1, resid2;
double junk;
//
// Pre-conditions:
// - a is randomized.
// - w is randomized.
// - x is randomized.
// - y is randomized.
// - z is randomized.
// - at is an alias to a.
// Note:
// - alpha and beta 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
//
// y := beta * y_orig + alpha * conjat(A^T) * conjw(w)
// z := z_orig + alpha * conja(A) * conjx(x)
//
// is functioning correctly if
//
// normf( y - v )
//
// and
//
// normf( z - q )
//
// are negligible, where v and q contain y and z as computed by repeated
// calls to dotxv and axpyv, respectively.
//
bli_obj_scalar_init_detached( dt_real, &norm );
bli_obj_scalar_init_detached( dt, &alpha_chi1 );
bli_obj_create( dt, b_n, 1, 0, 0, &v );
bli_obj_create( dt, m, 1, 0, 0, &q );
bli_copyv( y_orig, &v );
bli_copyv( z_orig, &q );
// v := beta * v + alpha * conjat(at) * conjw(w)
for ( i = 0; i < b_n; ++i )
{
bli_acquire_mpart_l2r( BLIS_SUBPART1, i, 1, at, &a1 );
bli_acquire_vpart_f2b( BLIS_SUBPART1, i, 1, &v, &psi1 );
bli_dotxv( alpha, &a1, w, beta, &psi1 );
}
// q := q + alpha * conja(a) * conjx(x)
for ( i = 0; i < b_n; ++i )
{
bli_acquire_mpart_l2r( BLIS_SUBPART1, i, 1, a, &a1 );
bli_acquire_vpart_f2b( BLIS_SUBPART1, i, 1, x, &chi1 );
bli_copysc( &chi1, &alpha_chi1 );
bli_mulsc( alpha, &alpha_chi1 );
bli_axpyv( &alpha_chi1, &a1, &q );
}
bli_subv( y, &v );
bli_normfv( &v, &norm );
bli_getsc( &norm, &resid1, &junk );
bli_subv( z, &q );
bli_normfv( &q, &norm );
bli_getsc( &norm, &resid2, &junk );
*resid = bli_fmaxabs( resid1, resid2 );
bli_obj_free( &v );
bli_obj_free( &q );
}
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_gemmtrsm_ukr_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, n, k;
char sc_a = 'c';
char sc_b = 'r';
side_t side = BLIS_LEFT;
uplo_t uploa;
obj_t kappa;
obj_t alpha;
obj_t a_big, a, b;
obj_t b11, c11;
obj_t ap, bp;
obj_t a1xp, a11p, bx1p, b11p;
obj_t c11_save;
// Map the dimension specifier to actual dimensions.
k = libblis_test_get_dim_from_prob_size( op->dim_spec[0], p_cur );
// Fix m and n to MR and NR, respectively.
m = bli_blksz_for_type( datatype, gemm_mr );
n = bli_blksz_for_type( datatype, gemm_nr );
// Store the register blocksizes so that the driver can retrieve the
// values later when printing results.
op->dim_aux[0] = m;
op->dim_aux[1] = n;
// Map parameter characters to BLIS constants.
bli_param_map_char_to_blis_uplo( pc_str[0], &uploa );
// Create test scalars.
bli_obj_scalar_init_detached( datatype, &kappa );
bli_obj_scalar_init_detached( datatype, &alpha );
// Create test operands (vectors and/or matrices).
libblis_test_mobj_create( params, datatype, BLIS_NO_TRANSPOSE,
sc_a, k+m, k+m, &a_big );
libblis_test_mobj_create( params, datatype, BLIS_NO_TRANSPOSE,
sc_b, k+m, n, &b );
libblis_test_mobj_create( params, datatype, BLIS_NO_TRANSPOSE,
sc_str[0], m, n, &c11 );
libblis_test_mobj_create( params, datatype, BLIS_NO_TRANSPOSE,
sc_str[0], m, n, &c11_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, uplo, and diagonal offset properties of A.
bli_obj_set_struc( BLIS_TRIANGULAR, a_big );
bli_obj_set_uplo( uploa, a_big );
// Randomize A and make it densely triangular.
bli_randm( &a_big );
// Normalize B and save.
bli_randm( &b );
bli_setsc( 1.0/( double )m, 0.0, &kappa );
bli_scalm( &kappa, &b );
// Use the last m rows of A_big as A.
bli_acquire_mpart_t2b( BLIS_SUBPART1, k, m, &a_big, &a );
// Locate the B11 block of B, copy to C11, and save.
if ( bli_obj_is_lower( a ) )
bli_acquire_mpart_t2b( BLIS_SUBPART1, k, m, &b, &b11 );
else
bli_acquire_mpart_t2b( BLIS_SUBPART1, 0, m, &b, &b11 );
bli_copym( &b11, &c11 );
bli_copym( &c11, &c11_save );
// Initialize pack objects.
bli_obj_init_pack( &ap );
bli_obj_init_pack( &bp );
// Create pack objects for a and b.
libblis_test_pobj_create( gemm_mr,
gemm_mr,
BLIS_INVERT_DIAG,
BLIS_PACKED_ROW_PANELS,
BLIS_BUFFER_FOR_A_BLOCK,
&a, &ap );
libblis_test_pobj_create( gemm_mr,
gemm_nr,
BLIS_NO_INVERT_DIAG,
BLIS_PACKED_COL_PANELS,
BLIS_BUFFER_FOR_B_PANEL,
&b, &bp );
// Pack the contents of a to ap.
bli_packm_blk_var3( &a, &ap );
// Pack the contents of b to bp.
bli_packm_blk_var2( &b, &bp );
// Create subpartitions from the a and b panels.
bli_gemmtrsm_ukr_make_subparts( k, &ap, &bp,
&a1xp, &a11p, &bx1p, &b11p );
// Repeat the experiment n_repeats times and record results.
for ( i = 0; i < n_repeats; ++i )
{
bli_copym( &c11_save, &c11 );
// Re-pack the contents of b to bp.
bli_packm_blk_var2( &b, &bp );
time = bli_clock();
libblis_test_gemmtrsm_ukr_impl( impl, side, &alpha,
&a1xp, &a11p, &bx1p, &b11p, &c11 );
time_min = bli_clock_min_diff( time_min, time );
}
// Estimate the performance of the best experiment repeat.
*perf = ( 2.0 * m * n * k + 1.0 * m * m * n ) / time_min / FLOPS_PER_UNIT_PERF;
if ( bli_obj_is_complex( b ) ) *perf *= 4.0;
// Perform checks.
libblis_test_gemmtrsm_ukr_check( side, &alpha,
&a1xp, &a11p, &bx1p, &b11p, &c11, &c11_save, resid );
// Zero out performance and residual if output matrix is empty.
//libblis_test_check_empty_problem( &c11, perf, resid );
// Release packing buffers within pack objects.
bli_obj_release_pack( &ap );
bli_obj_release_pack( &bp );
// Free the test objects.
bli_obj_free( &a_big );
bli_obj_free( &b );
bli_obj_free( &c11 );
bli_obj_free( &c11_save );
}
void libblis_test_her_check( obj_t* alpha,
obj_t* x,
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 xh, t, v, w;
obj_t tau, rho, norm;
double junk;
//
// Pre-conditions:
// - x is randomized.
// - a is randomized and Hermitian.
// Note:
// - alpha must be real-valued.
//
// Under these conditions, we assume that the implementation for
//
// A := A_orig + alpha * conjx(x) * conjx(x)^H
//
// is functioning correctly if
//
// normf( v - w )
//
// is negligible, where
//
// v = A * t
// w = ( A_orig + alpha * conjx(x) * conjx(x)^H ) * t
// = A_orig * t + alpha * conjx(x) * conjx(x)^H * t
// = A_orig * t + alpha * conjx(x) * rho
// = A_orig * t + w
//
bli_mkherm( a );
bli_mkherm( 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, &w );
bli_obj_alias_with_conj( BLIS_CONJUGATE, *x, xh );
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( &xh, &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 );
}
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 );
}
void libblis_test_gemmtrsm_ukr_check( side_t side,
obj_t* alpha,
obj_t* a1x,
obj_t* a11,
obj_t* bx1,
obj_t* b11,
obj_t* c11,
obj_t* c11_orig,
double* resid )
{
num_t dt = bli_obj_datatype( *b11 );
num_t dt_real = bli_obj_datatype_proj_to_real( *b11 );
dim_t m = bli_obj_length( *b11 );
dim_t n = bli_obj_width( *b11 );
dim_t k = bli_obj_width( *a1x );
obj_t kappa, norm;
obj_t t, v, w, z;
double junk;
//
// Pre-conditions:
// - a1x, a11, bx1, c11_orig are randomized; a11 is triangular.
// - contents of b11 == contents of c11.
// - side == BLIS_LEFT.
//
// Under these conditions, we assume that the implementation for
//
// B := inv(A11) * ( alpha * B11 - A1x * Bx1 ) (side = left)
//
// is functioning correctly if
//
// fnorm( v - z )
//
// is negligible, where
//
// v = B11 * t
//
// z = ( inv(A11) * ( alpha * B11_orig - A1x * Bx1 ) ) * t
// = inv(A11) * ( alpha * B11_orig * t - A1x * Bx1 * t )
// = inv(A11) * ( alpha * B11_orig * t - A1x * w )
//
bli_obj_scalar_init_detached( dt, &kappa );
bli_obj_scalar_init_detached( dt_real, &norm );
if ( bli_is_left( side ) )
{
bli_obj_create( dt, n, 1, 0, 0, &t );
bli_obj_create( dt, m, 1, 0, 0, &v );
bli_obj_create( dt, k, 1, 0, 0, &w );
bli_obj_create( dt, m, 1, 0, 0, &z );
}
else // else if ( bli_is_left( side ) )
{
// BLIS does not currently support right-side micro-kernels.
bli_check_error_code( BLIS_NOT_YET_IMPLEMENTED );
}
bli_randv( &t );
bli_setsc( 1.0/( double )n, 0.0, &kappa );
bli_scalv( &kappa, &t );
bli_gemv( &BLIS_ONE, b11, &t, &BLIS_ZERO, &v );
// Restore the diagonal of a11 to its original, un-inverted state
// (needed for trsv).
bli_invertd( a11 );
if ( bli_is_left( side ) )
{
bli_gemv( &BLIS_ONE, bx1, &t, &BLIS_ZERO, &w );
bli_gemv( alpha, c11_orig, &t, &BLIS_ZERO, &z );
bli_gemv( &BLIS_MINUS_ONE, a1x, &w, &BLIS_ONE, &z );
bli_trsv( &BLIS_ONE, a11, &z );
}
else // else if ( bli_is_left( side ) )
{
// BLIS does not currently support right-side micro-kernels.
bli_check_error_code( BLIS_NOT_YET_IMPLEMENTED );
}
bli_subv( &z, &v );
bli_fnormv( &v, &norm );
bli_getsc( &norm, resid, &junk );
bli_obj_free( &t );
bli_obj_free( &v );
bli_obj_free( &w );
bli_obj_free( &z );
}
void libblis_test_her2k_check
(
test_params_t* params,
obj_t* alpha,
obj_t* a,
obj_t* b,
obj_t* beta,
obj_t* c,
obj_t* c_orig,
double* resid
)
{
num_t dt = bli_obj_datatype( *c );
num_t dt_real = bli_obj_datatype_proj_to_real( *c );
dim_t m = bli_obj_length( *c );
dim_t k = bli_obj_width_after_trans( *a );
obj_t alphac, ah, bh;
obj_t norm;
obj_t t, v, w1, w2, z;
double junk;
//
// Pre-conditions:
// - a is randomized.
// - b is randomized.
// - c_orig is randomized and Hermitian.
// Note:
// - alpha should have a non-zero imaginary component in the
// complex cases in order to more fully exercise the implementation.
// - beta must be real-valued.
//
// Under these conditions, we assume that the implementation for
//
// C := beta * C_orig + alpha * transa(A) * transb(B)^H + conj(alpha) * transb(B) * transa(A)^H
//
// is functioning correctly if
//
// normf( v - z )
//
// is negligible, where
//
// v = C * t
// z = ( beta * C_orig + alpha * transa(A) * transb(B)^H + conj(alpha) * transb(B) * transa(A)^H ) * t
// = beta * C_orig * t + alpha * transa(A) * transb(B)^H * t + conj(alpha) * transb(B) * transa(A)^H * t
// = beta * C_orig * t + alpha * transa(A) * transb(B)^H * t + conj(alpha) * transb(B) * w2
// = beta * C_orig * t + alpha * transa(A) * w1 + conj(alpha) * transb(B) * w2
// = beta * C_orig * t + alpha * transa(A) * w1 + z
// = beta * C_orig * t + z
//
bli_obj_alias_with_trans( BLIS_CONJ_TRANSPOSE, *a, ah );
bli_obj_alias_with_trans( BLIS_CONJ_TRANSPOSE, *b, bh );
bli_obj_scalar_init_detached( dt_real, &norm );
bli_obj_scalar_init_detached_copy_of( dt, BLIS_CONJUGATE, alpha, &alphac );
bli_obj_create( dt, m, 1, 0, 0, &t );
bli_obj_create( dt, m, 1, 0, 0, &v );
bli_obj_create( dt, k, 1, 0, 0, &w1 );
bli_obj_create( dt, k, 1, 0, 0, &w2 );
bli_obj_create( dt, m, 1, 0, 0, &z );
libblis_test_vobj_randomize( params, TRUE, &t );
bli_hemv( &BLIS_ONE, c, &t, &BLIS_ZERO, &v );
bli_gemv( &BLIS_ONE, &ah, &t, &BLIS_ZERO, &w2 );
bli_gemv( &BLIS_ONE, &bh, &t, &BLIS_ZERO, &w1 );
bli_gemv( alpha, a, &w1, &BLIS_ZERO, &z );
bli_gemv( &alphac, b, &w2, &BLIS_ONE, &z );
bli_hemv( beta, c_orig, &t, &BLIS_ONE, &z );
bli_subv( &z, &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 );
bli_obj_free( &z );
}
void libblis_test_fnormv_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;
num_t dt_real = bli_datatype_proj_to_real( datatype );
double time_min = 1e9;
double time;
dim_t m;
obj_t beta, norm;
obj_t x;
// 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.
// Create test scalars.
bli_obj_scalar_init_detached( datatype, &beta );
bli_obj_scalar_init_detached( dt_real, &norm );
// Create test operands (vectors and/or matrices).
libblis_test_vobj_create( params, datatype, sc_str[0], m, &x );
// Initialize beta to 2 - 2i.
bli_setsc( 2.0, -2.0, &beta );
// Set all elements of x to beta.
bli_setv( &beta, &x );
// Repeat the experiment n_repeats times and record results.
for ( i = 0; i < n_repeats; ++i )
{
time = bli_clock();
libblis_test_fnormv_impl( impl, &x, &norm );
time_min = bli_clock_min_diff( time_min, time );
}
// Estimate the performance of the best experiment repeat.
*perf = ( 2.0 * m ) / time_min / FLOPS_PER_UNIT_PERF;
if ( bli_obj_is_complex( x ) ) *perf *= 2.0;
// Perform checks.
libblis_test_fnormv_check( &beta, &x, &norm, resid );
// Zero out performance and residual if input vector is empty.
libblis_test_check_empty_problem( &x, perf, resid );
// Free the test objects.
bli_obj_free( &x );
}
void libblis_test_axpyf_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, b_n;
conj_t conja, conjx;
obj_t alpha, a, x, 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 );
// Query the operation's fusing factor for the current datatype.
b_n = bli_axpyf_fusefac( datatype );
// Store the fusing factor so that the driver can retrieve the value
// later when printing results.
op->dim_aux[0] = b_n;
// Map parameter characters to BLIS constants.
bli_param_map_char_to_blis_conj( pc_str[0], &conja );
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_mobj_create( params, datatype, BLIS_NO_TRANSPOSE,
sc_str[0], m, b_n, &a );
libblis_test_vobj_create( params, datatype, sc_str[1], b_n, &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.
if ( bli_obj_is_real( y ) )
{
bli_setsc( -1.0, 0.0, &alpha );
}
else
{
bli_setsc( 0.0, -1.0, &alpha );
}
// Randomize A, x, and y, and save y.
bli_randm( &a );
bli_randv( &x );
bli_randv( &y );
bli_copyv( &y, &y_save );
// Apply the 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_copyv( &y_save, &y );
time = bli_clock();
libblis_test_axpyf_impl( iface, &alpha, &a, &x, &y );
time_min = bli_clock_min_diff( time_min, time );
}
// Estimate the performance of the best experiment repeat.
*perf = ( 2.0 * m * b_n ) / time_min / FLOPS_PER_UNIT_PERF;
if ( bli_obj_is_complex( y ) ) *perf *= 4.0;
// Perform checks.
libblis_test_axpyf_check( &alpha, &a, &x, &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_dotxaxpyf_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, b_n;
conj_t conjat, conja, conjw, conjx;
obj_t alpha, at, a, w, x, beta, y, z;
obj_t y_save, z_save;
cntx_t cntx;
// Initialize a context.
bli_dotxaxpyf_cntx_init( &cntx );
// Map the dimension specifier to an actual dimension.
m = libblis_test_get_dim_from_prob_size( op->dim_spec[0], p_cur );
// Query the operation's fusing factor for the current datatype.
b_n = bli_cntx_get_blksz_def_dt( datatype, BLIS_XF, &cntx );
// Store the fusing factor so that the driver can retrieve the value
// later when printing results.
op->dim_aux[0] = b_n;
// Map parameter characters to BLIS constants.
bli_param_map_char_to_blis_conj( pc_str[0], &conjat );
bli_param_map_char_to_blis_conj( pc_str[1], &conja );
bli_param_map_char_to_blis_conj( pc_str[2], &conjw );
bli_param_map_char_to_blis_conj( pc_str[3], &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, b_n, &a );
libblis_test_vobj_create( params, datatype, sc_str[1], m, &w );
libblis_test_vobj_create( params, datatype, sc_str[2], b_n, &x );
libblis_test_vobj_create( params, datatype, sc_str[3], b_n, &y );
libblis_test_vobj_create( params, datatype, sc_str[3], b_n, &y_save );
libblis_test_vobj_create( params, datatype, sc_str[4], m, &z );
libblis_test_vobj_create( params, datatype, sc_str[4], m, &z_save );
// Set alpha.
if ( bli_obj_is_real( y ) )
{
bli_setsc( 1.2, 0.0, &alpha );
bli_setsc( -1.0, 0.0, &beta );
}
else
{
bli_setsc( 1.2, 0.1, &alpha );
bli_setsc( -1.0, -0.1, &beta );
}
// Randomize A, w, x, y, and z, and save y and z.
libblis_test_mobj_randomize( params, FALSE, &a );
libblis_test_vobj_randomize( params, FALSE, &w );
libblis_test_vobj_randomize( params, FALSE, &x );
libblis_test_vobj_randomize( params, FALSE, &y );
libblis_test_vobj_randomize( params, FALSE, &z );
bli_copyv( &y, &y_save );
bli_copyv( &z, &z_save );
// Create an alias to a for at. (Note that it should NOT actually be
// marked for transposition since the transposition is part of the dotxf
// subproblem.)
bli_obj_alias_to( a, at );
// Apply the parameters.
bli_obj_set_conj( conjat, at );
bli_obj_set_conj( conja, a );
bli_obj_set_conj( conjw, w );
bli_obj_set_conj( conjx, x );
// Repeat the experiment n_repeats times and record results.
for ( i = 0; i < n_repeats; ++i )
{
bli_copyv( &y_save, &y );
bli_copyv( &z_save, &z );
time = bli_clock();
libblis_test_dotxaxpyf_impl( iface,
&alpha, &at, &a, &w, &x, &beta, &y, &z,
&cntx );
time_min = bli_clock_min_diff( time_min, time );
}
// Estimate the performance of the best experiment repeat.
*perf = ( 2.0 * m * b_n + 2.0 * m * b_n ) / time_min / FLOPS_PER_UNIT_PERF;
if ( bli_obj_is_complex( y ) ) *perf *= 4.0;
// Perform checks.
libblis_test_dotxaxpyf_check( params, &alpha, &at, &a, &w, &x, &beta, &y, &z, &y_save, &z_save, resid );
// Zero out performance and residual if either output vector is empty.
libblis_test_check_empty_problem( &y, perf, resid );
libblis_test_check_empty_problem( &z, perf, resid );
// Free the test objects.
bli_obj_free( &a );
bli_obj_free( &w );
bli_obj_free( &x );
bli_obj_free( &y );
bli_obj_free( &z );
bli_obj_free( &y_save );
bli_obj_free( &z_save );
// Finalize the context.
bli_dotxaxpyf_cntx_finalize( &cntx );
}
void libblis_test_axpyf_check( obj_t* alpha,
obj_t* a,
obj_t* x,
obj_t* y,
obj_t* y_orig,
double* resid )
{
num_t dt = bli_obj_datatype( *y );
num_t dt_real = bli_obj_datatype_proj_to_real( *y );
dim_t m = bli_obj_vector_dim( *y );
dim_t b_n = bli_obj_width( *a );
dim_t i;
obj_t a1, chi1, v;
obj_t alpha_chi1;
obj_t norm;
double junk;
//
// Pre-conditions:
// - a is randomized.
// - x is randomized.
// - y 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
//
// y := y_orig + alpha * conja(A) * conjx(x)
//
// is functioning correctly if
//
// normf( y - v )
//
// is negligible, where v contains y as computed by repeated calls to
// axpyv.
//
bli_obj_scalar_init_detached( dt_real, &norm );
bli_obj_scalar_init_detached( dt, &alpha_chi1 );
bli_obj_create( dt, m, 1, 0, 0, &v );
bli_copyv( y_orig, &v );
for ( i = 0; i < b_n; ++i )
{
bli_acquire_mpart_l2r( BLIS_SUBPART1, i, 1, a, &a1 );
bli_acquire_vpart_f2b( BLIS_SUBPART1, i, 1, x, &chi1 );
bli_copysc( &chi1, &alpha_chi1 );
bli_mulsc( alpha, &alpha_chi1 );
bli_axpyv( &alpha_chi1, &a1, &v );
}
bli_subv( y, &v );
bli_normfv( &v, &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_gemv_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;
trans_t transa;
conj_t conjx;
obj_t kappa;
obj_t alpha, a, x, beta, y;
obj_t y_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_trans( pc_str[0], &transa );
bli_param_map_char_to_blis_conj( pc_str[1], &conjx );
// 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, n, &a );
libblis_test_vobj_create( params, datatype,
sc_str[1], n, &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( 2.0, 0.0, &alpha );
bli_setsc( -1.0, 0.0, &beta );
}
else
{
bli_setsc( 0.0, 2.0, &alpha );
bli_setsc( 0.0, -1.0, &beta );
}
// Initialize diagonal of matrix A.
bli_setsc( 2.0, -1.0, &kappa );
bli_setm( &BLIS_ZERO, &a );
bli_setd( &kappa, &a );
// Randomize x and y, and save y.
bli_randv( &x );
bli_randv( &y );
bli_copyv( &y, &y_save );
// Apply the parameters.
bli_obj_set_conjtrans( transa, 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_gemv_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 = ( 2.0 * m * n ) / time_min / FLOPS_PER_UNIT_PERF;
if ( bli_obj_is_complex( y ) ) *perf *= 4.0;
// Perform checks.
libblis_test_gemv_check( &kappa, &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_trsm_check
(
test_params_t* params,
side_t side,
obj_t* alpha,
obj_t* a,
obj_t* b,
obj_t* b_orig,
double* resid
)
{
num_t dt = bli_obj_datatype( *b );
num_t dt_real = bli_obj_datatype_proj_to_real( *b );
dim_t m = bli_obj_length( *b );
dim_t n = bli_obj_width( *b );
obj_t norm;
obj_t t, v, w, z;
double junk;
//
// Pre-conditions:
// - a is randomized and triangular.
// - b_orig 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
//
// B := alpha * inv(transa(A)) * B_orig (side = left)
// B := alpha * B_orig * inv(transa(A)) (side = right)
//
// is functioning correctly if
//
// normf( v - z )
//
// is negligible, where
//
// v = B * t
//
// z = ( alpha * inv(transa(A)) * B ) * t (side = left)
// = alpha * inv(transa(A)) * B * t
// = alpha * inv(transa(A)) * w
//
// z = ( alpha * B * inv(transa(A)) ) * t (side = right)
// = alpha * B * tinv(ransa(A)) * t
// = alpha * B * w
bli_obj_scalar_init_detached( dt_real, &norm );
if ( bli_is_left( side ) )
{
bli_obj_create( dt, n, 1, 0, 0, &t );
bli_obj_create( dt, m, 1, 0, 0, &v );
bli_obj_create( dt, m, 1, 0, 0, &w );
bli_obj_create( dt, m, 1, 0, 0, &z );
}
else // else if ( bli_is_left( side ) )
{
bli_obj_create( dt, n, 1, 0, 0, &t );
bli_obj_create( dt, m, 1, 0, 0, &v );
bli_obj_create( dt, n, 1, 0, 0, &w );
bli_obj_create( dt, m, 1, 0, 0, &z );
}
libblis_test_vobj_randomize( params, TRUE, &t );
bli_gemv( &BLIS_ONE, b, &t, &BLIS_ZERO, &v );
if ( bli_is_left( side ) )
{
bli_gemv( alpha, b_orig, &t, &BLIS_ZERO, &w );
bli_trsv( &BLIS_ONE, a, &w );
bli_copyv( &w, &z );
}
else
{
bli_copyv( &t, &w );
bli_trsv( &BLIS_ONE, a, &w );
bli_gemv( alpha, b_orig, &w, &BLIS_ZERO, &z );
}
bli_subv( &z, &v );
bli_normfv( &v, &norm );
bli_getsc( &norm, resid, &junk );
bli_obj_free( &t );
bli_obj_free( &v );
bli_obj_free( &w );
bli_obj_free( &z );
}
void libblis_test_gemv_check( obj_t* kappa,
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_datatype( *y );
num_t dt_real = bli_obj_datatype_proj_to_real( *y );
conj_t conja = bli_obj_conj_status( *a );
dim_t n_x = bli_obj_vector_dim( *x );
dim_t m_y = bli_obj_vector_dim( *y );
dim_t min_m_n = bli_min( m_y, n_x );
obj_t x_temp, y_temp;
obj_t kappac, norm;
obj_t xT_temp, yT_temp, yT;
double junk;
//
// Pre-conditions:
// - a is initialized to kappa along the diagonal.
// - x is randomized.
// - y_orig is randomized.
// Note:
// - alpha, beta, and kappa 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 * transa(A) * conjx(x)
//
// is functioning correctly if
//
// normf( y - z )
//
// is negligible, where
//
// z = beta * y_orig + alpha * conja(kappa) * x
//
bli_obj_scalar_init_detached_copy_of( dt, conja, kappa, &kappac );
bli_obj_scalar_init_detached( dt_real, &norm );
bli_obj_create( dt, n_x, 1, 0, 0, &x_temp );
bli_obj_create( dt, m_y, 1, 0, 0, &y_temp );
bli_copyv( x, &x_temp );
bli_copyv( y_orig, &y_temp );
bli_acquire_vpart_f2b( BLIS_SUBPART1, 0, min_m_n,
&x_temp, &xT_temp );
bli_acquire_vpart_f2b( BLIS_SUBPART1, 0, min_m_n,
&y_temp, &yT_temp );
bli_acquire_vpart_f2b( BLIS_SUBPART1, 0, min_m_n,
y, &yT );
bli_scalv( &kappac, &xT_temp );
bli_scalv( beta, &yT_temp );
bli_axpyv( alpha, &xT_temp, &yT_temp );
bli_subv( &yT_temp, &yT );
bli_normfv( &yT, &norm );
bli_getsc( &norm, resid, &junk );
bli_obj_free( &x_temp );
bli_obj_free( &y_temp );
}
