void libblis_test_subv_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 aminusb;
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
//
// normfv(y) - sqrt( absqsc( beta - conjx(alpha) ) * m )
//
// is negligible.
//
bli_obj_scalar_init_detached( dt, &aminusb );
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_normfv( y, &norm_r );
bli_copysc( beta, &aminusb );
bli_subsc( &alpha_conj, &aminusb );
bli_setsc( ( double )m, 0.0, &m_r );
bli_absqsc( &aminusb, &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 bli_her2_basic_check( conj_t conjh,
obj_t* alpha,
obj_t* x,
obj_t* y,
obj_t* c )
{
err_t e_val;
// Check object datatypes.
e_val = bli_check_noninteger_object( alpha );
bli_check_error_code( e_val );
e_val = bli_check_floating_object( x );
bli_check_error_code( e_val );
e_val = bli_check_floating_object( y );
bli_check_error_code( e_val );
e_val = bli_check_floating_object( c );
bli_check_error_code( e_val );
// Check object dimensions.
e_val = bli_check_scalar_object( alpha );
bli_check_error_code( e_val );
e_val = bli_check_vector_object( x );
bli_check_error_code( e_val );
e_val = bli_check_vector_object( y );
bli_check_error_code( e_val );
e_val = bli_check_matrix_object( c );
bli_check_error_code( e_val );
e_val = bli_check_square_object( c );
bli_check_error_code( e_val );
e_val = bli_check_object_length_equals( c, bli_obj_vector_dim( *x ) );
bli_check_error_code( e_val );
e_val = bli_check_object_length_equals( c, bli_obj_vector_dim( *y ) );
bli_check_error_code( e_val );
}
void libblis_test_axpyv_check
(
test_params_t* params,
obj_t* alpha,
obj_t* x,
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 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 := y_orig + alpha * conjx(x)
//
// is functioning correctly if
//
// normf( y - ( 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_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 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 );
}
err_t bli_check_vector_dim_equals( obj_t* a, dim_t n )
{
err_t e_val = BLIS_SUCCESS;
if ( bli_obj_vector_dim( *a ) != n )
e_val = BLIS_UNEXPECTED_VECTOR_DIM;
return e_val;
}
void bli_dotxv_unb_var1( obj_t* alpha,
obj_t* x,
obj_t* y,
obj_t* beta,
obj_t* rho )
{
num_t dt_x = bli_obj_datatype( *x );
num_t dt_y = bli_obj_datatype( *y );
num_t dt_rho = bli_obj_datatype( *rho );
conj_t conjx = bli_obj_conj_status( *x );
conj_t conjy = bli_obj_conj_status( *y );
dim_t n = bli_obj_vector_dim( *x );
inc_t inc_x = bli_obj_vector_inc( *x );
void* buf_x = bli_obj_buffer_at_off( *x );
inc_t inc_y = bli_obj_vector_inc( *y );
void* buf_y = bli_obj_buffer_at_off( *y );
void* buf_rho = bli_obj_buffer_at_off( *rho );
num_t dt_alpha;
void* buf_alpha;
num_t dt_beta;
void* buf_beta;
FUNCPTR_T f;
// The datatype of alpha MUST be the type union of x and y. This is to
// prevent any unnecessary loss of information during computation.
dt_alpha = bli_datatype_union( dt_x, dt_y );
buf_alpha = bli_obj_scalar_buffer( dt_alpha, *alpha );
// The datatype of beta MUST be the same as the datatype of rho.
dt_beta = dt_rho;
buf_beta = bli_obj_scalar_buffer( dt_beta, *beta );
// Index into the type combination array to extract the correct
// function pointer.
f = ftypes[dt_x][dt_y][dt_rho];
// Invoke the function.
f( conjx,
conjy,
n,
buf_alpha,
buf_x, inc_x,
buf_y, inc_y,
buf_beta,
buf_rho );
}
void bli_dotaxpyv_kernel( obj_t* alpha,
obj_t* xt,
obj_t* x,
obj_t* y,
obj_t* rho,
obj_t* z )
{
num_t dt_x = bli_obj_datatype( *x );
num_t dt_y = bli_obj_datatype( *y );
num_t dt_z = bli_obj_datatype( *z );
conj_t conjxt = bli_obj_conj_status( *xt );
conj_t conjx = bli_obj_conj_status( *x );
conj_t conjy = bli_obj_conj_status( *y );
dim_t n = bli_obj_vector_dim( *x );
inc_t inc_x = bli_obj_vector_inc( *x );
void* buf_x = bli_obj_buffer_at_off( *x );
inc_t inc_y = bli_obj_vector_inc( *y );
void* buf_y = bli_obj_buffer_at_off( *y );
inc_t inc_z = bli_obj_vector_inc( *z );
void* buf_z = bli_obj_buffer_at_off( *z );
void* buf_rho = bli_obj_buffer_at_off( *rho );
num_t dt_alpha;
void* buf_alpha;
FUNCPTR_T f;
// If alpha is a scalar constant, use dt_x to extract the address of the
// corresponding constant value; otherwise, use the datatype encoded
// within the alpha object and extract the buffer at the alpha offset.
bli_set_scalar_dt_buffer( alpha, dt_x, dt_alpha, buf_alpha );
// Index into the type combination array to extract the correct
// function pointer.
f = ftypes[dt_x][dt_y][dt_z];
// Invoke the function.
f( conjxt,
conjx,
conjy,
n,
buf_alpha,
buf_x, inc_x,
buf_y, inc_y,
buf_rho,
buf_z, inc_z );
}
void bli_hemv_basic_check( obj_t* alpha,
obj_t* a,
obj_t* x,
obj_t* beta,
obj_t* y )
{
err_t e_val;
// Check object datatypes.
e_val = bli_check_noninteger_object( alpha );
bli_check_error_code( e_val );
e_val = bli_check_noninteger_object( beta );
bli_check_error_code( e_val );
e_val = bli_check_floating_object( a );
bli_check_error_code( e_val );
e_val = bli_check_floating_object( x );
bli_check_error_code( e_val );
e_val = bli_check_floating_object( y );
bli_check_error_code( e_val );
// Check object dimensions.
e_val = bli_check_scalar_object( alpha );
bli_check_error_code( e_val );
e_val = bli_check_scalar_object( beta );
bli_check_error_code( e_val );
e_val = bli_check_matrix_object( a );
bli_check_error_code( e_val );
e_val = bli_check_square_object( a );
bli_check_error_code( e_val );
e_val = bli_check_vector_object( x );
bli_check_error_code( e_val );
e_val = bli_check_vector_object( y );
bli_check_error_code( e_val );
e_val = bli_check_object_length_equals( a, bli_obj_vector_dim( *x ) );
bli_check_error_code( e_val );
e_val = bli_check_equal_vector_lengths( x, y );
bli_check_error_code( e_val );
}
void libblis_test_normfv_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_vector_dim( *x );
obj_t m_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 )
//
// where m is the length of x.
//
bli_obj_scalar_init_detached( dt_real, &temp_r );
bli_obj_scalar_init_detached( dt_real, &m_r );
bli_setsc( ( double )m, 0.0, &m_r );
bli_absqsc( beta, &temp_r );
bli_mulsc( &m_r, &temp_r );
bli_sqrtsc( &temp_r, &temp_r );
bli_subsc( &temp_r, norm );
bli_getsc( norm, resid, &junk );
}
void bli_randv_unb_var1( obj_t* x )
{
num_t dt_x = bli_obj_datatype( *x );
dim_t n = bli_obj_vector_dim( *x );
inc_t inc_x = bli_obj_vector_inc( *x );
void* buf_x = bli_obj_buffer_at_off( *x );
FUNCPTR_T f;
// Index into the type combination array to extract the correct
// function pointer.
f = ftypes[dt_x];
// Invoke the function.
f( n,
buf_x, inc_x );
}
void bli_axpy2v_ker( obj_t* alpha1,
obj_t* alpha2,
obj_t* x,
obj_t* y,
obj_t* z )
{
num_t dt = bli_obj_datatype( *z );
conj_t conjx = bli_obj_conj_status( *x );
conj_t conjy = bli_obj_conj_status( *y );
dim_t n = bli_obj_vector_dim( *x );
inc_t inc_x = bli_obj_vector_inc( *x );
void* buf_x = bli_obj_buffer_at_off( *x );
inc_t inc_y = bli_obj_vector_inc( *y );
void* buf_y = bli_obj_buffer_at_off( *y );
inc_t inc_z = bli_obj_vector_inc( *z );
void* buf_z = bli_obj_buffer_at_off( *z );
void* buf_alpha1 = bli_obj_buffer_for_1x1( dt, *alpha1 );
void* buf_alpha2 = bli_obj_buffer_for_1x1( dt, *alpha2 );
FUNCPTR_T f;
// Index into the type combination array to extract the correct
// function pointer.
f = ftypes[dt];
// Invoke the function.
f( conjx,
conjy,
n,
buf_alpha1,
buf_alpha2,
buf_x, inc_x,
buf_y, inc_y,
buf_z, inc_z );
}
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 bli_fprintv( FILE* file, char* s1, obj_t* x, char* format, char* s2 )
{
num_t dt_x = bli_obj_datatype( *x );
dim_t n = bli_obj_vector_dim( *x );
inc_t inc_x = bli_obj_vector_inc( *x );
void* buf_x = bli_obj_buffer_at_off( *x );
FUNCPTR_T f;
// Index into the type combination array to extract the correct
// function pointer.
f = ftypes[dt_x];
// Invoke the function.
f( file,
s1,
n,
buf_x, inc_x,
format,
s2 );
}
void bli_scal2v_unb_var1( obj_t* beta,
obj_t* x,
obj_t* y )
{
num_t dt_x = bli_obj_datatype( *x );
num_t dt_y = bli_obj_datatype( *y );
conj_t conjx = bli_obj_conj_status( *x );
dim_t n = bli_obj_vector_dim( *x );
inc_t inc_x = bli_obj_vector_inc( *x );
void* buf_x = bli_obj_buffer_at_off( *x );
inc_t inc_y = bli_obj_vector_inc( *y );
void* buf_y = bli_obj_buffer_at_off( *y );
num_t dt_beta;
void* buf_beta;
FUNCPTR_T f;
// If beta is a scalar constant, use dt_x to extract the address of the
// corresponding constant value; otherwise, use the datatype encoded
// within the beta object and extract the buffer at the beta offset.
bli_set_scalar_dt_buffer( beta, dt_x, dt_beta, buf_beta );
// Index into the type combination array to extract the correct
// function pointer.
f = ftypes[dt_beta][dt_x][dt_y];
// Invoke the function.
f( conjx,
n,
buf_beta,
buf_x, inc_x,
buf_y, inc_y );
}
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_randv_check( obj_t* x,
double* resid )
{
dim_t m_x = bli_obj_vector_dim( *x );
inc_t inc_x = bli_obj_vector_inc( *x );
void* buf_x = bli_obj_buffer_at_off( *x );
*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.
//
if ( bli_obj_is_float( *x ) )
{
float sum_x;
bli_sabsumv( m_x,
buf_x, inc_x,
&sum_x );
if ( sum_x == *bli_s0 ) *resid = 1.0;
else if ( sum_x >= 1.0 * m_x ) *resid = 2.0;
}
else if ( bli_obj_is_double( *x ) )
{
double sum_x;
bli_dabsumv( m_x,
buf_x, inc_x,
&sum_x );
if ( sum_x == *bli_d0 ) *resid = 1.0;
else if ( sum_x >= 1.0 * m_x ) *resid = 2.0;
}
else if ( bli_obj_is_scomplex( *x ) )
{
float sum_x;
bli_cabsumv( m_x,
buf_x, inc_x,
&sum_x );
if ( sum_x == *bli_s0 ) *resid = 1.0;
else if ( sum_x >= 2.0 * m_x ) *resid = 2.0;
}
else // if ( bli_obj_is_dcomplex( *x ) )
{
double sum_x;
bli_zabsumv( m_x,
buf_x, inc_x,
&sum_x );
if ( sum_x == *bli_d0 ) *resid = 1.0;
else if ( sum_x >= 2.0 * m_x ) *resid = 2.0;
}
}
//
// Define object-based interface.
//
void bli_dotxf( obj_t* alpha,
obj_t* a,
obj_t* x,
obj_t* beta,
obj_t* y )
{
num_t dt = bli_obj_datatype( *x );
conj_t conja = bli_obj_conj_status( *a );
conj_t conjx = bli_obj_conj_status( *x );
dim_t m = bli_obj_vector_dim( *y );
dim_t b_n = bli_obj_vector_dim( *x );
void* buf_a = bli_obj_buffer_at_off( *a );
inc_t rs_a = bli_obj_row_stride( *a );
inc_t cs_a = bli_obj_col_stride( *a );
void* buf_x = bli_obj_buffer_at_off( *x );
inc_t inc_x = bli_obj_vector_inc( *x );
void* buf_y = bli_obj_buffer_at_off( *y );
inc_t inc_y = bli_obj_vector_inc( *y );
obj_t alpha_local;
void* buf_alpha;
obj_t beta_local;
void* buf_beta;
FUNCPTR_T f = ftypes[dt];
if ( bli_error_checking_is_enabled() )
bli_dotxf_check( alpha, a, x, beta, y );
// Create local copy-casts of the scalars (and apply internal conjugation
// if needed).
bli_obj_scalar_init_detached_copy_of( dt,
BLIS_NO_CONJUGATE,
alpha,
&alpha_local );
bli_obj_scalar_init_detached_copy_of( dt,
BLIS_NO_CONJUGATE,
beta,
&beta_local );
// Extract the scalar buffers.
buf_alpha = bli_obj_buffer_for_1x1( dt, alpha_local );
buf_beta = bli_obj_buffer_for_1x1( dt, beta_local );
// Support cases where matrix A requires a transposition.
if ( bli_obj_has_trans( *a ) ) { bli_swap_incs( rs_a, cs_a ); }
// Invoke the void pointer-based function.
f( conja,
conjx,
m,
b_n,
buf_alpha,
buf_a, rs_a, cs_a,
buf_x, inc_x,
buf_beta,
buf_y, inc_y );
}
void libblis_test_symv_check
(
test_params_t* params,
obj_t* alpha,
obj_t* a,
obj_t* x,
obj_t* beta,
obj_t* y,
obj_t* y_orig,
double* resid
)
{
num_t dt = bli_obj_dt( y );
num_t dt_real = bli_obj_dt_proj_to_real( y );
dim_t m = bli_obj_vector_dim( y );
obj_t v;
obj_t norm;
double junk;
//
// Pre-conditions:
// - a is randomized and symmetric.
// - x is randomized.
// - y_orig is randomized.
// Note:
// - alpha and beta should have non-zero imaginary components in the
// complex cases in order to more fully exercise the implementation.
//
// Under these conditions, we assume that the implementation for
//
// y := beta * y_orig + alpha * conja(A) * conjx(x)
//
// is functioning correctly if
//
// normf( y - v )
//
// is negligible, where
//
// v = beta * y_orig + alpha * conja(A_dense) * x
//
bli_obj_scalar_init_detached( dt_real, &norm );
bli_obj_create( dt, m, 1, 0, 0, &v );
bli_copyv( y_orig, &v );
bli_mksymm( a );
bli_obj_set_struc( BLIS_GENERAL, a );
bli_obj_set_uplo( BLIS_DENSE, a );
bli_gemv( alpha, a, x, beta, &v );
bli_subv( &v, y );
bli_normfv( y, &norm );
bli_getsc( &norm, resid, &junk );
bli_obj_free( &v );
}
void libblis_test_dotaxpyv_check( obj_t* alpha,
obj_t* xt,
obj_t* x,
obj_t* y,
obj_t* rho,
obj_t* z,
obj_t* z_orig,
double* resid )
{
num_t dt = bli_obj_datatype( *z );
num_t dt_real = bli_obj_datatype_proj_to_real( *z );
dim_t m = bli_obj_vector_dim( *z );
obj_t rho_temp;
obj_t z_temp;
obj_t norm_z;
double resid1, resid2;
double junk;
//
// Pre-conditions:
// - x is randomized.
// - y is randomized.
// - z_orig is randomized.
// - xt is an alias to x.
// 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
//
// rho := conjxt(x^T) conjy(y)
// z := z_orig + alpha * conjx(x)
//
// is functioning correctly if
//
// ( rho - rho_temp )
//
// and
//
// normf( z - z_temp )
//
// are negligible, where rho_temp and z_temp contain rho and z as
// computed by dotv and axpyv, respectively.
//
bli_obj_scalar_init_detached( dt, &rho_temp );
bli_obj_scalar_init_detached( dt_real, &norm_z );
bli_obj_create( dt, m, 1, 0, 0, &z_temp );
bli_copyv( z_orig, &z_temp );
bli_dotv( xt, y, &rho_temp );
bli_axpyv( alpha, x, &z_temp );
bli_subsc( rho, &rho_temp );
bli_getsc( &rho_temp, &resid1, &junk );
bli_subv( &z_temp, z );
bli_normfv( z, &norm_z );
bli_getsc( &norm_z, &resid2, &junk );
*resid = bli_fmaxabs( resid1, resid2 );
bli_obj_free( &z_temp );
}
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_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_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 );
}
void bli_dotxaxpyf_check( obj_t* alpha,
obj_t* at,
obj_t* a,
obj_t* w,
obj_t* x,
obj_t* beta,
obj_t* y,
obj_t* z )
{
err_t e_val;
// Check object datatypes.
e_val = bli_check_noninteger_object( alpha );
bli_check_error_code( e_val );
e_val = bli_check_floating_object( at );
bli_check_error_code( e_val );
e_val = bli_check_floating_object( a );
bli_check_error_code( e_val );
e_val = bli_check_floating_object( w );
bli_check_error_code( e_val );
e_val = bli_check_floating_object( x );
bli_check_error_code( e_val );
e_val = bli_check_noninteger_object( beta );
bli_check_error_code( e_val );
e_val = bli_check_floating_object( y );
bli_check_error_code( e_val );
e_val = bli_check_floating_object( z );
bli_check_error_code( e_val );
// Check object dimensions.
e_val = bli_check_scalar_object( alpha );
bli_check_error_code( e_val );
e_val = bli_check_matrix_object( at );
bli_check_error_code( e_val );
e_val = bli_check_matrix_object( a );
bli_check_error_code( e_val );
e_val = bli_check_vector_object( w );
bli_check_error_code( e_val );
e_val = bli_check_vector_object( x );
bli_check_error_code( e_val );
e_val = bli_check_scalar_object( beta );
bli_check_error_code( e_val );
e_val = bli_check_vector_object( y );
bli_check_error_code( e_val );
e_val = bli_check_vector_object( z );
bli_check_error_code( e_val );
e_val = bli_check_equal_vector_lengths( w, z );
bli_check_error_code( e_val );
e_val = bli_check_equal_vector_lengths( x, y );
bli_check_error_code( e_val );
e_val = bli_check_conformal_dims( at, a );
bli_check_error_code( e_val );
e_val = bli_check_object_length_equals( at, bli_obj_vector_dim( *w ) );
bli_check_error_code( e_val );
e_val = bli_check_object_width_equals( at, bli_obj_vector_dim( *y ) );
bli_check_error_code( e_val );
e_val = bli_check_object_length_equals( a, bli_obj_vector_dim( *z ) );
bli_check_error_code( e_val );
e_val = bli_check_object_width_equals( a, bli_obj_vector_dim( *x ) );
bli_check_error_code( e_val );
// Check object aliases.
e_val = bli_check_object_alias_of( at, a );
bli_check_error_code( e_val );
// Check object buffers (for non-NULLness).
e_val = bli_check_object_buffer( alpha );
bli_check_error_code( e_val );
e_val = bli_check_object_buffer( at );
bli_check_error_code( e_val );
e_val = bli_check_object_buffer( a );
bli_check_error_code( e_val );
e_val = bli_check_object_buffer( w );
bli_check_error_code( e_val );
e_val = bli_check_object_buffer( x );
bli_check_error_code( e_val );
e_val = bli_check_object_buffer( beta );
bli_check_error_code( e_val );
e_val = bli_check_object_buffer( y );
bli_check_error_code( e_val );
e_val = bli_check_object_buffer( z );
bli_check_error_code( e_val );
}
siz_t bli_packv_init_pack
(
pack_t schema,
bszid_t bmult_id,
obj_t* a,
obj_t* p,
cntx_t* cntx
)
{
num_t dt = bli_obj_dt( a );
dim_t dim_a = bli_obj_vector_dim( a );
dim_t bmult = bli_cntx_get_blksz_def_dt( dt, bmult_id, cntx );
membrk_t* membrk = bli_cntx_membrk( cntx );
#if 0
mem_t* mem_p;
#endif
dim_t m_p_pad;
siz_t size_p;
inc_t rs_p, cs_p;
void* buf;
// We begin by copying the basic fields of c.
bli_obj_alias_to( a, p );
// Update the dimensions.
bli_obj_set_dims( dim_a, 1, p );
// Reset the view offsets to (0,0).
bli_obj_set_offs( 0, 0, p );
// Set the pack schema in the p object to the value in the control tree
// node.
bli_obj_set_pack_schema( schema, p );
// Compute the dimensions padded by the dimension multiples.
m_p_pad = bli_align_dim_to_mult( bli_obj_vector_dim( p ), bmult );
// Compute the size of the packed buffer.
size_p = m_p_pad * 1 * bli_obj_elem_size( p );
#if 0
// Extract the address of the mem_t object within p that will track
// properties of the packed buffer.
mem_p = bli_obj_pack_mem( *p );
if ( bli_mem_is_unalloc( mem_p ) )
{
// If the mem_t object of p has not yet been allocated, then acquire
// a memory block suitable for a vector.
bli_membrk_acquire_v( membrk,
size_p,
mem_p );
}
else
{
// If the mem_t object has already been allocated, then release and
// re-acquire the memory so there is sufficient space.
if ( bli_mem_size( mem_p ) < size_p )
{
bli_membrk_release( mem_p );
bli_membrk_acquire_v( membrk,
size_p,
mem_p );
}
}
// Grab the buffer address from the mem_t object and copy it to the
// main object buffer field. (Sometimes this buffer address will be
// copied when the value is already up-to-date, because it persists
// in the main object buffer field across loop iterations.)
buf = bli_mem_buffer( mem_p );
bli_obj_set_buffer( buf, p );
#endif
// Save the padded (packed) dimensions into the packed object.
bli_obj_set_padded_dims( m_p_pad, 1, p );
// Set the row and column strides of p based on the pack schema.
if ( schema == BLIS_PACKED_VECTOR )
{
// Set the strides to reflect a column-stored vector. Note that the
// column stride may never be used, and is only useful to determine
// how much space beyond the vector would need to be zero-padded, if
// zero-padding was needed.
rs_p = 1;
cs_p = bli_obj_padded_length( p );
bli_obj_set_strides( rs_p, cs_p, p );
}
return size_p;
}
void libblis_test_setv_check( obj_t* beta,
obj_t* x,
double* resid )
{
num_t dt_x = bli_obj_datatype( *x );
dim_t m_x = bli_obj_vector_dim( *x );
inc_t inc_x = bli_obj_vector_inc( *x );
void* buf_x = bli_obj_buffer_at_off( *x );
void* buf_beta = bli_obj_buffer_for_1x1( dt_x, *beta );
dim_t i;
*resid = 0.0;
//
// The easiest way to check that setv was successful is to confirm
// that each element of x is equal to beta.
//
if ( bli_obj_is_float( *x ) )
{
float* chi1 = buf_x;
float* beta_cast = buf_beta;
for ( i = 0; i < m_x; ++i )
{
if ( !bli_seq( *chi1, *beta_cast ) ) { *resid = 1.0; return; }
chi1 += inc_x;
}
}
else if ( bli_obj_is_double( *x ) )
{
double* chi1 = buf_x;
double* beta_cast = buf_beta;
for ( i = 0; i < m_x; ++i )
{
if ( !bli_deq( *chi1, *beta_cast ) ) { *resid = 1.0; return; }
chi1 += inc_x;
}
}
else if ( bli_obj_is_scomplex( *x ) )
{
scomplex* chi1 = buf_x;
scomplex* beta_cast = buf_beta;
for ( i = 0; i < m_x; ++i )
{
if ( !bli_ceq( *chi1, *beta_cast ) ) { *resid = 1.0; return; }
chi1 += inc_x;
}
}
else // if ( bli_obj_is_dcomplex( *x ) )
{
dcomplex* chi1 = buf_x;
dcomplex* beta_cast = buf_beta;
for ( i = 0; i < m_x; ++i )
{
if ( !bli_zeq( *chi1, *beta_cast ) ) { *resid = 1.0; return; }
chi1 += inc_x;
}
}
}
void libblis_test_axpy2v_check( obj_t* alpha1,
obj_t* alpha2,
obj_t* x,
obj_t* y,
obj_t* z,
obj_t* z_orig,
double* resid )
{
num_t dt = bli_obj_datatype( *z );
num_t dt_real = bli_obj_datatype_proj_to_real( *z );
dim_t m = bli_obj_vector_dim( *z );
obj_t x_temp, y_temp, z_temp;
obj_t norm;
double junk;
//
// Pre-conditions:
// - x is randomized.
// - y is randomized.
// - z_orig is randomized.
// Note:
// - alpha1, alpha2 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
//
// z := z_orig + alpha1 * conjx(x) + alpha2 * conjy(y)
//
// is functioning correctly if
//
// normf( z - v )
//
// is negligible, where v contains z as computed by two calls to axpyv.
//
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_obj_create( dt, m, 1, 0, 0, &z_temp );
bli_copyv( x, &x_temp );
bli_copyv( y, &y_temp );
bli_copyv( z_orig, &z_temp );
bli_scalv( alpha1, &x_temp );
bli_scalv( alpha2, &y_temp );
bli_addv( &x_temp, &z_temp );
bli_addv( &y_temp, &z_temp );
bli_subv( &z_temp, z );
bli_normfv( z, &norm );
bli_getsc( &norm, resid, &junk );
bli_obj_free( &x_temp );
bli_obj_free( &y_temp );
bli_obj_free( &z_temp );
}
void bli_dotxaxpyf_unb_var2( obj_t* alpha,
obj_t* at,
obj_t* a,
obj_t* w,
obj_t* x,
obj_t* beta,
obj_t* y,
obj_t* z )
{
num_t dt_a = bli_obj_datatype( *a );
num_t dt_x = bli_obj_datatype( *x );
num_t dt_y = bli_obj_datatype( *y );
conj_t conjat = bli_obj_conj_status( *at );
conj_t conja = bli_obj_conj_status( *a );
conj_t conjw = bli_obj_conj_status( *w );
conj_t conjx = bli_obj_conj_status( *x );
dim_t m = bli_obj_vector_dim( *z );
dim_t b_n = bli_obj_vector_dim( *y );
void* buf_a = bli_obj_buffer_at_off( *a );
inc_t rs_a = bli_obj_row_stride( *a );
inc_t cs_a = bli_obj_col_stride( *a );
inc_t inc_w = bli_obj_vector_inc( *w );
void* buf_w = bli_obj_buffer_at_off( *w );
inc_t inc_x = bli_obj_vector_inc( *x );
void* buf_x = bli_obj_buffer_at_off( *x );
inc_t inc_y = bli_obj_vector_inc( *y );
void* buf_y = bli_obj_buffer_at_off( *y );
inc_t inc_z = bli_obj_vector_inc( *z );
void* buf_z = bli_obj_buffer_at_off( *z );
num_t dt_alpha;
void* buf_alpha;
num_t dt_beta;
void* buf_beta;
FUNCPTR_T f;
// The datatype of alpha MUST be the type union of a and x. This is to
// prevent any unnecessary loss of information during computation.
dt_alpha = bli_datatype_union( dt_a, dt_x );
buf_alpha = bli_obj_buffer_for_1x1( dt_alpha, *alpha );
// The datatype of beta MUST be the same as the datatype of y.
dt_beta = dt_y;
buf_beta = bli_obj_buffer_for_1x1( dt_beta, *beta );
// Index into the type combination array to extract the correct
// function pointer.
f = ftypes[dt_a][dt_x][dt_y];
// Invoke the function.
f( conjat,
conja,
conjw,
conjx,
m,
b_n,
buf_alpha,
buf_a, rs_a, cs_a,
buf_w, inc_w,
buf_x, inc_x,
buf_beta,
buf_y, inc_y,
buf_z, inc_z );
}