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 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_scalv_impl
(
iface_t iface,
obj_t* beta,
obj_t* y
)
{
switch ( iface )
{
case BLIS_TEST_SEQ_FRONT_END:
bli_scalv( beta, y );
break;
default:
libblis_test_printf_error( "Invalid interface type.\n" );
}
}
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 alpha, beta, gamma;
obj_t x, y, z, w, a;
num_t dt;
dim_t m, n;
inc_t rs, cs;
//
// This file demonstrates working with vector objects and the level-1v
// operations.
//
//
// Example 1: Create vector objects and then broadcast (copy) scalar
// values to all elements.
//
printf( "\n#\n# -- Example 1 --\n#\n\n" );
// Create a few vectors to work with. We make them all of the same length
// so that we can perform operations between them.
// NOTE: We've chosen to use row vectors here (1x4) instead of column
// vectors (4x1) to allow for easier reading of standard output (less
// scrolling).
dt = BLIS_DOUBLE;
m = 1; n = 4; rs = 0; cs = 0;
bli_obj_create( dt, m, n, rs, cs, &x );
bli_obj_create( dt, m, n, rs, cs, &y );
bli_obj_create( dt, m, n, rs, cs, &z );
bli_obj_create( dt, m, n, rs, cs, &w );
bli_obj_create( dt, m, n, rs, cs, &a );
// Let's also create and initialize some scalar objects.
bli_obj_create_1x1( dt, &alpha );
bli_obj_create_1x1( dt, &beta );
bli_obj_create_1x1( dt, &gamma );
bli_setsc( 2.0, 0.0, &alpha );
bli_setsc( 0.2, 0.0, &beta );
bli_setsc( 3.0, 0.0, &gamma );
bli_printm( "alpha:", &alpha, "%4.1f", "" );
bli_printm( "beta:", &beta, "%4.1f", "" );
bli_printm( "gamma:", &gamma, "%4.1f", "" );
// Vectors can set by "broadcasting" a constant to every element.
bli_setv( &BLIS_ONE, &x );
bli_setv( &alpha, &y );
bli_setv( &BLIS_ZERO, &z );
// Note that we can use printv or printm to print vectors since vectors
// are also matrices. We choose to use printm because it honors the
// orientation of the vector (row or column) when printing, whereas
// printv always prints vectors as column vectors regardless of their
// they are 1 x n or n x 1.
bli_printm( "x := 1.0", &x, "%4.1f", "" );
bli_printm( "y := alpha", &y, "%4.1f", "" );
bli_printm( "z := 0.0", &z, "%4.1f", "" );
//
// Example 2: Randomize a vector object.
//
printf( "\n#\n# -- Example 2 --\n#\n\n" );
// Set a vector to random values.
bli_randv( &w );
bli_printm( "w := randv()", &w, "%4.1f", "" );
//
// Example 3: Perform various element-wise operations on vector objects.
//
printf( "\n#\n# -- Example 3 --\n#\n\n" );
// Copy a vector.
bli_copyv( &w, &a );
bli_printm( "a := w", &a, "%4.1f", "" );
// Add and subtract vectors.
bli_addv( &y, &a );
bli_printm( "a := a + y", &a, "%4.1f", "" );
bli_subv( &w, &a );
bli_printm( "a := a - w", &a, "%4.1f", "" );
// Scale a vector (destructive).
bli_scalv( &beta, &a );
bli_printm( "a := beta * a", &a, "%4.1f", "" );
// Scale a vector (non-destructive).
bli_scal2v( &gamma, &a, &z );
bli_printm( "z := gamma * a", &z, "%4.1f", "" );
// Scale and accumulate between vectors.
bli_axpyv( &alpha, &w, &x );
bli_printm( "x := x + alpha * w", &x, "%4.1f", "" );
bli_xpbyv( &w, &BLIS_MINUS_ONE, &x );
bli_printm( "x := -1.0 * x + w", &x, "%4.1f", "" );
// Invert a vector element-wise.
bli_invertv( &y );
bli_printm( "y := 1 / y", &y, "%4.1f", "" );
// Swap two vectors.
bli_swapv( &x, &y );
bli_printm( "x (after swapping with y)", &x, "%4.1f", "" );
bli_printm( "y (after swapping with x)", &y, "%4.1f", "" );
//
// Example 4: Perform contraction-like operations on vector objects.
//
printf( "\n#\n# -- Example 4 --\n#\n\n" );
// Perform a dot product.
bli_dotv( &a, &z, &gamma );
bli_printm( "gamma := a * z (dot product)", &gamma, "%5.2f", "" );
// Perform an extended dot product.
bli_dotxv( &alpha, &a, &z, &BLIS_ONE, &gamma );
bli_printm( "gamma := 1.0 * gamma + alpha * a * z (accumulate scaled dot product)", &gamma, "%5.2f", "" );
// Free the objects.
bli_obj_free( &alpha );
bli_obj_free( &beta );
bli_obj_free( &gamma );
bli_obj_free( &x );
bli_obj_free( &y );
bli_obj_free( &z );
bli_obj_free( &w );
bli_obj_free( &a );
return 0;
}
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_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 libblis_test_trmm3_check( side_t side,
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 n = bli_obj_width( *c );
obj_t kappa, norm;
obj_t t, v, w, z;
double junk;
//
// Pre-conditions:
// - a is randomized and triangular.
// - b is randomized.
// - c_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
//
// C := beta * C_orig + alpha * transa(A) * transb(B) (side = left)
// C := beta * C_orig + alpha * transb(B) * transa(A) (side = right)
//
// is functioning correctly if
//
// fnorm( v - z )
//
// is negligible, where
//
// v = C * t
//
// z = ( beta * C_orig + alpha * transa(A) * transb(B) ) * t (side = left)
// = beta * C_orig * t + alpha * transa(A) * transb(B) * t
// = beta * C_orig * t + alpha * transa(A) * w
// = beta * C_orig * t + z
//
// z = ( beta * C_orig + alpha * transb(B) * transa(A) ) * t (side = right)
// = beta * C_orig * t + alpha * transb(B) * transa(A) * t
// = beta * C_orig * t + alpha * transb(B) * w
// = beta * C_orig * t + z
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, 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 );
}
bli_randv( &t );
bli_setsc( 1.0/( double )n, 0.0, &kappa );
bli_scalv( &kappa, &t );
bli_gemv( &BLIS_ONE, c, &t, &BLIS_ZERO, &v );
if ( bli_is_left( side ) )
{
bli_gemv( &BLIS_ONE, b, &t, &BLIS_ZERO, &w );
bli_trmv( alpha, a, &w );
bli_copyv( &w, &z );
}
else
{
bli_copyv( &t, &w );
bli_trmv( &BLIS_ONE, a, &w );
bli_gemv( alpha, b, &w, &BLIS_ZERO, &z );
}
bli_gemv( beta, c_orig, &t, &BLIS_ONE, &z );
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_syr2k_check( 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 at, bt;
obj_t kappa, norm;
obj_t t, v, w1, w2, z;
double junk;
//
// Pre-conditions:
// - a is randomized.
// - b is randomized.
// - c_orig is randomized and symmetric.
// 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
//
// C := beta * C_orig + alpha * transa(A) * transb(B)^T + alpha * transb(B) * transa(A)^T
//
// is functioning correctly if
//
// normf( v - z )
//
// is negligible, where
//
// v = C * t
// z = ( beta * C_orig + alpha * transa(A) * transb(B)^T + alpha * transb(B) * transa(A)^T ) * t
// = beta * C_orig * t + alpha * transa(A) * transb(B)^T * t + alpha * transb(B) * transa(A)^T * t
// = beta * C_orig * t + alpha * transa(A) * transb(B)^T * t + alpha * transb(B) * w2
// = beta * C_orig * t + alpha * transa(A) * w1 + alpha * transb(B) * w2
// = beta * C_orig * t + alpha * transa(A) * w1 + z
// = beta * C_orig * t + z
//
bli_obj_alias_with_trans( BLIS_TRANSPOSE, *a, at );
bli_obj_alias_with_trans( BLIS_TRANSPOSE, *b, bt );
bli_obj_scalar_init_detached( dt, &kappa );
bli_obj_scalar_init_detached( dt_real, &norm );
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 );
bli_randv( &t );
bli_setsc( 1.0/( double )m, 0.0, &kappa );
bli_scalv( &kappa, &t );
bli_symv( &BLIS_ONE, c, &t, &BLIS_ZERO, &v );
bli_gemv( &BLIS_ONE, &at, &t, &BLIS_ZERO, &w2 );
bli_gemv( &BLIS_ONE, &bt, &t, &BLIS_ZERO, &w1 );
bli_gemv( alpha, a, &w1, &BLIS_ZERO, &z );
bli_gemv( alpha, b, &w2, &BLIS_ONE, &z );
bli_symv( 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_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 libblis_test_herk_check( obj_t* alpha,
obj_t* a,
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 ah;
obj_t kappa, norm;
obj_t t, v, w, z;
double junk;
//
// Pre-conditions:
// - a is randomized.
// - c_orig is randomized and Hermitian.
// Note:
// - alpha and beta must be real-valued.
//
// Under these conditions, we assume that the implementation for
//
// C := beta * C_orig + alpha * transa(A) * transa(A)^H
//
// is functioning correctly if
//
// fnorm( v - z )
//
// is negligible, where
//
// v = C * t
// z = ( beta * C_orig + alpha * transa(A) * transa(A)^H ) * t
// = beta * C_orig * t + alpha * transa(A) * transa(A)^H * t
// = beta * C_orig * t + alpha * transa(A) * w
// = beta * C_orig * t + z
//
bli_obj_alias_with_trans( BLIS_CONJ_TRANSPOSE, *a, ah );
bli_obj_scalar_init_detached( dt, &kappa );
bli_obj_scalar_init_detached( dt_real, &norm );
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, &w );
bli_obj_create( dt, m, 1, 0, 0, &z );
bli_randv( &t );
bli_setsc( 1.0/( double )m, 0.0, &kappa );
bli_scalv( &kappa, &t );
bli_hemv( &BLIS_ONE, c, &t, &BLIS_ZERO, &v );
bli_gemv( &BLIS_ONE, &ah, &t, &BLIS_ZERO, &w );
bli_gemv( alpha, a, &w, &BLIS_ZERO, &z );
bli_hemv( beta, c_orig, &t, &BLIS_ONE, &z );
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_hemv_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 conja;
conj_t conjx;
obj_t kappa;
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 );
bli_obj_scalar_init_detached( 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[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( 0.0, 1.0, &alpha );
bli_setsc( 0.0, -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 x and y, and save y.
bli_randv( &x );
bli_randv( &y );
bli_copyv( &y, &y_save );
// Normalize vectors by m.
bli_setsc( 1.0/( double )m, 0.0, &kappa );
bli_scalv( &kappa, &x );
bli_scalv( &kappa, &y );
bli_scalv( &kappa, &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_hemv_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_hemv_check( &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 );
}