|
libflame
revision_anchor
|
Functions | |
| FLA_Error | FLA_Apply_H2_UT_l_opt_var1 (FLA_Obj tau, FLA_Obj u2, FLA_Obj a1t, FLA_Obj A2) |
| FLA_Error | FLA_Apply_H2_UT_l_ops_var1 (int m_u2_A2, int n_a1t, float *tau, float *u2, int inc_u2, float *a1t, int inc_a1t, float *A2, int rs_A2, int cs_A2) |
| FLA_Error | FLA_Apply_H2_UT_l_opd_var1 (int m_u2_A2, int n_a1t, double *tau, double *u2, int inc_u2, double *a1t, int inc_a1t, double *A2, int rs_A2, int cs_A2) |
| FLA_Error | FLA_Apply_H2_UT_l_opc_var1 (int m_u2_A2, int n_a1t, scomplex *tau, scomplex *u2, int inc_u2, scomplex *a1t, int inc_a1t, scomplex *A2, int rs_A2, int cs_A2) |
| FLA_Error | FLA_Apply_H2_UT_l_opz_var1 (int m_u2_A2, int n_a1t, dcomplex *tau, dcomplex *u2, int inc_u2, dcomplex *a1t, int inc_a1t, dcomplex *A2, int rs_A2, int cs_A2) |
| FLA_Error FLA_Apply_H2_UT_l_opc_var1 | ( | int | m_u2_A2, |
| int | n_a1t, | ||
| scomplex * | tau, | ||
| scomplex * | u2, | ||
| int | inc_u2, | ||
| scomplex * | a1t, | ||
| int | inc_a1t, | ||
| scomplex * | A2, | ||
| int | rs_A2, | ||
| int | cs_A2 | ||
| ) |
References bli_caxpyv(), bli_ccopyv(), bli_cgemv(), bli_cger(), bli_cinvscalv(), BLIS_CONJUGATE, BLIS_NO_CONJUGATE, BLIS_TRANSPOSE, FLA_free(), FLA_malloc(), FLA_MINUS_ONE, and FLA_ONE.
Referenced by FLA_Apply_H2_UT_l_opt_var1(), FLA_Bidiag_UT_u_step_opc_var1(), FLA_CAQR2_UT_opc_var1(), FLA_Hess_UT_step_opc_var1(), FLA_QR2_UT_opc_var1(), FLA_QR_UT_form_Q_opc_var1(), FLA_QR_UT_opc_var1(), and FLA_QR_UT_opc_var2().
{
scomplex* one_p = FLA_COMPLEX_PTR( FLA_ONE );
scomplex* minus_one_p = FLA_COMPLEX_PTR( FLA_MINUS_ONE );
int inc_w1t;
// FLA_Obj w1t;
scomplex* w1t;
// if ( FLA_Obj_has_zero_dim( a1t ) ) return FLA_SUCCESS;
if ( n_a1t == 0 ) return FLA_SUCCESS;
// FLA_Obj_create_conf_to( FLA_NO_TRANSPOSE, a1t, &w1t );
w1t = ( scomplex* ) FLA_malloc( n_a1t * sizeof( *a1t ) );
inc_w1t = 1;
// // w1t = a1t;
// FLA_Copy_external( a1t, w1t );
bli_ccopyv( BLIS_NO_CONJUGATE,
n_a1t,
a1t, inc_a1t,
w1t, inc_w1t );
// // w1t = w1t + u2' * A2;
// // w1t = w1t + A2^T * conj(u2);
// FLA_Gemvc_external( FLA_TRANSPOSE, FLA_CONJUGATE, FLA_ONE, A2, u2, FLA_ONE, w1t );
bli_cgemv( BLIS_TRANSPOSE,
BLIS_CONJUGATE,
m_u2_A2,
n_a1t,
one_p,
A2, rs_A2, cs_A2,
u2, inc_u2,
one_p,
w1t, inc_w1t );
// // w1t = w1t / tau;
// FLA_Inv_scalc_external( FLA_NO_CONJUGATE, tau, w1t );
bli_cinvscalv( BLIS_NO_CONJUGATE,
n_a1t,
tau,
w1t, inc_w1t );
// // a1t = - w1t + a1t;
// FLA_Axpy_external( FLA_MINUS_ONE, w1t, a1t );
bli_caxpyv( BLIS_NO_CONJUGATE,
n_a1t,
minus_one_p,
w1t, inc_w1t,
a1t, inc_a1t );
// // A2 = - u2 * w1t + A2;
// FLA_Ger_external( FLA_MINUS_ONE, u2, w1t, A2 );
bli_cger( BLIS_NO_CONJUGATE,
BLIS_NO_CONJUGATE,
m_u2_A2,
n_a1t,
minus_one_p,
u2, inc_u2,
w1t, inc_w1t,
A2, rs_A2, cs_A2 );
// FLA_Obj_free( &w1t );
FLA_free( w1t );
return FLA_SUCCESS;
}
| FLA_Error FLA_Apply_H2_UT_l_opd_var1 | ( | int | m_u2_A2, |
| int | n_a1t, | ||
| double * | tau, | ||
| double * | u2, | ||
| int | inc_u2, | ||
| double * | a1t, | ||
| int | inc_a1t, | ||
| double * | A2, | ||
| int | rs_A2, | ||
| int | cs_A2 | ||
| ) |
References bli_daxpyv(), bli_dcopyv(), bli_dgemv(), bli_dger(), bli_dinvscalv(), BLIS_CONJUGATE, BLIS_NO_CONJUGATE, BLIS_TRANSPOSE, FLA_free(), FLA_malloc(), FLA_MINUS_ONE, and FLA_ONE.
Referenced by FLA_Apply_H2_UT_l_opt_var1(), FLA_Bidiag_UT_u_step_opd_var1(), FLA_CAQR2_UT_opd_var1(), FLA_Hess_UT_step_opd_var1(), FLA_QR2_UT_opd_var1(), FLA_QR_UT_form_Q_opd_var1(), FLA_QR_UT_opd_var1(), and FLA_QR_UT_opd_var2().
{
double* one_p = FLA_DOUBLE_PTR( FLA_ONE );
double* minus_one_p = FLA_DOUBLE_PTR( FLA_MINUS_ONE );
int inc_w1t;
// FLA_Obj w1t;
double* w1t;
// if ( FLA_Obj_has_zero_dim( a1t ) ) return FLA_SUCCESS;
if ( n_a1t == 0 ) return FLA_SUCCESS;
// FLA_Obj_create_conf_to( FLA_NO_TRANSPOSE, a1t, &w1t );
w1t = ( double* ) FLA_malloc( n_a1t * sizeof( *a1t ) );
inc_w1t = 1;
// // w1t = a1t;
// FLA_Copy_external( a1t, w1t );
bli_dcopyv( BLIS_NO_CONJUGATE,
n_a1t,
a1t, inc_a1t,
w1t, inc_w1t );
// // w1t = w1t + u2' * A2;
// // w1t = w1t + A2^T * conj(u2);
// FLA_Gemvc_external( FLA_TRANSPOSE, FLA_CONJUGATE, FLA_ONE, A2, u2, FLA_ONE, w1t );
bli_dgemv( BLIS_TRANSPOSE,
BLIS_CONJUGATE,
m_u2_A2,
n_a1t,
one_p,
A2, rs_A2, cs_A2,
u2, inc_u2,
one_p,
w1t, inc_w1t );
// // w1t = w1t / tau;
// FLA_Inv_scalc_external( FLA_NO_CONJUGATE, tau, w1t );
bli_dinvscalv( BLIS_NO_CONJUGATE,
n_a1t,
tau,
w1t, inc_w1t );
// // a1t = - w1t + a1t;
// FLA_Axpy_external( FLA_MINUS_ONE, w1t, a1t );
bli_daxpyv( BLIS_NO_CONJUGATE,
n_a1t,
minus_one_p,
w1t, inc_w1t,
a1t, inc_a1t );
// // A2 = - u2 * w1t + A2;
// FLA_Ger_external( FLA_MINUS_ONE, u2, w1t, A2 );
bli_dger( BLIS_NO_CONJUGATE,
BLIS_NO_CONJUGATE,
m_u2_A2,
n_a1t,
minus_one_p,
u2, inc_u2,
w1t, inc_w1t,
A2, rs_A2, cs_A2 );
// FLA_Obj_free( &w1t );
FLA_free( w1t );
return FLA_SUCCESS;
}
| FLA_Error FLA_Apply_H2_UT_l_ops_var1 | ( | int | m_u2_A2, |
| int | n_a1t, | ||
| float * | tau, | ||
| float * | u2, | ||
| int | inc_u2, | ||
| float * | a1t, | ||
| int | inc_a1t, | ||
| float * | A2, | ||
| int | rs_A2, | ||
| int | cs_A2 | ||
| ) |
References bli_saxpyv(), bli_scopyv(), bli_sgemv(), bli_sger(), bli_sinvscalv(), BLIS_CONJUGATE, BLIS_NO_CONJUGATE, BLIS_TRANSPOSE, FLA_free(), FLA_malloc(), FLA_MINUS_ONE, and FLA_ONE.
Referenced by FLA_Apply_H2_UT_l_opt_var1(), FLA_Bidiag_UT_u_step_ops_var1(), FLA_CAQR2_UT_ops_var1(), FLA_Hess_UT_step_ops_var1(), FLA_QR2_UT_ops_var1(), FLA_QR_UT_form_Q_ops_var1(), FLA_QR_UT_ops_var1(), and FLA_QR_UT_ops_var2().
{
float* one_p = FLA_FLOAT_PTR( FLA_ONE );
float* minus_one_p = FLA_FLOAT_PTR( FLA_MINUS_ONE );
int inc_w1t;
// FLA_Obj w1t;
float* w1t;
// if ( FLA_Obj_has_zero_dim( a1t ) ) return FLA_SUCCESS;
if ( n_a1t == 0 ) return FLA_SUCCESS;
// FLA_Obj_create_conf_to( FLA_NO_TRANSPOSE, a1t, &w1t );
w1t = ( float* ) FLA_malloc( n_a1t * sizeof( *a1t ) );
inc_w1t = 1;
// // w1t = a1t;
// FLA_Copy_external( a1t, w1t );
bli_scopyv( BLIS_NO_CONJUGATE,
n_a1t,
a1t, inc_a1t,
w1t, inc_w1t );
// // w1t = w1t + u2' * A2;
// // w1t = w1t + A2^T * conj(u2);
// FLA_Gemvc_external( FLA_TRANSPOSE, FLA_CONJUGATE, FLA_ONE, A2, u2, FLA_ONE, w1t );
bli_sgemv( BLIS_TRANSPOSE,
BLIS_CONJUGATE,
m_u2_A2,
n_a1t,
one_p,
A2, rs_A2, cs_A2,
u2, inc_u2,
one_p,
w1t, inc_w1t );
// // w1t = w1t / tau;
// FLA_Inv_scalc_external( FLA_NO_CONJUGATE, tau, w1t );
bli_sinvscalv( BLIS_NO_CONJUGATE,
n_a1t,
tau,
w1t, inc_w1t );
// // a1t = - w1t + a1t;
// FLA_Axpy_external( FLA_MINUS_ONE, w1t, a1t );
bli_saxpyv( BLIS_NO_CONJUGATE,
n_a1t,
minus_one_p,
w1t, inc_w1t,
a1t, inc_a1t );
// // A2 = - u2 * w1t + A2;
// FLA_Ger_external( FLA_MINUS_ONE, u2, w1t, A2 );
bli_sger( BLIS_NO_CONJUGATE,
BLIS_NO_CONJUGATE,
m_u2_A2,
n_a1t,
minus_one_p,
u2, inc_u2,
w1t, inc_w1t,
A2, rs_A2, cs_A2 );
// FLA_Obj_free( &w1t );
FLA_free( w1t );
return FLA_SUCCESS;
}
References FLA_Apply_H2_UT_l_opc_var1(), FLA_Apply_H2_UT_l_opd_var1(), FLA_Apply_H2_UT_l_ops_var1(), FLA_Apply_H2_UT_l_opz_var1(), FLA_Obj_col_stride(), FLA_Obj_datatype(), FLA_Obj_has_zero_dim(), FLA_Obj_length(), FLA_Obj_row_stride(), FLA_Obj_vector_inc(), and FLA_Obj_width().
Referenced by FLA_Apply_H2_UT_internal().
{
FLA_Datatype datatype;
int m_u2_A2;
int n_a1t;
int inc_u2;
int inc_a1t;
int rs_A2;
int cs_A2;
if ( FLA_Obj_has_zero_dim( a1t ) ) return FLA_SUCCESS;
datatype = FLA_Obj_datatype( A2 );
m_u2_A2 = FLA_Obj_length( A2 );
n_a1t = FLA_Obj_width( a1t );
inc_u2 = FLA_Obj_vector_inc( u2 );
inc_a1t = FLA_Obj_vector_inc( a1t );
rs_A2 = FLA_Obj_row_stride( A2 );
cs_A2 = FLA_Obj_col_stride( A2 );
switch ( datatype )
{
case FLA_FLOAT:
{
float* tau_p = ( float* ) FLA_FLOAT_PTR( tau );
float* u2_p = ( float* ) FLA_FLOAT_PTR( u2 );
float* a1t_p = ( float* ) FLA_FLOAT_PTR( a1t );
float* A2_p = ( float* ) FLA_FLOAT_PTR( A2 );
FLA_Apply_H2_UT_l_ops_var1( m_u2_A2, n_a1t,
tau_p,
u2_p, inc_u2,
a1t_p, inc_a1t,
A2_p, rs_A2, cs_A2 );
break;
}
case FLA_DOUBLE:
{
double* tau_p = ( double* ) FLA_DOUBLE_PTR( tau );
double* u2_p = ( double* ) FLA_DOUBLE_PTR( u2 );
double* a1t_p = ( double* ) FLA_DOUBLE_PTR( a1t );
double* A2_p = ( double* ) FLA_DOUBLE_PTR( A2 );
FLA_Apply_H2_UT_l_opd_var1( m_u2_A2, n_a1t,
tau_p,
u2_p, inc_u2,
a1t_p, inc_a1t,
A2_p, rs_A2, cs_A2 );
break;
}
case FLA_COMPLEX:
{
scomplex* tau_p = ( scomplex* ) FLA_COMPLEX_PTR( tau );
scomplex* u2_p = ( scomplex* ) FLA_COMPLEX_PTR( u2 );
scomplex* a1t_p = ( scomplex* ) FLA_COMPLEX_PTR( a1t );
scomplex* A2_p = ( scomplex* ) FLA_COMPLEX_PTR( A2 );
FLA_Apply_H2_UT_l_opc_var1( m_u2_A2, n_a1t,
tau_p,
u2_p, inc_u2,
a1t_p, inc_a1t,
A2_p, rs_A2, cs_A2 );
break;
}
case FLA_DOUBLE_COMPLEX:
{
dcomplex* tau_p = ( dcomplex* ) FLA_DOUBLE_COMPLEX_PTR( tau );
dcomplex* u2_p = ( dcomplex* ) FLA_DOUBLE_COMPLEX_PTR( u2 );
dcomplex* a1t_p = ( dcomplex* ) FLA_DOUBLE_COMPLEX_PTR( a1t );
dcomplex* A2_p = ( dcomplex* ) FLA_DOUBLE_COMPLEX_PTR( A2 );
FLA_Apply_H2_UT_l_opz_var1( m_u2_A2, n_a1t,
tau_p,
u2_p, inc_u2,
a1t_p, inc_a1t,
A2_p, rs_A2, cs_A2 );
break;
}
}
return FLA_SUCCESS;
}
| FLA_Error FLA_Apply_H2_UT_l_opz_var1 | ( | int | m_u2_A2, |
| int | n_a1t, | ||
| dcomplex * | tau, | ||
| dcomplex * | u2, | ||
| int | inc_u2, | ||
| dcomplex * | a1t, | ||
| int | inc_a1t, | ||
| dcomplex * | A2, | ||
| int | rs_A2, | ||
| int | cs_A2 | ||
| ) |
References bli_zaxpyv(), bli_zcopyv(), bli_zgemv(), bli_zger(), bli_zinvscalv(), BLIS_CONJUGATE, BLIS_NO_CONJUGATE, BLIS_TRANSPOSE, FLA_free(), FLA_malloc(), FLA_MINUS_ONE, and FLA_ONE.
Referenced by FLA_Apply_H2_UT_l_opt_var1(), FLA_Bidiag_UT_u_step_opz_var1(), FLA_CAQR2_UT_opz_var1(), FLA_Hess_UT_step_opz_var1(), FLA_QR2_UT_opz_var1(), FLA_QR_UT_form_Q_opz_var1(), FLA_QR_UT_opz_var1(), and FLA_QR_UT_opz_var2().
{
dcomplex* one_p = FLA_DOUBLE_COMPLEX_PTR( FLA_ONE );
dcomplex* minus_one_p = FLA_DOUBLE_COMPLEX_PTR( FLA_MINUS_ONE );
int inc_w1t;
// FLA_Obj w1t;
dcomplex* w1t;
// if ( FLA_Obj_has_zero_dim( a1t ) ) return FLA_SUCCESS;
if ( n_a1t == 0 ) return FLA_SUCCESS;
// FLA_Obj_create_conf_to( FLA_NO_TRANSPOSE, a1t, &w1t );
w1t = ( dcomplex* ) FLA_malloc( n_a1t * sizeof( *a1t ) );
inc_w1t = 1;
// // w1t = a1t;
// FLA_Copy_external( a1t, w1t );
bli_zcopyv( BLIS_NO_CONJUGATE,
n_a1t,
a1t, inc_a1t,
w1t, inc_w1t );
// // w1t = w1t + u2' * A2;
// // w1t = w1t + A2^T * conj(u2);
// FLA_Gemvc_external( FLA_TRANSPOSE, FLA_CONJUGATE, FLA_ONE, A2, u2, FLA_ONE, w1t );
bli_zgemv( BLIS_TRANSPOSE,
BLIS_CONJUGATE,
m_u2_A2,
n_a1t,
one_p,
A2, rs_A2, cs_A2,
u2, inc_u2,
one_p,
w1t, inc_w1t );
// // w1t = w1t / tau;
// FLA_Inv_scalc_external( FLA_NO_CONJUGATE, tau, w1t );
bli_zinvscalv( BLIS_NO_CONJUGATE,
n_a1t,
tau,
w1t, inc_w1t );
// // a1t = - w1t + a1t;
// FLA_Axpy_external( FLA_MINUS_ONE, w1t, a1t );
bli_zaxpyv( BLIS_NO_CONJUGATE,
n_a1t,
minus_one_p,
w1t, inc_w1t,
a1t, inc_a1t );
// // A2 = - u2 * w1t + A2;
// FLA_Ger_external( FLA_MINUS_ONE, u2, w1t, A2 );
bli_zger( BLIS_NO_CONJUGATE,
BLIS_NO_CONJUGATE,
m_u2_A2,
n_a1t,
minus_one_p,
u2, inc_u2,
w1t, inc_w1t,
A2, rs_A2, cs_A2 );
// FLA_Obj_free( &w1t );
FLA_free( w1t );
return FLA_SUCCESS;
}
1.7.6.1