|
libflame
revision_anchor
|
Functions | |
| FLA_Error | FLA_LQ_UT_form_Q (FLA_Obj A, FLA_Obj T, FLA_Obj Q) |
| FLA_Error | FLA_LQ_UT_form_Q_blk_var1 (FLA_Obj A, FLA_Obj T, FLA_Obj W) |
| FLA_Error | FLA_LQ_UT_form_Q_opt_var1 (FLA_Obj A, FLA_Obj T) |
| FLA_Error | FLA_LQ_UT_form_Q_ops_var1 (int m_A, int n_A, float *buff_A, int rs_A, int cs_A, float *buff_T, int rs_T, int cs_T) |
| FLA_Error | FLA_LQ_UT_form_Q_opd_var1 (int m_A, int n_A, double *buff_A, int rs_A, int cs_A, double *buff_T, int rs_T, int cs_T) |
| FLA_Error | FLA_LQ_UT_form_Q_opc_var1 (int m_A, int n_A, scomplex *buff_A, int rs_A, int cs_A, scomplex *buff_T, int rs_T, int cs_T) |
| FLA_Error | FLA_LQ_UT_form_Q_opz_var1 (int m_A, int n_A, dcomplex *buff_A, int rs_A, int cs_A, dcomplex *buff_T, int rs_T, int cs_T) |
| FLA_Error FLA_LQ_UT_form_Q | ( | FLA_Obj | A, |
| FLA_Obj | T, | ||
| FLA_Obj | Q | ||
| ) |
References FLA_Apply_Q_UT_create_workspace(), FLA_Check_error_level(), FLA_Copyr(), FLA_LQ_UT_form_Q_blk_var1(), FLA_LQ_UT_form_Q_check(), FLA_Merge_1x2(), FLA_Obj_free(), FLA_Obj_is(), FLA_Obj_length(), FLA_ONE, FLA_Part_2x2(), FLA_Set_diag(), FLA_Setr(), and FLA_ZERO.
{
FLA_Error r_val = FLA_SUCCESS;
FLA_Obj QTL, QTR,
QBL, QBR;
FLA_Obj QT;
FLA_Obj W;
dim_t m_A;
if ( FLA_Check_error_level() >= FLA_MIN_ERROR_CHECKING )
FLA_LQ_UT_form_Q_check( A, T, Q );
// Zero out the lower triangle of Q.
FLA_Setr( FLA_LOWER_TRIANGULAR, FLA_ZERO, Q );
// Set the digaonal to one.
FLA_Set_diag( FLA_ONE, Q );
// If A and Q are different objects, copy the Householder vectors
// from A to QT, and zero out the upper triangle of QBR. If they
// are the same object, we don't need to do the copy, and don't
// need to zero anything out since the user should only have A and
// Q be the same object if A is square, since Q needs to be square
// (specifically, dim(Q) needs to equal n(A)).
if ( FLA_Obj_is( A, Q ) == FALSE )
{
m_A = FLA_Obj_length( A );
FLA_Part_2x2( Q, &QTL, &QTR,
&QBL, &QBR, m_A, m_A, FLA_TL );
FLA_Merge_1x2( QTL, QTR, &QT );
// Copy the Householder vectors in A to QT.
FLA_Copyr( FLA_UPPER_TRIANGULAR, A, QT );
// Zero out the lower triangle of QBR.
FLA_Setr( FLA_UPPER_TRIANGULAR, FLA_ZERO, QBR );
}
// Create workspace for applying the block Householder transforms.
FLA_Apply_Q_UT_create_workspace( T, Q, &W );
// Overwrite Q, which currently contains Householder vectors in the
// strictly lower triangle and identity in the upper triangle, with
// the unitary matrix associated with those Householder transforms.
r_val = FLA_LQ_UT_form_Q_blk_var1( Q, T, W );
// Free the temporary workspace.
FLA_Obj_free( &W );
return r_val;
}
| FLA_Error FLA_LQ_UT_form_Q_blk_var1 | ( | FLA_Obj | A, |
| FLA_Obj | T, | ||
| FLA_Obj | W | ||
| ) |
References FLA_Apply_Q_UT(), FLA_Cont_with_1x3_to_1x2(), FLA_Cont_with_3x3_to_2x2(), FLA_LQ_UT_form_Q_opt_var1(), FLA_Merge_1x2(), FLA_Obj_length(), FLA_Obj_min_dim(), FLA_Obj_width(), FLA_Part_1x2(), FLA_Part_2x1(), FLA_Part_2x2(), FLA_Repart_1x2_to_1x3(), and FLA_Repart_2x2_to_3x3().
Referenced by FLA_LQ_UT_form_Q().
{
FLA_Obj ATL, ATR, A00, A01, A02,
ABL, ABR, A10, A11, A12,
A20, A21, A22;
FLA_Obj TL, TR, T0, T1, T2;
FLA_Obj T1T,
T2B;
FLA_Obj WTL, WTR,
WBL, WBR;
FLA_Obj AR1, AR2;
dim_t b, b_alg;
dim_t m_BR, n_BR;
b_alg = FLA_Obj_length( T );
// If A is longer than T is wide, then we need to position ourseves
// carefully within the matrix for the initial partitioning.
if ( FLA_Obj_length( A ) > FLA_Obj_width( T ) )
{
m_BR = FLA_Obj_length( A ) - FLA_Obj_width( T );
n_BR = FLA_Obj_width( A ) - FLA_Obj_width( T );
}
else
{
m_BR = 0;
n_BR = 0;
}
FLA_Part_2x2( A, &ATL, &ATR,
&ABL, &ABR, m_BR, n_BR, FLA_BR );
FLA_Part_1x2( T, &TL, &TR, 0, FLA_RIGHT );
while ( FLA_Obj_length( ATL ) > 0 )
{
b = min( b_alg, FLA_Obj_min_dim( ATL ) );
// Since T was filled from left to right, and since we need to access them
// in reverse order, we need to handle the case where the last block is
// smaller than the other b x b blocks.
if ( FLA_Obj_width( TR ) == 0 && FLA_Obj_width( T ) % b_alg > 0 )
b = FLA_Obj_width( T ) % b_alg;
FLA_Repart_2x2_to_3x3( ATL, /**/ ATR, &A00, &A01, /**/ &A02,
&A10, &A11, /**/ &A12,
/* ************* */ /* ******************** */
ABL, /**/ ABR, &A20, &A21, /**/ &A22,
b, b, FLA_TL );
FLA_Repart_1x2_to_1x3( TL, /**/ TR, &T0, &T1, /**/ &T2,
b, FLA_LEFT );
/*------------------------------------------------------------*/
FLA_Part_2x1( T1, &T1T,
&T2B, b, FLA_TOP );
FLA_Part_2x2( W, &WTL, &WTR,
&WBL, &WBR, b, FLA_Obj_length( A21 ), FLA_TL );
// Use an unblocked algorithm for the first (or only) block.
if ( FLA_Obj_length( ABR ) == 0 )
{
FLA_LQ_UT_form_Q_opt_var1( A11, T1T );
}
else
{
FLA_Merge_1x2( A11, A12, &AR1 );
FLA_Merge_1x2( A21, A22, &AR2 );
// Apply the block Householder transforms to A21 and A22.
FLA_Apply_Q_UT( FLA_RIGHT, FLA_CONJ_TRANSPOSE, FLA_FORWARD, FLA_ROWWISE,
AR1, T1T, WTL, AR2 );
// Apply H to the current block panel consisting of A11 and A12.
FLA_LQ_UT_form_Q_opt_var1( AR1, T1T );
}
/*------------------------------------------------------------*/
FLA_Cont_with_3x3_to_2x2( &ATL, /**/ &ATR, A00, /**/ A01, A02,
/* ************** */ /* ****************** */
A10, /**/ A11, A12,
&ABL, /**/ &ABR, A20, /**/ A21, A22,
FLA_BR );
FLA_Cont_with_1x3_to_1x2( &TL, /**/ &TR, T0, /**/ T1, T2,
FLA_RIGHT );
}
return FLA_SUCCESS;
}
| FLA_Error FLA_LQ_UT_form_Q_opc_var1 | ( | int | m_A, |
| int | n_A, | ||
| scomplex * | buff_A, | ||
| int | rs_A, | ||
| int | cs_A, | ||
| scomplex * | buff_T, | ||
| int | rs_T, | ||
| int | cs_T | ||
| ) |
References bli_c0(), bli_c1(), bli_cconjv(), bli_cscalv(), BLIS_NO_CONJUGATE, FLA_Apply_H2_UT_r_opc_var1(), scomplex::imag, and scomplex::real.
Referenced by FLA_LQ_UT_form_Q_opt_var1().
{
scomplex zero = bli_c0();
scomplex one = bli_c1();
int min_m_n = min( m_A, n_A );
int i;
for ( i = min_m_n - 1; i >= 0; --i )
{
scomplex* alpha11 = buff_A + (i )*cs_A + (i )*rs_A;
scomplex* a21 = buff_A + (i )*cs_A + (i+1)*rs_A;
scomplex* a12t = buff_A + (i+1)*cs_A + (i )*rs_A;
scomplex* A22 = buff_A + (i+1)*cs_A + (i+1)*rs_A;
scomplex* tau11 = buff_T + (i )*cs_T + (i )*rs_T;
scomplex minus_inv_tau11;
int n_ahead = n_A - i - 1;
int m_ahead = m_A - i - 1;
FLA_Apply_H2_UT_r_opc_var1( m_ahead,
n_ahead,
tau11,
a12t, cs_A,
a21, rs_A,
A22, rs_A, cs_A );
minus_inv_tau11.real = -one.real / tau11->real;
minus_inv_tau11.imag = zero.imag;
alpha11->real = one.real + minus_inv_tau11.real;
alpha11->imag = zero.imag;
bli_cconjv( n_ahead,
a12t, cs_A );
bli_cscalv( BLIS_NO_CONJUGATE,
n_ahead,
&minus_inv_tau11,
a12t, cs_A );
}
return FLA_SUCCESS;
}
| FLA_Error FLA_LQ_UT_form_Q_opd_var1 | ( | int | m_A, |
| int | n_A, | ||
| double * | buff_A, | ||
| int | rs_A, | ||
| int | cs_A, | ||
| double * | buff_T, | ||
| int | rs_T, | ||
| int | cs_T | ||
| ) |
References bli_d1(), bli_dscalv(), BLIS_NO_CONJUGATE, and FLA_Apply_H2_UT_r_opd_var1().
Referenced by FLA_LQ_UT_form_Q_opt_var1().
{
double one = bli_d1();
int min_m_n = min( m_A, n_A );
int i;
for ( i = min_m_n - 1; i >= 0; --i )
{
double* alpha11 = buff_A + (i )*cs_A + (i )*rs_A;
double* a21 = buff_A + (i )*cs_A + (i+1)*rs_A;
double* a12t = buff_A + (i+1)*cs_A + (i )*rs_A;
double* A22 = buff_A + (i+1)*cs_A + (i+1)*rs_A;
double* tau11 = buff_T + (i )*cs_T + (i )*rs_T;
double minus_inv_tau11;
int n_ahead = n_A - i - 1;
int m_ahead = m_A - i - 1;
FLA_Apply_H2_UT_r_opd_var1( m_ahead,
n_ahead,
tau11,
a12t, cs_A,
a21, rs_A,
A22, rs_A, cs_A );
minus_inv_tau11 = -one / *tau11;
*alpha11 = one + minus_inv_tau11;
bli_dscalv( BLIS_NO_CONJUGATE,
n_ahead,
&minus_inv_tau11,
a12t, cs_A );
}
return FLA_SUCCESS;
}
| FLA_Error FLA_LQ_UT_form_Q_ops_var1 | ( | int | m_A, |
| int | n_A, | ||
| float * | buff_A, | ||
| int | rs_A, | ||
| int | cs_A, | ||
| float * | buff_T, | ||
| int | rs_T, | ||
| int | cs_T | ||
| ) |
References bli_d1(), bli_sscalv(), BLIS_NO_CONJUGATE, and FLA_Apply_H2_UT_r_ops_var1().
Referenced by FLA_LQ_UT_form_Q_opt_var1().
{
float one = bli_d1();
int min_m_n = min( m_A, n_A );
int i;
for ( i = min_m_n - 1; i >= 0; --i )
{
float* alpha11 = buff_A + (i )*cs_A + (i )*rs_A;
float* a21 = buff_A + (i )*cs_A + (i+1)*rs_A;
float* a12t = buff_A + (i+1)*cs_A + (i )*rs_A;
float* A22 = buff_A + (i+1)*cs_A + (i+1)*rs_A;
float* tau11 = buff_T + (i )*cs_T + (i )*rs_T;
float minus_inv_tau11;
int n_ahead = n_A - i - 1;
int m_ahead = m_A - i - 1;
FLA_Apply_H2_UT_r_ops_var1( m_ahead,
n_ahead,
tau11,
a12t, cs_A,
a21, rs_A,
A22, rs_A, cs_A );
minus_inv_tau11 = -one / *tau11;
*alpha11 = one + minus_inv_tau11;
bli_sscalv( BLIS_NO_CONJUGATE,
n_ahead,
&minus_inv_tau11,
a12t, cs_A );
}
return FLA_SUCCESS;
}
| FLA_Error FLA_LQ_UT_form_Q_opt_var1 | ( | FLA_Obj | A, |
| FLA_Obj | T | ||
| ) |
References FLA_LQ_UT_form_Q_opc_var1(), FLA_LQ_UT_form_Q_opd_var1(), FLA_LQ_UT_form_Q_ops_var1(), FLA_LQ_UT_form_Q_opz_var1(), FLA_Obj_col_stride(), FLA_Obj_datatype(), FLA_Obj_length(), FLA_Obj_row_stride(), and FLA_Obj_width().
Referenced by FLA_LQ_UT_form_Q_blk_var1().
{
FLA_Datatype datatype;
int m_A, n_A;
int rs_A, cs_A;
int rs_T, cs_T;
datatype = FLA_Obj_datatype( A );
m_A = FLA_Obj_length( A );
n_A = FLA_Obj_width( A );
rs_A = FLA_Obj_row_stride( A );
cs_A = FLA_Obj_col_stride( A );
rs_T = FLA_Obj_row_stride( T );
cs_T = FLA_Obj_col_stride( T );
switch ( datatype )
{
case FLA_FLOAT:
{
float* buff_A = ( float* ) FLA_FLOAT_PTR( A );
float* buff_T = ( float* ) FLA_FLOAT_PTR( T );
FLA_LQ_UT_form_Q_ops_var1( m_A,
n_A,
buff_A, rs_A, cs_A,
buff_T, rs_T, cs_T );
break;
}
case FLA_DOUBLE:
{
double* buff_A = ( double* ) FLA_DOUBLE_PTR( A );
double* buff_T = ( double* ) FLA_DOUBLE_PTR( T );
FLA_LQ_UT_form_Q_opd_var1( m_A,
n_A,
buff_A, rs_A, cs_A,
buff_T, rs_T, cs_T );
break;
}
case FLA_COMPLEX:
{
scomplex* buff_A = ( scomplex* ) FLA_COMPLEX_PTR( A );
scomplex* buff_T = ( scomplex* ) FLA_COMPLEX_PTR( T );
FLA_LQ_UT_form_Q_opc_var1( m_A,
n_A,
buff_A, rs_A, cs_A,
buff_T, rs_T, cs_T );
break;
}
case FLA_DOUBLE_COMPLEX:
{
dcomplex* buff_A = ( dcomplex* ) FLA_DOUBLE_COMPLEX_PTR( A );
dcomplex* buff_T = ( dcomplex* ) FLA_DOUBLE_COMPLEX_PTR( T );
FLA_LQ_UT_form_Q_opz_var1( m_A,
n_A,
buff_A, rs_A, cs_A,
buff_T, rs_T, cs_T );
break;
}
}
return FLA_SUCCESS;
}
| FLA_Error FLA_LQ_UT_form_Q_opz_var1 | ( | int | m_A, |
| int | n_A, | ||
| dcomplex * | buff_A, | ||
| int | rs_A, | ||
| int | cs_A, | ||
| dcomplex * | buff_T, | ||
| int | rs_T, | ||
| int | cs_T | ||
| ) |
References bli_z0(), bli_z1(), bli_zconjv(), bli_zscalv(), BLIS_NO_CONJUGATE, FLA_Apply_H2_UT_r_opz_var1(), dcomplex::imag, and dcomplex::real.
Referenced by FLA_LQ_UT_form_Q_opt_var1().
{
dcomplex zero = bli_z0();
dcomplex one = bli_z1();
int min_m_n = min( m_A, n_A );
int i;
for ( i = min_m_n - 1; i >= 0; --i )
{
dcomplex* alpha11 = buff_A + (i )*cs_A + (i )*rs_A;
dcomplex* a21 = buff_A + (i )*cs_A + (i+1)*rs_A;
dcomplex* a12t = buff_A + (i+1)*cs_A + (i )*rs_A;
dcomplex* A22 = buff_A + (i+1)*cs_A + (i+1)*rs_A;
dcomplex* tau11 = buff_T + (i )*cs_T + (i )*rs_T;
dcomplex minus_inv_tau11;
int n_ahead = n_A - i - 1;
int m_ahead = m_A - i - 1;
FLA_Apply_H2_UT_r_opz_var1( m_ahead,
n_ahead,
tau11,
a12t, cs_A,
a21, rs_A,
A22, rs_A, cs_A );
minus_inv_tau11.real = -one.real / tau11->real;
minus_inv_tau11.imag = zero.imag;
alpha11->real = one.real + minus_inv_tau11.real;
alpha11->imag = zero.imag;
bli_zconjv( n_ahead,
a12t, cs_A );
bli_zscalv( BLIS_NO_CONJUGATE,
n_ahead,
&minus_inv_tau11,
a12t, cs_A );
}
return FLA_SUCCESS;
}
1.7.6.1