FLA_Error FLA_Herk_un_unb_var4( FLA_Obj alpha, FLA_Obj A, FLA_Obj beta, FLA_Obj C )
{
FLA_Obj AT, A0,
AB, a1t,
A2;
FLA_Obj CTL, CTR, C00, c01, C02,
CBL, CBR, c10t, gamma11, c12t,
C20, c21, C22;
FLA_Scalr_external( FLA_UPPER_TRIANGULAR, beta, C );
FLA_Part_2x1( A, &AT,
&AB, 0, FLA_BOTTOM );
FLA_Part_2x2( C, &CTL, &CTR,
&CBL, &CBR, 0, 0, FLA_BR );
while ( FLA_Obj_length( AB ) < FLA_Obj_length( A ) ){
FLA_Repart_2x1_to_3x1( AT, &A0,
&a1t,
/* ** */ /* *** */
AB, &A2, 1, FLA_TOP );
FLA_Repart_2x2_to_3x3( CTL, /**/ CTR, &C00, &c01, /**/ &C02,
&c10t, &gamma11, /**/ &c12t,
/* ************* */ /* ************************** */
CBL, /**/ CBR, &C20, &c21, /**/ &C22,
1, 1, FLA_TL );
/*------------------------------------------------------------*/
/* c01 = c01 + A0 * a1t' */
FLA_Gemvc_external( FLA_NO_TRANSPOSE, FLA_CONJUGATE, alpha, A0, a1t, FLA_ONE, c01 );
/* gamma11 = gamma11 + a1t * a1t' */
FLA_Dotcs_external( FLA_CONJUGATE, alpha, a1t, a1t, FLA_ONE, gamma11 );
/*------------------------------------------------------------*/
FLA_Cont_with_3x1_to_2x1( &AT, A0,
/* ** */ /* *** */
a1t,
&AB, A2, FLA_BOTTOM );
FLA_Cont_with_3x3_to_2x2( &CTL, /**/ &CTR, C00, /**/ c01, C02,
/* ************** */ /* ************************ */
c10t, /**/ gamma11, c12t,
&CBL, /**/ &CBR, C20, /**/ c21, C22,
FLA_BR );
}
return FLA_SUCCESS;
}
FLA_Error FLA_Herk_lh_unb_var1( FLA_Obj alpha, FLA_Obj A, FLA_Obj beta, FLA_Obj C )
{
FLA_Obj AL, AR, A0, a1, A2;
FLA_Obj CTL, CTR, C00, c01, C02,
CBL, CBR, c10t, gamma11, c12t,
C20, c21, C22;
FLA_Scalr_external( FLA_LOWER_TRIANGULAR, beta, C );
FLA_Part_1x2( A, &AL, &AR, 0, FLA_LEFT );
FLA_Part_2x2( C, &CTL, &CTR,
&CBL, &CBR, 0, 0, FLA_TL );
while ( FLA_Obj_width( AL ) < FLA_Obj_width( A ) ){
FLA_Repart_1x2_to_1x3( AL, /**/ AR, &A0, /**/ &a1, &A2,
1, FLA_RIGHT );
FLA_Repart_2x2_to_3x3( CTL, /**/ CTR, &C00, /**/ &c01, &C02,
/* ************* */ /* ************************** */
&c10t, /**/ &gamma11, &c12t,
CBL, /**/ CBR, &C20, /**/ &c21, &C22,
1, 1, FLA_BR );
/*------------------------------------------------------------*/
/* C10 = C10 + A1' * A0 */
/* c10t = c10t + A0' * a1 */
FLA_Gemvc_external( FLA_TRANSPOSE, FLA_CONJUGATE, alpha, A0, a1, FLA_ONE, c10t );
/* gamma11 = gamma11 + a1' * a1 */
FLA_Dotcs_external( FLA_CONJUGATE, alpha, a1, a1, FLA_ONE, gamma11 );
/*------------------------------------------------------------*/
FLA_Cont_with_1x3_to_1x2( &AL, /**/ &AR, A0, a1, /**/ A2,
FLA_LEFT );
FLA_Cont_with_3x3_to_2x2( &CTL, /**/ &CTR, C00, c01, /**/ C02,
c10t, gamma11, /**/ c12t,
/* ************** */ /* ************************ */
&CBL, /**/ &CBR, C20, c21, /**/ C22,
FLA_TL );
}
return FLA_SUCCESS;
}
FLA_Error FLA_Chol_u_unb_var2( FLA_Obj A )
{
FLA_Obj ATL, ATR, A00, a01, A02,
ABL, ABR, a10t, alpha11, a12t,
A20, a21, A22;
int r_val = FLA_SUCCESS;
FLA_Part_2x2( A, &ATL, &ATR,
&ABL, &ABR, 0, 0, FLA_TL );
while ( FLA_Obj_length( ATL ) < FLA_Obj_length( A ) ){
FLA_Repart_2x2_to_3x3( ATL, /**/ ATR, &A00, /**/ &a01, &A02,
/* ************* */ /* ************************** */
&a10t, /**/ &alpha11, &a12t,
ABL, /**/ ABR, &A20, /**/ &a21, &A22,
1, 1, FLA_BR );
/*------------------------------------------------------------*/
// alpha11 = alpha11 - a01' * a01
FLA_Dotcs_external( FLA_CONJUGATE, FLA_MINUS_ONE, a01, a01, FLA_ONE, alpha11 );
// a12t = a12t - a01' * A02
// a12t' = a12t' - A02' * conj(a01)
FLA_Gemvc_external( FLA_TRANSPOSE, FLA_CONJUGATE, FLA_MINUS_ONE, A02, a01, FLA_ONE, a12t );
// alpha11 = sqrt( alpha11 )
r_val = FLA_Sqrt( alpha11 );
if ( r_val != FLA_SUCCESS )
return ( FLA_Obj_length( A00 ) );
// a12t = a12t / alpha11
FLA_Inv_scal_external( alpha11, a12t );
/*------------------------------------------------------------*/
FLA_Cont_with_3x3_to_2x2( &ATL, /**/ &ATR, A00, a01, /**/ A02,
a10t, alpha11, /**/ a12t,
/* ************** */ /* ************************ */
&ABL, /**/ &ABR, A20, a21, /**/ A22,
FLA_TL );
}
return r_val;
}
FLA_Error FLA_Chol_l_unb_var1( FLA_Obj A )
{
FLA_Obj ATL, ATR, A00, a01, A02,
ABL, ABR, a10t, alpha11, a12t,
A20, a21, A22;
int r_val = FLA_SUCCESS;
FLA_Part_2x2( A, &ATL, &ATR,
&ABL, &ABR, 0, 0, FLA_TL );
while ( FLA_Obj_length( ATL ) < FLA_Obj_length( A ) ){
FLA_Repart_2x2_to_3x3( ATL, /**/ ATR, &A00, /**/ &a01, &A02,
/* ************* */ /* ************************** */
&a10t, /**/ &alpha11, &a12t,
ABL, /**/ ABR, &A20, /**/ &a21, &A22,
1, 1, FLA_BR );
/*------------------------------------------------------------*/
// a10t = a10t * inv( tril( A00 )' )
// a10t' = inv( conj( tril( A00 ) ) ) * a10t'
FLA_Trsv_external( FLA_LOWER_TRIANGULAR, FLA_CONJ_NO_TRANSPOSE, FLA_NONUNIT_DIAG, A00, a10t );
// alpha11 = alpha11 - a10t * a10t'
FLA_Dotcs_external( FLA_CONJUGATE, FLA_MINUS_ONE, a10t, a10t, FLA_ONE, alpha11 );
// alpha11 = sqrt( alpha11 )
r_val = FLA_Sqrt( alpha11 );
if ( r_val != FLA_SUCCESS )
return ( FLA_Obj_length( A00 ) );
/*------------------------------------------------------------*/
FLA_Cont_with_3x3_to_2x2( &ATL, /**/ &ATR, A00, a01, /**/ A02,
a10t, alpha11, /**/ a12t,
/* ************** */ /* ************************ */
&ABL, /**/ &ABR, A20, a21, /**/ A22,
FLA_TL );
}
return r_val;
}
FLA_Error FLA_Ttmm_u_unb_var3( FLA_Obj A )
{
FLA_Obj ATL, ATR, A00, a01, A02,
ABL, ABR, a10t, alpha11, a12t,
A20, a21, A22;
FLA_Part_2x2( A, &ATL, &ATR,
&ABL, &ABR, 0, 0, FLA_TL );
while ( FLA_Obj_length( ATL ) < FLA_Obj_length( A ) ){
FLA_Repart_2x2_to_3x3( ATL, /**/ ATR, &A00, /**/ &a01, &A02,
/* ************* */ /* ************************** */
&a10t, /**/ &alpha11, &a12t,
ABL, /**/ ABR, &A20, /**/ &a21, &A22,
1, 1, FLA_BR );
/*------------------------------------------------------------*/
// alpha11 = alpha11 * alpha11'
FLA_Absolute_square( alpha11 );
// alpha11 = alpha11 + a12t * a12t'
FLA_Dotcs_external( FLA_CONJUGATE, FLA_ONE, a12t, a12t, FLA_ONE, alpha11 );
// a12t = a12t * triu( A22 )'
// a12t' = triu( A22 ) * a12t'
FLA_Trmv_external( FLA_UPPER_TRIANGULAR, FLA_CONJ_NO_TRANSPOSE, FLA_NONUNIT_DIAG, A22, a12t );
/*------------------------------------------------------------*/
FLA_Cont_with_3x3_to_2x2( &ATL, /**/ &ATR, A00, a01, /**/ A02,
a10t, alpha11, /**/ a12t,
/* ************** */ /* ************************ */
&ABL, /**/ &ABR, A20, a21, /**/ A22,
FLA_TL );
}
return FLA_SUCCESS;
}