FLA_Error FLA_Her2k_lh_unb_var2( FLA_Obj alpha, FLA_Obj A, FLA_Obj B, FLA_Obj beta, FLA_Obj C )
{
FLA_Obj AL, AR, A0, a1, A2;
FLA_Obj BL, BR, B0, b1, B2;
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_1x2( B, &BL, &BR, 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_1x2_to_1x3( BL, /**/ BR, &B0, /**/ &b1, &B2,
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 );
/*------------------------------------------------------------*/
/* c10t = c10t + b1' * A0 */
FLA_Gemvc_external( FLA_TRANSPOSE, FLA_CONJUGATE, alpha, A0, b1, FLA_ONE, c10t );
/* c21 = c21 + A2' * b1 */
FLA_Gemv_external( FLA_CONJ_TRANSPOSE, alpha, A2, b1, FLA_ONE, c21 );
/* gamma11 = gamma11 + a1' * b1 + b1' * a1 */
FLA_Dot2cs_external( FLA_CONJUGATE, alpha, a1, b1, FLA_ONE, gamma11 );
/*------------------------------------------------------------*/
FLA_Cont_with_1x3_to_1x2( &AL, /**/ &AR, A0, a1, /**/ A2,
FLA_LEFT );
FLA_Cont_with_1x3_to_1x2( &BL, /**/ &BR, B0, b1, /**/ B2,
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_Eig_gest_iu_unb_var2( FLA_Obj A, FLA_Obj Y, FLA_Obj B )
{
FLA_Obj ATL, ATR, A00, a01, A02,
ABL, ABR, a10t, alpha11, a12t,
A20, a21, A22;
FLA_Obj BTL, BTR, B00, b01, B02,
BBL, BBR, b10t, beta11, b12t,
B20, b21, B22;
FLA_Obj yT, y01,
yB, psi11,
y21;
FLA_Obj y01_l, y01_r;
FLA_Part_2x2( A, &ATL, &ATR,
&ABL, &ABR, 0, 0, FLA_TL );
FLA_Part_2x2( B, &BTL, &BTR,
&BBL, &BBR, 0, 0, FLA_TL );
FLA_Part_2x1( Y, &yT,
&yB, 0, FLA_TOP );
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 );
FLA_Repart_2x2_to_3x3( BTL, /**/ BTR, &B00, /**/ &b01, &B02,
/* ************* */ /* ************************* */
&b10t, /**/ &beta11, &b12t,
BBL, /**/ BBR, &B20, /**/ &b21, &B22,
1, 1, FLA_BR );
FLA_Repart_2x1_to_3x1( yT, &y01,
/* ** */ /* ***** */
&psi11,
yB, &y21, 1, FLA_BOTTOM );
/*------------------------------------------------------------*/
FLA_Part_1x2( y01, &y01_l, &y01_r, 1, FLA_LEFT );
// y01 = 1/2 * A00 * b01;
FLA_Hemvc_external( FLA_UPPER_TRIANGULAR, FLA_NO_CONJUGATE,
FLA_ONE_HALF, A00, b01, FLA_ZERO, y01_l );
// a01 = a01 - y01;
FLA_Axpy_external( FLA_MINUS_ONE, y01_l, a01 );
// alpha11 = alpha11 - a01' * b01 - b01' * a01;
FLA_Dot2cs_external( FLA_CONJUGATE, FLA_MINUS_ONE, a01, b01, FLA_ONE, alpha11 );
// alpha11 = inv(beta11) * alpha11 * inv(conj(beta11));
// = inv(beta11) * alpha11 * inv(beta11);
FLA_Inv_scal_external( beta11, alpha11 );
FLA_Inv_scal_external( beta11, alpha11 );
// a12t = a12t - b01' * A02;
// a12t^T = a12t^T - A02^T * conj(b01);
FLA_Gemvc_external( FLA_TRANSPOSE, FLA_CONJUGATE,
FLA_MINUS_ONE, A02, b01, FLA_ONE, a12t );
// a12t = inv(conj(beta11)) * a12t;
// a12t = inv(beta11) * a12t;
FLA_Inv_scal_external( beta11, a12t );
// a01 = a01 - y01;
FLA_Axpy_external( FLA_MINUS_ONE, y01_l, a01 );
// a01 = a01 * inv(beta11);
FLA_Inv_scal_external( beta11, a01 );
/*------------------------------------------------------------*/
FLA_Cont_with_3x3_to_2x2( &ATL, /**/ &ATR, A00, a01, /**/ A02,
a10t, alpha11, /**/ a12t,
/* ************** */ /* ************************ */
&ABL, /**/ &ABR, A20, a21, /**/ A22,
FLA_TL );
FLA_Cont_with_3x3_to_2x2( &BTL, /**/ &BTR, B00, b01, /**/ B02,
b10t, beta11, /**/ b12t,
/* ************** */ /* *********************** */
&BBL, /**/ &BBR, B20, b21, /**/ B22,
FLA_TL );
FLA_Cont_with_3x1_to_2x1( &yT, y01,
psi11,
/* ** */ /* ***** */
&yB, y21, FLA_TOP );
}
return FLA_SUCCESS;
}
FLA_Error FLA_Her2k_un_unb_var6( FLA_Obj alpha, FLA_Obj A, FLA_Obj B, FLA_Obj beta, FLA_Obj C )
{
FLA_Obj AT, A0,
AB, a1t,
A2;
FLA_Obj BT, B0,
BB, b1t,
B2;
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_2x1( B, &BT,
&BB, 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_2x1_to_3x1( BT, &B0,
&b1t,
/* ** */ /* ** */
BB, &B2, 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 * b1t' */
FLA_Gemvc_external( FLA_NO_TRANSPOSE, FLA_CONJUGATE, alpha, A0, b1t, FLA_ONE, c01 );
/* c12t = c12t + b1t * A2' */
FLA_Gemv_external( FLA_CONJ_NO_TRANSPOSE, alpha, A2, b1t, FLA_ONE, c12t );
/* gamma11 = gamma11 + a1t * b1t' + b1t * a1t' */
FLA_Dot2cs_external( FLA_CONJUGATE, alpha, a1t, b1t, FLA_ONE, gamma11 );
/*------------------------------------------------------------*/
FLA_Cont_with_3x1_to_2x1( &AT, A0,
/* ** */ /* ** */
a1t,
&AB, A2, FLA_BOTTOM );
FLA_Cont_with_3x1_to_2x1( &BT, B0,
/* ** */ /* ** */
b1t,
&BB, B2, 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_Dot2cs( FLA_Conj conj, FLA_Obj alpha, FLA_Obj x, FLA_Obj y, FLA_Obj beta, FLA_Obj rho )
{
return FLA_Dot2cs_external( conj, alpha, x, y, beta, rho );
}