|
libflame
revision_anchor
|
Go to the source code of this file.
Functions | |
| FLA_Error | FLA_Eig_gest_iu_blk_var1 (FLA_Obj A, FLA_Obj Y, FLA_Obj B, fla_eig_gest_t *cntl) |
| FLA_Error | FLA_Eig_gest_iu_blk_var2 (FLA_Obj A, FLA_Obj Y, FLA_Obj B, fla_eig_gest_t *cntl) |
| FLA_Error | FLA_Eig_gest_iu_blk_var3 (FLA_Obj A, FLA_Obj Y, FLA_Obj B, fla_eig_gest_t *cntl) |
| FLA_Error | FLA_Eig_gest_iu_blk_var4 (FLA_Obj A, FLA_Obj Y, FLA_Obj B, fla_eig_gest_t *cntl) |
| FLA_Error | FLA_Eig_gest_iu_blk_var5 (FLA_Obj A, FLA_Obj Y, FLA_Obj B, fla_eig_gest_t *cntl) |
| FLA_Error | FLA_Eig_gest_iu_unb_var1 (FLA_Obj A, FLA_Obj Y, FLA_Obj B) |
| FLA_Error | FLA_Eig_gest_iu_unb_var2 (FLA_Obj A, FLA_Obj Y, FLA_Obj B) |
| FLA_Error | FLA_Eig_gest_iu_unb_var3 (FLA_Obj A, FLA_Obj Y, FLA_Obj B) |
| FLA_Error | FLA_Eig_gest_iu_unb_var4 (FLA_Obj A, FLA_Obj Y, FLA_Obj B) |
| FLA_Error | FLA_Eig_gest_iu_unb_var5 (FLA_Obj A, FLA_Obj Y, FLA_Obj B) |
| FLA_Error | FLA_Eig_gest_iu_opt_var1 (FLA_Obj A, FLA_Obj Y, FLA_Obj B) |
| FLA_Error | FLA_Eig_gest_iu_ops_var1 (int m_AB, float *buff_A, int rs_A, int cs_A, float *buff_y, int inc_y, float *buff_B, int rs_B, int cs_B) |
| FLA_Error | FLA_Eig_gest_iu_opd_var1 (int m_AB, double *buff_A, int rs_A, int cs_A, double *buff_y, int inc_y, double *buff_B, int rs_B, int cs_B) |
| FLA_Error | FLA_Eig_gest_iu_opc_var1 (int m_AB, scomplex *buff_A, int rs_A, int cs_A, scomplex *buff_y, int inc_y, scomplex *buff_B, int rs_B, int cs_B) |
| FLA_Error | FLA_Eig_gest_iu_opz_var1 (int m_AB, dcomplex *buff_A, int rs_A, int cs_A, dcomplex *buff_y, int inc_y, dcomplex *buff_B, int rs_B, int cs_B) |
| FLA_Error | FLA_Eig_gest_iu_opt_var2 (FLA_Obj A, FLA_Obj Y, FLA_Obj B) |
| FLA_Error | FLA_Eig_gest_iu_ops_var2 (int m_AB, float *buff_A, int rs_A, int cs_A, float *buff_y, int inc_y, float *buff_B, int rs_B, int cs_B) |
| FLA_Error | FLA_Eig_gest_iu_opd_var2 (int m_AB, double *buff_A, int rs_A, int cs_A, double *buff_y, int inc_y, double *buff_B, int rs_B, int cs_B) |
| FLA_Error | FLA_Eig_gest_iu_opc_var2 (int m_AB, scomplex *buff_A, int rs_A, int cs_A, scomplex *buff_y, int inc_y, scomplex *buff_B, int rs_B, int cs_B) |
| FLA_Error | FLA_Eig_gest_iu_opz_var2 (int m_AB, dcomplex *buff_A, int rs_A, int cs_A, dcomplex *buff_y, int inc_y, dcomplex *buff_B, int rs_B, int cs_B) |
| FLA_Error | FLA_Eig_gest_iu_opt_var3 (FLA_Obj A, FLA_Obj Y, FLA_Obj B) |
| FLA_Error | FLA_Eig_gest_iu_ops_var3 (int m_AB, float *buff_A, int rs_A, int cs_A, float *buff_Y, int rs_Y, int cs_Y, float *buff_B, int rs_B, int cs_B) |
| FLA_Error | FLA_Eig_gest_iu_opd_var3 (int m_AB, double *buff_A, int rs_A, int cs_A, double *buff_Y, int rs_Y, int cs_Y, double *buff_B, int rs_B, int cs_B) |
| FLA_Error | FLA_Eig_gest_iu_opc_var3 (int m_AB, scomplex *buff_A, int rs_A, int cs_A, scomplex *buff_Y, int rs_Y, int cs_Y, scomplex *buff_B, int rs_B, int cs_B) |
| FLA_Error | FLA_Eig_gest_iu_opz_var3 (int m_AB, dcomplex *buff_A, int rs_A, int cs_A, dcomplex *buff_Y, int rs_Y, int cs_Y, dcomplex *buff_B, int rs_B, int cs_B) |
| FLA_Error | FLA_Eig_gest_iu_opt_var4 (FLA_Obj A, FLA_Obj Y, FLA_Obj B) |
| FLA_Error | FLA_Eig_gest_iu_ops_var4 (int m_AB, float *buff_A, int rs_A, int cs_A, float *buff_y, int inc_y, float *buff_B, int rs_B, int cs_B) |
| FLA_Error | FLA_Eig_gest_iu_opd_var4 (int m_AB, double *buff_A, int rs_A, int cs_A, double *buff_y, int inc_y, double *buff_B, int rs_B, int cs_B) |
| FLA_Error | FLA_Eig_gest_iu_opc_var4 (int m_AB, scomplex *buff_A, int rs_A, int cs_A, scomplex *buff_y, int inc_y, scomplex *buff_B, int rs_B, int cs_B) |
| FLA_Error | FLA_Eig_gest_iu_opz_var4 (int m_AB, dcomplex *buff_A, int rs_A, int cs_A, dcomplex *buff_y, int inc_y, dcomplex *buff_B, int rs_B, int cs_B) |
| FLA_Error | FLA_Eig_gest_iu_opt_var5 (FLA_Obj A, FLA_Obj Y, FLA_Obj B) |
| FLA_Error | FLA_Eig_gest_iu_ops_var5 (int m_AB, float *buff_A, int rs_A, int cs_A, float *buff_y, int inc_y, float *buff_B, int rs_B, int cs_B) |
| FLA_Error | FLA_Eig_gest_iu_opd_var5 (int m_AB, double *buff_A, int rs_A, int cs_A, double *buff_y, int inc_y, double *buff_B, int rs_B, int cs_B) |
| FLA_Error | FLA_Eig_gest_iu_opc_var5 (int m_AB, scomplex *buff_A, int rs_A, int cs_A, scomplex *buff_y, int inc_y, scomplex *buff_B, int rs_B, int cs_B) |
| FLA_Error | FLA_Eig_gest_iu_opz_var5 (int m_AB, dcomplex *buff_A, int rs_A, int cs_A, dcomplex *buff_y, int inc_y, dcomplex *buff_B, int rs_B, int cs_B) |
| FLA_Error FLA_Eig_gest_iu_blk_var1 | ( | FLA_Obj | A, |
| FLA_Obj | Y, | ||
| FLA_Obj | B, | ||
| fla_eig_gest_t * | cntl | ||
| ) |
References FLA_Axpy_internal(), FLA_Cont_with_3x1_to_2x1(), FLA_Cont_with_3x3_to_2x2(), FLA_Determine_blocksize(), FLA_Eig_gest_internal(), FLA_Hemm_internal(), FLA_Her2k_internal(), FLA_MINUS_ONE, FLA_MINUS_ONE_HALF, FLA_Obj_length(), FLA_ONE, FLA_Part_1x2(), FLA_Part_2x1(), FLA_Part_2x2(), FLA_Repart_2x1_to_3x1(), FLA_Repart_2x2_to_3x3(), FLA_Trsm_internal(), and FLA_ZERO.
Referenced by FLA_Eig_gest_iu().
{
FLA_Obj ATL, ATR, A00, A01, A02,
ABL, ABR, A10, A11, A12,
A20, A21, A22;
FLA_Obj BTL, BTR, B00, B01, B02,
BBL, BBR, B10, B11, B12,
B20, B21, B22;
FLA_Obj YT, Y01,
YB, Y11,
Y21;
FLA_Obj Y01_l, Y01_r;
dim_t b;
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 ) ){
b = FLA_Determine_blocksize( ABR, FLA_BR, FLA_Cntl_blocksize( cntl ) );
FLA_Repart_2x2_to_3x3( ATL, /**/ ATR, &A00, /**/ &A01, &A02,
/* ************* */ /* ******************** */
&A10, /**/ &A11, &A12,
ABL, /**/ ABR, &A20, /**/ &A21, &A22,
b, b, FLA_BR );
FLA_Repart_2x2_to_3x3( BTL, /**/ BTR, &B00, /**/ &B01, &B02,
/* ************* */ /* ******************** */
&B10, /**/ &B11, &B12,
BBL, /**/ BBR, &B20, /**/ &B21, &B22,
b, b, FLA_BR );
FLA_Repart_2x1_to_3x1( YT, &Y01,
/* ** */ /* *** */
&Y11,
YB, &Y21, b, FLA_BOTTOM );
/*------------------------------------------------------------*/
FLA_Part_1x2( Y01, &Y01_l, &Y01_r, b, FLA_LEFT );
// Y01 = A00 * B01;
FLA_Hemm_internal( FLA_LEFT, FLA_UPPER_TRIANGULAR,
FLA_ONE, A00, B01, FLA_ZERO, Y01_l,
FLA_Cntl_sub_hemm( cntl ) );
// A01 = inv( triu( B00 )' ) * A01;
FLA_Trsm_internal( FLA_LEFT, FLA_UPPER_TRIANGULAR,
FLA_CONJ_TRANSPOSE, FLA_NONUNIT_DIAG,
FLA_ONE, B00, A01,
FLA_Cntl_sub_trsm1( cntl ) );
// A01 = A01 - 1/2 * Y01;
FLA_Axpy_internal( FLA_MINUS_ONE_HALF, Y01_l, A01,
FLA_Cntl_sub_axpy1( cntl ) );
// A11 = A11 - A01' * B01 - B01' * A01;
FLA_Her2k_internal( FLA_UPPER_TRIANGULAR, FLA_CONJ_TRANSPOSE,
FLA_MINUS_ONE, A01, B01, FLA_ONE, A11,
FLA_Cntl_sub_her2k( cntl ) );
// A11 = inv( triu( B11 )' ) * A11 * inv( triu( B11 ) );
FLA_Eig_gest_internal( FLA_INVERSE, FLA_UPPER_TRIANGULAR,
A11, Y11, B11,
FLA_Cntl_sub_eig_gest( cntl ) );
// A01 = A01 - 1/2 * Y01;
FLA_Axpy_internal( FLA_MINUS_ONE_HALF, Y01_l, A01,
FLA_Cntl_sub_axpy2( cntl ) );
// A01 = A01 * inv( triu( B11 ) );
FLA_Trsm_internal( FLA_RIGHT, FLA_UPPER_TRIANGULAR,
FLA_NO_TRANSPOSE, FLA_NONUNIT_DIAG,
FLA_ONE, B11, A01,
FLA_Cntl_sub_trsm2( cntl ) );
/*------------------------------------------------------------*/
FLA_Cont_with_3x3_to_2x2( &ATL, /**/ &ATR, A00, A01, /**/ A02,
A10, A11, /**/ A12,
/* ************** */ /* ****************** */
&ABL, /**/ &ABR, A20, A21, /**/ A22,
FLA_TL );
FLA_Cont_with_3x3_to_2x2( &BTL, /**/ &BTR, B00, B01, /**/ B02,
B10, B11, /**/ B12,
/* ************** */ /* ****************** */
&BBL, /**/ &BBR, B20, B21, /**/ B22,
FLA_TL );
FLA_Cont_with_3x1_to_2x1( &YT, Y01,
Y11,
/* ** */ /* *** */
&YB, Y21, FLA_TOP );
}
return FLA_SUCCESS;
}
| FLA_Error FLA_Eig_gest_iu_blk_var2 | ( | FLA_Obj | A, |
| FLA_Obj | Y, | ||
| FLA_Obj | B, | ||
| fla_eig_gest_t * | cntl | ||
| ) |
References FLA_Axpy_internal(), FLA_Cont_with_3x1_to_2x1(), FLA_Cont_with_3x3_to_2x2(), FLA_Determine_blocksize(), FLA_Eig_gest_internal(), FLA_Gemm_internal(), FLA_Hemm_internal(), FLA_Her2k_internal(), FLA_MINUS_ONE, FLA_Obj_length(), FLA_ONE, FLA_ONE_HALF, FLA_Part_1x2(), FLA_Part_2x1(), FLA_Part_2x2(), FLA_Repart_2x1_to_3x1(), FLA_Repart_2x2_to_3x3(), FLA_Trsm_internal(), and FLA_ZERO.
Referenced by FLA_Eig_gest_iu().
{
FLA_Obj ATL, ATR, A00, A01, A02,
ABL, ABR, A10, A11, A12,
A20, A21, A22;
FLA_Obj BTL, BTR, B00, B01, B02,
BBL, BBR, B10, B11, B12,
B20, B21, B22;
FLA_Obj YT, Y01,
YB, Y11,
Y21;
FLA_Obj Y01_l, Y01_r;
dim_t b;
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 ) ){
b = FLA_Determine_blocksize( ABR, FLA_BR, FLA_Cntl_blocksize( cntl ) );
FLA_Repart_2x2_to_3x3( ATL, /**/ ATR, &A00, /**/ &A01, &A02,
/* ************* */ /* ******************** */
&A10, /**/ &A11, &A12,
ABL, /**/ ABR, &A20, /**/ &A21, &A22,
b, b, FLA_BR );
FLA_Repart_2x2_to_3x3( BTL, /**/ BTR, &B00, /**/ &B01, &B02,
/* ************* */ /* ******************** */
&B10, /**/ &B11, &B12,
BBL, /**/ BBR, &B20, /**/ &B21, &B22,
b, b, FLA_BR );
FLA_Repart_2x1_to_3x1( YT, &Y01,
/* ** */ /* *** */
&Y11,
YB, &Y21, b, FLA_BOTTOM );
/*------------------------------------------------------------*/
FLA_Part_1x2( Y01, &Y01_l, &Y01_r, b, FLA_LEFT );
// Y01 = 1/2 * A00 * B01;
FLA_Hemm_internal( FLA_LEFT, FLA_UPPER_TRIANGULAR,
FLA_ONE_HALF, A00, B01, FLA_ZERO, Y01_l,
FLA_Cntl_sub_hemm( cntl ) );
// A01 = A01 - Y01;
FLA_Axpy_internal( FLA_MINUS_ONE, Y01_l, A01,
FLA_Cntl_sub_axpy1( cntl ) );
// A11 = A11 - A01' * B01 - B01' * A01;
FLA_Her2k_internal( FLA_UPPER_TRIANGULAR, FLA_CONJ_TRANSPOSE,
FLA_MINUS_ONE, A01, B01, FLA_ONE, A11,
FLA_Cntl_sub_her2k( cntl ) );
// A11 = inv( triu( B11 )' ) * A11 * inv( triu( B11 ) );
FLA_Eig_gest_internal( FLA_INVERSE, FLA_UPPER_TRIANGULAR,
A11, Y11, B11,
FLA_Cntl_sub_eig_gest( cntl ) );
// A12 = A12 - B01' * A02;
FLA_Gemm_internal( FLA_CONJ_TRANSPOSE, FLA_NO_TRANSPOSE,
FLA_MINUS_ONE, B01, A02, FLA_ONE, A12,
FLA_Cntl_sub_gemm1( cntl ) );
// A12 = inv( triu( B11 )' ) * A12;
FLA_Trsm_internal( FLA_LEFT, FLA_UPPER_TRIANGULAR,
FLA_CONJ_TRANSPOSE, FLA_NONUNIT_DIAG,
FLA_ONE, B11, A12,
FLA_Cntl_sub_trsm1( cntl ) );
// A01 = A01 - Y01;
FLA_Axpy_internal( FLA_MINUS_ONE, Y01_l, A01,
FLA_Cntl_sub_axpy2( cntl ) );
// A01 = A01 * inv( triu( B11 ) );
FLA_Trsm_internal( FLA_RIGHT, FLA_UPPER_TRIANGULAR,
FLA_NO_TRANSPOSE, FLA_NONUNIT_DIAG,
FLA_ONE, B11, A01,
FLA_Cntl_sub_trsm2( cntl ) );
/*------------------------------------------------------------*/
FLA_Cont_with_3x3_to_2x2( &ATL, /**/ &ATR, A00, A01, /**/ A02,
A10, A11, /**/ A12,
/* ************** */ /* ****************** */
&ABL, /**/ &ABR, A20, A21, /**/ A22,
FLA_TL );
FLA_Cont_with_3x3_to_2x2( &BTL, /**/ &BTR, B00, B01, /**/ B02,
B10, B11, /**/ B12,
/* ************** */ /* ****************** */
&BBL, /**/ &BBR, B20, B21, /**/ B22,
FLA_TL );
FLA_Cont_with_3x1_to_2x1( &YT, Y01,
Y11,
/* ** */ /* *** */
&YB, Y21, FLA_TOP );
}
return FLA_SUCCESS;
}
| FLA_Error FLA_Eig_gest_iu_blk_var3 | ( | FLA_Obj | A, |
| FLA_Obj | Y, | ||
| FLA_Obj | B, | ||
| fla_eig_gest_t * | cntl | ||
| ) |
References FLA_Axpy_internal(), FLA_Cont_with_3x3_to_2x2(), FLA_Determine_blocksize(), FLA_Eig_gest_internal(), FLA_Gemm_internal(), FLA_Hemm_internal(), FLA_Her2k_internal(), FLA_MINUS_ONE, FLA_MINUS_ONE_HALF, FLA_Obj_length(), FLA_ONE, FLA_Part_2x2(), FLA_Repart_2x2_to_3x3(), FLA_Trsm_internal(), and FLA_ZERO.
Referenced by FLA_Eig_gest_iu().
{
FLA_Obj ATL, ATR, A00, A01, A02,
ABL, ABR, A10, A11, A12,
A20, A21, A22;
FLA_Obj BTL, BTR, B00, B01, B02,
BBL, BBR, B10, B11, B12,
B20, B21, B22;
FLA_Obj YTL, YTR, Y00, Y01, Y02,
YBL, YBR, Y10, Y11, Y12,
Y20, Y21, Y22;
dim_t b;
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_2x2( Y, &YTL, &YTR,
&YBL, &YBR, 0, 0, FLA_TL );
while ( FLA_Obj_length( ATL ) < FLA_Obj_length( A ) ){
b = FLA_Determine_blocksize( ABR, FLA_BR, FLA_Cntl_blocksize( cntl ) );
FLA_Repart_2x2_to_3x3( ATL, /**/ ATR, &A00, /**/ &A01, &A02,
/* ************* */ /* ******************** */
&A10, /**/ &A11, &A12,
ABL, /**/ ABR, &A20, /**/ &A21, &A22,
b, b, FLA_BR );
FLA_Repart_2x2_to_3x3( BTL, /**/ BTR, &B00, /**/ &B01, &B02,
/* ************* */ /* ******************** */
&B10, /**/ &B11, &B12,
BBL, /**/ BBR, &B20, /**/ &B21, &B22,
b, b, FLA_BR );
FLA_Repart_2x2_to_3x3( YTL, /**/ YTR, &Y00, /**/ &Y01, &Y02,
/* ************* */ /* ******************** */
&Y10, /**/ &Y11, &Y12,
YBL, /**/ YBR, &Y20, /**/ &Y21, &Y22,
b, b, FLA_BR );
/*------------------------------------------------------------*/
// A01 = A01 - 1/2 * Y01;
FLA_Axpy_internal( FLA_MINUS_ONE_HALF, Y01, A01,
FLA_Cntl_sub_axpy1( cntl ) );
// A11 = A11 - A01' * B01 - B01' * A01;
FLA_Her2k_internal( FLA_UPPER_TRIANGULAR, FLA_CONJ_TRANSPOSE,
FLA_MINUS_ONE, A01, B01, FLA_ONE, A11,
FLA_Cntl_sub_her2k( cntl ) );
// A11 = inv( triu( B11 )' ) * A11 * inv( triu( B11 ) );
FLA_Eig_gest_internal( FLA_INVERSE, FLA_UPPER_TRIANGULAR,
A11, Y11, B11,
FLA_Cntl_sub_eig_gest( cntl ) );
// A12 = A12 - B01' * A02;
FLA_Gemm_internal( FLA_CONJ_TRANSPOSE, FLA_NO_TRANSPOSE,
FLA_MINUS_ONE, B01, A02, FLA_ONE, A12,
FLA_Cntl_sub_gemm1( cntl ) );
// A12 = A12 * inv( triu( B11 )' );
FLA_Trsm_internal( FLA_LEFT, FLA_UPPER_TRIANGULAR,
FLA_CONJ_TRANSPOSE, FLA_NONUNIT_DIAG,
FLA_ONE, B11, A12,
FLA_Cntl_sub_trsm1( cntl ) );
// A01 = A01 - 1/2 * Y01;
FLA_Axpy_internal( FLA_MINUS_ONE_HALF, Y01, A01,
FLA_Cntl_sub_axpy2( cntl ) );
// A01 = A01 * inv( triu( B11 ) );
FLA_Trsm_internal( FLA_RIGHT, FLA_UPPER_TRIANGULAR,
FLA_NO_TRANSPOSE, FLA_NONUNIT_DIAG,
FLA_ONE, B11, A01,
FLA_Cntl_sub_trsm2( cntl ) );
// Y02 = Y02 + A01 * B12;
FLA_Gemm_internal( FLA_NO_TRANSPOSE, FLA_NO_TRANSPOSE,
FLA_ONE, A01, B12, FLA_ONE, Y02,
FLA_Cntl_sub_gemm2( cntl ) );
// Y12 = A11 * B12;
FLA_Hemm_internal( FLA_LEFT, FLA_UPPER_TRIANGULAR,
FLA_ONE, A11, B12, FLA_ZERO, Y12,
FLA_Cntl_sub_hemm( cntl ) );
// Y12 = Y12 + A01' * B02;
FLA_Gemm_internal( FLA_CONJ_TRANSPOSE, FLA_NO_TRANSPOSE,
FLA_ONE, A01, B02, FLA_ONE, Y12,
FLA_Cntl_sub_gemm3( cntl ) );
/*------------------------------------------------------------*/
FLA_Cont_with_3x3_to_2x2( &ATL, /**/ &ATR, A00, A01, /**/ A02,
A10, A11, /**/ A12,
/* ************** */ /* ****************** */
&ABL, /**/ &ABR, A20, A21, /**/ A22,
FLA_TL );
FLA_Cont_with_3x3_to_2x2( &BTL, /**/ &BTR, B00, B01, /**/ B02,
B10, B11, /**/ B12,
/* ************** */ /* ****************** */
&BBL, /**/ &BBR, B20, B21, /**/ B22,
FLA_TL );
FLA_Cont_with_3x3_to_2x2( &YTL, /**/ &YTR, Y00, Y01, /**/ Y02,
Y10, Y11, /**/ Y12,
/* ************** */ /* ****************** */
&YBL, /**/ &YBR, Y20, Y21, /**/ Y22,
FLA_TL );
}
return FLA_SUCCESS;
}
| FLA_Error FLA_Eig_gest_iu_blk_var4 | ( | FLA_Obj | A, |
| FLA_Obj | Y, | ||
| FLA_Obj | B, | ||
| fla_eig_gest_t * | cntl | ||
| ) |
References FLA_Axpy_internal(), FLA_Cont_with_1x3_to_1x2(), FLA_Cont_with_3x3_to_2x2(), FLA_Determine_blocksize(), FLA_Eig_gest_internal(), FLA_Gemm_internal(), FLA_Hemm_internal(), FLA_Her2k_internal(), FLA_MINUS_ONE, FLA_MINUS_ONE_HALF, FLA_Obj_length(), FLA_ONE, FLA_Part_1x2(), FLA_Part_2x1(), FLA_Part_2x2(), FLA_Repart_1x2_to_1x3(), FLA_Repart_2x2_to_3x3(), FLA_Trsm_internal(), and FLA_ZERO.
Referenced by FLA_Eig_gest_iu().
{
FLA_Obj ATL, ATR, A00, A01, A02,
ABL, ABR, A10, A11, A12,
A20, A21, A22;
FLA_Obj BTL, BTR, B00, B01, B02,
BBL, BBR, B10, B11, B12,
B20, B21, B22;
FLA_Obj YL, YR, Y10, Y11, Y12;
FLA_Obj Y12_t,
Y12_b;
dim_t b;
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_1x2( Y, &YL, &YR, 0, FLA_LEFT );
while ( FLA_Obj_length( ATL ) < FLA_Obj_length( A ) ){
b = FLA_Determine_blocksize( ABR, FLA_BR, FLA_Cntl_blocksize( cntl ) );
FLA_Repart_2x2_to_3x3( ATL, /**/ ATR, &A00, /**/ &A01, &A02,
/* ************* */ /* ******************** */
&A10, /**/ &A11, &A12,
ABL, /**/ ABR, &A20, /**/ &A21, &A22,
b, b, FLA_BR );
FLA_Repart_2x2_to_3x3( BTL, /**/ BTR, &B00, /**/ &B01, &B02,
/* ************* */ /* ******************** */
&B10, /**/ &B11, &B12,
BBL, /**/ BBR, &B20, /**/ &B21, &B22,
b, b, FLA_BR );
FLA_Repart_1x2_to_1x3( YL, /**/ YR, &Y10, /**/ &Y11, &Y12,
b, FLA_RIGHT );
/*------------------------------------------------------------*/
FLA_Part_2x1( Y12, &Y12_t,
&Y12_b, b, FLA_TOP );
// A01 = A01 * inv( triu( B11 ) );
FLA_Trsm_internal( FLA_RIGHT, FLA_UPPER_TRIANGULAR,
FLA_NO_TRANSPOSE, FLA_NONUNIT_DIAG,
FLA_ONE, B11, A01,
FLA_Cntl_sub_trsm1( cntl ) );
// A02 = A02 - A01 * B12;
FLA_Gemm_internal( FLA_NO_TRANSPOSE, FLA_NO_TRANSPOSE,
FLA_MINUS_ONE, A01, B12, FLA_ONE, A02,
FLA_Cntl_sub_gemm1( cntl ) );
// A11 = inv( triu( B11 )' ) * A11 * inv( triu( B11 ) );
FLA_Eig_gest_internal( FLA_INVERSE, FLA_UPPER_TRIANGULAR,
A11, Y11, B11,
FLA_Cntl_sub_eig_gest( cntl ) );
// Y12 = A11 * B12;
FLA_Hemm_internal( FLA_LEFT, FLA_UPPER_TRIANGULAR,
FLA_ONE, A11, B12, FLA_ZERO, Y12_t,
FLA_Cntl_sub_hemm( cntl ) );
// A12 = inv( triu( B11 )' ) * A12;
FLA_Trsm_internal( FLA_LEFT, FLA_UPPER_TRIANGULAR,
FLA_CONJ_TRANSPOSE, FLA_NONUNIT_DIAG,
FLA_ONE, B11, A12,
FLA_Cntl_sub_trsm2( cntl ) );
// A12 = A12 - 1/2 * Y12;
FLA_Axpy_internal( FLA_MINUS_ONE_HALF, Y12_t, A12,
FLA_Cntl_sub_axpy1( cntl ) );
// A22 = A22 - A12' * B12 - B12' * A12;
FLA_Her2k_internal( FLA_UPPER_TRIANGULAR, FLA_CONJ_TRANSPOSE,
FLA_MINUS_ONE, A12, B12, FLA_ONE, A22,
FLA_Cntl_sub_her2k( cntl ) );
// A12 = A12 - 1/2 * Y12;
FLA_Axpy_internal( FLA_MINUS_ONE_HALF, Y12_t, A12,
FLA_Cntl_sub_axpy2( cntl ) );
/*------------------------------------------------------------*/
FLA_Cont_with_3x3_to_2x2( &ATL, /**/ &ATR, A00, A01, /**/ A02,
A10, A11, /**/ A12,
/* ************** */ /* ****************** */
&ABL, /**/ &ABR, A20, A21, /**/ A22,
FLA_TL );
FLA_Cont_with_3x3_to_2x2( &BTL, /**/ &BTR, B00, B01, /**/ B02,
B10, B11, /**/ B12,
/* ************** */ /* ****************** */
&BBL, /**/ &BBR, B20, B21, /**/ B22,
FLA_TL );
FLA_Cont_with_1x3_to_1x2( &YL, /**/ &YR, Y10, Y11, /**/ Y12,
FLA_LEFT );
}
return FLA_SUCCESS;
}
| FLA_Error FLA_Eig_gest_iu_blk_var5 | ( | FLA_Obj | A, |
| FLA_Obj | Y, | ||
| FLA_Obj | B, | ||
| fla_eig_gest_t * | cntl | ||
| ) |
References FLA_Axpy_internal(), FLA_Cont_with_1x3_to_1x2(), FLA_Cont_with_3x3_to_2x2(), FLA_Determine_blocksize(), FLA_Eig_gest_internal(), FLA_Hemm_internal(), FLA_Her2k_internal(), FLA_MINUS_ONE, FLA_MINUS_ONE_HALF, FLA_Obj_length(), FLA_ONE, FLA_Part_1x2(), FLA_Part_2x1(), FLA_Part_2x2(), FLA_Repart_1x2_to_1x3(), FLA_Repart_2x2_to_3x3(), FLA_Trsm_internal(), and FLA_ZERO.
Referenced by FLA_Eig_gest_iu().
{
FLA_Obj ATL, ATR, A00, A01, A02,
ABL, ABR, A10, A11, A12,
A20, A21, A22;
FLA_Obj BTL, BTR, B00, B01, B02,
BBL, BBR, B10, B11, B12,
B20, B21, B22;
FLA_Obj YL, YR, Y10, Y11, Y12;
FLA_Obj Y12_t,
Y12_b;
dim_t b;
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_1x2( Y, &YL, &YR, 0, FLA_LEFT );
while ( FLA_Obj_length( ATL ) < FLA_Obj_length( A ) ){
b = FLA_Determine_blocksize( ABR, FLA_BR, FLA_Cntl_blocksize( cntl ) );
FLA_Repart_2x2_to_3x3( ATL, /**/ ATR, &A00, /**/ &A01, &A02,
/* ************* */ /* ******************** */
&A10, /**/ &A11, &A12,
ABL, /**/ ABR, &A20, /**/ &A21, &A22,
b, b, FLA_BR );
FLA_Repart_2x2_to_3x3( BTL, /**/ BTR, &B00, /**/ &B01, &B02,
/* ************* */ /* ******************** */
&B10, /**/ &B11, &B12,
BBL, /**/ BBR, &B20, /**/ &B21, &B22,
b, b, FLA_BR );
FLA_Repart_1x2_to_1x3( YL, /**/ YR, &Y10, /**/ &Y11, &Y12,
b, FLA_RIGHT );
/*------------------------------------------------------------*/
FLA_Part_2x1( Y12, &Y12_t,
&Y12_b, b, FLA_TOP );
// A11 = inv( triu( B11 )' ) * A11 * inv( triu( B11 ) );
FLA_Eig_gest_internal( FLA_INVERSE, FLA_UPPER_TRIANGULAR,
A11, Y11, B11,
FLA_Cntl_sub_eig_gest( cntl ) );
// Y12 = A11 * B12;
FLA_Hemm_internal( FLA_LEFT, FLA_UPPER_TRIANGULAR,
FLA_ONE, A11, B12, FLA_ZERO, Y12_t,
FLA_Cntl_sub_hemm( cntl ) );
// A12 = inv( triu( B11 )' ) * A12;
FLA_Trsm_internal( FLA_LEFT, FLA_UPPER_TRIANGULAR,
FLA_CONJ_TRANSPOSE, FLA_NONUNIT_DIAG,
FLA_ONE, B11, A12,
FLA_Cntl_sub_trsm1( cntl ) );
// A12 = A12 - 1/2 * Y12;
FLA_Axpy_internal( FLA_MINUS_ONE_HALF, Y12_t, A12,
FLA_Cntl_sub_axpy1( cntl ) );
// A22 = A22 - A12' * B12 - B12' * A12;
FLA_Her2k_internal( FLA_UPPER_TRIANGULAR, FLA_CONJ_TRANSPOSE,
FLA_MINUS_ONE, A12, B12, FLA_ONE, A22,
FLA_Cntl_sub_her2k( cntl ) );
// A12 = A12 - 1/2 * Y12;
FLA_Axpy_internal( FLA_MINUS_ONE_HALF, Y12_t, A12,
FLA_Cntl_sub_axpy2( cntl ) );
// A12 = A12 * inv( triu( B22 ) );
FLA_Trsm_internal( FLA_RIGHT, FLA_UPPER_TRIANGULAR,
FLA_NO_TRANSPOSE, FLA_NONUNIT_DIAG,
FLA_ONE, B22, A12,
FLA_Cntl_sub_trsm2( cntl ) );
/*------------------------------------------------------------*/
FLA_Cont_with_3x3_to_2x2( &ATL, /**/ &ATR, A00, A01, /**/ A02,
A10, A11, /**/ A12,
/* ************** */ /* ****************** */
&ABL, /**/ &ABR, A20, A21, /**/ A22,
FLA_TL );
FLA_Cont_with_3x3_to_2x2( &BTL, /**/ &BTR, B00, B01, /**/ B02,
B10, B11, /**/ B12,
/* ************** */ /* ****************** */
&BBL, /**/ &BBR, B20, B21, /**/ B22,
FLA_TL );
FLA_Cont_with_1x3_to_1x2( &YL, /**/ &YR, Y10, Y11, /**/ Y12,
FLA_LEFT );
}
return FLA_SUCCESS;
}
| FLA_Error FLA_Eig_gest_iu_opc_var1 | ( | int | m_AB, |
| scomplex * | buff_A, | ||
| int | rs_A, | ||
| int | cs_A, | ||
| scomplex * | buff_y, | ||
| int | inc_y, | ||
| scomplex * | buff_B, | ||
| int | rs_B, | ||
| int | cs_B | ||
| ) |
References bl1_caxpyv(), bl1_cdot2s(), bl1_chemv(), bl1_cinvscalv(), bl1_ctrsv(), BLIS1_CONJ_TRANSPOSE, BLIS1_CONJUGATE, BLIS1_NO_CONJUGATE, BLIS1_NONUNIT_DIAG, BLIS1_UPPER_TRIANGULAR, FLA_MINUS_ONE, FLA_MINUS_ONE_HALF, FLA_ONE, and FLA_ZERO.
Referenced by FLA_Eig_gest_iu_opt_var1().
{
scomplex* buff_1 = FLA_COMPLEX_PTR( FLA_ONE );
scomplex* buff_0 = FLA_COMPLEX_PTR( FLA_ZERO );
scomplex* buff_m1 = FLA_COMPLEX_PTR( FLA_MINUS_ONE );
scomplex* buff_m1h = FLA_COMPLEX_PTR( FLA_MINUS_ONE_HALF );
int i;
for ( i = 0; i < m_AB; ++i )
{
scomplex* A00 = buff_A + (0 )*cs_A + (0 )*rs_A;
scomplex* a01 = buff_A + (i )*cs_A + (0 )*rs_A;
scomplex* alpha11 = buff_A + (i )*cs_A + (i )*rs_A;
scomplex* y01 = buff_y + (0 )*inc_y;
scomplex* B00 = buff_B + (0 )*cs_B + (0 )*rs_B;
scomplex* b01 = buff_B + (i )*cs_B + (0 )*rs_B;
scomplex* beta11 = buff_B + (i )*cs_B + (i )*rs_B;
int m_behind = i;
/*------------------------------------------------------------*/
// FLA_Hemvc_external( FLA_UPPER_TRIANGULAR, FLA_NO_CONJUGATE,
// FLA_ONE, A00, b01, FLA_ZERO, y01_l );
bl1_chemv( BLIS1_UPPER_TRIANGULAR,
BLIS1_NO_CONJUGATE,
m_behind,
buff_1,
A00, rs_A, cs_A,
b01, rs_B,
buff_0,
y01, inc_y );
// FLA_Trsv_external( FLA_UPPER_TRIANGULAR, FLA_CONJ_TRANSPOSE, FLA_NONUNIT_DIAG,
// B00, a01 );
bl1_ctrsv( BLIS1_UPPER_TRIANGULAR,
BLIS1_CONJ_TRANSPOSE,
BLIS1_NONUNIT_DIAG,
m_behind,
B00, rs_B, cs_B,
a01, rs_A );
// FLA_Axpy_external( FLA_MINUS_ONE_HALF, y01_l, a01 );
bl1_caxpyv( BLIS1_NO_CONJUGATE,
m_behind,
buff_m1h,
y01, inc_y,
a01, rs_A );
// FLA_Dot2cs_external( FLA_CONJUGATE, FLA_MINUS_ONE, a01, b01, FLA_ONE, alpha11 );
bl1_cdot2s( BLIS1_CONJUGATE,
m_behind,
buff_m1,
a01, rs_A,
b01, rs_B,
buff_1,
alpha11 );
// FLA_Inv_scal_external( beta11, alpha11 );
// FLA_Inv_scal_external( beta11, alpha11 );
bl1_cinvscals( beta11, alpha11 );
bl1_cinvscals( beta11, alpha11 );
// FLA_Axpy_external( FLA_MINUS_ONE_HALF, y01_l, a01 );
bl1_caxpyv( BLIS1_NO_CONJUGATE,
m_behind,
buff_m1h,
y01, inc_y,
a01, rs_A );
// FLA_Inv_scal_external( beta11, a01 );
bl1_cinvscalv( BLIS1_NO_CONJUGATE,
m_behind,
beta11,
a01, rs_A );
/*------------------------------------------------------------*/
}
return FLA_SUCCESS;
}
| FLA_Error FLA_Eig_gest_iu_opc_var2 | ( | int | m_AB, |
| scomplex * | buff_A, | ||
| int | rs_A, | ||
| int | cs_A, | ||
| scomplex * | buff_y, | ||
| int | inc_y, | ||
| scomplex * | buff_B, | ||
| int | rs_B, | ||
| int | cs_B | ||
| ) |
References bl1_caxpyv(), bl1_cdot2s(), bl1_cgemv(), bl1_chemv(), bl1_cinvscalv(), BLIS1_CONJUGATE, BLIS1_NO_CONJUGATE, BLIS1_TRANSPOSE, BLIS1_UPPER_TRIANGULAR, FLA_MINUS_ONE, FLA_ONE, FLA_ONE_HALF, and FLA_ZERO.
Referenced by FLA_Eig_gest_iu_opt_var2().
{
scomplex* buff_1 = FLA_COMPLEX_PTR( FLA_ONE );
scomplex* buff_1h = FLA_COMPLEX_PTR( FLA_ONE_HALF );
scomplex* buff_0 = FLA_COMPLEX_PTR( FLA_ZERO );
scomplex* buff_m1 = FLA_COMPLEX_PTR( FLA_MINUS_ONE );
int i;
for ( i = 0; i < m_AB; ++i )
{
scomplex* A00 = buff_A + (0 )*cs_A + (0 )*rs_A;
scomplex* a01 = buff_A + (i )*cs_A + (0 )*rs_A;
scomplex* alpha11 = buff_A + (i )*cs_A + (i )*rs_A;
scomplex* A02 = buff_A + (i+1)*cs_A + (0 )*rs_A;
scomplex* a12t = buff_A + (i+1)*cs_A + (i )*rs_A;
scomplex* y01 = buff_y + (0 )*inc_y;
scomplex* b01 = buff_B + (i )*cs_B + (0 )*rs_B;
scomplex* beta11 = buff_B + (i )*cs_B + (i )*rs_B;
int m_ahead = m_AB - i - 1;
int m_behind = i;
/*------------------------------------------------------------*/
// FLA_Hemvc_external( FLA_UPPER_TRIANGULAR, FLA_NO_CONJUGATE,
// FLA_ONE_HALF, A00, b01, FLA_ZERO, y01_l );
bl1_chemv( BLIS1_UPPER_TRIANGULAR,
BLIS1_NO_CONJUGATE,
m_behind,
buff_1h,
A00, rs_A, cs_A,
b01, rs_B,
buff_0,
y01, inc_y );
// FLA_Axpy_external( FLA_MINUS_ONE, y01_l, a01 );
bl1_caxpyv( BLIS1_NO_CONJUGATE,
m_behind,
buff_m1,
y01, inc_y,
a01, rs_A );
// FLA_Dot2cs_external( FLA_CONJUGATE, FLA_MINUS_ONE, a01, b01, FLA_ONE, alpha11 );
bl1_cdot2s( BLIS1_CONJUGATE,
m_behind,
buff_m1,
a01, rs_A,
b01, rs_B,
buff_1,
alpha11 );
// FLA_Inv_scal_external( beta11, alpha11 );
// FLA_Inv_scal_external( beta11, alpha11 );
bl1_cinvscals( beta11, alpha11 );
bl1_cinvscals( beta11, alpha11 );
// FLA_Gemvc_external( FLA_TRANSPOSE, FLA_CONJUGATE,
// FLA_MINUS_ONE, A02, b01, FLA_ONE, a12t );
bl1_cgemv( BLIS1_TRANSPOSE,
BLIS1_CONJUGATE,
m_behind,
m_ahead,
buff_m1,
A02, rs_A, cs_A,
b01, rs_B,
buff_1,
a12t, cs_A );
// FLA_Inv_scal_external( beta11, a12t );
bl1_cinvscalv( BLIS1_NO_CONJUGATE,
m_ahead,
beta11,
a12t, cs_A );
// FLA_Axpy_external( FLA_MINUS_ONE, y01_l, a01 );
bl1_caxpyv( BLIS1_NO_CONJUGATE,
m_behind,
buff_m1,
y01, inc_y,
a01, rs_A );
// FLA_Inv_scal_external( beta11, a01 );
bl1_cinvscalv( BLIS1_NO_CONJUGATE,
m_behind,
beta11,
a01, rs_A );
/*------------------------------------------------------------*/
}
return FLA_SUCCESS;
}
| FLA_Error FLA_Eig_gest_iu_opc_var3 | ( | int | m_AB, |
| scomplex * | buff_A, | ||
| int | rs_A, | ||
| int | cs_A, | ||
| scomplex * | buff_Y, | ||
| int | rs_Y, | ||
| int | cs_Y, | ||
| scomplex * | buff_B, | ||
| int | rs_B, | ||
| int | cs_B | ||
| ) |
References bl1_caxpyv(), bl1_ccopyv(), bl1_cdot2s(), bl1_cgemv(), bl1_cger(), bl1_cinvscalv(), bl1_cscalv(), BLIS1_CONJUGATE, BLIS1_NO_CONJUGATE, BLIS1_TRANSPOSE, FLA_MINUS_ONE, FLA_MINUS_ONE_HALF, and FLA_ONE.
Referenced by FLA_Eig_gest_iu_opt_var3().
{
scomplex* buff_1 = FLA_COMPLEX_PTR( FLA_ONE );
scomplex* buff_m1 = FLA_COMPLEX_PTR( FLA_MINUS_ONE );
scomplex* buff_m1h = FLA_COMPLEX_PTR( FLA_MINUS_ONE_HALF );
int i;
for ( i = 0; i < m_AB; ++i )
{
scomplex* a01 = buff_A + (i )*cs_A + (0 )*rs_A;
scomplex* alpha11 = buff_A + (i )*cs_A + (i )*rs_A;
scomplex* A02 = buff_A + (i+1)*cs_A + (0 )*rs_A;
scomplex* a12t = buff_A + (i+1)*cs_A + (i )*rs_A;
scomplex* y01 = buff_Y + (i )*cs_Y + (0 )*rs_Y;
scomplex* Y02 = buff_Y + (i+1)*cs_Y + (0 )*rs_Y;
scomplex* y12t = buff_Y + (i+1)*cs_Y + (i )*rs_Y;
scomplex* b01 = buff_B + (i )*cs_B + (0 )*rs_B;
scomplex* beta11 = buff_B + (i )*cs_B + (i )*rs_B;
scomplex* B02 = buff_B + (i+1)*cs_B + (0 )*rs_B;
scomplex* b12t = buff_B + (i+1)*cs_B + (i )*rs_B;
int m_ahead = m_AB - i - 1;
int m_behind = i;
/*------------------------------------------------------------*/
// FLA_Axpy_external( FLA_MINUS_ONE_HALF, y01, a01 );
bl1_caxpyv( BLIS1_NO_CONJUGATE,
m_behind,
buff_m1h,
y01, rs_Y,
a01, rs_A );
// FLA_Dot2cs_external( FLA_CONJUGATE, FLA_MINUS_ONE, a01, b01, FLA_ONE, alpha11 );
bl1_cdot2s( BLIS1_CONJUGATE,
m_behind,
buff_m1,
a01, rs_A,
b01, rs_B,
buff_1,
alpha11 );
// FLA_Inv_scal_external( beta11, alpha11 );
// FLA_Inv_scal_external( beta11, alpha11 );
bl1_cinvscals( beta11, alpha11 );
bl1_cinvscals( beta11, alpha11 );
// FLA_Gemvc_external( FLA_TRANSPOSE, FLA_CONJUGATE,
// FLA_MINUS_ONE, A02, b01, FLA_ONE, a12t );
bl1_cgemv( BLIS1_TRANSPOSE,
BLIS1_CONJUGATE,
m_behind,
m_ahead,
buff_m1,
A02, rs_A, cs_A,
b01, rs_B,
buff_1,
a12t, cs_A );
// FLA_Inv_scal_external( beta11, a12t );
bl1_cinvscalv( BLIS1_NO_CONJUGATE,
m_ahead,
beta11,
a12t, cs_A );
// FLA_Axpy_external( FLA_MINUS_ONE_HALF, y01, a01 );
bl1_caxpyv( BLIS1_NO_CONJUGATE,
m_behind,
buff_m1h,
y01, rs_Y,
a01, rs_A );
// FLA_Inv_scal_external( beta11, a01 );
bl1_cinvscalv( BLIS1_NO_CONJUGATE,
m_behind,
beta11,
a01, rs_A );
// FLA_Ger_external( FLA_ONE, a01, b12t, Y02 );
bl1_cger( BLIS1_NO_CONJUGATE,
BLIS1_NO_CONJUGATE,
m_behind,
m_ahead,
buff_1,
a01, rs_A,
b12t, cs_B,
Y02, rs_Y, cs_Y );
// FLA_Copy_external( b12t, y12t );
// FLA_Scal_external( alpha11, y12t );
bl1_ccopyv( BLIS1_NO_CONJUGATE,
m_ahead,
b12t, cs_B,
y12t, cs_Y );
bl1_cscalv( BLIS1_NO_CONJUGATE,
m_ahead,
alpha11,
y12t, cs_Y );
// FLA_Gemvc_external( FLA_TRANSPOSE, FLA_CONJUGATE,
// FLA_ONE, B02, a01, FLA_ONE, y12t );
bl1_cgemv( BLIS1_TRANSPOSE,
BLIS1_CONJUGATE,
m_behind,
m_ahead,
buff_1,
B02, rs_B, cs_B,
a01, rs_A,
buff_1,
y12t, cs_Y );
/*------------------------------------------------------------*/
}
return FLA_SUCCESS;
}
| FLA_Error FLA_Eig_gest_iu_opc_var4 | ( | int | m_AB, |
| scomplex * | buff_A, | ||
| int | rs_A, | ||
| int | cs_A, | ||
| scomplex * | buff_y, | ||
| int | inc_y, | ||
| scomplex * | buff_B, | ||
| int | rs_B, | ||
| int | cs_B | ||
| ) |
References bl1_caxpyv(), bl1_cger(), bl1_cher2(), bl1_cinvscalv(), BLIS1_CONJUGATE, BLIS1_NO_CONJUGATE, BLIS1_UPPER_TRIANGULAR, FLA_MINUS_ONE, and FLA_MINUS_ONE_HALF.
Referenced by FLA_Eig_gest_iu_opt_var4().
{
scomplex* buff_m1 = FLA_COMPLEX_PTR( FLA_MINUS_ONE );
scomplex* buff_m1h = FLA_COMPLEX_PTR( FLA_MINUS_ONE_HALF );
int i;
for ( i = 0; i < m_AB; ++i )
{
scomplex* a01 = buff_A + (i )*cs_A + (0 )*rs_A;
scomplex* alpha11 = buff_A + (i )*cs_A + (i )*rs_A;
scomplex* A02 = buff_A + (i+1)*cs_A + (0 )*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* beta11 = buff_B + (i )*cs_B + (i )*rs_B;
scomplex* b12t = buff_B + (i+1)*cs_B + (i )*rs_B;
scomplex psi11;
int m_ahead = m_AB - i - 1;
int m_behind = i;
/*------------------------------------------------------------*/
// FLA_Inv_scal_external( beta11, a01 );
bl1_cinvscalv( BLIS1_NO_CONJUGATE,
m_behind,
beta11,
a01, rs_A );
// FLA_Ger_external( FLA_MINUS_ONE, a01, b12t, A02 );
bl1_cger( BLIS1_NO_CONJUGATE,
BLIS1_NO_CONJUGATE,
m_behind,
m_ahead,
buff_m1,
a01, rs_A,
b12t, cs_B,
A02, rs_A, cs_A );
// FLA_Inv_scal_external( beta11, alpha11 );
// FLA_Inv_scal_external( beta11, alpha11 );
bl1_cinvscals( beta11, alpha11 );
bl1_cinvscals( beta11, alpha11 );
// FLA_Copy_external( alpha11, psi11 );
// FLA_Scal_external( FLA_MINUS_ONE_HALF, psi11 );
bl1_cmult3( buff_m1h, alpha11, &psi11 );
// FLA_Inv_scal_external( beta11, a12t );
bl1_cinvscalv( BLIS1_NO_CONJUGATE,
m_ahead,
beta11,
a12t, cs_A );
// FLA_Axpy_external( psi11, b12t, a12t );
bl1_caxpyv( BLIS1_NO_CONJUGATE,
m_ahead,
&psi11,
b12t, cs_B,
a12t, cs_A );
// FLA_Her2c_external( FLA_UPPER_TRIANGULAR, FLA_CONJUGATE,
// FLA_MINUS_ONE, a12t, b12t, A22 );
bl1_cher2( BLIS1_UPPER_TRIANGULAR,
BLIS1_CONJUGATE,
m_ahead,
buff_m1,
a12t, cs_A,
b12t, cs_B,
A22, rs_A, cs_A );
// FLA_Axpy_external( psi11, b12t, a12t );
bl1_caxpyv( BLIS1_NO_CONJUGATE,
m_ahead,
&psi11,
b12t, cs_B,
a12t, cs_A );
/*------------------------------------------------------------*/
}
return FLA_SUCCESS;
}
| FLA_Error FLA_Eig_gest_iu_opc_var5 | ( | int | m_AB, |
| scomplex * | buff_A, | ||
| int | rs_A, | ||
| int | cs_A, | ||
| scomplex * | buff_y, | ||
| int | inc_y, | ||
| scomplex * | buff_B, | ||
| int | rs_B, | ||
| int | cs_B | ||
| ) |
References bl1_caxpyv(), bl1_cher2(), bl1_cinvscalv(), bl1_ctrsv(), BLIS1_CONJUGATE, BLIS1_NO_CONJUGATE, BLIS1_NONUNIT_DIAG, BLIS1_TRANSPOSE, BLIS1_UPPER_TRIANGULAR, FLA_MINUS_ONE, and FLA_MINUS_ONE_HALF.
Referenced by FLA_Eig_gest_iu_opt_var5().
{
scomplex* buff_m1 = FLA_COMPLEX_PTR( FLA_MINUS_ONE );
scomplex* buff_m1h = FLA_COMPLEX_PTR( FLA_MINUS_ONE_HALF );
int i;
for ( i = 0; i < m_AB; ++i )
{
scomplex* alpha11 = buff_A + (i )*cs_A + (i )*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* beta11 = buff_B + (i )*cs_B + (i )*rs_B;
scomplex* b12t = buff_B + (i+1)*cs_B + (i )*rs_B;
scomplex* B22 = buff_B + (i+1)*cs_B + (i+1)*rs_B;
scomplex psi11;
int m_ahead = m_AB - i - 1;
/*------------------------------------------------------------*/
// FLA_Inv_scal_external( beta11, alpha11 );
// FLA_Inv_scal_external( beta11, alpha11 );
bl1_cinvscals( beta11, alpha11 );
bl1_cinvscals( beta11, alpha11 );
// FLA_Copy_external( alpha11, psi11 );
// FLA_Scal_external( FLA_MINUS_ONE_HALF, psi11 );
bl1_cmult3( buff_m1h, alpha11, &psi11 );
// FLA_Inv_scal_external( beta11, a12t );
bl1_cinvscalv( BLIS1_NO_CONJUGATE,
m_ahead,
beta11,
a12t, cs_A );
// FLA_Axpy_external( psi11, b12t, a12t );
bl1_caxpyv( BLIS1_NO_CONJUGATE,
m_ahead,
&psi11,
b12t, cs_B,
a12t, cs_A );
// FLA_Her2c_external( FLA_UPPER_TRIANGULAR, FLA_CONJUGATE,
// FLA_MINUS_ONE, a12t, b12t, A22 );
bl1_cher2( BLIS1_UPPER_TRIANGULAR,
BLIS1_CONJUGATE,
m_ahead,
buff_m1,
a12t, cs_A,
b12t, cs_B,
A22, rs_A, cs_A );
// FLA_Axpy_external( psi11, b12t, a12t );
bl1_caxpyv( BLIS1_NO_CONJUGATE,
m_ahead,
&psi11,
b12t, cs_B,
a12t, cs_A );
// FLA_Trsv_external( FLA_UPPER_TRIANGULAR, FLA_TRANSPOSE, FLA_NONUNIT_DIAG,
// B22, a12t );
bl1_ctrsv( BLIS1_UPPER_TRIANGULAR,
BLIS1_TRANSPOSE,
BLIS1_NONUNIT_DIAG,
m_ahead,
B22, rs_B, cs_B,
a12t, cs_A );
/*------------------------------------------------------------*/
}
return FLA_SUCCESS;
}
| FLA_Error FLA_Eig_gest_iu_opd_var1 | ( | int | m_AB, |
| double * | buff_A, | ||
| int | rs_A, | ||
| int | cs_A, | ||
| double * | buff_y, | ||
| int | inc_y, | ||
| double * | buff_B, | ||
| int | rs_B, | ||
| int | cs_B | ||
| ) |
References bl1_daxpyv(), bl1_ddot2s(), bl1_dhemv(), bl1_dinvscalv(), bl1_dtrsv(), BLIS1_CONJ_TRANSPOSE, BLIS1_CONJUGATE, BLIS1_NO_CONJUGATE, BLIS1_NONUNIT_DIAG, BLIS1_UPPER_TRIANGULAR, FLA_MINUS_ONE, FLA_MINUS_ONE_HALF, FLA_ONE, and FLA_ZERO.
Referenced by FLA_Eig_gest_iu_opt_var1().
{
double* buff_1 = FLA_DOUBLE_PTR( FLA_ONE );
double* buff_0 = FLA_DOUBLE_PTR( FLA_ZERO );
double* buff_m1 = FLA_DOUBLE_PTR( FLA_MINUS_ONE );
double* buff_m1h = FLA_DOUBLE_PTR( FLA_MINUS_ONE_HALF );
int i;
for ( i = 0; i < m_AB; ++i )
{
double* A00 = buff_A + (0 )*cs_A + (0 )*rs_A;
double* a01 = buff_A + (i )*cs_A + (0 )*rs_A;
double* alpha11 = buff_A + (i )*cs_A + (i )*rs_A;
double* y01 = buff_y + (0 )*inc_y;
double* B00 = buff_B + (0 )*cs_B + (0 )*rs_B;
double* b01 = buff_B + (i )*cs_B + (0 )*rs_B;
double* beta11 = buff_B + (i )*cs_B + (i )*rs_B;
int m_behind = i;
/*------------------------------------------------------------*/
// FLA_Hemvc_external( FLA_UPPER_TRIANGULAR, FLA_NO_CONJUGATE,
// FLA_ONE, A00, b01, FLA_ZERO, y01_l );
bl1_dhemv( BLIS1_UPPER_TRIANGULAR,
BLIS1_NO_CONJUGATE,
m_behind,
buff_1,
A00, rs_A, cs_A,
b01, rs_B,
buff_0,
y01, inc_y );
// FLA_Trsv_external( FLA_UPPER_TRIANGULAR, FLA_CONJ_TRANSPOSE, FLA_NONUNIT_DIAG,
// B00, a01 );
bl1_dtrsv( BLIS1_UPPER_TRIANGULAR,
BLIS1_CONJ_TRANSPOSE,
BLIS1_NONUNIT_DIAG,
m_behind,
B00, rs_B, cs_B,
a01, rs_A );
// FLA_Axpy_external( FLA_MINUS_ONE_HALF, y01_l, a01 );
bl1_daxpyv( BLIS1_NO_CONJUGATE,
m_behind,
buff_m1h,
y01, inc_y,
a01, rs_A );
// FLA_Dot2cs_external( FLA_CONJUGATE, FLA_MINUS_ONE, a01, b01, FLA_ONE, alpha11 );
bl1_ddot2s( BLIS1_CONJUGATE,
m_behind,
buff_m1,
a01, rs_A,
b01, rs_B,
buff_1,
alpha11 );
// FLA_Inv_scal_external( beta11, alpha11 );
// FLA_Inv_scal_external( beta11, alpha11 );
bl1_dinvscals( beta11, alpha11 );
bl1_dinvscals( beta11, alpha11 );
// FLA_Axpy_external( FLA_MINUS_ONE_HALF, y01_l, a01 );
bl1_daxpyv( BLIS1_NO_CONJUGATE,
m_behind,
buff_m1h,
y01, inc_y,
a01, rs_A );
// FLA_Inv_scal_external( beta11, a01 );
bl1_dinvscalv( BLIS1_NO_CONJUGATE,
m_behind,
beta11,
a01, rs_A );
/*------------------------------------------------------------*/
}
return FLA_SUCCESS;
}
| FLA_Error FLA_Eig_gest_iu_opd_var2 | ( | int | m_AB, |
| double * | buff_A, | ||
| int | rs_A, | ||
| int | cs_A, | ||
| double * | buff_y, | ||
| int | inc_y, | ||
| double * | buff_B, | ||
| int | rs_B, | ||
| int | cs_B | ||
| ) |
References bl1_daxpyv(), bl1_ddot2s(), bl1_dgemv(), bl1_dhemv(), bl1_dinvscalv(), BLIS1_CONJUGATE, BLIS1_NO_CONJUGATE, BLIS1_TRANSPOSE, BLIS1_UPPER_TRIANGULAR, FLA_MINUS_ONE, FLA_ONE, FLA_ONE_HALF, and FLA_ZERO.
Referenced by FLA_Eig_gest_iu_opt_var2().
{
double* buff_1 = FLA_DOUBLE_PTR( FLA_ONE );
double* buff_1h = FLA_DOUBLE_PTR( FLA_ONE_HALF );
double* buff_0 = FLA_DOUBLE_PTR( FLA_ZERO );
double* buff_m1 = FLA_DOUBLE_PTR( FLA_MINUS_ONE );
int i;
for ( i = 0; i < m_AB; ++i )
{
double* A00 = buff_A + (0 )*cs_A + (0 )*rs_A;
double* a01 = buff_A + (i )*cs_A + (0 )*rs_A;
double* alpha11 = buff_A + (i )*cs_A + (i )*rs_A;
double* A02 = buff_A + (i+1)*cs_A + (0 )*rs_A;
double* a12t = buff_A + (i+1)*cs_A + (i )*rs_A;
double* y01 = buff_y + (0 )*inc_y;
double* b01 = buff_B + (i )*cs_B + (0 )*rs_B;
double* beta11 = buff_B + (i )*cs_B + (i )*rs_B;
int m_ahead = m_AB - i - 1;
int m_behind = i;
/*------------------------------------------------------------*/
// FLA_Hemvc_external( FLA_UPPER_TRIANGULAR, FLA_NO_CONJUGATE,
// FLA_ONE_HALF, A00, b01, FLA_ZERO, y01_l );
bl1_dhemv( BLIS1_UPPER_TRIANGULAR,
BLIS1_NO_CONJUGATE,
m_behind,
buff_1h,
A00, rs_A, cs_A,
b01, rs_B,
buff_0,
y01, inc_y );
// FLA_Axpy_external( FLA_MINUS_ONE, y01_l, a01 );
bl1_daxpyv( BLIS1_NO_CONJUGATE,
m_behind,
buff_m1,
y01, inc_y,
a01, rs_A );
// FLA_Dot2cs_external( FLA_CONJUGATE, FLA_MINUS_ONE, a01, b01, FLA_ONE, alpha11 );
bl1_ddot2s( BLIS1_CONJUGATE,
m_behind,
buff_m1,
a01, rs_A,
b01, rs_B,
buff_1,
alpha11 );
// FLA_Inv_scal_external( beta11, alpha11 );
// FLA_Inv_scal_external( beta11, alpha11 );
bl1_dinvscals( beta11, alpha11 );
bl1_dinvscals( beta11, alpha11 );
// FLA_Gemvc_external( FLA_TRANSPOSE, FLA_CONJUGATE,
// FLA_MINUS_ONE, A02, b01, FLA_ONE, a12t );
bl1_dgemv( BLIS1_TRANSPOSE,
BLIS1_CONJUGATE,
m_behind,
m_ahead,
buff_m1,
A02, rs_A, cs_A,
b01, rs_B,
buff_1,
a12t, cs_A );
// FLA_Inv_scal_external( beta11, a12t );
bl1_dinvscalv( BLIS1_NO_CONJUGATE,
m_ahead,
beta11,
a12t, cs_A );
// FLA_Axpy_external( FLA_MINUS_ONE, y01_l, a01 );
bl1_daxpyv( BLIS1_NO_CONJUGATE,
m_behind,
buff_m1,
y01, inc_y,
a01, rs_A );
// FLA_Inv_scal_external( beta11, a01 );
bl1_dinvscalv( BLIS1_NO_CONJUGATE,
m_behind,
beta11,
a01, rs_A );
/*------------------------------------------------------------*/
}
return FLA_SUCCESS;
}
| FLA_Error FLA_Eig_gest_iu_opd_var3 | ( | int | m_AB, |
| double * | buff_A, | ||
| int | rs_A, | ||
| int | cs_A, | ||
| double * | buff_Y, | ||
| int | rs_Y, | ||
| int | cs_Y, | ||
| double * | buff_B, | ||
| int | rs_B, | ||
| int | cs_B | ||
| ) |
References bl1_daxpyv(), bl1_dcopyv(), bl1_ddot2s(), bl1_dgemv(), bl1_dger(), bl1_dinvscalv(), bl1_dscalv(), BLIS1_CONJUGATE, BLIS1_NO_CONJUGATE, BLIS1_TRANSPOSE, FLA_MINUS_ONE, FLA_MINUS_ONE_HALF, and FLA_ONE.
Referenced by FLA_Eig_gest_iu_opt_var3().
{
double* buff_1 = FLA_DOUBLE_PTR( FLA_ONE );
double* buff_m1 = FLA_DOUBLE_PTR( FLA_MINUS_ONE );
double* buff_m1h = FLA_DOUBLE_PTR( FLA_MINUS_ONE_HALF );
int i;
for ( i = 0; i < m_AB; ++i )
{
double* a01 = buff_A + (i )*cs_A + (0 )*rs_A;
double* alpha11 = buff_A + (i )*cs_A + (i )*rs_A;
double* A02 = buff_A + (i+1)*cs_A + (0 )*rs_A;
double* a12t = buff_A + (i+1)*cs_A + (i )*rs_A;
double* y01 = buff_Y + (i )*cs_Y + (0 )*rs_Y;
double* Y02 = buff_Y + (i+1)*cs_Y + (0 )*rs_Y;
double* y12t = buff_Y + (i+1)*cs_Y + (i )*rs_Y;
double* b01 = buff_B + (i )*cs_B + (0 )*rs_B;
double* beta11 = buff_B + (i )*cs_B + (i )*rs_B;
double* B02 = buff_B + (i+1)*cs_B + (0 )*rs_B;
double* b12t = buff_B + (i+1)*cs_B + (i )*rs_B;
int m_ahead = m_AB - i - 1;
int m_behind = i;
/*------------------------------------------------------------*/
// FLA_Axpy_external( FLA_MINUS_ONE_HALF, y01, a01 );
bl1_daxpyv( BLIS1_NO_CONJUGATE,
m_behind,
buff_m1h,
y01, rs_Y,
a01, rs_A );
// FLA_Dot2cs_external( FLA_CONJUGATE, FLA_MINUS_ONE, a01, b01, FLA_ONE, alpha11 );
bl1_ddot2s( BLIS1_CONJUGATE,
m_behind,
buff_m1,
a01, rs_A,
b01, rs_B,
buff_1,
alpha11 );
// FLA_Inv_scal_external( beta11, alpha11 );
// FLA_Inv_scal_external( beta11, alpha11 );
bl1_dinvscals( beta11, alpha11 );
bl1_dinvscals( beta11, alpha11 );
// FLA_Gemvc_external( FLA_TRANSPOSE, FLA_CONJUGATE,
// FLA_MINUS_ONE, A02, b01, FLA_ONE, a12t );
bl1_dgemv( BLIS1_TRANSPOSE,
BLIS1_CONJUGATE,
m_behind,
m_ahead,
buff_m1,
A02, rs_A, cs_A,
b01, rs_B,
buff_1,
a12t, cs_A );
// FLA_Inv_scal_external( beta11, a12t );
bl1_dinvscalv( BLIS1_NO_CONJUGATE,
m_ahead,
beta11,
a12t, cs_A );
// FLA_Axpy_external( FLA_MINUS_ONE_HALF, y01, a01 );
bl1_daxpyv( BLIS1_NO_CONJUGATE,
m_behind,
buff_m1h,
y01, rs_Y,
a01, rs_A );
// FLA_Inv_scal_external( beta11, a01 );
bl1_dinvscalv( BLIS1_NO_CONJUGATE,
m_behind,
beta11,
a01, rs_A );
// FLA_Ger_external( FLA_ONE, a01, b12t, Y02 );
bl1_dger( BLIS1_NO_CONJUGATE,
BLIS1_NO_CONJUGATE,
m_behind,
m_ahead,
buff_1,
a01, rs_A,
b12t, cs_B,
Y02, rs_Y, cs_Y );
// FLA_Copy_external( b12t, y12t );
// FLA_Scal_external( alpha11, y12t );
bl1_dcopyv( BLIS1_NO_CONJUGATE,
m_ahead,
b12t, cs_B,
y12t, cs_Y );
bl1_dscalv( BLIS1_NO_CONJUGATE,
m_ahead,
alpha11,
y12t, cs_Y );
// FLA_Gemvc_external( FLA_TRANSPOSE, FLA_CONJUGATE,
// FLA_ONE, B02, a01, FLA_ONE, y12t );
bl1_dgemv( BLIS1_TRANSPOSE,
BLIS1_CONJUGATE,
m_behind,
m_ahead,
buff_1,
B02, rs_B, cs_B,
a01, rs_A,
buff_1,
y12t, cs_Y );
/*------------------------------------------------------------*/
}
return FLA_SUCCESS;
}
| FLA_Error FLA_Eig_gest_iu_opd_var4 | ( | int | m_AB, |
| double * | buff_A, | ||
| int | rs_A, | ||
| int | cs_A, | ||
| double * | buff_y, | ||
| int | inc_y, | ||
| double * | buff_B, | ||
| int | rs_B, | ||
| int | cs_B | ||
| ) |
References bl1_daxpyv(), bl1_dger(), bl1_dher2(), bl1_dinvscalv(), BLIS1_CONJUGATE, BLIS1_NO_CONJUGATE, BLIS1_UPPER_TRIANGULAR, FLA_MINUS_ONE, and FLA_MINUS_ONE_HALF.
Referenced by FLA_Eig_gest_iu_opt_var4().
{
double* buff_m1 = FLA_DOUBLE_PTR( FLA_MINUS_ONE );
double* buff_m1h = FLA_DOUBLE_PTR( FLA_MINUS_ONE_HALF );
int i;
for ( i = 0; i < m_AB; ++i )
{
double* a01 = buff_A + (i )*cs_A + (0 )*rs_A;
double* alpha11 = buff_A + (i )*cs_A + (i )*rs_A;
double* A02 = buff_A + (i+1)*cs_A + (0 )*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* beta11 = buff_B + (i )*cs_B + (i )*rs_B;
double* b12t = buff_B + (i+1)*cs_B + (i )*rs_B;
double psi11;
int m_ahead = m_AB - i - 1;
int m_behind = i;
/*------------------------------------------------------------*/
// FLA_Inv_scal_external( beta11, a01 );
bl1_dinvscalv( BLIS1_NO_CONJUGATE,
m_behind,
beta11,
a01, rs_A );
// FLA_Ger_external( FLA_MINUS_ONE, a01, b12t, A02 );
bl1_dger( BLIS1_NO_CONJUGATE,
BLIS1_NO_CONJUGATE,
m_behind,
m_ahead,
buff_m1,
a01, rs_A,
b12t, cs_B,
A02, rs_A, cs_A );
// FLA_Inv_scal_external( beta11, alpha11 );
// FLA_Inv_scal_external( beta11, alpha11 );
bl1_dinvscals( beta11, alpha11 );
bl1_dinvscals( beta11, alpha11 );
// FLA_Copy_external( alpha11, psi11 );
// FLA_Scal_external( FLA_MINUS_ONE_HALF, psi11 );
bl1_dmult3( buff_m1h, alpha11, &psi11 );
// FLA_Inv_scal_external( beta11, a12t );
bl1_dinvscalv( BLIS1_NO_CONJUGATE,
m_ahead,
beta11,
a12t, cs_A );
// FLA_Axpy_external( psi11, b12t, a12t );
bl1_daxpyv( BLIS1_NO_CONJUGATE,
m_ahead,
&psi11,
b12t, cs_B,
a12t, cs_A );
// FLA_Her2c_external( FLA_UPPER_TRIANGULAR, FLA_CONJUGATE,
// FLA_MINUS_ONE, a12t, b12t, A22 );
bl1_dher2( BLIS1_UPPER_TRIANGULAR,
BLIS1_CONJUGATE,
m_ahead,
buff_m1,
a12t, cs_A,
b12t, cs_B,
A22, rs_A, cs_A );
// FLA_Axpy_external( psi11, b12t, a12t );
bl1_daxpyv( BLIS1_NO_CONJUGATE,
m_ahead,
&psi11,
b12t, cs_B,
a12t, cs_A );
/*------------------------------------------------------------*/
}
return FLA_SUCCESS;
}
| FLA_Error FLA_Eig_gest_iu_opd_var5 | ( | int | m_AB, |
| double * | buff_A, | ||
| int | rs_A, | ||
| int | cs_A, | ||
| double * | buff_y, | ||
| int | inc_y, | ||
| double * | buff_B, | ||
| int | rs_B, | ||
| int | cs_B | ||
| ) |
References bl1_daxpyv(), bl1_dher2(), bl1_dinvscalv(), bl1_dtrsv(), BLIS1_CONJUGATE, BLIS1_NO_CONJUGATE, BLIS1_NONUNIT_DIAG, BLIS1_TRANSPOSE, BLIS1_UPPER_TRIANGULAR, FLA_MINUS_ONE, and FLA_MINUS_ONE_HALF.
Referenced by FLA_Eig_gest_iu_opt_var5().
{
double* buff_m1 = FLA_DOUBLE_PTR( FLA_MINUS_ONE );
double* buff_m1h = FLA_DOUBLE_PTR( FLA_MINUS_ONE_HALF );
int i;
for ( i = 0; i < m_AB; ++i )
{
double* alpha11 = buff_A + (i )*cs_A + (i )*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* beta11 = buff_B + (i )*cs_B + (i )*rs_B;
double* b12t = buff_B + (i+1)*cs_B + (i )*rs_B;
double* B22 = buff_B + (i+1)*cs_B + (i+1)*rs_B;
double psi11;
int m_ahead = m_AB - i - 1;
/*------------------------------------------------------------*/
// FLA_Inv_scal_external( beta11, alpha11 );
// FLA_Inv_scal_external( beta11, alpha11 );
bl1_dinvscals( beta11, alpha11 );
bl1_dinvscals( beta11, alpha11 );
// FLA_Copy_external( alpha11, psi11 );
// FLA_Scal_external( FLA_MINUS_ONE_HALF, psi11 );
bl1_dmult3( buff_m1h, alpha11, &psi11 );
// FLA_Inv_scal_external( beta11, a12t );
bl1_dinvscalv( BLIS1_NO_CONJUGATE,
m_ahead,
beta11,
a12t, cs_A );
// FLA_Axpy_external( psi11, b12t, a12t );
bl1_daxpyv( BLIS1_NO_CONJUGATE,
m_ahead,
&psi11,
b12t, cs_B,
a12t, cs_A );
// FLA_Her2c_external( FLA_UPPER_TRIANGULAR, FLA_CONJUGATE,
// FLA_MINUS_ONE, a12t, b12t, A22 );
bl1_dher2( BLIS1_UPPER_TRIANGULAR,
BLIS1_CONJUGATE,
m_ahead,
buff_m1,
a12t, cs_A,
b12t, cs_B,
A22, rs_A, cs_A );
// FLA_Axpy_external( psi11, b12t, a12t );
bl1_daxpyv( BLIS1_NO_CONJUGATE,
m_ahead,
&psi11,
b12t, cs_B,
a12t, cs_A );
// FLA_Trsv_external( FLA_UPPER_TRIANGULAR, FLA_TRANSPOSE, FLA_NONUNIT_DIAG,
// B22, a12t );
bl1_dtrsv( BLIS1_UPPER_TRIANGULAR,
BLIS1_TRANSPOSE,
BLIS1_NONUNIT_DIAG,
m_ahead,
B22, rs_B, cs_B,
a12t, cs_A );
/*------------------------------------------------------------*/
}
return FLA_SUCCESS;
}
| FLA_Error FLA_Eig_gest_iu_ops_var1 | ( | int | m_AB, |
| float * | buff_A, | ||
| int | rs_A, | ||
| int | cs_A, | ||
| float * | buff_y, | ||
| int | inc_y, | ||
| float * | buff_B, | ||
| int | rs_B, | ||
| int | cs_B | ||
| ) |
References bl1_saxpyv(), bl1_sdot2s(), bl1_shemv(), bl1_sinvscalv(), bl1_strsv(), BLIS1_CONJ_TRANSPOSE, BLIS1_CONJUGATE, BLIS1_NO_CONJUGATE, BLIS1_NONUNIT_DIAG, BLIS1_UPPER_TRIANGULAR, FLA_MINUS_ONE, FLA_MINUS_ONE_HALF, FLA_ONE, and FLA_ZERO.
Referenced by FLA_Eig_gest_iu_opt_var1().
{
float* buff_1 = FLA_FLOAT_PTR( FLA_ONE );
float* buff_0 = FLA_FLOAT_PTR( FLA_ZERO );
float* buff_m1 = FLA_FLOAT_PTR( FLA_MINUS_ONE );
float* buff_m1h = FLA_FLOAT_PTR( FLA_MINUS_ONE_HALF );
int i;
for ( i = 0; i < m_AB; ++i )
{
float* A00 = buff_A + (0 )*cs_A + (0 )*rs_A;
float* a01 = buff_A + (i )*cs_A + (0 )*rs_A;
float* alpha11 = buff_A + (i )*cs_A + (i )*rs_A;
float* y01 = buff_y + (0 )*inc_y;
float* B00 = buff_B + (0 )*cs_B + (0 )*rs_B;
float* b01 = buff_B + (i )*cs_B + (0 )*rs_B;
float* beta11 = buff_B + (i )*cs_B + (i )*rs_B;
int m_behind = i;
/*------------------------------------------------------------*/
// FLA_Hemvc_external( FLA_UPPER_TRIANGULAR, FLA_NO_CONJUGATE,
// FLA_ONE, A00, b01, FLA_ZERO, y01_l );
bl1_shemv( BLIS1_UPPER_TRIANGULAR,
BLIS1_NO_CONJUGATE,
m_behind,
buff_1,
A00, rs_A, cs_A,
b01, rs_B,
buff_0,
y01, inc_y );
// FLA_Trsv_external( FLA_UPPER_TRIANGULAR, FLA_CONJ_TRANSPOSE, FLA_NONUNIT_DIAG,
// B00, a01 );
bl1_strsv( BLIS1_UPPER_TRIANGULAR,
BLIS1_CONJ_TRANSPOSE,
BLIS1_NONUNIT_DIAG,
m_behind,
B00, rs_B, cs_B,
a01, rs_A );
// FLA_Axpy_external( FLA_MINUS_ONE_HALF, y01_l, a01 );
bl1_saxpyv( BLIS1_NO_CONJUGATE,
m_behind,
buff_m1h,
y01, inc_y,
a01, rs_A );
// FLA_Dot2cs_external( FLA_CONJUGATE, FLA_MINUS_ONE, a01, b01, FLA_ONE, alpha11 );
bl1_sdot2s( BLIS1_CONJUGATE,
m_behind,
buff_m1,
a01, rs_A,
b01, rs_B,
buff_1,
alpha11 );
// FLA_Inv_scal_external( beta11, alpha11 );
// FLA_Inv_scal_external( beta11, alpha11 );
bl1_sinvscals( beta11, alpha11 );
bl1_sinvscals( beta11, alpha11 );
// FLA_Axpy_external( FLA_MINUS_ONE_HALF, y01_l, a01 );
bl1_saxpyv( BLIS1_NO_CONJUGATE,
m_behind,
buff_m1h,
y01, inc_y,
a01, rs_A );
// FLA_Inv_scal_external( beta11, a01 );
bl1_sinvscalv( BLIS1_NO_CONJUGATE,
m_behind,
beta11,
a01, rs_A );
/*------------------------------------------------------------*/
}
return FLA_SUCCESS;
}
| FLA_Error FLA_Eig_gest_iu_ops_var2 | ( | int | m_AB, |
| float * | buff_A, | ||
| int | rs_A, | ||
| int | cs_A, | ||
| float * | buff_y, | ||
| int | inc_y, | ||
| float * | buff_B, | ||
| int | rs_B, | ||
| int | cs_B | ||
| ) |
References bl1_saxpyv(), bl1_sdot2s(), bl1_sgemv(), bl1_shemv(), bl1_sinvscalv(), BLIS1_CONJUGATE, BLIS1_NO_CONJUGATE, BLIS1_TRANSPOSE, BLIS1_UPPER_TRIANGULAR, FLA_MINUS_ONE, FLA_ONE, FLA_ONE_HALF, and FLA_ZERO.
Referenced by FLA_Eig_gest_iu_opt_var2().
{
float* buff_1 = FLA_FLOAT_PTR( FLA_ONE );
float* buff_1h = FLA_FLOAT_PTR( FLA_ONE_HALF );
float* buff_0 = FLA_FLOAT_PTR( FLA_ZERO );
float* buff_m1 = FLA_FLOAT_PTR( FLA_MINUS_ONE );
int i;
for ( i = 0; i < m_AB; ++i )
{
float* A00 = buff_A + (0 )*cs_A + (0 )*rs_A;
float* a01 = buff_A + (i )*cs_A + (0 )*rs_A;
float* alpha11 = buff_A + (i )*cs_A + (i )*rs_A;
float* A02 = buff_A + (i+1)*cs_A + (0 )*rs_A;
float* a12t = buff_A + (i+1)*cs_A + (i )*rs_A;
float* y01 = buff_y + (0 )*inc_y;
float* b01 = buff_B + (i )*cs_B + (0 )*rs_B;
float* beta11 = buff_B + (i )*cs_B + (i )*rs_B;
int m_ahead = m_AB - i - 1;
int m_behind = i;
/*------------------------------------------------------------*/
// FLA_Hemvc_external( FLA_UPPER_TRIANGULAR, FLA_NO_CONJUGATE,
// FLA_ONE_HALF, A00, b01, FLA_ZERO, y01_l );
bl1_shemv( BLIS1_UPPER_TRIANGULAR,
BLIS1_NO_CONJUGATE,
m_behind,
buff_1h,
A00, rs_A, cs_A,
b01, rs_B,
buff_0,
y01, inc_y );
// FLA_Axpy_external( FLA_MINUS_ONE, y01_l, a01 );
bl1_saxpyv( BLIS1_NO_CONJUGATE,
m_behind,
buff_m1,
y01, inc_y,
a01, rs_A );
// FLA_Dot2cs_external( FLA_CONJUGATE, FLA_MINUS_ONE, a01, b01, FLA_ONE, alpha11 );
bl1_sdot2s( BLIS1_CONJUGATE,
m_behind,
buff_m1,
a01, rs_A,
b01, rs_B,
buff_1,
alpha11 );
// FLA_Inv_scal_external( beta11, alpha11 );
// FLA_Inv_scal_external( beta11, alpha11 );
bl1_sinvscals( beta11, alpha11 );
bl1_sinvscals( beta11, alpha11 );
// FLA_Gemvc_external( FLA_TRANSPOSE, FLA_CONJUGATE,
// FLA_MINUS_ONE, A02, b01, FLA_ONE, a12t );
bl1_sgemv( BLIS1_TRANSPOSE,
BLIS1_CONJUGATE,
m_behind,
m_ahead,
buff_m1,
A02, rs_A, cs_A,
b01, rs_B,
buff_1,
a12t, cs_A );
// FLA_Inv_scal_external( beta11, a12t );
bl1_sinvscalv( BLIS1_NO_CONJUGATE,
m_ahead,
beta11,
a12t, cs_A );
// FLA_Axpy_external( FLA_MINUS_ONE, y01_l, a01 );
bl1_saxpyv( BLIS1_NO_CONJUGATE,
m_behind,
buff_m1,
y01, inc_y,
a01, rs_A );
// FLA_Inv_scal_external( beta11, a01 );
bl1_sinvscalv( BLIS1_NO_CONJUGATE,
m_behind,
beta11,
a01, rs_A );
/*------------------------------------------------------------*/
}
return FLA_SUCCESS;
}
| FLA_Error FLA_Eig_gest_iu_ops_var3 | ( | int | m_AB, |
| float * | buff_A, | ||
| int | rs_A, | ||
| int | cs_A, | ||
| float * | buff_Y, | ||
| int | rs_Y, | ||
| int | cs_Y, | ||
| float * | buff_B, | ||
| int | rs_B, | ||
| int | cs_B | ||
| ) |
References bl1_saxpyv(), bl1_scopyv(), bl1_sdot2s(), bl1_sgemv(), bl1_sger(), bl1_sinvscalv(), bl1_sscalv(), BLIS1_CONJUGATE, BLIS1_NO_CONJUGATE, BLIS1_TRANSPOSE, FLA_MINUS_ONE, FLA_MINUS_ONE_HALF, and FLA_ONE.
Referenced by FLA_Eig_gest_iu_opt_var3().
{
float* buff_1 = FLA_FLOAT_PTR( FLA_ONE );
float* buff_m1 = FLA_FLOAT_PTR( FLA_MINUS_ONE );
float* buff_m1h = FLA_FLOAT_PTR( FLA_MINUS_ONE_HALF );
int i;
for ( i = 0; i < m_AB; ++i )
{
float* a01 = buff_A + (i )*cs_A + (0 )*rs_A;
float* alpha11 = buff_A + (i )*cs_A + (i )*rs_A;
float* A02 = buff_A + (i+1)*cs_A + (0 )*rs_A;
float* a12t = buff_A + (i+1)*cs_A + (i )*rs_A;
float* y01 = buff_Y + (i )*cs_Y + (0 )*rs_Y;
float* Y02 = buff_Y + (i+1)*cs_Y + (0 )*rs_Y;
float* y12t = buff_Y + (i+1)*cs_Y + (i )*rs_Y;
float* b01 = buff_B + (i )*cs_B + (0 )*rs_B;
float* beta11 = buff_B + (i )*cs_B + (i )*rs_B;
float* B02 = buff_B + (i+1)*cs_B + (0 )*rs_B;
float* b12t = buff_B + (i+1)*cs_B + (i )*rs_B;
int m_ahead = m_AB - i - 1;
int m_behind = i;
/*------------------------------------------------------------*/
// FLA_Axpy_external( FLA_MINUS_ONE_HALF, y01, a01 );
bl1_saxpyv( BLIS1_NO_CONJUGATE,
m_behind,
buff_m1h,
y01, rs_Y,
a01, rs_A );
// FLA_Dot2cs_external( FLA_CONJUGATE, FLA_MINUS_ONE, a01, b01, FLA_ONE, alpha11 );
bl1_sdot2s( BLIS1_CONJUGATE,
m_behind,
buff_m1,
a01, rs_A,
b01, rs_B,
buff_1,
alpha11 );
// FLA_Inv_scal_external( beta11, alpha11 );
// FLA_Inv_scal_external( beta11, alpha11 );
bl1_sinvscals( beta11, alpha11 );
bl1_sinvscals( beta11, alpha11 );
// FLA_Gemvc_external( FLA_TRANSPOSE, FLA_CONJUGATE,
// FLA_MINUS_ONE, A02, b01, FLA_ONE, a12t );
bl1_sgemv( BLIS1_TRANSPOSE,
BLIS1_CONJUGATE,
m_behind,
m_ahead,
buff_m1,
A02, rs_A, cs_A,
b01, rs_B,
buff_1,
a12t, cs_A );
// FLA_Inv_scal_external( beta11, a12t );
bl1_sinvscalv( BLIS1_NO_CONJUGATE,
m_ahead,
beta11,
a12t, cs_A );
// FLA_Axpy_external( FLA_MINUS_ONE_HALF, y01, a01 );
bl1_saxpyv( BLIS1_NO_CONJUGATE,
m_behind,
buff_m1h,
y01, rs_Y,
a01, rs_A );
// FLA_Inv_scal_external( beta11, a01 );
bl1_sinvscalv( BLIS1_NO_CONJUGATE,
m_behind,
beta11,
a01, rs_A );
// FLA_Ger_external( FLA_ONE, a01, b12t, Y02 );
bl1_sger( BLIS1_NO_CONJUGATE,
BLIS1_NO_CONJUGATE,
m_behind,
m_ahead,
buff_1,
a01, rs_A,
b12t, cs_B,
Y02, rs_Y, cs_Y );
// FLA_Copy_external( b12t, y12t );
// FLA_Scal_external( alpha11, y12t );
bl1_scopyv( BLIS1_NO_CONJUGATE,
m_ahead,
b12t, cs_B,
y12t, cs_Y );
bl1_sscalv( BLIS1_NO_CONJUGATE,
m_ahead,
alpha11,
y12t, cs_Y );
// FLA_Gemvc_external( FLA_TRANSPOSE, FLA_CONJUGATE,
// FLA_ONE, B02, a01, FLA_ONE, y12t );
bl1_sgemv( BLIS1_TRANSPOSE,
BLIS1_CONJUGATE,
m_behind,
m_ahead,
buff_1,
B02, rs_B, cs_B,
a01, rs_A,
buff_1,
y12t, cs_Y );
/*------------------------------------------------------------*/
}
return FLA_SUCCESS;
}
| FLA_Error FLA_Eig_gest_iu_ops_var4 | ( | int | m_AB, |
| float * | buff_A, | ||
| int | rs_A, | ||
| int | cs_A, | ||
| float * | buff_y, | ||
| int | inc_y, | ||
| float * | buff_B, | ||
| int | rs_B, | ||
| int | cs_B | ||
| ) |
References bl1_saxpyv(), bl1_sger(), bl1_sher2(), bl1_sinvscalv(), BLIS1_CONJUGATE, BLIS1_NO_CONJUGATE, BLIS1_UPPER_TRIANGULAR, FLA_MINUS_ONE, and FLA_MINUS_ONE_HALF.
Referenced by FLA_Eig_gest_iu_opt_var4().
{
float* buff_m1 = FLA_FLOAT_PTR( FLA_MINUS_ONE );
float* buff_m1h = FLA_FLOAT_PTR( FLA_MINUS_ONE_HALF );
int i;
for ( i = 0; i < m_AB; ++i )
{
float* a01 = buff_A + (i )*cs_A + (0 )*rs_A;
float* alpha11 = buff_A + (i )*cs_A + (i )*rs_A;
float* A02 = buff_A + (i+1)*cs_A + (0 )*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* beta11 = buff_B + (i )*cs_B + (i )*rs_B;
float* b12t = buff_B + (i+1)*cs_B + (i )*rs_B;
float psi11;
int m_ahead = m_AB - i - 1;
int m_behind = i;
/*------------------------------------------------------------*/
// FLA_Inv_scal_external( beta11, a01 );
bl1_sinvscalv( BLIS1_NO_CONJUGATE,
m_behind,
beta11,
a01, rs_A );
// FLA_Ger_external( FLA_MINUS_ONE, a01, b12t, A02 );
bl1_sger( BLIS1_NO_CONJUGATE,
BLIS1_NO_CONJUGATE,
m_behind,
m_ahead,
buff_m1,
a01, rs_A,
b12t, cs_B,
A02, rs_A, cs_A );
// FLA_Inv_scal_external( beta11, alpha11 );
// FLA_Inv_scal_external( beta11, alpha11 );
bl1_sinvscals( beta11, alpha11 );
bl1_sinvscals( beta11, alpha11 );
// FLA_Copy_external( alpha11, psi11 );
// FLA_Scal_external( FLA_MINUS_ONE_HALF, psi11 );
bl1_smult3( buff_m1h, alpha11, &psi11 );
// FLA_Inv_scal_external( beta11, a12t );
bl1_sinvscalv( BLIS1_NO_CONJUGATE,
m_ahead,
beta11,
a12t, cs_A );
// FLA_Axpy_external( psi11, b12t, a12t );
bl1_saxpyv( BLIS1_NO_CONJUGATE,
m_ahead,
&psi11,
b12t, cs_B,
a12t, cs_A );
// FLA_Her2c_external( FLA_UPPER_TRIANGULAR, FLA_CONJUGATE,
// FLA_MINUS_ONE, a12t, b12t, A22 );
bl1_sher2( BLIS1_UPPER_TRIANGULAR,
BLIS1_CONJUGATE,
m_ahead,
buff_m1,
a12t, cs_A,
b12t, cs_B,
A22, rs_A, cs_A );
// FLA_Axpy_external( psi11, b12t, a12t );
bl1_saxpyv( BLIS1_NO_CONJUGATE,
m_ahead,
&psi11,
b12t, cs_B,
a12t, cs_A );
/*------------------------------------------------------------*/
}
return FLA_SUCCESS;
}
| FLA_Error FLA_Eig_gest_iu_ops_var5 | ( | int | m_AB, |
| float * | buff_A, | ||
| int | rs_A, | ||
| int | cs_A, | ||
| float * | buff_y, | ||
| int | inc_y, | ||
| float * | buff_B, | ||
| int | rs_B, | ||
| int | cs_B | ||
| ) |
References bl1_saxpyv(), bl1_sher2(), bl1_sinvscalv(), bl1_strsv(), BLIS1_CONJUGATE, BLIS1_NO_CONJUGATE, BLIS1_NONUNIT_DIAG, BLIS1_TRANSPOSE, BLIS1_UPPER_TRIANGULAR, FLA_MINUS_ONE, and FLA_MINUS_ONE_HALF.
Referenced by FLA_Eig_gest_iu_opt_var5().
{
float* buff_m1 = FLA_FLOAT_PTR( FLA_MINUS_ONE );
float* buff_m1h = FLA_FLOAT_PTR( FLA_MINUS_ONE_HALF );
int i;
for ( i = 0; i < m_AB; ++i )
{
float* alpha11 = buff_A + (i )*cs_A + (i )*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* beta11 = buff_B + (i )*cs_B + (i )*rs_B;
float* b12t = buff_B + (i+1)*cs_B + (i )*rs_B;
float* B22 = buff_B + (i+1)*cs_B + (i+1)*rs_B;
float psi11;
int m_ahead = m_AB - i - 1;
/*------------------------------------------------------------*/
// FLA_Inv_scal_external( beta11, alpha11 );
// FLA_Inv_scal_external( beta11, alpha11 );
bl1_sinvscals( beta11, alpha11 );
bl1_sinvscals( beta11, alpha11 );
// FLA_Copy_external( alpha11, psi11 );
// FLA_Scal_external( FLA_MINUS_ONE_HALF, psi11 );
bl1_smult3( buff_m1h, alpha11, &psi11 );
// FLA_Inv_scal_external( beta11, a12t );
bl1_sinvscalv( BLIS1_NO_CONJUGATE,
m_ahead,
beta11,
a12t, cs_A );
// FLA_Axpy_external( psi11, b12t, a12t );
bl1_saxpyv( BLIS1_NO_CONJUGATE,
m_ahead,
&psi11,
b12t, cs_B,
a12t, cs_A );
// FLA_Her2c_external( FLA_UPPER_TRIANGULAR, FLA_CONJUGATE,
// FLA_MINUS_ONE, a12t, b12t, A22 );
bl1_sher2( BLIS1_UPPER_TRIANGULAR,
BLIS1_CONJUGATE,
m_ahead,
buff_m1,
a12t, cs_A,
b12t, cs_B,
A22, rs_A, cs_A );
// FLA_Axpy_external( psi11, b12t, a12t );
bl1_saxpyv( BLIS1_NO_CONJUGATE,
m_ahead,
&psi11,
b12t, cs_B,
a12t, cs_A );
// FLA_Trsv_external( FLA_UPPER_TRIANGULAR, FLA_TRANSPOSE, FLA_NONUNIT_DIAG,
// B22, a12t );
bl1_strsv( BLIS1_UPPER_TRIANGULAR,
BLIS1_TRANSPOSE,
BLIS1_NONUNIT_DIAG,
m_ahead,
B22, rs_B, cs_B,
a12t, cs_A );
/*------------------------------------------------------------*/
}
return FLA_SUCCESS;
}
| FLA_Error FLA_Eig_gest_iu_opt_var1 | ( | FLA_Obj | A, |
| FLA_Obj | Y, | ||
| FLA_Obj | B | ||
| ) |
References FLA_Eig_gest_iu_opc_var1(), FLA_Eig_gest_iu_opd_var1(), FLA_Eig_gest_iu_ops_var1(), FLA_Eig_gest_iu_opz_var1(), FLA_Obj_col_stride(), FLA_Obj_datatype(), FLA_Obj_length(), FLA_Obj_row_stride(), FLA_Obj_vector_inc(), and FLA_Part_1x2().
Referenced by FLA_Eig_gest_iu().
{
FLA_Datatype datatype;
int m_AB;
int rs_A, cs_A;
int rs_B, cs_B;
int inc_y;
FLA_Obj yL, yR;
datatype = FLA_Obj_datatype( A );
m_AB = FLA_Obj_length( A );
rs_A = FLA_Obj_row_stride( A );
cs_A = FLA_Obj_col_stride( A );
rs_B = FLA_Obj_row_stride( B );
cs_B = FLA_Obj_col_stride( B );
FLA_Part_1x2( Y, &yL, &yR, 1, FLA_LEFT );
inc_y = FLA_Obj_vector_inc( yL );
switch ( datatype )
{
case FLA_FLOAT:
{
float* buff_A = FLA_FLOAT_PTR( A );
float* buff_y = FLA_FLOAT_PTR( yL );
float* buff_B = FLA_FLOAT_PTR( B );
FLA_Eig_gest_iu_ops_var1( m_AB,
buff_A, rs_A, cs_A,
buff_y, inc_y,
buff_B, rs_B, cs_B );
break;
}
case FLA_DOUBLE:
{
double* buff_A = FLA_DOUBLE_PTR( A );
double* buff_y = FLA_DOUBLE_PTR( yL );
double* buff_B = FLA_DOUBLE_PTR( B );
FLA_Eig_gest_iu_opd_var1( m_AB,
buff_A, rs_A, cs_A,
buff_y, inc_y,
buff_B, rs_B, cs_B );
break;
}
case FLA_COMPLEX:
{
scomplex* buff_A = FLA_COMPLEX_PTR( A );
scomplex* buff_y = FLA_COMPLEX_PTR( yL );
scomplex* buff_B = FLA_COMPLEX_PTR( B );
FLA_Eig_gest_iu_opc_var1( m_AB,
buff_A, rs_A, cs_A,
buff_y, inc_y,
buff_B, rs_B, cs_B );
break;
}
case FLA_DOUBLE_COMPLEX:
{
dcomplex* buff_A = FLA_DOUBLE_COMPLEX_PTR( A );
dcomplex* buff_y = FLA_DOUBLE_COMPLEX_PTR( yL );
dcomplex* buff_B = FLA_DOUBLE_COMPLEX_PTR( B );
FLA_Eig_gest_iu_opz_var1( m_AB,
buff_A, rs_A, cs_A,
buff_y, inc_y,
buff_B, rs_B, cs_B );
break;
}
}
return FLA_SUCCESS;
}
| FLA_Error FLA_Eig_gest_iu_opt_var2 | ( | FLA_Obj | A, |
| FLA_Obj | Y, | ||
| FLA_Obj | B | ||
| ) |
References FLA_Eig_gest_iu_opc_var2(), FLA_Eig_gest_iu_opd_var2(), FLA_Eig_gest_iu_ops_var2(), FLA_Eig_gest_iu_opz_var2(), FLA_Obj_col_stride(), FLA_Obj_datatype(), FLA_Obj_length(), FLA_Obj_row_stride(), FLA_Obj_vector_inc(), and FLA_Part_1x2().
Referenced by FLA_Eig_gest_iu().
{
FLA_Datatype datatype;
int m_AB;
int rs_A, cs_A;
int rs_B, cs_B;
int inc_y;
FLA_Obj yL, yR;
datatype = FLA_Obj_datatype( A );
m_AB = FLA_Obj_length( A );
rs_A = FLA_Obj_row_stride( A );
cs_A = FLA_Obj_col_stride( A );
rs_B = FLA_Obj_row_stride( B );
cs_B = FLA_Obj_col_stride( B );
FLA_Part_1x2( Y, &yL, &yR, 1, FLA_LEFT );
inc_y = FLA_Obj_vector_inc( yL );
switch ( datatype )
{
case FLA_FLOAT:
{
float* buff_A = FLA_FLOAT_PTR( A );
float* buff_y = FLA_FLOAT_PTR( yL );
float* buff_B = FLA_FLOAT_PTR( B );
FLA_Eig_gest_iu_ops_var2( m_AB,
buff_A, rs_A, cs_A,
buff_y, inc_y,
buff_B, rs_B, cs_B );
break;
}
case FLA_DOUBLE:
{
double* buff_A = FLA_DOUBLE_PTR( A );
double* buff_y = FLA_DOUBLE_PTR( yL );
double* buff_B = FLA_DOUBLE_PTR( B );
FLA_Eig_gest_iu_opd_var2( m_AB,
buff_A, rs_A, cs_A,
buff_y, inc_y,
buff_B, rs_B, cs_B );
break;
}
case FLA_COMPLEX:
{
scomplex* buff_A = FLA_COMPLEX_PTR( A );
scomplex* buff_y = FLA_COMPLEX_PTR( yL );
scomplex* buff_B = FLA_COMPLEX_PTR( B );
FLA_Eig_gest_iu_opc_var2( m_AB,
buff_A, rs_A, cs_A,
buff_y, inc_y,
buff_B, rs_B, cs_B );
break;
}
case FLA_DOUBLE_COMPLEX:
{
dcomplex* buff_A = FLA_DOUBLE_COMPLEX_PTR( A );
dcomplex* buff_y = FLA_DOUBLE_COMPLEX_PTR( yL );
dcomplex* buff_B = FLA_DOUBLE_COMPLEX_PTR( B );
FLA_Eig_gest_iu_opz_var2( m_AB,
buff_A, rs_A, cs_A,
buff_y, inc_y,
buff_B, rs_B, cs_B );
break;
}
}
return FLA_SUCCESS;
}
| FLA_Error FLA_Eig_gest_iu_opt_var3 | ( | FLA_Obj | A, |
| FLA_Obj | Y, | ||
| FLA_Obj | B | ||
| ) |
References FLA_Eig_gest_iu_opc_var3(), FLA_Eig_gest_iu_opd_var3(), FLA_Eig_gest_iu_ops_var3(), FLA_Eig_gest_iu_opz_var3(), FLA_Obj_col_stride(), FLA_Obj_datatype(), FLA_Obj_length(), and FLA_Obj_row_stride().
Referenced by FLA_Eig_gest_iu().
{
FLA_Datatype datatype;
int m_AB;
int rs_A, cs_A;
int rs_B, cs_B;
int rs_Y, cs_Y;
datatype = FLA_Obj_datatype( A );
m_AB = FLA_Obj_length( A );
rs_A = FLA_Obj_row_stride( A );
cs_A = FLA_Obj_col_stride( A );
rs_B = FLA_Obj_row_stride( B );
cs_B = FLA_Obj_col_stride( B );
rs_Y = FLA_Obj_row_stride( Y );
cs_Y = FLA_Obj_col_stride( Y );
switch ( datatype )
{
case FLA_FLOAT:
{
float* buff_A = FLA_FLOAT_PTR( A );
float* buff_Y = FLA_FLOAT_PTR( Y );
float* buff_B = FLA_FLOAT_PTR( B );
FLA_Eig_gest_iu_ops_var3( m_AB,
buff_A, rs_A, cs_A,
buff_Y, rs_Y, cs_Y,
buff_B, rs_B, cs_B );
break;
}
case FLA_DOUBLE:
{
double* buff_A = FLA_DOUBLE_PTR( A );
double* buff_Y = FLA_DOUBLE_PTR( Y );
double* buff_B = FLA_DOUBLE_PTR( B );
FLA_Eig_gest_iu_opd_var3( m_AB,
buff_A, rs_A, cs_A,
buff_Y, rs_Y, cs_Y,
buff_B, rs_B, cs_B );
break;
}
case FLA_COMPLEX:
{
scomplex* buff_A = FLA_COMPLEX_PTR( A );
scomplex* buff_Y = FLA_COMPLEX_PTR( Y );
scomplex* buff_B = FLA_COMPLEX_PTR( B );
FLA_Eig_gest_iu_opc_var3( m_AB,
buff_A, rs_A, cs_A,
buff_Y, rs_Y, cs_Y,
buff_B, rs_B, cs_B );
break;
}
case FLA_DOUBLE_COMPLEX:
{
dcomplex* buff_A = FLA_DOUBLE_COMPLEX_PTR( A );
dcomplex* buff_Y = FLA_DOUBLE_COMPLEX_PTR( Y );
dcomplex* buff_B = FLA_DOUBLE_COMPLEX_PTR( B );
FLA_Eig_gest_iu_opz_var3( m_AB,
buff_A, rs_A, cs_A,
buff_Y, rs_Y, cs_Y,
buff_B, rs_B, cs_B );
break;
}
}
return FLA_SUCCESS;
}
| FLA_Error FLA_Eig_gest_iu_opt_var4 | ( | FLA_Obj | A, |
| FLA_Obj | Y, | ||
| FLA_Obj | B | ||
| ) |
References FLA_Eig_gest_iu_opc_var4(), FLA_Eig_gest_iu_opd_var4(), FLA_Eig_gest_iu_ops_var4(), FLA_Eig_gest_iu_opz_var4(), FLA_Obj_col_stride(), FLA_Obj_datatype(), FLA_Obj_length(), FLA_Obj_row_stride(), FLA_Obj_vector_inc(), and FLA_Part_1x2().
Referenced by FLA_Eig_gest_iu().
{
FLA_Datatype datatype;
int m_AB;
int rs_A, cs_A;
int rs_B, cs_B;
int inc_y;
FLA_Obj yL, yR;
datatype = FLA_Obj_datatype( A );
m_AB = FLA_Obj_length( A );
rs_A = FLA_Obj_row_stride( A );
cs_A = FLA_Obj_col_stride( A );
rs_B = FLA_Obj_row_stride( B );
cs_B = FLA_Obj_col_stride( B );
FLA_Part_1x2( Y, &yL, &yR, 1, FLA_LEFT );
inc_y = FLA_Obj_vector_inc( yL );
switch ( datatype )
{
case FLA_FLOAT:
{
float* buff_A = FLA_FLOAT_PTR( A );
float* buff_y = FLA_FLOAT_PTR( yL );
float* buff_B = FLA_FLOAT_PTR( B );
FLA_Eig_gest_iu_ops_var4( m_AB,
buff_A, rs_A, cs_A,
buff_y, inc_y,
buff_B, rs_B, cs_B );
break;
}
case FLA_DOUBLE:
{
double* buff_A = FLA_DOUBLE_PTR( A );
double* buff_y = FLA_DOUBLE_PTR( yL );
double* buff_B = FLA_DOUBLE_PTR( B );
FLA_Eig_gest_iu_opd_var4( m_AB,
buff_A, rs_A, cs_A,
buff_y, inc_y,
buff_B, rs_B, cs_B );
break;
}
case FLA_COMPLEX:
{
scomplex* buff_A = FLA_COMPLEX_PTR( A );
scomplex* buff_y = FLA_COMPLEX_PTR( yL );
scomplex* buff_B = FLA_COMPLEX_PTR( B );
FLA_Eig_gest_iu_opc_var4( m_AB,
buff_A, rs_A, cs_A,
buff_y, inc_y,
buff_B, rs_B, cs_B );
break;
}
case FLA_DOUBLE_COMPLEX:
{
dcomplex* buff_A = FLA_DOUBLE_COMPLEX_PTR( A );
dcomplex* buff_y = FLA_DOUBLE_COMPLEX_PTR( yL );
dcomplex* buff_B = FLA_DOUBLE_COMPLEX_PTR( B );
FLA_Eig_gest_iu_opz_var4( m_AB,
buff_A, rs_A, cs_A,
buff_y, inc_y,
buff_B, rs_B, cs_B );
break;
}
}
return FLA_SUCCESS;
}
| FLA_Error FLA_Eig_gest_iu_opt_var5 | ( | FLA_Obj | A, |
| FLA_Obj | Y, | ||
| FLA_Obj | B | ||
| ) |
References FLA_Eig_gest_iu_opc_var5(), FLA_Eig_gest_iu_opd_var5(), FLA_Eig_gest_iu_ops_var5(), FLA_Eig_gest_iu_opz_var5(), FLA_Obj_col_stride(), FLA_Obj_datatype(), FLA_Obj_length(), FLA_Obj_row_stride(), FLA_Obj_vector_inc(), and FLA_Part_1x2().
Referenced by FLA_Eig_gest_iu().
{
FLA_Datatype datatype;
int m_AB;
int rs_A, cs_A;
int rs_B, cs_B;
int inc_y;
FLA_Obj yL, yR;
datatype = FLA_Obj_datatype( A );
m_AB = FLA_Obj_length( A );
rs_A = FLA_Obj_row_stride( A );
cs_A = FLA_Obj_col_stride( A );
rs_B = FLA_Obj_row_stride( B );
cs_B = FLA_Obj_col_stride( B );
FLA_Part_1x2( Y, &yL, &yR, 1, FLA_LEFT );
inc_y = FLA_Obj_vector_inc( yL );
switch ( datatype )
{
case FLA_FLOAT:
{
float* buff_A = FLA_FLOAT_PTR( A );
float* buff_y = FLA_FLOAT_PTR( yL );
float* buff_B = FLA_FLOAT_PTR( B );
FLA_Eig_gest_iu_ops_var5( m_AB,
buff_A, rs_A, cs_A,
buff_y, inc_y,
buff_B, rs_B, cs_B );
break;
}
case FLA_DOUBLE:
{
double* buff_A = FLA_DOUBLE_PTR( A );
double* buff_y = FLA_DOUBLE_PTR( yL );
double* buff_B = FLA_DOUBLE_PTR( B );
FLA_Eig_gest_iu_opd_var5( m_AB,
buff_A, rs_A, cs_A,
buff_y, inc_y,
buff_B, rs_B, cs_B );
break;
}
case FLA_COMPLEX:
{
scomplex* buff_A = FLA_COMPLEX_PTR( A );
scomplex* buff_y = FLA_COMPLEX_PTR( yL );
scomplex* buff_B = FLA_COMPLEX_PTR( B );
FLA_Eig_gest_iu_opc_var5( m_AB,
buff_A, rs_A, cs_A,
buff_y, inc_y,
buff_B, rs_B, cs_B );
break;
}
case FLA_DOUBLE_COMPLEX:
{
dcomplex* buff_A = FLA_DOUBLE_COMPLEX_PTR( A );
dcomplex* buff_y = FLA_DOUBLE_COMPLEX_PTR( yL );
dcomplex* buff_B = FLA_DOUBLE_COMPLEX_PTR( B );
FLA_Eig_gest_iu_opz_var5( m_AB,
buff_A, rs_A, cs_A,
buff_y, inc_y,
buff_B, rs_B, cs_B );
break;
}
}
return FLA_SUCCESS;
}
| FLA_Error FLA_Eig_gest_iu_opz_var1 | ( | int | m_AB, |
| dcomplex * | buff_A, | ||
| int | rs_A, | ||
| int | cs_A, | ||
| dcomplex * | buff_y, | ||
| int | inc_y, | ||
| dcomplex * | buff_B, | ||
| int | rs_B, | ||
| int | cs_B | ||
| ) |
References bl1_zaxpyv(), bl1_zdot2s(), bl1_zhemv(), bl1_zinvscalv(), bl1_ztrsv(), BLIS1_CONJ_TRANSPOSE, BLIS1_CONJUGATE, BLIS1_NO_CONJUGATE, BLIS1_NONUNIT_DIAG, BLIS1_UPPER_TRIANGULAR, FLA_MINUS_ONE, FLA_MINUS_ONE_HALF, FLA_ONE, and FLA_ZERO.
Referenced by FLA_Eig_gest_iu_opt_var1().
{
dcomplex* buff_1 = FLA_DOUBLE_COMPLEX_PTR( FLA_ONE );
dcomplex* buff_0 = FLA_DOUBLE_COMPLEX_PTR( FLA_ZERO );
dcomplex* buff_m1 = FLA_DOUBLE_COMPLEX_PTR( FLA_MINUS_ONE );
dcomplex* buff_m1h = FLA_DOUBLE_COMPLEX_PTR( FLA_MINUS_ONE_HALF );
int i;
for ( i = 0; i < m_AB; ++i )
{
dcomplex* A00 = buff_A + (0 )*cs_A + (0 )*rs_A;
dcomplex* a01 = buff_A + (i )*cs_A + (0 )*rs_A;
dcomplex* alpha11 = buff_A + (i )*cs_A + (i )*rs_A;
dcomplex* y01 = buff_y + (0 )*inc_y;
dcomplex* B00 = buff_B + (0 )*cs_B + (0 )*rs_B;
dcomplex* b01 = buff_B + (i )*cs_B + (0 )*rs_B;
dcomplex* beta11 = buff_B + (i )*cs_B + (i )*rs_B;
int m_behind = i;
/*------------------------------------------------------------*/
// FLA_Hemvc_external( FLA_UPPER_TRIANGULAR, FLA_NO_CONJUGATE,
// FLA_ONE, A00, b01, FLA_ZERO, y01_l );
bl1_zhemv( BLIS1_UPPER_TRIANGULAR,
BLIS1_NO_CONJUGATE,
m_behind,
buff_1,
A00, rs_A, cs_A,
b01, rs_B,
buff_0,
y01, inc_y );
// FLA_Trsv_external( FLA_UPPER_TRIANGULAR, FLA_CONJ_TRANSPOSE, FLA_NONUNIT_DIAG,
// B00, a01 );
bl1_ztrsv( BLIS1_UPPER_TRIANGULAR,
BLIS1_CONJ_TRANSPOSE,
BLIS1_NONUNIT_DIAG,
m_behind,
B00, rs_B, cs_B,
a01, rs_A );
// FLA_Axpy_external( FLA_MINUS_ONE_HALF, y01_l, a01 );
bl1_zaxpyv( BLIS1_NO_CONJUGATE,
m_behind,
buff_m1h,
y01, inc_y,
a01, rs_A );
// FLA_Dot2cs_external( FLA_CONJUGATE, FLA_MINUS_ONE, a01, b01, FLA_ONE, alpha11 );
bl1_zdot2s( BLIS1_CONJUGATE,
m_behind,
buff_m1,
a01, rs_A,
b01, rs_B,
buff_1,
alpha11 );
// FLA_Inv_scal_external( beta11, alpha11 );
// FLA_Inv_scal_external( beta11, alpha11 );
bl1_zinvscals( beta11, alpha11 );
bl1_zinvscals( beta11, alpha11 );
// FLA_Axpy_external( FLA_MINUS_ONE_HALF, y01_l, a01 );
bl1_zaxpyv( BLIS1_NO_CONJUGATE,
m_behind,
buff_m1h,
y01, inc_y,
a01, rs_A );
// FLA_Inv_scal_external( beta11, a01 );
bl1_zinvscalv( BLIS1_NO_CONJUGATE,
m_behind,
beta11,
a01, rs_A );
/*------------------------------------------------------------*/
}
return FLA_SUCCESS;
}
| FLA_Error FLA_Eig_gest_iu_opz_var2 | ( | int | m_AB, |
| dcomplex * | buff_A, | ||
| int | rs_A, | ||
| int | cs_A, | ||
| dcomplex * | buff_y, | ||
| int | inc_y, | ||
| dcomplex * | buff_B, | ||
| int | rs_B, | ||
| int | cs_B | ||
| ) |
References bl1_zaxpyv(), bl1_zdot2s(), bl1_zgemv(), bl1_zhemv(), bl1_zinvscalv(), BLIS1_CONJUGATE, BLIS1_NO_CONJUGATE, BLIS1_TRANSPOSE, BLIS1_UPPER_TRIANGULAR, FLA_MINUS_ONE, FLA_ONE, FLA_ONE_HALF, and FLA_ZERO.
Referenced by FLA_Eig_gest_iu_opt_var2().
{
dcomplex* buff_1 = FLA_DOUBLE_COMPLEX_PTR( FLA_ONE );
dcomplex* buff_1h = FLA_DOUBLE_COMPLEX_PTR( FLA_ONE_HALF );
dcomplex* buff_0 = FLA_DOUBLE_COMPLEX_PTR( FLA_ZERO );
dcomplex* buff_m1 = FLA_DOUBLE_COMPLEX_PTR( FLA_MINUS_ONE );
int i;
for ( i = 0; i < m_AB; ++i )
{
dcomplex* A00 = buff_A + (0 )*cs_A + (0 )*rs_A;
dcomplex* a01 = buff_A + (i )*cs_A + (0 )*rs_A;
dcomplex* alpha11 = buff_A + (i )*cs_A + (i )*rs_A;
dcomplex* A02 = buff_A + (i+1)*cs_A + (0 )*rs_A;
dcomplex* a12t = buff_A + (i+1)*cs_A + (i )*rs_A;
dcomplex* y01 = buff_y + (0 )*inc_y;
dcomplex* b01 = buff_B + (i )*cs_B + (0 )*rs_B;
dcomplex* beta11 = buff_B + (i )*cs_B + (i )*rs_B;
int m_ahead = m_AB - i - 1;
int m_behind = i;
/*------------------------------------------------------------*/
// FLA_Hemvc_external( FLA_UPPER_TRIANGULAR, FLA_NO_CONJUGATE,
// FLA_ONE_HALF, A00, b01, FLA_ZERO, y01_l );
bl1_zhemv( BLIS1_UPPER_TRIANGULAR,
BLIS1_NO_CONJUGATE,
m_behind,
buff_1h,
A00, rs_A, cs_A,
b01, rs_B,
buff_0,
y01, inc_y );
// FLA_Axpy_external( FLA_MINUS_ONE, y01_l, a01 );
bl1_zaxpyv( BLIS1_NO_CONJUGATE,
m_behind,
buff_m1,
y01, inc_y,
a01, rs_A );
// FLA_Dot2cs_external( FLA_CONJUGATE, FLA_MINUS_ONE, a01, b01, FLA_ONE, alpha11 );
bl1_zdot2s( BLIS1_CONJUGATE,
m_behind,
buff_m1,
a01, rs_A,
b01, rs_B,
buff_1,
alpha11 );
// FLA_Inv_scal_external( beta11, alpha11 );
// FLA_Inv_scal_external( beta11, alpha11 );
bl1_zinvscals( beta11, alpha11 );
bl1_zinvscals( beta11, alpha11 );
// FLA_Gemvc_external( FLA_TRANSPOSE, FLA_CONJUGATE,
// FLA_MINUS_ONE, A02, b01, FLA_ONE, a12t );
bl1_zgemv( BLIS1_TRANSPOSE,
BLIS1_CONJUGATE,
m_behind,
m_ahead,
buff_m1,
A02, rs_A, cs_A,
b01, rs_B,
buff_1,
a12t, cs_A );
// FLA_Inv_scal_external( beta11, a12t );
bl1_zinvscalv( BLIS1_NO_CONJUGATE,
m_ahead,
beta11,
a12t, cs_A );
// FLA_Axpy_external( FLA_MINUS_ONE, y01_l, a01 );
bl1_zaxpyv( BLIS1_NO_CONJUGATE,
m_behind,
buff_m1,
y01, inc_y,
a01, rs_A );
// FLA_Inv_scal_external( beta11, a01 );
bl1_zinvscalv( BLIS1_NO_CONJUGATE,
m_behind,
beta11,
a01, rs_A );
/*------------------------------------------------------------*/
}
return FLA_SUCCESS;
}
| FLA_Error FLA_Eig_gest_iu_opz_var3 | ( | int | m_AB, |
| dcomplex * | buff_A, | ||
| int | rs_A, | ||
| int | cs_A, | ||
| dcomplex * | buff_Y, | ||
| int | rs_Y, | ||
| int | cs_Y, | ||
| dcomplex * | buff_B, | ||
| int | rs_B, | ||
| int | cs_B | ||
| ) |
References bl1_zaxpyv(), bl1_zcopyv(), bl1_zdot2s(), bl1_zgemv(), bl1_zger(), bl1_zinvscalv(), bl1_zscalv(), BLIS1_CONJUGATE, BLIS1_NO_CONJUGATE, BLIS1_TRANSPOSE, FLA_MINUS_ONE, FLA_MINUS_ONE_HALF, and FLA_ONE.
Referenced by FLA_Eig_gest_iu_opt_var3().
{
dcomplex* buff_1 = FLA_DOUBLE_COMPLEX_PTR( FLA_ONE );
dcomplex* buff_m1 = FLA_DOUBLE_COMPLEX_PTR( FLA_MINUS_ONE );
dcomplex* buff_m1h = FLA_DOUBLE_COMPLEX_PTR( FLA_MINUS_ONE_HALF );
int i;
for ( i = 0; i < m_AB; ++i )
{
dcomplex* a01 = buff_A + (i )*cs_A + (0 )*rs_A;
dcomplex* alpha11 = buff_A + (i )*cs_A + (i )*rs_A;
dcomplex* A02 = buff_A + (i+1)*cs_A + (0 )*rs_A;
dcomplex* a12t = buff_A + (i+1)*cs_A + (i )*rs_A;
dcomplex* y01 = buff_Y + (i )*cs_Y + (0 )*rs_Y;
dcomplex* Y02 = buff_Y + (i+1)*cs_Y + (0 )*rs_Y;
dcomplex* y12t = buff_Y + (i+1)*cs_Y + (i )*rs_Y;
dcomplex* b01 = buff_B + (i )*cs_B + (0 )*rs_B;
dcomplex* beta11 = buff_B + (i )*cs_B + (i )*rs_B;
dcomplex* B02 = buff_B + (i+1)*cs_B + (0 )*rs_B;
dcomplex* b12t = buff_B + (i+1)*cs_B + (i )*rs_B;
int m_ahead = m_AB - i - 1;
int m_behind = i;
/*------------------------------------------------------------*/
// FLA_Axpy_external( FLA_MINUS_ONE_HALF, y01, a01 );
bl1_zaxpyv( BLIS1_NO_CONJUGATE,
m_behind,
buff_m1h,
y01, rs_Y,
a01, rs_A );
// FLA_Dot2cs_external( FLA_CONJUGATE, FLA_MINUS_ONE, a01, b01, FLA_ONE, alpha11 );
bl1_zdot2s( BLIS1_CONJUGATE,
m_behind,
buff_m1,
a01, rs_A,
b01, rs_B,
buff_1,
alpha11 );
// FLA_Inv_scal_external( beta11, alpha11 );
// FLA_Inv_scal_external( beta11, alpha11 );
bl1_zinvscals( beta11, alpha11 );
bl1_zinvscals( beta11, alpha11 );
// FLA_Gemvc_external( FLA_TRANSPOSE, FLA_CONJUGATE,
// FLA_MINUS_ONE, A02, b01, FLA_ONE, a12t );
bl1_zgemv( BLIS1_TRANSPOSE,
BLIS1_CONJUGATE,
m_behind,
m_ahead,
buff_m1,
A02, rs_A, cs_A,
b01, rs_B,
buff_1,
a12t, cs_A );
// FLA_Inv_scal_external( beta11, a12t );
bl1_zinvscalv( BLIS1_NO_CONJUGATE,
m_ahead,
beta11,
a12t, cs_A );
// FLA_Axpy_external( FLA_MINUS_ONE_HALF, y01, a01 );
bl1_zaxpyv( BLIS1_NO_CONJUGATE,
m_behind,
buff_m1h,
y01, rs_Y,
a01, rs_A );
// FLA_Inv_scal_external( beta11, a01 );
bl1_zinvscalv( BLIS1_NO_CONJUGATE,
m_behind,
beta11,
a01, rs_A );
// FLA_Ger_external( FLA_ONE, a01, b12t, Y02 );
bl1_zger( BLIS1_NO_CONJUGATE,
BLIS1_NO_CONJUGATE,
m_behind,
m_ahead,
buff_1,
a01, rs_A,
b12t, cs_B,
Y02, rs_Y, cs_Y );
// FLA_Copy_external( b12t, y12t );
// FLA_Scal_external( alpha11, y12t );
bl1_zcopyv( BLIS1_NO_CONJUGATE,
m_ahead,
b12t, cs_B,
y12t, cs_Y );
bl1_zscalv( BLIS1_NO_CONJUGATE,
m_ahead,
alpha11,
y12t, cs_Y );
// FLA_Gemvc_external( FLA_TRANSPOSE, FLA_CONJUGATE,
// FLA_ONE, B02, a01, FLA_ONE, y12t );
bl1_zgemv( BLIS1_TRANSPOSE,
BLIS1_CONJUGATE,
m_behind,
m_ahead,
buff_1,
B02, rs_B, cs_B,
a01, rs_A,
buff_1,
y12t, cs_Y );
/*------------------------------------------------------------*/
}
return FLA_SUCCESS;
}
| FLA_Error FLA_Eig_gest_iu_opz_var4 | ( | int | m_AB, |
| dcomplex * | buff_A, | ||
| int | rs_A, | ||
| int | cs_A, | ||
| dcomplex * | buff_y, | ||
| int | inc_y, | ||
| dcomplex * | buff_B, | ||
| int | rs_B, | ||
| int | cs_B | ||
| ) |
References bl1_zaxpyv(), bl1_zger(), bl1_zher2(), bl1_zinvscalv(), BLIS1_CONJUGATE, BLIS1_NO_CONJUGATE, BLIS1_UPPER_TRIANGULAR, FLA_MINUS_ONE, and FLA_MINUS_ONE_HALF.
Referenced by FLA_Eig_gest_iu_opt_var4().
{
dcomplex* buff_m1 = FLA_DOUBLE_COMPLEX_PTR( FLA_MINUS_ONE );
dcomplex* buff_m1h = FLA_DOUBLE_COMPLEX_PTR( FLA_MINUS_ONE_HALF );
int i;
for ( i = 0; i < m_AB; ++i )
{
dcomplex* a01 = buff_A + (i )*cs_A + (0 )*rs_A;
dcomplex* alpha11 = buff_A + (i )*cs_A + (i )*rs_A;
dcomplex* A02 = buff_A + (i+1)*cs_A + (0 )*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* beta11 = buff_B + (i )*cs_B + (i )*rs_B;
dcomplex* b12t = buff_B + (i+1)*cs_B + (i )*rs_B;
dcomplex psi11;
int m_ahead = m_AB - i - 1;
int m_behind = i;
/*------------------------------------------------------------*/
// FLA_Inv_scal_external( beta11, a01 );
bl1_zinvscalv( BLIS1_NO_CONJUGATE,
m_behind,
beta11,
a01, rs_A );
// FLA_Ger_external( FLA_MINUS_ONE, a01, b12t, A02 );
bl1_zger( BLIS1_NO_CONJUGATE,
BLIS1_NO_CONJUGATE,
m_behind,
m_ahead,
buff_m1,
a01, rs_A,
b12t, cs_B,
A02, rs_A, cs_A );
// FLA_Inv_scal_external( beta11, alpha11 );
// FLA_Inv_scal_external( beta11, alpha11 );
bl1_zinvscals( beta11, alpha11 );
bl1_zinvscals( beta11, alpha11 );
// FLA_Copy_external( alpha11, psi11 );
// FLA_Scal_external( FLA_MINUS_ONE_HALF, psi11 );
bl1_zmult3( buff_m1h, alpha11, &psi11 );
// FLA_Inv_scal_external( beta11, a12t );
bl1_zinvscalv( BLIS1_NO_CONJUGATE,
m_ahead,
beta11,
a12t, cs_A );
// FLA_Axpy_external( psi11, b12t, a12t );
bl1_zaxpyv( BLIS1_NO_CONJUGATE,
m_ahead,
&psi11,
b12t, cs_B,
a12t, cs_A );
// FLA_Her2c_external( FLA_UPPER_TRIANGULAR, FLA_CONJUGATE,
// FLA_MINUS_ONE, a12t, b12t, A22 );
bl1_zher2( BLIS1_UPPER_TRIANGULAR,
BLIS1_CONJUGATE,
m_ahead,
buff_m1,
a12t, cs_A,
b12t, cs_B,
A22, rs_A, cs_A );
// FLA_Axpy_external( psi11, b12t, a12t );
bl1_zaxpyv( BLIS1_NO_CONJUGATE,
m_ahead,
&psi11,
b12t, cs_B,
a12t, cs_A );
/*------------------------------------------------------------*/
}
return FLA_SUCCESS;
}
| FLA_Error FLA_Eig_gest_iu_opz_var5 | ( | int | m_AB, |
| dcomplex * | buff_A, | ||
| int | rs_A, | ||
| int | cs_A, | ||
| dcomplex * | buff_y, | ||
| int | inc_y, | ||
| dcomplex * | buff_B, | ||
| int | rs_B, | ||
| int | cs_B | ||
| ) |
References bl1_zaxpyv(), bl1_zher2(), bl1_zinvscalv(), bl1_ztrsv(), BLIS1_CONJUGATE, BLIS1_NO_CONJUGATE, BLIS1_NONUNIT_DIAG, BLIS1_TRANSPOSE, BLIS1_UPPER_TRIANGULAR, FLA_MINUS_ONE, and FLA_MINUS_ONE_HALF.
Referenced by FLA_Eig_gest_iu_opt_var5().
{
dcomplex* buff_m1 = FLA_DOUBLE_COMPLEX_PTR( FLA_MINUS_ONE );
dcomplex* buff_m1h = FLA_DOUBLE_COMPLEX_PTR( FLA_MINUS_ONE_HALF );
int i;
for ( i = 0; i < m_AB; ++i )
{
dcomplex* alpha11 = buff_A + (i )*cs_A + (i )*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* beta11 = buff_B + (i )*cs_B + (i )*rs_B;
dcomplex* b12t = buff_B + (i+1)*cs_B + (i )*rs_B;
dcomplex* B22 = buff_B + (i+1)*cs_B + (i+1)*rs_B;
dcomplex psi11;
int m_ahead = m_AB - i - 1;
/*------------------------------------------------------------*/
// FLA_Inv_scal_external( beta11, alpha11 );
// FLA_Inv_scal_external( beta11, alpha11 );
bl1_zinvscals( beta11, alpha11 );
bl1_zinvscals( beta11, alpha11 );
// FLA_Copy_external( alpha11, psi11 );
// FLA_Scal_external( FLA_MINUS_ONE_HALF, psi11 );
bl1_zmult3( buff_m1h, alpha11, &psi11 );
// FLA_Inv_scal_external( beta11, a12t );
bl1_zinvscalv( BLIS1_NO_CONJUGATE,
m_ahead,
beta11,
a12t, cs_A );
// FLA_Axpy_external( psi11, b12t, a12t );
bl1_zaxpyv( BLIS1_NO_CONJUGATE,
m_ahead,
&psi11,
b12t, cs_B,
a12t, cs_A );
// FLA_Her2c_external( FLA_UPPER_TRIANGULAR, FLA_CONJUGATE,
// FLA_MINUS_ONE, a12t, b12t, A22 );
bl1_zher2( BLIS1_UPPER_TRIANGULAR,
BLIS1_CONJUGATE,
m_ahead,
buff_m1,
a12t, cs_A,
b12t, cs_B,
A22, rs_A, cs_A );
// FLA_Axpy_external( psi11, b12t, a12t );
bl1_zaxpyv( BLIS1_NO_CONJUGATE,
m_ahead,
&psi11,
b12t, cs_B,
a12t, cs_A );
// FLA_Trsv_external( FLA_UPPER_TRIANGULAR, FLA_TRANSPOSE, FLA_NONUNIT_DIAG,
// B22, a12t );
bl1_ztrsv( BLIS1_UPPER_TRIANGULAR,
BLIS1_TRANSPOSE,
BLIS1_NONUNIT_DIAG,
m_ahead,
B22, rs_B, cs_B,
a12t, cs_A );
/*------------------------------------------------------------*/
}
return FLA_SUCCESS;
}
| FLA_Error FLA_Eig_gest_iu_unb_var1 | ( | FLA_Obj | A, |
| FLA_Obj | Y, | ||
| FLA_Obj | B | ||
| ) |
References FLA_Axpy_external(), FLA_Cont_with_3x1_to_2x1(), FLA_Cont_with_3x3_to_2x2(), FLA_Dot2cs_external(), FLA_Hemvc_external(), FLA_Inv_scal_external(), FLA_MINUS_ONE, FLA_MINUS_ONE_HALF, FLA_Obj_length(), FLA_ONE, FLA_Part_1x2(), FLA_Part_2x1(), FLA_Part_2x2(), FLA_Repart_2x1_to_3x1(), FLA_Repart_2x2_to_3x3(), FLA_Trsv_external(), and FLA_ZERO.
Referenced by FLA_Eig_gest_iu().
{
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 = A00 * b01;
FLA_Hemvc_external( FLA_UPPER_TRIANGULAR, FLA_NO_CONJUGATE,
FLA_ONE, A00, b01, FLA_ZERO, y01_l );
// a01 = inv( triu( B00 )' ) * a01;
FLA_Trsv_external( FLA_UPPER_TRIANGULAR, FLA_CONJ_TRANSPOSE, FLA_NONUNIT_DIAG,
B00, a01 );
// a01 = a01 - 1/2 * y01;
FLA_Axpy_external( FLA_MINUS_ONE_HALF, 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 );
// a01 = a01 - 1/2 * y01;
FLA_Axpy_external( FLA_MINUS_ONE_HALF, 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_Eig_gest_iu_unb_var2 | ( | FLA_Obj | A, |
| FLA_Obj | Y, | ||
| FLA_Obj | B | ||
| ) |
References FLA_Axpy_external(), FLA_Cont_with_3x1_to_2x1(), FLA_Cont_with_3x3_to_2x2(), FLA_Dot2cs_external(), FLA_Gemvc_external(), FLA_Hemvc_external(), FLA_Inv_scal_external(), FLA_MINUS_ONE, FLA_Obj_length(), FLA_ONE, FLA_ONE_HALF, FLA_Part_1x2(), FLA_Part_2x1(), FLA_Part_2x2(), FLA_Repart_2x1_to_3x1(), FLA_Repart_2x2_to_3x3(), and FLA_ZERO.
Referenced by FLA_Eig_gest_iu().
{
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_Eig_gest_iu_unb_var3 | ( | FLA_Obj | A, |
| FLA_Obj | Y, | ||
| FLA_Obj | B | ||
| ) |
References FLA_Axpy_external(), FLA_Cont_with_3x3_to_2x2(), FLA_Copy_external(), FLA_Dot2cs_external(), FLA_Gemvc_external(), FLA_Ger_external(), FLA_Inv_scal_external(), FLA_MINUS_ONE, FLA_MINUS_ONE_HALF, FLA_Obj_length(), FLA_ONE, FLA_Part_2x2(), FLA_Repart_2x2_to_3x3(), and FLA_Scal_external().
Referenced by FLA_Eig_gest_iu().
{
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 YTL, YTR, Y00, y01, Y02,
YBL, YBR, y10t, psi11, y12t,
Y20, y21, Y22;
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_2x2( Y, &YTL, &YTR,
&YBL, &YBR, 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 );
FLA_Repart_2x2_to_3x3( BTL, /**/ BTR, &B00, /**/ &b01, &B02,
/* ************* */ /* ************************* */
&b10t, /**/ &beta11, &b12t,
BBL, /**/ BBR, &B20, /**/ &b21, &B22,
1, 1, FLA_BR );
FLA_Repart_2x2_to_3x3( YTL, /**/ YTR, &Y00, /**/ &y01, &Y02,
/* ************* */ /* ************************ */
&y10t, /**/ &psi11, &y12t,
YBL, /**/ YBR, &Y20, /**/ &y21, &Y22,
1, 1, FLA_BR );
/*------------------------------------------------------------*/
// a01 = a01 - 1/2 * y01;
FLA_Axpy_external( FLA_MINUS_ONE_HALF, y01, 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 - 1/2 * y01;
FLA_Axpy_external( FLA_MINUS_ONE_HALF, y01, a01 );
// a01 = a01 * inv(beta11);
FLA_Inv_scal_external( beta11, a01 );
// Y02 = Y02 + a01 * b12t;
FLA_Ger_external( FLA_ONE, a01, b12t, Y02 );
// y12t = alpha11 * b12t;
FLA_Copy_external( b12t, y12t );
FLA_Scal_external( alpha11, y12t );
// y12t = y12t + a01' * B02;
// y12t^T = y12t^T + B02^T * conj(a01);
FLA_Gemvc_external( FLA_TRANSPOSE, FLA_CONJUGATE,
FLA_ONE, B02, a01, FLA_ONE, y12t );
/*------------------------------------------------------------*/
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_3x3_to_2x2( &YTL, /**/ &YTR, Y00, y01, /**/ Y02,
y10t, psi11, /**/ y12t,
/* ************** */ /* ********************** */
&YBL, /**/ &YBR, Y20, y21, /**/ Y22,
FLA_TL );
}
return FLA_SUCCESS;
}
| FLA_Error FLA_Eig_gest_iu_unb_var4 | ( | FLA_Obj | A, |
| FLA_Obj | Y, | ||
| FLA_Obj | B | ||
| ) |
References FLA_Axpy_external(), FLA_Cont_with_3x3_to_2x2(), FLA_Copy_external(), FLA_Ger_external(), FLA_Her2c_external(), FLA_Inv_scal_external(), FLA_MINUS_ONE, FLA_MINUS_ONE_HALF, FLA_Obj_length(), FLA_Part_2x2(), FLA_Repart_2x2_to_3x3(), and FLA_Scal_external().
Referenced by FLA_Eig_gest_iu().
{
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 yL, yR, y10t, psi11, y12t;
//FLA_Obj y12t_t,
// y12t_b;
FLA_Obj psi11, y12t,
y21, Y22;
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_1x2( Y, &yL, &yR, 0, FLA_LEFT );
FLA_Part_2x2( Y, &psi11, &y12t,
&y21, &Y22, 1, 1, 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 );
FLA_Repart_2x2_to_3x3( BTL, /**/ BTR, &B00, /**/ &b01, &B02,
/* ************* */ /* ************************* */
&b10t, /**/ &beta11, &b12t,
BBL, /**/ BBR, &B20, /**/ &b21, &B22,
1, 1, FLA_BR );
//FLA_Repart_1x2_to_1x3( yL, /**/ yR, &y10t, /**/ &psi11, &y12t,
// 1, FLA_RIGHT );
/*------------------------------------------------------------*/
//FLA_Part_2x1( y12t, &y12t_t,
// &y12t_b, 1, FLA_TOP );
// a01 = a01 * inv(beta11);
FLA_Inv_scal_external( beta11, a01 );
// A02 = A02 - a01 * b12t;
FLA_Ger_external( FLA_MINUS_ONE, a01, b12t, A02 );
// alpha11 = inv(beta11) * alpha11 * inv(conj(beta11));
// = inv(beta11) * alpha11 * inv(beta11);
FLA_Inv_scal_external( beta11, alpha11 );
FLA_Inv_scal_external( beta11, alpha11 );
//// y12t = alpha11 * b12t;
//FLA_Copy_external( b12t, y12t_t );
//FLA_Scal_external( alpha11, y12t_t );
// psi11 = - 1/2 * alpha11;
FLA_Copy_external( alpha11, psi11 );
FLA_Scal_external( FLA_MINUS_ONE_HALF, psi11 );
// a12t = inv(conj(beta11)) * a12t;
// = inv(beta11) * a12t;
FLA_Inv_scal_external( beta11, a12t );
//// a12t = a12t - 1/2 * y12t;
//FLA_Axpy_external( FLA_MINUS_ONE_HALF, y12t_t, a12t );
// a12t = a12t - 1/2 * alpha11 * b12t;
FLA_Axpy_external( psi11, b12t, a12t );
// A22 = A22 - a12t' * b12t - b12t' * a12t;
FLA_Her2c_external( FLA_UPPER_TRIANGULAR, FLA_CONJUGATE,
FLA_MINUS_ONE, a12t, b12t, A22 );
//// a12t = a12t - 1/2 * y12t;
//FLA_Axpy_external( FLA_MINUS_ONE_HALF, y12t_t, a12t );
// a12t = a12t - 1/2 * alpha11 * b12t;
FLA_Axpy_external( psi11, b12t, a12t );
/*------------------------------------------------------------*/
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_1x3_to_1x2( &yL, /**/ &yR, y10t, psi11, /**/ y12t,
// FLA_LEFT );
}
return FLA_SUCCESS;
}
| FLA_Error FLA_Eig_gest_iu_unb_var5 | ( | FLA_Obj | A, |
| FLA_Obj | Y, | ||
| FLA_Obj | B | ||
| ) |
References FLA_Axpy_external(), FLA_Cont_with_3x3_to_2x2(), FLA_Copy_external(), FLA_Her2c_external(), FLA_Inv_scal_external(), FLA_MINUS_ONE, FLA_MINUS_ONE_HALF, FLA_Obj_length(), FLA_Part_2x2(), FLA_Repart_2x2_to_3x3(), FLA_Scal_external(), and FLA_Trsv_external().
Referenced by FLA_Eig_gest_iu().
{
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 yL, yR, y10t, psi11, y12t;
//FLA_Obj y12t_t,
// y12t_b;
FLA_Obj psi11, y12t,
y21, Y22;
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_1x2( Y, &yL, &yR, 0, FLA_LEFT );
FLA_Part_2x2( Y, &psi11, &y12t,
&y21, &Y22, 1, 1, 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 );
FLA_Repart_2x2_to_3x3( BTL, /**/ BTR, &B00, /**/ &b01, &B02,
/* ************* */ /* ************************* */
&b10t, /**/ &beta11, &b12t,
BBL, /**/ BBR, &B20, /**/ &b21, &B22,
1, 1, FLA_BR );
//FLA_Repart_1x2_to_1x3( yL, /**/ yR, &y10t, /**/ &psi11, &y12t,
// 1, FLA_RIGHT );
/*------------------------------------------------------------*/
//FLA_Part_2x1( y12t, &y12t_t,
// &y12t_b, 1, FLA_TOP );
// alpha11 = inv(beta11) * alpha11 * inv(conj(beta11));
// = inv(beta11) * alpha11 * inv(beta11);
FLA_Inv_scal_external( beta11, alpha11 );
FLA_Inv_scal_external( beta11, alpha11 );
//// y12t = alpha11 * b12t;
//FLA_Copy_external( b12t, y12t_t );
//FLA_Scal_external( alpha11, y12t_t );
// psi11 = - 1/2 * alpha11;
FLA_Copy_external( alpha11, psi11 );
FLA_Scal_external( FLA_MINUS_ONE_HALF, psi11 );
// a12t = inv(conj(beta11)) * a12t;
// = inv(beta11) * a12t;
FLA_Inv_scal_external( beta11, a12t );
//// a12t = a12t - 1/2 * y12t;
//FLA_Axpy_external( FLA_MINUS_ONE_HALF, y12t_t, a12t );
// a12t = a12t - 1/2 * alpha11 * b12t;
FLA_Axpy_external( psi11, b12t, a12t );
// A22 = A22 - a12t' * b12t - b12t' * a12t;
FLA_Her2c_external( FLA_UPPER_TRIANGULAR, FLA_CONJUGATE,
FLA_MINUS_ONE, a12t, b12t, A22 );
//// a12t = a12t - 1/2 * y12t;
//FLA_Axpy_external( FLA_MINUS_ONE_HALF, y12t_t, a12t );
// a12t = a12t - 1/2 * alpha11 * b12t;
FLA_Axpy_external( psi11, b12t, a12t );
// a12t = a12t * inv( tril( B22 ) );
// a12t^T = inv( triu( B22 )^T ) * a12t^T;
FLA_Trsv_external( FLA_UPPER_TRIANGULAR, FLA_TRANSPOSE, FLA_NONUNIT_DIAG,
B22, a12t );
/*------------------------------------------------------------*/
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_1x3_to_1x2( &yL, /**/ &yR, y10t, psi11, /**/ y12t,
// FLA_LEFT );
}
return FLA_SUCCESS;
}
1.7.6.1