libflame revision_anchor
|
Go to the source code of this file.
Functions | |
FLA_Error | FLA_Eig_gest_il_blk_var1 (FLA_Obj A, FLA_Obj Y, FLA_Obj B, fla_eig_gest_t *cntl) |
FLA_Error | FLA_Eig_gest_il_blk_var2 (FLA_Obj A, FLA_Obj Y, FLA_Obj B, fla_eig_gest_t *cntl) |
FLA_Error | FLA_Eig_gest_il_blk_var3 (FLA_Obj A, FLA_Obj Y, FLA_Obj B, fla_eig_gest_t *cntl) |
FLA_Error | FLA_Eig_gest_il_blk_var4 (FLA_Obj A, FLA_Obj Y, FLA_Obj B, fla_eig_gest_t *cntl) |
FLA_Error | FLA_Eig_gest_il_blk_var5 (FLA_Obj A, FLA_Obj Y, FLA_Obj B, fla_eig_gest_t *cntl) |
FLA_Error | FLA_Eig_gest_il_unb_var1 (FLA_Obj A, FLA_Obj Y, FLA_Obj B) |
FLA_Error | FLA_Eig_gest_il_unb_var2 (FLA_Obj A, FLA_Obj Y, FLA_Obj B) |
FLA_Error | FLA_Eig_gest_il_unb_var3 (FLA_Obj A, FLA_Obj Y, FLA_Obj B) |
FLA_Error | FLA_Eig_gest_il_unb_var4 (FLA_Obj A, FLA_Obj Y, FLA_Obj B) |
FLA_Error | FLA_Eig_gest_il_unb_var5 (FLA_Obj A, FLA_Obj Y, FLA_Obj B) |
FLA_Error | FLA_Eig_gest_il_opt_var1 (FLA_Obj A, FLA_Obj Y, FLA_Obj B) |
FLA_Error | FLA_Eig_gest_il_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_il_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_il_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_il_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_il_opt_var2 (FLA_Obj A, FLA_Obj Y, FLA_Obj B) |
FLA_Error | FLA_Eig_gest_il_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_il_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_il_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_il_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_il_opt_var3 (FLA_Obj A, FLA_Obj Y, FLA_Obj B) |
FLA_Error | FLA_Eig_gest_il_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_il_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_il_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_il_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_il_opt_var4 (FLA_Obj A, FLA_Obj Y, FLA_Obj B) |
FLA_Error | FLA_Eig_gest_il_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_il_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_il_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_il_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_il_opt_var5 (FLA_Obj A, FLA_Obj Y, FLA_Obj B) |
FLA_Error | FLA_Eig_gest_il_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_il_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_il_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_il_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_il_blk_var1 | ( | 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_il().
{ 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 Y10_t, Y10_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( Y10, &Y10_t, &Y10_b, b, FLA_TOP ); // Y10 = B10 * A00; FLA_Hemm_internal( FLA_RIGHT, FLA_LOWER_TRIANGULAR, FLA_ONE, A00, B10, FLA_ZERO, Y10_t, FLA_Cntl_sub_hemm( cntl ) ); // A10 = A10 * inv( tril( B00 )' ); FLA_Trsm_internal( FLA_RIGHT, FLA_LOWER_TRIANGULAR, FLA_CONJ_TRANSPOSE, FLA_NONUNIT_DIAG, FLA_ONE, B00, A10, FLA_Cntl_sub_trsm1( cntl ) ); // A10 = A10 - 1/2 * Y10; FLA_Axpy_internal( FLA_MINUS_ONE_HALF, Y10_t, A10, FLA_Cntl_sub_axpy1( cntl ) ); // A11 = A11 - A10 * B10' - B10 * A10'; FLA_Her2k_internal( FLA_LOWER_TRIANGULAR, FLA_NO_TRANSPOSE, FLA_MINUS_ONE, A10, B10, FLA_ONE, A11, FLA_Cntl_sub_her2k( cntl ) ); // A11 = inv( tril( B11 ) ) * A11 * inv( tril( B11 )' ); FLA_Eig_gest_internal( FLA_INVERSE, FLA_LOWER_TRIANGULAR, A11, Y11, B11, FLA_Cntl_sub_eig_gest( cntl ) ); // A10 = A10 - 1/2 * Y10; FLA_Axpy_internal( FLA_MINUS_ONE_HALF, Y10_t, A10, FLA_Cntl_sub_axpy2( cntl ) ); // A10 = inv( tril( B11 ) ) * A10; FLA_Trsm_internal( FLA_LEFT, FLA_LOWER_TRIANGULAR, FLA_NO_TRANSPOSE, FLA_NONUNIT_DIAG, FLA_ONE, B11, A10, 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_il_blk_var2 | ( | 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_Obj_length(), FLA_ONE, FLA_ONE_HALF, 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_il().
{ 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 Y10_t, Y10_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( Y10, &Y10_t, &Y10_b, b, FLA_TOP ); // Y10 = 1/2 * B10 * A00; FLA_Hemm_internal( FLA_RIGHT, FLA_LOWER_TRIANGULAR, FLA_ONE_HALF, A00, B10, FLA_ZERO, Y10_t, FLA_Cntl_sub_hemm( cntl ) ); // A10 = A10 - Y10; FLA_Axpy_internal( FLA_MINUS_ONE, Y10_t, A10, FLA_Cntl_sub_axpy1( cntl ) ); // A11 = A11 - A10 * B10' - B10 * A10'; FLA_Her2k_internal( FLA_LOWER_TRIANGULAR, FLA_NO_TRANSPOSE, FLA_MINUS_ONE, A10, B10, FLA_ONE, A11, FLA_Cntl_sub_her2k( cntl ) ); // A11 = inv( tril( B11 ) ) * A11 * inv( tril( B11 )' ); FLA_Eig_gest_internal( FLA_INVERSE, FLA_LOWER_TRIANGULAR, A11, Y11, B11, FLA_Cntl_sub_eig_gest( cntl ) ); // A21 = A21 - A20 * B10'; FLA_Gemm_internal( FLA_NO_TRANSPOSE, FLA_CONJ_TRANSPOSE, FLA_MINUS_ONE, A20, B10, FLA_ONE, A21, FLA_Cntl_sub_gemm1( cntl ) ); // A21 = A21 * inv( tril( B11 )' ); FLA_Trsm_internal( FLA_RIGHT, FLA_LOWER_TRIANGULAR, FLA_CONJ_TRANSPOSE, FLA_NONUNIT_DIAG, FLA_ONE, B11, A21, FLA_Cntl_sub_trsm1( cntl ) ); // A10 = A10 - Y10; FLA_Axpy_internal( FLA_MINUS_ONE, Y10_t, A10, FLA_Cntl_sub_axpy2( cntl ) ); // A10 = inv( tril( B11 ) ) * A10; FLA_Trsm_internal( FLA_LEFT, FLA_LOWER_TRIANGULAR, FLA_NO_TRANSPOSE, FLA_NONUNIT_DIAG, FLA_ONE, B11, A10, 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_il_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_il().
{ 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 ); /*------------------------------------------------------------*/ // A10 = A10 - 1/2 * Y10; FLA_Axpy_internal( FLA_MINUS_ONE_HALF, Y10, A10, FLA_Cntl_sub_axpy1( cntl ) ); // A11 = A11 - A10 * B10' - B10 * A10'; FLA_Her2k_internal( FLA_LOWER_TRIANGULAR, FLA_NO_TRANSPOSE, FLA_MINUS_ONE, A10, B10, FLA_ONE, A11, FLA_Cntl_sub_her2k( cntl ) ); // A11 = inv( tril( B11 ) ) * A11 * inv( tril( B11 )' ); FLA_Eig_gest_internal( FLA_INVERSE, FLA_LOWER_TRIANGULAR, A11, Y11, B11, FLA_Cntl_sub_eig_gest( cntl ) ); // A21 = A21 - A20 * B10'; FLA_Gemm_internal( FLA_NO_TRANSPOSE, FLA_CONJ_TRANSPOSE, FLA_MINUS_ONE, A20, B10, FLA_ONE, A21, FLA_Cntl_sub_gemm1( cntl ) ); // A21 = A21 * inv( tril( B11 )' ); FLA_Trsm_internal( FLA_RIGHT, FLA_LOWER_TRIANGULAR, FLA_CONJ_TRANSPOSE, FLA_NONUNIT_DIAG, FLA_ONE, B11, A21, FLA_Cntl_sub_trsm1( cntl ) ); // A10 = A10 - 1/2 * Y10; FLA_Axpy_internal( FLA_MINUS_ONE_HALF, Y10, A10, FLA_Cntl_sub_axpy2( cntl ) ); // A10 = inv( tril( B11 ) ) * A10; FLA_Trsm_internal( FLA_LEFT, FLA_LOWER_TRIANGULAR, FLA_NO_TRANSPOSE, FLA_NONUNIT_DIAG, FLA_ONE, B11, A10, FLA_Cntl_sub_trsm2( cntl ) ); // Y20 = Y20 + B21 * A10; FLA_Gemm_internal( FLA_NO_TRANSPOSE, FLA_NO_TRANSPOSE, FLA_ONE, B21, A10, FLA_ONE, Y20, FLA_Cntl_sub_gemm2( cntl ) ); // Y21 = B21 * A11; FLA_Hemm_internal( FLA_RIGHT, FLA_LOWER_TRIANGULAR, FLA_ONE, A11, B21, FLA_ZERO, Y21, FLA_Cntl_sub_hemm( cntl ) ); // Y21 = Y21 + B20 * A10'; FLA_Gemm_internal( FLA_NO_TRANSPOSE, FLA_CONJ_TRANSPOSE, FLA_ONE, B20, A10, FLA_ONE, Y21, 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_il_blk_var4 | ( | 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_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_il().
{ 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 Y21_l, Y21_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( Y21, &Y21_l, &Y21_r, b, FLA_LEFT ); // A10 = inv( tril( B11 ) ) * A10; FLA_Trsm_internal( FLA_LEFT, FLA_LOWER_TRIANGULAR, FLA_NO_TRANSPOSE, FLA_NONUNIT_DIAG, FLA_ONE, B11, A10, FLA_Cntl_sub_trsm1( cntl ) ); // A20 = A20 - B21 * A10; FLA_Gemm_internal( FLA_NO_TRANSPOSE, FLA_NO_TRANSPOSE, FLA_MINUS_ONE, B21, A10, FLA_ONE, A20, FLA_Cntl_sub_gemm1( cntl ) ); // A11 = inv( tril( B11 ) ) * A11 * inv( tril( B11 )' ); FLA_Eig_gest_internal( FLA_INVERSE, FLA_LOWER_TRIANGULAR, A11, Y11, B11, FLA_Cntl_sub_eig_gest( cntl ) ); // Y21 = B21 * A11; FLA_Hemm_internal( FLA_RIGHT, FLA_LOWER_TRIANGULAR, FLA_ONE, A11, B21, FLA_ZERO, Y21_l, FLA_Cntl_sub_hemm( cntl ) ); // A21 = A21 * inv( tril( B11 )' ); FLA_Trsm_internal( FLA_RIGHT, FLA_LOWER_TRIANGULAR, FLA_CONJ_TRANSPOSE, FLA_NONUNIT_DIAG, FLA_ONE, B11, A21, FLA_Cntl_sub_trsm2( cntl ) ); // A21 = A21 - 1/2 * Y21; FLA_Axpy_internal( FLA_MINUS_ONE_HALF, Y21_l, A21, FLA_Cntl_sub_axpy1( cntl ) ); // A22 = A22 - A21 * B21' - B21 * A21'; FLA_Her2k_internal( FLA_LOWER_TRIANGULAR, FLA_NO_TRANSPOSE, FLA_MINUS_ONE, A21, B21, FLA_ONE, A22, FLA_Cntl_sub_her2k( cntl ) ); // A21 = A21 - 1/2 * Y21; FLA_Axpy_internal( FLA_MINUS_ONE_HALF, Y21_l, A21, 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_3x1_to_2x1( &YT, Y01, Y11, /* ** */ /* *** */ &YB, Y21, FLA_TOP ); } return FLA_SUCCESS; }
FLA_Error FLA_Eig_gest_il_blk_var5 | ( | 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_il().
{ 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 Y21_l, Y21_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( Y21, &Y21_l, &Y21_r, b, FLA_LEFT ); // A11 = inv( tril( B11 ) ) * A11 * inv( tril( B11 )' ); FLA_Eig_gest_internal( FLA_INVERSE, FLA_LOWER_TRIANGULAR, A11, Y11, B11, FLA_Cntl_sub_eig_gest( cntl ) ); // Y21 = B21 * A11; FLA_Hemm_internal( FLA_RIGHT, FLA_LOWER_TRIANGULAR, FLA_ONE, A11, B21, FLA_ZERO, Y21_l, FLA_Cntl_sub_hemm( cntl ) ); // A21 = A21 * inv( tril( B11 )' ); FLA_Trsm_internal( FLA_RIGHT, FLA_LOWER_TRIANGULAR, FLA_CONJ_TRANSPOSE, FLA_NONUNIT_DIAG, FLA_ONE, B11, A21, FLA_Cntl_sub_trsm1( cntl ) ); // A21 = A21 - 1/2 * Y21; FLA_Axpy_internal( FLA_MINUS_ONE_HALF, Y21_l, A21, FLA_Cntl_sub_axpy1( cntl ) ); // A22 = A22 - A21 * B21' - B21 * A21'; FLA_Her2k_internal( FLA_LOWER_TRIANGULAR, FLA_NO_TRANSPOSE, FLA_MINUS_ONE, A21, B21, FLA_ONE, A22, FLA_Cntl_sub_her2k( cntl ) ); // A21 = A21 - 1/2 * Y21; FLA_Axpy_internal( FLA_MINUS_ONE_HALF, Y21_l, A21, FLA_Cntl_sub_axpy2( cntl ) ); // A21 = inv( tril( B22 ) ) * A21; FLA_Trsm_internal( FLA_LEFT, FLA_LOWER_TRIANGULAR, FLA_NO_TRANSPOSE, FLA_NONUNIT_DIAG, FLA_ONE, B22, A21, 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_il_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 bli_caxpyv(), bli_cdot2s(), bli_chemv(), bli_cinvscalv(), bli_ctrsv(), FLA_MINUS_ONE, FLA_MINUS_ONE_HALF, FLA_ONE, and FLA_ZERO.
Referenced by FLA_Eig_gest_il_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* a10t = buff_A + (0 )*cs_A + (i )*rs_A; scomplex* alpha11 = buff_A + (i )*cs_A + (i )*rs_A; scomplex* y10t = buff_y + (0 )*inc_y; scomplex* B00 = buff_B + (0 )*cs_B + (0 )*rs_B; scomplex* b10t = buff_B + (0 )*cs_B + (i )*rs_B; scomplex* beta11 = buff_B + (i )*cs_B + (i )*rs_B; int m_behind = i; /*------------------------------------------------------------*/ // FLA_Hemvc_external( FLA_LOWER_TRIANGULAR, FLA_CONJUGATE, // FLA_ONE, A00, b10t, FLA_ZERO, y10t_t ); bli_chemv( BLIS_LOWER_TRIANGULAR, BLIS_CONJUGATE, m_behind, buff_1, A00, rs_A, cs_A, b10t, cs_B, buff_0, y10t, inc_y ); // FLA_Trsv_external( FLA_LOWER_TRIANGULAR, FLA_CONJ_NO_TRANSPOSE, FLA_NONUNIT_DIAG, // B00, a10t ); bli_ctrsv( BLIS_LOWER_TRIANGULAR, BLIS_CONJ_NO_TRANSPOSE, BLIS_NONUNIT_DIAG, m_behind, B00, rs_B, cs_B, a10t, cs_A ); // FLA_Axpy_external( FLA_MINUS_ONE_HALF, y10t_t, a10t ); bli_caxpyv( BLIS_NO_CONJUGATE, m_behind, buff_m1h, y10t, inc_y, a10t, cs_A ); // FLA_Dot2cs_external( FLA_CONJUGATE, FLA_MINUS_ONE, a10t, b10t, FLA_ONE, alpha11 ); bli_cdot2s( BLIS_CONJUGATE, m_behind, buff_m1, a10t, cs_A, b10t, cs_B, buff_1, alpha11 ); // FLA_Inv_scal_external( beta11, alpha11 ); // FLA_Inv_scal_external( beta11, alpha11 ); bli_cinvscals( beta11, alpha11 ); bli_cinvscals( beta11, alpha11 ); // FLA_Axpy_external( FLA_MINUS_ONE_HALF, y10t_t, a10t ); bli_caxpyv( BLIS_NO_CONJUGATE, m_behind, buff_m1h, y10t, inc_y, a10t, cs_A ); // FLA_Inv_scal_external( beta11, a10t ); bli_cinvscalv( BLIS_NO_CONJUGATE, m_behind, beta11, a10t, cs_A ); /*------------------------------------------------------------*/ } return FLA_SUCCESS; }
FLA_Error FLA_Eig_gest_il_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 bli_caxpyv(), bli_cdot2s(), bli_cgemv(), bli_chemv(), bli_cinvscalv(), FLA_MINUS_ONE, FLA_ONE, FLA_ONE_HALF, and FLA_ZERO.
Referenced by FLA_Eig_gest_il_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* a10t = buff_A + (0 )*cs_A + (i )*rs_A; scomplex* A20 = buff_A + (0 )*cs_A + (i+1)*rs_A; scomplex* alpha11 = buff_A + (i )*cs_A + (i )*rs_A; scomplex* a21 = buff_A + (i )*cs_A + (i+1)*rs_A; scomplex* y10t = buff_y + (0 )*inc_y; scomplex* b10t = buff_B + (0 )*cs_B + (i )*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_LOWER_TRIANGULAR, FLA_CONJUGATE, // FLA_ONE_HALF, A00, b10t, FLA_ZERO, y10t_t ); bli_chemv( BLIS_LOWER_TRIANGULAR, BLIS_CONJUGATE, m_behind, buff_1h, A00, rs_A, cs_A, b10t, cs_B, buff_0, y10t, inc_y ); // FLA_Axpy_external( FLA_MINUS_ONE, y10t_t, a10t ); bli_caxpyv( BLIS_NO_CONJUGATE, m_behind, buff_m1, y10t, inc_y, a10t, cs_A ); // FLA_Dot2cs_external( FLA_CONJUGATE, FLA_MINUS_ONE, a10t, b10t, FLA_ONE, alpha11 ); bli_cdot2s( BLIS_CONJUGATE, m_behind, buff_m1, a10t, cs_A, b10t, cs_B, buff_1, alpha11 ); // FLA_Inv_scal_external( beta11, alpha11 ); // FLA_Inv_scal_external( beta11, alpha11 ); bli_cinvscals( beta11, alpha11 ); bli_cinvscals( beta11, alpha11 ); // FLA_Gemvc_external( FLA_NO_TRANSPOSE, FLA_CONJUGATE, // FLA_MINUS_ONE, A20, b10t, FLA_ONE, a21 ); bli_cgemv( BLIS_NO_TRANSPOSE, BLIS_CONJUGATE, m_ahead, m_behind, buff_m1, A20, rs_A, cs_A, b10t, cs_B, buff_1, a21, rs_A ); // FLA_Inv_scal_external( beta11, a21 ); bli_cinvscalv( BLIS_NO_CONJUGATE, m_ahead, beta11, a21, rs_A ); // FLA_Axpy_external( FLA_MINUS_ONE, y10t_t, a10t ); bli_caxpyv( BLIS_NO_CONJUGATE, m_behind, buff_m1, y10t, inc_y, a10t, cs_A ); // FLA_Inv_scal_external( beta11, a10t ); bli_cinvscalv( BLIS_NO_CONJUGATE, m_behind, beta11, a10t, cs_A ); /*------------------------------------------------------------*/ } return FLA_SUCCESS; }
FLA_Error FLA_Eig_gest_il_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 bli_caxpyv(), bli_ccopyv(), bli_cdot2s(), bli_cgemv(), bli_cger(), bli_cinvscalv(), bli_cscalv(), FLA_MINUS_ONE, FLA_MINUS_ONE_HALF, and FLA_ONE.
Referenced by FLA_Eig_gest_il_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* a10t = buff_A + (0 )*cs_A + (i )*rs_A; scomplex* A20 = buff_A + (0 )*cs_A + (i+1)*rs_A; scomplex* alpha11 = buff_A + (i )*cs_A + (i )*rs_A; scomplex* a21 = buff_A + (i )*cs_A + (i+1)*rs_A; scomplex* y10t = buff_Y + (0 )*cs_Y + (i )*rs_Y; scomplex* Y20 = buff_Y + (0 )*cs_Y + (i+1)*rs_Y; scomplex* y21 = buff_Y + (i )*cs_Y + (i+1)*rs_Y; scomplex* b10t = buff_B + (0 )*cs_B + (i )*rs_B; scomplex* B20 = buff_B + (0 )*cs_B + (i+1)*rs_B; scomplex* beta11 = buff_B + (i )*cs_B + (i )*rs_B; scomplex* b21 = buff_B + (i )*cs_B + (i+1)*rs_B; int m_ahead = m_AB - i - 1; int m_behind = i; /*------------------------------------------------------------*/ // FLA_Axpy_external( FLA_MINUS_ONE_HALF, y10t, a10t ); bli_caxpyv( BLIS_NO_CONJUGATE, m_behind, buff_m1h, y10t, cs_Y, a10t, cs_A ); // FLA_Dot2cs_external( FLA_CONJUGATE, FLA_MINUS_ONE, a10t, b10t, FLA_ONE, alpha11 ); bli_cdot2s( BLIS_CONJUGATE, m_behind, buff_m1, a10t, cs_A, b10t, cs_B, buff_1, alpha11 ); // FLA_Inv_scal_external( beta11, alpha11 ); // FLA_Inv_scal_external( beta11, alpha11 ); bli_cinvscals( beta11, alpha11 ); bli_cinvscals( beta11, alpha11 ); // FLA_Gemvc_external( FLA_NO_TRANSPOSE, FLA_CONJUGATE, // FLA_MINUS_ONE, A20, b10t, FLA_ONE, a21 ); bli_cgemv( BLIS_NO_TRANSPOSE, BLIS_CONJUGATE, m_ahead, m_behind, buff_m1, A20, rs_A, cs_A, b10t, cs_B, buff_1, a21, rs_A ); // FLA_Inv_scal_external( beta11, a21 ); bli_cinvscalv( BLIS_NO_CONJUGATE, m_ahead, beta11, a21, rs_A ); // FLA_Axpy_external( FLA_MINUS_ONE_HALF, y10t, a10t ); bli_caxpyv( BLIS_NO_CONJUGATE, m_behind, buff_m1h, y10t, cs_Y, a10t, cs_A ); // FLA_Inv_scal_external( beta11, a10t ); bli_cinvscalv( BLIS_NO_CONJUGATE, m_behind, beta11, a10t, cs_A ); // FLA_Ger_external( FLA_ONE, b21, a10t, Y20 ); bli_cger( BLIS_NO_CONJUGATE, BLIS_NO_CONJUGATE, m_ahead, m_behind, buff_1, b21, rs_B, a10t, cs_A, Y20, rs_Y, cs_Y ); // FLA_Copy_external( b21, y21 ); // FLA_Scal_external( alpha11, y21 ); bli_ccopyv( BLIS_NO_CONJUGATE, m_ahead, b21, rs_B, y21, rs_Y ); bli_cscalv( BLIS_NO_CONJUGATE, m_ahead, alpha11, y21, rs_Y ); // FLA_Gemvc_external( FLA_NO_TRANSPOSE, FLA_CONJUGATE, // FLA_ONE, B20, a10t, FLA_ONE, y21 ); bli_cgemv( BLIS_NO_TRANSPOSE, BLIS_CONJUGATE, m_ahead, m_behind, buff_1, B20, rs_B, cs_B, a10t, cs_A, buff_1, y21, rs_Y ); /*------------------------------------------------------------*/ } return FLA_SUCCESS; }
FLA_Error FLA_Eig_gest_il_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 bli_caxpyv(), bli_cger(), bli_cher2(), bli_cinvscalv(), FLA_MINUS_ONE, and FLA_MINUS_ONE_HALF.
Referenced by FLA_Eig_gest_il_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* a10t = buff_A + (0 )*cs_A + (i )*rs_A; scomplex* A20 = buff_A + (0 )*cs_A + (i+1)*rs_A; scomplex* alpha11 = buff_A + (i )*cs_A + (i )*rs_A; scomplex* a21 = buff_A + (i )*cs_A + (i+1)*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* b21 = buff_B + (i )*cs_B + (i+1)*rs_B; scomplex psi11; int m_ahead = m_AB - i - 1; int m_behind = i; /*------------------------------------------------------------*/ // FLA_Inv_scal_external( beta11, a10t ); bli_cinvscalv( BLIS_NO_CONJUGATE, m_behind, beta11, a10t, cs_A ); // FLA_Ger_external( FLA_MINUS_ONE, b21, a10t, A20 ); bli_cger( BLIS_NO_CONJUGATE, BLIS_NO_CONJUGATE, m_ahead, m_behind, buff_m1, b21, rs_B, a10t, cs_A, A20, rs_A, cs_A ); // FLA_Inv_scal_external( beta11, alpha11 ); // FLA_Inv_scal_external( beta11, alpha11 ); bli_cinvscals( beta11, alpha11 ); bli_cinvscals( beta11, alpha11 ); // FLA_Copy_external( alpha11, psi11 ); // FLA_Scal_external( FLA_MINUS_ONE_HALF, psi11 ); bli_cmult3( buff_m1h, alpha11, &psi11 ); // FLA_Inv_scal_external( beta11, a21 ); bli_cinvscalv( BLIS_NO_CONJUGATE, m_ahead, beta11, a21, rs_A ); // FLA_Axpy_external( psi11, b21, a21 ); bli_caxpyv( BLIS_NO_CONJUGATE, m_ahead, &psi11, b21, rs_B, a21, rs_A ); // FLA_Her2c_external( FLA_LOWER_TRIANGULAR, FLA_NO_CONJUGATE, // FLA_MINUS_ONE, a21, b21, A22 ); bli_cher2( BLIS_LOWER_TRIANGULAR, BLIS_NO_CONJUGATE, m_ahead, buff_m1, a21, rs_A, b21, rs_B, A22, rs_A, cs_A ); // FLA_Axpy_external( psi11, b21, a21 ); bli_caxpyv( BLIS_NO_CONJUGATE, m_ahead, &psi11, b21, rs_B, a21, rs_A ); /*------------------------------------------------------------*/ } return FLA_SUCCESS; }
FLA_Error FLA_Eig_gest_il_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 bli_caxpyv(), bli_cher2(), bli_cinvscalv(), bli_ctrsv(), FLA_MINUS_ONE, and FLA_MINUS_ONE_HALF.
Referenced by FLA_Eig_gest_il_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* a21 = buff_A + (i )*cs_A + (i+1)*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* b21 = buff_B + (i )*cs_B + (i+1)*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 ); bli_cinvscals( beta11, alpha11 ); bli_cinvscals( beta11, alpha11 ); // FLA_Copy_external( alpha11, psi11 ); // FLA_Scal_external( FLA_MINUS_ONE_HALF, psi11 ); bli_cmult3( buff_m1h, alpha11, &psi11 ); // FLA_Inv_scal_external( beta11, a21 ); bli_cinvscalv( BLIS_NO_CONJUGATE, m_ahead, beta11, a21, rs_A ); // FLA_Axpy_external( psi11, b21, a21 ); bli_caxpyv( BLIS_NO_CONJUGATE, m_ahead, &psi11, b21, rs_B, a21, rs_A ); // FLA_Her2c_external( FLA_LOWER_TRIANGULAR, FLA_NO_CONJUGATE, // FLA_MINUS_ONE, a21, b21, A22 ); bli_cher2( BLIS_LOWER_TRIANGULAR, BLIS_NO_CONJUGATE, m_ahead, buff_m1, a21, rs_A, b21, rs_B, A22, rs_A, cs_A ); // FLA_Axpy_external( psi11, b21, a21 ); bli_caxpyv( BLIS_NO_CONJUGATE, m_ahead, &psi11, b21, rs_B, a21, rs_A ); // FLA_Trsv_external( FLA_LOWER_TRIANGULAR, FLA_NO_TRANSPOSE, FLA_NONUNIT_DIAG, // B22, a21 ); bli_ctrsv( BLIS_LOWER_TRIANGULAR, BLIS_NO_TRANSPOSE, BLIS_NONUNIT_DIAG, m_ahead, B22, rs_B, cs_B, a21, rs_A ); /*------------------------------------------------------------*/ } return FLA_SUCCESS; }
FLA_Error FLA_Eig_gest_il_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 bli_daxpyv(), bli_ddot2s(), bli_dhemv(), bli_dinvscalv(), bli_dtrsv(), FLA_MINUS_ONE, FLA_MINUS_ONE_HALF, FLA_ONE, and FLA_ZERO.
Referenced by FLA_Eig_gest_il_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* a10t = buff_A + (0 )*cs_A + (i )*rs_A; double* alpha11 = buff_A + (i )*cs_A + (i )*rs_A; double* y10t = buff_y + (0 )*inc_y; double* B00 = buff_B + (0 )*cs_B + (0 )*rs_B; double* b10t = buff_B + (0 )*cs_B + (i )*rs_B; double* beta11 = buff_B + (i )*cs_B + (i )*rs_B; int m_behind = i; /*------------------------------------------------------------*/ // FLA_Hemvc_external( FLA_LOWER_TRIANGULAR, FLA_CONJUGATE, // FLA_ONE, A00, b10t, FLA_ZERO, y10t_t ); bli_dhemv( BLIS_LOWER_TRIANGULAR, BLIS_CONJUGATE, m_behind, buff_1, A00, rs_A, cs_A, b10t, cs_B, buff_0, y10t, inc_y ); // FLA_Trsv_external( FLA_LOWER_TRIANGULAR, FLA_CONJ_NO_TRANSPOSE, FLA_NONUNIT_DIAG, // B00, a10t ); bli_dtrsv( BLIS_LOWER_TRIANGULAR, BLIS_CONJ_NO_TRANSPOSE, BLIS_NONUNIT_DIAG, m_behind, B00, rs_B, cs_B, a10t, cs_A ); // FLA_Axpy_external( FLA_MINUS_ONE_HALF, y10t_t, a10t ); bli_daxpyv( BLIS_NO_CONJUGATE, m_behind, buff_m1h, y10t, inc_y, a10t, cs_A ); // FLA_Dot2cs_external( FLA_CONJUGATE, FLA_MINUS_ONE, a10t, b10t, FLA_ONE, alpha11 ); bli_ddot2s( BLIS_CONJUGATE, m_behind, buff_m1, a10t, cs_A, b10t, cs_B, buff_1, alpha11 ); // FLA_Inv_scal_external( beta11, alpha11 ); // FLA_Inv_scal_external( beta11, alpha11 ); bli_dinvscals( beta11, alpha11 ); bli_dinvscals( beta11, alpha11 ); // FLA_Axpy_external( FLA_MINUS_ONE_HALF, y10t_t, a10t ); bli_daxpyv( BLIS_NO_CONJUGATE, m_behind, buff_m1h, y10t, inc_y, a10t, cs_A ); // FLA_Inv_scal_external( beta11, a10t ); bli_dinvscalv( BLIS_NO_CONJUGATE, m_behind, beta11, a10t, cs_A ); /*------------------------------------------------------------*/ } return FLA_SUCCESS; }
FLA_Error FLA_Eig_gest_il_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 bli_daxpyv(), bli_ddot2s(), bli_dgemv(), bli_dhemv(), bli_dinvscalv(), FLA_MINUS_ONE, FLA_ONE, FLA_ONE_HALF, and FLA_ZERO.
Referenced by FLA_Eig_gest_il_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* a10t = buff_A + (0 )*cs_A + (i )*rs_A; double* A20 = buff_A + (0 )*cs_A + (i+1)*rs_A; double* alpha11 = buff_A + (i )*cs_A + (i )*rs_A; double* a21 = buff_A + (i )*cs_A + (i+1)*rs_A; double* y10t = buff_y + (0 )*inc_y; double* b10t = buff_B + (0 )*cs_B + (i )*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_LOWER_TRIANGULAR, FLA_CONJUGATE, // FLA_ONE_HALF, A00, b10t, FLA_ZERO, y10t_t ); bli_dhemv( BLIS_LOWER_TRIANGULAR, BLIS_CONJUGATE, m_behind, buff_1h, A00, rs_A, cs_A, b10t, cs_B, buff_0, y10t, inc_y ); // FLA_Axpy_external( FLA_MINUS_ONE, y10t_t, a10t ); bli_daxpyv( BLIS_NO_CONJUGATE, m_behind, buff_m1, y10t, inc_y, a10t, cs_A ); // FLA_Dot2cs_external( FLA_CONJUGATE, FLA_MINUS_ONE, a10t, b10t, FLA_ONE, alpha11 ); bli_ddot2s( BLIS_CONJUGATE, m_behind, buff_m1, a10t, cs_A, b10t, cs_B, buff_1, alpha11 ); // FLA_Inv_scal_external( beta11, alpha11 ); // FLA_Inv_scal_external( beta11, alpha11 ); bli_dinvscals( beta11, alpha11 ); bli_dinvscals( beta11, alpha11 ); // FLA_Gemvc_external( FLA_NO_TRANSPOSE, FLA_CONJUGATE, // FLA_MINUS_ONE, A20, b10t, FLA_ONE, a21 ); bli_dgemv( BLIS_NO_TRANSPOSE, BLIS_CONJUGATE, m_ahead, m_behind, buff_m1, A20, rs_A, cs_A, b10t, cs_B, buff_1, a21, rs_A ); // FLA_Inv_scal_external( beta11, a21 ); bli_dinvscalv( BLIS_NO_CONJUGATE, m_ahead, beta11, a21, rs_A ); // FLA_Axpy_external( FLA_MINUS_ONE, y10t_t, a10t ); bli_daxpyv( BLIS_NO_CONJUGATE, m_behind, buff_m1, y10t, inc_y, a10t, cs_A ); // FLA_Inv_scal_external( beta11, a10t ); bli_dinvscalv( BLIS_NO_CONJUGATE, m_behind, beta11, a10t, cs_A ); /*------------------------------------------------------------*/ } return FLA_SUCCESS; }
FLA_Error FLA_Eig_gest_il_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 bli_daxpyv(), bli_dcopyv(), bli_ddot2s(), bli_dgemv(), bli_dger(), bli_dinvscalv(), bli_dscalv(), FLA_MINUS_ONE, FLA_MINUS_ONE_HALF, and FLA_ONE.
Referenced by FLA_Eig_gest_il_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* a10t = buff_A + (0 )*cs_A + (i )*rs_A; double* A20 = buff_A + (0 )*cs_A + (i+1)*rs_A; double* alpha11 = buff_A + (i )*cs_A + (i )*rs_A; double* a21 = buff_A + (i )*cs_A + (i+1)*rs_A; double* y10t = buff_Y + (0 )*cs_Y + (i )*rs_Y; double* Y20 = buff_Y + (0 )*cs_Y + (i+1)*rs_Y; double* y21 = buff_Y + (i )*cs_Y + (i+1)*rs_Y; double* b10t = buff_B + (0 )*cs_B + (i )*rs_B; double* B20 = buff_B + (0 )*cs_B + (i+1)*rs_B; double* beta11 = buff_B + (i )*cs_B + (i )*rs_B; double* b21 = buff_B + (i )*cs_B + (i+1)*rs_B; int m_ahead = m_AB - i - 1; int m_behind = i; /*------------------------------------------------------------*/ // FLA_Axpy_external( FLA_MINUS_ONE_HALF, y10t, a10t ); bli_daxpyv( BLIS_NO_CONJUGATE, m_behind, buff_m1h, y10t, cs_Y, a10t, cs_A ); // FLA_Dot2cs_external( FLA_CONJUGATE, FLA_MINUS_ONE, a10t, b10t, FLA_ONE, alpha11 ); bli_ddot2s( BLIS_CONJUGATE, m_behind, buff_m1, a10t, cs_A, b10t, cs_B, buff_1, alpha11 ); // FLA_Inv_scal_external( beta11, alpha11 ); // FLA_Inv_scal_external( beta11, alpha11 ); bli_dinvscals( beta11, alpha11 ); bli_dinvscals( beta11, alpha11 ); // FLA_Gemvc_external( FLA_NO_TRANSPOSE, FLA_CONJUGATE, // FLA_MINUS_ONE, A20, b10t, FLA_ONE, a21 ); bli_dgemv( BLIS_NO_TRANSPOSE, BLIS_CONJUGATE, m_ahead, m_behind, buff_m1, A20, rs_A, cs_A, b10t, cs_B, buff_1, a21, rs_A ); // FLA_Inv_scal_external( beta11, a21 ); bli_dinvscalv( BLIS_NO_CONJUGATE, m_ahead, beta11, a21, rs_A ); // FLA_Axpy_external( FLA_MINUS_ONE_HALF, y10t, a10t ); bli_daxpyv( BLIS_NO_CONJUGATE, m_behind, buff_m1h, y10t, cs_Y, a10t, cs_A ); // FLA_Inv_scal_external( beta11, a10t ); bli_dinvscalv( BLIS_NO_CONJUGATE, m_behind, beta11, a10t, cs_A ); // FLA_Ger_external( FLA_ONE, b21, a10t, Y20 ); bli_dger( BLIS_NO_CONJUGATE, BLIS_NO_CONJUGATE, m_ahead, m_behind, buff_1, b21, rs_B, a10t, cs_A, Y20, rs_Y, cs_Y ); // FLA_Copy_external( b21, y21 ); // FLA_Scal_external( alpha11, y21 ); bli_dcopyv( BLIS_NO_CONJUGATE, m_ahead, b21, rs_B, y21, rs_Y ); bli_dscalv( BLIS_NO_CONJUGATE, m_ahead, alpha11, y21, rs_Y ); // FLA_Gemvc_external( FLA_NO_TRANSPOSE, FLA_CONJUGATE, // FLA_ONE, B20, a10t, FLA_ONE, y21 ); bli_dgemv( BLIS_NO_TRANSPOSE, BLIS_CONJUGATE, m_ahead, m_behind, buff_1, B20, rs_B, cs_B, a10t, cs_A, buff_1, y21, rs_Y ); /*------------------------------------------------------------*/ } return FLA_SUCCESS; }
FLA_Error FLA_Eig_gest_il_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 bli_daxpyv(), bli_dger(), bli_dher2(), bli_dinvscalv(), FLA_MINUS_ONE, and FLA_MINUS_ONE_HALF.
Referenced by FLA_Eig_gest_il_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* a10t = buff_A + (0 )*cs_A + (i )*rs_A; double* A20 = buff_A + (0 )*cs_A + (i+1)*rs_A; double* alpha11 = buff_A + (i )*cs_A + (i )*rs_A; double* a21 = buff_A + (i )*cs_A + (i+1)*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* b21 = buff_B + (i )*cs_B + (i+1)*rs_B; double psi11; int m_ahead = m_AB - i - 1; int m_behind = i; /*------------------------------------------------------------*/ // FLA_Inv_scal_external( beta11, a10t ); bli_dinvscalv( BLIS_NO_CONJUGATE, m_behind, beta11, a10t, cs_A ); // FLA_Ger_external( FLA_MINUS_ONE, b21, a10t, A20 ); bli_dger( BLIS_NO_CONJUGATE, BLIS_NO_CONJUGATE, m_ahead, m_behind, buff_m1, b21, rs_B, a10t, cs_A, A20, rs_A, cs_A ); // FLA_Inv_scal_external( beta11, alpha11 ); // FLA_Inv_scal_external( beta11, alpha11 ); bli_dinvscals( beta11, alpha11 ); bli_dinvscals( beta11, alpha11 ); // FLA_Copy_external( alpha11, psi11 ); // FLA_Scal_external( FLA_MINUS_ONE_HALF, psi11 ); bli_dmult3( buff_m1h, alpha11, &psi11 ); // FLA_Inv_scal_external( beta11, a21 ); bli_dinvscalv( BLIS_NO_CONJUGATE, m_ahead, beta11, a21, rs_A ); // FLA_Axpy_external( psi11, b21, a21 ); bli_daxpyv( BLIS_NO_CONJUGATE, m_ahead, &psi11, b21, rs_B, a21, rs_A ); // FLA_Her2c_external( FLA_LOWER_TRIANGULAR, FLA_NO_CONJUGATE, // FLA_MINUS_ONE, a21, b21, A22 ); bli_dher2( BLIS_LOWER_TRIANGULAR, BLIS_NO_CONJUGATE, m_ahead, buff_m1, a21, rs_A, b21, rs_B, A22, rs_A, cs_A ); // FLA_Axpy_external( psi11, b21, a21 ); bli_daxpyv( BLIS_NO_CONJUGATE, m_ahead, &psi11, b21, rs_B, a21, rs_A ); /*------------------------------------------------------------*/ } return FLA_SUCCESS; }
FLA_Error FLA_Eig_gest_il_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 bli_daxpyv(), bli_dher2(), bli_dinvscalv(), bli_dtrsv(), FLA_MINUS_ONE, and FLA_MINUS_ONE_HALF.
Referenced by FLA_Eig_gest_il_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* a21 = buff_A + (i )*cs_A + (i+1)*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* b21 = buff_B + (i )*cs_B + (i+1)*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 ); bli_dinvscals( beta11, alpha11 ); bli_dinvscals( beta11, alpha11 ); // FLA_Copy_external( alpha11, psi11 ); // FLA_Scal_external( FLA_MINUS_ONE_HALF, psi11 ); bli_dmult3( buff_m1h, alpha11, &psi11 ); // FLA_Inv_scal_external( beta11, a21 ); bli_dinvscalv( BLIS_NO_CONJUGATE, m_ahead, beta11, a21, rs_A ); // FLA_Axpy_external( psi11, b21, a21 ); bli_daxpyv( BLIS_NO_CONJUGATE, m_ahead, &psi11, b21, rs_B, a21, rs_A ); // FLA_Her2c_external( FLA_LOWER_TRIANGULAR, FLA_NO_CONJUGATE, // FLA_MINUS_ONE, a21, b21, A22 ); bli_dher2( BLIS_LOWER_TRIANGULAR, BLIS_NO_CONJUGATE, m_ahead, buff_m1, a21, rs_A, b21, rs_B, A22, rs_A, cs_A ); // FLA_Axpy_external( psi11, b21, a21 ); bli_daxpyv( BLIS_NO_CONJUGATE, m_ahead, &psi11, b21, rs_B, a21, rs_A ); // FLA_Trsv_external( FLA_LOWER_TRIANGULAR, FLA_NO_TRANSPOSE, FLA_NONUNIT_DIAG, // B22, a21 ); bli_dtrsv( BLIS_LOWER_TRIANGULAR, BLIS_NO_TRANSPOSE, BLIS_NONUNIT_DIAG, m_ahead, B22, rs_B, cs_B, a21, rs_A ); /*------------------------------------------------------------*/ } return FLA_SUCCESS; }
FLA_Error FLA_Eig_gest_il_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 bli_saxpyv(), bli_sdot2s(), bli_shemv(), bli_sinvscalv(), bli_strsv(), FLA_MINUS_ONE, FLA_MINUS_ONE_HALF, FLA_ONE, and FLA_ZERO.
Referenced by FLA_Eig_gest_il_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* a10t = buff_A + (0 )*cs_A + (i )*rs_A; float* alpha11 = buff_A + (i )*cs_A + (i )*rs_A; float* y10t = buff_y + (0 )*inc_y; float* B00 = buff_B + (0 )*cs_B + (0 )*rs_B; float* b10t = buff_B + (0 )*cs_B + (i )*rs_B; float* beta11 = buff_B + (i )*cs_B + (i )*rs_B; int m_behind = i; /*------------------------------------------------------------*/ // FLA_Hemvc_external( FLA_LOWER_TRIANGULAR, FLA_CONJUGATE, // FLA_ONE, A00, b10t, FLA_ZERO, y10t_t ); bli_shemv( BLIS_LOWER_TRIANGULAR, BLIS_CONJUGATE, m_behind, buff_1, A00, rs_A, cs_A, b10t, cs_B, buff_0, y10t, inc_y ); // FLA_Trsv_external( FLA_LOWER_TRIANGULAR, FLA_CONJ_NO_TRANSPOSE, FLA_NONUNIT_DIAG, // B00, a10t ); bli_strsv( BLIS_LOWER_TRIANGULAR, BLIS_CONJ_NO_TRANSPOSE, BLIS_NONUNIT_DIAG, m_behind, B00, rs_B, cs_B, a10t, cs_A ); // FLA_Axpy_external( FLA_MINUS_ONE_HALF, y10t_t, a10t ); bli_saxpyv( BLIS_NO_CONJUGATE, m_behind, buff_m1h, y10t, inc_y, a10t, cs_A ); // FLA_Dot2cs_external( FLA_CONJUGATE, FLA_MINUS_ONE, a10t, b10t, FLA_ONE, alpha11 ); bli_sdot2s( BLIS_CONJUGATE, m_behind, buff_m1, a10t, cs_A, b10t, cs_B, buff_1, alpha11 ); // FLA_Inv_scal_external( beta11, alpha11 ); // FLA_Inv_scal_external( beta11, alpha11 ); bli_sinvscals( beta11, alpha11 ); bli_sinvscals( beta11, alpha11 ); // FLA_Axpy_external( FLA_MINUS_ONE_HALF, y10t_t, a10t ); bli_saxpyv( BLIS_NO_CONJUGATE, m_behind, buff_m1h, y10t, inc_y, a10t, cs_A ); // FLA_Inv_scal_external( beta11, a10t ); bli_sinvscalv( BLIS_NO_CONJUGATE, m_behind, beta11, a10t, cs_A ); /*------------------------------------------------------------*/ } return FLA_SUCCESS; }
FLA_Error FLA_Eig_gest_il_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 bli_saxpyv(), bli_sdot2s(), bli_sgemv(), bli_shemv(), bli_sinvscalv(), FLA_MINUS_ONE, FLA_ONE, FLA_ONE_HALF, and FLA_ZERO.
Referenced by FLA_Eig_gest_il_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* a10t = buff_A + (0 )*cs_A + (i )*rs_A; float* A20 = buff_A + (0 )*cs_A + (i+1)*rs_A; float* alpha11 = buff_A + (i )*cs_A + (i )*rs_A; float* a21 = buff_A + (i )*cs_A + (i+1)*rs_A; float* y10t = buff_y + (0 )*inc_y; float* b10t = buff_B + (0 )*cs_B + (i )*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_LOWER_TRIANGULAR, FLA_CONJUGATE, // FLA_ONE_HALF, A00, b10t, FLA_ZERO, y10t_t ); bli_shemv( BLIS_LOWER_TRIANGULAR, BLIS_CONJUGATE, m_behind, buff_1h, A00, rs_A, cs_A, b10t, cs_B, buff_0, y10t, inc_y ); // FLA_Axpy_external( FLA_MINUS_ONE, y10t_t, a10t ); bli_saxpyv( BLIS_NO_CONJUGATE, m_behind, buff_m1, y10t, inc_y, a10t, cs_A ); // FLA_Dot2cs_external( FLA_CONJUGATE, FLA_MINUS_ONE, a10t, b10t, FLA_ONE, alpha11 ); bli_sdot2s( BLIS_CONJUGATE, m_behind, buff_m1, a10t, cs_A, b10t, cs_B, buff_1, alpha11 ); // FLA_Inv_scal_external( beta11, alpha11 ); // FLA_Inv_scal_external( beta11, alpha11 ); bli_sinvscals( beta11, alpha11 ); bli_sinvscals( beta11, alpha11 ); // FLA_Gemvc_external( FLA_NO_TRANSPOSE, FLA_CONJUGATE, // FLA_MINUS_ONE, A20, b10t, FLA_ONE, a21 ); bli_sgemv( BLIS_NO_TRANSPOSE, BLIS_CONJUGATE, m_ahead, m_behind, buff_m1, A20, rs_A, cs_A, b10t, cs_B, buff_1, a21, rs_A ); // FLA_Inv_scal_external( beta11, a21 ); bli_sinvscalv( BLIS_NO_CONJUGATE, m_ahead, beta11, a21, rs_A ); // FLA_Axpy_external( FLA_MINUS_ONE, y10t_t, a10t ); bli_saxpyv( BLIS_NO_CONJUGATE, m_behind, buff_m1, y10t, inc_y, a10t, cs_A ); // FLA_Inv_scal_external( beta11, a10t ); bli_sinvscalv( BLIS_NO_CONJUGATE, m_behind, beta11, a10t, cs_A ); /*------------------------------------------------------------*/ } return FLA_SUCCESS; }
FLA_Error FLA_Eig_gest_il_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 bli_saxpyv(), bli_scopyv(), bli_sdot2s(), bli_sgemv(), bli_sger(), bli_sinvscalv(), bli_sscalv(), FLA_MINUS_ONE, FLA_MINUS_ONE_HALF, and FLA_ONE.
Referenced by FLA_Eig_gest_il_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* a10t = buff_A + (0 )*cs_A + (i )*rs_A; float* A20 = buff_A + (0 )*cs_A + (i+1)*rs_A; float* alpha11 = buff_A + (i )*cs_A + (i )*rs_A; float* a21 = buff_A + (i )*cs_A + (i+1)*rs_A; float* y10t = buff_Y + (0 )*cs_Y + (i )*rs_Y; float* Y20 = buff_Y + (0 )*cs_Y + (i+1)*rs_Y; float* y21 = buff_Y + (i )*cs_Y + (i+1)*rs_Y; float* b10t = buff_B + (0 )*cs_B + (i )*rs_B; float* B20 = buff_B + (0 )*cs_B + (i+1)*rs_B; float* beta11 = buff_B + (i )*cs_B + (i )*rs_B; float* b21 = buff_B + (i )*cs_B + (i+1)*rs_B; int m_ahead = m_AB - i - 1; int m_behind = i; /*------------------------------------------------------------*/ // FLA_Axpy_external( FLA_MINUS_ONE_HALF, y10t, a10t ); bli_saxpyv( BLIS_NO_CONJUGATE, m_behind, buff_m1h, y10t, cs_Y, a10t, cs_A ); // FLA_Dot2cs_external( FLA_CONJUGATE, FLA_MINUS_ONE, a10t, b10t, FLA_ONE, alpha11 ); bli_sdot2s( BLIS_CONJUGATE, m_behind, buff_m1, a10t, cs_A, b10t, cs_B, buff_1, alpha11 ); // FLA_Inv_scal_external( beta11, alpha11 ); // FLA_Inv_scal_external( beta11, alpha11 ); bli_sinvscals( beta11, alpha11 ); bli_sinvscals( beta11, alpha11 ); // FLA_Gemvc_external( FLA_NO_TRANSPOSE, FLA_CONJUGATE, // FLA_MINUS_ONE, A20, b10t, FLA_ONE, a21 ); bli_sgemv( BLIS_NO_TRANSPOSE, BLIS_CONJUGATE, m_ahead, m_behind, buff_m1, A20, rs_A, cs_A, b10t, cs_B, buff_1, a21, rs_A ); // FLA_Inv_scal_external( beta11, a21 ); bli_sinvscalv( BLIS_NO_CONJUGATE, m_ahead, beta11, a21, rs_A ); // FLA_Axpy_external( FLA_MINUS_ONE_HALF, y10t, a10t ); bli_saxpyv( BLIS_NO_CONJUGATE, m_behind, buff_m1h, y10t, cs_Y, a10t, cs_A ); // FLA_Inv_scal_external( beta11, a10t ); bli_sinvscalv( BLIS_NO_CONJUGATE, m_behind, beta11, a10t, cs_A ); // FLA_Ger_external( FLA_ONE, b21, a10t, Y20 ); bli_sger( BLIS_NO_CONJUGATE, BLIS_NO_CONJUGATE, m_ahead, m_behind, buff_1, b21, rs_B, a10t, cs_A, Y20, rs_Y, cs_Y ); // FLA_Copy_external( b21, y21 ); // FLA_Scal_external( alpha11, y21 ); bli_scopyv( BLIS_NO_CONJUGATE, m_ahead, b21, rs_B, y21, rs_Y ); bli_sscalv( BLIS_NO_CONJUGATE, m_ahead, alpha11, y21, rs_Y ); // FLA_Gemvc_external( FLA_NO_TRANSPOSE, FLA_CONJUGATE, // FLA_ONE, B20, a10t, FLA_ONE, y21 ); bli_sgemv( BLIS_NO_TRANSPOSE, BLIS_CONJUGATE, m_ahead, m_behind, buff_1, B20, rs_B, cs_B, a10t, cs_A, buff_1, y21, rs_Y ); /*------------------------------------------------------------*/ } return FLA_SUCCESS; }
FLA_Error FLA_Eig_gest_il_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 bli_saxpyv(), bli_sger(), bli_sher2(), bli_sinvscalv(), FLA_MINUS_ONE, and FLA_MINUS_ONE_HALF.
Referenced by FLA_Eig_gest_il_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* a10t = buff_A + (0 )*cs_A + (i )*rs_A; float* A20 = buff_A + (0 )*cs_A + (i+1)*rs_A; float* alpha11 = buff_A + (i )*cs_A + (i )*rs_A; float* a21 = buff_A + (i )*cs_A + (i+1)*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* b21 = buff_B + (i )*cs_B + (i+1)*rs_B; float psi11; int m_ahead = m_AB - i - 1; int m_behind = i; /*------------------------------------------------------------*/ // FLA_Inv_scal_external( beta11, a10t ); bli_sinvscalv( BLIS_NO_CONJUGATE, m_behind, beta11, a10t, cs_A ); // FLA_Ger_external( FLA_MINUS_ONE, b21, a10t, A20 ); bli_sger( BLIS_NO_CONJUGATE, BLIS_NO_CONJUGATE, m_ahead, m_behind, buff_m1, b21, rs_B, a10t, cs_A, A20, rs_A, cs_A ); // FLA_Inv_scal_external( beta11, alpha11 ); // FLA_Inv_scal_external( beta11, alpha11 ); bli_sinvscals( beta11, alpha11 ); bli_sinvscals( beta11, alpha11 ); // FLA_Copy_external( alpha11, psi11 ); // FLA_Scal_external( FLA_MINUS_ONE_HALF, psi11 ); bli_smult3( buff_m1h, alpha11, &psi11 ); // FLA_Inv_scal_external( beta11, a21 ); bli_sinvscalv( BLIS_NO_CONJUGATE, m_ahead, beta11, a21, rs_A ); // FLA_Axpy_external( psi11, b21, a21 ); bli_saxpyv( BLIS_NO_CONJUGATE, m_ahead, &psi11, b21, rs_B, a21, rs_A ); // FLA_Her2c_external( FLA_LOWER_TRIANGULAR, FLA_NO_CONJUGATE, // FLA_MINUS_ONE, a21, b21, A22 ); bli_sher2( BLIS_LOWER_TRIANGULAR, BLIS_NO_CONJUGATE, m_ahead, buff_m1, a21, rs_A, b21, rs_B, A22, rs_A, cs_A ); // FLA_Axpy_external( psi11, b21, a21 ); bli_saxpyv( BLIS_NO_CONJUGATE, m_ahead, &psi11, b21, rs_B, a21, rs_A ); /*------------------------------------------------------------*/ } return FLA_SUCCESS; }
FLA_Error FLA_Eig_gest_il_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 bli_saxpyv(), bli_sher2(), bli_sinvscalv(), bli_strsv(), FLA_MINUS_ONE, and FLA_MINUS_ONE_HALF.
Referenced by FLA_Eig_gest_il_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* a21 = buff_A + (i )*cs_A + (i+1)*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* b21 = buff_B + (i )*cs_B + (i+1)*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 ); bli_sinvscals( beta11, alpha11 ); bli_sinvscals( beta11, alpha11 ); // FLA_Copy_external( alpha11, psi11 ); // FLA_Scal_external( FLA_MINUS_ONE_HALF, psi11 ); bli_smult3( buff_m1h, alpha11, &psi11 ); // FLA_Inv_scal_external( beta11, a21 ); bli_sinvscalv( BLIS_NO_CONJUGATE, m_ahead, beta11, a21, rs_A ); // FLA_Axpy_external( psi11, b21, a21 ); bli_saxpyv( BLIS_NO_CONJUGATE, m_ahead, &psi11, b21, rs_B, a21, rs_A ); // FLA_Her2c_external( FLA_LOWER_TRIANGULAR, FLA_NO_CONJUGATE, // FLA_MINUS_ONE, a21, b21, A22 ); bli_sher2( BLIS_LOWER_TRIANGULAR, BLIS_NO_CONJUGATE, m_ahead, buff_m1, a21, rs_A, b21, rs_B, A22, rs_A, cs_A ); // FLA_Axpy_external( psi11, b21, a21 ); bli_saxpyv( BLIS_NO_CONJUGATE, m_ahead, &psi11, b21, rs_B, a21, rs_A ); // FLA_Trsv_external( FLA_LOWER_TRIANGULAR, FLA_NO_TRANSPOSE, FLA_NONUNIT_DIAG, // B22, a21 ); bli_strsv( BLIS_LOWER_TRIANGULAR, BLIS_NO_TRANSPOSE, BLIS_NONUNIT_DIAG, m_ahead, B22, rs_B, cs_B, a21, rs_A ); /*------------------------------------------------------------*/ } return FLA_SUCCESS; }
References FLA_Eig_gest_il_opc_var1(), FLA_Eig_gest_il_opd_var1(), FLA_Eig_gest_il_ops_var1(), FLA_Eig_gest_il_opz_var1(), FLA_Obj_col_stride(), FLA_Obj_datatype(), FLA_Obj_length(), FLA_Obj_row_stride(), FLA_Obj_vector_inc(), and FLA_Part_2x1().
Referenced by FLA_Eig_gest_il().
{ FLA_Datatype datatype; int m_AB; int rs_A, cs_A; int rs_B, cs_B; int inc_y; FLA_Obj yT, yB; 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_2x1( Y, &yT, &yB, 1, FLA_TOP ); inc_y = FLA_Obj_vector_inc( yT ); switch ( datatype ) { case FLA_FLOAT: { float* buff_A = FLA_FLOAT_PTR( A ); float* buff_y = FLA_FLOAT_PTR( yT ); float* buff_B = FLA_FLOAT_PTR( B ); FLA_Eig_gest_il_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( yT ); double* buff_B = FLA_DOUBLE_PTR( B ); FLA_Eig_gest_il_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( yT ); scomplex* buff_B = FLA_COMPLEX_PTR( B ); FLA_Eig_gest_il_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( yT ); dcomplex* buff_B = FLA_DOUBLE_COMPLEX_PTR( B ); FLA_Eig_gest_il_opz_var1( m_AB, buff_A, rs_A, cs_A, buff_y, inc_y, buff_B, rs_B, cs_B ); break; } } return FLA_SUCCESS; }
References FLA_Eig_gest_il_opc_var2(), FLA_Eig_gest_il_opd_var2(), FLA_Eig_gest_il_ops_var2(), FLA_Eig_gest_il_opz_var2(), FLA_Obj_col_stride(), FLA_Obj_datatype(), FLA_Obj_length(), FLA_Obj_row_stride(), FLA_Obj_vector_inc(), and FLA_Part_2x1().
Referenced by FLA_Eig_gest_il().
{ FLA_Datatype datatype; int m_AB; int rs_A, cs_A; int rs_B, cs_B; int inc_y; FLA_Obj yT, yB; 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_2x1( Y, &yT, &yB, 1, FLA_TOP ); inc_y = FLA_Obj_vector_inc( yT ); switch ( datatype ) { case FLA_FLOAT: { float* buff_A = FLA_FLOAT_PTR( A ); float* buff_y = FLA_FLOAT_PTR( yT ); float* buff_B = FLA_FLOAT_PTR( B ); FLA_Eig_gest_il_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( yT ); double* buff_B = FLA_DOUBLE_PTR( B ); FLA_Eig_gest_il_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( yT ); scomplex* buff_B = FLA_COMPLEX_PTR( B ); FLA_Eig_gest_il_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( yT ); dcomplex* buff_B = FLA_DOUBLE_COMPLEX_PTR( B ); FLA_Eig_gest_il_opz_var2( m_AB, buff_A, rs_A, cs_A, buff_y, inc_y, buff_B, rs_B, cs_B ); break; } } return FLA_SUCCESS; }
References FLA_Eig_gest_il_opc_var3(), FLA_Eig_gest_il_opd_var3(), FLA_Eig_gest_il_ops_var3(), FLA_Eig_gest_il_opz_var3(), FLA_Obj_col_stride(), FLA_Obj_datatype(), FLA_Obj_length(), and FLA_Obj_row_stride().
Referenced by FLA_Eig_gest_il().
{ FLA_Datatype datatype; int m_AB; int rs_A, cs_A; int rs_Y, cs_Y; int rs_B, cs_B; 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_Y = FLA_Obj_row_stride( Y ); cs_Y = FLA_Obj_col_stride( Y ); rs_B = FLA_Obj_row_stride( B ); cs_B = FLA_Obj_col_stride( B ); 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_il_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_il_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_il_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_il_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; }
References FLA_Eig_gest_il_opc_var4(), FLA_Eig_gest_il_opd_var4(), FLA_Eig_gest_il_ops_var4(), FLA_Eig_gest_il_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_il().
{ 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_il_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_il_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_il_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_il_opz_var4( m_AB, buff_A, rs_A, cs_A, buff_y, inc_y, buff_B, rs_B, cs_B ); break; } } return FLA_SUCCESS; }
References FLA_Eig_gest_il_opc_var5(), FLA_Eig_gest_il_opd_var5(), FLA_Eig_gest_il_ops_var5(), FLA_Eig_gest_il_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_il().
{ 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_il_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_il_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_il_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_il_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_il_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 bli_zaxpyv(), bli_zdot2s(), bli_zhemv(), bli_zinvscalv(), bli_ztrsv(), FLA_MINUS_ONE, FLA_MINUS_ONE_HALF, FLA_ONE, and FLA_ZERO.
Referenced by FLA_Eig_gest_il_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* a10t = buff_A + (0 )*cs_A + (i )*rs_A; dcomplex* alpha11 = buff_A + (i )*cs_A + (i )*rs_A; dcomplex* y10t = buff_y + (0 )*inc_y; dcomplex* B00 = buff_B + (0 )*cs_B + (0 )*rs_B; dcomplex* b10t = buff_B + (0 )*cs_B + (i )*rs_B; dcomplex* beta11 = buff_B + (i )*cs_B + (i )*rs_B; int m_behind = i; /*------------------------------------------------------------*/ // FLA_Hemvc_external( FLA_LOWER_TRIANGULAR, FLA_CONJUGATE, // FLA_ONE, A00, b10t, FLA_ZERO, y10t_t ); bli_zhemv( BLIS_LOWER_TRIANGULAR, BLIS_CONJUGATE, m_behind, buff_1, A00, rs_A, cs_A, b10t, cs_B, buff_0, y10t, inc_y ); // FLA_Trsv_external( FLA_LOWER_TRIANGULAR, FLA_CONJ_NO_TRANSPOSE, FLA_NONUNIT_DIAG, // B00, a10t ); bli_ztrsv( BLIS_LOWER_TRIANGULAR, BLIS_CONJ_NO_TRANSPOSE, BLIS_NONUNIT_DIAG, m_behind, B00, rs_B, cs_B, a10t, cs_A ); // FLA_Axpy_external( FLA_MINUS_ONE_HALF, y10t_t, a10t ); bli_zaxpyv( BLIS_NO_CONJUGATE, m_behind, buff_m1h, y10t, inc_y, a10t, cs_A ); // FLA_Dot2cs_external( FLA_CONJUGATE, FLA_MINUS_ONE, a10t, b10t, FLA_ONE, alpha11 ); bli_zdot2s( BLIS_CONJUGATE, m_behind, buff_m1, a10t, cs_A, b10t, cs_B, buff_1, alpha11 ); // FLA_Inv_scal_external( beta11, alpha11 ); // FLA_Inv_scal_external( beta11, alpha11 ); bli_zinvscals( beta11, alpha11 ); bli_zinvscals( beta11, alpha11 ); // FLA_Axpy_external( FLA_MINUS_ONE_HALF, y10t_t, a10t ); bli_zaxpyv( BLIS_NO_CONJUGATE, m_behind, buff_m1h, y10t, inc_y, a10t, cs_A ); // FLA_Inv_scal_external( beta11, a10t ); bli_zinvscalv( BLIS_NO_CONJUGATE, m_behind, beta11, a10t, cs_A ); /*------------------------------------------------------------*/ } return FLA_SUCCESS; }
FLA_Error FLA_Eig_gest_il_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 bli_zaxpyv(), bli_zdot2s(), bli_zgemv(), bli_zhemv(), bli_zinvscalv(), FLA_MINUS_ONE, FLA_ONE, FLA_ONE_HALF, and FLA_ZERO.
Referenced by FLA_Eig_gest_il_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* a10t = buff_A + (0 )*cs_A + (i )*rs_A; dcomplex* A20 = buff_A + (0 )*cs_A + (i+1)*rs_A; dcomplex* alpha11 = buff_A + (i )*cs_A + (i )*rs_A; dcomplex* a21 = buff_A + (i )*cs_A + (i+1)*rs_A; dcomplex* y10t = buff_y + (0 )*inc_y; dcomplex* b10t = buff_B + (0 )*cs_B + (i )*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_LOWER_TRIANGULAR, FLA_CONJUGATE, // FLA_ONE_HALF, A00, b10t, FLA_ZERO, y10t_t ); bli_zhemv( BLIS_LOWER_TRIANGULAR, BLIS_CONJUGATE, m_behind, buff_1h, A00, rs_A, cs_A, b10t, cs_B, buff_0, y10t, inc_y ); // FLA_Axpy_external( FLA_MINUS_ONE, y10t_t, a10t ); bli_zaxpyv( BLIS_NO_CONJUGATE, m_behind, buff_m1, y10t, inc_y, a10t, cs_A ); // FLA_Dot2cs_external( FLA_CONJUGATE, FLA_MINUS_ONE, a10t, b10t, FLA_ONE, alpha11 ); bli_zdot2s( BLIS_CONJUGATE, m_behind, buff_m1, a10t, cs_A, b10t, cs_B, buff_1, alpha11 ); // FLA_Inv_scal_external( beta11, alpha11 ); // FLA_Inv_scal_external( beta11, alpha11 ); bli_zinvscals( beta11, alpha11 ); bli_zinvscals( beta11, alpha11 ); // FLA_Gemvc_external( FLA_NO_TRANSPOSE, FLA_CONJUGATE, // FLA_MINUS_ONE, A20, b10t, FLA_ONE, a21 ); bli_zgemv( BLIS_NO_TRANSPOSE, BLIS_CONJUGATE, m_ahead, m_behind, buff_m1, A20, rs_A, cs_A, b10t, cs_B, buff_1, a21, rs_A ); // FLA_Inv_scal_external( beta11, a21 ); bli_zinvscalv( BLIS_NO_CONJUGATE, m_ahead, beta11, a21, rs_A ); // FLA_Axpy_external( FLA_MINUS_ONE, y10t_t, a10t ); bli_zaxpyv( BLIS_NO_CONJUGATE, m_behind, buff_m1, y10t, inc_y, a10t, cs_A ); // FLA_Inv_scal_external( beta11, a10t ); bli_zinvscalv( BLIS_NO_CONJUGATE, m_behind, beta11, a10t, cs_A ); /*------------------------------------------------------------*/ } return FLA_SUCCESS; }
FLA_Error FLA_Eig_gest_il_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 bli_zaxpyv(), bli_zcopyv(), bli_zdot2s(), bli_zgemv(), bli_zger(), bli_zinvscalv(), bli_zscalv(), FLA_MINUS_ONE, FLA_MINUS_ONE_HALF, and FLA_ONE.
Referenced by FLA_Eig_gest_il_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* a10t = buff_A + (0 )*cs_A + (i )*rs_A; dcomplex* A20 = buff_A + (0 )*cs_A + (i+1)*rs_A; dcomplex* alpha11 = buff_A + (i )*cs_A + (i )*rs_A; dcomplex* a21 = buff_A + (i )*cs_A + (i+1)*rs_A; dcomplex* y10t = buff_Y + (0 )*cs_Y + (i )*rs_Y; dcomplex* Y20 = buff_Y + (0 )*cs_Y + (i+1)*rs_Y; dcomplex* y21 = buff_Y + (i )*cs_Y + (i+1)*rs_Y; dcomplex* b10t = buff_B + (0 )*cs_B + (i )*rs_B; dcomplex* B20 = buff_B + (0 )*cs_B + (i+1)*rs_B; dcomplex* beta11 = buff_B + (i )*cs_B + (i )*rs_B; dcomplex* b21 = buff_B + (i )*cs_B + (i+1)*rs_B; int m_ahead = m_AB - i - 1; int m_behind = i; /*------------------------------------------------------------*/ // FLA_Axpy_external( FLA_MINUS_ONE_HALF, y10t, a10t ); bli_zaxpyv( BLIS_NO_CONJUGATE, m_behind, buff_m1h, y10t, cs_Y, a10t, cs_A ); // FLA_Dot2cs_external( FLA_CONJUGATE, FLA_MINUS_ONE, a10t, b10t, FLA_ONE, alpha11 ); bli_zdot2s( BLIS_CONJUGATE, m_behind, buff_m1, a10t, cs_A, b10t, cs_B, buff_1, alpha11 ); // FLA_Inv_scal_external( beta11, alpha11 ); // FLA_Inv_scal_external( beta11, alpha11 ); bli_zinvscals( beta11, alpha11 ); bli_zinvscals( beta11, alpha11 ); // FLA_Gemvc_external( FLA_NO_TRANSPOSE, FLA_CONJUGATE, // FLA_MINUS_ONE, A20, b10t, FLA_ONE, a21 ); bli_zgemv( BLIS_NO_TRANSPOSE, BLIS_CONJUGATE, m_ahead, m_behind, buff_m1, A20, rs_A, cs_A, b10t, cs_B, buff_1, a21, rs_A ); // FLA_Inv_scal_external( beta11, a21 ); bli_zinvscalv( BLIS_NO_CONJUGATE, m_ahead, beta11, a21, rs_A ); // FLA_Axpy_external( FLA_MINUS_ONE_HALF, y10t, a10t ); bli_zaxpyv( BLIS_NO_CONJUGATE, m_behind, buff_m1h, y10t, cs_Y, a10t, cs_A ); // FLA_Inv_scal_external( beta11, a10t ); bli_zinvscalv( BLIS_NO_CONJUGATE, m_behind, beta11, a10t, cs_A ); // FLA_Ger_external( FLA_ONE, b21, a10t, Y20 ); bli_zger( BLIS_NO_CONJUGATE, BLIS_NO_CONJUGATE, m_ahead, m_behind, buff_1, b21, rs_B, a10t, cs_A, Y20, rs_Y, cs_Y ); // FLA_Copy_external( b21, y21 ); // FLA_Scal_external( alpha11, y21 ); bli_zcopyv( BLIS_NO_CONJUGATE, m_ahead, b21, rs_B, y21, rs_Y ); bli_zscalv( BLIS_NO_CONJUGATE, m_ahead, alpha11, y21, rs_Y ); // FLA_Gemvc_external( FLA_NO_TRANSPOSE, FLA_CONJUGATE, // FLA_ONE, B20, a10t, FLA_ONE, y21 ); bli_zgemv( BLIS_NO_TRANSPOSE, BLIS_CONJUGATE, m_ahead, m_behind, buff_1, B20, rs_B, cs_B, a10t, cs_A, buff_1, y21, rs_Y ); /*------------------------------------------------------------*/ } return FLA_SUCCESS; }
FLA_Error FLA_Eig_gest_il_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 bli_zaxpyv(), bli_zger(), bli_zher2(), bli_zinvscalv(), FLA_MINUS_ONE, and FLA_MINUS_ONE_HALF.
Referenced by FLA_Eig_gest_il_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* a10t = buff_A + (0 )*cs_A + (i )*rs_A; dcomplex* A20 = buff_A + (0 )*cs_A + (i+1)*rs_A; dcomplex* alpha11 = buff_A + (i )*cs_A + (i )*rs_A; dcomplex* a21 = buff_A + (i )*cs_A + (i+1)*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* b21 = buff_B + (i )*cs_B + (i+1)*rs_B; dcomplex psi11; int m_ahead = m_AB - i - 1; int m_behind = i; /*------------------------------------------------------------*/ // FLA_Inv_scal_external( beta11, a10t ); bli_zinvscalv( BLIS_NO_CONJUGATE, m_behind, beta11, a10t, cs_A ); // FLA_Ger_external( FLA_MINUS_ONE, b21, a10t, A20 ); bli_zger( BLIS_NO_CONJUGATE, BLIS_NO_CONJUGATE, m_ahead, m_behind, buff_m1, b21, rs_B, a10t, cs_A, A20, rs_A, cs_A ); // FLA_Inv_scal_external( beta11, alpha11 ); // FLA_Inv_scal_external( beta11, alpha11 ); bli_zinvscals( beta11, alpha11 ); bli_zinvscals( beta11, alpha11 ); // FLA_Copy_external( alpha11, psi11 ); // FLA_Scal_external( FLA_MINUS_ONE_HALF, psi11 ); bli_zmult3( buff_m1h, alpha11, &psi11 ); // FLA_Inv_scal_external( beta11, a21 ); bli_zinvscalv( BLIS_NO_CONJUGATE, m_ahead, beta11, a21, rs_A ); // FLA_Axpy_external( psi11, b21, a21 ); bli_zaxpyv( BLIS_NO_CONJUGATE, m_ahead, &psi11, b21, rs_B, a21, rs_A ); // FLA_Her2c_external( FLA_LOWER_TRIANGULAR, FLA_NO_CONJUGATE, // FLA_MINUS_ONE, a21, b21, A22 ); bli_zher2( BLIS_LOWER_TRIANGULAR, BLIS_NO_CONJUGATE, m_ahead, buff_m1, a21, rs_A, b21, rs_B, A22, rs_A, cs_A ); // FLA_Axpy_external( psi11, b21, a21 ); bli_zaxpyv( BLIS_NO_CONJUGATE, m_ahead, &psi11, b21, rs_B, a21, rs_A ); /*------------------------------------------------------------*/ } return FLA_SUCCESS; }
FLA_Error FLA_Eig_gest_il_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 bli_zaxpyv(), bli_zher2(), bli_zinvscalv(), bli_ztrsv(), FLA_MINUS_ONE, and FLA_MINUS_ONE_HALF.
Referenced by FLA_Eig_gest_il_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* a21 = buff_A + (i )*cs_A + (i+1)*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* b21 = buff_B + (i )*cs_B + (i+1)*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 ); bli_zinvscals( beta11, alpha11 ); bli_zinvscals( beta11, alpha11 ); // FLA_Copy_external( alpha11, psi11 ); // FLA_Scal_external( FLA_MINUS_ONE_HALF, psi11 ); bli_zmult3( buff_m1h, alpha11, &psi11 ); // FLA_Inv_scal_external( beta11, a21 ); bli_zinvscalv( BLIS_NO_CONJUGATE, m_ahead, beta11, a21, rs_A ); // FLA_Axpy_external( psi11, b21, a21 ); bli_zaxpyv( BLIS_NO_CONJUGATE, m_ahead, &psi11, b21, rs_B, a21, rs_A ); // FLA_Her2c_external( FLA_LOWER_TRIANGULAR, FLA_NO_CONJUGATE, // FLA_MINUS_ONE, a21, b21, A22 ); bli_zher2( BLIS_LOWER_TRIANGULAR, BLIS_NO_CONJUGATE, m_ahead, buff_m1, a21, rs_A, b21, rs_B, A22, rs_A, cs_A ); // FLA_Axpy_external( psi11, b21, a21 ); bli_zaxpyv( BLIS_NO_CONJUGATE, m_ahead, &psi11, b21, rs_B, a21, rs_A ); // FLA_Trsv_external( FLA_LOWER_TRIANGULAR, FLA_NO_TRANSPOSE, FLA_NONUNIT_DIAG, // B22, a21 ); bli_ztrsv( BLIS_LOWER_TRIANGULAR, BLIS_NO_TRANSPOSE, BLIS_NONUNIT_DIAG, m_ahead, B22, rs_B, cs_B, a21, rs_A ); /*------------------------------------------------------------*/ } return FLA_SUCCESS; }
References FLA_Axpy_external(), FLA_Cont_with_1x3_to_1x2(), 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_1x2_to_1x3(), FLA_Repart_2x2_to_3x3(), FLA_Trsv_external(), and FLA_ZERO.
Referenced by FLA_Eig_gest_il().
{ 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 y10t_t, y10t_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 ) ){ 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( y10t, &y10t_t, &y10t_b, 1, FLA_TOP ); // y10t = b10t * A00; // y10t^T = A00^T * b10t^T; // = conj(A00) * b10t^T; FLA_Hemvc_external( FLA_LOWER_TRIANGULAR, FLA_CONJUGATE, FLA_ONE, A00, b10t, FLA_ZERO, y10t_t ); // a10t = a10t * inv( tril( B00 )' ); // a10t^T = inv( conj(tril( B00 )) ) * a10t^T; FLA_Trsv_external( FLA_LOWER_TRIANGULAR, FLA_CONJ_NO_TRANSPOSE, FLA_NONUNIT_DIAG, B00, a10t ); // a10t = a10t - 1/2 * y10t; FLA_Axpy_external( FLA_MINUS_ONE_HALF, y10t_t, a10t ); // alpha11 = alpha11 - a10t * b10t' - b10t * a10t'; FLA_Dot2cs_external( FLA_CONJUGATE, FLA_MINUS_ONE, a10t, b10t, 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 ); // a10t = a10t - 1/2 * y10t; FLA_Axpy_external( FLA_MINUS_ONE_HALF, y10t_t, a10t ); // a10t = inv(beta11) * a10t; FLA_Inv_scal_external( beta11, a10t ); /*------------------------------------------------------------*/ 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; }
References FLA_Axpy_external(), FLA_Cont_with_1x3_to_1x2(), 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_1x2_to_1x3(), FLA_Repart_2x2_to_3x3(), and FLA_ZERO.
Referenced by FLA_Eig_gest_il().
{ 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 y10t_t, y10t_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 ) ){ 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( y10t, &y10t_t, &y10t_b, 1, FLA_TOP ); // y10t = 1/2 * b10t * A00; // y10t^T = 1/2 * A00^T * b10t^T; // = 1/2 * conj(A00) * b10t^T; FLA_Hemvc_external( FLA_LOWER_TRIANGULAR, FLA_CONJUGATE, FLA_ONE_HALF, A00, b10t, FLA_ZERO, y10t_t ); // a10t = a10t - y10t; FLA_Axpy_external( FLA_MINUS_ONE, y10t_t, a10t ); // alpha11 = alpha11 - a10t * b10t' - b10t * a10t'; FLA_Dot2cs_external( FLA_CONJUGATE, FLA_MINUS_ONE, a10t, b10t, 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 ); // a21 = a21 - A20 * b10t'; FLA_Gemvc_external( FLA_NO_TRANSPOSE, FLA_CONJUGATE, FLA_MINUS_ONE, A20, b10t, FLA_ONE, a21 ); // a21 = a21 * inv(conj(beta11)); // = a21 * inv(beta11); FLA_Inv_scal_external( beta11, a21 ); // a10t = a10t - y10t; FLA_Axpy_external( FLA_MINUS_ONE, y10t_t, a10t ); // a10t = inv(beta11) * a10t; FLA_Inv_scal_external( beta11, a10t ); /*------------------------------------------------------------*/ 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; }
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_il().
{ 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 ); /*------------------------------------------------------------*/ // a10t = a10t - 1/2 * y10t; FLA_Axpy_external( FLA_MINUS_ONE_HALF, y10t, a10t ); // alpha11 = alpha11 - a10t * b10t' - b10t * a10t'; FLA_Dot2cs_external( FLA_CONJUGATE, FLA_MINUS_ONE, a10t, b10t, 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 ); // a21 = a21 - A20 * b10t'; FLA_Gemvc_external( FLA_NO_TRANSPOSE, FLA_CONJUGATE, FLA_MINUS_ONE, A20, b10t, FLA_ONE, a21 ); // a21 = a21 * inv(conj(beta11)); // = a21 * inv(beta11); FLA_Inv_scal_external( beta11, a21 ); // a10t = a10t - 1/2 * y10t; FLA_Axpy_external( FLA_MINUS_ONE_HALF, y10t, a10t ); // a10t = inv(beta11) * a10t; FLA_Inv_scal_external( beta11, a10t ); // Y20 = Y20 + b21 * a10t; FLA_Ger_external( FLA_ONE, b21, a10t, Y20 ); // y21 = b21 * alpha11; FLA_Copy_external( b21, y21 ); FLA_Scal_external( alpha11, y21 ); // y21 = y21 + B20 * a10t'; FLA_Gemvc_external( FLA_NO_TRANSPOSE, FLA_CONJUGATE, FLA_ONE, B20, a10t, FLA_ONE, y21 ); /*------------------------------------------------------------*/ 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; }
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_il().
{ 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 y21_l, y21_r; 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_2x1( Y, &yT, // &yB, 0, FLA_TOP ); 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_2x1_to_3x1( yT, &y01, // /* ** */ /* ***** */ // &psi11, // yB, &y21, 1, FLA_BOTTOM ); /*------------------------------------------------------------*/ //FLA_Part_1x2( y21, &y21_l, &y21_r, 1, FLA_LEFT ); // a10t = inv(beta11) * a10t; FLA_Inv_scal_external( beta11, a10t ); // A20 = A20 - b21 * a10t; FLA_Ger_external( FLA_MINUS_ONE, b21, a10t, A20 ); // alpha11 = inv(beta11) * alpha11 * inv(conj(beta11)); // = inv(beta11) * alpha11 * inv(beta11); FLA_Inv_scal_external( beta11, alpha11 ); FLA_Inv_scal_external( beta11, alpha11 ); //// y21 = b21 * alpha11; //FLA_Copy_external( b21, y21_l ); //FLA_Scal_external( alpha11, y21_l ); // psi11 = - 1/2 * alpha11; FLA_Copy_external( alpha11, psi11 ); FLA_Scal_external( FLA_MINUS_ONE_HALF, psi11 ); // a21 = a21 * inv(conj(beta11)); // = a21 * inv(beta11); FLA_Inv_scal_external( beta11, a21 ); //// a21 = a21 - 1/2 * y21; //FLA_Axpy_external( FLA_MINUS_ONE_HALF, y21_l, a21 ); // a21 = a21 - 1/2 * alpha11 * b21; FLA_Axpy_external( psi11, b21, a21 ); // A22 = A22 - a21 * b21' - b21 * a21'; FLA_Her2c_external( FLA_LOWER_TRIANGULAR, FLA_NO_CONJUGATE, FLA_MINUS_ONE, a21, b21, A22 ); //// a21 = a21 - 1/2 * y21; //FLA_Axpy_external( FLA_MINUS_ONE_HALF, y21_l, a21 ); // a21 = a21 - 1/2 * alpha11 * b21; FLA_Axpy_external( psi11, b21, a21 ); /*------------------------------------------------------------*/ 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; }
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_il().
{ 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 y21_l, y21_r; 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_2x1( Y, &yT, // &yB, 0, FLA_TOP ); 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_2x1_to_3x1( yT, &y01, // /* ** */ /* ***** */ // &psi11, // yB, &y21, 1, FLA_BOTTOM ); /*------------------------------------------------------------*/ //FLA_Part_1x2( y21, &y21_l, &y21_r, 1, FLA_LEFT ); // alpha11 = inv(beta11) * alpha11 * inv(conj(beta11)); // = inv(beta11) * alpha11 * inv(beta11); FLA_Inv_scal_external( beta11, alpha11 ); FLA_Inv_scal_external( beta11, alpha11 ); //// y21 = b21 * alpha11; //FLA_Copy_external( b21, y21_l ); //FLA_Scal_external( alpha11, y21_l ); // psi11 = - 1/2 * alpha11; FLA_Copy_external( alpha11, psi11 ); FLA_Scal_external( FLA_MINUS_ONE_HALF, psi11 ); // a21 = a21 * inv(conj(beta11)); // = a21 * inv(beta11); FLA_Inv_scal_external( beta11, a21 ); //// a21 = a21 - 1/2 * y21; //FLA_Axpy_external( FLA_MINUS_ONE_HALF, y21_l, a21 ); // a21 = a21 - 1/2 * alpha11 * b21; FLA_Axpy_external( psi11, b21, a21 ); // A22 = A22 - a21 * b21' - b21 * a21'; FLA_Her2c_external( FLA_LOWER_TRIANGULAR, FLA_NO_CONJUGATE, FLA_MINUS_ONE, a21, b21, A22 ); //// a21 = a21 - 1/2 * y21; //FLA_Axpy_external( FLA_MINUS_ONE_HALF, y21_l, a21 ); // a21 = a21 - 1/2 * alpha11 * b21; FLA_Axpy_external( psi11, b21, a21 ); // a21 = inv( tril( B22 ) ) * a21; FLA_Trsv_external( FLA_LOWER_TRIANGULAR, FLA_NO_TRANSPOSE, FLA_NONUNIT_DIAG, B22, a21 ); /*------------------------------------------------------------*/ 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; }