libflame revision_anchor
|
Go to the source code of this file.
FLA_Error FLA_Chol_l_blk_var1 | ( | FLA_Obj | A, |
fla_chol_t * | cntl | ||
) |
References FLA_Chol_internal(), FLA_Cont_with_3x3_to_2x2(), FLA_Determine_blocksize(), FLA_Herk_internal(), FLA_MINUS_ONE, FLA_Obj_length(), FLA_ONE, FLA_Part_2x2(), FLA_Repart_2x2_to_3x3(), and FLA_Trsm_internal().
Referenced by FLA_Chol_l().
{ FLA_Obj ATL, ATR, A00, A01, A02, ABL, ABR, A10, A11, A12, A20, A21, A22; dim_t b; int value = 0; FLA_Part_2x2( A, &ATL, &ATR, &ABL, &ABR, 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 ); /*------------------------------------------------------------*/ // A10 = A10 * inv( tril( A00 )' ) FLA_Trsm_internal( FLA_RIGHT, FLA_LOWER_TRIANGULAR, FLA_CONJ_TRANSPOSE, FLA_NONUNIT_DIAG, FLA_ONE, A00, A10, FLA_Cntl_sub_trsm( cntl ) ); // A11 = A11 - A10 * A10' FLA_Herk_internal( FLA_LOWER_TRIANGULAR, FLA_NO_TRANSPOSE, FLA_MINUS_ONE, A10, FLA_ONE, A11, FLA_Cntl_sub_herk( cntl ) ); // A11 = chol( A11 ) value = FLA_Chol_internal( FLA_LOWER_TRIANGULAR, A11, FLA_Cntl_sub_chol( cntl ) ); if ( value != FLA_SUCCESS ) return ( FLA_Obj_length( A00 ) + value ); /*------------------------------------------------------------*/ FLA_Cont_with_3x3_to_2x2( &ATL, /**/ &ATR, A00, A01, /**/ A02, A10, A11, /**/ A12, /* ************** */ /* ****************** */ &ABL, /**/ &ABR, A20, A21, /**/ A22, FLA_TL ); } return value; }
FLA_Error FLA_Chol_l_blk_var2 | ( | FLA_Obj | A, |
fla_chol_t * | cntl | ||
) |
References FLA_Chol_internal(), FLA_Cont_with_3x3_to_2x2(), FLA_Determine_blocksize(), FLA_Gemm_internal(), FLA_Herk_internal(), FLA_MINUS_ONE, FLA_Obj_length(), FLA_ONE, FLA_Part_2x2(), FLA_Repart_2x2_to_3x3(), and FLA_Trsm_internal().
Referenced by FLA_Chol_l().
{ FLA_Obj ATL, ATR, A00, A01, A02, ABL, ABR, A10, A11, A12, A20, A21, A22; dim_t b; int value = 0; FLA_Part_2x2( A, &ATL, &ATR, &ABL, &ABR, 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 ); /*------------------------------------------------------------*/ // A11 = A11 - A10 * A10' FLA_Herk_internal( FLA_LOWER_TRIANGULAR, FLA_NO_TRANSPOSE, FLA_MINUS_ONE, A10, FLA_ONE, A11, FLA_Cntl_sub_herk( cntl ) ); // A21 = A21 - A20 * A10' FLA_Gemm_internal( FLA_NO_TRANSPOSE, FLA_CONJ_TRANSPOSE, FLA_MINUS_ONE, A20, A10, FLA_ONE, A21, FLA_Cntl_sub_gemm( cntl ) ); // A11 = chol( A11 ) value = FLA_Chol_internal( FLA_LOWER_TRIANGULAR, A11, FLA_Cntl_sub_chol( cntl ) ); if ( value != FLA_SUCCESS ) return ( FLA_Obj_length( A00 ) + value ); // A21 = A21 * inv( tril( A11 )' ) FLA_Trsm_internal( FLA_RIGHT, FLA_LOWER_TRIANGULAR, FLA_CONJ_TRANSPOSE, FLA_NONUNIT_DIAG, FLA_ONE, A11, A21, FLA_Cntl_sub_trsm( cntl ) ); /*------------------------------------------------------------*/ FLA_Cont_with_3x3_to_2x2( &ATL, /**/ &ATR, A00, A01, /**/ A02, A10, A11, /**/ A12, /* ************** */ /* ****************** */ &ABL, /**/ &ABR, A20, A21, /**/ A22, FLA_TL ); } return value; }
FLA_Error FLA_Chol_l_blk_var3 | ( | FLA_Obj | A, |
fla_chol_t * | cntl | ||
) |
References FLA_Chol_internal(), FLA_Cont_with_3x3_to_2x2(), FLA_Determine_blocksize(), FLA_Herk_internal(), FLA_MINUS_ONE, FLA_Obj_length(), FLA_ONE, FLA_Part_2x2(), FLA_Repart_2x2_to_3x3(), and FLA_Trsm_internal().
Referenced by FLA_Chol_l().
{ FLA_Obj ATL, ATR, A00, A01, A02, ABL, ABR, A10, A11, A12, A20, A21, A22; dim_t b; int value = 0; FLA_Part_2x2( A, &ATL, &ATR, &ABL, &ABR, 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 ); /*------------------------------------------------------------*/ // A11 = chol( A11 ) value = FLA_Chol_internal( FLA_LOWER_TRIANGULAR, A11, FLA_Cntl_sub_chol( cntl ) ); if ( value != FLA_SUCCESS ) return ( FLA_Obj_length( A00 ) + value ); // A21 = A21 * inv( tril( A11 )' ) FLA_Trsm_internal( FLA_RIGHT, FLA_LOWER_TRIANGULAR, FLA_CONJ_TRANSPOSE, FLA_NONUNIT_DIAG, FLA_ONE, A11, A21, FLA_Cntl_sub_trsm( cntl ) ); // A22 = A22 - A21 * A21' FLA_Herk_internal( FLA_LOWER_TRIANGULAR, FLA_NO_TRANSPOSE, FLA_MINUS_ONE, A21, FLA_ONE, A22, FLA_Cntl_sub_herk( cntl ) ); /*------------------------------------------------------------*/ FLA_Cont_with_3x3_to_2x2( &ATL, /**/ &ATR, A00, A01, /**/ A02, A10, A11, /**/ A12, /* ************** */ /* ****************** */ &ABL, /**/ &ABR, A20, A21, /**/ A22, FLA_TL ); } return value; }
References bli_cdots(), bli_ctrsv(), FLA_MINUS_ONE, and FLA_ONE.
Referenced by FLA_Chol_l_opt_var1().
{ scomplex* buff_1 = FLA_COMPLEX_PTR( FLA_ONE ); scomplex* buff_m1 = FLA_COMPLEX_PTR( FLA_MINUS_ONE ); int i; FLA_Error e_val; for ( i = 0; i < mn_A; ++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; int mn_behind = i; /*------------------------------------------------------------*/ // FLA_Trsv_external( FLA_LOWER_TRIANGULAR, FLA_CONJ_NO_TRANSPOSE, FLA_NONUNIT_DIAG, A00, a10t ); bli_ctrsv( BLIS_LOWER_TRIANGULAR, BLIS_CONJ_NO_TRANSPOSE, BLIS_NONUNIT_DIAG, mn_behind, A00, rs_A, cs_A, a10t, cs_A ); // FLA_Dotcs_external( FLA_CONJUGATE, FLA_MINUS_ONE, a10t, a10t, FLA_ONE, alpha11 ); bli_cdots( BLIS_CONJUGATE, mn_behind, buff_m1, a10t, cs_A, a10t, cs_A, buff_1, alpha11 ); // value = FLA_Sqrt( alpha11 ); // if ( value != FLA_SUCCESS ) // return ( FLA_Obj_length( A00 ) + 1 ); bli_csqrte( alpha11, &e_val ); if ( e_val != FLA_SUCCESS ) return mn_behind; /*------------------------------------------------------------*/ } return FLA_SUCCESS; }
References bli_cdots(), bli_cgemv(), bli_cinvscalv(), FLA_MINUS_ONE, and FLA_ONE.
Referenced by FLA_Chol_l_opt_var2().
{ scomplex* buff_1 = FLA_COMPLEX_PTR( FLA_ONE ); scomplex* buff_m1 = FLA_COMPLEX_PTR( FLA_MINUS_ONE ); int i; FLA_Error e_val; for ( i = 0; i < mn_A; ++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; int mn_ahead = mn_A - i - 1; int mn_behind = i; /*------------------------------------------------------------*/ // FLA_Dotcs_external( FLA_CONJUGATE, FLA_MINUS_ONE, a10t, a10t, FLA_ONE, alpha11 ); bli_cdots( BLIS_CONJUGATE, mn_behind, buff_m1, a10t, cs_A, a10t, cs_A, buff_1, alpha11 ); // FLA_Gemvc_external( FLA_NO_TRANSPOSE, FLA_CONJUGATE, FLA_MINUS_ONE, A20, a10t, FLA_ONE, a21 ); bli_cgemv( BLIS_NO_TRANSPOSE, BLIS_CONJUGATE, mn_ahead, mn_behind, buff_m1, A20, rs_A, cs_A, a10t, cs_A, buff_1, a21, rs_A ); // value = FLA_Sqrt( alpha11 ); // if ( value != FLA_SUCCESS ) // return ( FLA_Obj_length( A00 ) + 1 ); bli_csqrte( alpha11, &e_val ); if ( e_val != FLA_SUCCESS ) return mn_behind; // FLA_Inv_scal_external( alpha11, a21 ); bli_cinvscalv( BLIS_NO_CONJUGATE, mn_ahead, alpha11, a21, rs_A ); /*------------------------------------------------------------*/ } return FLA_SUCCESS; }
References bli_cher(), bli_cinvscalv(), and FLA_MINUS_ONE.
Referenced by FLA_Chol_l_opt_var3().
{ float* buff_m1 = FLA_FLOAT_PTR( FLA_MINUS_ONE ); int i; FLA_Error e_val; for ( i = 0; i < mn_A; ++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; int mn_ahead = mn_A - i - 1; int mn_behind = i; /*------------------------------------------------------------*/ // value = FLA_Sqrt( alpha11 ); // if ( value != FLA_SUCCESS ) // return ( FLA_Obj_length( A00 ) + 1 ); bli_csqrte( alpha11, &e_val ); if ( e_val != FLA_SUCCESS ) return mn_behind; // FLA_Inv_scal_external( alpha11, a21 ); bli_cinvscalv( BLIS_NO_CONJUGATE, mn_ahead, alpha11, a21, rs_A ); // FLA_Her_external( FLA_LOWER_TRIANGULAR, FLA_MINUS_ONE, a21, A22 ); bli_cher( BLIS_LOWER_TRIANGULAR, BLIS_NO_CONJUGATE, mn_ahead, buff_m1, a21, rs_A, A22, rs_A, cs_A ); /*------------------------------------------------------------*/ } return FLA_SUCCESS; }
FLA_Error FLA_Chol_l_opd_var1 | ( | int | mn_A, |
double * | A, | ||
int | rs_A, | ||
int | cs_A | ||
) |
References bli_ddots(), bli_dtrsv(), FLA_MINUS_ONE, and FLA_ONE.
Referenced by FLA_Chol_l_opt_var1().
{ double* buff_1 = FLA_DOUBLE_PTR( FLA_ONE ); double* buff_m1 = FLA_DOUBLE_PTR( FLA_MINUS_ONE ); int i; FLA_Error e_val; for ( i = 0; i < mn_A; ++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; int mn_behind = i; /*------------------------------------------------------------*/ // FLA_Trsv_external( FLA_LOWER_TRIANGULAR, FLA_CONJ_NO_TRANSPOSE, FLA_NONUNIT_DIAG, A00, a10t ); bli_dtrsv( BLIS_LOWER_TRIANGULAR, BLIS_CONJ_NO_TRANSPOSE, BLIS_NONUNIT_DIAG, mn_behind, A00, rs_A, cs_A, a10t, cs_A ); // FLA_Dotcs_external( FLA_CONJUGATE, FLA_MINUS_ONE, a10t, a10t, FLA_ONE, alpha11 ); bli_ddots( BLIS_CONJUGATE, mn_behind, buff_m1, a10t, cs_A, a10t, cs_A, buff_1, alpha11 ); // value = FLA_Sqrt( alpha11 ); // if ( value != FLA_SUCCESS ) // return ( FLA_Obj_length( A00 ) + 1 ); bli_dsqrte( alpha11, &e_val ); if ( e_val != FLA_SUCCESS ) return mn_behind; /*------------------------------------------------------------*/ } return FLA_SUCCESS; }
FLA_Error FLA_Chol_l_opd_var2 | ( | int | mn_A, |
double * | A, | ||
int | rs_A, | ||
int | cs_A | ||
) |
References bli_ddots(), bli_dgemv(), bli_dinvscalv(), FLA_MINUS_ONE, and FLA_ONE.
Referenced by FLA_Chol_l_opt_var2().
{ double* buff_1 = FLA_DOUBLE_PTR( FLA_ONE ); double* buff_m1 = FLA_DOUBLE_PTR( FLA_MINUS_ONE ); int i; FLA_Error e_val; for ( i = 0; i < mn_A; ++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; int mn_ahead = mn_A - i - 1; int mn_behind = i; /*------------------------------------------------------------*/ // FLA_Dotcs_external( FLA_CONJUGATE, FLA_MINUS_ONE, a10t, a10t, FLA_ONE, alpha11 ); bli_ddots( BLIS_CONJUGATE, mn_behind, buff_m1, a10t, cs_A, a10t, cs_A, buff_1, alpha11 ); // FLA_Gemvc_external( FLA_NO_TRANSPOSE, FLA_CONJUGATE, FLA_MINUS_ONE, A20, a10t, FLA_ONE, a21 ); bli_dgemv( BLIS_NO_TRANSPOSE, BLIS_CONJUGATE, mn_ahead, mn_behind, buff_m1, A20, rs_A, cs_A, a10t, cs_A, buff_1, a21, rs_A ); // value = FLA_Sqrt( alpha11 ); // if ( value != FLA_SUCCESS ) // return ( FLA_Obj_length( A00 ) + 1 ); bli_dsqrte( alpha11, &e_val ); if ( e_val != FLA_SUCCESS ) return mn_behind; // FLA_Inv_scal_external( alpha11, a21 ); bli_dinvscalv( BLIS_NO_CONJUGATE, mn_ahead, alpha11, a21, rs_A ); /*------------------------------------------------------------*/ } return FLA_SUCCESS; }
FLA_Error FLA_Chol_l_opd_var3 | ( | int | mn_A, |
double * | A, | ||
int | rs_A, | ||
int | cs_A | ||
) |
References bli_dinvscalv(), bli_dsyr(), and FLA_MINUS_ONE.
Referenced by FLA_Chol_l_opt_var3().
{ double* buff_m1 = FLA_DOUBLE_PTR( FLA_MINUS_ONE ); int i; FLA_Error e_val; for ( i = 0; i < mn_A; ++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; int mn_ahead = mn_A - i - 1; int mn_behind = i; /*------------------------------------------------------------*/ // value = FLA_Sqrt( alpha11 ); // if ( value != FLA_SUCCESS ) // return ( FLA_Obj_length( A00 ) + 1 ); bli_dsqrte( alpha11, &e_val ); if ( e_val != FLA_SUCCESS ) return mn_behind; // FLA_Inv_scal_external( alpha11, a21 ); bli_dinvscalv( BLIS_NO_CONJUGATE, mn_ahead, alpha11, a21, rs_A ); // FLA_Her_external( FLA_LOWER_TRIANGULAR, FLA_MINUS_ONE, a21, A22 ); bli_dsyr( BLIS_LOWER_TRIANGULAR, mn_ahead, buff_m1, a21, rs_A, A22, rs_A, cs_A ); /*------------------------------------------------------------*/ } return FLA_SUCCESS; }
FLA_Error FLA_Chol_l_ops_var1 | ( | int | mn_A, |
float * | A, | ||
int | rs_A, | ||
int | cs_A | ||
) |
References bli_sdots(), bli_strsv(), FLA_MINUS_ONE, and FLA_ONE.
Referenced by FLA_Chol_l_opt_var1().
{ float* buff_1 = FLA_FLOAT_PTR( FLA_ONE ); float* buff_m1 = FLA_FLOAT_PTR( FLA_MINUS_ONE ); int i; FLA_Error e_val; for ( i = 0; i < mn_A; ++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; int mn_behind = i; /*------------------------------------------------------------*/ // FLA_Trsv_external( FLA_LOWER_TRIANGULAR, FLA_CONJ_NO_TRANSPOSE, FLA_NONUNIT_DIAG, A00, a10t ); bli_strsv( BLIS_LOWER_TRIANGULAR, BLIS_CONJ_NO_TRANSPOSE, BLIS_NONUNIT_DIAG, mn_behind, A00, rs_A, cs_A, a10t, cs_A ); // FLA_Dotcs_external( FLA_CONJUGATE, FLA_MINUS_ONE, a10t, a10t, FLA_ONE, alpha11 ); bli_sdots( BLIS_CONJUGATE, mn_behind, buff_m1, a10t, cs_A, a10t, cs_A, buff_1, alpha11 ); // value = FLA_Sqrt( alpha11 ); // if ( value != FLA_SUCCESS ) // return ( FLA_Obj_length( A00 ) + 1 ); bli_ssqrte( alpha11, &e_val ); if ( e_val != FLA_SUCCESS ) return mn_behind; /*------------------------------------------------------------*/ } return FLA_SUCCESS; }
FLA_Error FLA_Chol_l_ops_var2 | ( | int | mn_A, |
float * | A, | ||
int | rs_A, | ||
int | cs_A | ||
) |
References bli_sdots(), bli_sgemv(), bli_sinvscalv(), FLA_MINUS_ONE, and FLA_ONE.
Referenced by FLA_Chol_l_opt_var2().
{ float* buff_1 = FLA_FLOAT_PTR( FLA_ONE ); float* buff_m1 = FLA_FLOAT_PTR( FLA_MINUS_ONE ); int i; FLA_Error e_val; for ( i = 0; i < mn_A; ++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; int mn_ahead = mn_A - i - 1; int mn_behind = i; /*------------------------------------------------------------*/ // FLA_Dotcs_external( FLA_CONJUGATE, FLA_MINUS_ONE, a10t, a10t, FLA_ONE, alpha11 ); bli_sdots( BLIS_CONJUGATE, mn_behind, buff_m1, a10t, cs_A, a10t, cs_A, buff_1, alpha11 ); // FLA_Gemvc_external( FLA_NO_TRANSPOSE, FLA_CONJUGATE, FLA_MINUS_ONE, A20, a10t, FLA_ONE, a21 ); bli_sgemv( BLIS_NO_TRANSPOSE, BLIS_CONJUGATE, mn_ahead, mn_behind, buff_m1, A20, rs_A, cs_A, a10t, cs_A, buff_1, a21, rs_A ); // value = FLA_Sqrt( alpha11 ); // if ( value != FLA_SUCCESS ) // return ( FLA_Obj_length( A00 ) + 1 ); bli_ssqrte( alpha11, &e_val ); if ( e_val != FLA_SUCCESS ) return mn_behind; // FLA_Inv_scal_external( alpha11, a21 ); bli_sinvscalv( BLIS_NO_CONJUGATE, mn_ahead, alpha11, a21, rs_A ); /*------------------------------------------------------------*/ } return FLA_SUCCESS; }
FLA_Error FLA_Chol_l_ops_var3 | ( | int | mn_A, |
float * | A, | ||
int | rs_A, | ||
int | cs_A | ||
) |
References bli_sinvscalv(), bli_ssyr(), and FLA_MINUS_ONE.
Referenced by FLA_Chol_l_opt_var3().
{ float* buff_m1 = FLA_FLOAT_PTR( FLA_MINUS_ONE ); int i; FLA_Error e_val; for ( i = 0; i < mn_A; ++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; int mn_ahead = mn_A - i - 1; int mn_behind = i; /*------------------------------------------------------------*/ // value = FLA_Sqrt( alpha11 ); // if ( value != FLA_SUCCESS ) // return ( FLA_Obj_length( A00 ) + 1 ); bli_ssqrte( alpha11, &e_val ); if ( e_val != FLA_SUCCESS ) return mn_behind; // FLA_Inv_scal_external( alpha11, a21 ); bli_sinvscalv( BLIS_NO_CONJUGATE, mn_ahead, alpha11, a21, rs_A ); // FLA_Her_external( FLA_LOWER_TRIANGULAR, FLA_MINUS_ONE, a21, A22 ); bli_ssyr( BLIS_LOWER_TRIANGULAR, mn_ahead, buff_m1, a21, rs_A, A22, rs_A, cs_A ); /*------------------------------------------------------------*/ } return FLA_SUCCESS; }
References FLA_Chol_l_opc_var1(), FLA_Chol_l_opd_var1(), FLA_Chol_l_ops_var1(), FLA_Chol_l_opz_var1(), FLA_Obj_col_stride(), FLA_Obj_datatype(), FLA_Obj_length(), and FLA_Obj_row_stride().
Referenced by FLA_Chol_l().
{ FLA_Datatype datatype; int mn_A; int rs_A, cs_A; datatype = FLA_Obj_datatype( A ); mn_A = FLA_Obj_length( A ); rs_A = FLA_Obj_row_stride( A ); cs_A = FLA_Obj_col_stride( A ); switch ( datatype ) { case FLA_FLOAT: { float* buff_A = FLA_FLOAT_PTR( A ); FLA_Chol_l_ops_var1( mn_A, buff_A, rs_A, cs_A ); break; } case FLA_DOUBLE: { double* buff_A = FLA_DOUBLE_PTR( A ); FLA_Chol_l_opd_var1( mn_A, buff_A, rs_A, cs_A ); break; } case FLA_COMPLEX: { scomplex* buff_A = FLA_COMPLEX_PTR( A ); FLA_Chol_l_opc_var1( mn_A, buff_A, rs_A, cs_A ); break; } case FLA_DOUBLE_COMPLEX: { dcomplex* buff_A = FLA_DOUBLE_COMPLEX_PTR( A ); FLA_Chol_l_opz_var1( mn_A, buff_A, rs_A, cs_A ); break; } } return FLA_SUCCESS; }
References FLA_Chol_l_opc_var2(), FLA_Chol_l_opd_var2(), FLA_Chol_l_ops_var2(), FLA_Chol_l_opz_var2(), FLA_Obj_col_stride(), FLA_Obj_datatype(), FLA_Obj_length(), and FLA_Obj_row_stride().
Referenced by FLA_Chol_l().
{ FLA_Datatype datatype; int mn_A; int rs_A, cs_A; datatype = FLA_Obj_datatype( A ); mn_A = FLA_Obj_length( A ); rs_A = FLA_Obj_row_stride( A ); cs_A = FLA_Obj_col_stride( A ); switch ( datatype ) { case FLA_FLOAT: { float* buff_A = FLA_FLOAT_PTR( A ); FLA_Chol_l_ops_var2( mn_A, buff_A, rs_A, cs_A ); break; } case FLA_DOUBLE: { double* buff_A = FLA_DOUBLE_PTR( A ); FLA_Chol_l_opd_var2( mn_A, buff_A, rs_A, cs_A ); break; } case FLA_COMPLEX: { scomplex* buff_A = FLA_COMPLEX_PTR( A ); FLA_Chol_l_opc_var2( mn_A, buff_A, rs_A, cs_A ); break; } case FLA_DOUBLE_COMPLEX: { dcomplex* buff_A = FLA_DOUBLE_COMPLEX_PTR( A ); FLA_Chol_l_opz_var2( mn_A, buff_A, rs_A, cs_A ); break; } } return FLA_SUCCESS; }
References FLA_Chol_l_opc_var3(), FLA_Chol_l_opd_var3(), FLA_Chol_l_ops_var3(), FLA_Chol_l_opz_var3(), FLA_Obj_col_stride(), FLA_Obj_datatype(), FLA_Obj_length(), and FLA_Obj_row_stride().
Referenced by FLA_Chol_l().
{ FLA_Datatype datatype; int mn_A; int rs_A, cs_A; datatype = FLA_Obj_datatype( A ); mn_A = FLA_Obj_length( A ); rs_A = FLA_Obj_row_stride( A ); cs_A = FLA_Obj_col_stride( A ); switch ( datatype ) { case FLA_FLOAT: { float* buff_A = FLA_FLOAT_PTR( A ); FLA_Chol_l_ops_var3( mn_A, buff_A, rs_A, cs_A ); break; } case FLA_DOUBLE: { double* buff_A = FLA_DOUBLE_PTR( A ); FLA_Chol_l_opd_var3( mn_A, buff_A, rs_A, cs_A ); break; } case FLA_COMPLEX: { scomplex* buff_A = FLA_COMPLEX_PTR( A ); FLA_Chol_l_opc_var3( mn_A, buff_A, rs_A, cs_A ); break; } case FLA_DOUBLE_COMPLEX: { dcomplex* buff_A = FLA_DOUBLE_COMPLEX_PTR( A ); FLA_Chol_l_opz_var3( mn_A, buff_A, rs_A, cs_A ); break; } } return FLA_SUCCESS; }
References bli_zdots(), bli_ztrsv(), FLA_MINUS_ONE, and FLA_ONE.
Referenced by FLA_Chol_l_opt_var1().
{ dcomplex* buff_1 = FLA_DOUBLE_COMPLEX_PTR( FLA_ONE ); dcomplex* buff_m1 = FLA_DOUBLE_COMPLEX_PTR( FLA_MINUS_ONE ); int i; FLA_Error e_val; for ( i = 0; i < mn_A; ++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; int mn_behind = i; /*------------------------------------------------------------*/ // FLA_Trsv_external( FLA_LOWER_TRIANGULAR, FLA_CONJ_NO_TRANSPOSE, FLA_NONUNIT_DIAG, A00, a10t ); bli_ztrsv( BLIS_LOWER_TRIANGULAR, BLIS_CONJ_NO_TRANSPOSE, BLIS_NONUNIT_DIAG, mn_behind, A00, rs_A, cs_A, a10t, cs_A ); // FLA_Dotcs_external( FLA_CONJUGATE, FLA_MINUS_ONE, a10t, a10t, FLA_ONE, alpha11 ); bli_zdots( BLIS_CONJUGATE, mn_behind, buff_m1, a10t, cs_A, a10t, cs_A, buff_1, alpha11 ); // value = FLA_Sqrt( alpha11 ); // if ( value != FLA_SUCCESS ) // return ( FLA_Obj_length( A00 ) + 1 ); bli_zsqrte( alpha11, &e_val ); if ( e_val != FLA_SUCCESS ) return mn_behind; /*------------------------------------------------------------*/ } return FLA_SUCCESS; }
References bli_zdots(), bli_zgemv(), bli_zinvscalv(), FLA_MINUS_ONE, and FLA_ONE.
Referenced by FLA_Chol_l_opt_var2().
{ dcomplex* buff_1 = FLA_DOUBLE_COMPLEX_PTR( FLA_ONE ); dcomplex* buff_m1 = FLA_DOUBLE_COMPLEX_PTR( FLA_MINUS_ONE ); int i; FLA_Error e_val; for ( i = 0; i < mn_A; ++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; int mn_ahead = mn_A - i - 1; int mn_behind = i; /*------------------------------------------------------------*/ // FLA_Dotcs_external( FLA_CONJUGATE, FLA_MINUS_ONE, a10t, a10t, FLA_ONE, alpha11 ); bli_zdots( BLIS_CONJUGATE, mn_behind, buff_m1, a10t, cs_A, a10t, cs_A, buff_1, alpha11 ); // FLA_Gemvc_external( FLA_NO_TRANSPOSE, FLA_CONJUGATE, FLA_MINUS_ONE, A20, a10t, FLA_ONE, a21 ); bli_zgemv( BLIS_NO_TRANSPOSE, BLIS_CONJUGATE, mn_ahead, mn_behind, buff_m1, A20, rs_A, cs_A, a10t, cs_A, buff_1, a21, rs_A ); // value = FLA_Sqrt( alpha11 ); // if ( value != FLA_SUCCESS ) // return ( FLA_Obj_length( A00 ) + 1 ); bli_zsqrte( alpha11, &e_val ); if ( e_val != FLA_SUCCESS ) return mn_behind; // FLA_Inv_scal_external( alpha11, a21 ); bli_zinvscalv( BLIS_NO_CONJUGATE, mn_ahead, alpha11, a21, rs_A ); /*------------------------------------------------------------*/ } return FLA_SUCCESS; }
References bli_zher(), bli_zinvscalv(), and FLA_MINUS_ONE.
Referenced by FLA_Chol_l_opt_var3().
{ double* buff_m1 = FLA_DOUBLE_PTR( FLA_MINUS_ONE ); int i; FLA_Error e_val; for ( i = 0; i < mn_A; ++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; int mn_ahead = mn_A - i - 1; int mn_behind = i; /*------------------------------------------------------------*/ // value = FLA_Sqrt( alpha11 ); // if ( value != FLA_SUCCESS ) // return ( FLA_Obj_length( A00 ) + 1 ); bli_zsqrte( alpha11, &e_val ); if ( e_val != FLA_SUCCESS ) return mn_behind; // FLA_Inv_scal_external( alpha11, a21 ); bli_zinvscalv( BLIS_NO_CONJUGATE, mn_ahead, alpha11, a21, rs_A ); // FLA_Her_external( FLA_LOWER_TRIANGULAR, FLA_MINUS_ONE, a21, A22 ); bli_zher( BLIS_LOWER_TRIANGULAR, BLIS_NO_CONJUGATE, mn_ahead, buff_m1, a21, rs_A, A22, rs_A, cs_A ); /*------------------------------------------------------------*/ } return FLA_SUCCESS; }
References FLA_Cont_with_3x3_to_2x2(), FLA_Dotcs_external(), FLA_MINUS_ONE, FLA_Obj_length(), FLA_ONE, FLA_Part_2x2(), FLA_Repart_2x2_to_3x3(), FLA_Sqrt(), and FLA_Trsv_external().
Referenced by FLA_Chol_l().
{ FLA_Obj ATL, ATR, A00, a01, A02, ABL, ABR, a10t, alpha11, a12t, A20, a21, A22; int value = 0; FLA_Part_2x2( A, &ATL, &ATR, &ABL, &ABR, 0, 0, FLA_TL ); while ( FLA_Obj_length( ATL ) < FLA_Obj_length( A ) ){ FLA_Repart_2x2_to_3x3( ATL, /**/ ATR, &A00, /**/ &a01, &A02, /* ************* */ /* ************************** */ &a10t, /**/ &alpha11, &a12t, ABL, /**/ ABR, &A20, /**/ &a21, &A22, 1, 1, FLA_BR ); /*------------------------------------------------------------*/ // a10t = a10t * inv( tril( A00 )' ) // a10t' = inv( conj( tril( A00 ) ) ) * a10t' FLA_Trsv_external( FLA_LOWER_TRIANGULAR, FLA_CONJ_NO_TRANSPOSE, FLA_NONUNIT_DIAG, A00, a10t ); // alpha11 = alpha11 - a10t * a10t' FLA_Dotcs_external( FLA_CONJUGATE, FLA_MINUS_ONE, a10t, a10t, FLA_ONE, alpha11 ); // alpha11 = sqrt( alpha11 ) value = FLA_Sqrt( alpha11 ); if ( value != FLA_SUCCESS ) return ( FLA_Obj_length( A00 ) + 1 ); /*------------------------------------------------------------*/ FLA_Cont_with_3x3_to_2x2( &ATL, /**/ &ATR, A00, a01, /**/ A02, a10t, alpha11, /**/ a12t, /* ************** */ /* ************************ */ &ABL, /**/ &ABR, A20, a21, /**/ A22, FLA_TL ); } return value; }
References FLA_Cont_with_3x3_to_2x2(), FLA_Dotcs_external(), FLA_Gemvc_external(), FLA_Inv_scal_external(), FLA_MINUS_ONE, FLA_Obj_length(), FLA_ONE, FLA_Part_2x2(), FLA_Repart_2x2_to_3x3(), and FLA_Sqrt().
Referenced by FLA_Chol_l().
{ FLA_Obj ATL, ATR, A00, a01, A02, ABL, ABR, a10t, alpha11, a12t, A20, a21, A22; int value = 0; FLA_Part_2x2( A, &ATL, &ATR, &ABL, &ABR, 0, 0, FLA_TL ); while ( FLA_Obj_length( ATL ) < FLA_Obj_length( A ) ){ FLA_Repart_2x2_to_3x3( ATL, /**/ ATR, &A00, /**/ &a01, &A02, /* ************* */ /* ************************** */ &a10t, /**/ &alpha11, &a12t, ABL, /**/ ABR, &A20, /**/ &a21, &A22, 1, 1, FLA_BR ); /*------------------------------------------------------------*/ // alpha11 = alpha11 - a10t * a10t' FLA_Dotcs_external( FLA_CONJUGATE, FLA_MINUS_ONE, a10t, a10t, FLA_ONE, alpha11 ); // a21 = a21 - A20 * a10t' FLA_Gemvc_external( FLA_NO_TRANSPOSE, FLA_CONJUGATE, FLA_MINUS_ONE, A20, a10t, FLA_ONE, a21 ); // alpha11 = sqrt( alpha11 ) value = FLA_Sqrt( alpha11 ); if ( value != FLA_SUCCESS ) return ( FLA_Obj_length( A00 ) + 1 ); // a21 = a21 / alpha11 FLA_Inv_scal_external( alpha11, a21 ); /*------------------------------------------------------------*/ FLA_Cont_with_3x3_to_2x2( &ATL, /**/ &ATR, A00, a01, /**/ A02, a10t, alpha11, /**/ a12t, /* ************** */ /* ************************ */ &ABL, /**/ &ABR, A20, a21, /**/ A22, FLA_TL ); } return value; }
References FLA_Cont_with_3x3_to_2x2(), FLA_Her_external(), FLA_Inv_scal_external(), FLA_MINUS_ONE, FLA_Obj_length(), FLA_Part_2x2(), FLA_Repart_2x2_to_3x3(), and FLA_Sqrt().
Referenced by FLA_Chol_l().
{ FLA_Obj ATL, ATR, A00, a01, A02, ABL, ABR, a10t, alpha11, a12t, A20, a21, A22; int value = 0; FLA_Part_2x2( A, &ATL, &ATR, &ABL, &ABR, 0, 0, FLA_TL ); while ( FLA_Obj_length( ATL ) < FLA_Obj_length( A ) ){ FLA_Repart_2x2_to_3x3( ATL, /**/ ATR, &A00, /**/ &a01, &A02, /* ************* */ /* ************************** */ &a10t, /**/ &alpha11, &a12t, ABL, /**/ ABR, &A20, /**/ &a21, &A22, 1, 1, FLA_BR ); /*------------------------------------------------------------*/ // alpha11 = sqrt( alpha11 ) value = FLA_Sqrt( alpha11 ); if ( value != FLA_SUCCESS ) return ( FLA_Obj_length( A00 ) + 1 ); // a21 = a21 / alpha11 FLA_Inv_scal_external( alpha11, a21 ); // A22 = A22 - a21 * a21' FLA_Her_external( FLA_LOWER_TRIANGULAR, FLA_MINUS_ONE, a21, A22 ); /*------------------------------------------------------------*/ FLA_Cont_with_3x3_to_2x2( &ATL, /**/ &ATR, A00, a01, /**/ A02, a10t, alpha11, /**/ a12t, /* ************** */ /* ************************ */ &ABL, /**/ &ABR, A20, a21, /**/ A22, FLA_TL ); } return value; }