FLA_Apply_Q_UT_UD_lhc.h File Reference

(r)

Go to the source code of this file.

Functions

FLA_Error FLA_Apply_Q_UT_UD_lhc_blk_var1 (FLA_Obj D, FLA_Obj T, FLA_Obj W1, FLA_Obj C, FLA_Obj E, fla_apqutud_t *cntl)
FLA_Error FLA_Apply_Q_UT_UD_lhc_blk_var2 (FLA_Obj D, FLA_Obj T, FLA_Obj W1, FLA_Obj C, FLA_Obj E, fla_apqutud_t *cntl)
FLA_Error FLA_Apply_Q_UT_UD_lhc_blk_var3 (FLA_Obj D, FLA_Obj T, FLA_Obj W, FLA_Obj C, FLA_Obj E, fla_apqutud_t *cntl)


Function Documentation

FLA_Error FLA_Apply_Q_UT_UD_lhc_blk_var1 ( FLA_Obj  D,
FLA_Obj  T,
FLA_Obj  W1,
FLA_Obj  C,
FLA_Obj  E,
fla_apqutud_t cntl 
)

References FLA_Axpy_internal(), FLA_Cont_with_1x3_to_1x2(), FLA_Cont_with_3x1_to_2x1(), FLA_Copy_internal(), FLA_Determine_blocksize(), FLA_Gemm_internal(), FLA_MINUS_ONE, FLA_Obj_width(), FLA_ONE, FLA_Part_1x2(), FLA_Part_2x1(), FLA_Part_2x2(), FLA_Repart_1x2_to_1x3(), FLA_Repart_2x1_to_3x1(), and FLA_Trsm_internal().

Referenced by FLA_Apply_Q_UT_UD_lhc().

00037 {
00038   FLA_Obj DL,    DR,       D0,  D1,  D2;
00039 
00040   FLA_Obj TL,    TR,       T0,  T1,  T2;
00041 
00042   FLA_Obj T1T,
00043           T2B;
00044 
00045   FLA_Obj CT,              C0,
00046           CB,              C1,
00047                            C2;
00048 
00049   FLA_Obj W1TL,  W1TR,
00050           W1BL,  W1BR;
00051 
00052   dim_t b;
00053 
00054   FLA_Part_1x2( D,    &DL,  &DR,      0, FLA_LEFT );
00055 
00056   FLA_Part_1x2( T,    &TL,  &TR,      0, FLA_LEFT );
00057 
00058   FLA_Part_2x1( C,    &CT, 
00059                       &CB,            0, FLA_TOP );
00060 
00061   while ( FLA_Obj_width( DL ) < FLA_Obj_width( D ) ){
00062 
00063     b = FLA_Determine_blocksize( DR, FLA_RIGHT, FLA_Cntl_blocksize( cntl ) );
00064 
00065     FLA_Repart_1x2_to_1x3( DL,  /**/ DR,        &D0, /**/ &D1, &D2,
00066                            b, FLA_RIGHT );
00067 
00068     FLA_Repart_1x2_to_1x3( TL,  /**/ TR,        &T0, /**/ &T1, &T2,
00069                            b, FLA_RIGHT );
00070 
00071     FLA_Repart_2x1_to_3x1( CT,                &C0, 
00072                         /* ** */            /* ** */
00073                                               &C1, 
00074                            CB,                &C2,        b, FLA_BOTTOM );
00075 
00076     /*------------------------------------------------------------*/
00077 
00078     FLA_Part_2x1( T1,    &T1T, 
00079                          &T2B, b, FLA_TOP );
00080 
00081     // W1T  = FLA_Top_part( W1, b );
00082     // W1TL = FLA_Left_part( W1T, FLA_Obj_width( C1 ) );
00083     FLA_Part_2x2( W1,    &W1TL, &W1TR,
00084                          &W1BL, &W1BR,     b, FLA_Obj_width( C1 ), FLA_TL );
00085 
00086     // W1TL = C1;
00087     FLA_Copy_internal( C1, W1TL,
00088                        FLA_Cntl_sub_copy( cntl ) );
00089 
00090     // W1TL = inv( triu( T1T ) )' * ( C1 + D1' * E );
00091 
00092     FLA_Gemm_internal( FLA_CONJ_TRANSPOSE, FLA_NO_TRANSPOSE, 
00093                        FLA_ONE, D1, E, FLA_ONE, W1TL,
00094                        FLA_Cntl_sub_gemm1( cntl ) );
00095 
00096     FLA_Trsm_internal( FLA_LEFT, FLA_UPPER_TRIANGULAR,
00097                        FLA_CONJ_TRANSPOSE, FLA_NONUNIT_DIAG,
00098                        FLA_ONE, T1T, W1TL,
00099                        FLA_Cntl_sub_trsm( cntl ) );
00100 
00101     // C1 = C1 - W1TL;
00102     // E  = E  - D1 * W1TL;
00103 
00104     FLA_Axpy_internal( FLA_MINUS_ONE, W1TL, C1,
00105                        FLA_Cntl_sub_axpy( cntl ) );
00106 
00107     FLA_Gemm_internal( FLA_NO_TRANSPOSE, FLA_NO_TRANSPOSE,
00108                        FLA_MINUS_ONE, D1, W1TL, FLA_ONE, E,
00109                        FLA_Cntl_sub_gemm2( cntl ) );
00110 
00111     /*------------------------------------------------------------*/
00112 
00113     FLA_Cont_with_1x3_to_1x2( &DL,  /**/ &DR,        D0, D1, /**/ D2,
00114                               FLA_LEFT );
00115 
00116     FLA_Cont_with_1x3_to_1x2( &TL,  /**/ &TR,        T0, T1, /**/ T2,
00117                               FLA_LEFT );
00118 
00119     FLA_Cont_with_3x1_to_2x1( &CT,                C0, 
00120                                                   C1, 
00121                             /* ** */           /* ** */
00122                               &CB,                C2,     FLA_TOP );
00123   }
00124 
00125   return FLA_SUCCESS;
00126 }

FLA_Error FLA_Apply_Q_UT_UD_lhc_blk_var2 ( FLA_Obj  D,
FLA_Obj  T,
FLA_Obj  W1,
FLA_Obj  C,
FLA_Obj  E,
fla_apqutud_t cntl 
)

References FLA_Apply_Q_UT_UD_internal(), FLA_Cont_with_3x1_to_2x1(), FLA_Determine_blocksize(), FLA_Obj_length(), FLA_Part_2x1(), and FLA_Repart_2x1_to_3x1().

Referenced by FLA_Apply_Q_UT_UD_lhc().

00037 {
00038   FLA_Obj DT,              D0,
00039           DB,              D1,
00040                            D2;
00041 
00042   FLA_Obj TT,              T0,
00043           TB,              T1,
00044                            T2;
00045 
00046   FLA_Obj ET,              E0,
00047           EB,              E1,
00048                            E2;
00049 
00050   dim_t b;
00051 
00052   FLA_Part_2x1( D,    &DT, 
00053                       &DB,            0, FLA_TOP );
00054 
00055   FLA_Part_2x1( T,    &TT, 
00056                       &TB,            0, FLA_TOP );
00057 
00058   FLA_Part_2x1( E,    &ET, 
00059                       &EB,            0, FLA_TOP );
00060 
00061   while ( FLA_Obj_length( DT ) < FLA_Obj_length( D ) ){
00062 
00063     b = FLA_Determine_blocksize( DB, FLA_BOTTOM, FLA_Cntl_blocksize( cntl ) );
00064 
00065     FLA_Repart_2x1_to_3x1( DT,                &D0, 
00066                         /* ** */            /* ** */
00067                                               &D1, 
00068                            DB,                &D2,        b, FLA_BOTTOM );
00069 
00070     FLA_Repart_2x1_to_3x1( TT,                &T0, 
00071                         /* ** */            /* ** */
00072                                               &T1, 
00073                            TB,                &T2,        b, FLA_BOTTOM );
00074 
00075     FLA_Repart_2x1_to_3x1( ET,                &E0, 
00076                         /* ** */            /* ** */
00077                                               &E1, 
00078                            EB,                &E2,        b, FLA_BOTTOM );
00079 
00080     /*------------------------------------------------------------*/
00081 
00082     /*
00083         / C  \ = Q^H / C  \
00084         \ E1 /       \ E1 /
00085                  where Q is formed from D1 and T1.
00086     */
00087 
00088     FLA_Apply_Q_UT_UD_internal( FLA_LEFT, FLA_CONJ_TRANSPOSE, FLA_COLUMNWISE,
00089                                 D1, T1, W1, C,
00090                                             E1, FLA_Cntl_sub_apqutud( cntl ) );
00091 
00092     /*------------------------------------------------------------*/
00093 
00094     FLA_Cont_with_3x1_to_2x1( &DT,                D0, 
00095                                                   D1, 
00096                             /* ** */           /* ** */
00097                               &DB,                D2,     FLA_TOP );
00098 
00099     FLA_Cont_with_3x1_to_2x1( &TT,                T0, 
00100                                                   T1, 
00101                             /* ** */           /* ** */
00102                               &TB,                T2,     FLA_TOP );
00103 
00104     FLA_Cont_with_3x1_to_2x1( &ET,                E0, 
00105                                                   E1, 
00106                             /* ** */           /* ** */
00107                               &EB,                E2,     FLA_TOP );
00108   }
00109 
00110   return FLA_SUCCESS;
00111 }

FLA_Error FLA_Apply_Q_UT_UD_lhc_blk_var3 ( FLA_Obj  D,
FLA_Obj  T,
FLA_Obj  W,
FLA_Obj  C,
FLA_Obj  E,
fla_apqutud_t cntl 
)

References FLA_Apply_Q_UT_UD_internal(), FLA_Cont_with_1x3_to_1x2(), FLA_Determine_blocksize(), FLA_Obj_width(), FLA_Part_1x2(), and FLA_Repart_1x2_to_1x3().

Referenced by FLA_Apply_Q_UT_UD_lhc().

00037 {
00038   FLA_Obj WL,    WR,       W0,  W1,  W2;
00039 
00040   FLA_Obj CL,    CR,       C0,  C1,  C2;
00041 
00042   FLA_Obj EL,    ER,       E0,  E1,  E2;
00043 
00044   dim_t b;
00045 
00046   FLA_Part_1x2( W,    &WL,  &WR,      0, FLA_LEFT );
00047 
00048   FLA_Part_1x2( C,    &CL,  &CR,      0, FLA_LEFT );
00049 
00050   FLA_Part_1x2( E,    &EL,  &ER,      0, FLA_LEFT );
00051 
00052   while ( FLA_Obj_width( CL ) < FLA_Obj_width( C ) ){
00053 
00054     b = FLA_Determine_blocksize( CR, FLA_RIGHT, FLA_Cntl_blocksize( cntl ) );
00055 
00056     FLA_Repart_1x2_to_1x3( WL,  /**/ WR,        &W0, /**/ &W1, &W2,
00057                            b, FLA_RIGHT );
00058 
00059     FLA_Repart_1x2_to_1x3( CL,  /**/ CR,        &C0, /**/ &C1, &C2,
00060                            b, FLA_RIGHT );
00061 
00062     FLA_Repart_1x2_to_1x3( EL,  /**/ ER,        &E0, /**/ &E1, &E2,
00063                            b, FLA_RIGHT );
00064 
00065     /*------------------------------------------------------------*/
00066 
00067     /*
00068         / C1 \ = Q^H / C1 \
00069         \ E1 /       \ E1 /
00070                  where Q is formed from D and T.
00071     */
00072 
00073     FLA_Apply_Q_UT_UD_internal( FLA_LEFT, FLA_CONJ_TRANSPOSE, FLA_COLUMNWISE,
00074                                 D, T, W1, C1,
00075                                           E1, FLA_Cntl_sub_apqutud( cntl ) );
00076 
00077     /*------------------------------------------------------------*/
00078 
00079     FLA_Cont_with_1x3_to_1x2( &WL,  /**/ &WR,        W0, W1, /**/ W2,
00080                               FLA_LEFT );
00081 
00082     FLA_Cont_with_1x3_to_1x2( &CL,  /**/ &CR,        C0, C1, /**/ C2,
00083                               FLA_LEFT );
00084 
00085     FLA_Cont_with_1x3_to_1x2( &EL,  /**/ &ER,        E0, E1, /**/ E2,
00086                               FLA_LEFT );
00087   }
00088 
00089   return FLA_SUCCESS;
00090 }


Generated on Mon Jul 6 05:45:57 2009 for libflame by  doxygen 1.5.9