FLA_Trsv_uc.h File Reference

(r)

Go to the source code of this file.

Functions

FLA_Error FLA_Trsv_uc_blk_var1 (FLA_Diag diagA, FLA_Obj A, FLA_Obj x, fla_trsv_t *cntl)
FLA_Error FLA_Trsv_uc_blk_var2 (FLA_Diag diagA, FLA_Obj A, FLA_Obj x, fla_trsv_t *cntl)


Function Documentation

FLA_Error FLA_Trsv_uc_blk_var1 ( FLA_Diag  diagA,
FLA_Obj  A,
FLA_Obj  x,
fla_trsv_t cntl 
)

References FLA_Cont_with_3x1_to_2x1(), FLA_Cont_with_3x3_to_2x2(), FLA_Determine_blocksize(), FLA_Gemv_internal(), FLA_MINUS_ONE, FLA_Obj_length(), FLA_ONE, FLA_Part_2x1(), FLA_Part_2x2(), FLA_Repart_2x1_to_3x1(), FLA_Repart_2x2_to_3x3(), and FLA_Trsv_internal().

Referenced by FLA_Trsv_uc().

00038 {
00039   FLA_Obj ATL,   ATR,      A00, A01, A02, 
00040           ABL,   ABR,      A10, A11, A12,
00041                            A20, A21, A22;
00042 
00043   FLA_Obj xT,              x0,
00044           xB,              x1,
00045                            x2;
00046 
00047   dim_t b;
00048 
00049   FLA_Part_2x2( A,    &ATL, &ATR,
00050                       &ABL, &ABR,     0, 0, FLA_TL );
00051 
00052   FLA_Part_2x1( x,    &xT, 
00053                       &xB,            0, FLA_TOP );
00054 
00055   while ( FLA_Obj_length( ATL ) < FLA_Obj_length( A ) ){
00056 
00057     b = FLA_Determine_blocksize( ABR, FLA_BR, FLA_Cntl_blocksize( cntl ) );
00058 
00059     FLA_Repart_2x2_to_3x3( ATL, /**/ ATR,       &A00, /**/ &A01, &A02,
00060                         /* ************* */   /* ******************** */
00061                                                 &A10, /**/ &A11, &A12,
00062                            ABL, /**/ ABR,       &A20, /**/ &A21, &A22,
00063                            b, b, FLA_BR );
00064 
00065     FLA_Repart_2x1_to_3x1( xT,                &x0, 
00066                         /* ** */            /* ** */
00067                                               &x1, 
00068                            xB,                &x2,        b, FLA_BOTTOM );
00069 
00070     /*------------------------------------------------------------*/
00071 
00072     /* x1 = x1 - A01' * x0 */
00073     FLA_Gemv_internal( FLA_CONJ_TRANSPOSE,
00074                        FLA_MINUS_ONE, A01, x0, FLA_ONE, x1,
00075                        FLA_Cntl_sub_gemv( cntl ) );
00076 
00077     /* x1 = triu( A11' ) \ x1 */
00078     FLA_Trsv_internal( FLA_UPPER_TRIANGULAR, FLA_CONJ_TRANSPOSE, diagA,
00079                        A11, x1,
00080                        FLA_Cntl_sub_trsv( cntl ) );
00081 
00082     /*------------------------------------------------------------*/
00083 
00084     FLA_Cont_with_3x3_to_2x2( &ATL, /**/ &ATR,       A00, A01, /**/ A02,
00085                                                      A10, A11, /**/ A12,
00086                             /* ************** */  /* ****************** */
00087                               &ABL, /**/ &ABR,       A20, A21, /**/ A22,
00088                               FLA_TL );
00089 
00090     FLA_Cont_with_3x1_to_2x1( &xT,                x0, 
00091                                                   x1, 
00092                             /* ** */           /* ** */
00093                               &xB,                x2,     FLA_TOP );
00094 
00095   }
00096 
00097   return FLA_SUCCESS;
00098 }

FLA_Error FLA_Trsv_uc_blk_var2 ( FLA_Diag  diagA,
FLA_Obj  A,
FLA_Obj  x,
fla_trsv_t cntl 
)

References FLA_Cont_with_3x1_to_2x1(), FLA_Cont_with_3x3_to_2x2(), FLA_Determine_blocksize(), FLA_Gemv_internal(), FLA_MINUS_ONE, FLA_Obj_length(), FLA_ONE, FLA_Part_2x1(), FLA_Part_2x2(), FLA_Repart_2x1_to_3x1(), FLA_Repart_2x2_to_3x3(), and FLA_Trsv_internal().

Referenced by FLA_Trsv_uc().

00036 {
00037   FLA_Obj ATL,   ATR,      A00, A01, A02, 
00038           ABL,   ABR,      A10, A11, A12,
00039                            A20, A21, A22;
00040 
00041   FLA_Obj xT,              x0,
00042           xB,              x1,
00043                            x2;
00044 
00045   dim_t b;
00046 
00047   FLA_Part_2x2( A,    &ATL, &ATR,
00048                       &ABL, &ABR,     0, 0, FLA_TL );
00049 
00050   FLA_Part_2x1( x,    &xT, 
00051                       &xB,            0, FLA_TOP );
00052 
00053   while ( FLA_Obj_length( ATL ) < FLA_Obj_length( A ) ){
00054 
00055     b = FLA_Determine_blocksize( ABR, FLA_BR, FLA_Cntl_blocksize( cntl ) );
00056 
00057     FLA_Repart_2x2_to_3x3( ATL, /**/ ATR,       &A00, /**/ &A01, &A02,
00058                         /* ************* */   /* ******************** */
00059                                                 &A10, /**/ &A11, &A12,
00060                            ABL, /**/ ABR,       &A20, /**/ &A21, &A22,
00061                            b, b, FLA_BR );
00062 
00063     FLA_Repart_2x1_to_3x1( xT,                &x0, 
00064                         /* ** */            /* ** */
00065                                               &x1, 
00066                            xB,                &x2,        b, FLA_BOTTOM );
00067 
00068     /*------------------------------------------------------------*/
00069 
00070     /* x1 = triu( A11' ) \ x1 */
00071     FLA_Trsv_internal( FLA_UPPER_TRIANGULAR, FLA_CONJ_TRANSPOSE, diagA,
00072                        A11, x1,
00073                        FLA_Cntl_sub_trsv( cntl ) );
00074 
00075     /* x2 = x2 - A12' * x1 */
00076     FLA_Gemv_internal( FLA_CONJ_TRANSPOSE,
00077                        FLA_MINUS_ONE, A12, x1, FLA_ONE, x2,
00078                        FLA_Cntl_sub_gemv( cntl ) );
00079 
00080     /*------------------------------------------------------------*/
00081 
00082     FLA_Cont_with_3x3_to_2x2( &ATL, /**/ &ATR,       A00, A01, /**/ A02,
00083                                                      A10, A11, /**/ A12,
00084                             /* ************** */  /* ****************** */
00085                               &ABL, /**/ &ABR,       A20, A21, /**/ A22,
00086                               FLA_TL );
00087 
00088     FLA_Cont_with_3x1_to_2x1( &xT,                x0, 
00089                                                   x1, 
00090                             /* ** */           /* ** */
00091                               &xB,                x2,     FLA_TOP );
00092 
00093   }
00094 
00095   return FLA_SUCCESS;
00096 }


Generated on Mon Jul 6 05:45:53 2009 for libflame by  doxygen 1.5.9