FLA_LU_nopiv_vars.h File Reference

(r)

Go to the source code of this file.

Functions

FLA_Error FLA_LU_nopiv_blk_var1 (FLA_Obj A, fla_lu_t *cntl)
FLA_Error FLA_LU_nopiv_blk_var2 (FLA_Obj A, fla_lu_t *cntl)
FLA_Error FLA_LU_nopiv_blk_var3 (FLA_Obj A, fla_lu_t *cntl)
FLA_Error FLA_LU_nopiv_blk_var4 (FLA_Obj A, fla_lu_t *cntl)
FLA_Error FLA_LU_nopiv_blk_var5 (FLA_Obj A, fla_lu_t *cntl)
FLA_Error FLA_LU_nopiv_unb_var1 (FLA_Obj A)
FLA_Error FLA_LU_nopiv_unb_var2 (FLA_Obj A)
FLA_Error FLA_LU_nopiv_unb_var3 (FLA_Obj A)
FLA_Error FLA_LU_nopiv_unb_var4 (FLA_Obj A)
FLA_Error FLA_LU_nopiv_unb_var5 (FLA_Obj A)


Function Documentation

FLA_Error FLA_LU_nopiv_blk_var1 ( FLA_Obj  A,
fla_lu_t cntl 
)

References FLA_Cont_with_3x3_to_2x2(), FLA_Determine_blocksize(), FLA_Gemm_internal(), FLA_LU_nopiv_internal(), FLA_MINUS_ONE, FLA_Obj_length(), FLA_Obj_width(), FLA_ONE, FLA_Part_2x2(), FLA_Repart_2x2_to_3x3(), FLA_Trsm_external(), and FLA_Trsm_internal().

Referenced by FLA_LU_nopiv_internal().

00038 {
00039   FLA_Obj ATL,   ATR,      A00, A01, A02, 
00040           ABL,   ABR,      A10, A11, A12,
00041                            A20, A21, A22;
00042 
00043   dim_t b;
00044 
00045   FLA_Part_2x2( A,    &ATL, &ATR,
00046                       &ABL, &ABR,     0, 0, FLA_TL );
00047 
00048 
00049   while ( FLA_Obj_length( ATL ) < FLA_Obj_length( A ) &&
00050           FLA_Obj_width( ATL ) < FLA_Obj_width( A )){
00051 
00052     b = FLA_Determine_blocksize( ABR, FLA_BR, FLA_Cntl_blocksize( cntl ) );
00053 
00054     FLA_Repart_2x2_to_3x3( ATL, /**/ ATR,       &A00, /**/ &A01, &A02,
00055                         /* ************* */   /* ******************** */
00056                                                 &A10, /**/ &A11, &A12,
00057                            ABL, /**/ ABR,       &A20, /**/ &A21, &A22,
00058                            b, b, FLA_BR );
00059 
00060     /*------------------------------------------------------------*/
00061 
00062     /* A01 = trilu( A00 ) \ A01 */ 
00063     FLA_Trsm_internal( FLA_LEFT, FLA_LOWER_TRIANGULAR, 
00064                        FLA_NO_TRANSPOSE, FLA_UNIT_DIAG,
00065                        FLA_ONE, A00, A01,
00066                        FLA_Cntl_sub_trsm1( cntl ) );
00067 
00068     /* A10 = A10 / triu( A00 ) */ 
00069     FLA_Trsm_internal( FLA_RIGHT, FLA_UPPER_TRIANGULAR, 
00070                        FLA_NO_TRANSPOSE, FLA_NONUNIT_DIAG,
00071                        FLA_ONE, A00, A10,
00072                        FLA_Cntl_sub_trsm2( cntl ) );
00073 
00074     /* A11 = LU( A11 - A10 * A01 ) */
00075     FLA_Gemm_internal( FLA_NO_TRANSPOSE, FLA_NO_TRANSPOSE,
00076                        FLA_MINUS_ONE, A10, A01, FLA_ONE, A11,
00077                        FLA_Cntl_sub_gemm1( cntl ) );
00078 
00079     FLA_LU_nopiv_internal( A11, FLA_Cntl_sub_lu( cntl ) );
00080 
00081     /*------------------------------------------------------------*/
00082 
00083     FLA_Cont_with_3x3_to_2x2( &ATL, /**/ &ATR,       A00, A01, /**/ A02,
00084                                                      A10, A11, /**/ A12,
00085                             /* ************** */  /* ****************** */
00086                               &ABL, /**/ &ABR,       A20, A21, /**/ A22,
00087                               FLA_TL );
00088   }
00089 
00090   if ( FLA_Obj_length( ABL ) > 0 )
00091     /* ABL = ABL / triu( ATL ) */ 
00092     FLA_Trsm_external( FLA_RIGHT, FLA_UPPER_TRIANGULAR, 
00093                        FLA_NO_TRANSPOSE, FLA_NONUNIT_DIAG,
00094                        FLA_ONE, ATL, ABL );
00095   else if ( FLA_Obj_width( ATR ) > 0 )
00096     /* ATR = trilu( ATL ) \ ATR */ 
00097     FLA_Trsm_external( FLA_LEFT, FLA_LOWER_TRIANGULAR, 
00098                        FLA_NO_TRANSPOSE, FLA_UNIT_DIAG,
00099                        FLA_ONE, ATL, ATR );
00100 
00101   return FLA_SUCCESS;
00102 }

FLA_Error FLA_LU_nopiv_blk_var2 ( FLA_Obj  A,
fla_lu_t cntl 
)

References FLA_Cont_with_3x3_to_2x2(), FLA_Determine_blocksize(), FLA_Gemm_internal(), FLA_LU_nopiv_internal(), FLA_MINUS_ONE, FLA_Obj_length(), FLA_Obj_width(), FLA_ONE, FLA_Part_2x2(), FLA_Repart_2x2_to_3x3(), FLA_Trsm_external(), and FLA_Trsm_internal().

Referenced by FLA_LU_nopiv_internal().

00038 {
00039   FLA_Obj ATL,   ATR,      A00, A01, A02, 
00040           ABL,   ABR,      A10, A11, A12,
00041                            A20, A21, A22;
00042 
00043   dim_t b;
00044 
00045   FLA_Part_2x2( A,    &ATL, &ATR,
00046                       &ABL, &ABR,     0, 0, FLA_TL );
00047 
00048   while ( FLA_Obj_length( ATL ) < FLA_Obj_length( A ) &&
00049           FLA_Obj_width( ATL ) < FLA_Obj_width( A )){
00050 
00051     b = FLA_Determine_blocksize( ABR, FLA_BR, FLA_Cntl_blocksize( cntl ) );
00052 
00053     FLA_Repart_2x2_to_3x3( ATL, /**/ ATR,       &A00, /**/ &A01, &A02,
00054                         /* ************* */   /* ******************** */
00055                                                 &A10, /**/ &A11, &A12,
00056                            ABL, /**/ ABR,       &A20, /**/ &A21, &A22,
00057                            b, b, FLA_BR );
00058 
00059     /*------------------------------------------------------------*/
00060 
00061     /* A10 = A10 / triu( A00 ) */ 
00062     FLA_Trsm_internal( FLA_RIGHT, FLA_UPPER_TRIANGULAR, 
00063                        FLA_NO_TRANSPOSE, FLA_NONUNIT_DIAG,
00064                        FLA_ONE, A00, A10,
00065                        FLA_Cntl_sub_trsm1( cntl ) );
00066 
00067     /* A11 = LU( A11 - A10 * A01 ) */
00068     FLA_Gemm_internal( FLA_NO_TRANSPOSE, FLA_NO_TRANSPOSE,
00069                        FLA_MINUS_ONE, A10, A01, FLA_ONE, A11,
00070                        FLA_Cntl_sub_gemm1( cntl ) );
00071 
00072     FLA_LU_nopiv_internal( A11, FLA_Cntl_sub_lu( cntl ) );
00073 
00074     /* A12 = A12 - A10 * A02 */
00075     FLA_Gemm_internal( FLA_NO_TRANSPOSE, FLA_NO_TRANSPOSE,
00076                        FLA_MINUS_ONE, A10, A02, FLA_ONE, A12,
00077                        FLA_Cntl_sub_gemm2( cntl ) );
00078 
00079     /* A12 = trilu( A11 ) \ A12 */
00080     FLA_Trsm_internal( FLA_LEFT, FLA_LOWER_TRIANGULAR,
00081                        FLA_NO_TRANSPOSE, FLA_UNIT_DIAG,
00082                        FLA_ONE, A11, A12,
00083                        FLA_Cntl_sub_trsm2( cntl ) );
00084 
00085     /*------------------------------------------------------------*/
00086 
00087     FLA_Cont_with_3x3_to_2x2( &ATL, /**/ &ATR,       A00, A01, /**/ A02,
00088                                                      A10, A11, /**/ A12,
00089                             /* ************** */  /* ****************** */
00090                               &ABL, /**/ &ABR,       A20, A21, /**/ A22,
00091                               FLA_TL );
00092   }
00093 
00094   if ( FLA_Obj_length( ABL ) > 0 )
00095     /* ABL = ABL / triu( ATL ) */ 
00096     FLA_Trsm_external( FLA_RIGHT, FLA_UPPER_TRIANGULAR, 
00097                        FLA_NO_TRANSPOSE, FLA_NONUNIT_DIAG,
00098                        FLA_ONE, ATL, ABL );
00099 
00100   return FLA_SUCCESS;
00101 }

FLA_Error FLA_LU_nopiv_blk_var3 ( FLA_Obj  A,
fla_lu_t cntl 
)

References FLA_Cont_with_3x3_to_2x2(), FLA_Determine_blocksize(), FLA_Gemm_internal(), FLA_LU_nopiv_internal(), FLA_MINUS_ONE, FLA_Obj_length(), FLA_Obj_width(), FLA_ONE, FLA_Part_2x2(), FLA_Repart_2x2_to_3x3(), FLA_Trsm_external(), and FLA_Trsm_internal().

Referenced by FLA_LU_nopiv_internal().

00038 {
00039   FLA_Obj ATL,   ATR,      A00, A01, A02, 
00040           ABL,   ABR,      A10, A11, A12,
00041                            A20, A21, A22;
00042 
00043   dim_t b;
00044 
00045   FLA_Part_2x2( A,    &ATL, &ATR,
00046                       &ABL, &ABR,     0, 0, FLA_TL );
00047 
00048   while ( FLA_Obj_length( ATL ) < FLA_Obj_length( A ) &&
00049           FLA_Obj_width( ATL ) < FLA_Obj_width( A )){
00050 
00051     b = FLA_Determine_blocksize( ABR, FLA_BR, FLA_Cntl_blocksize( cntl ) );
00052 
00053     FLA_Repart_2x2_to_3x3( ATL, /**/ ATR,       &A00, /**/ &A01, &A02,
00054                         /* ************* */   /* ******************** */
00055                                                 &A10, /**/ &A11, &A12,
00056                            ABL, /**/ ABR,       &A20, /**/ &A21, &A22,
00057                            b, b, FLA_BR );
00058 
00059     /*------------------------------------------------------------*/
00060 
00061     /* A01 = trilu( A00 ) \ A10 */ 
00062     FLA_Trsm_internal( FLA_LEFT, FLA_LOWER_TRIANGULAR, 
00063                        FLA_NO_TRANSPOSE, FLA_UNIT_DIAG,
00064                        FLA_ONE, A00, A01,
00065                        FLA_Cntl_sub_trsm1( cntl ) );
00066 
00067     /* A11 = LU( A11 - A10 * A01 ) */
00068     FLA_Gemm_internal( FLA_NO_TRANSPOSE, FLA_NO_TRANSPOSE,
00069                        FLA_MINUS_ONE, A10, A01, FLA_ONE, A11,
00070                        FLA_Cntl_sub_gemm1( cntl ) );
00071 
00072     FLA_LU_nopiv_internal( A11, FLA_Cntl_sub_lu( cntl ) );
00073 
00074     /* A21 = A21 - A20 * A01 */
00075     FLA_Gemm_internal( FLA_NO_TRANSPOSE, FLA_NO_TRANSPOSE,
00076                        FLA_MINUS_ONE, A20, A01, FLA_ONE, A21,
00077                        FLA_Cntl_sub_gemm2( cntl ) );
00078 
00079     /* A21 = A21 / triu( A11 ) */ 
00080     FLA_Trsm_internal( FLA_RIGHT, FLA_UPPER_TRIANGULAR, 
00081                        FLA_NO_TRANSPOSE, FLA_NONUNIT_DIAG,
00082                        FLA_ONE, A11, A21,
00083                        FLA_Cntl_sub_trsm2( cntl ) );
00084 
00085     /*------------------------------------------------------------*/
00086 
00087     FLA_Cont_with_3x3_to_2x2( &ATL, /**/ &ATR,       A00, A01, /**/ A02,
00088                                                      A10, A11, /**/ A12,
00089                             /* ************** */  /* ****************** */
00090                               &ABL, /**/ &ABR,       A20, A21, /**/ A22,
00091                               FLA_TL );
00092   }
00093 
00094   if ( FLA_Obj_width( ATR ) > 0 )
00095     /* ATR = trilu( ATL ) \ ATR */ 
00096     FLA_Trsm_external( FLA_LEFT, FLA_LOWER_TRIANGULAR, 
00097                        FLA_NO_TRANSPOSE, FLA_UNIT_DIAG,
00098                        FLA_ONE, ATL, ATR );
00099 
00100   return FLA_SUCCESS;
00101 }

FLA_Error FLA_LU_nopiv_blk_var4 ( FLA_Obj  A,
fla_lu_t cntl 
)

References FLA_Cont_with_3x3_to_2x2(), FLA_Determine_blocksize(), FLA_Gemm_internal(), FLA_LU_nopiv_internal(), FLA_MINUS_ONE, FLA_Obj_length(), FLA_Obj_width(), FLA_ONE, FLA_Part_2x2(), FLA_Repart_2x2_to_3x3(), and FLA_Trsm_internal().

Referenced by FLA_LU_nopiv_internal().

00038 {
00039   FLA_Obj ATL,   ATR,      A00, A01, A02, 
00040           ABL,   ABR,      A10, A11, A12,
00041                            A20, A21, A22;
00042 
00043   dim_t b;
00044 
00045   FLA_Part_2x2( A,    &ATL, &ATR,
00046                       &ABL, &ABR,     0, 0, FLA_TL );
00047 
00048   while ( FLA_Obj_length( ATL ) < FLA_Obj_length( A ) &&
00049           FLA_Obj_width( ATL ) < FLA_Obj_width( A )){
00050 
00051     b = FLA_Determine_blocksize( ABR, FLA_BR, FLA_Cntl_blocksize( cntl ) );
00052 
00053     FLA_Repart_2x2_to_3x3( ATL, /**/ ATR,       &A00, /**/ &A01, &A02,
00054                         /* ************* */   /* ******************** */
00055                                                 &A10, /**/ &A11, &A12,
00056                            ABL, /**/ ABR,       &A20, /**/ &A21, &A22,
00057                            b, b, FLA_BR );
00058 
00059     /*------------------------------------------------------------*/
00060 
00061     /* A11 = LU( A11 - A10 * A01 ) */
00062     FLA_Gemm_internal( FLA_NO_TRANSPOSE, FLA_NO_TRANSPOSE,
00063                        FLA_MINUS_ONE, A10, A01, FLA_ONE, A11,
00064                        FLA_Cntl_sub_gemm1( cntl ) );
00065 
00066     FLA_LU_nopiv_internal( A11, FLA_Cntl_sub_lu( cntl ) );
00067 
00068     /* A12 = A12 - A10 * A02 */
00069     FLA_Gemm_internal( FLA_NO_TRANSPOSE, FLA_NO_TRANSPOSE,
00070                        FLA_MINUS_ONE, A10, A02, FLA_ONE, A12,
00071                        FLA_Cntl_sub_gemm2( cntl ) );
00072 
00073     /* A12 = trilu( A11 ) \ A12 */ 
00074     FLA_Trsm_internal( FLA_LEFT, FLA_LOWER_TRIANGULAR, 
00075                        FLA_NO_TRANSPOSE, FLA_UNIT_DIAG,
00076                        FLA_ONE, A11, A12,
00077                        FLA_Cntl_sub_trsm1( cntl ) );
00078 
00079     /* A21 = A21 - A20 * A01 */
00080     FLA_Gemm_internal( FLA_NO_TRANSPOSE, FLA_NO_TRANSPOSE,
00081                        FLA_MINUS_ONE, A20, A01, FLA_ONE, A21,
00082                        FLA_Cntl_sub_gemm3( cntl ) );
00083 
00084     /* A21 = A21 / triu( A11 ) */ 
00085     FLA_Trsm_internal( FLA_RIGHT, FLA_UPPER_TRIANGULAR, 
00086                        FLA_NO_TRANSPOSE, FLA_NONUNIT_DIAG,
00087                        FLA_ONE, A11, A21,
00088                        FLA_Cntl_sub_trsm2( cntl ) );
00089 
00090     /*------------------------------------------------------------*/
00091 
00092     FLA_Cont_with_3x3_to_2x2( &ATL, /**/ &ATR,       A00, A01, /**/ A02,
00093                                                      A10, A11, /**/ A12,
00094                             /* ************** */  /* ****************** */
00095                               &ABL, /**/ &ABR,       A20, A21, /**/ A22,
00096                               FLA_TL );
00097   }
00098 
00099   return FLA_SUCCESS;
00100 }

FLA_Error FLA_LU_nopiv_blk_var5 ( FLA_Obj  A,
fla_lu_t cntl 
)

References FLA_Cont_with_3x3_to_2x2(), FLA_Determine_blocksize(), FLA_Gemm_internal(), FLA_LU_nopiv_internal(), FLA_MINUS_ONE, FLA_Obj_length(), FLA_Obj_width(), FLA_ONE, FLA_Part_2x2(), FLA_Repart_2x2_to_3x3(), and FLA_Trsm_internal().

Referenced by FLA_LU_nopiv_internal().

00036 {
00037   FLA_Obj ATL,   ATR,      A00, A01, A02, 
00038           ABL,   ABR,      A10, A11, A12,
00039                            A20, A21, A22;
00040 
00041   dim_t b;
00042 
00043   FLA_Part_2x2( A,    &ATL, &ATR,
00044                       &ABL, &ABR,     0, 0, FLA_TL );
00045 
00046   while ( FLA_Obj_length( ATL ) < FLA_Obj_length( A ) &&
00047           FLA_Obj_width( ATL ) < FLA_Obj_width( A )){
00048 
00049     b = FLA_Determine_blocksize( ABR, FLA_BR, FLA_Cntl_blocksize( cntl ) );
00050 
00051     FLA_Repart_2x2_to_3x3( ATL, /**/ ATR,       &A00, /**/ &A01, &A02,
00052                         /* ************* */   /* ******************** */
00053                                                 &A10, /**/ &A11, &A12,
00054                            ABL, /**/ ABR,       &A20, /**/ &A21, &A22,
00055                            b, b, FLA_BR );
00056 
00057     /*------------------------------------------------------------*/
00058 
00059     /* A11 = LU( A11 ) */
00060     FLA_LU_nopiv_internal( A11, FLA_Cntl_sub_lu( cntl ) );
00061 
00062     /* A12 = trilu( A11 ) \ A12 */ 
00063     FLA_Trsm_internal( FLA_LEFT, FLA_LOWER_TRIANGULAR, 
00064                        FLA_NO_TRANSPOSE, FLA_UNIT_DIAG,
00065                        FLA_ONE, A11, A12,
00066                        FLA_Cntl_sub_trsm1( cntl ) );
00067 
00068     /* A21 = A21 / triu( A11 ) */ 
00069     FLA_Trsm_internal( FLA_RIGHT, FLA_UPPER_TRIANGULAR, 
00070                        FLA_NO_TRANSPOSE, FLA_NONUNIT_DIAG,
00071                        FLA_ONE, A11, A21,
00072                        FLA_Cntl_sub_trsm2( cntl ) );
00073 
00074     /* A22 = A22 - A21 * A12 */
00075     FLA_Gemm_internal( FLA_NO_TRANSPOSE, FLA_NO_TRANSPOSE,
00076                        FLA_MINUS_ONE, A21, A12, FLA_ONE, A22,
00077                        FLA_Cntl_sub_gemm1( cntl ) );
00078 
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 
00089   return FLA_SUCCESS;
00090 }

FLA_Error FLA_LU_nopiv_unb_var1 ( FLA_Obj  A  ) 

References FLA_Cont_with_3x3_to_2x2(), FLA_Dots_external(), FLA_MINUS_ONE, FLA_Obj_length(), FLA_Obj_width(), FLA_ONE, FLA_Part_2x2(), FLA_Repart_2x2_to_3x3(), FLA_Trsm_external(), and FLA_Trsv_external().

00038 {
00039   FLA_Obj ATL,   ATR,      A00,  a01,     A02, 
00040           ABL,   ABR,      a10t, alpha11, a12t,
00041                            A20,  a21,     A22;
00042 
00043   FLA_Part_2x2( A,    &ATL, &ATR,
00044                       &ABL, &ABR,     0, 0, FLA_TL );
00045 
00046   while ( FLA_Obj_length( ATL ) < FLA_Obj_length( A ) &&
00047           FLA_Obj_width( ATL ) < FLA_Obj_width( A )){
00048 
00049     FLA_Repart_2x2_to_3x3( ATL, /**/ ATR,       &A00,  /**/ &a01,     &A02,
00050                         /* ************* */   /* ************************** */
00051                                                 &a10t, /**/ &alpha11, &a12t,
00052                            ABL, /**/ ABR,       &A20,  /**/ &a21,     &A22,
00053                            1, 1, FLA_BR );
00054 
00055     /*------------------------------------------------------------*/
00056 
00057     /* a01 = trilu( A00 ) \ a01 */ 
00058     FLA_Trsv_external( FLA_LOWER_TRIANGULAR, FLA_NO_TRANSPOSE, FLA_UNIT_DIAG, A00, a01 );
00059 
00060     /* a10t = a10t / triu( A00 ) */ 
00061     FLA_Trsv_external( FLA_UPPER_TRIANGULAR, FLA_TRANSPOSE, FLA_NONUNIT_DIAG, A00, a10t );
00062 
00063     /* alpha11 = alpha11 - a10t * a01 */
00064     FLA_Dots_external( FLA_MINUS_ONE, a10t, a01, FLA_ONE, alpha11 );
00065 
00066     /*------------------------------------------------------------*/
00067 
00068     FLA_Cont_with_3x3_to_2x2( &ATL, /**/ &ATR,       A00,  a01,     /**/ A02,
00069                                                      a10t, alpha11, /**/ a12t,
00070                             /* ************** */  /* ************************ */
00071                               &ABL, /**/ &ABR,       A20,  a21,     /**/ A22,
00072                               FLA_TL );
00073 
00074   }
00075 
00076   if ( FLA_Obj_length( ABL ) > 0 )
00077     /* ABL = ABL / triu( ATL ) */ 
00078     FLA_Trsm_external( FLA_RIGHT, FLA_UPPER_TRIANGULAR, 
00079                        FLA_NO_TRANSPOSE, FLA_NONUNIT_DIAG,
00080                        FLA_ONE, ATL, ABL );
00081   else if ( FLA_Obj_width( ATR ) > 0 )
00082     /* ATR = trilu( ATL ) \ ATR */ 
00083     FLA_Trsm_external( FLA_LEFT, FLA_LOWER_TRIANGULAR, 
00084                        FLA_NO_TRANSPOSE, FLA_UNIT_DIAG,
00085                        FLA_ONE, ATL, ATR );
00086 
00087   return FLA_SUCCESS;
00088 }

FLA_Error FLA_LU_nopiv_unb_var2 ( FLA_Obj  A  ) 

References FLA_Cont_with_3x3_to_2x2(), FLA_Dots_external(), FLA_Gemv_external(), FLA_MINUS_ONE, FLA_Obj_length(), FLA_Obj_width(), FLA_ONE, FLA_Part_2x2(), FLA_Repart_2x2_to_3x3(), FLA_Trsm_external(), and FLA_Trsv_external().

00038 {
00039   FLA_Obj ATL,   ATR,      A00,  a01,     A02, 
00040           ABL,   ABR,      a10t, alpha11, a12t,
00041                            A20,  a21,     A22;
00042 
00043   FLA_Part_2x2( A,    &ATL, &ATR,
00044                       &ABL, &ABR,     0, 0, FLA_TL );
00045 
00046   while ( FLA_Obj_length( ATL ) < FLA_Obj_length( A ) &&
00047           FLA_Obj_width( ATL ) < FLA_Obj_width( A )){
00048 
00049     FLA_Repart_2x2_to_3x3( ATL, /**/ ATR,       &A00,  /**/ &a01,     &A02,
00050                         /* ************* */   /* ************************** */
00051                                                 &a10t, /**/ &alpha11, &a12t,
00052                            ABL, /**/ ABR,       &A20,  /**/ &a21,     &A22,
00053                            1, 1, FLA_BR );
00054 
00055     /*------------------------------------------------------------*/
00056 
00057     /* a10t = a10t / triu( A00 ) */ 
00058     FLA_Trsv_external( FLA_UPPER_TRIANGULAR, FLA_TRANSPOSE, FLA_NONUNIT_DIAG, A00, a10t );
00059 
00060     /* alpha11 = alpha11 - a10t * a01 */
00061     FLA_Dots_external( FLA_MINUS_ONE, a10t, a01, FLA_ONE, alpha11 );
00062 
00063     /* a12t = a12t - a10t * A02 */
00064     FLA_Gemv_external( FLA_TRANSPOSE, FLA_MINUS_ONE, A02, a10t, FLA_ONE, a12t );
00065 
00066     /*------------------------------------------------------------*/
00067 
00068     FLA_Cont_with_3x3_to_2x2( &ATL, /**/ &ATR,       A00,  a01,     /**/ A02,
00069                                                      a10t, alpha11, /**/ a12t,
00070                             /* ************** */  /* ************************ */
00071                               &ABL, /**/ &ABR,       A20,  a21,     /**/ A22,
00072                               FLA_TL );
00073 
00074   }
00075 
00076   if ( FLA_Obj_length( ABL ) > 0 )
00077     /* ABL = ABL / triu( ATL ) */ 
00078     FLA_Trsm_external( FLA_RIGHT, FLA_UPPER_TRIANGULAR, 
00079                        FLA_NO_TRANSPOSE, FLA_NONUNIT_DIAG,
00080                        FLA_ONE, ATL, ABL );
00081 
00082   return FLA_SUCCESS;
00083 }

FLA_Error FLA_LU_nopiv_unb_var3 ( FLA_Obj  A  ) 

References FLA_Cont_with_3x3_to_2x2(), FLA_Dots_external(), FLA_Gemv_external(), FLA_Inv_scal_external(), FLA_MINUS_ONE, FLA_Obj_length(), FLA_Obj_width(), FLA_ONE, FLA_Part_2x2(), FLA_Repart_2x2_to_3x3(), FLA_Trsm_external(), and FLA_Trsv_external().

00038 {
00039   FLA_Obj ATL,   ATR,      A00,  a01,     A02, 
00040           ABL,   ABR,      a10t, alpha11, a12t,
00041                            A20,  a21,     A22;
00042 
00043   FLA_Part_2x2( A,    &ATL, &ATR,
00044                       &ABL, &ABR,     0, 0, FLA_TL );
00045 
00046   while ( FLA_Obj_length( ATL ) < FLA_Obj_length( A ) &&
00047           FLA_Obj_width( ATL ) < FLA_Obj_width( A )){
00048 
00049     FLA_Repart_2x2_to_3x3( ATL, /**/ ATR,       &A00,  /**/ &a01,     &A02,
00050                         /* ************* */   /* ************************** */
00051                                                 &a10t, /**/ &alpha11, &a12t,
00052                            ABL, /**/ ABR,       &A20,  /**/ &a21,     &A22,
00053                            1, 1, FLA_BR );
00054 
00055     /*------------------------------------------------------------*/
00056 
00057     /* a01 = trilu( A00 ) \ a10 */ 
00058     FLA_Trsv_external( FLA_LOWER_TRIANGULAR, FLA_NO_TRANSPOSE, FLA_UNIT_DIAG, A00, a01 );
00059 
00060     /* alpha11 = alpha11 - a10t * a01 */
00061     FLA_Dots_external( FLA_MINUS_ONE, a10t, a01, FLA_ONE, alpha11 );
00062 
00063     /* a21 = a21 - A20 * a01 */
00064     FLA_Gemv_external( FLA_NO_TRANSPOSE, FLA_MINUS_ONE, A20, a01, FLA_ONE, a21 );
00065 
00066     /* a21 = a21 / alpha11 */ 
00067     FLA_Inv_scal_external( alpha11, a21 );
00068 
00069     /*------------------------------------------------------------*/
00070 
00071     FLA_Cont_with_3x3_to_2x2( &ATL, /**/ &ATR,       A00,  a01,     /**/ A02,
00072                                                      a10t, alpha11, /**/ a12t,
00073                             /* ************** */  /* ************************ */
00074                               &ABL, /**/ &ABR,       A20,  a21,     /**/ A22,
00075                               FLA_TL );
00076 
00077   }
00078 
00079   if ( FLA_Obj_width( ATR ) > 0 )
00080     /* ATR = trilu( ATL ) \ ATR */ 
00081     FLA_Trsm_external( FLA_LEFT, FLA_LOWER_TRIANGULAR, 
00082                        FLA_NO_TRANSPOSE, FLA_UNIT_DIAG,
00083                        FLA_ONE, ATL, ATR );
00084 
00085   return FLA_SUCCESS;
00086 }

FLA_Error FLA_LU_nopiv_unb_var4 ( FLA_Obj  A  ) 

References FLA_Cont_with_3x3_to_2x2(), FLA_Dots_external(), FLA_Gemv_external(), FLA_Inv_scal_external(), FLA_MINUS_ONE, FLA_Obj_length(), FLA_Obj_width(), FLA_ONE, FLA_Part_2x2(), and FLA_Repart_2x2_to_3x3().

00038 {
00039   FLA_Obj ATL,   ATR,      A00,  a01,     A02, 
00040           ABL,   ABR,      a10t, alpha11, a12t,
00041                            A20,  a21,     A22;
00042 
00043   FLA_Part_2x2( A,    &ATL, &ATR,
00044                       &ABL, &ABR,     0, 0, FLA_TL );
00045 
00046   while ( FLA_Obj_length( ATL ) < FLA_Obj_length( A ) &&
00047           FLA_Obj_width( ATL ) < FLA_Obj_width( A )){
00048 
00049     FLA_Repart_2x2_to_3x3( ATL, /**/ ATR,       &A00,  /**/ &a01,     &A02,
00050                         /* ************* */   /* ************************** */
00051                                                 &a10t, /**/ &alpha11, &a12t,
00052                            ABL, /**/ ABR,       &A20,  /**/ &a21,     &A22,
00053                            1, 1, FLA_BR );
00054 
00055     /*------------------------------------------------------------*/
00056 
00057     /* alpha11 = alpha11 - a10t * a01 */
00058     FLA_Dots_external( FLA_MINUS_ONE, a10t, a01, FLA_ONE, alpha11 );
00059 
00060     /* a12t = a12t - a10t * A02 */
00061     FLA_Gemv_external( FLA_TRANSPOSE, FLA_MINUS_ONE, A02, a10t, FLA_ONE, a12t );
00062 
00063     /* a21 = a21 - A20 * a01 */
00064     FLA_Gemv_external( FLA_NO_TRANSPOSE, FLA_MINUS_ONE, A20, a01, FLA_ONE, a21 );
00065 
00066     /* a21 = a21 / alpha11 */ 
00067     FLA_Inv_scal_external( alpha11, a21 );
00068 
00069     /*------------------------------------------------------------*/
00070 
00071     FLA_Cont_with_3x3_to_2x2( &ATL, /**/ &ATR,       A00,  a01,     /**/ A02,
00072                                                      a10t, alpha11, /**/ a12t,
00073                             /* ************** */  /* ************************ */
00074                               &ABL, /**/ &ABR,       A20,  a21,     /**/ A22,
00075                               FLA_TL );
00076 
00077   }
00078 
00079   return FLA_SUCCESS;
00080 }

FLA_Error FLA_LU_nopiv_unb_var5 ( FLA_Obj  A  ) 

References FLA_Cont_with_3x3_to_2x2(), FLA_Ger_external(), FLA_Inv_scal_external(), FLA_MINUS_ONE, FLA_Obj_length(), FLA_Obj_width(), FLA_Part_2x2(), and FLA_Repart_2x2_to_3x3().

00038 {
00039   FLA_Obj ATL,   ATR,      A00,  a01,     A02, 
00040           ABL,   ABR,      a10t, alpha11, a12t,
00041                            A20,  a21,     A22;
00042 
00043   FLA_Part_2x2( A,    &ATL, &ATR,
00044                       &ABL, &ABR,     0, 0, FLA_TL );
00045 
00046   while ( FLA_Obj_length( ATL ) < FLA_Obj_length( A ) &&
00047           FLA_Obj_width( ATL ) < FLA_Obj_width( A )){
00048 
00049     FLA_Repart_2x2_to_3x3( ATL, /**/ ATR,       &A00,  /**/ &a01,     &A02,
00050                         /* ************* */   /* ************************** */
00051                                                 &a10t, /**/ &alpha11, &a12t,
00052                            ABL, /**/ ABR,       &A20,  /**/ &a21,     &A22,
00053                            1, 1, FLA_BR );
00054 
00055     /*------------------------------------------------------------*/
00056 
00057     /* a21 = a21 / alpha11 */ 
00058     FLA_Inv_scal_external( alpha11, a21 );
00059 
00060     /* A22 = A22 - a21 * a12t */
00061     FLA_Ger_external( FLA_MINUS_ONE, a21, a12t, A22 );
00062 
00063     /*------------------------------------------------------------*/
00064 
00065     FLA_Cont_with_3x3_to_2x2( &ATL, /**/ &ATR,       A00,  a01,     /**/ A02,
00066                                                      a10t, alpha11, /**/ a12t,
00067                             /* ************** */  /* ************************ */
00068                               &ABL, /**/ &ABR,       A20,  a21,     /**/ A22,
00069                               FLA_TL );
00070 
00071   }
00072 
00073   return FLA_SUCCESS;
00074 }


Generated on Mon Jul 6 05:45:56 2009 for libflame by  doxygen 1.5.9