FLA_Ttmm_l.h File Reference

(r)

Go to the source code of this file.

Functions

FLA_Error FLA_Ttmm_l_blk_var1 (FLA_Obj A, fla_ttmm_t *cntl)
FLA_Error FLA_Ttmm_l_blk_var2 (FLA_Obj A, fla_ttmm_t *cntl)
FLA_Error FLA_Ttmm_l_blk_var3 (FLA_Obj A, fla_ttmm_t *cntl)
FLA_Error FLA_Ttmm_l_unb_var1 (FLA_Obj A)
FLA_Error FLA_Ttmm_l_unb_var2 (FLA_Obj A)
FLA_Error FLA_Ttmm_l_unb_var3 (FLA_Obj A)


Function Documentation

FLA_Error FLA_Ttmm_l_blk_var1 ( FLA_Obj  A,
fla_ttmm_t cntl 
)

References FLA_Cont_with_3x3_to_2x2(), FLA_Determine_blocksize(), FLA_Herk_internal(), FLA_Obj_is_real(), FLA_Obj_length(), FLA_ONE, FLA_Part_2x2(), FLA_Repart_2x2_to_3x3(), FLA_Syrk_internal(), FLA_Trmm_internal(), and FLA_Ttmm_internal().

Referenced by FLA_Ttmm_l().

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 
00048     b = FLA_Determine_blocksize( ABR, FLA_BR, FLA_Cntl_blocksize( cntl ) );
00049 
00050     FLA_Repart_2x2_to_3x3( ATL, /**/ ATR,       &A00, /**/ &A01, &A02,
00051                         /* ************* */   /* ******************** */
00052                                                 &A10, /**/ &A11, &A12,
00053                            ABL, /**/ ABR,       &A20, /**/ &A21, &A22,
00054                            b, b, FLA_BR );
00055 
00056     /*------------------------------------------------------------*/
00057 
00058     if ( FLA_Obj_is_real( A ) )
00059     {
00060       /* A00 = A00 + A10' * A10; */
00061       FLA_Syrk_internal( FLA_LOWER_TRIANGULAR, FLA_TRANSPOSE,
00062                          FLA_ONE, A10, FLA_ONE, A00,
00063                          FLA_Cntl_sub_syrk( cntl ) );
00064 
00065       /* A10 = tril( A11 )' * A10; */
00066       FLA_Trmm_internal( FLA_LEFT, FLA_LOWER_TRIANGULAR, FLA_TRANSPOSE, FLA_NONUNIT_DIAG,
00067                          FLA_ONE, A11, A10,
00068                          FLA_Cntl_sub_trmm( cntl ) );
00069 
00070       /* A11 = tril( A11 )' * tril( A11 ); */
00071       FLA_Ttmm_internal( FLA_LOWER_TRIANGULAR, A11,
00072                          FLA_Cntl_sub_ttmm( cntl ) );
00073     }
00074     else
00075     {
00076       /* A00 = A00 + A10' * A10; */
00077       FLA_Herk_internal( FLA_LOWER_TRIANGULAR, FLA_CONJ_TRANSPOSE,
00078                          FLA_ONE, A10, FLA_ONE, A00,
00079                          FLA_Cntl_sub_herk( cntl ) );
00080 
00081       /* A10 = tril( A11 )' * A10; */
00082       FLA_Trmm_internal( FLA_LEFT, FLA_LOWER_TRIANGULAR, FLA_CONJ_TRANSPOSE, FLA_NONUNIT_DIAG,
00083                          FLA_ONE, A11, A10,
00084                          FLA_Cntl_sub_trmm( cntl ) );
00085 
00086       /* A11 = tril( A11 )' * tril( A11 ); */
00087       FLA_Ttmm_internal( FLA_LOWER_TRIANGULAR, A11,
00088                          FLA_Cntl_sub_ttmm( cntl ) );
00089     }
00090 
00091     /*------------------------------------------------------------*/
00092 
00093     FLA_Cont_with_3x3_to_2x2( &ATL, /**/ &ATR,       A00, A01, /**/ A02,
00094                                                      A10, A11, /**/ A12,
00095                             /* ************** */  /* ****************** */
00096                               &ABL, /**/ &ABR,       A20, A21, /**/ A22,
00097                               FLA_TL );
00098 
00099   }
00100 
00101   return FLA_SUCCESS;
00102 }

FLA_Error FLA_Ttmm_l_blk_var2 ( FLA_Obj  A,
fla_ttmm_t cntl 
)

References FLA_Cont_with_3x3_to_2x2(), FLA_Determine_blocksize(), FLA_Gemm_internal(), FLA_Herk_internal(), FLA_Obj_is_real(), FLA_Obj_length(), FLA_ONE, FLA_Part_2x2(), FLA_Repart_2x2_to_3x3(), FLA_Syrk_internal(), FLA_Trmm_internal(), and FLA_Ttmm_internal().

Referenced by FLA_Ttmm_l().

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 
00050     b = FLA_Determine_blocksize( ABR, FLA_BR, FLA_Cntl_blocksize( cntl ) );
00051 
00052     FLA_Repart_2x2_to_3x3( ATL, /**/ ATR,       &A00, /**/ &A01, &A02,
00053                         /* ************* */   /* ******************** */
00054                                                 &A10, /**/ &A11, &A12,
00055                            ABL, /**/ ABR,       &A20, /**/ &A21, &A22,
00056                            b, b, FLA_BR );
00057 
00058     /*------------------------------------------------------------*/
00059 
00060     if ( FLA_Obj_is_real( A ) )
00061     {
00062       /* A10 = tril( A11 )' * A10; */
00063       FLA_Trmm_internal( FLA_LEFT, FLA_LOWER_TRIANGULAR, FLA_TRANSPOSE, FLA_NONUNIT_DIAG,
00064                          FLA_ONE, A11, A10,
00065                          FLA_Cntl_sub_trmm( cntl ) );
00066 
00067       /* A10 = A10 + A21' * A20; */
00068       FLA_Gemm_internal( FLA_TRANSPOSE, FLA_NO_TRANSPOSE,
00069                          FLA_ONE, A21, A20, FLA_ONE, A10,
00070                          FLA_Cntl_sub_gemm( cntl ) );
00071 
00072       /* A11 = tril( A11 )' * tril( A11 ); */
00073       FLA_Ttmm_internal( FLA_LOWER_TRIANGULAR, A11,
00074                          FLA_Cntl_sub_ttmm( cntl ) );
00075 
00076       /* A11 = A11 + A21' * A21; */
00077       FLA_Syrk_internal( FLA_LOWER_TRIANGULAR, FLA_TRANSPOSE,
00078                          FLA_ONE, A21, FLA_ONE, A11,
00079                          FLA_Cntl_sub_syrk( cntl ) );
00080     }
00081     else
00082     {
00083       /* A10 = tril( A11 )' * A10; */
00084       FLA_Trmm_internal( FLA_LEFT, FLA_LOWER_TRIANGULAR, FLA_CONJ_TRANSPOSE, FLA_NONUNIT_DIAG,
00085                          FLA_ONE, A11, A10,
00086                          FLA_Cntl_sub_trmm( cntl ) );
00087 
00088       /* A10 = A10 + A21' * A20; */
00089       FLA_Gemm_internal( FLA_CONJ_TRANSPOSE, FLA_NO_TRANSPOSE,
00090                          FLA_ONE, A21, A20, FLA_ONE, A10,
00091                          FLA_Cntl_sub_gemm( cntl ) );
00092 
00093       /* A11 = tril( A11 )' * tril( A11 ); */
00094       FLA_Ttmm_internal( FLA_LOWER_TRIANGULAR, A11,
00095                          FLA_Cntl_sub_ttmm( cntl ) );
00096 
00097       /* A11 = A11 + A21' * A21; */
00098       FLA_Herk_internal( FLA_LOWER_TRIANGULAR, FLA_CONJ_TRANSPOSE,
00099                          FLA_ONE, A21, FLA_ONE, A11,
00100                          FLA_Cntl_sub_herk( cntl ) );
00101     }
00102 
00103     /*------------------------------------------------------------*/
00104 
00105     FLA_Cont_with_3x3_to_2x2( &ATL, /**/ &ATR,       A00, A01, /**/ A02,
00106                                                      A10, A11, /**/ A12,
00107                             /* ************** */  /* ****************** */
00108                               &ABL, /**/ &ABR,       A20, A21, /**/ A22,
00109                               FLA_TL );
00110 
00111   }
00112 
00113   return FLA_SUCCESS;
00114 }

FLA_Error FLA_Ttmm_l_blk_var3 ( FLA_Obj  A,
fla_ttmm_t cntl 
)

References FLA_Cont_with_3x3_to_2x2(), FLA_Determine_blocksize(), FLA_Herk_internal(), FLA_Obj_is_real(), FLA_Obj_length(), FLA_ONE, FLA_Part_2x2(), FLA_Repart_2x2_to_3x3(), FLA_Syrk_internal(), FLA_Trmm_internal(), and FLA_Ttmm_internal().

Referenced by FLA_Ttmm_l().

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 
00050     b = FLA_Determine_blocksize( ABR, FLA_BR, FLA_Cntl_blocksize( cntl ) );
00051 
00052     FLA_Repart_2x2_to_3x3( ATL, /**/ ATR,       &A00, /**/ &A01, &A02,
00053                         /* ************* */   /* ******************** */
00054                                                 &A10, /**/ &A11, &A12,
00055                            ABL, /**/ ABR,       &A20, /**/ &A21, &A22,
00056                            b, b, FLA_BR );
00057 
00058     /*------------------------------------------------------------*/
00059 
00060     if ( FLA_Obj_is_real( A ) )
00061     {
00062       /* A11 = tril( A11 )' * tril( A11 ); */
00063       FLA_Ttmm_internal( FLA_LOWER_TRIANGULAR, A11,
00064                          FLA_Cntl_sub_ttmm( cntl ) );
00065 
00066       /* A11 = A11 + A21' * A21; */
00067       FLA_Syrk_internal( FLA_LOWER_TRIANGULAR, FLA_TRANSPOSE,
00068                          FLA_ONE, A21, FLA_ONE, A11,
00069                          FLA_Cntl_sub_syrk( cntl ) );
00070 
00071       /* A21 = tril( A22 )' * A21; */
00072       FLA_Trmm_internal( FLA_LEFT, FLA_LOWER_TRIANGULAR, FLA_TRANSPOSE, FLA_NONUNIT_DIAG,
00073                          FLA_ONE, A22, A21,
00074                          FLA_Cntl_sub_trmm( cntl ) );
00075     }
00076     else
00077     {
00078       /* A11 = tril( A11 )' * tril( A11 ); */
00079       FLA_Ttmm_internal( FLA_LOWER_TRIANGULAR, A11,
00080                          FLA_Cntl_sub_ttmm( cntl ) );
00081 
00082       /* A11 = A11 + A21' * A21; */
00083       FLA_Herk_internal( FLA_LOWER_TRIANGULAR, FLA_CONJ_TRANSPOSE,
00084                          FLA_ONE, A21, FLA_ONE, A11,
00085                          FLA_Cntl_sub_herk( cntl ) );
00086 
00087       /* A21 = tril( A22 )' * A21; */
00088       FLA_Trmm_internal( FLA_LEFT, FLA_LOWER_TRIANGULAR, FLA_CONJ_TRANSPOSE, FLA_NONUNIT_DIAG,
00089                          FLA_ONE, A22, A21,
00090                          FLA_Cntl_sub_trmm( cntl ) );
00091     }
00092 
00093     /*------------------------------------------------------------*/
00094 
00095     FLA_Cont_with_3x3_to_2x2( &ATL, /**/ &ATR,       A00, A01, /**/ A02,
00096                                                      A10, A11, /**/ A12,
00097                             /* ************** */  /* ****************** */
00098                               &ABL, /**/ &ABR,       A20, A21, /**/ A22,
00099                               FLA_TL );
00100 
00101   }
00102 
00103   return FLA_SUCCESS;
00104 }

FLA_Error FLA_Ttmm_l_unb_var1 ( FLA_Obj  A  ) 

References FLA_Absolute_square(), FLA_Cont_with_3x3_to_2x2(), FLA_Herc_external(), FLA_Obj_is_real(), FLA_Obj_length(), FLA_ONE, FLA_Part_2x2(), FLA_Repart_2x2_to_3x3(), FLA_Scal_external(), and FLA_Syr_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 
00048     FLA_Repart_2x2_to_3x3( ATL, /**/ ATR,       &A00,  /**/ &a01,     &A02,
00049                         /* ************* */   /* ************************** */
00050                                                 &a10t, /**/ &alpha11, &a12t,
00051                            ABL, /**/ ABR,       &A20,  /**/ &a21,     &A22,
00052                            1, 1, FLA_BR );
00053 
00054     /*------------------------------------------------------------*/
00055 
00056     if ( FLA_Obj_is_real( A ) )
00057     {
00058       /* A00 = A00 + a10t' * a10t; */
00059       FLA_Syr_external( FLA_LOWER_TRIANGULAR, FLA_ONE, a10t, A00 );
00060 
00061       /* a10t = alpha11 * a10t; */
00062       FLA_Scal_external( alpha11, a10t );
00063 
00064       /* alpha11 = alpha11 * alpha11; */
00065       FLA_Absolute_square( alpha11 );
00066     }
00067     else
00068     {
00069       /* A00 = A00 + a10t' * a10t; */
00070       FLA_Herc_external( FLA_LOWER_TRIANGULAR, FLA_CONJUGATE, FLA_ONE, a10t, A00 );
00071 
00072       /* a10t = alpha11 * a10t; */
00073       FLA_Scal_external( alpha11, a10t );
00074 
00075       /* alpha11 = alpha11' * alpha11; */
00076       FLA_Absolute_square( alpha11 );
00077     }
00078 
00079     /*------------------------------------------------------------*/
00080 
00081     FLA_Cont_with_3x3_to_2x2( &ATL, /**/ &ATR,       A00,  a01,     /**/ A02,
00082                                                      a10t, alpha11, /**/ a12t,
00083                             /* ************** */  /* ************************ */
00084                               &ABL, /**/ &ABR,       A20,  a21,     /**/ A22,
00085                               FLA_TL );
00086 
00087   }
00088 
00089   return FLA_SUCCESS;
00090 }

FLA_Error FLA_Ttmm_l_unb_var2 ( FLA_Obj  A  ) 

References FLA_Absolute_square(), FLA_Cont_with_3x3_to_2x2(), FLA_Dotcs_external(), FLA_Dots_external(), FLA_Gemv_external(), FLA_Gemvc_external(), FLA_Obj_is_real(), FLA_Obj_length(), FLA_ONE, FLA_Part_2x2(), FLA_Repart_2x2_to_3x3(), and FLA_Scal_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 
00048     FLA_Repart_2x2_to_3x3( ATL, /**/ ATR,       &A00,  /**/ &a01,     &A02,
00049                         /* ************* */   /* ************************** */
00050                                                 &a10t, /**/ &alpha11, &a12t,
00051                            ABL, /**/ ABR,       &A20,  /**/ &a21,     &A22,
00052                            1, 1, FLA_BR );
00053 
00054     /*------------------------------------------------------------*/
00055 
00056     if ( FLA_Obj_is_real( A ) )
00057     {
00058       /* a10t = alpha11 * a10t; */
00059       FLA_Scal_external( alpha11, a10t );
00060 
00061       /* a10t  = a10t  + a21' * A20; */
00062       /* a10t' = a10t' + A20' * a21; */
00063       FLA_Gemv_external( FLA_TRANSPOSE, FLA_ONE, A20, a21, FLA_ONE, a10t );
00064 
00065       /* alpha11 = alpha11 * alpha11; */
00066       FLA_Absolute_square( alpha11 );
00067 
00068       /* alpha11 = alpha11 + a21' * a21; */
00069       FLA_Dots_external( FLA_ONE, a21, a21, FLA_ONE, alpha11 );
00070     }
00071     else
00072     {
00073       /* a10t = alpha11 * a10t; */
00074       FLA_Scal_external( alpha11, a10t );
00075 
00076       /* a10t  = a10t  + a21' * A20; */
00077       /* a10t' = a10t' + A20' * a21; */
00078       FLA_Gemvc_external( FLA_TRANSPOSE, FLA_CONJUGATE, FLA_ONE, A20, a21, FLA_ONE, a10t );
00079 
00080       /* alpha11 = alpha11' * alpha11; */
00081       FLA_Absolute_square( alpha11 );
00082 
00083       /* alpha11 = alpha11 + a21' * a21; */
00084       FLA_Dotcs_external( FLA_CONJUGATE, FLA_ONE, a21, a21, FLA_ONE, alpha11 );
00085     }
00086 
00087     /*------------------------------------------------------------*/
00088 
00089     FLA_Cont_with_3x3_to_2x2( &ATL, /**/ &ATR,       A00,  a01,     /**/ A02,
00090                                                      a10t, alpha11, /**/ a12t,
00091                             /* ************** */  /* ************************ */
00092                               &ABL, /**/ &ABR,       A20,  a21,     /**/ A22,
00093                               FLA_TL );
00094 
00095   }
00096 
00097   return FLA_SUCCESS;
00098 }

FLA_Error FLA_Ttmm_l_unb_var3 ( FLA_Obj  A  ) 

References FLA_Absolute_square(), FLA_Cont_with_3x3_to_2x2(), FLA_Dotcs_external(), FLA_Dots_external(), FLA_Obj_is_real(), FLA_Obj_length(), FLA_ONE, FLA_Part_2x2(), FLA_Repart_2x2_to_3x3(), and FLA_Trmv_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 
00048     FLA_Repart_2x2_to_3x3( ATL, /**/ ATR,       &A00,  /**/ &a01,     &A02,
00049                         /* ************* */   /* ************************** */
00050                                                 &a10t, /**/ &alpha11, &a12t,
00051                            ABL, /**/ ABR,       &A20,  /**/ &a21,     &A22,
00052                            1, 1, FLA_BR );
00053 
00054     /*------------------------------------------------------------*/
00055 
00056     if ( FLA_Obj_is_real( A ) )
00057     {
00058       /* alpha11 = alpha11 * alpha11; */
00059       FLA_Absolute_square( alpha11 );
00060 
00061       /* alpha11 = alpha11 + a21' * a21; */
00062       FLA_Dots_external( FLA_ONE, a21, a21, FLA_ONE, alpha11 );
00063 
00064       /* a21 = tril( A22 )' * a21; */
00065       FLA_Trmv_external( FLA_LOWER_TRIANGULAR, FLA_TRANSPOSE, FLA_NONUNIT_DIAG,
00066                 A22, a21 );
00067     }
00068     else
00069     {
00070       /* alpha11 = alpha11' * alpha11; */
00071       FLA_Absolute_square( alpha11 );
00072 
00073       /* alpha11 = alpha11 + a21' * a21; */
00074       FLA_Dotcs_external( FLA_CONJUGATE, FLA_ONE, a21, a21, FLA_ONE, alpha11 );
00075 
00076       /* a21 = tril( A22 )' * a21; */
00077       FLA_Trmv_external( FLA_LOWER_TRIANGULAR, FLA_CONJ_TRANSPOSE, FLA_NONUNIT_DIAG,
00078                 A22, a21 );
00079     }
00080 
00081     /*------------------------------------------------------------*/
00082 
00083     FLA_Cont_with_3x3_to_2x2( &ATL, /**/ &ATR,       A00,  a01,     /**/ A02,
00084                                                      a10t, alpha11, /**/ a12t,
00085                             /* ************** */  /* ************************ */
00086                               &ABL, /**/ &ABR,       A20,  a21,     /**/ A22,
00087                               FLA_TL );
00088 
00089   }
00090 
00091   return FLA_SUCCESS;
00092 }


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