FLA_Herk_un_blk_var4.c File Reference

(r)


Functions

FLA_Error FLA_Herk_un_blk_var4 (FLA_Obj alpha, FLA_Obj A, FLA_Obj beta, FLA_Obj C, fla_herk_t *cntl)

Function Documentation

FLA_Error FLA_Herk_un_blk_var4 ( FLA_Obj  alpha,
FLA_Obj  A,
FLA_Obj  beta,
FLA_Obj  C,
fla_herk_t cntl 
)

References FLA_Cont_with_3x1_to_2x1(), FLA_Cont_with_3x3_to_2x2(), FLA_Determine_blocksize(), FLA_Gemm_internal(), FLA_Herk_internal(), FLA_Obj_length(), FLA_Part_2x1(), FLA_Part_2x2(), FLA_Repart_2x1_to_3x1(), and FLA_Repart_2x2_to_3x3().

Referenced by FLA_Herk_un().

00038 {
00039   FLA_Obj AT,              A0,
00040           AB,              A1,
00041                            A2;
00042 
00043   FLA_Obj CTL,   CTR,      C00, C01, C02, 
00044           CBL,   CBR,      C10, C11, C12,
00045                            C20, C21, C22;
00046 
00047   dim_t b;
00048 
00049   FLA_Part_2x1( A,    &AT, 
00050                       &AB,            0, FLA_BOTTOM );
00051 
00052   FLA_Part_2x2( C,    &CTL, &CTR,
00053                       &CBL, &CBR,     0, 0, FLA_BR );
00054 
00055   while ( FLA_Obj_length( AB ) < FLA_Obj_length( A ) ){
00056 
00057     b = FLA_Determine_blocksize( AT, FLA_TOP, FLA_Cntl_blocksize( cntl ) );
00058 
00059     FLA_Repart_2x1_to_3x1( AT,                &A0, 
00060                                               &A1, 
00061                         /* ** */            /* ** */
00062                            AB,                &A2,        b, FLA_TOP );
00063 
00064     FLA_Repart_2x2_to_3x3( CTL, /**/ CTR,       &C00, &C01, /**/ &C02,
00065                                                 &C10, &C11, /**/ &C12,
00066                         /* ************* */   /* ******************** */
00067                            CBL, /**/ CBR,       &C20, &C21, /**/ &C22,
00068                            b, b, FLA_TL );
00069 
00070     /*------------------------------------------------------------*/
00071 
00072     /* C01 = C01 + A0 * A1' */
00073     FLA_Gemm_internal( FLA_NO_TRANSPOSE, FLA_CONJ_TRANSPOSE, 
00074                        alpha, A0, A1, beta, C01,
00075                        FLA_Cntl_sub_gemm( cntl ) );
00076 
00077     /* C11 = C11 + A1 * A1' */
00078     FLA_Herk_internal( FLA_UPPER_TRIANGULAR, FLA_NO_TRANSPOSE, 
00079                        alpha, A1, beta, C11,
00080                        FLA_Cntl_sub_herk( cntl ) );
00081 
00082     /*------------------------------------------------------------*/
00083 
00084     FLA_Cont_with_3x1_to_2x1( &AT,                A0, 
00085                             /* ** */           /* ** */
00086                                                   A1, 
00087                               &AB,                A2,     FLA_BOTTOM );
00088 
00089     FLA_Cont_with_3x3_to_2x2( &CTL, /**/ &CTR,       C00, /**/ C01, C02,
00090                             /* ************** */  /* ****************** */
00091                                                      C10, /**/ C11, C12,
00092                               &CBL, /**/ &CBR,       C20, /**/ C21, C22,
00093                               FLA_BR );
00094 
00095   }
00096 
00097   return FLA_SUCCESS;
00098 }


Generated on Mon Jul 6 05:45:55 2009 for libflame by  doxygen 1.5.9