FLA_Hemm_lu_blk_var4.c File Reference

(r)


Functions

FLA_Error FLA_Hemm_lu_blk_var4 (FLA_Obj alpha, FLA_Obj A, FLA_Obj B, FLA_Obj beta, FLA_Obj C, fla_hemm_t *cntl)

Function Documentation

FLA_Error FLA_Hemm_lu_blk_var4 ( FLA_Obj  alpha,
FLA_Obj  A,
FLA_Obj  B,
FLA_Obj  beta,
FLA_Obj  C,
fla_hemm_t cntl 
)

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

Referenced by FLA_Hemm_lu().

00036 {
00037   FLA_Obj ATL,   ATR,      A00, A01, A02, 
00038           ABL,   ABR,      A10, A11, A12,
00039                            A20, A21, A22;
00040 
00041   FLA_Obj BT,              B0,
00042           BB,              B1,
00043                            B2;
00044 
00045   FLA_Obj CT,              C0,
00046           CB,              C1,
00047                            C2;
00048 
00049   dim_t b;
00050 
00051   FLA_Part_2x2( A,    &ATL, &ATR,
00052                       &ABL, &ABR,     0, 0, FLA_TL );
00053 
00054   FLA_Part_2x1( B,    &BT, 
00055                       &BB,            0, FLA_TOP );
00056 
00057   FLA_Part_2x1( C,    &CT, 
00058                       &CB,            0, FLA_TOP );
00059 
00060   while ( FLA_Obj_length( ATL ) < FLA_Obj_length( A ) ){
00061 
00062     b = FLA_Determine_blocksize( ABR, FLA_BR, FLA_Cntl_blocksize( cntl ) );
00063 
00064     FLA_Repart_2x2_to_3x3( ATL, /**/ ATR,       &A00, /**/ &A01, &A02,
00065                         /* ************* */   /* ******************** */
00066                                                 &A10, /**/ &A11, &A12,
00067                            ABL, /**/ ABR,       &A20, /**/ &A21, &A22,
00068                            b, b, FLA_BR );
00069 
00070     FLA_Repart_2x1_to_3x1( BT,                &B0, 
00071                         /* ** */            /* ** */
00072                                               &B1, 
00073                            BB,                &B2,        b, FLA_BOTTOM );
00074 
00075     FLA_Repart_2x1_to_3x1( CT,                &C0, 
00076                         /* ** */            /* ** */
00077                                               &C1, 
00078                            CB,                &C2,        b, FLA_BOTTOM );
00079 
00080     /*------------------------------------------------------------*/
00081 
00082     /* C0 = C0 + A01 * B1 */
00083     FLA_Gemm_internal( FLA_NO_TRANSPOSE, FLA_NO_TRANSPOSE, 
00084                        alpha, A01, B1, FLA_ONE, C0,
00085                        FLA_Cntl_sub_gemm1( cntl ) );
00086 
00087     /* C1 = C1 + A11 * B1 */
00088     FLA_Hemm_internal( FLA_LEFT, FLA_UPPER_TRIANGULAR, 
00089                        alpha, A11, B1, beta, C1,
00090                        FLA_Cntl_sub_hemm( cntl ) );
00091 
00092     /* C2 = C2 + A12' * B1 */
00093     FLA_Gemm_internal( FLA_CONJ_TRANSPOSE, FLA_NO_TRANSPOSE, 
00094                        alpha, A12, B1, FLA_ONE, C2,
00095                        FLA_Cntl_sub_gemm2( cntl ) );
00096 
00097     /*------------------------------------------------------------*/
00098 
00099     FLA_Cont_with_3x3_to_2x2( &ATL, /**/ &ATR,       A00, A01, /**/ A02,
00100                                                      A10, A11, /**/ A12,
00101                             /* ************** */  /* ****************** */
00102                               &ABL, /**/ &ABR,       A20, A21, /**/ A22,
00103                               FLA_TL );
00104 
00105     FLA_Cont_with_3x1_to_2x1( &BT,                B0, 
00106                                                   B1, 
00107                             /* ** */           /* ** */
00108                               &BB,                B2,     FLA_TOP );
00109 
00110     FLA_Cont_with_3x1_to_2x1( &CT,                C0, 
00111                                                   C1, 
00112                             /* ** */           /* ** */
00113                               &CB,                C2,     FLA_TOP );
00114 
00115   }
00116 
00117   return FLA_SUCCESS;
00118 }


Generated on Mon Jul 6 05:45:54 2009 for libflame by  doxygen 1.5.9