FLA_Hemm_rl_blk_var1.c File Reference

(r)


Functions

FLA_Error FLA_Hemm_rl_blk_var1 (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_rl_blk_var1 ( FLA_Obj  alpha,
FLA_Obj  A,
FLA_Obj  B,
FLA_Obj  beta,
FLA_Obj  C,
fla_hemm_t cntl 
)

References FLA_Cont_with_1x3_to_1x2(), FLA_Cont_with_3x3_to_2x2(), FLA_Determine_blocksize(), FLA_Gemm_internal(), FLA_Hemm_internal(), FLA_Obj_length(), FLA_ONE, FLA_Part_1x2(), FLA_Part_2x2(), FLA_Repart_1x2_to_1x3(), and FLA_Repart_2x2_to_3x3().

Referenced by FLA_Hemm_rl().

00038 {
00039   FLA_Obj ATL,   ATR,      A00, A01, A02, 
00040           ABL,   ABR,      A10, A11, A12,
00041                            A20, A21, A22;
00042 
00043   FLA_Obj BL,    BR,       B0,  B1,  B2;
00044 
00045   FLA_Obj CL,    CR,       C0,  C1,  C2;
00046 
00047   dim_t b;
00048 
00049   FLA_Part_2x2( A,    &ATL, &ATR,
00050                       &ABL, &ABR,     0, 0, FLA_TL );
00051 
00052   FLA_Part_1x2( B,    &BL,  &BR,      0, FLA_LEFT );
00053 
00054   FLA_Part_1x2( C,    &CL,  &CR,      0, FLA_LEFT );
00055 
00056   while ( FLA_Obj_length( ATL ) < FLA_Obj_length( A ) ){
00057 
00058     b = FLA_Determine_blocksize( ABR, FLA_BR, FLA_Cntl_blocksize( cntl ) );
00059 
00060     FLA_Repart_2x2_to_3x3( ATL, /**/ ATR,       &A00, /**/ &A01, &A02,
00061                         /* ************* */   /* ******************** */
00062                                                 &A10, /**/ &A11, &A12,
00063                            ABL, /**/ ABR,       &A20, /**/ &A21, &A22,
00064                            b, b, FLA_BR );
00065 
00066     FLA_Repart_1x2_to_1x3( BL,  /**/ BR,        &B0, /**/ &B1, &B2,
00067                            b, FLA_RIGHT );
00068 
00069     FLA_Repart_1x2_to_1x3( CL,  /**/ CR,        &C0, /**/ &C1, &C2,
00070                            b, FLA_RIGHT );
00071 
00072     /*------------------------------------------------------------*/
00073 
00074     /* C0 = C0 + B1 * A10 */
00075     FLA_Gemm_internal( FLA_NO_TRANSPOSE, FLA_NO_TRANSPOSE, 
00076                        alpha, B1, A10, FLA_ONE, C0,
00077                        FLA_Cntl_sub_gemm1( cntl ) );
00078 
00079     /* C1 = C1 + B0 * A10' */
00080     FLA_Gemm_internal( FLA_NO_TRANSPOSE, FLA_CONJ_TRANSPOSE, 
00081                        alpha, B0, A10, FLA_ONE, C1,
00082                        FLA_Cntl_sub_gemm2( cntl ) );
00083 
00084     /* C1 = C1 + B1 * A11 */
00085     FLA_Hemm_internal( FLA_RIGHT, FLA_LOWER_TRIANGULAR, 
00086                        alpha, A11, B1, beta, C1,
00087                        FLA_Cntl_sub_hemm( cntl ) );
00088 
00089     /*------------------------------------------------------------*/
00090 
00091     FLA_Cont_with_3x3_to_2x2( &ATL, /**/ &ATR,       A00, A01, /**/ A02,
00092                                                      A10, A11, /**/ A12,
00093                             /* ************** */  /* ****************** */
00094                               &ABL, /**/ &ABR,       A20, A21, /**/ A22,
00095                               FLA_TL );
00096 
00097     FLA_Cont_with_1x3_to_1x2( &BL,  /**/ &BR,        B0, B1, /**/ B2,
00098                               FLA_LEFT );
00099 
00100     FLA_Cont_with_1x3_to_1x2( &CL,  /**/ &CR,        C0, C1, /**/ C2,
00101                               FLA_LEFT );
00102 
00103   }
00104 
00105   return FLA_SUCCESS;
00106 }


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