Functions | |
FLA_Error | FLA_Herk_ln_blk_var2 (FLA_Obj alpha, FLA_Obj A, FLA_Obj beta, FLA_Obj C, fla_herk_t *cntl) |
FLA_Error FLA_Herk_ln_blk_var2 | ( | 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_ln().
00036 { 00037 FLA_Obj AT, A0, 00038 AB, A1, 00039 A2; 00040 00041 FLA_Obj CTL, CTR, C00, C01, C02, 00042 CBL, CBR, C10, C11, C12, 00043 C20, C21, C22; 00044 00045 dim_t b; 00046 00047 FLA_Part_2x1( A, &AT, 00048 &AB, 0, FLA_TOP ); 00049 00050 FLA_Part_2x2( C, &CTL, &CTR, 00051 &CBL, &CBR, 0, 0, FLA_TL ); 00052 00053 while ( FLA_Obj_length( AT ) < FLA_Obj_length( A ) ){ 00054 00055 b = FLA_Determine_blocksize( AB, FLA_BOTTOM, FLA_Cntl_blocksize( cntl ) ); 00056 00057 FLA_Repart_2x1_to_3x1( AT, &A0, 00058 /* ** */ /* ** */ 00059 &A1, 00060 AB, &A2, b, FLA_BOTTOM ); 00061 00062 FLA_Repart_2x2_to_3x3( CTL, /**/ CTR, &C00, /**/ &C01, &C02, 00063 /* ************* */ /* ******************** */ 00064 &C10, /**/ &C11, &C12, 00065 CBL, /**/ CBR, &C20, /**/ &C21, &C22, 00066 b, b, FLA_BR ); 00067 00068 /*------------------------------------------------------------*/ 00069 00070 /* C21 = C21 + A2 * A1' */ 00071 FLA_Gemm_internal( FLA_NO_TRANSPOSE, FLA_CONJ_TRANSPOSE, 00072 alpha, A2, A1, beta, C21, 00073 FLA_Cntl_sub_gemm( cntl ) ); 00074 00075 /* C11 = C11 + A1 * A1' */ 00076 FLA_Herk_internal( FLA_LOWER_TRIANGULAR, FLA_NO_TRANSPOSE, 00077 alpha, A1, beta, C11, 00078 FLA_Cntl_sub_herk( cntl ) ); 00079 00080 /*------------------------------------------------------------*/ 00081 00082 FLA_Cont_with_3x1_to_2x1( &AT, A0, 00083 A1, 00084 /* ** */ /* ** */ 00085 &AB, A2, FLA_TOP ); 00086 00087 FLA_Cont_with_3x3_to_2x2( &CTL, /**/ &CTR, C00, C01, /**/ C02, 00088 C10, C11, /**/ C12, 00089 /* ************** */ /* ****************** */ 00090 &CBL, /**/ &CBR, C20, C21, /**/ C22, 00091 FLA_TL ); 00092 00093 } 00094 00095 return FLA_SUCCESS; 00096 }