Functions | |
FLA_Error | FLA_Her2k_ln_blk_var3 (FLA_Obj alpha, FLA_Obj A, FLA_Obj B, FLA_Obj beta, FLA_Obj C, fla_her2k_t *cntl) |
FLA_Error FLA_Her2k_ln_blk_var3 | ( | FLA_Obj | alpha, | |
FLA_Obj | A, | |||
FLA_Obj | B, | |||
FLA_Obj | beta, | |||
FLA_Obj | C, | |||
fla_her2k_t * | cntl | |||
) |
References FLA_Cont_with_3x1_to_2x1(), FLA_Cont_with_3x3_to_2x2(), FLA_Determine_blocksize(), FLA_Gemm_internal(), FLA_Her2k_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_Her2k_ln().
00036 { 00037 FLA_Obj AT, A0, 00038 AB, A1, 00039 A2; 00040 00041 FLA_Obj BT, B0, 00042 BB, B1, 00043 B2; 00044 00045 FLA_Obj CTL, CTR, C00, C01, C02, 00046 CBL, CBR, C10, C11, C12, 00047 C20, C21, C22; 00048 00049 dim_t b; 00050 00051 FLA_Part_2x1( A, &AT, 00052 &AB, 0, FLA_TOP ); 00053 00054 FLA_Part_2x1( B, &BT, 00055 &BB, 0, FLA_TOP ); 00056 00057 FLA_Part_2x2( C, &CTL, &CTR, 00058 &CBL, &CBR, 0, 0, FLA_TL ); 00059 00060 while ( FLA_Obj_length( AT ) < FLA_Obj_length( A ) ){ 00061 00062 b = FLA_Determine_blocksize( AB, FLA_BOTTOM, FLA_Cntl_blocksize( cntl ) ); 00063 00064 FLA_Repart_2x1_to_3x1( AT, &A0, 00065 /* ** */ /* ** */ 00066 &A1, 00067 AB, &A2, b, FLA_BOTTOM ); 00068 00069 FLA_Repart_2x1_to_3x1( BT, &B0, 00070 /* ** */ /* ** */ 00071 &B1, 00072 BB, &B2, b, FLA_BOTTOM ); 00073 00074 FLA_Repart_2x2_to_3x3( CTL, /**/ CTR, &C00, /**/ &C01, &C02, 00075 /* ************* */ /* ******************** */ 00076 &C10, /**/ &C11, &C12, 00077 CBL, /**/ CBR, &C20, /**/ &C21, &C22, 00078 b, b, FLA_BR ); 00079 00080 /*------------------------------------------------------------*/ 00081 00082 /* C10 = C10 + A1 * B0' */ 00083 FLA_Gemm_internal( FLA_NO_TRANSPOSE, FLA_CONJ_TRANSPOSE, 00084 alpha, A1, B0, beta, C10, 00085 FLA_Cntl_sub_gemm1( cntl ) ); 00086 00087 /* C21 = C21 + B2 * A1' */ 00088 FLA_Gemm_internal( FLA_NO_TRANSPOSE, FLA_CONJ_TRANSPOSE, 00089 alpha, B2, A1, FLA_ONE, C21, 00090 FLA_Cntl_sub_gemm2( cntl ) ); 00091 00092 /* C11 = C11 + A1 * B1' + B1 * A1' */ 00093 FLA_Her2k_internal( FLA_LOWER_TRIANGULAR, FLA_NO_TRANSPOSE, 00094 alpha, A1, B1, beta, C11, 00095 FLA_Cntl_sub_her2k( cntl ) ); 00096 00097 /*------------------------------------------------------------*/ 00098 00099 FLA_Cont_with_3x1_to_2x1( &AT, A0, 00100 A1, 00101 /* ** */ /* ** */ 00102 &AB, A2, FLA_TOP ); 00103 00104 FLA_Cont_with_3x1_to_2x1( &BT, B0, 00105 B1, 00106 /* ** */ /* ** */ 00107 &BB, B2, FLA_TOP ); 00108 00109 FLA_Cont_with_3x3_to_2x2( &CTL, /**/ &CTR, C00, C01, /**/ C02, 00110 C10, C11, /**/ C12, 00111 /* ************** */ /* ****************** */ 00112 &CBL, /**/ &CBR, C20, C21, /**/ C22, 00113 FLA_TL ); 00114 00115 } 00116 00117 return FLA_SUCCESS; 00118 }