Functions | |
FLA_Error | FLA_Her2k_lh_blk_var1 (FLA_Obj alpha, FLA_Obj A, FLA_Obj B, FLA_Obj beta, FLA_Obj C, fla_her2k_t *cntl) |
FLA_Error FLA_Her2k_lh_blk_var1 | ( | FLA_Obj | alpha, | |
FLA_Obj | A, | |||
FLA_Obj | B, | |||
FLA_Obj | beta, | |||
FLA_Obj | C, | |||
fla_her2k_t * | cntl | |||
) |
References FLA_Cont_with_1x3_to_1x2(), FLA_Cont_with_3x3_to_2x2(), FLA_Determine_blocksize(), FLA_Gemm_internal(), FLA_Her2k_internal(), FLA_Obj_width(), FLA_ONE, FLA_Part_1x2(), FLA_Part_2x2(), FLA_Repart_1x2_to_1x3(), and FLA_Repart_2x2_to_3x3().
Referenced by FLA_Her2k_lh().
00038 { 00039 FLA_Obj AL, AR, A0, A1, A2; 00040 00041 FLA_Obj BL, BR, B0, B1, B2; 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_1x2( A, &AL, &AR, 0, FLA_LEFT ); 00050 00051 FLA_Part_1x2( B, &BL, &BR, 0, FLA_LEFT ); 00052 00053 FLA_Part_2x2( C, &CTL, &CTR, 00054 &CBL, &CBR, 0, 0, FLA_TL ); 00055 00056 while ( FLA_Obj_width( AL ) < FLA_Obj_width( A ) ){ 00057 00058 b = FLA_Determine_blocksize( AR, FLA_RIGHT, FLA_Cntl_blocksize( cntl ) ); 00059 00060 FLA_Repart_1x2_to_1x3( AL, /**/ AR, &A0, /**/ &A1, &A2, 00061 b, FLA_RIGHT ); 00062 00063 FLA_Repart_1x2_to_1x3( BL, /**/ BR, &B0, /**/ &B1, &B2, 00064 b, FLA_RIGHT ); 00065 00066 FLA_Repart_2x2_to_3x3( CTL, /**/ CTR, &C00, /**/ &C01, &C02, 00067 /* ************* */ /* ******************** */ 00068 &C10, /**/ &C11, &C12, 00069 CBL, /**/ CBR, &C20, /**/ &C21, &C22, 00070 b, b, FLA_BR ); 00071 00072 /*------------------------------------------------------------*/ 00073 00074 /* C10 = C10 + A1' * B0 */ 00075 FLA_Gemm_internal( FLA_CONJ_TRANSPOSE, FLA_NO_TRANSPOSE, 00076 alpha, A1, B0, beta, C10, 00077 FLA_Cntl_sub_gemm1( cntl ) ); 00078 00079 /* C10 = C10 + B1' * A0 */ 00080 FLA_Gemm_internal( FLA_CONJ_TRANSPOSE, FLA_NO_TRANSPOSE, 00081 alpha, B1, A0, FLA_ONE, C10, 00082 FLA_Cntl_sub_gemm2( cntl ) ); 00083 00084 /* C11 = C11 + A1' * B1 + B1' * A1 */ 00085 FLA_Her2k_internal( FLA_LOWER_TRIANGULAR, FLA_CONJ_TRANSPOSE, 00086 alpha, A1, B1, beta, C11, 00087 FLA_Cntl_sub_her2k( cntl ) ); 00088 00089 /*------------------------------------------------------------*/ 00090 00091 FLA_Cont_with_1x3_to_1x2( &AL, /**/ &AR, A0, A1, /**/ A2, 00092 FLA_LEFT ); 00093 00094 FLA_Cont_with_1x3_to_1x2( &BL, /**/ &BR, B0, B1, /**/ B2, 00095 FLA_LEFT ); 00096 00097 FLA_Cont_with_3x3_to_2x2( &CTL, /**/ &CTR, C00, C01, /**/ C02, 00098 C10, C11, /**/ C12, 00099 /* ************** */ /* ****************** */ 00100 &CBL, /**/ &CBR, C20, C21, /**/ C22, 00101 FLA_TL ); 00102 00103 } 00104 00105 return FLA_SUCCESS; 00106 }