Functions | |
FLA_Error | FLA_Gemm_hh_blk_var6 (FLA_Obj alpha, FLA_Obj A, FLA_Obj B, FLA_Obj beta, FLA_Obj C, fla_gemm_t *cntl) |
FLA_Error FLA_Gemm_hh_blk_var6 | ( | FLA_Obj | alpha, | |
FLA_Obj | A, | |||
FLA_Obj | B, | |||
FLA_Obj | beta, | |||
FLA_Obj | C, | |||
fla_gemm_t * | cntl | |||
) |
References FLA_Cont_with_1x3_to_1x2(), FLA_Cont_with_3x1_to_2x1(), FLA_Determine_blocksize(), FLA_Gemm_internal(), FLA_Obj_length(), FLA_Part_1x2(), FLA_Part_2x1(), FLA_Repart_1x2_to_1x3(), and FLA_Repart_2x1_to_3x1().
Referenced by FLA_Gemm_hh().
00038 { 00039 FLA_Obj AT, A0, 00040 AB, A1, 00041 A2; 00042 00043 FLA_Obj BL, BR, B0, B1, B2; 00044 00045 dim_t b; 00046 00047 FLA_Part_2x1( A, &AT, 00048 &AB, 0, FLA_BOTTOM ); 00049 00050 FLA_Part_1x2( B, &BL, &BR, 0, FLA_RIGHT ); 00051 00052 while ( FLA_Obj_length( AB ) < FLA_Obj_length( A ) ){ 00053 00054 b = FLA_Determine_blocksize( AT, FLA_TOP, FLA_Cntl_blocksize( cntl ) ); 00055 00056 FLA_Repart_2x1_to_3x1( AT, &A0, 00057 &A1, 00058 /* ** */ /* ** */ 00059 AB, &A2, b, FLA_TOP ); 00060 00061 FLA_Repart_1x2_to_1x3( BL, /**/ BR, &B0, &B1, /**/ &B2, 00062 b, FLA_LEFT ); 00063 00064 /*------------------------------------------------------------*/ 00065 00066 /* C = alpha * A1' * B1' + C; */ 00067 FLA_Gemm_internal( FLA_CONJ_TRANSPOSE, FLA_CONJ_TRANSPOSE, 00068 alpha, A1, B1, beta, C, 00069 FLA_Cntl_sub_gemm( cntl ) ); 00070 00071 /*------------------------------------------------------------*/ 00072 00073 FLA_Cont_with_3x1_to_2x1( &AT, A0, 00074 /* ** */ /* ** */ 00075 A1, 00076 &AB, A2, FLA_BOTTOM ); 00077 00078 FLA_Cont_with_1x3_to_1x2( &BL, /**/ &BR, B0, /**/ B1, B2, 00079 FLA_RIGHT ); 00080 00081 } 00082 00083 return FLA_SUCCESS; 00084 }