Functions | |
FLA_Error | FLA_Gemm_tt_blk_var2 (FLA_Obj alpha, FLA_Obj A, FLA_Obj B, FLA_Obj beta, FLA_Obj C, fla_gemm_t *cntl) |
FLA_Error FLA_Gemm_tt_blk_var2 | ( | 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_width(), FLA_Part_1x2(), FLA_Part_2x1(), FLA_Repart_1x2_to_1x3(), and FLA_Repart_2x1_to_3x1().
Referenced by FLA_Gemm_tt().
00038 { 00039 FLA_Obj AL, AR, A0, A1, A2; 00040 00041 FLA_Obj CT, C0, 00042 CB, C1, 00043 C2; 00044 00045 dim_t b; 00046 00047 FLA_Part_1x2( A, &AL, &AR, 0, FLA_RIGHT ); 00048 00049 FLA_Part_2x1( C, &CT, 00050 &CB, 0, FLA_BOTTOM ); 00051 00052 while ( FLA_Obj_width( AR ) < FLA_Obj_width( A ) ){ 00053 00054 b = FLA_Determine_blocksize( AL, FLA_LEFT, FLA_Cntl_blocksize( cntl ) ); 00055 00056 FLA_Repart_1x2_to_1x3( AL, /**/ AR, &A0, &A1, /**/ &A2, 00057 b, FLA_LEFT ); 00058 00059 FLA_Repart_2x1_to_3x1( CT, &C0, 00060 &C1, 00061 /* ** */ /* ** */ 00062 CB, &C2, b, FLA_TOP ); 00063 00064 /*------------------------------------------------------------*/ 00065 00066 /* C1 = alpha * A1' * B' + C1; */ 00067 FLA_Gemm_internal( FLA_TRANSPOSE, FLA_TRANSPOSE, 00068 alpha, A1, B, beta, C1, 00069 FLA_Cntl_sub_gemm( cntl ) ); 00070 00071 /*------------------------------------------------------------*/ 00072 00073 FLA_Cont_with_1x3_to_1x2( &AL, /**/ &AR, A0, /**/ A1, A2, 00074 FLA_RIGHT ); 00075 00076 FLA_Cont_with_3x1_to_2x1( &CT, C0, 00077 /* ** */ /* ** */ 00078 C1, 00079 &CB, C2, FLA_BOTTOM ); 00080 00081 } 00082 00083 return FLA_SUCCESS; 00084 }