Functions | |
FLA_Error | FLA_Gemm_tt_blk_var4 (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_var4 | ( | 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_tt().
00038 { 00039 FLA_Obj BT, B0, 00040 BB, B1, 00041 B2; 00042 00043 FLA_Obj CL, CR, C0, C1, C2; 00044 00045 dim_t b; 00046 00047 FLA_Part_2x1( B, &BT, 00048 &BB, 0, FLA_BOTTOM ); 00049 00050 FLA_Part_1x2( C, &CL, &CR, 0, FLA_RIGHT ); 00051 00052 while ( FLA_Obj_length( BB ) < FLA_Obj_length( B ) ){ 00053 00054 b = FLA_Determine_blocksize( BT, FLA_TOP, FLA_Cntl_blocksize( cntl ) ); 00055 00056 FLA_Repart_2x1_to_3x1( BT, &B0, 00057 &B1, 00058 /* ** */ /* ** */ 00059 BB, &B2, b, FLA_TOP ); 00060 00061 FLA_Repart_1x2_to_1x3( CL, /**/ CR, &C0, &C1, /**/ &C2, 00062 b, FLA_LEFT ); 00063 00064 /*------------------------------------------------------------*/ 00065 00066 /* C1 = alpha * A' * B1' + C1; */ 00067 FLA_Gemm_internal( FLA_TRANSPOSE, FLA_TRANSPOSE, 00068 alpha, A, B1, beta, C1, 00069 FLA_Cntl_sub_gemm( cntl ) ); 00070 00071 /*------------------------------------------------------------*/ 00072 00073 FLA_Cont_with_3x1_to_2x1( &BT, B0, 00074 /* ** */ /* ** */ 00075 B1, 00076 &BB, B2, FLA_BOTTOM ); 00077 00078 FLA_Cont_with_1x3_to_1x2( &CL, /**/ &CR, C0, /**/ C1, C2, 00079 FLA_RIGHT ); 00080 00081 } 00082 00083 return FLA_SUCCESS; 00084 }