Functions | |
FLA_Error | FLA_Gemm_nt_blk_var3 (FLA_Obj alpha, FLA_Obj A, FLA_Obj B, FLA_Obj beta, FLA_Obj C, fla_gemm_t *cntl) |
FLA_Error FLA_Gemm_nt_blk_var3 | ( | 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_nt().
00036 { 00037 FLA_Obj BT, B0, 00038 BB, B1, 00039 B2; 00040 00041 FLA_Obj CL, CR, C0, C1, C2; 00042 00043 dim_t b; 00044 00045 FLA_Part_2x1( B, &BT, 00046 &BB, 0, FLA_TOP ); 00047 00048 FLA_Part_1x2( C, &CL, &CR, 0, FLA_LEFT ); 00049 00050 while ( FLA_Obj_length( BT ) < FLA_Obj_length( B ) ){ 00051 00052 b = FLA_Determine_blocksize( BB, FLA_BOTTOM, FLA_Cntl_blocksize( cntl ) ); 00053 00054 FLA_Repart_2x1_to_3x1( BT, &B0, 00055 /* ** */ /* ** */ 00056 &B1, 00057 BB, &B2, b, FLA_BOTTOM ); 00058 00059 FLA_Repart_1x2_to_1x3( CL, /**/ CR, &C0, /**/ &C1, &C2, 00060 b, FLA_RIGHT ); 00061 00062 /*------------------------------------------------------------*/ 00063 00064 /* C1 = alpha * A * B1' + C1; */ 00065 FLA_Gemm_internal( FLA_NO_TRANSPOSE, FLA_TRANSPOSE, 00066 alpha, A, B1, beta, C1, 00067 FLA_Cntl_sub_gemm( cntl ) ); 00068 00069 /*------------------------------------------------------------*/ 00070 00071 FLA_Cont_with_3x1_to_2x1( &BT, B0, 00072 B1, 00073 /* ** */ /* ** */ 00074 &BB, B2, FLA_TOP ); 00075 00076 FLA_Cont_with_1x3_to_1x2( &CL, /**/ &CR, C0, C1, /**/ C2, 00077 FLA_LEFT ); 00078 00079 } 00080 00081 return FLA_SUCCESS; 00082 }