Functions | |
FLA_Error | FLA_Sylv_tt_blk_var15 (FLA_Obj isgn, FLA_Obj A, FLA_Obj B, FLA_Obj C, FLA_Obj scale, fla_sylv_t *cntl) |
FLA_Error FLA_Sylv_tt_blk_var15 | ( | FLA_Obj | isgn, | |
FLA_Obj | A, | |||
FLA_Obj | B, | |||
FLA_Obj | C, | |||
FLA_Obj | scale, | |||
fla_sylv_t * | cntl | |||
) |
References FLA_Cont_with_3x1_to_2x1(), FLA_Cont_with_3x3_to_2x2(), FLA_Determine_blocksize(), FLA_Gemm_internal(), FLA_MINUS_ONE, FLA_Obj_length(), FLA_ONE, FLA_Part_2x1(), FLA_Part_2x2(), FLA_Repart_2x1_to_3x1(), FLA_Repart_2x2_to_3x3(), and FLA_Sylv_internal().
Referenced by FLA_Sylv_tt().
00036 { 00037 FLA_Obj ATL, ATR, A00, A01, A02, 00038 ABL, ABR, A10, A11, A12, 00039 A20, A21, A22; 00040 00041 FLA_Obj CT, C0, 00042 CB, C1, 00043 C2; 00044 00045 dim_t b; 00046 00047 FLA_Part_2x2( A, &ATL, &ATR, 00048 &ABL, &ABR, 0, 0, FLA_TL ); 00049 00050 FLA_Part_2x1( C, &CT, 00051 &CB, 0, FLA_TOP ); 00052 00053 while ( FLA_Obj_length( ATL ) < FLA_Obj_length( A ) ){ 00054 00055 b = FLA_Determine_blocksize( CB, FLA_BOTTOM, FLA_Cntl_blocksize( cntl ) ); 00056 00057 FLA_Repart_2x2_to_3x3( ATL, /**/ ATR, &A00, /**/ &A01, &A02, 00058 /* ************* */ /* ******************** */ 00059 &A10, /**/ &A11, &A12, 00060 ABL, /**/ ABR, &A20, /**/ &A21, &A22, 00061 b, b, FLA_BR ); 00062 00063 FLA_Repart_2x1_to_3x1( CT, &C0, 00064 /* ** */ /* ** */ 00065 &C1, 00066 CB, &C2, b, FLA_BOTTOM ); 00067 00068 // Loop Invariant: 00069 // CT = sylv( ATL', B', CT ) 00070 // CB = CB 00071 00072 /*------------------------------------------------------------*/ 00073 00074 // C1 = sylv( A11', B', C1 - A01' * C0 ); 00075 FLA_Gemm_internal( FLA_TRANSPOSE, FLA_NO_TRANSPOSE, 00076 FLA_MINUS_ONE, A01, C0, FLA_ONE, C1, 00077 FLA_Cntl_sub_gemm1( cntl ) ); 00078 00079 FLA_Sylv_internal( FLA_TRANSPOSE, FLA_TRANSPOSE, 00080 isgn, A11, B, C1, scale, 00081 FLA_Cntl_sub_sylv1( cntl ) ); 00082 00083 /*------------------------------------------------------------*/ 00084 00085 FLA_Cont_with_3x3_to_2x2( &ATL, /**/ &ATR, A00, A01, /**/ A02, 00086 A10, A11, /**/ A12, 00087 /* ************** */ /* ****************** */ 00088 &ABL, /**/ &ABR, A20, A21, /**/ A22, 00089 FLA_TL ); 00090 00091 FLA_Cont_with_3x1_to_2x1( &CT, C0, 00092 C1, 00093 /* ** */ /* ** */ 00094 &CB, C2, FLA_TOP ); 00095 00096 } 00097 00098 return FLA_SUCCESS; 00099 }