Functions | |
FLA_Error | FLA_Trsm_lut_blk_var1 (FLA_Diag diagA, FLA_Obj alpha, FLA_Obj A, FLA_Obj B, fla_trsm_t *cntl) |
FLA_Error FLA_Trsm_lut_blk_var1 | ( | FLA_Diag | diagA, | |
FLA_Obj | alpha, | |||
FLA_Obj | A, | |||
FLA_Obj | B, | |||
fla_trsm_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_Trsm_internal().
Referenced by FLA_Trsm_lut().
00038 { 00039 FLA_Obj ATL, ATR, A00, A01, A02, 00040 ABL, ABR, A10, A11, A12, 00041 A20, A21, A22; 00042 00043 FLA_Obj BT, B0, 00044 BB, B1, 00045 B2; 00046 00047 dim_t b; 00048 00049 FLA_Part_2x2( A, &ATL, &ATR, 00050 &ABL, &ABR, 0, 0, FLA_TL ); 00051 00052 FLA_Part_2x1( B, &BT, 00053 &BB, 0, FLA_TOP ); 00054 00055 while ( FLA_Obj_length( ATL ) < FLA_Obj_length( A ) ){ 00056 00057 b = FLA_Determine_blocksize( ABR, FLA_BR, FLA_Cntl_blocksize( cntl ) ); 00058 00059 FLA_Repart_2x2_to_3x3( ATL, /**/ ATR, &A00, /**/ &A01, &A02, 00060 /* ************* */ /* ******************** */ 00061 &A10, /**/ &A11, &A12, 00062 ABL, /**/ ABR, &A20, /**/ &A21, &A22, 00063 b, b, FLA_BR ); 00064 00065 FLA_Repart_2x1_to_3x1( BT, &B0, 00066 /* ** */ /* ** */ 00067 &B1, 00068 BB, &B2, b, FLA_BOTTOM ); 00069 00070 /*------------------------------------------------------------*/ 00071 00072 /* B1 = B1 - A01' * B0 */ 00073 FLA_Gemm_internal( FLA_TRANSPOSE, FLA_NO_TRANSPOSE, 00074 FLA_MINUS_ONE, A01, B0, FLA_ONE, B1, 00075 FLA_Cntl_sub_gemm( cntl ) ); 00076 00077 /* B1 = triu( A11' ) \ B1 */ 00078 FLA_Trsm_internal( FLA_LEFT, FLA_UPPER_TRIANGULAR, FLA_TRANSPOSE, diagA, 00079 alpha, A11, B1, 00080 FLA_Cntl_sub_trsm( cntl ) ); 00081 00082 /*------------------------------------------------------------*/ 00083 00084 FLA_Cont_with_3x3_to_2x2( &ATL, /**/ &ATR, A00, A01, /**/ A02, 00085 A10, A11, /**/ A12, 00086 /* ************** */ /* ****************** */ 00087 &ABL, /**/ &ABR, A20, A21, /**/ A22, 00088 FLA_TL ); 00089 00090 FLA_Cont_with_3x1_to_2x1( &BT, B0, 00091 B1, 00092 /* ** */ /* ** */ 00093 &BB, B2, FLA_TOP ); 00094 00095 } 00096 00097 return FLA_SUCCESS; 00098 }