Functions | |
FLA_Error | FLA_LU_piv_blk_var5 (FLA_Obj A, FLA_Obj p, fla_lu_t *cntl) |
References FLA_Apply_pivots(), FLA_Cont_with_3x1_to_2x1(), FLA_Cont_with_3x3_to_2x2(), FLA_Determine_blocksize(), FLA_Gemm_internal(), FLA_LU_piv_internal(), FLA_Merge_2x1(), FLA_MINUS_ONE, FLA_Obj_length(), FLA_Obj_width(), 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_LU_piv_internal().
00036 { 00037 FLA_Obj ATL, ATR, A00, A01, A02, 00038 ABL, ABR, A10, A11, A12, 00039 A20, A21, A22; 00040 00041 FLA_Obj pT, p0, 00042 pB, p1, 00043 p2; 00044 00045 FLA_Obj AB0, AB1, AB2; 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( p, &pT, 00053 &pB, 0, FLA_TOP ); 00054 00055 while ( FLA_Obj_length( ATL ) < FLA_Obj_length( A ) && 00056 FLA_Obj_width( ATL ) < FLA_Obj_width( A )){ 00057 00058 b = FLA_Determine_blocksize( ABR, FLA_BR, FLA_Cntl_blocksize( cntl ) ); 00059 00060 FLA_Repart_2x2_to_3x3( ATL, /**/ ATR, &A00, /**/ &A01, &A02, 00061 /* ************* */ /* ******************** */ 00062 &A10, /**/ &A11, &A12, 00063 ABL, /**/ ABR, &A20, /**/ &A21, &A22, 00064 b, b, FLA_BR ); 00065 00066 FLA_Repart_2x1_to_3x1( pT, &p0, 00067 /* ** */ /* ** */ 00068 &p1, 00069 pB, &p2, b, FLA_BOTTOM ); 00070 00071 /*------------------------------------------------------------*/ 00072 00073 /* AB1 = / A11 \ 00074 \ A21 / */ 00075 FLA_Merge_2x1( A11, 00076 A21, &AB1 ); 00077 00078 /* AB1, p1 = LU_piv( AB1 ) */ 00079 FLA_LU_piv_internal( AB1, p1, 00080 FLA_Cntl_sub_lu( cntl ) ); 00081 00082 /* AB0 = / A10 \ 00083 \ A20 / */ 00084 FLA_Merge_2x1( A10, 00085 A20, &AB0 ); 00086 00087 /* Apply computed pivots to AB0 */ 00088 FLA_Apply_pivots( FLA_LEFT, FLA_NO_TRANSPOSE, p1, AB0 ); 00089 00090 /* AB2 = / A12 \ 00091 \ A22 / */ 00092 FLA_Merge_2x1( A12, 00093 A22, &AB2 ); 00094 00095 /* Apply computed pivots to AB2 */ 00096 FLA_Apply_pivots( FLA_LEFT, FLA_NO_TRANSPOSE, p1, AB2 ); 00097 00098 /* A12 = trilu( A11 ) \ A12 */ 00099 FLA_Trsm_internal( FLA_LEFT, FLA_LOWER_TRIANGULAR, 00100 FLA_NO_TRANSPOSE, FLA_UNIT_DIAG, 00101 FLA_ONE, A11, A12, 00102 FLA_Cntl_sub_trsm1( cntl ) ); 00103 00104 /* A22 = A22 - A21 * A12 */ 00105 FLA_Gemm_internal( FLA_NO_TRANSPOSE, FLA_NO_TRANSPOSE, 00106 FLA_MINUS_ONE, A21, A12, FLA_ONE, A22, 00107 FLA_Cntl_sub_gemm1( cntl ) ); 00108 00109 /*------------------------------------------------------------*/ 00110 00111 FLA_Cont_with_3x3_to_2x2( &ATL, /**/ &ATR, A00, A01, /**/ A02, 00112 A10, A11, /**/ A12, 00113 /* ************** */ /* ****************** */ 00114 &ABL, /**/ &ABR, A20, A21, /**/ A22, 00115 FLA_TL ); 00116 00117 FLA_Cont_with_3x1_to_2x1( &pT, p0, 00118 p1, 00119 /* ** */ /* ** */ 00120 &pB, p2, FLA_TOP ); 00121 00122 } 00123 00124 return FLA_SUCCESS; 00125 }