Functions | |
FLA_Error | FLA_LU_piv_blk_var4 (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().
00038 { 00039 FLA_Obj ATL, ATR, A00, A01, A02, 00040 ABL, ABR, A10, A11, A12, 00041 A20, A21, A22; 00042 00043 FLA_Obj pT, p0, 00044 pB, p1, 00045 p2; 00046 00047 FLA_Obj AB0, AB1, AB2; 00048 00049 dim_t b; 00050 00051 FLA_Part_2x2( A, &ATL, &ATR, 00052 &ABL, &ABR, 0, 0, FLA_TL ); 00053 00054 FLA_Part_2x1( p, &pT, 00055 &pB, 0, FLA_TOP ); 00056 00057 while ( FLA_Obj_length( ATL ) < FLA_Obj_length( A ) && 00058 FLA_Obj_width( ATL ) < FLA_Obj_width( A )){ 00059 00060 b = FLA_Determine_blocksize( ABR, FLA_BR, FLA_Cntl_blocksize( cntl ) ); 00061 00062 FLA_Repart_2x2_to_3x3( ATL, /**/ ATR, &A00, /**/ &A01, &A02, 00063 /* ************* */ /* ******************** */ 00064 &A10, /**/ &A11, &A12, 00065 ABL, /**/ ABR, &A20, /**/ &A21, &A22, 00066 b, b, FLA_BR ); 00067 00068 FLA_Repart_2x1_to_3x1( pT, &p0, 00069 /* ** */ /* ** */ 00070 &p1, 00071 pB, &p2, b, FLA_BOTTOM ); 00072 00073 /*------------------------------------------------------------*/ 00074 00075 /* A11 = A11 - A10 * A01 */ 00076 FLA_Gemm_internal( FLA_NO_TRANSPOSE, FLA_NO_TRANSPOSE, 00077 FLA_MINUS_ONE, A10, A01, FLA_ONE, A11, 00078 FLA_Cntl_sub_gemm1( cntl ) ); 00079 00080 /* A21 = A21 - A20 * A01 */ 00081 FLA_Gemm_internal( FLA_NO_TRANSPOSE, FLA_NO_TRANSPOSE, 00082 FLA_MINUS_ONE, A20, A01, FLA_ONE, A21, 00083 FLA_Cntl_sub_gemm3( cntl ) ); 00084 00085 /* AB1 = / A11 \ 00086 \ A21 / */ 00087 FLA_Merge_2x1( A11, 00088 A21, &AB1 ); 00089 00090 /* AB1, p1 = LU_piv( AB1 ) */ 00091 FLA_LU_piv_internal( AB1, p1, 00092 FLA_Cntl_sub_lu( cntl ) ); 00093 00094 /* AB0 = / A10 \ 00095 \ A20 / */ 00096 FLA_Merge_2x1( A10, 00097 A20, &AB0 ); 00098 00099 /* AB2 = / A12 \ 00100 \ A22 / */ 00101 FLA_Merge_2x1( A12, 00102 A22, &AB2 ); 00103 00104 /* Apply pivots to remaining columns */ 00105 FLA_Apply_pivots( FLA_LEFT, FLA_NO_TRANSPOSE, p1, AB0 ); 00106 FLA_Apply_pivots( FLA_LEFT, FLA_NO_TRANSPOSE, p1, AB2 ); 00107 00108 /* A12 = A12 - A10 * A02 */ 00109 FLA_Gemm_internal( FLA_NO_TRANSPOSE, FLA_NO_TRANSPOSE, 00110 FLA_MINUS_ONE, A10, A02, FLA_ONE, A12, 00111 FLA_Cntl_sub_gemm2( cntl ) ); 00112 00113 /* A12 = trilu( A11 ) \ A12 */ 00114 FLA_Trsm_internal( FLA_LEFT, FLA_LOWER_TRIANGULAR, 00115 FLA_NO_TRANSPOSE, FLA_UNIT_DIAG, 00116 FLA_ONE, A11, A12, 00117 FLA_Cntl_sub_trsm1( cntl ) ); 00118 00119 /*------------------------------------------------------------*/ 00120 00121 FLA_Cont_with_3x3_to_2x2( &ATL, /**/ &ATR, A00, A01, /**/ A02, 00122 A10, A11, /**/ A12, 00123 /* ************** */ /* ****************** */ 00124 &ABL, /**/ &ABR, A20, A21, /**/ A22, 00125 FLA_TL ); 00126 00127 FLA_Cont_with_3x1_to_2x1( &pT, p0, 00128 p1, 00129 /* ** */ /* ** */ 00130 &pB, p2, FLA_TOP ); 00131 00132 } 00133 00134 return FLA_SUCCESS; 00135 }