Functions | |
FLA_Error | FLA_QR_UT_Accum_T_blk_var1 (FLA_Obj A, FLA_Obj T, fla_qrut_t *cntl) |
FLA_Error FLA_QR_UT_Accum_T_blk_var1 | ( | FLA_Obj | A, | |
FLA_Obj | T, | |||
fla_qrut_t * | cntl | |||
) |
References FLA_Axpy_external(), FLA_Cont_with_3x3_to_2x2(), FLA_Copy_external(), FLA_Copyt_external(), FLA_Determine_blocksize(), FLA_Gemm_external(), FLA_Merge_2x1(), FLA_MINUS_ONE, FLA_Obj_min_dim(), FLA_ONE, FLA_Part_2x2(), FLA_QR_UT_internal(), FLA_Repart_2x2_to_3x3(), FLA_Trmm_external(), and FLA_Trsm_external().
Referenced by FLA_QR_UT_internal().
00036 { 00037 FLA_Obj ATL, ATR, A00, A01, A02, 00038 ABL, ABR, A10, A11, A12, 00039 A20, A21, A22; 00040 00041 FLA_Obj TTL, TTR, T00, T01, T02, 00042 TBL, TBR, T10, T11, W12, 00043 T20, T21, T22; 00044 00045 FLA_Obj AB1; 00046 00047 dim_t b; 00048 00049 FLA_Part_2x2( A, &ATL, &ATR, 00050 &ABL, &ABR, 0, 0, FLA_TL ); 00051 00052 FLA_Part_2x2( T, &TTL, &TTR, 00053 &TBL, &TBR, 0, 0, FLA_TL ); 00054 00055 while ( FLA_Obj_min_dim( ABR ) > 0 ){ 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_2x2_to_3x3( TTL, /**/ TTR, &T00, /**/ &T01, &T02, 00066 /* ************* */ /* ******************** */ 00067 &T10, /**/ &T11, &W12, 00068 TBL, /**/ TBR, &T20, /**/ &T21, &T22, 00069 b, b, FLA_BR ); 00070 00071 /*------------------------------------------------------------*/ 00072 00073 /* 00074 [ U1, T11 ] = FLA_QR_UT_internal( [ A11 00075 A21 ], T11 ); 00076 */ 00077 00078 FLA_Merge_2x1( A11, 00079 A21, &AB1 ); 00080 00081 FLA_QR_UT_internal( AB1, T11, 00082 FLA_Cntl_sub_qrut( cntl ) ); 00083 00084 /* 00085 U11 = trilu( A11 ); 00086 U21 = A21; 00087 00088 W12 = triu( inv(T11) )' * ( U11' * A12 + U21' * A22 ); 00089 */ 00090 00091 FLA_Copy_external( A12, W12 ); 00092 FLA_Trmm_external( FLA_LEFT, FLA_LOWER_TRIANGULAR, 00093 FLA_CONJ_TRANSPOSE, FLA_UNIT_DIAG, 00094 FLA_ONE, A11, W12 ); 00095 FLA_Gemm_external( FLA_CONJ_TRANSPOSE, FLA_NO_TRANSPOSE, 00096 FLA_ONE, A21, A22, FLA_ONE, W12 ); 00097 FLA_Trsm_external( FLA_LEFT, FLA_UPPER_TRIANGULAR, 00098 FLA_CONJ_TRANSPOSE, FLA_NONUNIT_DIAG, 00099 FLA_ONE, T11, W12 ); 00100 00101 /* 00102 A22 = A22 - U21 * W12; 00103 A12 = A12 - U11 * W12; 00104 */ 00105 00106 FLA_Gemm_external( FLA_NO_TRANSPOSE, FLA_NO_TRANSPOSE, 00107 FLA_MINUS_ONE, A21, W12, FLA_ONE, A22 ); 00108 FLA_Trmm_external( FLA_LEFT, FLA_LOWER_TRIANGULAR, 00109 FLA_NO_TRANSPOSE, FLA_UNIT_DIAG, 00110 FLA_MINUS_ONE, A11, W12 ); 00111 FLA_Axpy_external( FLA_ONE, W12, A12 ); 00112 00113 /* 00114 Update T 00115 00116 T01 = A10' * U11 + A20' * U21; 00117 00118 Recall: U11 = trilu( A11 ); 00119 U21 = A21; 00120 */ 00121 00122 FLA_Copyt_external( FLA_CONJ_TRANSPOSE, A10, T01 ); 00123 FLA_Trmm_external( FLA_RIGHT, FLA_LOWER_TRIANGULAR, 00124 FLA_NO_TRANSPOSE, FLA_UNIT_DIAG, 00125 FLA_ONE, A11, T01 ); 00126 FLA_Gemm_external( FLA_CONJ_TRANSPOSE, FLA_NO_TRANSPOSE, 00127 FLA_ONE, A20, A21, FLA_ONE, T01 ); 00128 00129 /*------------------------------------------------------------*/ 00130 00131 FLA_Cont_with_3x3_to_2x2( &ATL, /**/ &ATR, A00, A01, /**/ A02, 00132 A10, A11, /**/ A12, 00133 /* ************** */ /* ****************** */ 00134 &ABL, /**/ &ABR, A20, A21, /**/ A22, 00135 FLA_TL ); 00136 00137 FLA_Cont_with_3x3_to_2x2( &TTL, /**/ &TTR, T00, T01, /**/ T02, 00138 T10, T11, /**/ W12, 00139 /* ************** */ /* ****************** */ 00140 &TBL, /**/ &TBR, T20, T21, /**/ T22, 00141 FLA_TL ); 00142 00143 } 00144 00145 return FLA_SUCCESS; 00146 }