Functions | |
FLA_Error | FLA_QR_UT_blk_var2 (FLA_Obj A, FLA_Obj T, fla_qrut_t *cntl) |
FLA_Error FLA_QR_UT_blk_var2 | ( | FLA_Obj | A, | |
FLA_Obj | T, | |||
fla_qrut_t * | cntl | |||
) |
References FLA_Axpy_external(), FLA_Cont_with_1x3_to_1x2(), FLA_Cont_with_3x3_to_2x2(), FLA_Copy_external(), FLA_Determine_blocksize(), FLA_Gemm_external(), FLA_Merge_2x1(), FLA_MINUS_ONE, FLA_Obj_min_dim(), FLA_Obj_width(), FLA_ONE, FLA_Part_1x2(), FLA_Part_2x1(), FLA_Part_2x2(), FLA_QR_UT_internal(), FLA_Repart_1x2_to_1x3(), 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 TL, TR, T0, T1, W12; 00042 00043 FLA_Obj W12TL, W12TR, 00044 W12BL, W12BR; 00045 00046 FLA_Obj T1T, T2B, AB1; 00047 00048 dim_t b; 00049 00050 FLA_Part_2x2( A, &ATL, &ATR, 00051 &ABL, &ABR, 0, 0, FLA_TL ); 00052 00053 FLA_Part_1x2( T, &TL, &TR, 0, FLA_LEFT ); 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_1x2_to_1x3( TL, /**/ TR, &T0, /**/ &T1, &W12, 00066 b, FLA_RIGHT ); 00067 00068 /*------------------------------------------------------------*/ 00069 00070 /* 00071 T1T = FLA_Top_part( T1, b ); 00072 */ 00073 00074 FLA_Part_2x1( T1, &T1T, 00075 &T2B, b, FLA_TOP ); 00076 00077 /* 00078 [ U1, T1 ] = FLA_QR_UT_internal( [ A11 00079 A21 ], t1, T1T ); 00080 */ 00081 00082 FLA_Merge_2x1( A11, 00083 A21, &AB1 ); 00084 00085 FLA_QR_UT_internal( AB1, T1T, 00086 FLA_Cntl_sub_qrut( cntl ) ); 00087 00088 00089 if ( FLA_Obj_width( A12 ) > 0 ) 00090 { 00091 /* 00092 W12T = FLA_Top_part( W12, b ); 00093 W12TL = FLA_Left_part( W12T, FLA_Obj_width( A12 ) ); 00094 */ 00095 00096 FLA_Part_2x2( W12, &W12TL, &W12TR, 00097 &W12BL, &W12BR, b, FLA_Obj_width( A12 ), FLA_TL ); 00098 00099 /* 00100 U11 = trilu( A11 ); 00101 U21 = A21; 00102 00103 W12TL = triu( inv(T1T) )' * ( U11' * A12 + U21' * A22 ); 00104 */ 00105 00106 FLA_Copy_external( A12, W12TL ); 00107 FLA_Trmm_external( FLA_LEFT, FLA_LOWER_TRIANGULAR, 00108 FLA_CONJ_TRANSPOSE, FLA_UNIT_DIAG, 00109 FLA_ONE, A11, W12TL ); 00110 FLA_Gemm_external( FLA_CONJ_TRANSPOSE, FLA_NO_TRANSPOSE, 00111 FLA_ONE, A21, A22, FLA_ONE, W12TL ); 00112 FLA_Trsm_external( FLA_LEFT, FLA_UPPER_TRIANGULAR, 00113 FLA_CONJ_TRANSPOSE, FLA_NONUNIT_DIAG, 00114 FLA_ONE, T1T, W12TL ); 00115 00116 /* 00117 A22 = A22 - U21 * W12TL; 00118 A12 = A12 - U11 * W12TL; 00119 */ 00120 00121 FLA_Gemm_external( FLA_NO_TRANSPOSE, FLA_NO_TRANSPOSE, 00122 FLA_MINUS_ONE, A21, W12TL, FLA_ONE, A22 ); 00123 FLA_Trmm_external( FLA_LEFT, FLA_LOWER_TRIANGULAR, 00124 FLA_NO_TRANSPOSE, FLA_UNIT_DIAG, 00125 FLA_MINUS_ONE, A11, W12TL ); 00126 FLA_Axpy_external( FLA_ONE, W12TL, A12 ); 00127 } 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_1x3_to_1x2( &TL, /**/ &TR, T0, T1, /**/ W12, 00138 FLA_LEFT ); 00139 00140 } 00141 00142 return FLA_SUCCESS; 00143 }