Go to the source code of this file.
Functions | |
FLA_Error | FLA_LQ_UT_Accum_T_blk_var1 (FLA_Obj A, FLA_Obj T, fla_lqut_t *cntl) |
FLA_Error | FLA_LQ_UT_blk_var2 (FLA_Obj A, FLA_Obj T, fla_lqut_t *cntl) |
FLA_Error FLA_LQ_UT_Accum_T_blk_var1 | ( | FLA_Obj | A, | |
FLA_Obj | T, | |||
fla_lqut_t * | cntl | |||
) |
References FLA_Axpyt_external(), FLA_Cont_with_3x3_to_2x2(), FLA_Copy_external(), FLA_Copyt_external(), FLA_Determine_blocksize(), FLA_Gemm_external(), FLA_LQ_UT_internal(), FLA_Merge_1x2(), FLA_MINUS_ONE, FLA_Obj_min_dim(), FLA_ONE, FLA_Part_2x2(), FLA_Repart_2x2_to_3x3(), FLA_Trmm_external(), and FLA_Trsm_external().
Referenced by FLA_LQ_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_LQ_UT_internal( [ A11 A12 ], T11 ); 00075 */ 00076 00077 FLA_Merge_1x2( A11, A12, &AB1 ); 00078 00079 FLA_LQ_UT_internal( AB1, T11, 00080 FLA_Cntl_sub_lqut( cntl ) ); 00081 00082 /* 00083 U11 = triuu( A11 )'; 00084 U21 = A12'; 00085 00086 W12 = triu( inv(T11) )' * ( U11' * A21' + U21' * A22' ); 00087 */ 00088 00089 FLA_Copyt_external( FLA_CONJ_TRANSPOSE, A21, W12 ); 00090 FLA_Trmm_external( FLA_LEFT, FLA_UPPER_TRIANGULAR, 00091 FLA_NO_TRANSPOSE, FLA_UNIT_DIAG, 00092 FLA_ONE, A11, W12 ); 00093 FLA_Gemm_external( FLA_NO_TRANSPOSE, FLA_CONJ_TRANSPOSE, 00094 FLA_ONE, A12, A22, FLA_ONE, W12 ); 00095 FLA_Trsm_external( FLA_LEFT, FLA_UPPER_TRIANGULAR, 00096 FLA_CONJ_TRANSPOSE, FLA_NONUNIT_DIAG, 00097 FLA_ONE, T11, W12 ); 00098 00099 /* 00100 A22 = A22 - W12' * U21'; 00101 A21 = A21 - W12' * U11'; 00102 */ 00103 00104 FLA_Gemm_external( FLA_CONJ_TRANSPOSE, FLA_NO_TRANSPOSE, 00105 FLA_MINUS_ONE, W12, A12, FLA_ONE, A22 ); 00106 FLA_Trmm_external( FLA_LEFT, FLA_UPPER_TRIANGULAR, 00107 FLA_CONJ_TRANSPOSE, FLA_UNIT_DIAG, 00108 FLA_MINUS_ONE, A11, W12 ); 00109 FLA_Axpyt_external( FLA_CONJ_TRANSPOSE, FLA_ONE, W12, A21 ); 00110 00111 /* 00112 Update T 00113 00114 T01 = A01 * U11 + A02 * U21; 00115 00116 Recall: U11 = triuu( A11 )'; 00117 U21 = A12'; 00118 */ 00119 00120 FLA_Copy_external( A01, T01 ); 00121 FLA_Trmm_external( FLA_RIGHT, FLA_UPPER_TRIANGULAR, 00122 FLA_CONJ_TRANSPOSE, FLA_UNIT_DIAG, 00123 FLA_ONE, A11, T01 ); 00124 FLA_Gemm_external( FLA_NO_TRANSPOSE, FLA_CONJ_TRANSPOSE, 00125 FLA_ONE, A02, A12, FLA_ONE, T01 ); 00126 00127 /*------------------------------------------------------------*/ 00128 00129 FLA_Cont_with_3x3_to_2x2( &ATL, /**/ &ATR, A00, A01, /**/ A02, 00130 A10, A11, /**/ A12, 00131 /* ************** */ /* ****************** */ 00132 &ABL, /**/ &ABR, A20, A21, /**/ A22, 00133 FLA_TL ); 00134 00135 FLA_Cont_with_3x3_to_2x2( &TTL, /**/ &TTR, T00, T01, /**/ T02, 00136 T10, T11, /**/ W12, 00137 /* ************** */ /* ****************** */ 00138 &TBL, /**/ &TBR, T20, T21, /**/ T22, 00139 FLA_TL ); 00140 00141 } 00142 00143 return FLA_SUCCESS; 00144 }
FLA_Error FLA_LQ_UT_blk_var2 | ( | FLA_Obj | A, | |
FLA_Obj | T, | |||
fla_lqut_t * | cntl | |||
) |
References FLA_Axpy_external(), FLA_Cont_with_3x1_to_2x1(), FLA_Cont_with_3x3_to_2x2(), FLA_Copy_external(), FLA_Determine_blocksize(), FLA_Gemm_external(), FLA_LQ_UT_internal(), FLA_Merge_1x2(), FLA_MINUS_ONE, FLA_Obj_length(), FLA_Obj_min_dim(), FLA_ONE, FLA_Part_1x2(), FLA_Part_2x1(), FLA_Part_2x2(), FLA_Repart_2x1_to_3x1(), FLA_Repart_2x2_to_3x3(), FLA_Trmm_external(), and FLA_Trsm_external().
Referenced by FLA_LQ_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 TT, T0, 00042 TB, T1, 00043 W21; 00044 00045 FLA_Obj W21TL, W21TR, 00046 W21BL, W21BR; 00047 00048 FLA_Obj T1L, T2R, AB1; 00049 00050 dim_t b; 00051 00052 FLA_Part_2x2( A, &ATL, &ATR, 00053 &ABL, &ABR, 0, 0, FLA_TL ); 00054 00055 FLA_Part_2x1( T, &TT, 00056 &TB, 0, FLA_TOP ); 00057 00058 while ( FLA_Obj_min_dim( ABR ) > 0 ){ 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( TT, &T0, 00069 /* ** */ /* ** */ 00070 &T1, 00071 TB, &W21, b, FLA_BOTTOM ); 00072 00073 /*------------------------------------------------------------*/ 00074 00075 /* 00076 T1L = FLA_Left_part( T1, b ); 00077 */ 00078 00079 FLA_Part_1x2( T1, &T1L, &T2R, b, FLA_LEFT ); 00080 00081 /* 00082 [ U1, T1L ] = FLA_LQ_UT_internal( [ A11 A12 ], T1L ); 00083 */ 00084 00085 FLA_Merge_1x2( A11, A12, &AB1 ); 00086 00087 FLA_LQ_UT_internal( AB1, T1L, 00088 FLA_Cntl_sub_lqut( cntl )); 00089 00090 00091 if ( FLA_Obj_length( A21 ) > 0 ) 00092 { 00093 /* 00094 W21L = FLA_Left_part( W21, b ); 00095 W21TL = FLA_Top_part( W21L, FLA_Obj_length( A21 ) ); 00096 */ 00097 00098 FLA_Part_2x2( W21, &W21TL, &W21TR, 00099 &W21BL, &W21BR, FLA_Obj_length( A21 ), b, FLA_TL ); 00100 00101 /* 00102 U11 = triuu( A11 ); 00103 U12 = A12; 00104 00105 W21TL = ( A21 * U11' + A22 * U12' ) * triu( inv(T1L) ); 00106 */ 00107 00108 FLA_Copy_external( A21, W21TL ); 00109 FLA_Trmm_external( FLA_RIGHT, FLA_UPPER_TRIANGULAR, 00110 FLA_CONJ_TRANSPOSE, FLA_UNIT_DIAG, 00111 FLA_ONE, A11, W21TL ); 00112 FLA_Gemm_external( FLA_NO_TRANSPOSE, FLA_CONJ_TRANSPOSE, 00113 FLA_ONE, A22, A12, FLA_ONE, W21TL ); 00114 FLA_Trsm_external( FLA_RIGHT, FLA_UPPER_TRIANGULAR, 00115 FLA_NO_TRANSPOSE, FLA_NONUNIT_DIAG, 00116 FLA_ONE, T1L, W21TL ); 00117 00118 /* 00119 A22 = A22 - W21TL * U12; 00120 A21 = A21 - W21TL * U11; 00121 */ 00122 00123 FLA_Gemm_external( FLA_NO_TRANSPOSE, FLA_NO_TRANSPOSE, 00124 FLA_MINUS_ONE, W21TL, A12, FLA_ONE, A22 ); 00125 FLA_Trmm_external( FLA_RIGHT, FLA_UPPER_TRIANGULAR, 00126 FLA_NO_TRANSPOSE, FLA_UNIT_DIAG, 00127 FLA_MINUS_ONE, A11, W21TL ); 00128 FLA_Axpy_external( FLA_ONE, W21TL, A21 ); 00129 } 00130 00131 /*------------------------------------------------------------*/ 00132 00133 FLA_Cont_with_3x3_to_2x2( &ATL, /**/ &ATR, A00, A01, /**/ A02, 00134 A10, A11, /**/ A12, 00135 /* ************** */ /* ****************** */ 00136 &ABL, /**/ &ABR, A20, A21, /**/ A22, 00137 FLA_TL ); 00138 00139 FLA_Cont_with_3x1_to_2x1( &TT, T0, 00140 T1, 00141 /* ** */ /* ** */ 00142 &TB, W21, FLA_TOP ); 00143 00144 } 00145 00146 return FLA_SUCCESS; 00147 }