Functions | |
FLA_Error | FLA_Apply_Q_UT_lhc_blk_var1 (FLA_Obj A, FLA_Obj T, FLA_Obj W1, FLA_Obj B, fla_apqut_t *cntl) |
FLA_Error FLA_Apply_Q_UT_lhc_blk_var1 | ( | FLA_Obj | A, | |
FLA_Obj | T, | |||
FLA_Obj | W1, | |||
FLA_Obj | B, | |||
fla_apqut_t * | cntl | |||
) |
References FLA_Axpy_internal(), FLA_Cont_with_1x3_to_1x2(), FLA_Cont_with_3x1_to_2x1(), FLA_Cont_with_3x3_to_2x2(), FLA_Copy_internal(), FLA_Determine_blocksize(), FLA_Gemm_internal(), FLA_MINUS_ONE, FLA_Obj_min_dim(), FLA_Obj_width(), FLA_ONE, FLA_Part_1x2(), FLA_Part_2x1(), FLA_Part_2x2(), FLA_Repart_1x2_to_1x3(), FLA_Repart_2x1_to_3x1(), FLA_Repart_2x2_to_3x3(), FLA_Trmm_internal(), and FLA_Trsm_internal().
Referenced by FLA_Apply_Q_UT_lhc().
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, T2; 00042 00043 FLA_Obj T1T, 00044 T2B; 00045 00046 FLA_Obj W1TL, W1TR, 00047 W1BL, W1BR; 00048 00049 FLA_Obj BT, B0, 00050 BB, B1, 00051 B2; 00052 00053 dim_t b; 00054 00055 FLA_Part_2x2( A, &ATL, &ATR, 00056 &ABL, &ABR, 0, 0, FLA_TL ); 00057 00058 FLA_Part_1x2( T, &TL, &TR, 0, FLA_LEFT ); 00059 00060 FLA_Part_2x1( B, &BT, 00061 &BB, 0, FLA_TOP ); 00062 00063 while ( FLA_Obj_min_dim( ABR ) > 0 ){ 00064 00065 b = FLA_Determine_blocksize( ABR, FLA_BR, FLA_Cntl_blocksize( cntl ) ); 00066 00067 FLA_Repart_2x2_to_3x3( ATL, /**/ ATR, &A00, /**/ &A01, &A02, 00068 /* ************* */ /* ******************** */ 00069 &A10, /**/ &A11, &A12, 00070 ABL, /**/ ABR, &A20, /**/ &A21, &A22, 00071 b, b, FLA_BR ); 00072 00073 FLA_Repart_1x2_to_1x3( TL, /**/ TR, &T0, /**/ &T1, &T2, 00074 b, FLA_RIGHT ); 00075 00076 FLA_Repart_2x1_to_3x1( BT, &B0, 00077 /* ** */ /* ** */ 00078 &B1, 00079 BB, &B2, b, FLA_BOTTOM ); 00080 00081 /*------------------------------------------------------------*/ 00082 00083 FLA_Part_2x1( T1, &T1T, 00084 &T2B, b, FLA_TOP ); 00085 00086 // W1T = FLA_Top_part( W1, b ); 00087 // W1TL = FLA_Left_part( W1T, FLA_Obj_width( B1 ) ); 00088 FLA_Part_2x2( W1, &W1TL, &W1TR, 00089 &W1BL, &W1BR, b, FLA_Obj_width( B1 ), FLA_TL ); 00090 00091 // W1TL = B1; 00092 FLA_Copy_internal( B1, W1TL, 00093 FLA_Cntl_sub_copy( cntl ) ); 00094 00095 // U11 = trilu( A11 ); 00096 // U21 = A21; 00097 00098 // W1TL = inv( triu( T1T ) )' * ( U11' * B1 + U21' * B2 ); 00099 00100 FLA_Trmm_internal( FLA_LEFT, FLA_LOWER_TRIANGULAR, 00101 FLA_CONJ_TRANSPOSE, FLA_UNIT_DIAG, 00102 FLA_ONE, A11, W1TL, 00103 FLA_Cntl_sub_trmm1( cntl ) ); 00104 00105 FLA_Gemm_internal( FLA_CONJ_TRANSPOSE, FLA_NO_TRANSPOSE, 00106 FLA_ONE, A21, B2, FLA_ONE, W1TL, 00107 FLA_Cntl_sub_gemm1( cntl ) ); 00108 00109 FLA_Trsm_internal( FLA_LEFT, FLA_UPPER_TRIANGULAR, 00110 FLA_CONJ_TRANSPOSE, FLA_NONUNIT_DIAG, 00111 FLA_ONE, T1T, W1TL, 00112 FLA_Cntl_sub_trsm( cntl ) ); 00113 00114 // B2 = B2 - U21 * W1TL; 00115 // B1 = B1 - U11 * W1TL; 00116 00117 FLA_Gemm_internal( FLA_NO_TRANSPOSE, FLA_NO_TRANSPOSE, 00118 FLA_MINUS_ONE, A21, W1TL, FLA_ONE, B2, 00119 FLA_Cntl_sub_gemm2( cntl ) ); 00120 00121 FLA_Trmm_internal( FLA_LEFT, FLA_LOWER_TRIANGULAR, 00122 FLA_NO_TRANSPOSE, FLA_UNIT_DIAG, 00123 FLA_MINUS_ONE, A11, W1TL, 00124 FLA_Cntl_sub_trmm2( cntl ) ); 00125 00126 FLA_Axpy_internal( FLA_ONE, W1TL, B1, 00127 FLA_Cntl_sub_axpy( cntl ) ); 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, /**/ T2, 00138 FLA_LEFT ); 00139 00140 FLA_Cont_with_3x1_to_2x1( &BT, B0, 00141 B1, 00142 /* ** */ /* ** */ 00143 &BB, B2, FLA_TOP ); 00144 } 00145 00146 return FLA_SUCCESS; 00147 }