| 1 |
/*! \file DissectionMatrix.hpp |
| 2 |
\brief management of threads for factorization and Fw/Bw substitution |
| 3 |
\author Atsushi Suzuki, Laboratoire Jacques-Louis Lions |
| 4 |
\date Jun. 20th 2014 |
| 5 |
\date Jul. 12th 2015 |
| 6 |
\date Nov. 30th 2016 |
| 7 |
*/ |
| 8 |
|
| 9 |
// This file is part of Dissection |
| 10 |
// |
| 11 |
// Dissection is free software: you can redistribute it and/or modify |
| 12 |
// it under the terms of the GNU General Public License as published by |
| 13 |
// the Free Software Foundation, either version 3 of the License, or |
| 14 |
// (at your option) any later version. |
| 15 |
// |
| 16 |
// Linking Dissection statically or dynamically with other modules is making |
| 17 |
// a combined work based on Disssection. Thus, the terms and conditions of |
| 18 |
// the GNU General Public License cover the whole combination. |
| 19 |
// |
| 20 |
// As a special exception, the copyright holders of Dissection give you |
| 21 |
// permission to combine Dissection program with free software programs or |
| 22 |
// libraries that are released under the GNU LGPL and with independent modules |
| 23 |
// that communicate with Dissection solely through the Dissection-fortran |
| 24 |
// interface. You may copy and distribute such a system following the terms of |
| 25 |
// the GNU GPL for Dissection and the licenses of the other code concerned, |
| 26 |
// provided that you include the source code of that other code when and as |
| 27 |
// the GNU GPL requires distribution of source code and provided that you do |
| 28 |
// not modify the Dissection-fortran interface. |
| 29 |
// |
| 30 |
// Note that people who make modified versions of Dissection are not obligated |
| 31 |
// to grant this special exception for their modified versions; it is their |
| 32 |
// choice whether to do so. The GNU General Public License gives permission to |
| 33 |
// release a modified version without this exception; this exception also makes |
| 34 |
// it possible to release a modified version which carries forward this |
| 35 |
// exception. If you modify the Dissection-fortran interface, this exception |
| 36 |
// does not apply to your modified version of Dissection, and you must remove |
| 37 |
// this exception when you distribute your modified version. |
| 38 |
// |
| 39 |
// This exception is an additional permission under section 7 of the GNU |
| 40 |
// General Public License, version 3 ("GPLv3") |
| 41 |
// |
| 42 |
// Dissection is distributed in the hope that it will be useful, |
| 43 |
// but WITHOUT ANY WARRANTY; without even the implied warranty of |
| 44 |
// MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the |
| 45 |
// GNU General Public License for more details. |
| 46 |
// |
| 47 |
// You should have received a copy of the GNU General Public License |
| 48 |
// along with Dissection. If not, see <http://www.gnu.org/licenses/>. |
| 49 |
// |
| 50 |
|
| 51 |
# ifndef _DRIVER_DISSECTIONMATRIX_ |
| 52 |
# define _DRIVER_DISSECTIONMATRIX_ |
| 53 |
#include <complex> |
| 54 |
#include <vector> |
| 55 |
#include "Splitters/BisectionTree.hpp" |
| 56 |
#include "Algebra/PlainMatrix.hpp" |
| 57 |
#include "Algebra/SquareMatrix.hpp" |
| 58 |
#include "Algebra/SquareBlockMatrix.hpp" |
| 59 |
#include "Algebra/RectBlockMatrix.hpp" |
| 60 |
#include "Algebra/SparseMatrix.hpp" |
| 61 |
#include "Algebra/ColumnMatrix.hpp" |
| 62 |
#include "Driver/TridiagBlockMatrix.hpp" |
| 63 |
#include "Driver/C_threads_tasks.hpp" |
| 64 |
|
| 65 |
using std::vector; |
| 66 |
|
| 67 |
template<typename T, typename U = T> |
| 68 |
class DissectionMatrix |
| 69 |
{ |
| 70 |
public: |
| 71 |
|
| 72 |
DissectionMatrix(Dissection::Tree *btree, |
| 73 |
const int nb, const bool isSym, const bool verbose, |
| 74 |
FILE *fp); |
| 75 |
|
| 76 |
DissectionMatrix(const DissectionMatrix<T, U> & im) |
| 77 |
{ |
| 78 |
(*this) = im; |
| 79 |
} |
| 80 |
|
| 81 |
~DissectionMatrix() |
| 82 |
{ |
| 83 |
if (_islast) { |
| 84 |
delete [] _color_mask; |
| 85 |
for (int i = 0; i < _colors; i++) { |
| 86 |
_tridiag[i]->free(); |
| 87 |
delete _tridiag[i]; |
| 88 |
} |
| 89 |
delete [] _tridiag; |
| 90 |
} |
| 91 |
_diag->free(); // free() is compatible to zero-sized matrix |
| 92 |
_lower->free(); |
| 93 |
_upper->free(); |
| 94 |
_localSchur->free(); // this may be redundant : 26 Jun.2014 |
| 95 |
delete _factorize_LDLt; |
| 96 |
delete _diag; |
| 97 |
delete _lower; |
| 98 |
delete _upper; |
| 99 |
delete _localSchur; |
| 100 |
} |
| 101 |
|
| 102 |
DissectionMatrix& operator = ( const DissectionMatrix<T, U> &im) |
| 103 |
{ |
| 104 |
if ( &im != this) { |
| 105 |
_nb = im._nb; |
| 106 |
_level = im._level; |
| 107 |
_nrow = im._nrow; |
| 108 |
_ncol_offdiag = im._ncol_offdiag; |
| 109 |
_nop = im._nop; |
| 110 |
_diag = im._diag->clone(); |
| 111 |
if (!_isSym) { |
| 112 |
_lower->copy(*im._lower); |
| 113 |
} |
| 114 |
_upper->copy(*im._upper); |
| 115 |
_isSym = im._isSym; |
| 116 |
_localSchur = &im.localSchurBlock(); |
| 117 |
_factorize_LDLt = im._factorize_LDLt; |
| 118 |
// _factorize_LDLt_diag = im._factorize_LDLt_diag; |
| 119 |
_csr_diag = im._csr_diag; |
| 120 |
_csr_offdiag = im._csr_offdiag; |
| 121 |
_alignedFather = im._alignedFather; |
| 122 |
} |
| 123 |
return *this; |
| 124 |
} |
| 125 |
int ColorTridiagBlockMatrix() { return _colors ; } |
| 126 |
void setColorTridiagBlockMatrix(int colors) { _colors = colors; } |
| 127 |
int nrow() { return _nrow; } |
| 128 |
SquareBlockMatrix<T>* addrdiagBlock() { return _diag; } |
| 129 |
SquareBlockMatrix<T>*& paddrdiagBlock() { return _diag; } |
| 130 |
SquareBlockMatrix<T>& diagBlock() { return *_diag; } |
| 131 |
const SquareBlockMatrix<T>& diagBlock() const { return *_diag;} |
| 132 |
TridiagBlockMatrix<T, U> **addrtridiagBlock() { return _tridiag; } |
| 133 |
TridiagBlockMatrix<T, U> **&paddrtridiagBlock() { return _tridiag; } |
| 134 |
RectBlockMatrix<T>* addrupperBlock() { return _upper; } |
| 135 |
RectBlockMatrix<T>*& paddrupperBlock() { return _upper; } |
| 136 |
RectBlockMatrix<T>& upperBlock() { return *_upper; } |
| 137 |
const RectBlockMatrix<T>& upperBlock() const { return *_upper; } |
| 138 |
RectBlockMatrix<T>* addrlowerBlock() { return _lower; } |
| 139 |
RectBlockMatrix<T>*& paddrlowerBlock() { return _lower; } |
| 140 |
RectBlockMatrix<T>& lowerBlock() { return *_lower; } |
| 141 |
const RectBlockMatrix<T>& lowerBlock() const { return *_lower; } |
| 142 |
bool isSym() const { return _isSym; } |
| 143 |
void setAlignedFather() { _alignedFather = true; } |
| 144 |
bool isAlignedFather() const { return _alignedFather; } |
| 145 |
bool isFactorized() const { return _diag->isFactorized(); } |
| 146 |
PlainMatrix<T>& loclSchurBlock(); |
| 147 |
const vector<int>& singIdxPermute() const { return _diag->getSingIdx(); } |
| 148 |
vector<int>& singIdxPermute() { return _diag->getSingIdx(); } |
| 149 |
const vector<int>& singIdx() const { return _diag->getSingIdx0(); } |
| 150 |
vector<int>& singIdx() { return _diag->getSingIdx0(); } |
| 151 |
int KernelDetected() const { return _diag->KernelDetected(); } |
| 152 |
int KernelDim() const { return _diag->dim_kern(); } |
| 153 |
void SetKernelDim(int nsing) { _diag->set_dim_kernel(nsing); } |
| 154 |
SquareMatrix<T>& localSchurBlock() { return *_localSchur; } |
| 155 |
const SquareMatrix<T>& localSchurBlock() const { return *_localSchur; } |
| 156 |
T getLastPivot() const { return _diag->lastPivot(); } |
| 157 |
void setLastPivot(T pivot_val) { _diag->set_lastPivot(pivot_val); } |
| 158 |
|
| 159 |
void C_SparseSymbFact_queue(vector<C_task*>& queue, |
| 160 |
Dissection::Tree *btree, |
| 161 |
const bool verbose, |
| 162 |
FILE **fp); |
| 163 |
|
| 164 |
void C_SparseNumFact_queue(vector<C_task*>& queue, |
| 165 |
Dissection::Tree *btree, |
| 166 |
int nnz, |
| 167 |
T *coefs, |
| 168 |
double *eps_pivot, |
| 169 |
double *pivot, |
| 170 |
bool *kernel_detection, |
| 171 |
bool *higher_precision, |
| 172 |
int *aug_dim, |
| 173 |
U *eps_machine, |
| 174 |
vector<C_task*>& task_q, |
| 175 |
const bool verbose, |
| 176 |
FILE **fp); |
| 177 |
|
| 178 |
void C_SparseLocalSchur_queue(vector<C_task*>& queue, |
| 179 |
Dissection::Tree *btree, |
| 180 |
int nnz, |
| 181 |
T *cofes, |
| 182 |
vector<C_task*>& task_p, |
| 183 |
const bool verbose, |
| 184 |
FILE **fp); |
| 185 |
|
| 186 |
void C_FillMatrix_queue(vector<C_task*>& queue, |
| 187 |
int nnz, |
| 188 |
T *cofes, |
| 189 |
const bool verbose, |
| 190 |
FILE **fp); |
| 191 |
|
| 192 |
int C_DFullLDLt_queue(vector<C_task*>& queue, |
| 193 |
vector<int>& task_indcol, |
| 194 |
vector<int>& task_ptr, |
| 195 |
double *eps_piv, |
| 196 |
bool *kernel_detection, |
| 197 |
int *aug_dim, |
| 198 |
U *eps_machine, |
| 199 |
double *pivot, |
| 200 |
double *pivot0, |
| 201 |
double *pivot1, |
| 202 |
vector<C_task*>& task_p, |
| 203 |
const bool isChldrnAlgnd, |
| 204 |
const bool verbose, |
| 205 |
FILE **fp); |
| 206 |
|
| 207 |
int C_DTRSMScale_queue(vector<C_task*> &queue, |
| 208 |
vector<list<int> > &queue_parents_index, |
| 209 |
vector<int> &queue_index, |
| 210 |
vector<int>& indcol, |
| 211 |
vector<int>& ptr, |
| 212 |
Dissection::Tree *btree, |
| 213 |
vector<C_task*>& task_o, |
| 214 |
vector<int>& task_indcol, |
| 215 |
vector<int>& task_ptr, |
| 216 |
vector<C_task*>& task_p, |
| 217 |
const bool verbose, |
| 218 |
FILE **fp); |
| 219 |
|
| 220 |
void C_DTRSMScale_rearrange(vector<C_task*> &queue, |
| 221 |
vector<list<int> > &queue_parents_index, |
| 222 |
vector<int> &queue_index, |
| 223 |
vector<int>& indcol, |
| 224 |
vector<int>& ptr, |
| 225 |
const bool verbose, |
| 226 |
FILE *fp); |
| 227 |
|
| 228 |
int C_DGEMM_local_queue(vector<C_task*> &queue, |
| 229 |
vector<int> &indcol, |
| 230 |
RectBlockMatrix<T> *upper1, |
| 231 |
RectBlockMatrix<T> *lower1, |
| 232 |
bool isSkip, |
| 233 |
bool isDirect, |
| 234 |
SquareBlockMatrix<T>* fdiag, |
| 235 |
vector<C_task *> &task_p, |
| 236 |
vector<int> &task_p_index, |
| 237 |
vector<int> &task_p_indcol, |
| 238 |
vector<int> &task_p_ptr, |
| 239 |
vector<C_task *> &task_q, |
| 240 |
vector<int> &task_q_index, |
| 241 |
vector<int> &task_q_indcol, |
| 242 |
vector<int> &task_q_ptr, |
| 243 |
vector<C_task *> *task_s, |
| 244 |
const bool verbose, |
| 245 |
FILE **fp); |
| 246 |
|
| 247 |
void C_deallocLocalSchur_queue(vector<C_task*> &queue, |
| 248 |
vector<int> &indcol, |
| 249 |
const bool verbose, |
| 250 |
FILE **fp); |
| 251 |
|
| 252 |
void ChildContrib(list<child_contribution<T> > *child_contribs, |
| 253 |
Dissection::Tree *btree, |
| 254 |
vector<DissectionMatrix<T, U>* >& dissectionMatrix, |
| 255 |
const bool verbose, |
| 256 |
FILE **fp); |
| 257 |
|
| 258 |
void deallocLower_queue(C_task*& queue, |
| 259 |
bool isSym, |
| 260 |
vector<C_task*> &task_p, |
| 261 |
bool isDirect, |
| 262 |
vector<C_task*> &task_q); |
| 263 |
|
| 264 |
int getUpperNRow() const { return upperBlock().nbRows(); } |
| 265 |
int getUpperNCol() const { return upperBlock().nbColumns(); } |
| 266 |
|
| 267 |
unsigned nb() const { return _nb; } |
| 268 |
unsigned level() const { return _level; } |
| 269 |
int ncol_offdiag() const {return _ncol_offdiag; } |
| 270 |
int nop() const { return _nop; } |
| 271 |
bool islast() const { return _islast; } |
| 272 |
bool alignedFather() const { return _alignedFather; } |
| 273 |
|
| 274 |
void setNb(unsigned nb) { _nb = nb; } |
| 275 |
void setLevel(unsigned level) { _level = level; } |
| 276 |
void setNrow(int nrow) { _nrow = nrow; } |
| 277 |
void setNcol_offdiag(int ncol_offdiag) { _ncol_offdiag = ncol_offdiag; } |
| 278 |
void setNop(int nop) { _nop = nop; } |
| 279 |
void setIsSym(bool isSym) { _isSym = isSym; } |
| 280 |
void setIslast(bool islast) { _islast = islast; } |
| 281 |
void setAlignedFather(bool alignedFather) { _alignedFather = alignedFather; } |
| 282 |
|
| 283 |
private: |
| 284 |
unsigned _nb; |
| 285 |
unsigned _level; |
| 286 |
int _nrow; |
| 287 |
int _ncol_offdiag; |
| 288 |
int _nop; // complexity of factorization |
| 289 |
SquareBlockMatrix<T>* _diag; // pointer to the diagonal block |
| 290 |
TridiagBlockMatrix<T, U>** _tridiag; // pointer to the tridiagonal blocks |
| 291 |
RectBlockMatrix<T>* _lower; // pointer to lower off-diagonal block |
| 292 |
RectBlockMatrix<T>* _upper; // pointer to upper off-diagonal block |
| 293 |
bool _isSym; |
| 294 |
bool _islast; |
| 295 |
bool _alignedFather; |
| 296 |
int _colors; |
| 297 |
int *_color_mask; |
| 298 |
|
| 299 |
SquareBlockMatrix<T> *_localSchur; |
| 300 |
ColumnMatrix<T> *_factorize_LDLt; |
| 301 |
// ColumnMatrix<T> *_factorize_LDLt_diag; |
| 302 |
const CSR_indirect *_csr_diag; |
| 303 |
const CSR_indirect *_csr_offdiag; |
| 304 |
}; |
| 305 |
|
| 306 |
template<typename T> |
| 307 |
class SchurMatrix |
| 308 |
{ |
| 309 |
public: |
| 310 |
SchurMatrix() { _full_pivoting = false;} |
| 311 |
~SchurMatrix() { } |
| 312 |
void free() { |
| 313 |
_sldu.free(); |
| 314 |
_sldu_list.clear(); // |
| 315 |
_sldu_list_left.clear(); |
| 316 |
delete _arow; |
| 317 |
delete _acol; |
| 318 |
_scol.free(); |
| 319 |
_schur.free(); |
| 320 |
} |
| 321 |
SubSquareMatrix<T>& getSldu() { return _sldu; } |
| 322 |
vector<int>& getSlduList() { return _sldu_list; } |
| 323 |
vector<int>& getSlduListLeft() { |
| 324 |
return _full_pivoting ? _sldu_list_left : _sldu_list; |
| 325 |
} |
| 326 |
SparseMatrix<T>*& getArow() { return _arow; } |
| 327 |
SparseMatrix<T>*& getAcol() { return _acol; } |
| 328 |
ColumnMatrix<T>& getScol() { return _scol; } |
| 329 |
ColumnMatrix<T>& getSchur() { return _schur; } // for debugging |
| 330 |
bool isFullPivoting() { return _full_pivoting; } |
| 331 |
void setFullPivoting(const bool full_pivoting) { |
| 332 |
_full_pivoting = full_pivoting; |
| 333 |
} |
| 334 |
private: |
| 335 |
SubSquareMatrix<T> _sldu; |
| 336 |
vector<int> _sldu_list; // _sldu_list_right; |
| 337 |
vector<int> _sldu_list_left; |
| 338 |
SparseMatrix<T>* _arow; |
| 339 |
SparseMatrix<T>* _acol; |
| 340 |
ColumnMatrix<T> _scol; |
| 341 |
ColumnMatrix<T> _schur; |
| 342 |
bool _full_pivoting; |
| 343 |
}; |
| 344 |
|
| 345 |
template<typename T> |
| 346 |
class KernelMatrix |
| 347 |
{ |
| 348 |
public: |
| 349 |
KernelMatrix(int dimension = 0) : _dimension(dimension), |
| 350 |
_full_pivoting(false) { } |
| 351 |
~KernelMatrix() { } |
| 352 |
void free() |
| 353 |
{ |
| 354 |
_singIdx.clear(); |
| 355 |
_kern_list_eq.clear(); |
| 356 |
_kern_basis.free(); |
| 357 |
_tkern_basis.free(); |
| 358 |
_kern_proj.free(); |
| 359 |
_tkern_proj.free(); |
| 360 |
_ntkern_proj.free(); |
| 361 |
} |
| 362 |
int dimension() const { return _dimension; } |
| 363 |
void set_dimension(int dimension){ _dimension = dimension; } |
| 364 |
vector<int>& getSingIdx() { return _singIdx; } |
| 365 |
vector<int>& getKernListEq() { |
| 366 |
return _kern_list_eq; |
| 367 |
} |
| 368 |
vector<int>& getKernListEqLeft() { |
| 369 |
return _full_pivoting ? _kern_list_eq_left : _kern_list_eq; |
| 370 |
} |
| 371 |
ColumnMatrix<T>& getKernBasis() {return _kern_basis; } |
| 372 |
ColumnMatrix<T>& getTKernBasis() { return _tkern_basis; } |
| 373 |
SquareMatrix<T>& getKernProj() { return _kern_proj; } |
| 374 |
SquareMatrix<T>& getTKernProj() { return _tkern_proj; } |
| 375 |
SquareMatrix<T>& getNTKernProj() { return _ntkern_proj; } |
| 376 |
void setFullPivoting(const bool full_pivoting) { |
| 377 |
_full_pivoting = full_pivoting; |
| 378 |
} |
| 379 |
private: |
| 380 |
int _dimension; |
| 381 |
vector<int> _singIdx; |
| 382 |
vector<int> _kern_list_eq; |
| 383 |
vector<int> _kern_list_eq_left; |
| 384 |
ColumnMatrix<T> _kern_basis; |
| 385 |
ColumnMatrix<T> _tkern_basis; |
| 386 |
SquareMatrix<T> _kern_proj; |
| 387 |
SquareMatrix<T> _tkern_proj; |
| 388 |
SquareMatrix<T> _ntkern_proj; |
| 389 |
bool _full_pivoting; |
| 390 |
}; |
| 391 |
|
| 392 |
#endif |