1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
|
// Copyright (C) 2016 EDF
// All Rights Reserved
// This code is published under the GNU Lesser General Public License (GNU LGPL)
#ifdef USE_MPI
#include <functional>
#include <memory>
#include <boost/mpi.hpp>
#ifdef _OPENMP
#include <omp.h>
#include "StOpt/core/utils/OpenmpException.h"
#endif
#include <Eigen/Dense>
#include "geners/BinaryFileArchive.hh"
#include "geners/Record.hh"
#include "geners/vectorIO.hh"
#include "StOpt/core/parallelism/ParallelComputeGridSplitting.h"
#include "StOpt/core/grids/GridIterator.h"
#include "StOpt/core/utils/eigenGeners.h"
#include "StOpt/regression/ContinuationValue.h"
#include "StOpt/regression/ContinuationValueGeners.h"
#include "StOpt/regression/GridAndRegressedValue.h"
#include "StOpt/regression/GridAndRegressedValueGeners.h"
#include "StOpt/dp/TransitionStepRegressionDPDist.h"
#include "StOpt/core/parallelism/GridReach.h"
using namespace StOpt;
using namespace Eigen;
using namespace std;
TransitionStepRegressionDPDist::TransitionStepRegressionDPDist(const shared_ptr<FullGrid> &p_pGridCurrent,
const shared_ptr<FullGrid> &p_pGridPrevious,
const shared_ptr<OptimizerDPBase > &p_pOptimize,
const boost::mpi::communicator &p_world): TransitionStepBaseDist(p_pGridCurrent, p_pGridPrevious, p_pOptimize, p_world) {}
pair< vector< shared_ptr< ArrayXXd >>, vector< shared_ptr< ArrayXXd > > > TransitionStepRegressionDPDist::oneStep(const vector< shared_ptr< ArrayXXd > > &p_phiIn,
const shared_ptr< BaseRegression> &p_condExp) const
{
// number of regimes at current time
int nbRegimes = m_pOptimize->getNbRegime();
vector< shared_ptr< ArrayXXd > > phiOut(nbRegimes);
int nbControl = m_pOptimize->getNbControl();
vector< shared_ptr< ArrayXXd > > controlOut(nbControl);
// only if the processor is working
vector < shared_ptr< ArrayXXd > > phiInExtended(p_phiIn.size());
// Organize the data splitting : spread the incoming values on an extended grid
for (size_t iReg = 0; iReg < p_phiIn.size() ; ++iReg)
{
// utilitary
ArrayXXd emptyArray;
if (p_phiIn[iReg])
{
phiInExtended[iReg] = make_shared< ArrayXXd >(m_paral->runOneStep(*p_phiIn[iReg])) ;
}
else
phiInExtended[iReg] = make_shared< ArrayXXd >(m_paral->runOneStep(emptyArray)) ;
}
if (m_gridCurrentProc->getNbPoints() > 0)
{
// allocate for solution
for (int iReg = 0; iReg < nbRegimes; ++iReg)
phiOut[iReg] = make_shared< ArrayXXd >(p_condExp->getNbSimul(), m_gridCurrentProc->getNbPoints());
for (int iCont = 0; iCont < nbControl; ++iCont)
controlOut[iCont] = make_shared< ArrayXXd >(p_condExp->getNbSimul(), m_gridCurrentProc->getNbPoints());
// create continuation values on extended grid
vector< ContinuationValue > contVal(p_phiIn.size());
for (size_t iReg = 0; iReg < p_phiIn.size(); ++iReg)
contVal[iReg] = ContinuationValue(m_gridExtendPreviousStep, p_condExp, *phiInExtended[iReg]);
// number of thread
#ifdef _OPENMP
int nbThreads = omp_get_max_threads();
#else
int nbThreads = 1;
#endif
// create iterator on current grid treated for processor
int iThread = 0 ;
#ifdef _OPENMP
OpenmpException excep; // deal with exception in openmp
#pragma omp parallel for private(iThread)
#endif
for (iThread = 0; iThread < nbThreads; ++iThread)
{
#ifdef _OPENMP
excep.run([&]
{
#endif
shared_ptr< GridIterator > iterGridPoint = m_gridCurrentProc->getGridIterator();
// account fo threads
iterGridPoint->jumpToAndInc(0, 1, iThread);
// iterates on points of the grid
while (iterGridPoint->isValid())
{
ArrayXd pointCoord = iterGridPoint->getCoordinate();
// optimize the current point and the set of regimes
pair< ArrayXXd, ArrayXXd> solutionAndControl = static_pointer_cast<OptimizerDPBase>(m_pOptimize)->stepOptimize(m_gridExtendPreviousStep, pointCoord, contVal, phiInExtended);
// copie solution
for (int iReg = 0; iReg < nbRegimes; ++iReg)
(*phiOut[iReg]).col(iterGridPoint->getCount()) = solutionAndControl.first.col(iReg);
for (int iCont = 0; iCont < nbControl; ++iCont)
(*controlOut[iCont]).col(iterGridPoint->getCount()) = solutionAndControl.second.col(iCont);
iterGridPoint->nextInc(nbThreads);
}
#ifdef _OPENMP
});
#endif
}
#ifdef _OPENMP
excep.rethrow();
#endif
}
return make_pair(phiOut, controlOut);
}
void TransitionStepRegressionDPDist::dumpContinuationValues(shared_ptr<gs::BinaryFileArchive> p_ar, const string &p_name, const int &p_iStep,
const vector< shared_ptr< ArrayXXd > > &p_phiInPrev, const vector< shared_ptr< ArrayXXd > > &p_control,
const shared_ptr<BaseRegression> &p_condExp,
const bool &p_bOneFile) const
{
string stepString = boost::lexical_cast<string>(p_iStep) ;
ArrayXi initialDimensionPrev = m_pGridPrevious->getDimensions();
ArrayXi initialDimension = m_pGridCurrent->getDimensions();
if (!p_bOneFile)
{
Array< array<int, 2 >, Dynamic, 1 > gridLocalPrev = m_paral->getPreviousCalculationGrid();
shared_ptr<FullGrid> gridPrevious = m_pGridPrevious->getSubGrid(gridLocalPrev);
Array< array<int, 2 >, Dynamic, 1 > gridLocal = m_paral->getCurrentCalculationGrid();
shared_ptr<FullGrid> gridCurrent = m_pGridCurrent->getSubGrid(gridLocal);
// dump caracteristics of the splitting
// organize the hypercube splitting for parallel
vector<int> vecPrev(initialDimensionPrev.data(), initialDimensionPrev.data() + initialDimensionPrev.size());
*p_ar << gs::Record(vecPrev, "initialSizeOfMeshPrev", stepString.c_str()) ;
vector<int> vecCurrent(initialDimension.data(), initialDimension.data() + initialDimension.size());
*p_ar << gs::Record(vecCurrent, "initialSizeOfMesh", stepString.c_str()) ;
// store regressor
*p_ar << gs::Record(dynamic_cast<const BaseRegression &>(*p_condExp), "regressor", stepString.c_str()) ;
vector<ArrayXXd> regressedValues(p_phiInPrev.size());
if (m_world.rank() < m_paral->getNbProcessorUsedPrev())
{
// regresse the values
for (size_t iReg = 0; iReg < p_phiInPrev.size(); ++iReg)
{
ArrayXXd transposeCont = p_phiInPrev[iReg]->transpose();
regressedValues[iReg] = p_condExp->getCoordBasisFunctionMultiple(transposeCont).transpose();
}
}
*p_ar << gs::Record(regressedValues, (p_name + "Values").c_str(), stepString.c_str()) ;
vector<ArrayXXd> contValues(p_control.size());
if (m_world.rank() < m_paral->getNbProcessorUsed())
{
for (size_t iCont = 0; iCont < p_control.size(); ++iCont)
{
ArrayXXd transposeCont = p_control[iCont]->transpose();
contValues[iCont] = p_condExp->getCoordBasisFunctionMultiple(transposeCont).transpose();
}
}
*p_ar << gs::Record(contValues, (p_name + "Control").c_str(), stepString.c_str()) ;
}
else
{
// utilitary
Array< array<int, 2 >, Dynamic, 1 > gridOnProc0Prev(initialDimensionPrev.size());
for (int id = 0; id < initialDimensionPrev.size(); ++id)
{
gridOnProc0Prev(id)[0] = 0 ;
gridOnProc0Prev(id)[1] = initialDimensionPrev(id) ;
}
ArrayXi splittingRatioPrev = paraOptimalSplitting(initialDimensionPrev, m_pOptimize->getDimensionToSplit(), m_world);
ParallelComputeGridSplitting paralObjectPrev(initialDimensionPrev, splittingRatioPrev, m_world);
vector< GridAndRegressedValue> contVal(p_phiInPrev.size());
for (size_t iReg = 0; iReg < p_phiInPrev.size(); ++iReg)
{
ArrayXXd reconstructedArray ;
if (m_world.rank() < m_paral->getNbProcessorUsedPrev())
reconstructedArray = paralObjectPrev.reconstruct(*p_phiInPrev[iReg], gridOnProc0Prev);
if (m_world.rank() == 0)
contVal[iReg] = GridAndRegressedValue(m_pGridPrevious, p_condExp, reconstructedArray);
}
if (m_world.rank() == 0)
{
*p_ar << gs::Record(contVal, (p_name + "Values").c_str(), stepString.c_str()) ;
}
// now the control
Array< array<int, 2 >, Dynamic, 1 > gridOnProc0(initialDimension.size());
for (int id = 0; id < initialDimension.size(); ++id)
{
gridOnProc0(id)[0] = 0 ;
gridOnProc0(id)[1] = initialDimension(id) ;
}
ArrayXi splittingRatio = paraOptimalSplitting(initialDimension, m_pOptimize->getDimensionToSplit(), m_world);
ParallelComputeGridSplitting paralObject(initialDimension, splittingRatio, m_world);
vector< GridAndRegressedValue > control(p_control.size());
for (size_t iCont = 0; iCont < p_control.size(); ++iCont)
{
ArrayXXd reconstructedArray ;
if (m_world.rank() < m_paral->getNbProcessorUsed())
reconstructedArray = paralObject.reconstruct(*p_control[iCont], gridOnProc0);
if (m_world.rank() == 0)
control[iCont] = GridAndRegressedValue(m_pGridCurrent, p_condExp, reconstructedArray);
}
if (m_world.rank() == 0)
*p_ar << gs::Record(control, (p_name + "Control").c_str(), stepString.c_str()) ;
}
if (m_world.rank() == 0)
p_ar->flush() ; // necessary for python mapping
m_world.barrier() ; // onlyt to prevent the reading in simualtion before the end of writting
}
void TransitionStepRegressionDPDist::dumpBellmanValues(shared_ptr<gs::BinaryFileArchive> p_ar, const string &p_name, const int &p_iStep,
const vector< shared_ptr< ArrayXXd > > &p_phiIn,
const shared_ptr<BaseRegression> &p_condExp,
const bool &p_bOneFile) const
{
string stepString = boost::lexical_cast<string>(p_iStep) ;
ArrayXi initialDimension = m_pGridCurrent->getDimensions();
if (!p_bOneFile)
{
Array< array<int, 2 >, Dynamic, 1 > gridLocal = m_paral->getCurrentCalculationGrid();
shared_ptr<FullGrid> gridCurrent = m_pGridCurrent->getSubGrid(gridLocal);
// dump caracteristics of the splitting
// organize the hypercube splitting for parallel
vector<int> vecCurrent(initialDimension.data(), initialDimension.data() + initialDimension.size());
*p_ar << gs::Record(vecCurrent, "initialSizeOfMesh", stepString.c_str()) ;
// store regressor
*p_ar << gs::Record(dynamic_cast<const BaseRegression &>(*p_condExp), "regressor", stepString.c_str()) ;
vector<ArrayXXd> regressedValues(p_phiIn.size());
if (m_world.rank() < m_paral->getNbProcessorUsed())
{
// regresse the values
for (size_t iReg = 0; iReg < p_phiIn.size(); ++iReg)
{
ArrayXXd transposeCont = p_phiIn[iReg]->transpose();
regressedValues[iReg] = p_condExp->getCoordBasisFunctionMultiple(transposeCont).transpose();
}
}
*p_ar << gs::Record(regressedValues, (p_name + "Values").c_str(), stepString.c_str()) ;
}
else
{
// utilitary
Array< array<int, 2 >, Dynamic, 1 > gridOnProc0(initialDimension.size());
for (int id = 0; id < initialDimension.size(); ++id)
{
gridOnProc0(id)[0] = 0 ;
gridOnProc0(id)[1] = initialDimension(id) ;
}
ArrayXi splittingRatio = paraOptimalSplitting(initialDimension, m_pOptimize->getDimensionToSplit(), m_world);
ParallelComputeGridSplitting paralObject(initialDimension, splittingRatio, m_world);
vector< GridAndRegressedValue> bellVal(p_phiIn.size());
for (size_t iReg = 0; iReg < p_phiIn.size(); ++iReg)
{
ArrayXXd reconstructedArray ;
if (m_world.rank() < m_paral->getNbProcessorUsed())
reconstructedArray = paralObject.reconstruct(*p_phiIn[iReg], gridOnProc0);
if (m_world.rank() == 0)
bellVal[iReg] = GridAndRegressedValue(m_pGridCurrent, p_condExp, reconstructedArray);
}
if (m_world.rank() == 0)
{
*p_ar << gs::Record(bellVal, (p_name + "Values").c_str(), stepString.c_str()) ;
}
}
if (m_world.rank() == 0)
p_ar->flush() ; // necessary for python mapping
m_world.barrier() ; // onlyt to prevent the reading in simualtion before the end of writting
}
#endif
|
