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
277
278
279
280
281
282
283
284
285
286
287
288
289
290
|
// Copyright (C) 2016 EDF
// All Rights Reserved
// This code is published under the GNU Lesser General Public License (GNU LGPL)
#include <memory>
#ifdef USE_MPI
#include "boost/mpi.hpp"
#include "StOpt/core/parallelism/all_gatherv.hpp"
#endif
#ifdef _OPENMP
#include <omp.h>
#include "StOpt/core/utils/OpenmpException.h"
#endif
#include "geners/Record.hh"
#include "StOpt/core/grids/SparseGridIterator.h"
#include "StOpt/dp/TransitionStepRegressionDPSparse.h"
#include "StOpt/regression/ContinuationValue.h"
#include "StOpt/regression/ContinuationValueGeners.h"
#include "StOpt/regression/GridAndRegressedValue.h"
#include "StOpt/regression/GridAndRegressedValueGeners.h"
using namespace Eigen;
using namespace StOpt;
using namespace std;
TransitionStepRegressionDPSparse::TransitionStepRegressionDPSparse(const shared_ptr<SparseSpaceGrid> &p_pGridCurrent,
const shared_ptr<SparseSpaceGrid> &p_pGridPrevious,
const shared_ptr<OptimizerDPBase > &p_pOptimize
#ifdef USE_MPI
, const boost::mpi::communicator &p_world
#endif
):
m_pGridCurrent(p_pGridCurrent), m_pGridPrevious(p_pGridPrevious), m_pOptimize(p_pOptimize)
#ifdef USE_MPI
, m_world(p_world)
#endif
{
}
std::pair< vector< shared_ptr< ArrayXXd > >, vector< shared_ptr< ArrayXXd > > > TransitionStepRegressionDPSparse::oneStep(const vector< shared_ptr< ArrayXXd > > &p_phiIn,
const shared_ptr< BaseRegression> &p_condExp) const
{
int rank = 0 ;
int nbProc = 1;
int nbSimul = p_condExp->getNbSimul();
#ifdef USE_MPI
rank = m_world.rank();
nbProc = m_world.size();
#endif
// number of regimes at current time
int nbRegimes = m_pOptimize->getNbRegime();
vector< shared_ptr< ArrayXXd > > phiOut(nbRegimes);
// to store hierarchical values of cash
vector< shared_ptr< ArrayXXd > > cashHierar(p_phiIn.size());
// parallelism
int nsimPProc = (int)(nbSimul / nbProc);
int nRestSim = nbSimul % nbProc;
int iFirstSim = rank * nsimPProc + (rank < nRestSim ? rank : nRestSim);
int iLastSim = iFirstSim + nsimPProc + (rank < nRestSim ? 1 : 0);
for (int iReg = 0; iReg < nbRegimes; ++iReg)
{
cashHierar[iReg] = make_shared<ArrayXXd>(p_phiIn[iReg]->rows(), p_phiIn[iReg]->cols());
ArrayXXd valHierar(p_phiIn[iReg]->cols(), iLastSim - iFirstSim);
ArrayXd valHierarCol(p_phiIn[iReg]->cols());
for (int is = iFirstSim; is < iLastSim; ++is)
{
// Hierarchize
valHierarCol = p_phiIn[iReg]->row(is).transpose();
m_pGridPrevious->toHierarchize(valHierarCol);
valHierar.col(is - iFirstSim) = valHierarCol;
}
#ifdef USE_MPI
if (m_world.size() > 1)
{
ArrayXXd valHierarShared(p_phiIn[iReg]->cols(), nbSimul);
boost::mpi::all_gatherv<double>(m_world, valHierar.data(), valHierar.size(), valHierarShared.data());
*cashHierar[iReg] = valHierarShared.transpose();
}
else
#endif
*cashHierar[iReg] = valHierar.transpose();
}
int nbControl = m_pOptimize->getNbControl();
vector< shared_ptr< ArrayXXd > > controlOut(nbControl);
// only if the processor is working
if (m_pGridCurrent->getNbPoints() > 0)
{
// number of thread
#ifdef _OPENMP
int nbThreads = omp_get_max_threads();
#else
int nbThreads = 1;
#endif
// create continuation values on Hierarchical values
vector< ContinuationValue > contVal(p_phiIn.size());
int nbPointsPrev = m_pGridPrevious->getNbPoints();
int npointPProcPrev = (int)(nbPointsPrev / nbProc);
int nRestPointPrev = nbPointsPrev % nbProc;
int iFirstPointPrev = rank * npointPProcPrev + (rank < nRestPointPrev ? rank : nRestPointPrev);
int iLastPointPrev = iFirstPointPrev + npointPProcPrev + (rank < nRestPointPrev ? 1 : 0);
for (size_t iReg = 0; iReg < p_phiIn.size(); ++iReg)
{
ArrayXXd regressed = p_condExp->getCoordBasisFunctionMultiple(p_phiIn[iReg]->block(0, iFirstPointPrev, p_phiIn[iReg]->rows(), iLastPointPrev - iFirstPointPrev).transpose());
#ifdef USE_MPI
if (m_world.size() > 1)
{
ArrayXXd regressedShared(regressed.cols(), nbPointsPrev);
ArrayXXd regrTranspose = regressed.transpose();
boost::mpi::all_gatherv<double>(m_world, regrTranspose.data(), regrTranspose.size(), regressedShared.data());
regressed = regressedShared.transpose();
}
#endif
// don't parallelize because hierarchize the regression coefficients and not the simulations
ArrayXd vHierarLoc(regressed.cols());
for (int iFonc = 0 ; iFonc < regressed.cols(); ++iFonc)
{
vHierarLoc = regressed.col(iFonc);
m_pGridPrevious->toHierarchize(vHierarLoc);
regressed.col(iFonc) = vHierarLoc;
}
contVal[iReg].loadForSimulation(m_pGridPrevious, p_condExp, regressed.transpose());
}
// allocate for solution
#ifdef USE_MPI
int nbPointsCur = m_pGridCurrent->getNbPoints();
int npointPProcCur = (int)(nbPointsCur / nbProc);
int nRestPointCur = nbPointsCur % nbProc;
int iFirstPointCur = rank * npointPProcCur + (rank < nRestPointCur ? rank : nRestPointCur);
int iLastPointCur = iFirstPointCur + npointPProcCur + (rank < nRestPointCur ? 1 : 0);
std::vector< ArrayXXd> phiOutLoc(nbRegimes), controlOutLoc(nbControl);
ArrayXi ilocToGLobal(iLastPointCur - iFirstPointCur);
for (int iReg = 0; iReg < nbRegimes; ++iReg)
phiOutLoc[iReg].resize(p_condExp->getNbSimul(), iLastPointCur - iFirstPointCur);
for (int iCont = 0; iCont < nbControl; ++iCont)
controlOutLoc[iCont].resize(p_condExp->getNbSimul(), iLastPointCur - iFirstPointCur);
#endif
for (int iReg = 0; iReg < nbRegimes; ++iReg)
phiOut[iReg] = make_shared< ArrayXXd >(p_condExp->getNbSimul(), m_pGridCurrent->getNbPoints());
for (int iCont = 0; iCont < nbControl; ++iCont)
controlOut[iCont] = make_shared< ArrayXXd >(p_condExp->getNbSimul(), m_pGridCurrent->getNbPoints());
// 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_pGridCurrent->getGridIterator();
// account for mpi and threads
#ifdef USE_MPI
iterGridPoint->jumpToAndInc(rank, nbProc, iThread);
#else
iterGridPoint->jumpToAndInc(0, 1, iThread);
#endif
// iterates on points of the grid
while (iterGridPoint->isValid())
{
ArrayXd pointCoord = iterGridPoint->getCoordinate();
// optimize the current point and the set of regimes
std::pair< ArrayXXd, ArrayXXd> solutionAndControl = m_pOptimize->stepOptimize(m_pGridPrevious, pointCoord, contVal, cashHierar);
// copie solution
#ifdef USE_MPI
int iposArray = iterGridPoint->getRelativePosition();
ilocToGLobal(iposArray) = iterGridPoint->getCount();
for (int iReg = 0; iReg < nbRegimes; ++iReg)
phiOutLoc[iReg].col(iposArray) = solutionAndControl.first.col(iReg);
for (int iCont = 0; iCont < nbControl; ++iCont)
controlOutLoc[iCont].col(iposArray) = solutionAndControl.second.col(iCont);
#else
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);
#endif
iterGridPoint->nextInc(nbThreads);
}
#ifdef _OPENMP
});
#endif
}
#ifdef _OPENMP
excep.rethrow();
#endif
#ifdef USE_MPI
ArrayXi ilocToGLobalGlob(nbPointsCur);
boost::mpi::all_gatherv<int>(m_world, ilocToGLobal.data(), ilocToGLobal.size(), ilocToGLobalGlob.data());
ArrayXXd storeGlob(p_condExp->getNbSimul(), nbPointsCur);
for (int iReg = 0; iReg < nbRegimes; ++iReg)
{
boost::mpi::all_gatherv<double>(m_world, phiOutLoc[iReg].data(), phiOutLoc[iReg].size(), storeGlob.data());
for (int ipos = 0; ipos < ilocToGLobalGlob.size(); ++ipos)
(*phiOut[iReg]).col(ilocToGLobalGlob(ipos)) = storeGlob.col(ipos);
}
for (int iCont = 0 ; iCont < nbControl; ++iCont)
{
boost::mpi::all_gatherv<double>(m_world, controlOutLoc[iCont].data(), controlOutLoc[iCont].size(), storeGlob.data());
for (int ipos = 0; ipos < ilocToGLobalGlob.size(); ++ipos)
(*controlOut[iCont]).col(ilocToGLobalGlob(ipos)) = storeGlob.col(ipos);
}
#endif
}
return make_pair(phiOut, controlOut);
}
void TransitionStepRegressionDPSparse::dumpContinuationValues(shared_ptr<gs::BinaryFileArchive> p_ar, const string &p_name, const int &p_iStep, const vector< shared_ptr< ArrayXXd > > &p_phiIn,
const vector< shared_ptr< ArrayXXd > > &p_control, const shared_ptr<BaseRegression> &p_condExp) const
{
int nbProc = 1;
int rank = 0 ;
#ifdef USE_MPI
rank = m_world.rank();
nbProc = m_world.size();
#endif
// hierarchize all values
vector< GridAndRegressedValue > contVal(p_phiIn.size());
int nbPoints = m_pGridPrevious->getNbPoints();
int npointPProc = (int)(nbPoints / nbProc);
int nRestPoint = nbPoints % nbProc;
int iFirstPoint = rank * npointPProc + (rank < nRestPoint ? rank : nRestPoint);
int iLastPoint = iFirstPoint + npointPProc + (rank < nRestPoint ? 1 : 0);
for (size_t iReg = 0; iReg < p_phiIn.size(); ++iReg)
{
// first regress the cash values on all stock points
ArrayXXd regressed = p_condExp->getCoordBasisFunctionMultiple(p_phiIn[iReg]->block(0, iFirstPoint, p_phiIn[iReg]->rows(), iLastPoint - iFirstPoint).transpose());
#ifdef USE_MPI
if (m_world.size() > 1)
{
ArrayXXd regressedShared(regressed.cols(), nbPoints);
ArrayXXd regrTranspose = regressed.transpose();
boost::mpi::all_gatherv<double>(m_world, regrTranspose.data(), regrTranspose.size(), regressedShared.data());
regressed = regressedShared.transpose();
}
#endif
contVal[iReg] = GridAndRegressedValue(m_pGridPrevious, p_condExp);
contVal[iReg].setRegressedValues(regressed.transpose());
}
string stepString = boost::lexical_cast<string>(p_iStep) ;
#ifdef USE_MPI
if (rank == 0)
#endif
*p_ar << gs::Record(contVal, (p_name + "Values").c_str(), stepString.c_str()) ;
// Hierarchize control
nbPoints = m_pGridCurrent->getNbPoints();
npointPProc = (int)(nbPoints / nbProc);
nRestPoint = nbPoints % nbProc;
iFirstPoint = rank * npointPProc + (rank < nRestPoint ? rank : nRestPoint);
iLastPoint = iFirstPoint + npointPProc + (rank < nRestPoint ? 1 : 0);
// number of controls
int nbControl = p_control.size();
vector< GridAndRegressedValue > controlVal(nbControl);
for (int iCont = 0; iCont < nbControl; ++iCont)
{
// first regress the cash values on all stock points
ArrayXXd contRegressed = p_condExp->getCoordBasisFunctionMultiple(p_control[iCont]->block(0, iFirstPoint, p_control[iCont]->rows(), iLastPoint - iFirstPoint).transpose());
#ifdef USE_MPI
if (m_world.size() > 1)
{
ArrayXXd regressedShared(contRegressed.cols(), nbPoints);
ArrayXXd regrTranspose = contRegressed.transpose();
boost::mpi::all_gatherv<double>(m_world, regrTranspose.data(), regrTranspose.size(), regressedShared.data());
contRegressed = regressedShared.transpose();
}
#endif
controlVal[iCont] = GridAndRegressedValue(m_pGridCurrent, p_condExp);
controlVal[iCont].setRegressedValues(contRegressed.transpose());
}
#ifdef USE_MPI
if (rank == 0)
{
#endif
*p_ar << gs::Record(controlVal, (p_name + "Control").c_str(), stepString.c_str()) ;
p_ar->flush() ; // necessary for python mapping
#ifdef USE_MPI
}
#endif
}
|
