MueLu Version of the Day
Loading...
Searching...
No Matches
MueLu_GeneralGeometricPFactory_def.hpp
Go to the documentation of this file.
1// @HEADER
2//
3// ***********************************************************************
4//
5// MueLu: A package for multigrid based preconditioning
6// Copyright 2012 Sandia Corporation
7//
8// Under the terms of Contract DE-AC04-94AL85000 with Sandia Corporation,
9// the U.S. Government retains certain rights in this software.
10//
11// Redistribution and use in source and binary forms, with or without
12// modification, are permitted provided that the following conditions are
13// met:
14//
15// 1. Redistributions of source code must retain the above copyright
16// notice, this list of conditions and the following disclaimer.
17//
18// 2. Redistributions in binary form must reproduce the above copyright
19// notice, this list of conditions and the following disclaimer in the
20// documentation and/or other materials provided with the distribution.
21//
22// 3. Neither the name of the Corporation nor the names of the
23// contributors may be used to endorse or promote products derived from
24// this software without specific prior written permission.
25//
26// THIS SOFTWARE IS PROVIDED BY SANDIA CORPORATION "AS IS" AND ANY
27// EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
28// IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR
29// PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL SANDIA CORPORATION OR THE
30// CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL,
31// EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO,
32// PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR
33// PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF
34// LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING
35// NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS
36// SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
37//
38// Questions? Contact
39// Jonathan Hu (jhu@sandia.gov)
40// Andrey Prokopenko (aprokop@sandia.gov)
41// Ray Tuminaro (rstumin@sandia.gov)
42//
43// ***********************************************************************
44//
45// @HEADER
46#ifndef MUELU_GENERALGEOMETRICPFACTORY_DEF_HPP
47#define MUELU_GENERALGEOMETRICPFACTORY_DEF_HPP
48
49#include <stdlib.h>
50#include <iomanip>
51
52
53// #include <Teuchos_LAPACK.hpp>
54#include <Teuchos_SerialDenseMatrix.hpp>
55#include <Teuchos_SerialDenseVector.hpp>
56#include <Teuchos_SerialDenseSolver.hpp>
57
58#include <Xpetra_CrsMatrixWrap.hpp>
59#include <Xpetra_ImportFactory.hpp>
60#include <Xpetra_Matrix.hpp>
61#include <Xpetra_MapFactory.hpp>
62#include <Xpetra_MultiVectorFactory.hpp>
63#include <Xpetra_VectorFactory.hpp>
64
65#include <Xpetra_IO.hpp>
66
67#include "MueLu_GeneralGeometricPFactory_decl.hpp"
68
69#include "MueLu_Monitor.hpp"
70
71namespace MueLu {
72
73 template <class Scalar, class LocalOrdinal, class GlobalOrdinal, class Node>
74 RCP<const ParameterList> GeneralGeometricPFactory<Scalar, LocalOrdinal, GlobalOrdinal, Node>::GetValidParameterList() const {
75 RCP<ParameterList> validParamList = rcp(new ParameterList());
76
77 // Coarsen can come in two forms, either a single char that will be interpreted as an integer
78 // which is used as the coarsening rate in every spatial dimentions, or it can be a longer
79 // string that will then be interpreted as an array of integers.
80 // By default coarsen is set as "{2}", hence a coarsening rate of 2 in every spatial dimension
81 // is the default setting!
82 validParamList->set<std::string > ("Coarsen", "{2}",
83 "Coarsening rate in all spatial dimensions");
84 validParamList->set<int> ("order", 1,
85 "Order of the interpolation scheme used");
86 validParamList->set<RCP<const FactoryBase> >("A", Teuchos::null,
87 "Generating factory of the matrix A");
88 validParamList->set<RCP<const FactoryBase> >("Nullspace", Teuchos::null,
89 "Generating factory of the nullspace");
90 validParamList->set<RCP<const FactoryBase> >("Coordinates", Teuchos::null,
91 "Generating factory for coorindates");
92 validParamList->set<RCP<const FactoryBase> >("gNodesPerDim", Teuchos::null,
93 "Number of nodes per spatial dimmension provided by CoordinatesTransferFactory.");
94 validParamList->set<RCP<const FactoryBase> >("lNodesPerDim", Teuchos::null,
95 "Number of nodes per spatial dimmension provided by CoordinatesTransferFactory.");
96 validParamList->set<std::string > ("meshLayout", "Global Lexicographic",
97 "Type of mesh ordering");
98 validParamList->set<RCP<const FactoryBase> >("meshData", Teuchos::null,
99 "Mesh ordering associated data");
100
101 return validParamList;
102 }
103
104 template <class Scalar,class LocalOrdinal, class GlobalOrdinal, class Node>
105 void GeneralGeometricPFactory<Scalar, LocalOrdinal, GlobalOrdinal, Node>::
106 DeclareInput(Level& fineLevel, Level& /* coarseLevel */) const {
107 Input(fineLevel, "A");
108 Input(fineLevel, "Nullspace");
109 Input(fineLevel, "Coordinates");
110 // Request the global number of nodes per dimensions
111 if(fineLevel.GetLevelID() == 0) {
112 if(fineLevel.IsAvailable("gNodesPerDim", NoFactory::get())) {
113 fineLevel.DeclareInput("gNodesPerDim", NoFactory::get(), this);
114 } else {
115 TEUCHOS_TEST_FOR_EXCEPTION(fineLevel.IsAvailable("gNodesPerDim", NoFactory::get()),
116 Exceptions::RuntimeError,
117 "gNodesPerDim was not provided by the user on level0!");
118 }
119 } else {
120 Input(fineLevel, "gNodesPerDim");
121 }
122
123 // Request the local number of nodes per dimensions
124 if(fineLevel.GetLevelID() == 0) {
125 if(fineLevel.IsAvailable("lNodesPerDim", NoFactory::get())) {
126 fineLevel.DeclareInput("lNodesPerDim", NoFactory::get(), this);
127 } else {
128 TEUCHOS_TEST_FOR_EXCEPTION(fineLevel.IsAvailable("lNodesPerDim", NoFactory::get()),
129 Exceptions::RuntimeError,
130 "lNodesPerDim was not provided by the user on level0!");
131 }
132 } else {
133 Input(fineLevel, "lNodesPerDim");
134 }
135 }
136
137 template <class Scalar,class LocalOrdinal, class GlobalOrdinal, class Node>
138 void GeneralGeometricPFactory<Scalar, LocalOrdinal, GlobalOrdinal, Node>::Build(Level& fineLevel,
139 Level& coarseLevel) const {
140 return BuildP(fineLevel, coarseLevel);
141 }
142
143 template <class Scalar,class LocalOrdinal, class GlobalOrdinal, class Node>
144 void GeneralGeometricPFactory<Scalar, LocalOrdinal, GlobalOrdinal, Node>::BuildP(Level& fineLevel,
145 Level& coarseLevel) const {
146 FactoryMonitor m(*this, "Build", coarseLevel);
147
148 // Obtain general variables
149 using xdMV = Xpetra::MultiVector<typename Teuchos::ScalarTraits<Scalar>::coordinateType,LO,GO,NO>;
150 RCP<Matrix> A = Get< RCP<Matrix> > (fineLevel, "A");
151 RCP<MultiVector> fineNullspace = Get< RCP<MultiVector> > (fineLevel, "Nullspace");
152 RCP<xdMV> fineCoords = Get< RCP<xdMV> >(fineLevel, "Coordinates");
153 RCP<xdMV> coarseCoords;
154
155 // Get user-provided coarsening rate parameter (constant over all levels)
156 const ParameterList& pL = GetParameterList();
157
158 // collect general input data
159 const LO blkSize = A->GetFixedBlockSize();
160 RCP<const Map> rowMap = A->getRowMap();
161 RCP<GeometricData> myGeometry = rcp(new GeometricData{});
162
163 // Load the mesh layout type and the associated mesh data
164 myGeometry->meshLayout = pL.get<std::string>("meshLayout");
165 if(fineLevel.GetLevelID() == 0) {
166 if(myGeometry->meshLayout == "Local Lexicographic") {
167 Array<GO> meshData;
168 meshData = fineLevel.Get<Array<GO> >("meshData", NoFactory::get());
169 TEUCHOS_TEST_FOR_EXCEPTION(meshData.empty() == true, Exceptions::RuntimeError,
170 "The meshData array is empty, somehow the input for geometric"
171 " multigrid are not captured correctly.");
172 myGeometry->meshData.resize(rowMap->getComm()->getSize());
173 for(int i = 0; i < rowMap->getComm()->getSize(); ++i) {
174 myGeometry->meshData[i].resize(10);
175 for(int j = 0; j < 10; ++j) {
176 myGeometry->meshData[i][j] = meshData[10*i + j];
177 }
178 }
179 }
180 }
181
182 TEUCHOS_TEST_FOR_EXCEPTION(fineCoords == Teuchos::null, Exceptions::RuntimeError,
183 "Coordinates cannot be accessed from fine level!");
184 myGeometry->numDimensions = fineCoords->getNumVectors();
185
186 // Get the number of points in each direction
187 if(fineLevel.GetLevelID() == 0) {
188 myGeometry->gFineNodesPerDir = fineLevel.Get<Array<GO> >("gNodesPerDim", NoFactory::get());
189 myGeometry->lFineNodesPerDir = fineLevel.Get<Array<LO> >("lNodesPerDim", NoFactory::get());
190 } else {
191 // Loading global number of nodes per diretions
192 myGeometry->gFineNodesPerDir = Get<Array<GO> >(fineLevel, "gNodesPerDim");
193
194 // Loading local number of nodes per diretions
195 myGeometry->lFineNodesPerDir = Get<Array<LO> >(fineLevel, "lNodesPerDim");
196 }
197 myGeometry->gNumFineNodes10 = myGeometry->gFineNodesPerDir[1]*myGeometry->gFineNodesPerDir[0];
198 myGeometry->gNumFineNodes = myGeometry->gFineNodesPerDir[2]*myGeometry->gNumFineNodes10;
199 myGeometry->lNumFineNodes10 = myGeometry->lFineNodesPerDir[1]*myGeometry->lFineNodesPerDir[0];
200 myGeometry->lNumFineNodes = myGeometry->lFineNodesPerDir[2]*myGeometry->lNumFineNodes10;
201
202 TEUCHOS_TEST_FOR_EXCEPTION(fineCoords->getLocalLength()
203 != static_cast<size_t>(myGeometry->lNumFineNodes),
204 Exceptions::RuntimeError,
205 "The local number of elements in Coordinates is not equal to the"
206 " number of nodes given by: lNodesPerDim!");
207 TEUCHOS_TEST_FOR_EXCEPTION(fineCoords->getGlobalLength()
208 != static_cast<size_t>(myGeometry->gNumFineNodes),
209 Exceptions::RuntimeError,
210 "The global number of elements in Coordinates is not equal to the"
211 " number of nodes given by: gNodesPerDim!");
212
213 // Get the coarsening rate
214 std::string coarsenRate = pL.get<std::string>("Coarsen");
215 Teuchos::Array<LO> crates;
216 try {
217 crates = Teuchos::fromStringToArray<LO>(coarsenRate);
218 } catch(const Teuchos::InvalidArrayStringRepresentation& e) {
219 GetOStream(Errors,-1) << " *** Coarsen must be a string convertible into an array! *** "
220 << std::endl;
221 throw e;
222 }
223 TEUCHOS_TEST_FOR_EXCEPTION((crates.size() > 1) && (crates.size() < myGeometry->numDimensions),
224 Exceptions::RuntimeError,
225 "Coarsen must have at least as many components as the number of"
226 " spatial dimensions in the problem.");
227
228 for(LO i = 0; i < 3; ++i) {
229 if(i < myGeometry->numDimensions) {
230 if(crates.size()==1) {
231 myGeometry->coarseRate[i] = crates[0];
232 } else if(crates.size() == myGeometry->numDimensions) {
233 myGeometry->coarseRate[i] = crates[i];
234 }
235 } else {
236 myGeometry->coarseRate[i] = 1;
237 }
238 }
239
240 int interpolationOrder = pL.get<int>("order");
241 TEUCHOS_TEST_FOR_EXCEPTION((interpolationOrder < 0) || (interpolationOrder > 1),
242 Exceptions::RuntimeError,
243 "The interpolation order can only be set to 0 or 1.");
244
245 // Get the axis permutation from Global axis to Local axis
246 Array<LO> mapDirG2L(3), mapDirL2G(3);
247 for(LO i = 0; i < myGeometry->numDimensions; ++i) {
248 mapDirG2L[i] = i;
249 }
250 for(LO i = 0; i < myGeometry->numDimensions; ++i) {
251 TEUCHOS_TEST_FOR_EXCEPTION(mapDirG2L[i] > myGeometry->numDimensions,
252 Exceptions::RuntimeError,
253 "axis permutation values must all be less than"
254 " the number of spatial dimensions.");
255 mapDirL2G[mapDirG2L[i]] = i;
256 }
257 RCP<const Map> fineCoordsMap = fineCoords->getMap();
258
259 // This struct stores PIDs, LIDs and GIDs on the fine mesh and GIDs on the coarse mesh.
260 RCP<NodesIDs> ghostedCoarseNodes = rcp(new NodesIDs{});
261 Array<Array<GO> > lCoarseNodesGIDs(2);
262 if((fineLevel.GetLevelID() == 0) && (myGeometry->meshLayout != "Global Lexicographic")) {
263 MeshLayoutInterface(interpolationOrder, blkSize, fineCoordsMap, myGeometry,
264 ghostedCoarseNodes, lCoarseNodesGIDs);
265 } else {
266 // This function expects perfect global lexicographic ordering of nodes and will not process
267 // data correctly otherwise. These restrictions allow for the simplest and most efficient
268 // processing of the levels (hopefully at least).
269 GetCoarsePoints(interpolationOrder, blkSize, fineCoordsMap, myGeometry, ghostedCoarseNodes,
270 lCoarseNodesGIDs);
271 }
272
273 // All that is left to do is loop over NCpts and:
274 // - extract coarse points coordiate for coarseCoords
275 // - get coordinates for current stencil computation
276 // - compute current stencil
277 // - compute row and column indices for stencil entries
278 RCP<const Map> stridedDomainMapP;
279 RCP<Matrix> P;
280 // Fancy formula for the number of non-zero terms
281 // All coarse points are injected, other points are using polynomial interpolation
282 // and have contribution from (interpolationOrder + 1)^numDimensions
283 // Noticebly this leads to 1 when the order is zero, hence fancy MatMatMatMult can be used.
284 GO lnnzP = Teuchos::as<LO>(std::pow(interpolationOrder + 1, myGeometry->numDimensions))
285 *(myGeometry->lNumFineNodes - myGeometry->lNumCoarseNodes) + myGeometry->lNumCoarseNodes;
286 lnnzP = lnnzP*blkSize;
287 GO gnnzP = Teuchos::as<LO>(std::pow(interpolationOrder + 1, myGeometry->numDimensions))
288 *(myGeometry->gNumFineNodes - myGeometry->gNumCoarseNodes) + myGeometry->gNumCoarseNodes;
289 gnnzP = gnnzP*blkSize;
290
291 GetOStream(Runtime1) << "P size = " << blkSize*myGeometry->gNumFineNodes
292 << " x " << blkSize*myGeometry->gNumCoarseNodes << std::endl;
293 GetOStream(Runtime1) << "P Fine grid : " << myGeometry->gFineNodesPerDir[0] << " -- "
294 << myGeometry->gFineNodesPerDir[1] << " -- "
295 << myGeometry->gFineNodesPerDir[2] << std::endl;
296 GetOStream(Runtime1) << "P Coarse grid : " << myGeometry->gCoarseNodesPerDir[0] << " -- "
297 << myGeometry->gCoarseNodesPerDir[1] << " -- "
298 << myGeometry->gCoarseNodesPerDir[2] << std::endl;
299 GetOStream(Runtime1) << "P nnz estimate: " << gnnzP << std::endl;
300
301 MakeGeneralGeometricP(myGeometry, fineCoords, lnnzP, blkSize, stridedDomainMapP,
302 A, P, coarseCoords, ghostedCoarseNodes, lCoarseNodesGIDs,
303 interpolationOrder);
304
305 // set StridingInformation of P
306 if (A->IsView("stridedMaps") == true) {
307 P->CreateView("stridedMaps", A->getRowMap("stridedMaps"), stridedDomainMapP);
308 } else {
309 P->CreateView("stridedMaps", P->getRangeMap(), stridedDomainMapP);
310 }
311
312 // store the transfer operator and node coordinates on coarse level
313 Set(coarseLevel, "P", P);
314 Set(coarseLevel, "coarseCoordinates", coarseCoords);
315 Set<Array<GO> >(coarseLevel, "gCoarseNodesPerDim", myGeometry->gCoarseNodesPerDir);
316 Set<Array<LO> >(coarseLevel, "lCoarseNodesPerDim", myGeometry->lCoarseNodesPerDir);
317
318 // rst: null space might get scaled here ... do we care. We could just inject at the cpoints,
319 // but I don't feel that this is needed.
320 RCP<MultiVector> coarseNullspace = MultiVectorFactory::Build(P->getDomainMap(),
321 fineNullspace->getNumVectors());
322 P->apply(*fineNullspace, *coarseNullspace, Teuchos::TRANS, Teuchos::ScalarTraits<SC>::one(),
323 Teuchos::ScalarTraits<SC>::zero());
324 Set(coarseLevel, "Nullspace", coarseNullspace);
325
326 }
327
328 template <class Scalar, class LocalOrdinal, class GlobalOrdinal, class Node>
329 void GeneralGeometricPFactory<Scalar, LocalOrdinal, GlobalOrdinal, Node>::
330 MeshLayoutInterface(const int /* interpolationOrder */, const LO /* blkSize */, RCP<const Map> fineCoordsMap,
331 RCP<GeometricData> myGeo, RCP<NodesIDs> ghostedCoarseNodes,
332 Array<Array<GO> >& lCoarseNodesGIDs) const{
333 // The goal here is to produce maps that globally labels the mesh lexicographically.
334 // These maps will replace the current maps of A, the coordinate vector and the nullspace.
335 // Ideally if the user provides the necessary maps then nothing needs to be done, otherwise
336 // it could be advantageous to allow the user to register a re-labeling function. Ultimately
337 // for common ordering schemes, some re-labeling can be directly implemented here.
338
339 int numRanks = fineCoordsMap->getComm()->getSize();
340 int myRank = fineCoordsMap->getComm()->getRank();
341
342 myGeo->myBlock = myGeo->meshData[myRank][2];
343 myGeo->startIndices[0] = myGeo->meshData[myRank][3];
344 myGeo->startIndices[1] = myGeo->meshData[myRank][5];
345 myGeo->startIndices[2] = myGeo->meshData[myRank][7];
346 myGeo->startIndices[3] = myGeo->meshData[myRank][4];
347 myGeo->startIndices[4] = myGeo->meshData[myRank][6];
348 myGeo->startIndices[5] = myGeo->meshData[myRank][8];
349 std::sort(myGeo->meshData.begin(), myGeo->meshData.end(),
350 [](const std::vector<GO>& a, const std::vector<GO>& b)->bool {
351 // The below function sorts ranks by blockID, kmin, jmin and imin
352 if(a[2] < b[2]) {
353 return true;
354 } else if(a[2] == b[2]) {
355 if(a[7] < b[7]) {
356 return true;
357 } else if(a[7] == b[7]) {
358 if(a[5] < b[5]) {
359 return true;
360 } else if(a[5] == b[5]) {
361 if(a[3] < b[3]) {return true;}
362 }
363 }
364 }
365 return false;
366 });
367
368 myGeo->numBlocks = myGeo->meshData[numRanks - 1][2] + 1;
369 // Find the range of the current block
370 auto myBlockStart = std::lower_bound(myGeo->meshData.begin(), myGeo->meshData.end(),
371 myGeo->myBlock - 1,
372 [](const std::vector<GO>& vec, const GO val)->bool{
373 return (vec[2] < val) ? true : false;
374 });
375 auto myBlockEnd = std::upper_bound(myGeo->meshData.begin(), myGeo->meshData.end(),
376 myGeo->myBlock,
377 [](const GO val, const std::vector<GO>& vec)->bool{
378 return (val < vec[2]) ? true : false;
379 });
380 // Assuming that i,j,k and ranges are split in pi, pj and pk processors
381 // we search for these numbers as they will allow us to find quickly the PID of processors
382 // owning ghost nodes.
383 auto myKEnd = std::upper_bound(myBlockStart, myBlockEnd, (*myBlockStart)[3],
384 [](const GO val, const std::vector<GO>& vec)->bool{
385 return (val < vec[7]) ? true : false;
386 });
387 auto myJEnd = std::upper_bound(myBlockStart, myKEnd, (*myBlockStart)[3],
388 [](const GO val, const std::vector<GO>& vec)->bool{
389 return (val < vec[5]) ? true : false;
390 });
391 LO pi = std::distance(myBlockStart, myJEnd);
392 LO pj = std::distance(myBlockStart, myKEnd) / pi;
393 LO pk = std::distance(myBlockStart, myBlockEnd) / (pj*pi);
394
395 // We also look for the index of the local rank in the current block.
396 LO myRankIndex = std::distance(myGeo->meshData.begin(),
397 std::find_if(myBlockStart, myBlockEnd,
398 [myRank](const std::vector<GO>& vec)->bool{
399 return (vec[0] == myRank) ? true : false;
400 })
401 );
402
403 for(int dim = 0; dim < 3; ++dim) {
404 if(dim < myGeo->numDimensions) {
405 myGeo->offsets[dim]= Teuchos::as<LO>(myGeo->startIndices[dim]) % myGeo->coarseRate[dim];
406 myGeo->offsets[dim + 3]= Teuchos::as<LO>(myGeo->startIndices[dim]) % myGeo->coarseRate[dim];
407 }
408 }
409
410 // Check if the partition contains nodes on a boundary, if so that boundary (face, line or
411 // point) will not require ghost nodes.
412 for(int dim = 0; dim < 3; ++dim) {
413 if(dim < myGeo->numDimensions && (myGeo->startIndices[dim] % myGeo->coarseRate[dim] != 0 ||
414 myGeo->startIndices[dim] == myGeo->gFineNodesPerDir[dim]-1)) {
415 myGeo->ghostInterface[2*dim] = true;
416 }
417 if(dim < myGeo->numDimensions
418 && myGeo->startIndices[dim + 3] != myGeo->gFineNodesPerDir[dim] - 1
419 && (myGeo->lFineNodesPerDir[dim] == 1 ||
420 myGeo->startIndices[dim + 3] % myGeo->coarseRate[dim] != 0)) {
421 myGeo->ghostInterface[2*dim+1] = true;
422 }
423 }
424
425 // Here one element can represent either the degenerate case of one node or the more general
426 // case of two nodes, i.e. x---x is a 1D element with two nodes and x is a 1D element with one
427 // node. This helps generating a 3D space from tensorial products...
428 // A good way to handle this would be to generalize the algorithm to take into account the
429 // discretization order used in each direction, at least in the FEM sense, since a 0 degree
430 // discretization will have a unique node per element. This way 1D discretization can be viewed
431 // as a 3D problem with one 0 degree element in the y direction and one 0 degre element in the z
432 // direction.
433 // !!! Operations below are aftecting both local and global values that have two different !!!
434 // orientations. Orientations can be interchanged using mapDirG2L and mapDirL2G. coarseRate,
435 // endRate and offsets are in the global basis, as well as all the variables starting with a g,
436 // !!! while the variables starting with an l are in the local basis. !!!
437 for(int i = 0; i < 3; ++i) {
438 if(i < myGeo->numDimensions) {
439 // This array is passed to the RAPFactory and eventually becomes gFineNodePerDir on the next
440 // level.
441 myGeo->gCoarseNodesPerDir[i] = (myGeo->gFineNodesPerDir[i] - 1) / myGeo->coarseRate[i];
442 myGeo->endRate[i] = Teuchos::as<LO>((myGeo->gFineNodesPerDir[i] - 1) %myGeo->coarseRate[i]);
443 if(myGeo->endRate[i] == 0) {
444 myGeo->endRate[i] = myGeo->coarseRate[i];
445 ++myGeo->gCoarseNodesPerDir[i];
446 } else {
447 myGeo->gCoarseNodesPerDir[i] += 2;
448 }
449 } else {
450 myGeo->endRate[i] = 1;
451 myGeo->gCoarseNodesPerDir[i] = 1;
452 }
453 }
454
455 myGeo->gNumCoarseNodes = myGeo->gCoarseNodesPerDir[0]*myGeo->gCoarseNodesPerDir[1]
456 *myGeo->gCoarseNodesPerDir[2];
457
458 for(LO i = 0; i < 3; ++i) {
459 if(i < myGeo->numDimensions) {
460 // Check whether the partition includes the "end" of the mesh which means that endRate will
461 // apply. Also make sure that endRate is not 0 which means that the mesh does not require a
462 // particular treatment at the boundaries.
463 if( (myGeo->startIndices[i] + myGeo->lFineNodesPerDir[i]) == myGeo->gFineNodesPerDir[i] ) {
464 myGeo->lCoarseNodesPerDir[i] = (myGeo->lFineNodesPerDir[i] - myGeo->endRate[i]
465 + myGeo->offsets[i] - 1) / myGeo->coarseRate[i] + 1;
466 if(myGeo->offsets[i] == 0) {++myGeo->lCoarseNodesPerDir[i];}
467 } else {
468 myGeo->lCoarseNodesPerDir[i] = (myGeo->lFineNodesPerDir[i] + myGeo->offsets[i] - 1)
469 / myGeo->coarseRate[i];
470 if(myGeo->offsets[i] == 0) {++myGeo->lCoarseNodesPerDir[i];}
471 }
472 } else {
473 myGeo->lCoarseNodesPerDir[i] = 1;
474 }
475 // This would happen if the rank does not own any nodes but in that case a subcommunicator
476 // should be used so this should really not be a concern.
477 if(myGeo->lFineNodesPerDir[i] < 1) {myGeo->lCoarseNodesPerDir[i] = 0;}
478 }
479
480 // Assuming linear interpolation, each fine point has contribution from 8 coarse points
481 // and each coarse point value gets injected.
482 // For systems of PDEs we assume that all dofs have the same P operator.
483 myGeo->lNumCoarseNodes = myGeo->lCoarseNodesPerDir[0]*myGeo->lCoarseNodesPerDir[1]
484 *myGeo->lCoarseNodesPerDir[2];
485
486 // For each direction, determine how many points (including ghosts) are required.
487 for(int dim = 0; dim < 3; ++dim) {
488 // The first branch of this if-statement will be used if the rank contains only one layer
489 // of nodes in direction i, that layer must also coincide with the boundary of the mesh
490 // and coarseRate[i] == endRate[i]...
491 if(myGeo->startIndices[dim] == myGeo->gFineNodesPerDir[dim] - 1 &&
492 myGeo->startIndices[dim] % myGeo->coarseRate[dim] == 0) {
493 myGeo->startGhostedCoarseNode[dim] = myGeo->startIndices[dim] / myGeo->coarseRate[dim] - 1;
494 } else {
495 myGeo->startGhostedCoarseNode[dim] = myGeo->startIndices[dim] / myGeo->coarseRate[dim];
496 }
497 myGeo->ghostedCoarseNodesPerDir[dim] = myGeo->lCoarseNodesPerDir[dim];
498 // Check whether face *low needs ghost nodes
499 if(myGeo->ghostInterface[2*dim]) {myGeo->ghostedCoarseNodesPerDir[dim] += 1;}
500 // Check whether face *hi needs ghost nodes
501 if(myGeo->ghostInterface[2*dim + 1]) {myGeo->ghostedCoarseNodesPerDir[dim] += 1;}
502 }
503 myGeo->lNumGhostedNodes = myGeo->ghostedCoarseNodesPerDir[2]*myGeo->ghostedCoarseNodesPerDir[1]
504 *myGeo->ghostedCoarseNodesPerDir[0];
505 myGeo->lNumGhostNodes = myGeo->lNumGhostedNodes - myGeo->lNumCoarseNodes;
506 ghostedCoarseNodes->PIDs.resize(myGeo->lNumGhostedNodes);
507 ghostedCoarseNodes->LIDs.resize(myGeo->lNumGhostedNodes);
508 ghostedCoarseNodes->GIDs.resize(myGeo->lNumGhostedNodes);
509 ghostedCoarseNodes->coarseGIDs.resize(myGeo->lNumGhostedNodes);
510 ghostedCoarseNodes->colInds.resize(myGeo->lNumGhostedNodes);
511 lCoarseNodesGIDs[0].resize(myGeo->lNumCoarseNodes);
512 lCoarseNodesGIDs[1].resize(myGeo->lNumCoarseNodes);
513
514 // Now the tricky part starts, the coarse nodes / ghosted coarse nodes need to be imported.
515 // This requires finding what their GID on the fine mesh is. They need to be ordered
516 // lexicographically to allow for fast sweeps through the mesh.
517
518 // We loop over all ghosted coarse nodes by increasing global lexicographic order
519 Array<LO> coarseNodeCoarseIndices(3), coarseNodeFineIndices(3);
520 LO currentIndex = -1, countCoarseNodes = 0;
521 for(int k = 0; k < myGeo->ghostedCoarseNodesPerDir[2]; ++k) {
522 for(int j = 0; j < myGeo->ghostedCoarseNodesPerDir[1]; ++j) {
523 for(int i = 0; i < myGeo->ghostedCoarseNodesPerDir[0]; ++i) {
524 currentIndex = k*myGeo->ghostedCoarseNodesPerDir[1]*myGeo->ghostedCoarseNodesPerDir[0]
525 + j*myGeo->ghostedCoarseNodesPerDir[0] + i;
526 coarseNodeCoarseIndices[0] = myGeo->startGhostedCoarseNode[0] + i;
527 coarseNodeFineIndices[0] = coarseNodeCoarseIndices[0]*myGeo->coarseRate[0];
528 if(coarseNodeFineIndices[0] > myGeo->gFineNodesPerDir[0] - 1) {
529 coarseNodeFineIndices[0] = myGeo->gFineNodesPerDir[0] - 1;
530 }
531 coarseNodeCoarseIndices[1] = myGeo->startGhostedCoarseNode[1] + j;
532 coarseNodeFineIndices[1] = coarseNodeCoarseIndices[1]*myGeo->coarseRate[1];
533 if(coarseNodeFineIndices[1] > myGeo->gFineNodesPerDir[1] - 1) {
534 coarseNodeFineIndices[1] = myGeo->gFineNodesPerDir[1] - 1;
535 }
536 coarseNodeCoarseIndices[2] = myGeo->startGhostedCoarseNode[2] + k;
537 coarseNodeFineIndices[2] = coarseNodeCoarseIndices[2]*myGeo->coarseRate[2];
538 if(coarseNodeFineIndices[2] > myGeo->gFineNodesPerDir[2] - 1) {
539 coarseNodeFineIndices[2] = myGeo->gFineNodesPerDir[2] - 1;
540 }
541 GO myGID = -1, myCoarseGID = -1;
542 LO myLID = -1, myPID = -1;
543 GetGIDLocalLexicographic(i, j, k, coarseNodeFineIndices, myGeo, myRankIndex, pi, pj, pk,
544 myBlockStart, myBlockEnd, myGID, myPID, myLID);
545 myCoarseGID = coarseNodeCoarseIndices[0]
546 + coarseNodeCoarseIndices[1]*myGeo->gCoarseNodesPerDir[0]
547 + coarseNodeCoarseIndices[2]*myGeo->gCoarseNodesPerDir[1]*myGeo->gCoarseNodesPerDir[0];
548 ghostedCoarseNodes->PIDs[currentIndex] = myPID;
549 ghostedCoarseNodes->LIDs[currentIndex] = myLID;
550 ghostedCoarseNodes->GIDs[currentIndex] = myGID;
551 ghostedCoarseNodes->coarseGIDs[currentIndex] = myCoarseGID;
552 if(myPID == myRank){
553 lCoarseNodesGIDs[0][countCoarseNodes] = myCoarseGID;
554 lCoarseNodesGIDs[1][countCoarseNodes] = myGID;
555 ++countCoarseNodes;
556 }
557 }
558 }
559 }
560
561 } // End MeshLayoutInterface
562
563 template <class Scalar, class LocalOrdinal, class GlobalOrdinal, class Node>
564 void GeneralGeometricPFactory<Scalar, LocalOrdinal, GlobalOrdinal, Node>::
565 GetCoarsePoints(const int /* interpolationOrder */, const LO /* blkSize */, RCP<const Map> fineCoordsMap,
566 RCP<GeometricData> myGeo, RCP<NodesIDs> ghostedCoarseNodes,
567 Array<Array<GO> >& lCoarseNodesGIDs) const{
568 // Assuming perfect global lexicographic ordering of the mesh, produce two arrays:
569 // 1) lGhostNodesIDs that stores PID, LID, GID and coarseGID associated with the coarse nodes
570 // need to compute the local part of the prolongator.
571 // 2) lCoarseNodesGIDs that stores the GIDs associated with the local nodes needed to create
572 // the map of the MultiVector of coarse node coordinates.
573
574 {
575 GO tmp = 0;
576 myGeo->startIndices[2] = fineCoordsMap->getMinGlobalIndex()
577 / (myGeo->gFineNodesPerDir[1]*myGeo->gFineNodesPerDir[0]);
578 tmp = fineCoordsMap->getMinGlobalIndex()
579 % (myGeo->gFineNodesPerDir[1]*myGeo->gFineNodesPerDir[0]);
580 myGeo->startIndices[1] = tmp / myGeo->gFineNodesPerDir[0];
581 myGeo->startIndices[0] = tmp % myGeo->gFineNodesPerDir[0];
582 } // End of scope for tmp
583 for(int dim = 0; dim < 3; ++dim) {
584 myGeo->startIndices[dim + 3] = myGeo->startIndices[dim] + myGeo->lFineNodesPerDir[dim] - 1;
585 }
586
587 for(int dim = 0; dim < 3; ++dim) {
588 if(dim < myGeo->numDimensions) {
589 myGeo->offsets[dim]= Teuchos::as<LO>(myGeo->startIndices[dim]) % myGeo->coarseRate[dim];
590 myGeo->offsets[dim + 3]= Teuchos::as<LO>(myGeo->startIndices[dim]) % myGeo->coarseRate[dim];
591 }
592 }
593
594 // Check if the partition contains nodes on a boundary, if so that boundary (face, line or
595 // point) will not require ghost nodes, unless there is only one node in that direction.
596 for(int dim = 0; dim < 3; ++dim) {
597 if(dim < myGeo->numDimensions && (myGeo->startIndices[dim] % myGeo->coarseRate[dim] != 0 ||
598 myGeo->startIndices[dim] == myGeo->gFineNodesPerDir[dim]-1)) {
599 myGeo->ghostInterface[2*dim] = true;
600 }
601 if(dim < myGeo->numDimensions
602 && myGeo->startIndices[dim + 3] != myGeo->gFineNodesPerDir[dim] - 1
603 && (myGeo->lFineNodesPerDir[dim] == 1 ||
604 myGeo->startIndices[dim + 3] % myGeo->coarseRate[dim] != 0)) {
605 myGeo->ghostInterface[2*dim + 1] = true;
606 }
607 }
608
609 // Here one element can represent either the degenerate case of one node or the more general
610 // case of two nodes, i.e. x---x is a 1D element with two nodes and x is a 1D element with one
611 // node. This helps generating a 3D space from tensorial products...
612 // A good way to handle this would be to generalize the algorithm to take into account the
613 // discretization order used in each direction, at least in the FEM sense, since a 0 degree
614 // discretization will have a unique node per element. This way 1D discretization can be viewed
615 // as a 3D problem with one 0 degree element in the y direction and one 0 degre element in the z
616 // direction.
617 // !!! Operations below are aftecting both local and global values that have two different !!!
618 // orientations. Orientations can be interchanged using mapDirG2L and mapDirL2G. coarseRate,
619 // endRate and offsets are in the global basis, as well as all the variables starting with a g,
620 // !!! while the variables starting with an l are in the local basis. !!!
621 for(int i = 0; i < 3; ++i) {
622 if(i < myGeo->numDimensions) {
623 // This array is passed to the RAPFactory and eventually becomes gFineNodePerDir on the next
624 // level.
625 myGeo->gCoarseNodesPerDir[i] = (myGeo->gFineNodesPerDir[i] - 1) / myGeo->coarseRate[i];
626 myGeo->endRate[i] = Teuchos::as<LO>((myGeo->gFineNodesPerDir[i] - 1) %myGeo->coarseRate[i]);
627 if(myGeo->endRate[i] == 0) {
628 myGeo->endRate[i] = myGeo->coarseRate[i];
629 ++myGeo->gCoarseNodesPerDir[i];
630 } else {
631 myGeo->gCoarseNodesPerDir[i] += 2;
632 }
633 } else {
634 myGeo->endRate[i] = 1;
635 myGeo->gCoarseNodesPerDir[i] = 1;
636 }
637 }
638
639 myGeo->gNumCoarseNodes = myGeo->gCoarseNodesPerDir[0]*myGeo->gCoarseNodesPerDir[1]
640 *myGeo->gCoarseNodesPerDir[2];
641
642 for(LO i = 0; i < 3; ++i) {
643 if(i < myGeo->numDimensions) {
644 // Check whether the partition includes the "end" of the mesh which means that endRate will
645 // apply. Also make sure that endRate is not 0 which means that the mesh does not require a
646 // particular treatment at the boundaries.
647 if( (myGeo->startIndices[i] + myGeo->lFineNodesPerDir[i]) == myGeo->gFineNodesPerDir[i] ) {
648 myGeo->lCoarseNodesPerDir[i] = (myGeo->lFineNodesPerDir[i] - myGeo->endRate[i]
649 + myGeo->offsets[i] - 1) / myGeo->coarseRate[i] + 1;
650 if(myGeo->offsets[i] == 0) {++myGeo->lCoarseNodesPerDir[i];}
651 } else {
652 myGeo->lCoarseNodesPerDir[i] = (myGeo->lFineNodesPerDir[i] + myGeo->offsets[i] - 1)
653 / myGeo->coarseRate[i];
654 if(myGeo->offsets[i] == 0) {++myGeo->lCoarseNodesPerDir[i];}
655 }
656 } else {
657 myGeo->lCoarseNodesPerDir[i] = 1;
658 }
659 // This would happen if the rank does not own any nodes but in that case a subcommunicator
660 // should be used so this should really not be a concern.
661 if(myGeo->lFineNodesPerDir[i] < 1) {myGeo->lCoarseNodesPerDir[i] = 0;}
662 }
663
664 // Assuming linear interpolation, each fine point has contribution from 8 coarse points
665 // and each coarse point value gets injected.
666 // For systems of PDEs we assume that all dofs have the same P operator.
667 myGeo->lNumCoarseNodes = myGeo->lCoarseNodesPerDir[0]*myGeo->lCoarseNodesPerDir[1]
668 *myGeo->lCoarseNodesPerDir[2];
669
670 // For each direction, determine how many points (including ghosts) are required.
671 bool ghostedDir[6] = {false};
672 for(int dim = 0; dim < 3; ++dim) {
673 // The first branch of this if-statement will be used if the rank contains only one layer
674 // of nodes in direction i, that layer must also coincide with the boundary of the mesh
675 // and coarseRate[i] == endRate[i]...
676 if(myGeo->startIndices[dim] == myGeo->gFineNodesPerDir[dim] - 1 &&
677 myGeo->startIndices[dim] % myGeo->coarseRate[dim] == 0) {
678 myGeo->startGhostedCoarseNode[dim] = myGeo->startIndices[dim] / myGeo->coarseRate[dim] - 1;
679 } else {
680 myGeo->startGhostedCoarseNode[dim] = myGeo->startIndices[dim] / myGeo->coarseRate[dim];
681 }
682 myGeo->ghostedCoarseNodesPerDir[dim] = myGeo->lCoarseNodesPerDir[dim];
683 // Check whether face *low needs ghost nodes
684 if(myGeo->ghostInterface[2*dim]) {
685 myGeo->ghostedCoarseNodesPerDir[dim] += 1;
686 ghostedDir[2*dim] = true;
687 }
688 // Check whether face *hi needs ghost nodes
689 if(myGeo->ghostInterface[2*dim + 1]) {
690 myGeo->ghostedCoarseNodesPerDir[dim] += 1;
691 ghostedDir[2*dim + 1] = true;
692 }
693 }
694 myGeo->lNumGhostedNodes = myGeo->ghostedCoarseNodesPerDir[2]*myGeo->ghostedCoarseNodesPerDir[1]
695 *myGeo->ghostedCoarseNodesPerDir[0];
696 myGeo->lNumGhostNodes = myGeo->lNumGhostedNodes - myGeo->lNumCoarseNodes;
697 ghostedCoarseNodes->PIDs.resize(myGeo->lNumGhostedNodes);
698 ghostedCoarseNodes->LIDs.resize(myGeo->lNumGhostedNodes);
699 ghostedCoarseNodes->GIDs.resize(myGeo->lNumGhostedNodes);
700 ghostedCoarseNodes->coarseGIDs.resize(myGeo->lNumGhostedNodes);
701 ghostedCoarseNodes->colInds.resize(myGeo->lNumGhostedNodes);
702 lCoarseNodesGIDs[0].resize(myGeo->lNumCoarseNodes);
703 lCoarseNodesGIDs[1].resize(myGeo->lNumCoarseNodes);
704
705 // Now the tricky part starts, the coarse nodes / ghosted coarse nodes need to be imported.
706 // This requires finding what their GID on the fine mesh is. They need to be ordered
707 // lexicographically to allow for fast sweeps through the mesh.
708
709 // We loop over all ghosted coarse nodes by increasing global lexicographic order
710 Array<LO> coarseNodeCoarseIndices(3), coarseNodeFineIndices(3), ijk(3);
711 LO currentIndex = -1, countCoarseNodes = 0;
712 for(ijk[2] = 0; ijk[2] < myGeo->ghostedCoarseNodesPerDir[2]; ++ijk[2]) {
713 for(ijk[1] = 0; ijk[1] < myGeo->ghostedCoarseNodesPerDir[1]; ++ijk[1]) {
714 for(ijk[0] = 0; ijk[0] < myGeo->ghostedCoarseNodesPerDir[0]; ++ijk[0]) {
715 currentIndex = ijk[2]*myGeo->ghostedCoarseNodesPerDir[1]*myGeo->ghostedCoarseNodesPerDir[0]
716 + ijk[1]*myGeo->ghostedCoarseNodesPerDir[0] + ijk[0];
717 coarseNodeCoarseIndices[0] = myGeo->startGhostedCoarseNode[0] + ijk[0];
718 coarseNodeFineIndices[0] = coarseNodeCoarseIndices[0]*myGeo->coarseRate[0];
719 if(coarseNodeFineIndices[0] > myGeo->gFineNodesPerDir[0] - 1) {
720 coarseNodeFineIndices[0] = myGeo->gFineNodesPerDir[0] - 1;
721 }
722 coarseNodeCoarseIndices[1] = myGeo->startGhostedCoarseNode[1] + ijk[1];
723 coarseNodeFineIndices[1] = coarseNodeCoarseIndices[1]*myGeo->coarseRate[1];
724 if(coarseNodeFineIndices[1] > myGeo->gFineNodesPerDir[1] - 1) {
725 coarseNodeFineIndices[1] = myGeo->gFineNodesPerDir[1] - 1;
726 }
727 coarseNodeCoarseIndices[2] = myGeo->startGhostedCoarseNode[2] + ijk[2];
728 coarseNodeFineIndices[2] = coarseNodeCoarseIndices[2]*myGeo->coarseRate[2];
729 if(coarseNodeFineIndices[2] > myGeo->gFineNodesPerDir[2] - 1) {
730 coarseNodeFineIndices[2] = myGeo->gFineNodesPerDir[2] - 1;
731 }
732 GO myGID = 0, myCoarseGID = -1;
733 GO factor[3] = {};
734 factor[2] = myGeo->gNumFineNodes10;
735 factor[1] = myGeo->gFineNodesPerDir[0];
736 factor[0] = 1;
737 for(int dim = 0; dim < 3; ++dim) {
738 if(dim < myGeo->numDimensions) {
739 if(myGeo->startIndices[dim] - myGeo->offsets[dim] + ijk[dim]*myGeo->coarseRate[dim]
740 < myGeo->gFineNodesPerDir[dim] - 1) {
741 myGID += (myGeo->startIndices[dim] - myGeo->offsets[dim]
742 + ijk[dim]*myGeo->coarseRate[dim])*factor[dim];
743 } else {
744 myGID += (myGeo->startIndices[dim] - myGeo->offsets[dim]
745 + (ijk[dim] - 1)*myGeo->coarseRate[dim] + myGeo->endRate[dim])*factor[dim];
746 }
747 }
748 }
749 myCoarseGID = coarseNodeCoarseIndices[0]
750 + coarseNodeCoarseIndices[1]*myGeo->gCoarseNodesPerDir[0]
751 + coarseNodeCoarseIndices[2]*myGeo->gCoarseNodesPerDir[1]*myGeo->gCoarseNodesPerDir[0];
752 ghostedCoarseNodes->GIDs[currentIndex] = myGID;
753 ghostedCoarseNodes->coarseGIDs[currentIndex] = myCoarseGID;
754 if((!ghostedDir[0] || ijk[0] != 0)
755 && (!ghostedDir[2] || ijk[1] != 0)
756 && (!ghostedDir[4] || ijk[2] != 0)
757 && (!ghostedDir[1] || ijk[0] != myGeo->ghostedCoarseNodesPerDir[0] - 1)
758 && (!ghostedDir[3] || ijk[1] != myGeo->ghostedCoarseNodesPerDir[1] - 1)
759 && (!ghostedDir[5] || ijk[2] != myGeo->ghostedCoarseNodesPerDir[2] - 1)){
760 lCoarseNodesGIDs[0][countCoarseNodes] = myCoarseGID;
761 lCoarseNodesGIDs[1][countCoarseNodes] = myGID;
762 ++countCoarseNodes;
763 }
764 }
765 }
766 }
767 Array<int> ghostsPIDs(myGeo->lNumGhostedNodes);
768 Array<LO> ghostsLIDs(myGeo->lNumGhostedNodes);
769 fineCoordsMap->getRemoteIndexList(ghostedCoarseNodes->GIDs(),
770 ghostedCoarseNodes->PIDs(),
771 ghostedCoarseNodes->LIDs());
772 } // End GetCoarsePoint
773
774 template <class Scalar, class LocalOrdinal, class GlobalOrdinal, class Node>
775 void GeneralGeometricPFactory<Scalar, LocalOrdinal, GlobalOrdinal, Node>::
776 MakeGeneralGeometricP(RCP<GeometricData> myGeo,
777 const RCP<Xpetra::MultiVector<typename Teuchos::ScalarTraits<Scalar>::coordinateType,LO,GO,Node> >& fineCoords,
778 const LO nnzP, const LO dofsPerNode,
779 RCP<const Map>& stridedDomainMapP,RCP<Matrix> & Amat, RCP<Matrix>& P,
780 RCP<Xpetra::MultiVector<typename Teuchos::ScalarTraits<Scalar>::coordinateType,LO,GO,Node> >& coarseCoords,
781 RCP<NodesIDs> ghostedCoarseNodes, Array<Array<GO> > coarseNodesGIDs,
782 int interpolationOrder) const {
783
784 /* On termination, return the number of local prolongator columns owned by
785 * this processor.
786 *
787 * Input
788 * =====
789 * nNodes Number of fine level Blk Rows owned by this processor
790 * coarseRate Rate of coarsening in each spatial direction.
791 * endRate Rate of coarsening in each spatial direction for the last
792 * nodes in the mesh where an adaptive coarsening rate is
793 * required.
794 * nTerms Number of nonzero entries in the prolongation matrix.
795 * dofsPerNode Number of degrees-of-freedom per mesh node.
796 *
797 * Output
798 * =====
799 * So far nothing...
800 */
801
802 using xdMV = Xpetra::MultiVector<typename Teuchos::ScalarTraits<Scalar>::coordinateType,LO,GO,NO>;
803 Xpetra::global_size_t OTI = Teuchos::OrdinalTraits<Xpetra::global_size_t>::invalid();
804
805 LO myRank = Amat->getRowMap()->getComm()->getRank();
806 GO numGloCols = dofsPerNode*myGeo->gNumCoarseNodes;
807
808 // Build maps necessary to create and fill complete the prolongator
809 // note: rowMapP == rangeMapP and colMapP != domainMapP
810 RCP<const Map> rowMapP = Amat->getDomainMap();
811
812 RCP<const Map> domainMapP;
813
814 RCP<const Map> colMapP;
815
816 // At this point we need to create the column map which is a delicate operation.
817 // The entries in that map need to be ordered as follows:
818 // 1) first owned entries ordered by LID
819 // 2) second order the remaining entries by PID
820 // 3) entries with the same remote PID are ordered by GID.
821 // One piece of good news: myGeo->lNumCoarseNodes is the number of ownedNodes and
822 // myGeo->lNumGhostNodes is the number of remote nodes. The sorting can be limited to remote
823 // nodes as the owned ones are alreaded ordered by LID!
824
825 {
826 std::vector<NodeID> colMapOrdering(myGeo->lNumGhostedNodes);
827 for(LO ind = 0; ind < myGeo->lNumGhostedNodes; ++ind) {
828 colMapOrdering[ind].GID = ghostedCoarseNodes->GIDs[ind];
829 if(ghostedCoarseNodes->PIDs[ind] == myRank) {
830 colMapOrdering[ind].PID = -1;
831 } else {
832 colMapOrdering[ind].PID = ghostedCoarseNodes->PIDs[ind];
833 }
834 colMapOrdering[ind].LID = ghostedCoarseNodes->LIDs[ind];
835 colMapOrdering[ind].lexiInd = ind;
836 }
837 std::sort(colMapOrdering.begin(), colMapOrdering.end(),
838 [](NodeID a, NodeID b)->bool{
839 return ( (a.PID < b.PID) || ((a.PID == b.PID) && (a.GID < b.GID)) );
840 });
841
842 Array<GO> colGIDs(dofsPerNode*myGeo->lNumGhostedNodes);
843 for(LO ind = 0; ind < myGeo->lNumGhostedNodes; ++ind) {
844 // Save the permutation calculated to go from Lexicographic indexing to column map indexing
845 ghostedCoarseNodes->colInds[colMapOrdering[ind].lexiInd] = ind;
846 for(LO dof = 0; dof < dofsPerNode; ++dof) {
847 colGIDs[dofsPerNode*ind + dof] = dofsPerNode*colMapOrdering[ind].GID + dof;
848 }
849 }
850 domainMapP = Xpetra::MapFactory<LO,GO,NO>::Build(rowMapP->lib(),
851 numGloCols,
852 colGIDs.view(0,dofsPerNode*
853 myGeo->lNumCoarseNodes),
854 rowMapP->getIndexBase(),
855 rowMapP->getComm());
856 colMapP = Xpetra::MapFactory<LO,GO,NO>::Build(rowMapP->lib(),
857 OTI,
858 colGIDs.view(0, colGIDs.size()),
859 rowMapP->getIndexBase(),
860 rowMapP->getComm());
861 } // End of scope for colMapOrdering and colGIDs
862
863 std::vector<size_t> strideInfo(1);
864 strideInfo[0] = dofsPerNode;
865 stridedDomainMapP = Xpetra::StridedMapFactory<LO,GO,NO>::Build(domainMapP, strideInfo);
866
867 // Build the map for the coarse level coordinates, create the associated MultiVector and fill it
868 // with an import from the fine coordinates MultiVector. As data is local this should not create
869 // communications during the importer creation.
870 RCP<const Map> coarseCoordsMap = MapFactory::Build (fineCoords->getMap()->lib(),
871 myGeo->gNumCoarseNodes,
872 coarseNodesGIDs[0](),
873 fineCoords->getMap()->getIndexBase(),
874 fineCoords->getMap()->getComm());
875 RCP<const Map> coarseCoordsFineMap = MapFactory::Build (fineCoords->getMap()->lib(),
876 myGeo->gNumCoarseNodes,
877 coarseNodesGIDs[1](),
878 fineCoords->getMap()->getIndexBase(),
879 fineCoords->getMap()->getComm());
880
881 RCP<const Import> coarseImporter = ImportFactory::Build(fineCoords->getMap(),
882 coarseCoordsFineMap);
883 coarseCoords = Xpetra::MultiVectorFactory<typename Teuchos::ScalarTraits<Scalar>::coordinateType,LO,GO,NO>::Build(coarseCoordsFineMap,
884 myGeo->numDimensions);
885 coarseCoords->doImport(*fineCoords, *coarseImporter, Xpetra::INSERT);
886 coarseCoords->replaceMap(coarseCoordsMap);
887
888 // Do the actual import using the fineCoords->getMap()
889 RCP<const Map> ghostMap = Xpetra::MapFactory<LO,GO,NO>::Build(fineCoords->getMap()->lib(),
890 OTI,
891 ghostedCoarseNodes->GIDs(),
892 fineCoords->getMap()->getIndexBase(),
893 fineCoords->getMap()->getComm());
894 RCP<const Import> ghostImporter = ImportFactory::Build(fineCoords->getMap(), ghostMap);
895 RCP<xdMV> ghostCoords = Xpetra::MultiVectorFactory<typename Teuchos::ScalarTraits<Scalar>::coordinateType,LO,GO,NO>::Build(ghostMap,
896 myGeo->numDimensions);
897 ghostCoords->doImport(*fineCoords, *ghostImporter, Xpetra::INSERT);
898
899 P = rcp(new CrsMatrixWrap(rowMapP, colMapP, 0));
900 RCP<CrsMatrix> PCrs = rcp_dynamic_cast<CrsMatrixWrap>(P)->getCrsMatrix();
901
902 ArrayRCP<size_t> iaP;
903 ArrayRCP<LO> jaP;
904 ArrayRCP<SC> valP;
905
906 PCrs->allocateAllValues(nnzP, iaP, jaP, valP);
907
908 ArrayView<size_t> ia = iaP();
909 ArrayView<LO> ja = jaP();
910 ArrayView<SC> val = valP();
911 ia[0] = 0;
912
913 Array<ArrayRCP<typename Teuchos::ScalarTraits<Scalar>::coordinateType> > ghostedCoords(3);
914 {
915 ArrayRCP<typename Teuchos::ScalarTraits<Scalar>::coordinateType> tmp(ghostCoords->getLocalLength(), 0.0);
916 for(int dim = 0; dim < 3; ++dim) {
917 if(dim < myGeo->numDimensions) {
918 ghostedCoords[dim] = ghostCoords->getDataNonConst(dim);
919 } else {
920 ghostedCoords[dim] = tmp;
921 }
922 }
923 }
924
925 // Declaration and assignment of fineCoords which holds the coordinates of the fine nodes in 3D.
926 // To do so we pull the nD coordinates from fineCoords and pad the rest with zero vectors...
927 RCP<Xpetra::Vector<typename Teuchos::ScalarTraits<Scalar>::coordinateType,LO,GO,NO> > zeros
928 = Xpetra::VectorFactory<typename Teuchos::ScalarTraits<Scalar>::coordinateType,LO,GO,NO>::Build(fineCoords->getMap(), true);
929 ArrayRCP< ArrayRCP<typename Teuchos::ScalarTraits<Scalar>::coordinateType> > lFineCoords(3);
930 for(int dim = 0; dim < 3; ++dim) {
931 if(dim < myGeo->numDimensions) {
932 lFineCoords[dim] = fineCoords->getDataNonConst(dim);
933 } else {
934 lFineCoords[dim] = zeros->getDataNonConst(0);
935 }
936 }
937
938 GO tStencil = 0;
939 for(int currentIndex = 0; currentIndex < myGeo->lNumFineNodes; ++currentIndex) {
940 Array<GO> ghostedIndices(3), firstInterpolationIndices(3);
941 Array<LO> interpolationPIDs(8), interpolationLIDs(8), interpolationGIDs(8);
942 Array<Array<typename Teuchos::ScalarTraits<Scalar>::coordinateType> > interpolationCoords(9);
943 interpolationCoords[0].resize(3);
944 GO firstInterpolationNodeIndex;
945 int nStencil = 0;
946 for(int dim = 0; dim < 3; ++dim) {
947 interpolationCoords[0][dim] = lFineCoords[dim][currentIndex];
948 }
949
950 // Compute the ghosted (i,j,k) of the current node, that assumes (I,J,K)=(0,0,0) to be
951 // associated with the first node in ghostCoords
952 {// Scope for tmp
953 ghostedIndices[2] = currentIndex / (myGeo->lFineNodesPerDir[1]*myGeo->lFineNodesPerDir[0]);
954 LO tmp = currentIndex % (myGeo->lFineNodesPerDir[1]*myGeo->lFineNodesPerDir[0]);
955 ghostedIndices[1] = tmp / myGeo->lFineNodesPerDir[0];
956 ghostedIndices[0] = tmp % myGeo->lFineNodesPerDir[0];
957 for(int dim = 0; dim < 3; ++dim) {
958 ghostedIndices[dim] += myGeo->offsets[dim];
959 }
960 // A special case appears when the mesh is really coarse: it is possible for a rank to
961 // have a single coarse node in a given direction. If this happens on the highest i, j or k
962 // we need to "grab" a coarse node with a lower i, j, or k which leads us to add to the
963 // value of ghostedIndices
964 }
965 // No we find the indices of the coarse nodes used for interpolation simply by integer
966 // division.
967 for(int dim = 0; dim < 3; ++dim) {
968 firstInterpolationIndices[dim] = ghostedIndices[dim] / myGeo->coarseRate[dim];
969 // If you are on the edge of the local domain go back one coarse node, unless there is only
970 // one node on the local domain...
971 if(firstInterpolationIndices[dim] == myGeo->ghostedCoarseNodesPerDir[dim] - 1
972 && myGeo->ghostedCoarseNodesPerDir[dim] > 1) {
973 firstInterpolationIndices[dim] -= 1;
974 }
975 }
976 firstInterpolationNodeIndex =
977 firstInterpolationIndices[2]*myGeo->ghostedCoarseNodesPerDir[1]*myGeo->ghostedCoarseNodesPerDir[0]
978 + firstInterpolationIndices[1]*myGeo->ghostedCoarseNodesPerDir[0]
979 + firstInterpolationIndices[0];
980
981 // We extract the coordinates and PIDs associated with each coarse node used during
982 // inteprolation in order to fill the prolongator correctly
983 {
984 LO ind = -1;
985 for(int k = 0; k < 2; ++k) {
986 for(int j = 0; j < 2; ++j) {
987 for(int i = 0; i < 2; ++i) {
988 ind = k*4 + j*2 + i;
989 interpolationLIDs[ind] = firstInterpolationNodeIndex
990 + k*myGeo->ghostedCoarseNodesPerDir[1]*myGeo->ghostedCoarseNodesPerDir[0]
991 + j*myGeo->ghostedCoarseNodesPerDir[0] + i;
992 if(ghostedCoarseNodes->PIDs[interpolationLIDs[ind]] == rowMapP->getComm()->getRank()){
993 interpolationPIDs[ind] = -1;
994 } else {
995 interpolationPIDs[ind] = ghostedCoarseNodes->PIDs[interpolationLIDs[ind]];
996 }
997 interpolationGIDs[ind] = ghostedCoarseNodes->coarseGIDs[interpolationLIDs[ind]];
998
999 interpolationCoords[ind + 1].resize(3);
1000 for(int dim = 0; dim < 3; ++dim) {
1001 interpolationCoords[ind + 1][dim]
1002 = ghostedCoords[dim][interpolationLIDs[ind]];
1003 }
1004 }
1005 }
1006 }
1007 } // End of ind scope
1008
1009 // Compute the actual geometric interpolation stencil
1010 // LO stencilLength = static_cast<LO>(std::pow(interpolationOrder + 1, 3));
1011 std::vector<double> stencil(8);
1012 Array<GO> firstCoarseNodeFineIndices(3);
1013 int rate[3] = {};
1014 for(int dim = 0; dim < 3; ++dim) {
1015 firstCoarseNodeFineIndices[dim] = firstInterpolationIndices[dim]*myGeo->coarseRate[dim];
1016 if((myGeo->startIndices[dim + 3] == myGeo->gFineNodesPerDir[dim] - 1)
1017 && (ghostedIndices[dim] >=
1018 (myGeo->ghostedCoarseNodesPerDir[dim] - 2)*myGeo->coarseRate[dim])) {
1019 rate[dim] = myGeo->endRate[dim];
1020 } else {
1021 rate[dim] = myGeo->coarseRate[dim];
1022 }
1023 }
1024 Array<GO> trueGhostedIndices(3);
1025 // This handles the case of a rank having a single node that also happens to be the last node
1026 // in any direction. It might be more efficient to re-write the algo so that this is
1027 // incorporated in the definition of ghostedIndices directly...
1028 for(int dim = 0; dim < 3; ++dim) {
1029 if (myGeo->startIndices[dim] == myGeo->gFineNodesPerDir[dim] - 1) {
1030 trueGhostedIndices[dim] = ghostedIndices[dim] + rate[dim];
1031 } else {
1032 trueGhostedIndices[dim] = ghostedIndices[dim];
1033 }
1034 }
1035 ComputeStencil(myGeo->numDimensions, trueGhostedIndices, firstCoarseNodeFineIndices, rate,
1036 interpolationCoords, interpolationOrder, stencil);
1037
1038 // Finally check whether the fine node is on a coarse: node, edge or face
1039 // and select accordingly the non-zero values from the stencil and the
1040 // corresponding column indices
1041 Array<LO> nzIndStencil(8), permutation(8);
1042 for(LO k = 0; k < 8; ++k) {permutation[k] = k;}
1043 if(interpolationOrder == 0) {
1044 nStencil = 1;
1045 for(LO k = 0; k < 8; ++k) {
1046 nzIndStencil[k] = static_cast<LO>(stencil[0]);
1047 }
1048 stencil[0] = 0.0;
1049 stencil[nzIndStencil[0]] = 1.0;
1050 } else if(interpolationOrder == 1) {
1051 Array<GO> currentNodeGlobalFineIndices(3);
1052 for(int dim = 0; dim < 3; ++dim) {
1053 currentNodeGlobalFineIndices[dim] = ghostedIndices[dim] - myGeo->offsets[dim]
1054 + myGeo->startIndices[dim];
1055 }
1056 if( ((ghostedIndices[0] % myGeo->coarseRate[0] == 0)
1057 || currentNodeGlobalFineIndices[0] == myGeo->gFineNodesPerDir[0] - 1)
1058 && ((ghostedIndices[1] % myGeo->coarseRate[1] == 0)
1059 || currentNodeGlobalFineIndices[1] == myGeo->gFineNodesPerDir[1] - 1)
1060 && ((ghostedIndices[2] % myGeo->coarseRate[2] == 0)
1061 || currentNodeGlobalFineIndices[2] == myGeo->gFineNodesPerDir[2] - 1) ) {
1062 if((currentNodeGlobalFineIndices[0] == myGeo->gFineNodesPerDir[0] - 1) ||
1063 (ghostedIndices[0] / myGeo->coarseRate[0] == myGeo->ghostedCoarseNodesPerDir[0] - 1)) {
1064 nzIndStencil[0] += 1;
1065 }
1066 if(((currentNodeGlobalFineIndices[1] == myGeo->gFineNodesPerDir[1] - 1) ||
1067 (ghostedIndices[1] / myGeo->coarseRate[1] == myGeo->ghostedCoarseNodesPerDir[1] - 1))
1068 && (myGeo->numDimensions > 1)){
1069 nzIndStencil[0] += 2;
1070 }
1071 if(((currentNodeGlobalFineIndices[2] == myGeo->gFineNodesPerDir[2] - 1) ||
1072 (ghostedIndices[2] / myGeo->coarseRate[2] == myGeo->ghostedCoarseNodesPerDir[2] - 1))
1073 && myGeo->numDimensions > 2) {
1074 nzIndStencil[0] += 4;
1075 }
1076 nStencil = 1;
1077 for(LO k = 0; k < 8; ++k) {
1078 nzIndStencil[k] = nzIndStencil[0];
1079 }
1080 } else {
1081 nStencil = 8;
1082 for(LO k = 0; k < 8; ++k) {
1083 nzIndStencil[k] = k;
1084 }
1085 }
1086 }
1087
1088 // Here the values are filled in the Crs matrix arrays
1089 // This is basically the only place these variables are modified
1090 // Hopefully this makes handling system of PDEs easy!
1091
1092 // Loop on dofsPerNode and process each row for the current Node
1093
1094
1095 // Sort nodes by PIDs using stable sort to keep sublist ordered by LIDs and GIDs
1096 sh_sort2(interpolationPIDs.begin(),interpolationPIDs.end(),
1097 permutation.begin(), permutation.end());
1098
1099 GO colInd;
1100 for(LO j = 0; j < dofsPerNode; ++j) {
1101 ia[currentIndex*dofsPerNode + j + 1] = ia[currentIndex*dofsPerNode + j] + nStencil;
1102 for(LO k = 0; k < nStencil; ++k) {
1103 colInd = ghostedCoarseNodes->colInds[interpolationLIDs[nzIndStencil[permutation[k]]]];
1104 ja[ia[currentIndex*dofsPerNode + j] + k] = colInd*dofsPerNode + j;
1105 val[ia[currentIndex*dofsPerNode + j] + k] = stencil[nzIndStencil[permutation[k]]];
1106 }
1107 // Add the stencil for each degree of freedom.
1108 tStencil += nStencil;
1109 }
1110 } // End loop over fine nodes
1111
1112 if (rowMapP->lib() == Xpetra::UseTpetra) {
1113 // - Cannot resize for Epetra, as it checks for same pointers
1114 // - Need to resize for Tpetra, as it check ().size() == ia[numRows]
1115 // NOTE: these invalidate ja and val views
1116 jaP .resize(tStencil);
1117 valP.resize(tStencil);
1118 }
1119
1120 // Set the values of the prolongation operators into the CrsMatrix P and call FillComplete
1121 PCrs->setAllValues(iaP, jaP, valP);
1122 PCrs->expertStaticFillComplete(domainMapP,rowMapP);
1123 } // End MakeGeneralGeometricP
1124
1125 // template <class Scalar, class LocalOrdinal, class GlobalOrdinal, class Node>
1126 // void BlackBoxPFactory<Scalar, LocalOrdinal, GlobalOrdinal, Node>::GetGeometricData(
1127 // RCP<Xpetra::MultiVector<typename Teuchos::ScalarTraits<Scalar>::coordinateType,LO,GO,NO> >& coordinates, const Array<LO> coarseRate,
1128 // const Array<GO> gFineNodesPerDir, const Array<LO> lFineNodesPerDir, const LO BlkSize,
1129 // Array<GO>& gIndices, Array<LO>& myOffset, Array<bool>& ghostInterface, Array<LO>& endRate,
1130 // Array<GO>& gCoarseNodesPerDir, Array<LO>& lCoarseNodesPerDir, Array<GO>& ghostGIDs,
1131 // Array<GO>& coarseNodesGIDs, Array<GO>& colGIDs, GO& gNumCoarseNodes, LO& lNumCoarseNodes,
1132 // ArrayRCP<Array<double> > coarseNodes) const {
1133
1134 // RCP<const Map> coordinatesMap = coordinates->getMap();
1135 // LO numDimensions = coordinates->getNumVectors();
1136
1137 // // Using the coarsening rate and the fine level data,
1138 // // compute coarse level data
1139
1140 // // Phase 1 //
1141 // // ------------------------------------------------------------------ //
1142 // // We first start by finding small informations on the mesh such as //
1143 // // the number of coarse nodes (local and global) and the number of //
1144 // // ghost nodes / the end rate of coarsening. //
1145 // // ------------------------------------------------------------------ //
1146 // GO minGlobalIndex = coordinatesMap->getMinGlobalIndex();
1147 // {
1148 // GO tmp;
1149 // gIndices[2] = minGlobalIndex / (gFineNodesPerDir[1]*gFineNodesPerDir[0]);
1150 // tmp = minGlobalIndex % (gFineNodesPerDir[1]*gFineNodesPerDir[0]);
1151 // gIndices[1] = tmp / gFineNodesPerDir[0];
1152 // gIndices[0] = tmp % gFineNodesPerDir[0];
1153
1154 // myOffset[2] = gIndices[2] % coarseRate[2];
1155 // myOffset[1] = gIndices[1] % coarseRate[1];
1156 // myOffset[0] = gIndices[0] % coarseRate[0];
1157 // }
1158
1159 // // Check whether ghost nodes are needed in each direction
1160 // for(LO i=0; i < numDimensions; ++i) {
1161 // if((gIndices[i] != 0) && (gIndices[i] % coarseRate[i] > 0)) {
1162 // ghostInterface[2*i] = true;
1163 // }
1164 // if(((gIndices[i] + lFineNodesPerDir[i]) != gFineNodesPerDir[i]) && ((gIndices[i] + lFineNodesPerDir[i] - 1) % coarseRate[i] > 0)) {
1165 // ghostInterface[2*i + 1] = true;
1166 // }
1167 // }
1168
1169 // for(LO i = 0; i < 3; ++i) {
1170 // if(i < numDimensions) {
1171 // lCoarseNodesPerDir[i] = (lFineNodesPerDir[i] + myOffset[i] - 1) / coarseRate[i];
1172 // if(myOffset[i] == 0) { ++lCoarseNodesPerDir[i]; }
1173 // gCoarseNodesPerDir[i] = (gFineNodesPerDir[i] - 1) / coarseRate[i];
1174 // endRate[i] = (gFineNodesPerDir[i] - 1) % coarseRate[i];
1175 // if(endRate[i] == 0) {
1176 // ++gCoarseNodesPerDir[i];
1177 // endRate[i] = coarseRate[i];
1178 // }
1179 // } else {
1180 // // Most quantities need to be set to 1 for extra dimensions
1181 // // this is rather logical, an x-y plane is like a single layer
1182 // // of nodes in the z direction...
1183 // gCoarseNodesPerDir[i] = 1;
1184 // lCoarseNodesPerDir[i] = 1;
1185 // endRate[i] = 1;
1186 // }
1187 // }
1188
1189 // gNumCoarseNodes = gCoarseNodesPerDir[0]*gCoarseNodesPerDir[1]*gCoarseNodesPerDir[2];
1190 // lNumCoarseNodes = lCoarseNodesPerDir[0]*lCoarseNodesPerDir[1]*lCoarseNodesPerDir[2];
1191
1192 // // For each direction, determine how many ghost points are required.
1193 // LO lNumGhostNodes = 0;
1194 // {
1195 // const int complementaryIndices[3][2] = {{1,2}, {0,2}, {0,1}};
1196 // LO tmp = 0;
1197 // for(LO i = 0; i < 3; ++i) {
1198 // // Check whether a face in direction i needs ghost nodes
1199 // if(ghostInterface[2*i] || ghostInterface[2*i+1]) {
1200 // if(i == 0) {tmp = lCoarseNodesPerDir[1]*lCoarseNodesPerDir[2];}
1201 // if(i == 1) {tmp = lCoarseNodesPerDir[0]*lCoarseNodesPerDir[2];}
1202 // if(i == 2) {tmp = lCoarseNodesPerDir[0]*lCoarseNodesPerDir[1];}
1203 // }
1204 // // If both faces in direction i need nodes, double the number of ghost nodes
1205 // if(ghostInterface[2*i] && ghostInterface[2*i+1]) {tmp = 2*tmp;}
1206 // lNumGhostNodes += tmp;
1207
1208 // // The corners and edges need to be checked in 2D / 3D to add more ghosts nodes
1209 // for(LO j = 0; j < 2; ++j) {
1210 // for(LO k = 0; k < 2; ++k) {
1211 // // Check if two adjoining faces need ghost nodes and then add their common edge
1212 // if(ghostInterface[2*complementaryIndices[i][0]+j] && ghostInterface[2*complementaryIndices[i][1]+k]) {
1213 // lNumGhostNodes += lCoarseNodesPerDir[i];
1214 // // Add corners if three adjoining faces need ghost nodes,
1215 // // but add them only once! Hence when i == 0.
1216 // if(ghostInterface[2*i] && (i == 0)) { lNumGhostNodes += 1; }
1217 // if(ghostInterface[2*i+1] && (i == 0)) { lNumGhostNodes += 1; }
1218 // }
1219 // }
1220 // }
1221 // tmp = 0;
1222 // }
1223 // } // end of scope for tmp and complementaryIndices
1224
1225 // // Phase 2 //
1226 // // ------------------------------------------------------------------ //
1227 // // Build a list of GIDs to import the required ghost nodes. //
1228 // // The ordering of the ghosts nodes will be as natural as possible, //
1229 // // i.e. it should follow the GID ordering of the mesh. //
1230 // // ------------------------------------------------------------------ //
1231
1232 // // Saddly we have to more or less redo what was just done to figure out the value of lNumGhostNodes,
1233 // // there might be some optimization possibility here...
1234 // ghostGIDs.resize(lNumGhostNodes);
1235 // LO countGhosts = 0;
1236 // // Get the GID of the first point on the current partition.
1237 // GO startingGID = minGlobalIndex;
1238 // Array<GO> startingIndices(3);
1239 // // We still want ghost nodes even if have with a 0 offset,
1240 // // except when we are on a boundary
1241 // if(ghostInterface[4] && (myOffset[2] == 0)) {
1242 // if(gIndices[2] + coarseRate[2] > gFineNodesPerDir[2]) {
1243 // startingGID -= endRate[2]*gFineNodesPerDir[1]*gFineNodesPerDir[0];
1244 // } else {
1245 // startingGID -= coarseRate[2]*gFineNodesPerDir[1]*gFineNodesPerDir[0];
1246 // }
1247 // } else {
1248 // startingGID -= myOffset[2]*gFineNodesPerDir[1]*gFineNodesPerDir[0];
1249 // }
1250 // if(ghostInterface[2] && (myOffset[1] == 0)) {
1251 // if(gIndices[1] + coarseRate[1] > gFineNodesPerDir[1]) {
1252 // startingGID -= endRate[1]*gFineNodesPerDir[0];
1253 // } else {
1254 // startingGID -= coarseRate[1]*gFineNodesPerDir[0];
1255 // }
1256 // } else {
1257 // startingGID -= myOffset[1]*gFineNodesPerDir[0];
1258 // }
1259 // if(ghostInterface[0] && (myOffset[0] == 0)) {
1260 // if(gIndices[0] + coarseRate[0] > gFineNodesPerDir[0]) {
1261 // startingGID -= endRate[0];
1262 // } else {
1263 // startingGID -= coarseRate[0];
1264 // }
1265 // } else {
1266 // startingGID -= myOffset[0];
1267 // }
1268
1269 // { // scope for tmp
1270 // GO tmp;
1271 // startingIndices[2] = startingGID / (gFineNodesPerDir[1]*gFineNodesPerDir[0]);
1272 // tmp = startingGID % (gFineNodesPerDir[1]*gFineNodesPerDir[0]);
1273 // startingIndices[1] = tmp / gFineNodesPerDir[0];
1274 // startingIndices[0] = tmp % gFineNodesPerDir[0];
1275 // }
1276
1277 // GO ghostOffset[3] = {0, 0, 0};
1278 // LO lengthZero = lCoarseNodesPerDir[0];
1279 // LO lengthOne = lCoarseNodesPerDir[1];
1280 // LO lengthTwo = lCoarseNodesPerDir[2];
1281 // if(ghostInterface[0]) {++lengthZero;}
1282 // if(ghostInterface[1]) {++lengthZero;}
1283 // if(ghostInterface[2]) {++lengthOne;}
1284 // if(ghostInterface[3]) {++lengthOne;}
1285 // if(ghostInterface[4]) {++lengthTwo;}
1286 // if(ghostInterface[5]) {++lengthTwo;}
1287
1288 // // First check the bottom face as it will have the lowest GIDs
1289 // if(ghostInterface[4]) {
1290 // ghostOffset[2] = startingGID;
1291 // for(LO j = 0; j < lengthOne; ++j) {
1292 // if( (j == lengthOne-1) && (startingIndices[1] + j*coarseRate[1] + 1 > gFineNodesPerDir[1]) ) {
1293 // ghostOffset[1] = ((j-1)*coarseRate[1] + endRate[1])*gFineNodesPerDir[0];
1294 // } else {
1295 // ghostOffset[1] = j*coarseRate[1]*gFineNodesPerDir[0];
1296 // }
1297 // for(LO k = 0; k < lengthZero; ++k) {
1298 // if( (k == lengthZero-1) && (startingIndices[0] + k*coarseRate[0] + 1 > gFineNodesPerDir[0]) ) {
1299 // ghostOffset[0] = (k-1)*coarseRate[0] + endRate[0];
1300 // } else {
1301 // ghostOffset[0] = k*coarseRate[0];
1302 // }
1303 // // If the partition includes a changed rate at the edge, ghost nodes need to be picked carefully.
1304 // // This if statement is repeated each time a ghost node is selected.
1305 // ghostGIDs[countGhosts] = ghostOffset[2] + ghostOffset[1] + ghostOffset[0];
1306 // ++countGhosts;
1307 // }
1308 // }
1309 // }
1310
1311 // // Sweep over the lCoarseNodesPerDir[2] coarse layers in direction 2 and gather necessary ghost nodes
1312 // // located on these layers.
1313 // for(LO i = 0; i < lengthTwo; ++i) {
1314 // // Exclude the cases where ghost nodes exists on the faces in directions 2, these faces are swept
1315 // // seperatly for efficiency.
1316 // if( !((i == lengthTwo-1) && ghostInterface[5]) && !((i == 0) && ghostInterface[4]) ) {
1317 // // Set the ghostOffset in direction 2 taking into account a possible endRate different
1318 // // from the regular coarseRate.
1319 // if( (i == lengthTwo-1) && !ghostInterface[5] ) {
1320 // ghostOffset[2] = startingGID + ((i-1)*coarseRate[2] + endRate[2])*gFineNodesPerDir[1]*gFineNodesPerDir[0];
1321 // } else {
1322 // ghostOffset[2] = startingGID + i*coarseRate[2]*gFineNodesPerDir[1]*gFineNodesPerDir[0];
1323 // }
1324 // for(LO j = 0; j < lengthOne; ++j) {
1325 // if( (j == 0) && ghostInterface[2] ) {
1326 // for(LO k = 0; k < lengthZero; ++k) {
1327 // if( (k == lengthZero-1) && (startingIndices[0] + k*coarseRate[0] + 1 > gFineNodesPerDir[0]) ) {
1328 // if(k == 0) {
1329 // ghostOffset[0] = endRate[0];
1330 // } else {
1331 // ghostOffset[0] = (k-1)*coarseRate[0] + endRate[0];
1332 // }
1333 // } else {
1334 // ghostOffset[0] = k*coarseRate[0];
1335 // }
1336 // // In this case j == 0 so ghostOffset[1] is zero and can be ignored
1337 // ghostGIDs[countGhosts] = ghostOffset[2] + ghostOffset[0];
1338 // ++countGhosts;
1339 // }
1340 // } else if( (j == lengthOne-1) && ghostInterface[3] ) {
1341 // // Set the ghostOffset in direction 1 taking into account a possible endRate different
1342 // // from the regular coarseRate.
1343 // if( (j == lengthOne-1) && (startingIndices[1] + j*coarseRate[1] + 1 > gFineNodesPerDir[1]) ) {
1344 // ghostOffset[1] = ((j-1)*coarseRate[1] + endRate[1])*gFineNodesPerDir[0];
1345 // } else {
1346 // ghostOffset[1] = j*coarseRate[1]*gFineNodesPerDir[0];
1347 // }
1348 // for(LO k = 0; k < lengthZero; ++k) {
1349 // if( (k == lengthZero-1) && (startingIndices[0] + k*coarseRate[0] + 1 > gFineNodesPerDir[0]) ) {
1350 // ghostOffset[0] = (k-1)*coarseRate[0] + endRate[0];
1351 // } else {
1352 // ghostOffset[0] = k*coarseRate[0];
1353 // }
1354 // ghostGIDs[countGhosts] = ghostOffset[2] + ghostOffset[1] + ghostOffset[0];
1355 // ++countGhosts;
1356 // }
1357 // } else {
1358 // // Set ghostOffset[1] for side faces sweep
1359 // if( (j == lengthOne-1) && (startingIndices[1] + j*coarseRate[1] + 1 > gFineNodesPerDir[1]) ) {
1360 // ghostOffset[1] = ( (j-1)*coarseRate[1] + endRate[1] )*gFineNodesPerDir[0];
1361 // } else {
1362 // ghostOffset[1] = j*coarseRate[1]*gFineNodesPerDir[0];
1363 // }
1364
1365 // // Set ghostOffset[0], ghostGIDs and countGhosts
1366 // if(ghostInterface[0]) { // In that case ghostOffset[0]==0, so we can ignore it
1367 // ghostGIDs[countGhosts] = ghostOffset[2] + ghostOffset[1];
1368 // ++countGhosts;
1369 // }
1370 // if(ghostInterface[1]) { // Grab ghost point at the end of direction 0.
1371 // if( (startingIndices[0] + (lengthZero-1)*coarseRate[0]) > gFineNodesPerDir[0] - 1 ) {
1372 // if(lengthZero > 2) {
1373 // ghostOffset[0] = (lengthZero-2)*coarseRate[0] + endRate[0];
1374 // } else {
1375 // ghostOffset[0] = endRate[0];
1376 // }
1377 // } else {
1378 // ghostOffset[0] = (lengthZero-1)*coarseRate[0];
1379 // }
1380 // ghostGIDs[countGhosts] = ghostOffset[2] + ghostOffset[1] + ghostOffset[0];
1381 // ++countGhosts;
1382 // }
1383 // }
1384 // }
1385 // }
1386 // }
1387
1388 // // Finally check the top face
1389 // if(ghostInterface[5]) {
1390 // if( startingIndices[2] + (lengthTwo-1)*coarseRate[2] + 1 > gFineNodesPerDir[2] ) {
1391 // ghostOffset[2] = startingGID + ((lengthTwo-2)*coarseRate[2] + endRate[2])*gFineNodesPerDir[1]*gFineNodesPerDir[0];
1392 // } else {
1393 // ghostOffset[2] = startingGID + (lengthTwo-1)*coarseRate[2]*gFineNodesPerDir[1]*gFineNodesPerDir[0];
1394 // }
1395 // for(LO j = 0; j < lengthOne; ++j) {
1396 // if( (j == lengthOne-1) && (startingIndices[1] + j*coarseRate[1] + 1 > gFineNodesPerDir[1]) ) { // && !ghostInterface[3] ) {
1397 // ghostOffset[1] = ( (j-1)*coarseRate[1] + endRate[1] )*gFineNodesPerDir[0];
1398 // } else {
1399 // ghostOffset[1] = j*coarseRate[1]*gFineNodesPerDir[0];
1400 // }
1401 // for(LO k = 0; k < lengthZero; ++k) {
1402 // if( (k == lengthZero-1) && (startingIndices[0] + k*coarseRate[0] + 1 > gFineNodesPerDir[0]) ) {
1403 // ghostOffset[0] = (k-1)*coarseRate[0] + endRate[0];
1404 // } else {
1405 // ghostOffset[0] = k*coarseRate[0];
1406 // }
1407 // ghostGIDs[countGhosts] = ghostOffset[2] + ghostOffset[1] + ghostOffset[0];
1408 // ++countGhosts;
1409 // }
1410 // }
1411 // }
1412
1413 // // Phase 3 //
1414 // // ------------------------------------------------------------------ //
1415 // // Final phase of this function, lists are being built to create the //
1416 // // column and domain maps of the projection as well as the map and //
1417 // // arrays for the coarse coordinates multivector. //
1418 // // ------------------------------------------------------------------ //
1419
1420 // Array<GO> gCoarseNodesGIDs(lNumCoarseNodes);
1421 // LO currentNode, offset2, offset1, offset0;
1422 // // Find the GIDs of the coarse nodes on the partition.
1423 // for(LO ind2 = 0; ind2 < lCoarseNodesPerDir[2]; ++ind2) {
1424 // if(myOffset[2] == 0) {
1425 // offset2 = startingIndices[2] + myOffset[2];
1426 // } else {
1427 // if(startingIndices[2] + endRate[2] == gFineNodesPerDir[2] - 1) {
1428 // offset2 = startingIndices[2] + endRate[2];
1429 // } else {
1430 // offset2 = startingIndices[2] + coarseRate[2];
1431 // }
1432 // }
1433 // if(offset2 + ind2*coarseRate[2] > gFineNodesPerDir[2] - 1) {
1434 // offset2 += (ind2 - 1)*coarseRate[2] + endRate[2];
1435 // } else {
1436 // offset2 += ind2*coarseRate[2];
1437 // }
1438 // offset2 = offset2*gFineNodesPerDir[1]*gFineNodesPerDir[0];
1439
1440 // for(LO ind1 = 0; ind1 < lCoarseNodesPerDir[1]; ++ind1) {
1441 // if(myOffset[1] == 0) {
1442 // offset1 = startingIndices[1] + myOffset[1];
1443 // } else {
1444 // if(startingIndices[1] + endRate[1] == gFineNodesPerDir[1] - 1) {
1445 // offset1 = startingIndices[1] + endRate[1];
1446 // } else {
1447 // offset1 = startingIndices[1] + coarseRate[1];
1448 // }
1449 // }
1450 // if(offset1 + ind1*coarseRate[1] > gFineNodesPerDir[1] - 1) {
1451 // offset1 += (ind1 - 1)*coarseRate[1] + endRate[1];
1452 // } else {
1453 // offset1 += ind1*coarseRate[1];
1454 // }
1455 // offset1 = offset1*gFineNodesPerDir[0];
1456 // for(LO ind0 = 0; ind0 < lCoarseNodesPerDir[0]; ++ind0) {
1457 // offset0 = startingIndices[0];
1458 // if(myOffset[0] == 0) {
1459 // offset0 += ind0*coarseRate[0];
1460 // } else {
1461 // offset0 += (ind0 + 1)*coarseRate[0];
1462 // }
1463 // if(offset0 > gFineNodesPerDir[0] - 1) {offset0 += endRate[0] - coarseRate[0];}
1464
1465 // currentNode = ind2*lCoarseNodesPerDir[1]*lCoarseNodesPerDir[0]
1466 // + ind1*lCoarseNodesPerDir[0]
1467 // + ind0;
1468 // gCoarseNodesGIDs[currentNode] = offset2 + offset1 + offset0;
1469 // }
1470 // }
1471 // }
1472
1473 // // Actual loop over all the coarse/ghost nodes to find their index on the coarse mesh
1474 // // and the corresponding dofs that will need to be added to colMapP.
1475 // colGIDs.resize(BlkSize*(lNumCoarseNodes+lNumGhostNodes));
1476 // coarseNodesGIDs.resize(lNumCoarseNodes);
1477 // for(LO i = 0; i < numDimensions; ++i) {coarseNodes[i].resize(lNumCoarseNodes);}
1478 // GO fineNodesPerCoarseSlab = coarseRate[2]*gFineNodesPerDir[1]*gFineNodesPerDir[0];
1479 // GO fineNodesEndCoarseSlab = endRate[2]*gFineNodesPerDir[1]*gFineNodesPerDir[0];
1480 // GO fineNodesPerCoarsePlane = coarseRate[1]*gFineNodesPerDir[0];
1481 // GO fineNodesEndCoarsePlane = endRate[1]*gFineNodesPerDir[0];
1482 // GO coarseNodesPerCoarseLayer = gCoarseNodesPerDir[1]*gCoarseNodesPerDir[0];
1483 // GO gCoarseNodeOnCoarseGridGID;
1484 // LO gInd[3], lCol;
1485 // Array<int> ghostPIDs (lNumGhostNodes);
1486 // Array<LO> ghostLIDs (lNumGhostNodes);
1487 // Array<LO> ghostPermut(lNumGhostNodes);
1488 // for(LO k = 0; k < lNumGhostNodes; ++k) {ghostPermut[k] = k;}
1489 // coordinatesMap->getRemoteIndexList(ghostGIDs, ghostPIDs, ghostLIDs);
1490 // sh_sort_permute(ghostPIDs.begin(),ghostPIDs.end(), ghostPermut.begin(),ghostPermut.end());
1491
1492 // { // scope for tmpInds, tmpVars and tmp.
1493 // GO tmpInds[3], tmpVars[2];
1494 // LO tmp;
1495 // // Loop over the coarse nodes of the partition and add them to colGIDs
1496 // // that will be used to construct the column and domain maps of P as well
1497 // // as to construct the coarse coordinates map.
1498 // // for(LO col = 0; col < lNumCoarseNodes; ++col) { // This should most likely be replaced by loops of lCoarseNodesPerDir[] to simplify arithmetics
1499 // LO col = 0;
1500 // LO firstCoarseNodeInds[3], currentCoarseNode;
1501 // for(LO dim = 0; dim < 3; ++dim) {
1502 // if(myOffset[dim] == 0) {
1503 // firstCoarseNodeInds[dim] = 0;
1504 // } else {
1505 // firstCoarseNodeInds[dim] = coarseRate[dim] - myOffset[dim];
1506 // }
1507 // }
1508 // Array<ArrayRCP<const typename Teuchos::ScalarTraits<Scalar>::coordinateType> > fineNodes(numDimensions);
1509 // for(LO dim = 0; dim < numDimensions; ++dim) {fineNodes[dim] = coordinates->getData(dim);}
1510 // for(LO k = 0; k < lCoarseNodesPerDir[2]; ++k) {
1511 // for(LO j = 0; j < lCoarseNodesPerDir[1]; ++j) {
1512 // for(LO i = 0; i < lCoarseNodesPerDir[0]; ++i) {
1513 // col = k*lCoarseNodesPerDir[1]*lCoarseNodesPerDir[0] + j*lCoarseNodesPerDir[0] + i;
1514
1515 // // Check for endRate
1516 // currentCoarseNode = 0;
1517 // if(firstCoarseNodeInds[0] + i*coarseRate[0] > lFineNodesPerDir[0] - 1) {
1518 // currentCoarseNode += firstCoarseNodeInds[0] + (i-1)*coarseRate[0] + endRate[0];
1519 // } else {
1520 // currentCoarseNode += firstCoarseNodeInds[0] + i*coarseRate[0];
1521 // }
1522 // if(firstCoarseNodeInds[1] + j*coarseRate[1] > lFineNodesPerDir[1] - 1) {
1523 // currentCoarseNode += (firstCoarseNodeInds[1] + (j-1)*coarseRate[1] + endRate[1])*lFineNodesPerDir[0];
1524 // } else {
1525 // currentCoarseNode += (firstCoarseNodeInds[1] + j*coarseRate[1])*lFineNodesPerDir[0];
1526 // }
1527 // if(firstCoarseNodeInds[2] + k*coarseRate[2] > lFineNodesPerDir[2] - 1) {
1528 // currentCoarseNode += (firstCoarseNodeInds[2] + (k-1)*coarseRate[2] + endRate[2])*lFineNodesPerDir[1]*lFineNodesPerDir[0];
1529 // } else {
1530 // currentCoarseNode += (firstCoarseNodeInds[2] + k*coarseRate[2])*lFineNodesPerDir[1]*lFineNodesPerDir[0];
1531 // }
1532 // // Load coordinates
1533 // for(LO dim = 0; dim < numDimensions; ++dim) {
1534 // coarseNodes[dim][col] = fineNodes[dim][currentCoarseNode];
1535 // }
1536
1537 // if((endRate[2] != coarseRate[2]) && (gCoarseNodesGIDs[col] > (gCoarseNodesPerDir[2] - 2)*fineNodesPerCoarseSlab + fineNodesEndCoarseSlab - 1)) {
1538 // tmpInds[2] = gCoarseNodesGIDs[col] / fineNodesPerCoarseSlab + 1;
1539 // tmpVars[0] = gCoarseNodesGIDs[col] - (tmpInds[2] - 1)*fineNodesPerCoarseSlab - fineNodesEndCoarseSlab;
1540 // } else {
1541 // tmpInds[2] = gCoarseNodesGIDs[col] / fineNodesPerCoarseSlab;
1542 // tmpVars[0] = gCoarseNodesGIDs[col] % fineNodesPerCoarseSlab;
1543 // }
1544 // if((endRate[1] != coarseRate[1]) && (tmpVars[0] > (gCoarseNodesPerDir[1] - 2)*fineNodesPerCoarsePlane + fineNodesEndCoarsePlane - 1)) {
1545 // tmpInds[1] = tmpVars[0] / fineNodesPerCoarsePlane + 1;
1546 // tmpVars[1] = tmpVars[0] - (tmpInds[1] - 1)*fineNodesPerCoarsePlane - fineNodesEndCoarsePlane;
1547 // } else {
1548 // tmpInds[1] = tmpVars[0] / fineNodesPerCoarsePlane;
1549 // tmpVars[1] = tmpVars[0] % fineNodesPerCoarsePlane;
1550 // }
1551 // if(tmpVars[1] == gFineNodesPerDir[0] - 1) {
1552 // tmpInds[0] = gCoarseNodesPerDir[0] - 1;
1553 // } else {
1554 // tmpInds[0] = tmpVars[1] / coarseRate[0];
1555 // }
1556 // gInd[2] = col / (lCoarseNodesPerDir[1]*lCoarseNodesPerDir[0]);
1557 // tmp = col % (lCoarseNodesPerDir[1]*lCoarseNodesPerDir[0]);
1558 // gInd[1] = tmp / lCoarseNodesPerDir[0];
1559 // gInd[0] = tmp % lCoarseNodesPerDir[0];
1560 // lCol = gInd[2]*(lCoarseNodesPerDir[1]*lCoarseNodesPerDir[0]) + gInd[1]*lCoarseNodesPerDir[0] + gInd[0];
1561 // gCoarseNodeOnCoarseGridGID = tmpInds[2]*coarseNodesPerCoarseLayer + tmpInds[1]*gCoarseNodesPerDir[0] + tmpInds[0];
1562 // coarseNodesGIDs[lCol] = gCoarseNodeOnCoarseGridGID;
1563 // for(LO dof = 0; dof < BlkSize; ++dof) {
1564 // colGIDs[BlkSize*lCol + dof] = BlkSize*gCoarseNodeOnCoarseGridGID + dof;
1565 // }
1566 // }
1567 // }
1568 // }
1569 // // Now loop over the ghost nodes of the partition to add them to colGIDs
1570 // // since they will need to be included in the column map of P
1571 // for(col = lNumCoarseNodes; col < lNumCoarseNodes + lNumGhostNodes; ++col) {
1572 // if((endRate[2] != coarseRate[2]) && (ghostGIDs[ghostPermut[col - lNumCoarseNodes]] > (gCoarseNodesPerDir[2] - 2)*fineNodesPerCoarseSlab + fineNodesEndCoarseSlab - 1)) {
1573 // tmpInds[2] = ghostGIDs[ghostPermut[col - lNumCoarseNodes]] / fineNodesPerCoarseSlab + 1;
1574 // tmpVars[0] = ghostGIDs[ghostPermut[col - lNumCoarseNodes]] - (tmpInds[2] - 1)*fineNodesPerCoarseSlab - fineNodesEndCoarseSlab;
1575 // } else {
1576 // tmpInds[2] = ghostGIDs[ghostPermut[col - lNumCoarseNodes]] / fineNodesPerCoarseSlab;
1577 // tmpVars[0] = ghostGIDs[ghostPermut[col - lNumCoarseNodes]] % fineNodesPerCoarseSlab;
1578 // }
1579 // if((endRate[1] != coarseRate[1]) && (tmpVars[0] > (gCoarseNodesPerDir[1] - 2)*fineNodesPerCoarsePlane + fineNodesEndCoarsePlane - 1)) {
1580 // tmpInds[1] = tmpVars[0] / fineNodesPerCoarsePlane + 1;
1581 // tmpVars[1] = tmpVars[0] - (tmpInds[1] - 1)*fineNodesPerCoarsePlane - fineNodesEndCoarsePlane;
1582 // } else {
1583 // tmpInds[1] = tmpVars[0] / fineNodesPerCoarsePlane;
1584 // tmpVars[1] = tmpVars[0] % fineNodesPerCoarsePlane;
1585 // }
1586 // if(tmpVars[1] == gFineNodesPerDir[0] - 1) {
1587 // tmpInds[0] = gCoarseNodesPerDir[0] - 1;
1588 // } else {
1589 // tmpInds[0] = tmpVars[1] / coarseRate[0];
1590 // }
1591 // gCoarseNodeOnCoarseGridGID = tmpInds[2]*coarseNodesPerCoarseLayer + tmpInds[1]*gCoarseNodesPerDir[0] + tmpInds[0];
1592 // for(LO dof = 0; dof < BlkSize; ++dof) {
1593 // colGIDs[BlkSize*col + dof] = BlkSize*gCoarseNodeOnCoarseGridGID + dof;
1594 // }
1595 // }
1596 // } // End of scope for tmpInds, tmpVars and tmp
1597
1598 // } // GetGeometricData()
1599
1600 template <class Scalar, class LocalOrdinal, class GlobalOrdinal, class Node>
1601 void GeneralGeometricPFactory<Scalar, LocalOrdinal, GlobalOrdinal, Node>::ComputeStencil(
1602 const LO numDimensions, const Array<GO> currentNodeIndices,
1603 const Array<GO> coarseNodeIndices, const LO rate[3],
1604 const Array<Array<typename Teuchos::ScalarTraits<Scalar>::coordinateType> > coord, const int interpolationOrder,
1605 std::vector<double>& stencil) const {
1606
1607 TEUCHOS_TEST_FOR_EXCEPTION((interpolationOrder > 1) || (interpolationOrder < 0),
1608 Exceptions::RuntimeError,
1609 "The interpolation order can be set to 0 or 1 only.");
1610
1611 if(interpolationOrder == 0) {
1612 ComputeConstantInterpolationStencil(numDimensions, currentNodeIndices, coarseNodeIndices,
1613 rate, stencil);
1614 } else if(interpolationOrder == 1) {
1615 ComputeLinearInterpolationStencil(numDimensions, coord, stencil);
1616 }
1617
1618 } // End ComputeStencil
1619
1620 template <class Scalar, class LocalOrdinal, class GlobalOrdinal, class Node>
1621 void GeneralGeometricPFactory<Scalar, LocalOrdinal, GlobalOrdinal, Node>::
1622 ComputeConstantInterpolationStencil(const LO numDimensions, const Array<GO> currentNodeIndices,
1623 const Array<GO> coarseNodeIndices, const LO rate[3],
1624 std::vector<double>& stencil) const {
1625
1626 LO coarseNode = 0;
1627 if(numDimensions > 2) {
1628 if((currentNodeIndices[2] - coarseNodeIndices[2]) > (rate[2] / 2)) {
1629 coarseNode += 4;
1630 }
1631 }
1632 if(numDimensions > 1) {
1633 if((currentNodeIndices[1] - coarseNodeIndices[1]) > (rate[1] / 2)) {
1634 coarseNode += 2;
1635 }
1636 }
1637 if((currentNodeIndices[0] - coarseNodeIndices[0]) > (rate[0] / 2)) {
1638 coarseNode += 1;
1639 }
1640 stencil[0] = coarseNode;
1641
1642 } // ComputeConstantInterpolationStencil
1643
1644 template <class Scalar, class LocalOrdinal, class GlobalOrdinal, class Node>
1645 void GeneralGeometricPFactory<Scalar, LocalOrdinal, GlobalOrdinal, Node>::
1646 ComputeLinearInterpolationStencil(const LO numDimensions, const Array<Array<typename Teuchos::ScalarTraits<Scalar>::coordinateType> > coord,
1647 std::vector<double>& stencil)
1648 const {
1649
1650 // 7 8 Find xi, eta and zeta such that
1651 // x---------x
1652 // /| /| Rx = x_p - sum N_i(xi,eta,zeta)x_i = 0
1653 // 5/ | 6/ | Ry = y_p - sum N_i(xi,eta,zeta)y_i = 0
1654 // x---------x | Rz = z_p - sum N_i(xi,eta,zeta)z_i = 0
1655 // | | *P | |
1656 // | x------|--x We can do this with a Newton solver:
1657 // | /3 | /4 We will start with initial guess (xi,eta,zeta) = (0,0,0)
1658 // |/ |/ We compute the Jacobian and iterate until convergence...
1659 // z y x---------x
1660 // | / 1 2 Once we have (xi,eta,zeta), we can evaluate all N_i which
1661 // |/ give us the weights for the interpolation stencil!
1662 // o---x
1663 //
1664
1665 Teuchos::SerialDenseMatrix<LO,double> Jacobian(numDimensions, numDimensions);
1666 Teuchos::SerialDenseVector<LO,double> residual(numDimensions);
1667 Teuchos::SerialDenseVector<LO,double> solutionDirection(numDimensions);
1668 Teuchos::SerialDenseVector<LO,double> paramCoords(numDimensions);
1669 Teuchos::SerialDenseSolver<LO,double> problem;
1670 LO numTerms = std::pow(2,numDimensions), iter = 0, max_iter = 5;
1671 double functions[4][8], norm_ref = 1, norm2 = 1, tol = 1e-5;
1672 paramCoords.size(numDimensions);
1673
1674 while( (iter < max_iter) && (norm2 > tol*norm_ref) ) {
1675 ++iter;
1676 norm2 = 0;
1677 solutionDirection.size(numDimensions);
1678 residual.size(numDimensions);
1679 Jacobian = 0.0;
1680
1681 // Compute Jacobian and Residual
1682 GetInterpolationFunctions(numDimensions, paramCoords, functions);
1683 for(LO i = 0; i < numDimensions; ++i) {
1684 residual(i) = coord[0][i]; // Add coordinates from point of interest
1685 for(LO k = 0; k < numTerms; ++k) {
1686 residual(i) -= functions[0][k]*coord[k+1][i]; //Remove contribution from all coarse points
1687 }
1688 if(iter == 1) {
1689 norm_ref += residual(i)*residual(i);
1690 if(i == numDimensions - 1) {
1691 norm_ref = std::sqrt(norm_ref);
1692 }
1693 }
1694
1695 for(LO j = 0; j < numDimensions; ++j) {
1696 for(LO k = 0; k < numTerms; ++k) {
1697 Jacobian(i,j) += functions[j+1][k]*coord[k+1][i];
1698 }
1699 }
1700 }
1701
1702 // Set Jacobian, Vectors and solve problem
1703 problem.setMatrix(Teuchos::rcp(&Jacobian, false));
1704 problem.setVectors(Teuchos::rcp(&solutionDirection, false), Teuchos::rcp(&residual, false));
1705 problem.factorWithEquilibration(true);
1706 problem.solve();
1707 problem.unequilibrateLHS();
1708
1709 for(LO i = 0; i < numDimensions; ++i) {
1710 paramCoords(i) = paramCoords(i) + solutionDirection(i);
1711 }
1712
1713 // Recompute Residual norm
1714 GetInterpolationFunctions(numDimensions, paramCoords, functions);
1715 for(LO i = 0; i < numDimensions; ++i) {
1716 double tmp = coord[0][i];
1717 for(LO k = 0; k < numTerms; ++k) {
1718 tmp -= functions[0][k]*coord[k+1][i];
1719 }
1720 norm2 += tmp*tmp;
1721 tmp = 0;
1722 }
1723 norm2 = std::sqrt(norm2);
1724 }
1725
1726 // Load the interpolation values onto the stencil.
1727 for(LO i = 0; i < 8; ++i) {
1728 stencil[i] = functions[0][i];
1729 }
1730
1731 } // End ComputeLinearInterpolationStencil
1732
1733 template <class Scalar, class LocalOrdinal, class GlobalOrdinal, class Node>
1734 void GeneralGeometricPFactory<Scalar, LocalOrdinal, GlobalOrdinal, Node>::
1735 GetInterpolationFunctions(const LO numDimensions,
1736 const Teuchos::SerialDenseVector<LO,double> parameters,
1737 double functions[4][8]) const {
1738 double xi = 0.0, eta = 0.0, zeta = 0.0, denominator = 0.0;
1739 if(numDimensions == 1) {
1740 xi = parameters[0];
1741 denominator = 2.0;
1742 } else if(numDimensions == 2) {
1743 xi = parameters[0];
1744 eta = parameters[1];
1745 denominator = 4.0;
1746 } else if(numDimensions == 3) {
1747 xi = parameters[0];
1748 eta = parameters[1];
1749 zeta = parameters[2];
1750 denominator = 8.0;
1751 }
1752
1753 functions[0][0] = (1.0 - xi)*(1.0 - eta)*(1.0 - zeta) / denominator;
1754 functions[0][1] = (1.0 + xi)*(1.0 - eta)*(1.0 - zeta) / denominator;
1755 functions[0][2] = (1.0 - xi)*(1.0 + eta)*(1.0 - zeta) / denominator;
1756 functions[0][3] = (1.0 + xi)*(1.0 + eta)*(1.0 - zeta) / denominator;
1757 functions[0][4] = (1.0 - xi)*(1.0 - eta)*(1.0 + zeta) / denominator;
1758 functions[0][5] = (1.0 + xi)*(1.0 - eta)*(1.0 + zeta) / denominator;
1759 functions[0][6] = (1.0 - xi)*(1.0 + eta)*(1.0 + zeta) / denominator;
1760 functions[0][7] = (1.0 + xi)*(1.0 + eta)*(1.0 + zeta) / denominator;
1761
1762 functions[1][0] = -(1.0 - eta)*(1.0 - zeta) / denominator;
1763 functions[1][1] = (1.0 - eta)*(1.0 - zeta) / denominator;
1764 functions[1][2] = -(1.0 + eta)*(1.0 - zeta) / denominator;
1765 functions[1][3] = (1.0 + eta)*(1.0 - zeta) / denominator;
1766 functions[1][4] = -(1.0 - eta)*(1.0 + zeta) / denominator;
1767 functions[1][5] = (1.0 - eta)*(1.0 + zeta) / denominator;
1768 functions[1][6] = -(1.0 + eta)*(1.0 + zeta) / denominator;
1769 functions[1][7] = (1.0 + eta)*(1.0 + zeta) / denominator;
1770
1771 functions[2][0] = -(1.0 - xi)*(1.0 - zeta) / denominator;
1772 functions[2][1] = -(1.0 + xi)*(1.0 - zeta) / denominator;
1773 functions[2][2] = (1.0 - xi)*(1.0 - zeta) / denominator;
1774 functions[2][3] = (1.0 + xi)*(1.0 - zeta) / denominator;
1775 functions[2][4] = -(1.0 - xi)*(1.0 + zeta) / denominator;
1776 functions[2][5] = -(1.0 + xi)*(1.0 + zeta) / denominator;
1777 functions[2][6] = (1.0 - xi)*(1.0 + zeta) / denominator;
1778 functions[2][7] = (1.0 + xi)*(1.0 + zeta) / denominator;
1779
1780 functions[3][0] = -(1.0 - xi)*(1.0 - eta) / denominator;
1781 functions[3][1] = -(1.0 + xi)*(1.0 - eta) / denominator;
1782 functions[3][2] = -(1.0 - xi)*(1.0 + eta) / denominator;
1783 functions[3][3] = -(1.0 + xi)*(1.0 + eta) / denominator;
1784 functions[3][4] = (1.0 - xi)*(1.0 - eta) / denominator;
1785 functions[3][5] = (1.0 + xi)*(1.0 - eta) / denominator;
1786 functions[3][6] = (1.0 - xi)*(1.0 + eta) / denominator;
1787 functions[3][7] = (1.0 + xi)*(1.0 + eta) / denominator;
1788
1789 } // End GetInterpolationFunctions
1790
1791 template <class Scalar, class LocalOrdinal, class GlobalOrdinal, class Node>
1792 void GeneralGeometricPFactory<Scalar, LocalOrdinal, GlobalOrdinal, Node>::sh_sort_permute(
1793 const typename Teuchos::Array<LocalOrdinal>::iterator& first1,
1794 const typename Teuchos::Array<LocalOrdinal>::iterator& last1,
1795 const typename Teuchos::Array<LocalOrdinal>::iterator& first2,
1796 const typename Teuchos::Array<LocalOrdinal>::iterator& /* last2 */) const
1797 {
1798 typedef typename std::iterator_traits<typename Teuchos::Array<LocalOrdinal>::iterator>::difference_type DT;
1799 DT n = last1 - first1;
1800 DT m = n / 2;
1801 DT z = Teuchos::OrdinalTraits<DT>::zero();
1802 while (m > z)
1803 {
1804 DT max = n - m;
1805 for (DT j = 0; j < max; j++)
1806 {
1807 for (DT k = j; k >= 0; k-=m)
1808 {
1809 if (first1[first2[k+m]] >= first1[first2[k]])
1810 break;
1811 std::swap(first2[k+m], first2[k]);
1812 }
1813 }
1814 m = m/2;
1815 }
1816 } // End sh_sort_permute
1817
1818 template <class Scalar, class LocalOrdinal, class GlobalOrdinal, class Node>
1819 void GeneralGeometricPFactory<Scalar, LocalOrdinal, GlobalOrdinal, Node>::sh_sort2(
1820 const typename Teuchos::Array<LocalOrdinal>::iterator& first1,
1821 const typename Teuchos::Array<LocalOrdinal>::iterator& last1,
1822 const typename Teuchos::Array<LocalOrdinal>::iterator& first2,
1823 const typename Teuchos::Array<LocalOrdinal>::iterator& /* last2 */) const
1824 {
1825 typedef typename std::iterator_traits<typename Teuchos::Array<LocalOrdinal>::iterator>::difference_type DT;
1826 DT n = last1 - first1;
1827 DT m = n / 2;
1828 DT z = Teuchos::OrdinalTraits<DT>::zero();
1829 while (m > z)
1830 {
1831 DT max = n - m;
1832 for (DT j = 0; j < max; j++)
1833 {
1834 for (DT k = j; k >= 0; k-=m)
1835 {
1836 if (first1[k+m] >= first1[k])
1837 break;
1838 std::swap(first1[k+m], first1[k]);
1839 std::swap(first2[k+m], first2[k]);
1840 }
1841 }
1842 m = m/2;
1843 }
1844 } // End sh_sort2
1845
1846 template <class Scalar, class LocalOrdinal, class GlobalOrdinal, class Node>
1847 void GeneralGeometricPFactory<Scalar, LocalOrdinal, GlobalOrdinal, Node>::
1848 GetGIDLocalLexicographic(const GO i, const GO j, const GO k,
1849 const Array<LO> coarseNodeFineIndices, const RCP<GeometricData> myGeo,
1850 const LO myRankIndex, const LO pi, const LO pj, const LO /* pk */,
1851 const typename std::vector<std::vector<GO> >::iterator blockStart,
1852 const typename std::vector<std::vector<GO> >::iterator blockEnd,
1853 GO& myGID, LO& myPID, LO& myLID) const {
1854
1855 LO ni = -1, nj = -1, li = -1, lj = -1, lk = -1;
1856 LO myRankGuess = myRankIndex;
1857 // We try to make a logical guess as to which PID owns the current coarse node
1858 if(i == 0 && myGeo->ghostInterface[0]) {
1859 --myRankGuess;
1860 } else if((i == myGeo->ghostedCoarseNodesPerDir[0] - 1) && myGeo->ghostInterface[1]) {
1861 ++myRankGuess;
1862 }
1863 if(j == 0 && myGeo->ghostInterface[2]) {
1864 myRankGuess -= pi;
1865 } else if((j == myGeo->ghostedCoarseNodesPerDir[1] - 1) && myGeo->ghostInterface[3]) {
1866 myRankGuess += pi;
1867 }
1868 if(k == 0 && myGeo->ghostInterface[4]) {
1869 myRankGuess -= pj*pi;
1870 } else if((k == myGeo->ghostedCoarseNodesPerDir[2] - 1) && myGeo->ghostInterface[5]) {
1871 myRankGuess += pj*pi;
1872 }
1873 if(coarseNodeFineIndices[0] >= myGeo->meshData[myRankGuess][3]
1874 && coarseNodeFineIndices[0] <= myGeo->meshData[myRankGuess][4]
1875 && coarseNodeFineIndices[1] >= myGeo->meshData[myRankGuess][5]
1876 && coarseNodeFineIndices[1] <= myGeo->meshData[myRankGuess][6]
1877 && coarseNodeFineIndices[2] >= myGeo->meshData[myRankGuess][7]
1878 && coarseNodeFineIndices[2] <= myGeo->meshData[myRankGuess][8]) {
1879 myPID = myGeo->meshData[myRankGuess][0];
1880 ni = myGeo->meshData[myRankGuess][4] - myGeo->meshData[myRankGuess][3] + 1;
1881 nj = myGeo->meshData[myRankGuess][6] - myGeo->meshData[myRankGuess][5] + 1;
1882 li = coarseNodeFineIndices[0] - myGeo->meshData[myRankGuess][3];
1883 lj = coarseNodeFineIndices[1] - myGeo->meshData[myRankGuess][5];
1884 lk = coarseNodeFineIndices[2] - myGeo->meshData[myRankGuess][7];
1885 myLID = lk*nj*ni + lj*ni + li;
1886 myGID = myGeo->meshData[myRankGuess][9] + myLID;
1887 } else { // The guess failed, let us use the heavy artilery: std::find_if()
1888 // It could be interesting to monitor how many times this branch of the code gets
1889 // used as it is far more expensive than the above one...
1890 auto nodeRank = std::find_if(blockStart, blockEnd,
1891 [coarseNodeFineIndices](const std::vector<GO>& vec){
1892 if(coarseNodeFineIndices[0] >= vec[3]
1893 && coarseNodeFineIndices[0] <= vec[4]
1894 && coarseNodeFineIndices[1] >= vec[5]
1895 && coarseNodeFineIndices[1] <= vec[6]
1896 && coarseNodeFineIndices[2] >= vec[7]
1897 && coarseNodeFineIndices[2] <= vec[8]) {
1898 return true;
1899 } else {
1900 return false;
1901 }
1902 });
1903 myPID = (*nodeRank)[0];
1904 ni = (*nodeRank)[4] - (*nodeRank)[3] + 1;
1905 nj = (*nodeRank)[6] - (*nodeRank)[5] + 1;
1906 li = coarseNodeFineIndices[0] - (*nodeRank)[3];
1907 lj = coarseNodeFineIndices[1] - (*nodeRank)[5];
1908 lk = coarseNodeFineIndices[2] - (*nodeRank)[7];
1909 myLID = lk*nj*ni + lj*ni + li;
1910 myGID = (*nodeRank)[9] + myLID;
1911 }
1912 } // End GetGIDLocalLexicographic
1913
1914} //namespace MueLu
1915
1916#define MUELU_GENERALGEOMETRICPFACTORY_SHORT
1917#endif // MUELU_GENERALGEOMETRICPFACTORY_DEF_HPP
MueLu_GeneralGeometricPFactory_decl.hpp
MueLu_Monitor.hpp
Node
MueLu::DefaultNode Node
Definition MueLu_UseDefaultTypes.hpp:52
MueLu::Exceptions::RuntimeError
Exception throws to report errors in the internal logical of the program.
Definition MueLu_Exceptions.hpp:70
MueLu::FactoryMonitor
Timer to be used in factories. Similar to Monitor but with additional timers.
Definition MueLu_Monitor.hpp:202
MueLu::Factory::Input
void Input(Level &level, const std::string &varName) const
Definition MueLu_Factory.hpp:146
MueLu::Factory::Get
T Get(Level &level, const std::string &varName) const
Definition MueLu_Factory.hpp:155
MueLu::Factory::Set
void Set(Level &level, const std::string &varName, const T &data) const
Definition MueLu_Factory.hpp:165
MueLu::GeneralGeometricPFactory::MakeGeneralGeometricP
void MakeGeneralGeometricP(RCP< GeometricData > myGeo, const RCP< Xpetra::MultiVector< typename Teuchos::ScalarTraits< Scalar >::coordinateType, LO, GO, NO > > &fCoords, const LO nnzP, const LO dofsPerNode, RCP< const Map > &stridedDomainMapP, RCP< Matrix > &Amat, RCP< Matrix > &P, RCP< Xpetra::MultiVector< typename Teuchos::ScalarTraits< Scalar >::coordinateType, LO, GO, NO > > &cCoords, RCP< NodesIDs > ghostedCoarseNodes, Array< Array< GO > > coarseNodesGIDs, int interpolationOrder) const
Definition MueLu_GeneralGeometricPFactory_def.hpp:776
MueLu::GeneralGeometricPFactory::GetValidParameterList
RCP< const ParameterList > GetValidParameterList() const
Return a const parameter list of valid parameters that setParameterList() will accept.
Definition MueLu_GeneralGeometricPFactory_def.hpp:74
MueLu::GeneralGeometricPFactory::DeclareInput
void DeclareInput(Level &fineLevel, Level &coarseLevel) const
Input.
Definition MueLu_GeneralGeometricPFactory_def.hpp:106
MueLu::GeneralGeometricPFactory::Build
void Build(Level &fineLevel, Level &coarseLevel) const
Build an object with this factory.
Definition MueLu_GeneralGeometricPFactory_def.hpp:138
MueLu::GeneralGeometricPFactory::sh_sort_permute
void sh_sort_permute(const typename Teuchos::Array< LocalOrdinal >::iterator &first1, const typename Teuchos::Array< LocalOrdinal >::iterator &last1, const typename Teuchos::Array< LocalOrdinal >::iterator &first2, const typename Teuchos::Array< LocalOrdinal >::iterator &last2) const
Definition MueLu_GeneralGeometricPFactory_def.hpp:1792
MueLu::GeneralGeometricPFactory::ComputeLinearInterpolationStencil
void ComputeLinearInterpolationStencil(const LO numDimension, const Array< Array< typename Teuchos::ScalarTraits< Scalar >::coordinateType > > coord, std::vector< double > &stencil) const
Definition MueLu_GeneralGeometricPFactory_def.hpp:1646
MueLu::GeneralGeometricPFactory::BuildP
void BuildP(Level &fineLevel, Level &coarseLevel) const
Abstract Build method.
Definition MueLu_GeneralGeometricPFactory_def.hpp:144
MueLu::GeneralGeometricPFactory::ComputeStencil
void ComputeStencil(const LO numDimension, const Array< GO > currentNodeIndices, const Array< GO > coarseNodeIndices, const LO rate[3], const Array< Array< typename Teuchos::ScalarTraits< Scalar >::coordinateType > > coord, const int interpolationOrder, std::vector< double > &stencil) const
Definition MueLu_GeneralGeometricPFactory_def.hpp:1601
MueLu::GeneralGeometricPFactory::GetCoarsePoints
void GetCoarsePoints(const int interpolationOrder, const LO blkSize, RCP< const Map > fineCoordsMap, RCP< GeometricData > myGeometry, RCP< NodesIDs > ghostedCoarseNodes, Array< Array< GO > > &lCoarseNodesGIDs) const
Definition MueLu_GeneralGeometricPFactory_def.hpp:565
MueLu::GeneralGeometricPFactory::GetInterpolationFunctions
void GetInterpolationFunctions(const LO numDimension, const Teuchos::SerialDenseVector< LO, double > parameters, double functions[4][8]) const
Definition MueLu_GeneralGeometricPFactory_def.hpp:1735
MueLu::GeneralGeometricPFactory::MeshLayoutInterface
void MeshLayoutInterface(const int interpolationOrder, const LO blkSize, RCP< const Map > fineCoordsMap, RCP< GeometricData > myGeometry, RCP< NodesIDs > ghostedCoarseNodes, Array< Array< GO > > &lCoarseNodesGIDs) const
Definition MueLu_GeneralGeometricPFactory_def.hpp:330
MueLu::GeneralGeometricPFactory::sh_sort2
void sh_sort2(const typename Teuchos::Array< LocalOrdinal >::iterator &first1, const typename Teuchos::Array< LocalOrdinal >::iterator &last1, const typename Teuchos::Array< LocalOrdinal >::iterator &first2, const typename Teuchos::Array< LocalOrdinal >::iterator &last2) const
Definition MueLu_GeneralGeometricPFactory_def.hpp:1819
MueLu::GeneralGeometricPFactory::ComputeConstantInterpolationStencil
void ComputeConstantInterpolationStencil(const LO numDimension, const Array< GO > currentNodeIndices, const Array< GO > coarseNodeIndices, const LO rate[3], std::vector< double > &stencil) const
Definition MueLu_GeneralGeometricPFactory_def.hpp:1622
MueLu::Level
Class that holds all level-specific information.
Definition MueLu_Level.hpp:99
MueLu::Level::IsAvailable
bool IsAvailable(const std::string &ename, const FactoryBase *factory=NoFactory::get()) const
Test whether a need's value has been saved.
Definition MueLu_Level.hpp:354
MueLu::Level::DeclareInput
void DeclareInput(const std::string &ename, const FactoryBase *factory, const FactoryBase *requestedBy=NoFactory::get())
Callback from FactoryBase::CallDeclareInput() and FactoryBase::DeclareInput().
Definition MueLu_Level.cpp:150
MueLu::Level::GetLevelID
int GetLevelID() const
Return level number.
Definition MueLu_Level.cpp:76
MueLu::Level::Get
T & Get(const std::string &ename, const FactoryBase *factory=NoFactory::get())
Get data without decrementing associated storage counter (i.e., read-only access)....
Definition MueLu_Level.hpp:191
MueLu::NoFactory::get
static const NoFactory * get()
Definition MueLu_NoFactory.cpp:59
MueLu::ParameterListAcceptorImpl::GetParameterList
virtual const Teuchos::ParameterList & GetParameterList() const
Definition MueLu_ParameterListAcceptor.cpp:41
MueLu::VerboseObject::GetOStream
Teuchos::FancyOStream & GetOStream(MsgType type, int thisProcRankOnly=0) const
Get an output stream for outputting the input message type.
Definition MueLu_VerboseObject.cpp:107
MueLu
Namespace for MueLu classes and methods.
Definition MueLu_BrickAggregationFactory_decl.hpp:78
MueLu::Runtime1
@ Runtime1
Description of what is happening (more verbose).
Definition MueLu_VerbosityLevel.hpp:63