FFmmAlgorithmThreadProc.hpp 76.6 KB
Newer Older
1
// See LICENCE file at project root
2 3
#ifndef FFMMALGORITHMTHREADPROC_HPP
#define FFMMALGORITHMTHREADPROC_HPP
4

5 6
#define SCALFMM_DISTRIBUTED_ALGORITHM

COULAUD Olivier's avatar
COULAUD Olivier committed
7
#include <omp.h>
8

COULAUD Olivier's avatar
COULAUD Olivier committed
9
//
10
#include "../Utils/FAssert.hpp"
11
#include "../Utils/FLog.hpp"
12

13
#include "../Utils/FTic.hpp"
14
#include "Utils/FAlgorithmTimers.hpp"
15

16 17
#include "../Utils/FGlobal.hpp"

18
#include "../Containers/FBoolArray.hpp"
19
#include "../Containers/FOctree.hpp"
berenger-bramas's avatar
berenger-bramas committed
20
#include "../Containers/FLightOctree.hpp"
21
#include "../Utils/FEnv.hpp"
22

23 24
#include "../Containers/FBufferWriter.hpp"
#include "../Containers/FBufferReader.hpp"
25 26
#include "../Containers/FMpiBufferWriter.hpp"
#include "../Containers/FMpiBufferReader.hpp"
27

28
#include "../Utils/FMpi.hpp"
29
#include <sys/time.h>
30

31
#include "FCoreCommon.hpp"
32
#include "FP2PExclusion.hpp"
33

34
#include <memory>
35
#include <vector>
36

37
/**
38
 * @author Berenger Bramas (berenger.bramas@inria.fr)
39
 *
40 41
 * Please read the license
 *
42 43 44 45
 * This class is a threaded FMM algorithm distributed using MPI. It iterates on
 * a tree and call the kernels with good arguments. It uses the inspector -
 * executor model : iterates on the tree and builds an array to work in parallel
 * on this array
46
 *
47
 * This class does not free pointers given in arguements.
48 49
 *
 * Threaded & based on the inspector-executor model
50 51 52 53 54 55
 *
 *     schedule(runtime) export OMP_NUM_THREADS=2
 *     export OMPI_CXX=`which g++-4.4`
 *     mpirun -np 2 valgrind --suppressions=/usr/share/openmpi/openmpi-valgrind.supp
 *        --tool=memcheck --leak-check=yes --show-reachable=yes --num-callers=20
 *        --track-fds=yes ./Tests/testFmmAlgorithmProc ../Data/testLoaderSmall.fma.tmp
56
 */
57
template<class OctreeClass, class CellClass, class ContainerClass, class KernelClass, class LeafClass, class P2PExclusionClass = FP2PMiddleExclusion>
58
class FFmmAlgorithmThreadProc : public FAbstractAlgorithm, public FAlgorithmTimers {
59 60 61
private:
    OctreeClass* const tree;     ///< The octree to work on
    KernelClass** kernels;       ///< The kernels
62

63 64
    const FMpi::FComm comm;      ///< MPI comm
    FMpi::FComm fcomCompute;
65

66
    /// Used to store pointers to cells/leafs to work with
67
    typename OctreeClass::Iterator* iterArray;
68 69
    /// Used to store pointers to cells/leafs to send/rcv
    typename OctreeClass::Iterator* iterArrayComm;
berenger-bramas's avatar
berenger-bramas committed
70

71 72
    int numberOfLeafs;           ///< To store the size at the previous level
    const int MaxThreads;        ///< Max number of thread allowed by openmp
73 74 75 76
    const int nbProcessOrig;         ///< Process count
    const int idProcessOrig;         ///< Current process id
    int nbProcess;        ///< Process count
    int idProcess;        ///< Current process id
77
    const int OctreeHeight;      ///< Tree height
78

79
    const int userChunkSize;
80
    const int leafLevelSeparationCriteria;
81

82
    /** An interval is the morton index interval
83
     * that a proc uses (i.e. it holds data in this interval) */
84
    struct Interval{
85 86
        MortonIndex leftIndex;
        MortonIndex rightIndex;
87
    };
88 89

    /// Current process interval
90
    Interval*const intervals;
91
    /// All processes intervals
92
    Interval*const workingIntervalsPerLevel;
93

94
    /// Get an interval from a process id and tree level
95
    Interval& getWorkingInterval( int level,  int proc){
96
        return workingIntervalsPerLevel[OctreeHeight * proc + level];
97
    }
98

99
    /// Get an interval from a process id and tree level
100
    const Interval& getWorkingInterval( int level,  int proc) const {
101
        return workingIntervalsPerLevel[OctreeHeight * proc + level];
102 103
    }

104 105
    /// Does \a procIdx have work at given \a idxLevel
    /** i.e. does it hold cells and is responsible of them ? */
106
    bool procHasWorkAtLevel(const int idxLevel , const int idxProc) const {
107
        return getWorkingInterval(idxLevel, idxProc).leftIndex <= getWorkingInterval(idxLevel, idxProc).rightIndex;
108 109
    }

110
    /** True if the \a idxProc left cell at \a idxLevel+1 has the same parent as us for our right cell */
111
    bool procCoversMyRightBorderCell(const int idxLevel , const int idxProc) const {
112
        return (getWorkingInterval((idxLevel+1) , idProcess).rightIndex>>3) == (getWorkingInterval((idxLevel+1) ,idxProc).leftIndex >>3);
113 114
    }

115
    /** True if the idxProc right cell at idxLevel+1 has the same parent as us for our left cell */
116
    bool procCoversMyLeftBorderCell(const int idxLevel , const int idxProc) const {
117
        return (getWorkingInterval((idxLevel+1) , idxProc).rightIndex >>3) == (getWorkingInterval((idxLevel+1) , idProcess).leftIndex>>3);
118 119
    }

120
public:
121
    /// Get current process interval at given \a level
122
    Interval& getWorkingInterval( int level){
123
        return getWorkingInterval(level, idProcess);
124 125
    }

126
    /// Does the current process has some work at this level ?
127
    bool hasWorkAtLevel( int level){
128
        return idProcess == 0 || (getWorkingInterval(level, idProcess - 1).rightIndex) < (getWorkingInterval(level, idProcess).rightIndex);
129 130
    }

131
    /**@brief Constructor
132 133
     * @param inTree the octree to work on
     * @param inKernels the kernels to call
134
     *
135 136
     * An assert is launched if one of the arguments is null
     */
137 138 139
    FFmmAlgorithmThreadProc(const FMpi::FComm& inComm, OctreeClass* const inTree,
                            KernelClass* const inKernels,
                            const int inUserChunkSize = 10, const int inLeafLevelSeperationCriteria = 1) :
140 141 142
        tree(inTree),
        kernels(nullptr),
        comm(inComm),
143
        fcomCompute(inComm),
144 145 146
        iterArray(nullptr),
        iterArrayComm(nullptr),
        numberOfLeafs(0),
147
        MaxThreads(FEnv::GetValue("SCALFMM_ALGO_NUM_THREADS",omp_get_max_threads())),
148 149 150 151
        nbProcessOrig(inComm.processCount()),
        idProcessOrig(inComm.processId()),
        nbProcess(0),
        idProcess(0),
152
        OctreeHeight(tree->getHeight()),
153
        userChunkSize(inUserChunkSize),
154
        leafLevelSeparationCriteria(inLeafLevelSeperationCriteria),
155 156
        intervals(new Interval[inComm.processCount()]),
        workingIntervalsPerLevel(new Interval[inComm.processCount() * tree->getHeight()]) {
157
        FAssertLF(tree, "tree cannot be null");
158
        FAssertLF(leafLevelSeparationCriteria < 3, "Separation criteria should be < 3");
159

160
        this->kernels = new KernelClass*[MaxThreads];
161
#pragma omp parallel num_threads(MaxThreads)
162
        {
163
#pragma omp critical (InitFFmmAlgorithmThreadProc)
164
            {
165
                this->kernels[omp_get_thread_num()] = new KernelClass(*inKernels);
166
            }
167
        }
168

169 170
        FAbstractAlgorithm::setNbLevelsInTree(tree->getHeight());

171
        FLOG(FLog::Controller << "FFmmAlgorithmThreadProc\n");
172
        FLOG(FLog::Controller << "Max threads = "  << MaxThreads << ", Procs = " << nbProcessOrig << ", I am " << idProcessOrig << ".\n");
173
        FLOG(FLog::Controller << "Chunck Size = " << userChunkSize << "\n");
174
    }
175 176

    /// Default destructor
177
    virtual ~FFmmAlgorithmThreadProc(){
178 179 180 181
        for(int idxThread = 0 ; idxThread < MaxThreads ; ++idxThread){
            delete this->kernels[idxThread];
        }
        delete [] this->kernels;
182

183 184
        delete [] intervals;
        delete [] workingIntervalsPerLevel;
185 186
    }

187
protected:
188 189 190 191
    /**
     * To execute the fmm algorithm
     * Call this function to run the complete algorithm
     */
192
    void executeCore(const unsigned operationsToProceed) override {
193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211
        // We are not involve if the tree is empty
        const int iHaveParticles = (!tree->isEmpty());

        std::unique_ptr<int[]> hasParticles(new int[comm.processCount()]);
        FMpi::Assert( MPI_Allgather(const_cast<int*>(&iHaveParticles), 1,MPI_INT,
                                    hasParticles.get(), 1, MPI_INT,
                                    comm.getComm()), __LINE__);

        fcomCompute = FMpi::FComm(comm);
        fcomCompute.groupReduce(hasParticles.get());

        if(iHaveParticles){

            nbProcess = fcomCompute.processCount();
            idProcess = fcomCompute.processId();

            FLOG(FLog::Controller << "Max threads = "  << MaxThreads << ", Procs = " << nbProcess << ", I am " << idProcess << ".\n");

            // Count leaf
212
#ifdef SCALFMM_TRACE_ALGO
213
            eztrace_resume();
214
#endif
215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245
            this->numberOfLeafs = 0;
            {
                Interval myFullInterval;
                {//Building the interval with the first and last leaves (and count the number of leaves)
                    typename OctreeClass::Iterator octreeIterator(tree);
                    octreeIterator.gotoBottomLeft();
                    myFullInterval.leftIndex = octreeIterator.getCurrentGlobalIndex();
                    do{
                        ++this->numberOfLeafs;
                    } while(octreeIterator.moveRight());
                    myFullInterval.rightIndex = octreeIterator.getCurrentGlobalIndex();
                }
                // Allocate a number to store the pointer of the cells at a level
                iterArray     = new typename OctreeClass::Iterator[numberOfLeafs];
                iterArrayComm = new typename OctreeClass::Iterator[numberOfLeafs];
                FAssertLF(iterArray,     "iterArray     bad alloc");
                FAssertLF(iterArrayComm, "iterArrayComm bad alloc");

                // We get the leftIndex/rightIndex indexes from each procs
                FMpi::MpiAssert( MPI_Allgather( &myFullInterval, sizeof(Interval), MPI_BYTE, intervals, sizeof(Interval), MPI_BYTE, fcomCompute.getComm()),  __LINE__ );

                // Build my intervals for all levels
                std::unique_ptr<Interval[]> myIntervals(new Interval[OctreeHeight]);
                // At leaf level we know it is the full interval
                myIntervals[OctreeHeight - 1] = myFullInterval;

                // We can estimate the interval for each level by using the parent/child relation
                for(int idxLevel = OctreeHeight - 2 ; idxLevel >= 0 ; --idxLevel){
                    myIntervals[idxLevel].leftIndex = myIntervals[idxLevel+1].leftIndex >> 3;
                    myIntervals[idxLevel].rightIndex = myIntervals[idxLevel+1].rightIndex >> 3;
                }
246

247 248 249 250 251 252 253
                // Process 0 uses the estimates as real intervals, but other processes
                // should remove cells that belong to others
                if(idProcess != 0){
                    //We test for each level if process on left (idProcess-1) own cell I thought I owned
                    typename OctreeClass::Iterator octreeIterator(tree);
                    octreeIterator.gotoBottomLeft();
                    octreeIterator.moveUp();
254

255 256
                    // At h-1 the working limit is the parent of the right cell of the proc on the left
                    MortonIndex workingLimitAtLevel = intervals[idProcess-1].rightIndex >> 3;
257

258 259
                    // We check if we have no more work to do
                    int nullIntervalFromLevel = 0;
260

261 262 263 264 265 266 267 268 269 270 271 272 273
                    for(int idxLevel = OctreeHeight - 2 ; idxLevel >= 1 && nullIntervalFromLevel == 0 ; --idxLevel){
                        while(octreeIterator.getCurrentGlobalIndex() <= workingLimitAtLevel){
                            if( !octreeIterator.moveRight() ){
                                // We cannot move right we are not owner of any more cell
                                nullIntervalFromLevel = idxLevel;
                                break;
                            }
                        }
                        // If we are responsible for some cells at this level keep the first index
                        if(nullIntervalFromLevel == 0){
                            myIntervals[idxLevel].leftIndex = octreeIterator.getCurrentGlobalIndex();
                            octreeIterator.moveUp();
                            workingLimitAtLevel >>= 3;
274 275
                        }
                    }
276 277 278
                    // In case we are not responsible for any cells we put the leftIndex = rightIndex+1
                    for(int idxLevel = nullIntervalFromLevel ; idxLevel >= 1 ; --idxLevel){
                        myIntervals[idxLevel].leftIndex = myIntervals[idxLevel].rightIndex + 1;
279 280 281
                    }
                }

282 283 284 285
                // We get the leftIndex/rightIndex indexes from each procs
                FMpi::MpiAssert( MPI_Allgather( myIntervals.get(), int(sizeof(Interval)) * OctreeHeight, MPI_BYTE,
                                                workingIntervalsPerLevel, int(sizeof(Interval)) * OctreeHeight, MPI_BYTE, fcomCompute.getComm()),  __LINE__ );
            }
COULAUD Olivier's avatar
COULAUD Olivier committed
286
#ifdef SCALFMM_TRACE_ALGO
287
            eztrace_enter_event("P2M", EZTRACE_YELLOW);
COULAUD Olivier's avatar
COULAUD Olivier committed
288
#endif
289 290 291
            Timers[P2MTimer].tic();
            if(operationsToProceed & FFmmP2M) bottomPass();
            Timers[P2MTimer].tac();
292

COULAUD Olivier's avatar
COULAUD Olivier committed
293
#ifdef SSCALFMM_TRACE_ALGO
294 295
            eztrace_leave_event();
            eztrace_enter_event("M2M", EZTRACE_PINK);
COULAUD Olivier's avatar
COULAUD Olivier committed
296 297
#endif

298 299 300
            Timers[M2MTimer].tic();
            if(operationsToProceed & FFmmM2M) upwardPass();
            Timers[M2MTimer].tac();
COULAUD Olivier's avatar
COULAUD Olivier committed
301 302

#ifdef SCALFMM_TRACE_ALGO
303 304
            eztrace_leave_event();
            eztrace_enter_event("M2L", EZTRACE_GREEN);
COULAUD Olivier's avatar
COULAUD Olivier committed
305
#endif
306

307 308 309
            Timers[M2LTimer].tic();
            if(operationsToProceed & FFmmM2L) transferPass();
            Timers[M2LTimer].tac();
310

311 312 313
#ifdef SCALFMM_TRACE_ALGO
            eztrace_leave_event();
            eztrace_enter_event("L2L", EZTRACE_PINK);
COULAUD Olivier's avatar
COULAUD Olivier committed
314 315
#endif

316 317 318
            Timers[L2LTimer].tic();
            if(operationsToProceed & FFmmL2L) downardPass();
            Timers[L2LTimer].tac();
319

COULAUD Olivier's avatar
COULAUD Olivier committed
320
#ifdef SCALFMM_TRACE_ALGO
321 322
            eztrace_leave_event();
            eztrace_enter_event("L2P+P2P", EZTRACE_BLUE);
COULAUD Olivier's avatar
COULAUD Olivier committed
323 324
#endif

325 326 327
            Timers[NearTimer].tic();
            if( (operationsToProceed & FFmmP2P) || (operationsToProceed & FFmmL2P) ) directPass((operationsToProceed & FFmmP2P),(operationsToProceed & FFmmL2P));
            Timers[NearTimer].tac();
328

329
#ifdef SCALFMM_TRACE_ALGO
330 331
            eztrace_leave_event();
            eztrace_stop();
332
#endif
333 334 335 336 337
            // delete array
            delete []     iterArray;
            delete []     iterArrayComm;
            iterArray          = nullptr;
            iterArrayComm = nullptr;
338
#ifdef SCALFMM_TRACE_ALGO
339
            eztrace_stop();
340
#endif
341 342 343 344
        }
        else{
            FLOG( FLog::Controller << "\tProcess = " << comm.processId() << " has zero particles.\n" );
        }
345
    }
346

347 348 349
    /////////////////////////////////////////////////////////////////////////////
    // P2M
    /////////////////////////////////////////////////////////////////////////////
350

351 352 353 354 355
    /**
     * P2M Bottom Pass
     * No communication are involved in the P2M.
     * It is similar to multi threaded version.
     */
356
    void bottomPass(){
357 358 359 360 361 362 363 364 365 366 367 368 369
        FLOG( FLog::Controller.write("\tStart Bottom Pass\n").write(FLog::Flush) );
        FLOG(FTic counterTime);
        FLOG(FTic computationCounter);
        typename OctreeClass::Iterator octreeIterator(tree);

        // Copy the ptr to leaves in array
        octreeIterator.gotoBottomLeft();
        int leafs = 0;
        do{
            iterArray[leafs++] = octreeIterator;
        } while(octreeIterator.moveRight());

        FLOG(computationCounter.tic());
370
#pragma omp parallel num_threads(MaxThreads)
371 372 373 374
        {
            // Each thread get its own kernel
            KernelClass * const myThreadkernels = kernels[omp_get_thread_num()];
            // Parallel iteration on the leaves
375
#pragma omp for nowait schedule(dynamic, userChunkSize)
376 377 378 379 380 381 382
            for(int idxLeafs = 0 ; idxLeafs < leafs ; ++idxLeafs){
                myThreadkernels->P2M( iterArray[idxLeafs].getCurrentCell() , iterArray[idxLeafs].getCurrentListSrc());
            }
        }
        FLOG(computationCounter.tac());
        FLOG( FLog::Controller << "\tFinished (@Bottom Pass (P2M) = "  << counterTime.tacAndElapsed() << " s)\n" );
        FLOG( FLog::Controller << "\t\t Computation : " << computationCounter.elapsed() << " s\n" );
BRAMAS Berenger's avatar
BRAMAS Berenger committed
383
        FLOG( FLog::Controller.flush());
384 385 386 387 388 389 390 391
    }

    /////////////////////////////////////////////////////////////////////////////
    // Upward
    /////////////////////////////////////////////////////////////////////////////

    /** M2M */
    void upwardPass(){
392 393 394 395 396 397 398 399 400 401
        FLOG( FLog::Controller.write("\tStart Upward Pass\n").write(FLog::Flush); );
        FLOG(FTic counterTime);
        FLOG(FTic computationCounter);
        FLOG(FTic singleCounter);
        FLOG(FTic parallelCounter);

        // Start from leal level (height-1)
        typename OctreeClass::Iterator octreeIterator(tree);
        octreeIterator.gotoBottomLeft();
        octreeIterator.moveUp();
402

403
        for(int idxLevel = OctreeHeight - 2 ; idxLevel > FAbstractAlgorithm::lowerWorkingLevel-1 ; --idxLevel){
404 405 406
            octreeIterator.moveUp();
        }

407 408 409 410 411 412 413 414 415 416
        typename OctreeClass::Iterator avoidGotoLeftIterator(octreeIterator);

        // The proc to send the shared cells to
        // Starting to the proc on the left this variable will go to 0
        int currentProcIdToSendTo = (idProcess - 1);

        // There are a maximum of 1 sends and 8-1 receptions
        MPI_Request requests[8];
        MPI_Status status[8];

417 418 419
        MPI_Request requestsSize[8];
        MPI_Status statusSize[8];

420
        FSize bufferSize;
421
        FMpiBufferWriter sendBuffer(1);// Max = 1 + sizeof(cell)*7
422
        std::unique_ptr<FMpiBufferReader[]> recvBuffer(new FMpiBufferReader[7]);
423
        FSize recvBufferSize[7];
424
        CellClass recvBufferCells[7];
425 426 427 428 429 430

        // The first proc that send to me a cell
        // This variable will go to nbProcess
        int firstProcThatSend = idProcess + 1;
        FLOG(computationCounter.tic());

431 432
        // for each levels
        for(int idxLevel = FMath::Min(OctreeHeight - 2, FAbstractAlgorithm::lowerWorkingLevel - 1) ; idxLevel >= FAbstractAlgorithm::upperWorkingLevel ; --idxLevel ){
433 434 435 436 437 438 439 440 441 442 443 444 445 446 447
            // Does my cells are covered by my neighbors working interval and so I have no more work?
            const bool noMoreWorkForMe = (idProcess != 0 && !procHasWorkAtLevel(idxLevel+1, idProcess));
            if(noMoreWorkForMe){
                FAssertLF(procHasWorkAtLevel(idxLevel, idProcess) == false);
                break;
            }

            // Copy and count ALL the cells (even the ones outside the working interval)
            int totalNbCellsAtLevel = 0;
            do{
                iterArray[totalNbCellsAtLevel++] = octreeIterator;
            } while(octreeIterator.moveRight());
            avoidGotoLeftIterator.moveUp();
            octreeIterator = avoidGotoLeftIterator;

448 449
            int iterMpiRequests       = 0; // The iterator for send/recv requests
            int iterMpiRequestsSize   = 0; // The iterator for send/recv requests
450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 472 473

            int nbCellsToSkip     = 0; // The number of cells to send
            // Skip all the cells that are out of my working interval
            while(nbCellsToSkip < totalNbCellsAtLevel && iterArray[nbCellsToSkip].getCurrentGlobalIndex() < getWorkingInterval(idxLevel, idProcess).leftIndex){
                ++nbCellsToSkip;
            }

            // We need to know if we will recv something in order to know if threads skip the last cell
            int nbCellsForThreads = totalNbCellsAtLevel; // totalNbCellsAtLevel or totalNbCellsAtLevel-1
            bool hasToReceive = false;
            if(idProcess != nbProcess-1 && procHasWorkAtLevel(idxLevel , idProcess)){
                // Find the first proc that may send to me
                while(firstProcThatSend < nbProcess && !procHasWorkAtLevel(idxLevel+1, firstProcThatSend) ){
                    firstProcThatSend += 1;
                }
                // Do we have to receive?
                if(firstProcThatSend < nbProcess && procHasWorkAtLevel(idxLevel+1, firstProcThatSend) && procCoversMyRightBorderCell(idxLevel, firstProcThatSend) ){
                    hasToReceive = true;
                    // Threads do not compute the last cell, we will do it once data are received
                    nbCellsForThreads -= 1;
                }
            }

            FLOG(parallelCounter.tic());
474
#pragma omp parallel num_threads(MaxThreads)
475
            {
476
                KernelClass* myThreadkernels = (kernels[omp_get_thread_num()]);
477
                //This single section post and receive the comms, and then do the M2M associated with it.
478
#pragma omp single nowait
479 480 481 482 483 484 485 486 487 488 489 490 491 492 493 494 495 496 497 498 499 500 501 502 503 504 505 506 507
                {
                    FLOG(singleCounter.tic());
                    // Master proc never send
                    if(idProcess != 0){
                        // Skip process that have no work at that level
                        while( currentProcIdToSendTo && !procHasWorkAtLevel(idxLevel, currentProcIdToSendTo)  ){
                            --currentProcIdToSendTo;
                        }
                        // Does the next proc that has work is sharing the parent of my left cell
                        if(procHasWorkAtLevel(idxLevel, currentProcIdToSendTo) && procCoversMyLeftBorderCell(idxLevel, currentProcIdToSendTo)){
                            FAssertLF(nbCellsToSkip != 0);

                            char packageFlags = 0;
                            sendBuffer.write(packageFlags);

                            // Only the cell the most on the right out of my working interval should be taken in
                            // consideration (at pos nbCellsToSkip-1) other (x < nbCellsToSkip-1) have already been sent
                            const CellClass* const* const child = iterArray[nbCellsToSkip-1].getCurrentChild();
                            for(int idxChild = 0 ; idxChild < 8 ; ++idxChild){
                                // Check if child exists and it was part of my working interval
                                if( child[idxChild] && getWorkingInterval((idxLevel+1), idProcess).leftIndex <= child[idxChild]->getMortonIndex() ){
                                    // Add the cell to the buffer
                                    child[idxChild]->serializeUp(sendBuffer);
                                    packageFlags = char(packageFlags | (0x1 << idxChild));
                                }
                            }
                            // Add the flag as first value
                            sendBuffer.writeAt(0,packageFlags);
                            // Post the message
508
                            bufferSize = sendBuffer.getSize();
509
                            MPI_Isend(&bufferSize, 1, FMpi::GetType(bufferSize), currentProcIdToSendTo,
510
                                      FMpi::TagFmmM2MSize + idxLevel, fcomCompute.getComm(), &requestsSize[iterMpiRequestsSize++]);
511
                            FAssertLF(sendBuffer.getSize() < std::numeric_limits<int>::max());
BRAMAS Berenger's avatar
BRAMAS Berenger committed
512
                            MPI_Isend(sendBuffer.data(), int(sendBuffer.getSize()), MPI_BYTE, currentProcIdToSendTo,
513
                                      FMpi::TagFmmM2M + idxLevel, fcomCompute.getComm(), &requests[iterMpiRequests++]);
514 515 516 517 518 519 520 521 522 523 524 525 526
                        }
                    }

                    //Post receive, Datas needed in several parts of the section
                    int nbProcThatSendToMe = 0;

                    if(hasToReceive){
                        //Test : if the firstProcThatSend father minimal value in interval is lesser than mine
                        int idProcSource = firstProcThatSend;
                        // Find the last proc that should send to me
                        while( idProcSource < nbProcess
                               && ( !procHasWorkAtLevel(idxLevel+1, idProcSource) || procCoversMyRightBorderCell(idxLevel, idProcSource) )){
                            if(procHasWorkAtLevel(idxLevel+1, idProcSource) && procCoversMyRightBorderCell(idxLevel, idProcSource)){
527
                                MPI_Irecv(&recvBufferSize[nbProcThatSendToMe], 1, FMpi::GetType(recvBufferSize[nbProcThatSendToMe]),
528
                                          idProcSource, FMpi::TagFmmM2MSize + idxLevel, fcomCompute.getComm(), &requestsSize[iterMpiRequestsSize++]);
529 530 531 532 533 534 535 536 537 538 539 540 541 542 543 544 545 546 547 548 549 550 551
                                nbProcThatSendToMe += 1;
                                FAssertLF(nbProcThatSendToMe <= 7);
                            }
                            ++idProcSource;
                        }
                    }

                    //Wait For the comms, and do the work
                    // Are we sending or waiting anything?
                    if(iterMpiRequestsSize){
                        FAssertLF(iterMpiRequestsSize <= 8);
                        MPI_Waitall( iterMpiRequestsSize, requestsSize, statusSize);
                    }

                    if(hasToReceive){
                        nbProcThatSendToMe = 0;
                        //Test : if the firstProcThatSend father minimal value in interval is lesser than mine
                        int idProcSource = firstProcThatSend;
                        // Find the last proc that should send to me
                        while( idProcSource < nbProcess
                               && ( !procHasWorkAtLevel(idxLevel+1, idProcSource) || procCoversMyRightBorderCell(idxLevel, idProcSource) )){
                            if(procHasWorkAtLevel(idxLevel+1, idProcSource) && procCoversMyRightBorderCell(idxLevel, idProcSource)){
                                recvBuffer[nbProcThatSendToMe].cleanAndResize(recvBufferSize[nbProcThatSendToMe]);
552
                                FAssertLF(recvBufferSize[nbProcThatSendToMe] < std::numeric_limits<int>::max());
BRAMAS Berenger's avatar
BRAMAS Berenger committed
553
                                MPI_Irecv(recvBuffer[nbProcThatSendToMe].data(), int(recvBufferSize[nbProcThatSendToMe]), MPI_BYTE,
554
                                          idProcSource, FMpi::TagFmmM2M + idxLevel, fcomCompute.getComm(), &requests[iterMpiRequests++]);
555 556 557 558 559 560 561 562 563 564 565 566
                                nbProcThatSendToMe += 1;
                                FAssertLF(nbProcThatSendToMe <= 7);
                            }
                            ++idProcSource;
                        }
                    }

                    //Wait For the comms, and do the work
                    // Are we sending or waiting anything?
                    if(iterMpiRequests){
                        FAssertLF(iterMpiRequests <= 8);
                        MPI_Waitall( iterMpiRequests, requests, status);
567
                    }
568

569 570 571 572 573 574 575
                    // We had received something so we need to proceed the last M2M
                    if( hasToReceive ){
                        FAssertLF(iterMpiRequests != 0);
                        CellClass* currentChild[8];
                        memcpy(currentChild, iterArray[totalNbCellsAtLevel - 1].getCurrentChild(), 8 * sizeof(CellClass*));

                        // Retreive data and merge my child and the child from others
576
                        int positionToInsert = 0;
577
                        for(int idxProc = 0 ; idxProc < nbProcThatSendToMe ; ++idxProc){
BRAMAS Berenger's avatar
BRAMAS Berenger committed
578
                            unsigned packageFlags = unsigned(recvBuffer[idxProc].getValue<unsigned char>());
579 580 581 582 583 584 585

                            int position = 0;
                            while( packageFlags && position < 8){
                                while(!(packageFlags & 0x1)){
                                    packageFlags >>= 1;
                                    ++position;
                                }
586 587
                                FAssertLF(positionToInsert < 7);
                                FAssertLF(position < 8);
588
                                FAssertLF(!currentChild[position], "Already has a cell here");
589 590
                                recvBufferCells[positionToInsert].deserializeUp(recvBuffer[idxProc]);
                                currentChild[position] = (CellClass*) &recvBufferCells[positionToInsert];
591 592

                                packageFlags >>= 1;
593 594
                                position += 1;
                                positionToInsert += 1;
595
                            }
596 597

                            recvBuffer[idxProc].seek(0);
598 599
                        }
                        // Finally compute
600
                        myThreadkernels->M2M( iterArray[totalNbCellsAtLevel - 1].getCurrentCell() , currentChild, idxLevel);
601 602 603 604 605 606 607 608
                        firstProcThatSend += nbProcThatSendToMe - 1;
                    }
                    // Reset buffer
                    sendBuffer.reset();
                    FLOG(singleCounter.tac());
                }//End Of Single section

                // All threads proceed the M2M
609
#pragma omp for nowait schedule(dynamic, userChunkSize)
610 611 612 613 614 615 616 617 618 619 620 621 622
                for( int idxCell = nbCellsToSkip ; idxCell < nbCellsForThreads ; ++idxCell){
                    myThreadkernels->M2M( iterArray[idxCell].getCurrentCell() , iterArray[idxCell].getCurrentChild(), idxLevel);
                }
            }//End of parallel section
            FLOG(parallelCounter.tac());
        }

        FLOG(counterTime.tac());
        FLOG(computationCounter.tac());
        FLOG( FLog::Controller << "\tFinished (@Upward Pass (M2M) = "  << counterTime.elapsed() << " s)\n" );
        FLOG( FLog::Controller << "\t\t Computation : " << computationCounter.elapsed() << " s\n" );
        FLOG( FLog::Controller << "\t\t Single : " << singleCounter.cumulated() << " s\n" );
        FLOG( FLog::Controller << "\t\t Parallel : " << parallelCounter.cumulated() << " s\n" );
BRAMAS Berenger's avatar
BRAMAS Berenger committed
623
        FLOG( FLog::Controller.flush());
624
    }
625

626 627 628
    /////////////////////////////////////////////////////////////////////////////
    // Downard
    /////////////////////////////////////////////////////////////////////////////
629 630


631
    void transferPass(){
632 633 634 635 636 637 638 639 640 641 642 643 644
        FLOG( FLog::Controller.write("\tStart Downward Pass (M2L)\n").write(FLog::Flush); );
        FLOG(FTic counterTime);
        FLOG(FTic computationCounter);
        FLOG(FTic sendCounter);
        FLOG(FTic receiveCounter);
        FLOG(FTic prepareCounter);
        FLOG(FTic gatherCounter);

        //////////////////////////////////////////////////////////////////
        // First know what to send to who
        //////////////////////////////////////////////////////////////////

        // pointer to send
645
        std::unique_ptr<FVector<typename OctreeClass::Iterator>[]> toSend(new FVector<typename OctreeClass::Iterator>[nbProcess * OctreeHeight]);
646
        // index
647 648
        long long int*const indexToSend = new long long int[nbProcess * OctreeHeight];
        memset(indexToSend, 0, sizeof(long long int) * nbProcess * OctreeHeight);
649 650 651 652 653 654
        // To know which one has need someone
        FBoolArray** const leafsNeedOther = new FBoolArray*[OctreeHeight];
        memset(leafsNeedOther, 0, sizeof(FBoolArray*) * OctreeHeight);

        // All process say to each others
        // what the will send to who
655 656
        long long int*const globalReceiveMap = new long long  int[nbProcess * nbProcess * OctreeHeight];
        memset(globalReceiveMap, 0, sizeof(long long  int) * nbProcess * nbProcess * OctreeHeight);
657 658 659 660 661 662

        FMpiBufferWriter**const sendBuffer = new FMpiBufferWriter*[nbProcess * OctreeHeight];
        memset(sendBuffer, 0, sizeof(FMpiBufferWriter*) * nbProcess * OctreeHeight);

        FMpiBufferReader**const recvBuffer = new FMpiBufferReader*[nbProcess * OctreeHeight];
        memset(recvBuffer, 0, sizeof(FMpiBufferReader*) * nbProcess * OctreeHeight);
663

664
#pragma omp parallel num_threads(MaxThreads)
665
        {
666
#pragma omp master
667 668 669 670 671 672 673 674 675 676 677 678
            {
                {
                    FLOG(prepareCounter.tic());

                    std::unique_ptr<typename OctreeClass::Iterator[]> iterArrayLocal(new typename OctreeClass::Iterator[numberOfLeafs]);

                    // To know if a leaf has been already sent to a proc
                    bool*const alreadySent = new bool[nbProcess];
                    memset(alreadySent, 0, sizeof(bool) * nbProcess);

                    typename OctreeClass::Iterator octreeIterator(tree);
                    octreeIterator.moveDown();
679

680
                    for(int idxLevel = 2 ; idxLevel < FAbstractAlgorithm::upperWorkingLevel ; ++idxLevel){
681 682 683
                        octreeIterator.moveDown();
                    }

684 685
                    typename OctreeClass::Iterator avoidGotoLeftIterator(octreeIterator);
                    // for each levels
686
                    for(int idxLevel = FAbstractAlgorithm::upperWorkingLevel ; idxLevel < FAbstractAlgorithm::lowerWorkingLevel ; ++idxLevel ){
687

688
                        const int separationCriteria = (idxLevel != FAbstractAlgorithm::lowerWorkingLevel-1 ? 1 : leafLevelSeparationCriteria);
689

690 691 692 693 694 695 696 697 698 699 700 701 702 703 704 705 706 707 708 709 710 711 712 713
                        if(!procHasWorkAtLevel(idxLevel, idProcess)){
                            avoidGotoLeftIterator.moveDown();
                            octreeIterator = avoidGotoLeftIterator;
                            continue;
                        }

                        int numberOfCells = 0;

                        while(octreeIterator.getCurrentGlobalIndex() <  getWorkingInterval(idxLevel , idProcess).leftIndex){
                            octreeIterator.moveRight();
                        }

                        // for each cells
                        do{
                            iterArrayLocal[numberOfCells] = octreeIterator;
                            ++numberOfCells;
                        } while(octreeIterator.moveRight());
                        avoidGotoLeftIterator.moveDown();
                        octreeIterator = avoidGotoLeftIterator;

                        leafsNeedOther[idxLevel] = new FBoolArray(numberOfCells);

                        // Which cell potentialy needs other data and in the same time
                        // are potentialy needed by other
714
                        MortonIndex neighborsIndexes[/*189+26+1*/216];
715 716
                        for(int idxCell = 0 ; idxCell < numberOfCells ; ++idxCell){
                            // Find the M2L neigbors of a cell
717
                            const int counter = iterArrayLocal[idxCell].getCurrentGlobalCoordinate().getInteractionNeighbors(idxLevel, neighborsIndexes, separationCriteria);
718 719 720 721 722 723 724 725 726 727 728 729 730 731 732 733 734 735 736 737 738 739 740 741

                            memset(alreadySent, false, sizeof(bool) * nbProcess);
                            bool needOther = false;
                            // Test each negibors to know which one do not belong to us
                            for(int idxNeigh = 0 ; idxNeigh < counter ; ++idxNeigh){
                                if(neighborsIndexes[idxNeigh] < getWorkingInterval(idxLevel , idProcess).leftIndex
                                        || (getWorkingInterval(idxLevel , idProcess).rightIndex) < neighborsIndexes[idxNeigh]){
                                    int procToReceive = idProcess;
                                    while( 0 != procToReceive && neighborsIndexes[idxNeigh] < getWorkingInterval(idxLevel , procToReceive).leftIndex ){
                                        --procToReceive;
                                    }
                                    while( procToReceive != nbProcess -1 && (getWorkingInterval(idxLevel , procToReceive).rightIndex) < neighborsIndexes[idxNeigh]){
                                        ++procToReceive;
                                    }
                                    // Maybe already sent to that proc?
                                    if( !alreadySent[procToReceive]
                                            && getWorkingInterval(idxLevel , procToReceive).leftIndex <= neighborsIndexes[idxNeigh]
                                            && neighborsIndexes[idxNeigh] <= getWorkingInterval(idxLevel , procToReceive).rightIndex){

                                        alreadySent[procToReceive] = true;

                                        needOther = true;

                                        toSend[idxLevel * nbProcess + procToReceive].push(iterArrayLocal[idxCell]);
742 743 744 745 746
                                        if(indexToSend[idxLevel * nbProcess + procToReceive] == 0){
                                            indexToSend[idxLevel * nbProcess + procToReceive] = sizeof(int);
                                        }
                                        indexToSend[idxLevel * nbProcess + procToReceive] += iterArrayLocal[idxCell].getCurrentCell()->getSavedSizeUp();
                                        indexToSend[idxLevel * nbProcess + procToReceive] += sizeof(MortonIndex);
747
                                        indexToSend[idxLevel * nbProcess + procToReceive] += sizeof(FSize);
748 749 750 751 752 753 754 755 756 757 758 759 760 761 762 763 764 765
                                    }
                                }
                            }
                            if(needOther){
                                leafsNeedOther[idxLevel]->set(idxCell,true);
                            }
                        }
                    }
                    FLOG(prepareCounter.tac());

                    delete[] alreadySent;
                }

                //////////////////////////////////////////////////////////////////
                // Gather this information
                //////////////////////////////////////////////////////////////////

                FLOG(gatherCounter.tic());
766
                FMpi::MpiAssert( MPI_Allgather( indexToSend, nbProcess * OctreeHeight, MPI_LONG_LONG_INT, globalReceiveMap, nbProcess * OctreeHeight, MPI_LONG_LONG_INT, fcomCompute.getComm()),  __LINE__ );
767 768 769 770 771 772 773 774 775
                FLOG(gatherCounter.tac());

                //////////////////////////////////////////////////////////////////
                // Send and receive for real
                //////////////////////////////////////////////////////////////////

                FLOG(sendCounter.tic());
                // Then they can send and receive (because they know what they will receive)
                // To send in asynchrone way
776 777
                std::vector<MPI_Request> requests;
                requests.reserve(2 * nbProcess * OctreeHeight);
778 779 780

                for(int idxLevel = 2 ; idxLevel < OctreeHeight ; ++idxLevel ){
                    for(int idxProc = 0 ; idxProc < nbProcess ; ++idxProc){
781
                        const long long int toSendAtProcAtLevel = indexToSend[idxLevel * nbProcess + idxProc];
782
                        if(toSendAtProcAtLevel != 0){
783
                            sendBuffer[idxLevel * nbProcess + idxProc] = new FMpiBufferWriter(toSendAtProcAtLevel);
784

785 786 787
                            sendBuffer[idxLevel * nbProcess + idxProc]->write(int(toSend[idxLevel * nbProcess + idxProc].getSize()));

                            for(int idxLeaf = 0 ; idxLeaf < toSend[idxLevel * nbProcess + idxProc].getSize(); ++idxLeaf){
788 789
                                const FSize currentTell = sendBuffer[idxLevel * nbProcess + idxProc]->getSize();
                                sendBuffer[idxLevel * nbProcess + idxProc]->write(currentTell);
790 791 792 793 794
                                const MortonIndex cellIndex = toSend[idxLevel * nbProcess + idxProc][idxLeaf].getCurrentGlobalIndex();
                                sendBuffer[idxLevel * nbProcess + idxProc]->write(cellIndex);
                                toSend[idxLevel * nbProcess + idxProc][idxLeaf].getCurrentCell()->serializeUp(*sendBuffer[idxLevel * nbProcess + idxProc]);
                            }

795 796
                            FAssertLF(sendBuffer[idxLevel * nbProcess + idxProc]->getSize() == toSendAtProcAtLevel);

BRAMAS Berenger's avatar
BRAMAS Berenger committed
797 798
                            FMpi::ISendSplit(sendBuffer[idxLevel * nbProcess + idxProc]->data(),
                                    sendBuffer[idxLevel * nbProcess + idxProc]->getSize(), idxProc,
799
                                    FMpi::TagLast + idxLevel*100, fcomCompute, &requests);
800 801
                        }

802
                        const long long int toReceiveFromProcAtLevel = globalReceiveMap[(idxProc * nbProcess * OctreeHeight) + idxLevel * nbProcess + idProcess];
803
                        if(toReceiveFromProcAtLevel){
804
                            recvBuffer[idxLevel * nbProcess + idxProc] = new FMpiBufferReader(toReceiveFromProcAtLevel);
805

BRAMAS Berenger's avatar
BRAMAS Berenger committed
806 807
                            FMpi::IRecvSplit(recvBuffer[idxLevel * nbProcess + idxProc]->data(),
                                    recvBuffer[idxLevel * nbProcess + idxProc]->getCapacity(), idxProc,
808
                                    FMpi::TagLast + idxLevel*100, fcomCompute, &requests);
809 810 811 812 813 814 815 816 817
                        }
                    }
                }

                //////////////////////////////////////////////////////////////////
                // Wait received data and compute
                //////////////////////////////////////////////////////////////////

                // Wait to receive every things (and send every things)
818
                FMpi::MpiAssert(MPI_Waitall(int(requests.size()), requests.data(), MPI_STATUS_IGNORE), __LINE__);
819 820 821 822 823 824 825

                FLOG(sendCounter.tac());
            }//End of Master region

            //////////////////////////////////////////////////////////////////
            // Do M2L
            //////////////////////////////////////////////////////////////////
PIACIBELLO Cyrille's avatar
PIACIBELLO Cyrille committed
826

827
#pragma omp single nowait
828 829 830
            {
                typename OctreeClass::Iterator octreeIterator(tree);
                octreeIterator.moveDown();
831

832
                for(int idxLevel = 2 ; idxLevel < FAbstractAlgorithm::upperWorkingLevel ; ++idxLevel){
833 834 835
                    octreeIterator.moveDown();
                }

836 837 838
                typename OctreeClass::Iterator avoidGotoLeftIterator(octreeIterator);
                // Now we can compute all the data
                // for each levels
839
                for(int idxLevel = FAbstractAlgorithm::upperWorkingLevel ; idxLevel < FAbstractAlgorithm::lowerWorkingLevel ; ++idxLevel ){
840
                    const int separationCriteria = (idxLevel != FAbstractAlgorithm::lowerWorkingLevel-1 ? 1 : leafLevelSeparationCriteria);
841

842 843 844 845 846 847 848 849 850 851 852 853 854 855 856 857 858 859 860 861
                    if(!procHasWorkAtLevel(idxLevel, idProcess)){
                        avoidGotoLeftIterator.moveDown();
                        octreeIterator = avoidGotoLeftIterator;
                        continue;
                    }

                    int numberOfCells = 0;
                    while(octreeIterator.getCurrentGlobalIndex() <  getWorkingInterval(idxLevel , idProcess).leftIndex){
                        octreeIterator.moveRight();
                    }
                    // for each cells
                    do{
                        iterArray[numberOfCells] = octreeIterator;
                        ++numberOfCells;
                    } while(octreeIterator.moveRight());
                    avoidGotoLeftIterator.moveDown();
                    octreeIterator = avoidGotoLeftIterator;

                    FLOG(computationCounter.tic());
                    {
862
                        const int chunckSize = userChunkSize;
863
                        for(int idxCell = 0 ; idxCell < numberOfCells ; idxCell += chunckSize){
864
#pragma omp task default(none) shared(numberOfCells,idxLevel) firstprivate(idxCell) //+ shared(chunckSize)
865 866
                            {
                                KernelClass * const myThreadkernels = kernels[omp_get_thread_num()];
867 868
                                const CellClass* neighbors[342];
                                int neighborPositions[342];
869 870 871

                                const int nbCellToCompute = FMath::Min(chunckSize, numberOfCells-idxCell);
                                for(int idxCellToCompute = idxCell ; idxCellToCompute < idxCell+nbCellToCompute ; ++idxCellToCompute){
872 873
                                    const int counter = tree->getInteractionNeighbors(neighbors,  neighborPositions, iterArray[idxCellToCompute].getCurrentGlobalCoordinate(), idxLevel, separationCriteria);
                                    if(counter) myThreadkernels->M2L( iterArray[idxCellToCompute].getCurrentCell() , neighbors, neighborPositions, counter, idxLevel);
874 875 876 877
                                }
                            }
                        }
                    }//End of task spawning
PIACIBELLO Cyrille's avatar
PIACIBELLO Cyrille committed
878

879
#pragma omp taskwait
PIACIBELLO Cyrille's avatar
PIACIBELLO Cyrille committed
880

881
                    for(int idxThread = 0 ; idxThread < omp_get_num_threads() ; ++idxThread){
882
#pragma omp task  default(none) firstprivate(idxThread,idxLevel)
883 884 885 886
                        {
                            kernels[idxThread]->finishedLevelM2L(idxLevel);
                        }
                    }
887
#pragma omp taskwait
888 889 890 891 892 893 894 895 896 897 898

                    FLOG(computationCounter.tac());
                }
            }
        }


        {
            FLOG(receiveCounter.tic());
            typename OctreeClass::Iterator octreeIterator(tree);
            octreeIterator.moveDown();
899

900
            for(int idxLevel = 2 ; idxLevel < FAbstractAlgorithm::upperWorkingLevel ; ++idxLevel){
901 902 903
                octreeIterator.moveDown();
            }

904 905 906
            typename OctreeClass::Iterator avoidGotoLeftIterator(octreeIterator);
            // compute the second time
            // for each levels
907
            for(int idxLevel = FAbstractAlgorithm::upperWorkingLevel ; idxLevel < FAbstractAlgorithm::lowerWorkingLevel ; ++idxLevel ){
908
                const int separationCriteria = (idxLevel != FAbstractAlgorithm::lowerWorkingLevel-1 ? 1 : leafLevelSeparationCriteria);
909

910 911 912 913 914 915 916 917 918
                if(!procHasWorkAtLevel(idxLevel, idProcess)){
                    avoidGotoLeftIterator.moveDown();
                    octreeIterator = avoidGotoLeftIterator;
                    continue;
                }

                // put the received data into a temporary tree
                FLightOctree<CellClass> tempTree;
                for(int idxProc = 0 ; idxProc < nbProcess ; ++idxProc){
919
                    if(recvBuffer[idxLevel * nbProcess + idxProc]){
920
                        const int toReceiveFromProcAtLevel = recvBuffer[idxLevel * nbProcess + idxProc]->template getValue<int>();
921

922
                        for(int idxCell = 0 ; idxCell < toReceiveFromProcAtLevel ; ++idxCell){
923 924 925 926
                            const FSize currentTell = recvBuffer[idxLevel * nbProcess + idxProc]->tell();
                            const FSize verifCurrentTell = recvBuffer[idxLevel * nbProcess + idxProc]->template getValue<FSize>();
                            FAssertLF(currentTell == verifCurrentTell, currentTell, " ", verifCurrentTell);

927
                            const MortonIndex cellIndex = recvBuffer[idxLevel * nbProcess + idxProc]->template getValue<MortonIndex>();
928

929 930 931
                            CellClass* const newCell = new CellClass;
                            newCell->setMortonIndex(cellIndex);
                            newCell->deserializeUp(*recvBuffer[idxLevel * nbProcess + idxProc]);
932

933 934 935 936 937
                            tempTree.insertCell(cellIndex, idxLevel, newCell);
                        }

                        FAssertLF(globalReceiveMap[(idxProc * nbProcess * OctreeHeight) + idxLevel * nbProcess + idProcess] ==
                                recvBuffer[idxLevel * nbProcess + idxProc]->tell());
938 939 940 941 942 943 944 945 946 947 948 949 950 951 952 953 954 955 956 957 958 959 960 961 962 963
                    }
                }

                // take cells from our octree only if they are
                // linked to received data
                int numberOfCells = 0;
                int realCellId = 0;

                while(octreeIterator.getCurrentGlobalIndex() <  getWorkingInterval(idxLevel , idProcess).leftIndex){
                    octreeIterator.moveRight();
                }
                // for each cells
                do{
                    // copy cells that need data from others
                    if(leafsNeedOther[idxLevel]->get(realCellId++)){
                        iterArray[numberOfCells++] = octreeIterator;
                    }
                } while(octreeIterator.moveRight());
                avoidGotoLeftIterator.moveDown();
                octreeIterator = avoidGotoLeftIterator;

                delete leafsNeedOther[idxLevel];
                leafsNeedOther[idxLevel] = nullptr;

                // Compute this cells
                FLOG(computationCounter.tic());
964
#pragma omp parallel num_threads(MaxThreads)
965 966
                {
                    KernelClass * const myThreadkernels = kernels[omp_get_thread_num()];
967 968
                    MortonIndex neighborsIndex[/*189+26+1*/216];
                    int neighborsPosition[/*189+26+1*/216];
969 970
                    const CellClass* neighbors[342];
                    int neighborPositions[342];
971

972
#pragma omp for  schedule(dynamic, userChunkSize) nowait
973 974
                    for(int idxCell = 0 ; idxCell < numberOfCells ; ++idxCell){
                        // compute indexes
975
                        const int counterNeighbors = iterArray[idxCell].getCurrentGlobalCoordinate().getInteractionNeighbors(idxLevel, neighborsIndex, neighborsPosition, separationCriteria);
976 977 978 979 980 981 982 983 984 985

                        int counter = 0;
                        // does we receive this index from someone?
                        for(int idxNeig = 0 ;idxNeig < counterNeighbors ; ++idxNeig){
                            if(neighborsIndex[idxNeig] < (getWorkingInterval(idxLevel , idProcess).leftIndex)
                                    || (getWorkingInterval(idxLevel , idProcess).rightIndex) < neighborsIndex[idxNeig]){

                                CellClass*const otherCell = tempTree.getCell(neighborsIndex[idxNeig], idxLevel);

                                if(otherCell){
986 987
                                    neighbors[counter] = otherCell;
                                    neighborPositions[counter] = neighborsPosition[idxNeig];
988 989 990 991 992 993
                                    ++counter;
                                }
                            }
                        }
                        // need to compute
                        if(counter){
994
                            myThreadkernels->M2L( iterArray[idxCell].getCurrentCell() , neighbors, neighborPositions, counter, idxLevel);
995 996 997 998 999 1000 1001 1002 1003 1004 1005 1006 1007 1008 1009 1010 1011 1012 1013 1014 1015 1016 1017 1018 1019 1020 1021 1022 1023 1024
                        }
                    }

                    myThreadkernels->finishedLevelM2L(idxLevel);
                }
                FLOG(computationCounter.tac());
            }
            FLOG(receiveCounter.tac());
        }

        for(int idxComm = 0 ; idxComm < nbProcess * OctreeHeight; ++idxComm){
            delete sendBuffer[idxComm];
            delete recvBuffer[idxComm];
        }
        for(int idxComm = 0 ; idxComm < OctreeHeight; ++idxComm){
            delete leafsNeedOther[idxComm];
        }
        delete[] sendBuffer;
        delete[] recvBuffer;
        delete[] indexToSend;
        delete[] leafsNeedOther;
        delete[] globalReceiveMap;


        FLOG( FLog::Controller << "\tFinished (@Downward Pass (M2L) = "  << counterTime.tacAndElapsed() << " s)\n" );
        FLOG( FLog::Controller << "\t\t Computation : " << computationCounter.cumulated() << " s\n" );
        FLOG( FLog::Controller << "\t\t Send : " << sendCounter.cumulated() << " s\n" );
        FLOG( FLog::Controller << "\t\t Receive : " << receiveCounter.cumulated() << " s\n" );
        FLOG( FLog::Controller << "\t\t Gather : " << gatherCounter.cumulated() << " s\n" );
        FLOG( FLog::Controller << "\t\t Prepare : " << prepareCounter.cumulated() << " s\n" );
BRAMAS Berenger's avatar
BRAMAS Berenger committed
1025
        FLOG( FLog::Controller.flush());
1026

1027
    }
1028

1029 1030 1031 1032
    //////////////////////////////////////////////////////////////////
    // ---------------- L2L ---------------
    //////////////////////////////////////////////////////////////////

1033
    void downardPass(){ // second L2L
1034 1035 1036 1037 1038 1039 1040 1041 1042
        FLOG( FLog::Controller.write("\tStart Downward Pass (L2L)\n").write(FLog::Flush); );
        FLOG(FTic counterTime);
        FLOG(FTic computationCounter);
        FLOG(FTic prepareCounter);
        FLOG(FTic waitCounter);

        // Start from leal level - 1
        typename OctreeClass::Iterator octreeIterator(tree);
        octreeIterator.moveDown();
1043

1044
        for(int idxLevel = 2 ; idxLevel < FAbstractAlgorithm::upperWorkingLevel ; ++idxLevel){
1045 1046 1047
            octreeIterator.moveDown();
        }

1048 1049 1050 1051
        typename OctreeClass::Iterator avoidGotoLeftIterator(octreeIterator);

        // Max 1 receive and 7 send (but 7 times the same data)
        MPI_Request*const requests = new MPI_Request[8];
1052
        MPI_Request*const requestsSize = new MPI_Request[8];
1053

1054
        const int heightMinusOne = FAbstractAlgorithm::lowerWorkingLevel - 1;
1055

1056 1057
        FMpiBufferWriter sendBuffer;
        FMpiBufferReader recvBuffer;
1058 1059 1060 1061

        int righestProcToSendTo   = nbProcess - 1;

        // for each levels exepted leaf level
1062
        for(int idxLevel = FAbstractAlgorithm::upperWorkingLevel ; idxLevel < heightMinusOne ; ++idxLevel ){
1063 1064 1065 1066 1067 1068 1069 1070 1071 1072 1073 1074 1075 1076 1077 1078 1079 1080 1081 1082 1083 1084 1085 1086 1087 1088 1089 1090 1091 1092 1093 1094 1095 1096
            // If nothing to do in the next level skip the current one
            if(idProcess != 0 && !procHasWorkAtLevel(idxLevel+1, idProcess) ){
                avoidGotoLeftIterator.moveDown();
                octreeIterator = avoidGotoLeftIterator;
                continue;
            }

            // Copy all the cells in an array even the one that are out of my working interval
            int totalNbCellsAtLevel = 0;
            do{
                iterArray[totalNbCellsAtLevel++] = octreeIterator;
            } while(octreeIterator.moveRight());
            avoidGotoLeftIterator.moveDown();
            octreeIterator = avoidGotoLeftIterator;

            // Count the number of cells that are out of my working interval
            int nbCellsToSkip = 0;
            while(nbCellsToSkip < totalNbCellsAtLevel && iterArray[nbCellsToSkip].getCurrentGlobalIndex() < getWorkingInterval(idxLevel , idProcess).leftIndex){
                nbCellsToSkip += 1;
            }

            // Check if someone will send a cell to me
            bool hasToReceive = false;
            int idxProcToReceive = idProcess - 1;
            if(idProcess != 0 && nbCellsToSkip){
                // Starting from my left neighbor stop at the first proc that has work to do (not null interval)
                while(idxProcToReceive && !procHasWorkAtLevel(idxLevel, idxProcToReceive) ){
                    idxProcToReceive -= 1;
                }
                // Check if we find such a proc and that it share a cell with us on the border
                if(procHasWorkAtLevel(idxLevel, idxProcToReceive) && procCoversMyLeftBorderCell(idxLevel, idxProcToReceive)){
                    hasToReceive = true;
                }
            }
1097

1098
#pragma omp parallel num_threads(MaxThreads)
1099
            {
1100
                KernelClass* myThreadkernels = (kernels[omp_get_thread_num()]);
1101
#pragma omp single nowait
1102 1103 1104
                {
                    FLOG(prepareCounter.tic());
                    int iterRequests = 0;
1105
                    int iterRequestsSize = 0;
1106 1107
                    FSize recvBufferSize = 0;
                    FSize sendBufferSize;
1108 1109
                    // Post the receive
                    if(hasToReceive){
1110
                        FMpi::MpiAssert( MPI_Irecv( &recvBufferSize, 1, FMpi::GetType(recvBufferSize), idxProcToReceive,
1111
                                                    FMpi::TagFmmL2LSize + idxLevel, fcomCompute.getComm(), &requestsSize[iterRequestsSize++]), __LINE__ );
1112 1113 1114 1115 1116 1117 1118 1119 1120 1121 1122 1123 1124 1125 1126
                    }

                    // We have to be sure that we are not sending if we have no work in the current level
                    if(idProcess != nbProcess - 1 && idProcess < righestProcToSendTo && procHasWorkAtLevel(idxLevel, idProcess)){
                        int idxProcSend = idProcess + 1;
                        int nbMessageSent = 0;
                        // From the proc on the right to righestProcToSendTo, check if we have to send something
                        while(idxProcSend <= righestProcToSendTo && ( !procHasWorkAtLevel(idxLevel+1, idxProcSend) || procCoversMyRightBorderCell(idxLevel, idxProcSend)) ){
                            // We know that if the proc has work at the next level it share a cell with us due to the while condition
                            if(procHasWorkAtLevel(idxLevel+1, idxProcSend)){
                                FAssertLF(procCoversMyRightBorderCell(idxLevel, idxProcSend));
                                // If first message then serialize the cell to send
                                if( nbMessageSent == 0 ){
                                    // We send our last cell
                                    iterArray[totalNbCellsAtLevel - 1].getCurrentCell()->serializeDown(sendBuffer);
1127
                                    sendBufferSize = sendBuffer.getSize();
1128 1129
                                }
                                // Post the send message
1130
                                FMpi::MpiAssert( MPI_Isend(&sendBufferSize, 1, FMpi::GetType(sendBufferSize), idxProcSend,
1131
                                                           FMpi::TagFmmL2LSize + idxLevel, fcomCompute.getComm(), &requestsSize[iterRequestsSize++]), __LINE__);
1132
                                FAssertLF(sendBuffer.getSize() < std::numeric_limits<int>::max());
BRAMAS Berenger's avatar
BRAMAS Berenger committed
1133
                                FMpi::MpiAssert( MPI_Isend(sendBuffer.data(), int(sendBuffer.getSize()), MPI_BYTE, idxProcSend,
1134
                                                           FMpi::TagFmmL2L + idxLevel, fcomCompute.getComm(), &requests[iterRequests++]), __LINE__);
1135 1136 1137 1138 1139 1140 1141
                                // Inc and check the counter
                                nbMessageSent += 1;
                                FAssertLF(nbMessageSent <= 7);
                            }
                            idxProcSend += 1;
                        }
                        // Next time we will not need to go further than idxProcSend
1142 1143 1144
                        if(idxProcSend < righestProcToSendTo){
                            righestProcToSendTo = idxProcSend;
                        }
1145
                    }
1146

1147
                    // Finalize the communication
1148 1149
                    if(iterRequestsSize){
                        FLOG(waitCounter.tic());
1150
                        FAssertLF(iterRequestsSize <= 8);
1151
                        FMpi::MpiAssert(MPI_Waitall( iterRequestsSize, requestsSize, MPI_STATUSES_IGNORE), __LINE__);
1152 1153 1154 1155
                        FLOG(waitCounter.tac());
                    }

                    if(hasToReceive){
1156
                        recvBuffer.cleanAndResize(recvBufferSize);
1157
                        FAssertLF(recvBuffer.getCapacity() < std::numeric_limits<int>::max());
BRAMAS Berenger's avatar
BRAMAS Berenger committed
1158
                        FMpi::MpiAssert( MPI_Irecv( recvBuffer.data(), int(recvBuffer.getCapacity()), MPI_BYTE, idxProcToReceive,
1159
                                                    FMpi::TagFmmL2L + idxLevel, fcomCompute.getComm(), &requests[iterRequests++]), __LINE__ );
1160 1161
                    }

1162 1163
                    if(iterRequests){
                        FLOG(waitCounter.tic());
1164
                        FAssertLF(iterRequests <= 8);
1165
                        FMpi::MpiAssert(MPI_Waitall( iterRequests, requests, MPI_STATUSES_IGNORE), __LINE__);
1166 1167
                        FLOG(waitCounter.tac());
                    }
1168

1169 1170 1171 1172 1173 1174
                    // If we receive something proceed the L2L
                    if(hasToReceive){
                        FAssertLF(iterRequests != 0);
                        // In this case we know that we have to perform the L2L with the last cell that are
                        // exclude from our working interval nbCellsToSkip-1
                        iterArray[nbCellsToSkip-1].getCurrentCell()->deserializeDown(recvBuffer);
1175
                        myThreadkernels->L2L( iterArray[nbCellsToSkip-1].getCurrentCell() , iterArray[nbCellsToSkip-1].getCurrentChild(), idxLevel);
1176 1177 1178
                    }
                    FLOG(prepareCounter.tac());
                }
1179

1180
#pragma omp single nowait
1181 1182 1183 1184
                {
                    FLOG(computationCounter.tic());
                }
                // Threads are working on all the cell of our working interval at that level
1185
#pragma omp for nowait  schedule(dynamic, userChunkSize)
1186 1187 1188 1189 1190 1191 1192 1193 1194 1195 1196
                for(int idxCell = nbCellsToSkip ; idxCell < totalNbCellsAtLevel ; ++idxCell){
                    myThreadkernels->L2L( iterArray[idxCell].getCurrentCell() , iterArray[idxCell].getCurrentChild(), idxLevel);
                }
            }
            FLOG(computationCounter.tac());

            sendBuffer.reset();
            recvBuffer.seek(0);
        }

        delete[] requests;
1197
        delete[] requestsSize;
1198 1199 1200 1201 1202

        FLOG( FLog::Controller << "\tFinished (@Downward Pass (L2L) = "  << counterTime.tacAndElapsed() << " s)\n" );
        FLOG( FLog::Controller << "\t\t Computation : " << computationCounter.cumulated() << " s\n" );
        FLOG( FLog::Controller << "\t\t Prepare : " << prepareCounter.cumulated() << " s\n" );
        FLOG( FLog::Controller << "\t\t Wait : " << waitCounter.cumulated() << " s\n" );
BRAMAS Berenger's avatar
BRAMAS Berenger committed
1203
        FLOG( FLog::Controller.flush());
1204 1205 1206 1207 1208 1209 1210
    }


    /////////////////////////////////////////////////////////////////////////////
    // Direct
    /////////////////////////////////////////////////////////////////////////////
    struct LeafData{
1211 1212 1213 1214
        FTreeCoordinate coord;
        CellClass* cell;
        ContainerClass* targets;
        ContainerClass* sources;
1215 1216 1217 1218
    };


    /** P2P */
1219
    void directPass(const bool p2pEnabled, const bool l2pEnabled){
1220 1221 1222 1223 1224 1225 1226 1227 1228 1229 1230 1231 1232 1233 1234
        FLOG( FLog::Controller.write("\tStart Direct Pass\n").write(FLog::Flush); );
        FLOG( FTic counterTime);
        FLOG( FTic prepareCounter);
        FLOG( FTic gatherCounter);
        FLOG( FTic waitCounter);
        FLOG(FTic computationCounter);
        FLOG(FTic computation2Counter);

        ///////////////////////////////////////////////////
        // Prepare data to send receive
        ///////////////////////////////////////////////////
        FLOG(prepareCounter.tic());

        FBoolArray leafsNeedOther(this->numberOfLeafs);
        int countNeedOther = 0;
PIACIBELLO Cyrille's avatar
PIACIBELLO Cyrille committed
1235

1236 1237
        // To store the result
        OctreeClass otherP2Ptree( tree->getHeight(), tree->getSubHeight(), tree->getBoxWidth(), tree->getBoxCenter() );
PIACIBELLO Cyrille's avatar
PIACIBELLO Cyrille committed
1238

1239
        // init
1240
        const int SizeShape = P2PExclusionClass::SizeShape;
PIACIBELLO Cyrille's avatar
PIACIBELLO Cyrille committed
1241

1242 1243
        int shapeLeaf[SizeShape];
        memset(shapeLeaf,0,SizeShape*sizeof(int));
PIACIBELLO Cyrille's avatar
PIACIBELLO Cyrille committed
1244

1245
        LeafData* const leafsDataArray = new LeafData[this->numberOfLeafs];
PIACIBELLO Cyrille's avatar
PIACIBELLO Cyrille committed
1246

1247
        FVector<LeafData> leafsNeedOtherData(countNeedOther);
PIACIBELLO Cyrille's avatar
PIACIBELLO Cyrille committed
1248

1249
        FVector<typename OctreeClass::Iterator>*const toSend = new FVector<typename OctreeClass::Iterator>[nbProcess];
1250 1251
        FSize partsToSend[nbProcess];
        memset(partsToSend, 0, sizeof(FSize) * nbProcess);
PIACIBELLO Cyrille's avatar
PIACIBELLO Cyrille committed
1252

1253
#pragma omp parallel num_threads(MaxThreads)
1254
        {
PIACIBELLO Cyrille's avatar
PIACIBELLO Cyrille committed
1255
#pragma omp master // MUST WAIT to fill leafsNeedOther
1256
            if(p2pEnabled){
1257 1258 1259 1260 1261 1262 1263 1264 1265 1266 1267 1268 1269 1270 1271 1272 1273 1274 1275 1276 1277 1278 1279 1280 1281 1282 1283 1284 1285 1286 1287 1288 1289 1290 1291 1292 1293 1294 1295 1296 1297 1298 1299 1300 1301 1302 1303 1304 1305 1306 1307 1308 1309 1310 1311
                // Copy leafs
                {
                    typename OctreeClass::Iterator octreeIterator(tree);
                    octreeIterator.gotoBottomLeft();
                    int idxLeaf = 0;
                    do{
                        this->iterArray[idxLeaf++] = octreeIterator;
                    } while(octreeIterator.moveRight());
                }

                int alreadySent[nbProcess];

                //Will store the indexes of the neighbors of current cell
                MortonIndex indexesNeighbors[26];

                for(int idxLeaf = 0 ; idxLeaf < this->numberOfLeafs ; ++idxLeaf){
                    memset(alreadySent, 0, sizeof(int) * nbProcess);
                    bool needOther = false;
                    //Get the neighbors of current cell in indexesNeighbors, and their number in neighCount
                    const int neighCount = (iterArray[idxLeaf].getCurrentGlobalCoordinate()).getNeighborsIndexes(OctreeHeight,indexesNeighbors);
                    //Loop over the neighbor leafs
                    for(int idxNeigh = 0 ; idxNeigh < neighCount ; ++idxNeigh){
                        //Test if leaf belongs to someone else (false if it's mine)
                        if(indexesNeighbors[idxNeigh] < (intervals[idProcess].leftIndex) || (intervals[idProcess].rightIndex) < indexesNeighbors[idxNeigh]){
                            needOther = true;

                            // find the proc that will need current leaf
                            int procToReceive = idProcess;
                            while( procToReceive != 0 && indexesNeighbors[idxNeigh] < intervals[procToReceive].leftIndex){
                                --procToReceive; //scroll process "before" current process
                            }

                            while( procToReceive != nbProcess - 1 && (intervals[procToReceive].rightIndex) < indexesNeighbors[idxNeigh]){
                                ++procToReceive;//scroll process "after" current process
                            }
                            //  Test : Not Already Send && be sure someone hold this interval
                            if( !alreadySent[procToReceive] && intervals[procToReceive].leftIndex <= indexesNeighbors[idxNeigh] && indexesNeighbors[idxNeigh] <= intervals[procToReceive].rightIndex){

                                alreadySent[procToReceive] = 1;
                                toSend[procToReceive].push( iterArray[idxLeaf] );
                                partsToSend[procToReceive] += iterArray[idxLeaf].getCurrentListSrc()->getSavedSize();
                                partsToSend[procToReceive] += int(sizeof(MortonIndex));
                            }
                        }
                    }

                    if(needOther){ //means that something need to be sent (or received)
                        leafsNeedOther.set(idxLeaf,true);
                        ++countNeedOther;
                    }
                }

                // No idea why it is mandatory there, could it be a few line before,
                for(int idxProc = 0 ; idxProc < nbProcess ; ++idxProc){
                    if(partsToSend[idxProc]){
1312
                        partsToSend[idxProc] += int(sizeof(FSize));