FGroupTaskStarpuMpiAlgorithm.hpp 121 KB
Newer Older
1 2 3 4 5
// Keep in private GIT
// @SCALFMM_PRIVATE
#ifndef FGROUPTASKSTARPUMPIALGORITHM_HPP
#define FGROUPTASKSTARPUMPIALGORITHM_HPP

6 7 8 9 10 11 12 13 14 15 16
#include "../../Utils/FGlobal.hpp"
#include "../../Core/FCoreCommon.hpp"
#include "../../Utils/FQuickSort.hpp"
#include "../../Containers/FTreeCoordinate.hpp"
#include "../../Utils/FLog.hpp"
#include "../../Utils/FTic.hpp"
#include "../../Utils/FAssert.hpp"
#include "../../Utils/FAlignedMemory.hpp"
#include "../../Utils/FAssert.hpp"

#include "../../Utils/FMpi.hpp"
17

18 19
#include "FOutOfBlockInteraction.hpp"

20 21 22 23 24 25
#include <vector>
#include <memory>

#include <omp.h>

#include <starpu.h>
26
#include <starpu_mpi.h>
27
#include "../StarPUUtils/FStarPUUtils.hpp"
28
#include "../StarPUUtils/FStarPUFmmPriorities.hpp"
29

30
#ifdef STARPU_USE_CPU
31
#include "../StarPUUtils/FStarPUCpuWrapper.hpp"
32
#endif
33
#ifdef SCALFMM_ENABLE_CUDA_KERNEL
34 35 36 37 38
#include "../StarPUUtils/FStarPUCudaWrapper.hpp"
#include "../Cuda/FCudaEmptyKernel.hpp"
#include "../Cuda/FCudaGroupAttachedLeaf.hpp"
#include "../Cuda/FCudaGroupOfParticles.hpp"
#include "../Cuda/FCudaGroupOfCells.hpp"
39
#include "../Cuda/FCudaEmptyCellSymb.hpp"
40
#endif
41
#ifdef SCALFMM_ENABLE_OPENCL_KERNEL
42 43
#include "../StarPUUtils/FStarPUOpenClWrapper.hpp"
#include "../OpenCl/FOpenCLDeviceWrapper.hpp"
44
#include "../OpenCl/FEmptyOpenCLCode.hpp"
45
#endif
46

BRAMAS Berenger's avatar
BRAMAS Berenger committed
47 48
#include "../StarPUUtils/FStarPUReduxCpu.hpp"

49

50
template <class OctreeClass, class CellContainerClass, class KernelClass, class ParticleGroupClass, class StarPUCpuWrapperClass
BRAMAS Berenger's avatar
BRAMAS Berenger committed
51 52 53 54 55 56 57
          #ifdef SCALFMM_ENABLE_CUDA_KERNEL
          , class StarPUCudaWrapperClass = FStarPUCudaWrapper<KernelClass, FCudaEmptyCellSymb, int, int, FCudaGroupOfCells<FCudaEmptyCellSymb, int, int>,
                                                              FCudaGroupOfParticles<int, 0, 0, int>, FCudaGroupAttachedLeaf<int, 0, 0, int>, FCudaEmptyKernel<int> >
          #endif
          #ifdef SCALFMM_ENABLE_OPENCL_KERNEL
          , class StarPUOpenClWrapperClass = FStarPUOpenClWrapper<KernelClass, FOpenCLDeviceWrapper<KernelClass>>
          #endif
58
          >
59
class FGroupTaskStarPUMpiAlgorithm : public FAbstractAlgorithm {
60
protected:
61
    typedef FGroupTaskStarPUMpiAlgorithm<OctreeClass, CellContainerClass, KernelClass, ParticleGroupClass, StarPUCpuWrapperClass
62
#ifdef SCALFMM_ENABLE_CUDA_KERNEL
BRAMAS Berenger's avatar
BRAMAS Berenger committed
63
    , StarPUCudaWrapperClass
64
#endif
65
#ifdef SCALFMM_ENABLE_OPENCL_KERNEL
66
    , StarPUOpenClWrapperClass
67 68
#endif
    > ThisClass;
69

70
    int getTag(const int inLevel, const MortonIndex mindex, const int mode) const{
71 72 73
        int shift = 0;
        int height = tree->getHeight();
        while(height) { shift += 1; height >>= 1; }
BRAMAS Berenger's avatar
BRAMAS Berenger committed
74
        return int((((mindex<<shift) + inLevel) << 5) + mode);
75 76
    }

77 78 79 80 81 82 83 84 85
    const FMpi::FComm& comm;

    template <class OtherBlockClass>
    struct BlockInteractions{
        OtherBlockClass* otherBlock;
        int otherBlockId;
        std::vector<OutOfBlockInteraction> interactions;
    };

86 87 88 89
    struct CellHandles{
        starpu_data_handle_t symb;
        starpu_data_handle_t up;
        starpu_data_handle_t down;
90
        int intervalSize;
91 92 93 94 95
    };

    struct ParticleHandles{
        starpu_data_handle_t symb;
        starpu_data_handle_t down;
96
        int intervalSize;
97 98
    };

99 100 101 102
    std::vector< std::vector< std::vector<BlockInteractions<CellContainerClass>>>> externalInteractionsAllLevel;
    std::vector< std::vector<BlockInteractions<ParticleGroupClass>>> externalInteractionsLeafLevel;

    OctreeClass*const tree;       //< The Tree
103
    KernelClass*const originalCpuKernel;
104

105 106
    std::vector<CellHandles>* cellHandles;
    std::vector<ParticleHandles> particleHandles;
107 108

    starpu_codelet p2m_cl;
BRAMAS Berenger's avatar
BRAMAS Berenger committed
109 110
    starpu_codelet m2m_cl;
    starpu_codelet l2l_cl;
BRAMAS Berenger's avatar
BRAMAS Berenger committed
111
    starpu_codelet l2l_cl_nocommute;
112 113 114 115
    starpu_codelet l2p_cl;

    starpu_codelet m2l_cl_in;
    starpu_codelet m2l_cl_inout;
116
    starpu_codelet m2l_cl_inout_mpi;
117 118 119

    starpu_codelet p2p_cl_in;
    starpu_codelet p2p_cl_inout;
120
    starpu_codelet p2p_cl_inout_mpi;
121

BRAMAS Berenger's avatar
BRAMAS Berenger committed
122 123 124 125 126 127 128

#ifdef STARPU_USE_REDUX
    starpu_codelet p2p_redux_init;
    starpu_codelet p2p_redux_perform;
    starpu_codelet p2p_redux_read;
#endif

129
#ifdef STARPU_USE_CPU
130
    StarPUCpuWrapperClass cpuWrapper;
131
#endif
132
#ifdef SCALFMM_ENABLE_CUDA_KERNEL
133 134
    StarPUCudaWrapperClass cudaWrapper;
#endif
135
#ifdef SCALFMM_ENABLE_OPENCL_KERNEL
136 137
    StarPUOpenClWrapperClass openclWrapper;
#endif
138 139 140

    FStarPUPtrInterface wrappers;
    FStarPUPtrInterface* wrapperptr;
141

142
#ifdef STARPU_SUPPORT_ARBITER
BRAMAS Berenger's avatar
BRAMAS Berenger committed
143
    starpu_arbiter_t arbiterGlobal;
144
#endif
BRAMAS Berenger's avatar
BRAMAS Berenger committed
145 146 147 148 149 150 151 152 153 154 155 156

#ifdef STARPU_USE_TASK_NAME
    std::vector<std::unique_ptr<char[]>> m2mTaskNames;
    std::vector<std::unique_ptr<char[]>> m2lTaskNames;
    std::vector<std::unique_ptr<char[]>> m2lOuterTaskNames;
    std::vector<std::unique_ptr<char[]>> l2lTaskNames;
    std::unique_ptr<char[]> p2mTaskNames;
    std::unique_ptr<char[]> l2pTaskNames;
    std::unique_ptr<char[]> p2pTaskNames;
    std::unique_ptr<char[]> p2pOuterTaskNames;
#endif

157
public:
158 159
    FGroupTaskStarPUMpiAlgorithm(const FMpi::FComm& inComm, OctreeClass*const inTree, KernelClass* inKernels)
        :   comm(inComm), tree(inTree), originalCpuKernel(inKernels),
160
          cellHandles(nullptr),
BRAMAS Berenger's avatar
BRAMAS Berenger committed
161 162 163 164 165 166 167 168 169 170
      #ifdef STARPU_USE_CPU
          cpuWrapper(tree->getHeight()),
      #endif
      #ifdef SCALFMM_ENABLE_CUDA_KERNEL
          cudaWrapper(tree->getHeight()),
      #endif
      #ifdef SCALFMM_ENABLE_OPENCL_KERNEL
          openclWrapper(tree->getHeight()),
      #endif
          wrapperptr(&wrappers){
171 172 173
        FAssertLF(tree, "tree cannot be null");
        FAssertLF(inKernels, "kernels cannot be null");

174 175
        FAbstractAlgorithm::setNbLevelsInTree(tree->getHeight());

176 177
        struct starpu_conf conf;
        FAssertLF(starpu_conf_init(&conf) == 0);
178
        FStarPUFmmPriorities::Controller().init(&conf, tree->getHeight(), inKernels);
179
        FAssertLF(starpu_init(&conf) == 0);
180
        FAssertLF(starpu_mpi_init ( 0, 0, 0 ) == 0);
181

BRAMAS Berenger's avatar
BRAMAS Berenger committed
182 183
        starpu_malloc_set_align(32);

184 185
        starpu_pthread_mutex_t initMutex;
        starpu_pthread_mutex_init(&initMutex, NULL);
186
#ifdef STARPU_USE_CPU
187 188 189 190 191
        FStarPUUtils::ExecOnWorkers(STARPU_CPU, [&](){
            starpu_pthread_mutex_lock(&initMutex);
            cpuWrapper.initKernel(starpu_worker_get_id(), inKernels);
            starpu_pthread_mutex_unlock(&initMutex);
        });
192
        wrappers.set(FSTARPU_CPU_IDX, &cpuWrapper);
193
#endif
194
#ifdef SCALFMM_ENABLE_CUDA_KERNEL
195 196 197 198 199 200 201
        FStarPUUtils::ExecOnWorkers(STARPU_CUDA, [&](){
            starpu_pthread_mutex_lock(&initMutex);
            cudaWrapper.initKernel(starpu_worker_get_id(), inKernels);
            starpu_pthread_mutex_unlock(&initMutex);
        });
        wrappers.set(FSTARPU_CUDA_IDX, &cudaWrapper);
#endif
202
#ifdef SCALFMM_ENABLE_OPENCL_KERNEL
203 204 205 206 207 208
        FStarPUUtils::ExecOnWorkers(STARPU_OPENCL, [&](){
            starpu_pthread_mutex_lock(&initMutex);
            openclWrapper.initKernel(starpu_worker_get_id(), inKernels);
            starpu_pthread_mutex_unlock(&initMutex);
        });
        wrappers.set(FSTARPU_OPENCL_IDX, &openclWrapper);
209
#endif
210 211
        starpu_pthread_mutex_destroy(&initMutex);

212 213
        starpu_pause();

214
        cellHandles   = new std::vector<CellHandles>[tree->getHeight()];
215

216
#ifdef STARPU_SUPPORT_ARBITER
BRAMAS Berenger's avatar
BRAMAS Berenger committed
217
        arbiterGlobal = starpu_arbiter_create();
218
#endif
219

BRAMAS Berenger's avatar
BRAMAS Berenger committed
220 221
        initCodelet();
        initCodeletMpi();
BRAMAS Berenger's avatar
BRAMAS Berenger committed
222
        rebuildInteractions();
BRAMAS Berenger's avatar
BRAMAS Berenger committed
223

BRAMAS Berenger's avatar
BRAMAS Berenger committed
224
        FLOG(FLog::Controller << "FGroupTaskStarPUAlgorithm (Max Worker " << starpu_worker_get_count() << ")\n");
BRAMAS Berenger's avatar
BRAMAS Berenger committed
225
#ifdef STARPU_USE_CPU
BRAMAS Berenger's avatar
BRAMAS Berenger committed
226
        FLOG(FLog::Controller << "FGroupTaskStarPUAlgorithm (Max CPU " << starpu_cpu_worker_get_count() << ")\n");
BRAMAS Berenger's avatar
BRAMAS Berenger committed
227
#endif
228
#ifdef SCALFMM_ENABLE_OPENCL_KERNEL
BRAMAS Berenger's avatar
BRAMAS Berenger committed
229 230
        FLOG(FLog::Controller << "FGroupTaskStarPUAlgorithm (Max OpenCL " << starpu_opencl_worker_get_count() << ")\n");
#endif
231
#ifdef SCALFMM_ENABLE_CUDA_KERNEL
BRAMAS Berenger's avatar
BRAMAS Berenger committed
232 233
        FLOG(FLog::Controller << "FGroupTaskStarPUAlgorithm (Max CUDA " << starpu_cuda_worker_get_count() << ")\n");
#endif
BRAMAS Berenger's avatar
BRAMAS Berenger committed
234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264

        buildTaskNames();
    }

    void buildTaskNames(){
#ifdef STARPU_USE_TASK_NAME
        const int namesLength = 128;
        m2mTaskNames.resize(tree->getHeight());
        m2lTaskNames.resize(tree->getHeight());
        m2lOuterTaskNames.resize(tree->getHeight());
        l2lTaskNames.resize(tree->getHeight());
        for(int idxLevel = 0 ; idxLevel < tree->getHeight() ; ++idxLevel){
            m2mTaskNames[idxLevel].reset(new char[namesLength]);
            snprintf(m2mTaskNames[idxLevel].get(), namesLength, "M2M-level-%d", idxLevel);
            m2lTaskNames[idxLevel].reset(new char[namesLength]);
            snprintf(m2lTaskNames[idxLevel].get(), namesLength, "M2L-level-%d", idxLevel);
            m2lOuterTaskNames[idxLevel].reset(new char[namesLength]);
            snprintf(m2lOuterTaskNames[idxLevel].get(), namesLength, "M2L-out-level-%d", idxLevel);
            l2lTaskNames[idxLevel].reset(new char[namesLength]);
            snprintf(l2lTaskNames[idxLevel].get(), namesLength, "L2L-level-%d", idxLevel);
        }

        p2mTaskNames.reset(new char[namesLength]);
        snprintf(p2mTaskNames.get(), namesLength, "P2M");
        l2pTaskNames.reset(new char[namesLength]);
        snprintf(l2pTaskNames.get(), namesLength, "L2P");
        p2pTaskNames.reset(new char[namesLength]);
        snprintf(p2pTaskNames.get(), namesLength, "P2P");
        p2pOuterTaskNames.reset(new char[namesLength]);
        snprintf(p2pOuterTaskNames.get(), namesLength, "P2P-out");
#endif
265 266 267
    }

    ~FGroupTaskStarPUMpiAlgorithm(){
BRAMAS Berenger's avatar
BRAMAS Berenger committed
268 269
        starpu_resume();

270
        cleanHandle();
271
        cleanHandleMpi();
272
        delete[] cellHandles;
273

BRAMAS Berenger's avatar
BRAMAS Berenger committed
274 275 276 277 278 279 280 281 282 283
        starpu_pthread_mutex_t releaseMutex;
        starpu_pthread_mutex_init(&releaseMutex, NULL);
#ifdef STARPU_USE_CPU
        FStarPUUtils::ExecOnWorkers(STARPU_CPU, [&](){
            starpu_pthread_mutex_lock(&releaseMutex);
            cpuWrapper.releaseKernel(starpu_worker_get_id());
            starpu_pthread_mutex_unlock(&releaseMutex);
        });
        wrappers.set(FSTARPU_CPU_IDX, &cpuWrapper);
#endif
284
#ifdef SCALFMM_ENABLE_CUDA_KERNEL
BRAMAS Berenger's avatar
BRAMAS Berenger committed
285 286 287 288 289 290 291
        FStarPUUtils::ExecOnWorkers(STARPU_CUDA, [&](){
            starpu_pthread_mutex_lock(&releaseMutex);
            cudaWrapper.releaseKernel(starpu_worker_get_id());
            starpu_pthread_mutex_unlock(&releaseMutex);
        });
        wrappers.set(FSTARPU_CUDA_IDX, &cudaWrapper);
#endif
292
#ifdef SCALFMM_ENABLE_OPENCL_KERNEL
BRAMAS Berenger's avatar
BRAMAS Berenger committed
293 294 295 296 297 298 299 300 301
        FStarPUUtils::ExecOnWorkers(STARPU_OPENCL, [&](){
            starpu_pthread_mutex_lock(&releaseMutex);
            openclWrapper.releaseKernel(starpu_worker_get_id());
            starpu_pthread_mutex_unlock(&releaseMutex);
        });
        wrappers.set(FSTARPU_OPENCL_IDX, &openclWrapper);
#endif
        starpu_pthread_mutex_destroy(&releaseMutex);

302 303

#ifdef STARPU_SUPPORT_ARBITER
BRAMAS Berenger's avatar
BRAMAS Berenger committed
304
        starpu_arbiter_destroy(arbiterGlobal);
305
#endif
306
        starpu_mpi_shutdown();
307 308 309
        starpu_shutdown();
    }

BRAMAS Berenger's avatar
BRAMAS Berenger committed
310
    void rebuildInteractions(){
311 312
        FAssertLF(getenv("OMP_WAIT_POLICY") == nullptr
                || strcmp(getenv("OMP_WAIT_POLICY"), "PASSIVE") != 0);
313

BRAMAS Berenger's avatar
BRAMAS Berenger committed
314 315
#pragma omp parallel
#pragma omp single
316 317 318
        buildExternalInteractionVecs();
        buildHandles();

BRAMAS Berenger's avatar
BRAMAS Berenger committed
319 320
#pragma omp parallel
#pragma omp single
321
        buildRemoteInteractionsAndHandles();
BRAMAS Berenger's avatar
BRAMAS Berenger committed
322 323 324 325 326 327 328 329 330 331 332 333 334 335
    }

protected:
    /**
      * Runs the complete algorithm.
      */
    void executeCore(const unsigned operationsToProceed) override {
        FLOG( FLog::Controller << "\tStart FGroupTaskStarPUMpiAlgorithm\n" );
        const bool directOnly = (tree->getHeight() <= 2);

#ifdef STARPU_USE_CPU
        FTIME_TASKS(cpuWrapper.taskTimeRecorder.start());
#endif

336
        starpu_resume();
337
        postRecvAllocatedBlocks();
338

339
        if( operationsToProceed & FFmmP2P ) insertParticlesSend();
340

341
        if(operationsToProceed & FFmmP2M && !directOnly) bottomPass();
342

BRAMAS Berenger's avatar
BRAMAS Berenger committed
343
        if(operationsToProceed & FFmmM2M && !directOnly) upwardPass();
344
        if(operationsToProceed & FFmmM2L && !directOnly) insertCellsSend();
345

BRAMAS Berenger's avatar
BRAMAS Berenger committed
346 347 348
        if(operationsToProceed & FFmmM2L && !directOnly) transferPass(FAbstractAlgorithm::upperWorkingLevel, FAbstractAlgorithm::lowerWorkingLevel-1 , true, true);
        if(operationsToProceed & FFmmM2L && !directOnly) transferPass(FAbstractAlgorithm::lowerWorkingLevel-1, FAbstractAlgorithm::lowerWorkingLevel, false, false);
        if(operationsToProceed & FFmmM2L && !directOnly) transferPassMpi();
349

BRAMAS Berenger's avatar
BRAMAS Berenger committed
350
        if(operationsToProceed & FFmmL2L && !directOnly) downardPass();
351

BRAMAS Berenger's avatar
BRAMAS Berenger committed
352
        if(operationsToProceed & FFmmM2L && !directOnly) transferPass(FAbstractAlgorithm::lowerWorkingLevel-1, FAbstractAlgorithm::lowerWorkingLevel, true, true);
BRAMAS Berenger's avatar
BRAMAS Berenger committed
353

354 355
        if( operationsToProceed & FFmmP2P ) directPass();
        if( operationsToProceed & FFmmP2P ) directPassMpi();
BRAMAS Berenger's avatar
BRAMAS Berenger committed
356

BRAMAS Berenger's avatar
BRAMAS Berenger committed
357 358 359 360
        if( operationsToProceed & FFmmL2P && !directOnly) mergePass();
#ifdef STARPU_USE_REDUX
        if( operationsToProceed & FFmmL2P && !directOnly) readParticle();
#endif
361 362 363

        starpu_task_wait_for_all();
        starpu_pause();
BRAMAS Berenger's avatar
BRAMAS Berenger committed
364 365 366 367 368

#ifdef STARPU_USE_CPU
        FTIME_TASKS(cpuWrapper.taskTimeRecorder.end());
        FTIME_TASKS(cpuWrapper.taskTimeRecorder.saveToDisk("/tmp/taskstime-FGroupTaskStarPUAlgorithm.txt"));
#endif
369 370
    }

371

372 373
    void initCodelet(){
        memset(&p2m_cl, 0, sizeof(p2m_cl));
374
#ifdef STARPU_USE_CPU
375
        if(originalCpuKernel->supportP2M(FSTARPU_CPU_IDX)){
376 377 378
            p2m_cl.cpu_funcs[0] = StarPUCpuWrapperClass::bottomPassCallback;
            p2m_cl.where |= STARPU_CPU;
        }
379
#endif
380
#ifdef SCALFMM_ENABLE_CUDA_KERNEL
381
        if(originalCpuKernel->supportP2M(FSTARPU_CUDA_IDX)){
382 383 384 385
            p2m_cl.cuda_funcs[0] = StarPUCudaWrapperClass::bottomPassCallback;
            p2m_cl.where |= STARPU_CUDA;
        }
#endif
386
#ifdef SCALFMM_ENABLE_OPENCL_KERNEL
387
        if(originalCpuKernel->supportP2M(FSTARPU_OPENCL_IDX)){
388 389 390
            p2m_cl.opencl_funcs[0] = StarPUOpenClWrapperClass::bottomPassCallback;
            p2m_cl.where |= STARPU_OPENCL;
        }
391
#endif
392 393 394 395
        p2m_cl.nbuffers = 3;
        p2m_cl.modes[0] = STARPU_R;
        p2m_cl.modes[1] = STARPU_RW;
        p2m_cl.modes[2] = STARPU_R;
396
        p2m_cl.name = "p2m_cl";
397

BRAMAS Berenger's avatar
BRAMAS Berenger committed
398
        memset(&m2m_cl, 0, sizeof(m2m_cl));
399
#ifdef STARPU_USE_CPU
BRAMAS Berenger's avatar
BRAMAS Berenger committed
400 401 402 403
        if(originalCpuKernel->supportM2M(FSTARPU_CPU_IDX)){
            m2m_cl.cpu_funcs[0] = StarPUCpuWrapperClass::upwardPassCallback;
            m2m_cl.where |= STARPU_CPU;
        }
404
#endif
405
#ifdef SCALFMM_ENABLE_CUDA_KERNEL
BRAMAS Berenger's avatar
BRAMAS Berenger committed
406 407 408 409
        if(originalCpuKernel->supportM2M(FSTARPU_CUDA_IDX)){
            m2m_cl.cuda_funcs[0] = StarPUCudaWrapperClass::upwardPassCallback;
            m2m_cl.where |= STARPU_CUDA;
        }
410
#endif
411
#ifdef SCALFMM_ENABLE_OPENCL_KERNEL
BRAMAS Berenger's avatar
BRAMAS Berenger committed
412 413 414 415
        if(originalCpuKernel->supportM2M(FSTARPU_OPENCL_IDX)){
            m2m_cl.opencl_funcs[0] = StarPUOpenClWrapperClass::upwardPassCallback;
            m2m_cl.where |= STARPU_OPENCL;
        }
416
#endif
BRAMAS Berenger's avatar
BRAMAS Berenger committed
417 418 419 420 421 422 423 424 425
        m2m_cl.nbuffers = 4;
        m2m_cl.dyn_modes = (starpu_data_access_mode*)malloc(m2m_cl.nbuffers*sizeof(starpu_data_access_mode));
        m2m_cl.dyn_modes[0] = STARPU_R;
        m2m_cl.dyn_modes[1] = starpu_data_access_mode(STARPU_RW|STARPU_COMMUTE_IF_SUPPORTED);
        m2m_cl.name = "m2m_cl";
        m2m_cl.dyn_modes[2] = STARPU_R;
        m2m_cl.dyn_modes[3] = STARPU_R;

        memset(&l2l_cl, 0, sizeof(l2l_cl));
426
#ifdef STARPU_USE_CPU
BRAMAS Berenger's avatar
BRAMAS Berenger committed
427 428 429 430
        if(originalCpuKernel->supportL2L(FSTARPU_CPU_IDX)){
            l2l_cl.cpu_funcs[0] = StarPUCpuWrapperClass::downardPassCallback;
            l2l_cl.where |= STARPU_CPU;
        }
431
#endif
432
#ifdef SCALFMM_ENABLE_CUDA_KERNEL
BRAMAS Berenger's avatar
BRAMAS Berenger committed
433 434 435 436
        if(originalCpuKernel->supportL2L(FSTARPU_CUDA_IDX)){
            l2l_cl.cuda_funcs[0] = StarPUCudaWrapperClass::downardPassCallback;
            l2l_cl.where |= STARPU_CUDA;
        }
437
#endif
438
#ifdef SCALFMM_ENABLE_OPENCL_KERNEL
BRAMAS Berenger's avatar
BRAMAS Berenger committed
439 440 441
        if(originalCpuKernel->supportL2L(FSTARPU_OPENCL_IDX)){
            l2l_cl.opencl_funcs[0] = StarPUOpenClWrapperClass::downardPassCallback;
            l2l_cl.where |= STARPU_OPENCL;
442
        }
BRAMAS Berenger's avatar
BRAMAS Berenger committed
443 444 445 446 447 448 449 450
#endif
        l2l_cl.nbuffers = 4;
        l2l_cl.dyn_modes = (starpu_data_access_mode*)malloc(l2l_cl.nbuffers*sizeof(starpu_data_access_mode));
        l2l_cl.dyn_modes[0] = STARPU_R;
        l2l_cl.dyn_modes[1] = STARPU_R;
        l2l_cl.name = "l2l_cl";
        l2l_cl.dyn_modes[2] = STARPU_R;
        l2l_cl.dyn_modes[3] = starpu_data_access_mode(STARPU_RW|STARPU_COMMUTE_IF_SUPPORTED);
451

BRAMAS Berenger's avatar
BRAMAS Berenger committed
452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 472 473 474 475 476 477 478
        memset(&l2l_cl_nocommute, 0, sizeof(l2l_cl_nocommute));
#ifdef STARPU_USE_CPU
        if(originalCpuKernel->supportL2L(FSTARPU_CPU_IDX)){
            l2l_cl_nocommute.cpu_funcs[0] = StarPUCpuWrapperClass::downardPassCallback;
            l2l_cl_nocommute.where |= STARPU_CPU;
        }
#endif
#ifdef SCALFMM_ENABLE_CUDA_KERNEL
        if(originalCpuKernel->supportL2L(FSTARPU_CUDA_IDX)){
            l2l_cl_nocommute.cuda_funcs[0] = StarPUCudaWrapperClass::downardPassCallback;
            l2l_cl_nocommute.where |= STARPU_CUDA;
        }
#endif
#ifdef SCALFMM_ENABLE_OPENCL_KERNEL
        if(originalCpuKernel->supportL2L(FSTARPU_OPENCL_IDX)){
            l2l_cl_nocommute.opencl_funcs[0] = StarPUOpenClWrapperClass::downardPassCallback;
            l2l_cl_nocommute.where |= STARPU_OPENCL;
        }
#endif
        l2l_cl_nocommute.nbuffers = 4;
        l2l_cl_nocommute.dyn_modes = (starpu_data_access_mode*)malloc(l2l_cl_nocommute.nbuffers*sizeof(starpu_data_access_mode));
        l2l_cl_nocommute.dyn_modes[0] = STARPU_R;
        l2l_cl_nocommute.dyn_modes[1] = STARPU_R;
        l2l_cl_nocommute.name = "l2l_cl";
        l2l_cl_nocommute.dyn_modes[2] = STARPU_R;
        l2l_cl_nocommute.dyn_modes[3] = STARPU_RW;

479
        memset(&l2p_cl, 0, sizeof(l2p_cl));
480
#ifdef STARPU_USE_CPU
481
        if(originalCpuKernel->supportL2P(FSTARPU_CPU_IDX)){
482 483 484
            l2p_cl.cpu_funcs[0] = StarPUCpuWrapperClass::mergePassCallback;
            l2p_cl.where |= STARPU_CPU;
        }
485
#endif
486
#ifdef SCALFMM_ENABLE_CUDA_KERNEL
487
        if(originalCpuKernel->supportL2P(FSTARPU_CUDA_IDX)){
488 489 490 491
            l2p_cl.cuda_funcs[0] = StarPUCudaWrapperClass::mergePassCallback;
            l2p_cl.where |= STARPU_CUDA;
        }
#endif
492
#ifdef SCALFMM_ENABLE_OPENCL_KERNEL
493
        if(originalCpuKernel->supportL2P(FSTARPU_OPENCL_IDX)){
494 495 496
            l2p_cl.opencl_funcs[0] = StarPUOpenClWrapperClass::mergePassCallback;
            l2p_cl.where |= STARPU_OPENCL;
        }
497
#endif
498
        l2p_cl.nbuffers = 4;
499
        l2p_cl.modes[0] = STARPU_R;
500 501
        l2p_cl.modes[1] = STARPU_R;
        l2p_cl.modes[2] = STARPU_R;
BRAMAS Berenger's avatar
BRAMAS Berenger committed
502 503 504
#ifdef STARPU_USE_REDUX
        l2p_cl.modes[3] = STARPU_REDUX;
#else
505
        l2p_cl.modes[3] = starpu_data_access_mode(STARPU_RW|STARPU_COMMUTE_IF_SUPPORTED);
BRAMAS Berenger's avatar
BRAMAS Berenger committed
506
#endif
507
        l2p_cl.name = "l2p_cl";
508 509

        memset(&p2p_cl_in, 0, sizeof(p2p_cl_in));
510
#ifdef STARPU_USE_CPU
511
        if(originalCpuKernel->supportP2P(FSTARPU_CPU_IDX)){
512 513 514
            p2p_cl_in.cpu_funcs[0] = StarPUCpuWrapperClass::directInPassCallback;
            p2p_cl_in.where |= STARPU_CPU;
        }
515
#endif
516
#ifdef SCALFMM_ENABLE_CUDA_KERNEL
517
        if(originalCpuKernel->supportP2P(FSTARPU_CUDA_IDX)){
518 519 520 521
            p2p_cl_in.cuda_funcs[0] = StarPUCudaWrapperClass::directInPassCallback;
            p2p_cl_in.where |= STARPU_CUDA;
        }
#endif
522
#ifdef SCALFMM_ENABLE_OPENCL_KERNEL
523
        if(originalCpuKernel->supportP2P(FSTARPU_OPENCL_IDX)){
524 525 526
            p2p_cl_in.opencl_funcs[0] = StarPUOpenClWrapperClass::directInPassCallback;
            p2p_cl_in.where |= STARPU_OPENCL;
        }
527
#endif
528 529
        p2p_cl_in.nbuffers = 2;
        p2p_cl_in.modes[0] = STARPU_R;
BRAMAS Berenger's avatar
BRAMAS Berenger committed
530 531 532
#ifdef STARPU_USE_REDUX
        p2p_cl_in.modes[1] = STARPU_REDUX;
#else
533
        p2p_cl_in.modes[1] = starpu_data_access_mode(STARPU_RW|STARPU_COMMUTE_IF_SUPPORTED);
BRAMAS Berenger's avatar
BRAMAS Berenger committed
534
#endif
535
        p2p_cl_in.name = "p2p_cl_in";
536
        memset(&p2p_cl_inout, 0, sizeof(p2p_cl_inout));
537
#ifdef STARPU_USE_CPU
538
        if(originalCpuKernel->supportP2PExtern(FSTARPU_CPU_IDX)){
539 540 541
            p2p_cl_inout.cpu_funcs[0] = StarPUCpuWrapperClass::directInoutPassCallback;
            p2p_cl_inout.where |= STARPU_CPU;
        }
542
#endif
543
#ifdef SCALFMM_ENABLE_CUDA_KERNEL
544
        if(originalCpuKernel->supportP2PExtern(FSTARPU_CUDA_IDX)){
545 546 547 548
            p2p_cl_inout.cuda_funcs[0] = StarPUCudaWrapperClass::directInoutPassCallback;
            p2p_cl_inout.where |= STARPU_CUDA;
        }
#endif
549
#ifdef SCALFMM_ENABLE_OPENCL_KERNEL
550
        if(originalCpuKernel->supportP2PExtern(FSTARPU_OPENCL_IDX)){
551 552 553
            p2p_cl_inout.opencl_funcs[0] = StarPUOpenClWrapperClass::directInoutPassCallback;
            p2p_cl_inout.where |= STARPU_OPENCL;
        }
554
#endif
555 556
        p2p_cl_inout.nbuffers = 4;
        p2p_cl_inout.modes[0] = STARPU_R;
BRAMAS Berenger's avatar
BRAMAS Berenger committed
557 558 559
#ifdef STARPU_USE_REDUX
        p2p_cl_inout.modes[1] = STARPU_REDUX;
#else
560
        p2p_cl_inout.modes[1] = starpu_data_access_mode(STARPU_RW|STARPU_COMMUTE_IF_SUPPORTED);
BRAMAS Berenger's avatar
BRAMAS Berenger committed
561
#endif
562
        p2p_cl_inout.modes[2] = STARPU_R;
BRAMAS Berenger's avatar
BRAMAS Berenger committed
563 564 565
#ifdef STARPU_USE_REDUX
        p2p_cl_inout.modes[3] = STARPU_REDUX;
#else
566
        p2p_cl_inout.modes[3] = starpu_data_access_mode(STARPU_RW|STARPU_COMMUTE_IF_SUPPORTED);
BRAMAS Berenger's avatar
BRAMAS Berenger committed
567
#endif
568
        p2p_cl_inout.name = "p2p_cl_inout";
569 570

        memset(&m2l_cl_in, 0, sizeof(m2l_cl_in));
571
#ifdef STARPU_USE_CPU
572
        if(originalCpuKernel->supportM2L(FSTARPU_CPU_IDX)){
573 574 575
            m2l_cl_in.cpu_funcs[0] = StarPUCpuWrapperClass::transferInPassCallback;
            m2l_cl_in.where |= STARPU_CPU;
        }
576
#endif
577
#ifdef SCALFMM_ENABLE_CUDA_KERNEL
578
        if(originalCpuKernel->supportM2L(FSTARPU_CUDA_IDX)){
579 580 581 582
            m2l_cl_in.cuda_funcs[0] = StarPUCudaWrapperClass::transferInPassCallback;
            m2l_cl_in.where |= STARPU_CUDA;
        }
#endif
583
#ifdef SCALFMM_ENABLE_OPENCL_KERNEL
584
        if(originalCpuKernel->supportM2L(FSTARPU_OPENCL_IDX)){
585 586 587
            m2l_cl_in.opencl_funcs[0] = StarPUOpenClWrapperClass::transferInPassCallback;
            m2l_cl_in.where |= STARPU_OPENCL;
        }
588
#endif
589 590
        m2l_cl_in.nbuffers = 3;
        m2l_cl_in.modes[0] = STARPU_R;
591
        m2l_cl_in.modes[1] = STARPU_R;
592
        m2l_cl_in.modes[2] = starpu_data_access_mode(STARPU_RW|STARPU_COMMUTE_IF_SUPPORTED);
593
        m2l_cl_in.name = "m2l_cl_in";
BRAMAS Berenger's avatar
BRAMAS Berenger committed
594

595
        memset(&m2l_cl_inout, 0, sizeof(m2l_cl_inout));
596
#ifdef STARPU_USE_CPU
597
        if(originalCpuKernel->supportM2LExtern(FSTARPU_CPU_IDX)){
598 599 600
            m2l_cl_inout.cpu_funcs[0] = StarPUCpuWrapperClass::transferInoutPassCallback;
            m2l_cl_inout.where |= STARPU_CPU;
        }
601
#endif
602
#ifdef SCALFMM_ENABLE_CUDA_KERNEL
603
        if(originalCpuKernel->supportM2LExtern(FSTARPU_CUDA_IDX)){
604 605 606 607
            m2l_cl_inout.cuda_funcs[0] = StarPUCudaWrapperClass::transferInoutPassCallback;
            m2l_cl_inout.where |= STARPU_CUDA;
        }
#endif
608
#ifdef SCALFMM_ENABLE_OPENCL_KERNEL
609
        if(originalCpuKernel->supportM2LExtern(FSTARPU_OPENCL_IDX)){
610 611 612
            m2l_cl_inout.opencl_funcs[0] = StarPUOpenClWrapperClass::transferInoutPassCallback;
            m2l_cl_inout.where |= STARPU_OPENCL;
        }
613
#endif
BRAMAS Berenger's avatar
BRAMAS Berenger committed
614
        m2l_cl_inout.nbuffers = 4;
615
        m2l_cl_inout.modes[0] = STARPU_R;
BRAMAS Berenger's avatar
BRAMAS Berenger committed
616 617
        m2l_cl_inout.modes[1] = starpu_data_access_mode(STARPU_RW|STARPU_COMMUTE_IF_SUPPORTED);
        m2l_cl_inout.modes[2] = STARPU_R;
618
        m2l_cl_inout.modes[3] = STARPU_R;
619
        m2l_cl_inout.name = "m2l_cl_inout";
BRAMAS Berenger's avatar
BRAMAS Berenger committed
620

621

BRAMAS Berenger's avatar
BRAMAS Berenger committed
622 623 624 625 626 627 628 629 630 631 632 633 634 635 636 637 638 639 640 641 642 643 644 645 646 647 648 649 650 651 652 653 654 655 656 657 658 659 660 661 662 663 664
#ifdef STARPU_USE_REDUX
        memset(&p2p_redux_init, 0, sizeof(p2p_redux_init));
#ifdef STARPU_USE_CPU
        p2p_redux_init.cpu_funcs[0] = FStarPUReduxCpu::InitData<typename ParticleGroupClass::ParticleDataType>;
        p2p_redux_init.where |= STARPU_CPU;
#endif
        p2p_redux_init.nbuffers = 1;
        p2p_redux_init.modes[0] = STARPU_RW;
        p2p_redux_init.name = "p2p_redux_init";

        memset(&p2p_redux_perform, 0, sizeof(p2p_redux_perform));
#ifdef STARPU_USE_CPU
        p2p_redux_perform.cpu_funcs[0] = FStarPUReduxCpu::ReduceData<typename ParticleGroupClass::ParticleDataType>;
        p2p_redux_perform.where |= STARPU_CPU;
#endif
        p2p_redux_perform.nbuffers = 2;
        p2p_redux_perform.modes[0] = STARPU_RW;
        p2p_redux_perform.modes[1] = STARPU_R;
        p2p_redux_perform.name = "p2p_redux_perform";

        memset(&p2p_redux_read, 0, sizeof(p2p_redux_read));
#ifdef STARPU_USE_CPU
        if(originalCpuKernel->supportL2P(FSTARPU_CPU_IDX)){
            p2p_redux_read.cpu_funcs[0] = FStarPUReduxCpu::EmptyCodelet<typename ParticleGroupClass::ParticleDataType>;
            p2p_redux_read.where |= STARPU_CPU;
        }
#endif
#ifdef SCALFMM_ENABLE_CUDA_KERNEL
        if(originalCpuKernel->supportL2P(FSTARPU_CUDA_IDX)){
            p2p_redux_read.cuda_funcs[0] = FStarPUReduxCpu::EmptyCodelet<typename ParticleGroupClass::ParticleDataType>;
            p2p_redux_read.where |= STARPU_CUDA;
        }
#endif
#ifdef SCALFMM_ENABLE_OPENCL_KERNEL
        if(originalCpuKernel->supportL2P(FSTARPU_OPENCL_IDX)){
            p2p_redux_read.opencl_funcs[0] = FStarPUReduxCpu::EmptyCodelet<typename ParticleGroupClass::ParticleDataType>;
            p2p_redux_read.where |= STARPU_OPENCL;
        }
#endif
        p2p_redux_read.nbuffers = 1;
        p2p_redux_read.modes[0] = STARPU_R;
        p2p_redux_read.name = "p2p_redux_read";
#endif
BRAMAS Berenger's avatar
BRAMAS Berenger committed
665
    }
BRAMAS Berenger's avatar
BRAMAS Berenger committed
666

667 668 669
    /** dealloc in a starpu way all the defined handles */
    void cleanHandle(){
        for(int idxLevel = 0 ; idxLevel < tree->getHeight() ; ++idxLevel){
670 671 672 673
            for(int idxHandle = 0 ; idxHandle < int(cellHandles[idxLevel].size()) ; ++idxHandle){
                starpu_data_unregister(cellHandles[idxLevel][idxHandle].symb);
                starpu_data_unregister(cellHandles[idxLevel][idxHandle].up);
                starpu_data_unregister(cellHandles[idxLevel][idxHandle].down);
674
            }
675
            cellHandles[idxLevel].clear();
676 677
        }
        {
678 679 680
            for(int idxHandle = 0 ; idxHandle < int(particleHandles.size()) ; ++idxHandle){
                starpu_data_unregister(particleHandles[idxHandle].symb);
                starpu_data_unregister(particleHandles[idxHandle].down);
681
            }
682
            particleHandles.clear();
683 684 685 686
        }
    }

    ////////////////////////////////////////////////////////////////////////////
687 688

    void initCodeletMpi(){
689
        memset(&p2p_cl_inout_mpi, 0, sizeof(p2p_cl_inout_mpi));
690
#ifdef STARPU_USE_CPU
691
        if(originalCpuKernel->supportP2PMpi(FSTARPU_CPU_IDX)){
692 693 694
            p2p_cl_inout_mpi.where |= STARPU_CPU;
            p2p_cl_inout_mpi.cpu_funcs[0] = StarPUCpuWrapperClass::directInoutPassCallbackMpi;
        }
695
#endif
696
#ifdef SCALFMM_ENABLE_CUDA_KERNEL
697
        if(originalCpuKernel->supportP2PMpi(FSTARPU_CUDA_IDX)){
698 699 700 701
            p2p_cl_inout_mpi.where |= STARPU_CUDA;
            p2p_cl_inout_mpi.cuda_funcs[0] = StarPUCudaWrapperClass::directInoutPassCallbackMpi;
        }
#endif
702
#ifdef SCALFMM_ENABLE_OPENCL_KERNEL
703
        if(originalCpuKernel->supportP2PMpi(FSTARPU_OPENCL_IDX)){
704 705 706
            p2p_cl_inout_mpi.where |= STARPU_OPENCL;
            p2p_cl_inout_mpi.opencl_funcs[0] = StarPUOpenClWrapperClass::directInoutPassCallbackMpi;
        }
707
#endif
708 709
        p2p_cl_inout_mpi.nbuffers = 3;
        p2p_cl_inout_mpi.modes[0] = STARPU_R;
710
        p2p_cl_inout_mpi.modes[1] = starpu_data_access_mode(STARPU_RW|STARPU_COMMUTE_IF_SUPPORTED);
711
        p2p_cl_inout_mpi.modes[2] = STARPU_R;
712
        p2p_cl_inout_mpi.name = "p2p_cl_inout_mpi";
713

714
        memset(&m2l_cl_inout_mpi, 0, sizeof(m2l_cl_inout_mpi));
715
#ifdef STARPU_USE_CPU
716
        if(originalCpuKernel->supportM2LMpi(FSTARPU_CPU_IDX)){
717 718 719
            m2l_cl_inout_mpi.where |= STARPU_CPU;
            m2l_cl_inout_mpi.cpu_funcs[0] = StarPUCpuWrapperClass::transferInoutPassCallbackMpi;
        }
720
#endif
721
#ifdef SCALFMM_ENABLE_CUDA_KERNEL
722
        if(originalCpuKernel->supportM2LMpi(FSTARPU_CUDA_IDX)){
723 724 725 726
            m2l_cl_inout_mpi.where |= STARPU_CUDA;
            m2l_cl_inout_mpi.cuda_funcs[0] = StarPUCudaWrapperClass::transferInoutPassCallbackMpi;
        }
#endif
727
#ifdef SCALFMM_ENABLE_OPENCL_KERNEL
728
        if(originalCpuKernel->supportM2LMpi(FSTARPU_OPENCL_IDX)){
729 730 731
            m2l_cl_inout_mpi.where |= STARPU_OPENCL;
            m2l_cl_inout_mpi.opencl_funcs[0] = StarPUOpenClWrapperClass::transferInoutPassCallbackMpi;
        }
732
#endif
733 734
        m2l_cl_inout_mpi.nbuffers = 4;
        m2l_cl_inout_mpi.modes[0] = STARPU_R;
735
        m2l_cl_inout_mpi.modes[1] = starpu_data_access_mode(STARPU_RW|STARPU_COMMUTE_IF_SUPPORTED);
736 737
        m2l_cl_inout_mpi.modes[2] = STARPU_R;
        m2l_cl_inout_mpi.modes[3] = STARPU_R;
738
        m2l_cl_inout_mpi.name = "m2l_cl_inout_mpi";
739 740
    }

741 742 743 744 745 746
    std::vector<std::pair<MortonIndex,MortonIndex>> processesIntervalPerLevels;
    struct BlockDescriptor{
        MortonIndex firstIndex;
        MortonIndex lastIndex;
        int owner;
        int bufferSize;
747 748 749 750
        size_t bufferSizeSymb;
        size_t bufferSizeUp;
        size_t bufferSizeDown;
        size_t leavesBufferSize;
751 752 753 754 755 756 757 758
    };
    std::vector<std::vector<BlockDescriptor>> processesBlockInfos;
    std::vector<int> nbBlocksPerLevelAll;
    std::vector<int> nbBlocksBeforeMinPerLevel;

    std::vector< std::vector< std::vector<BlockInteractions<CellContainerClass>>>> externalInteractionsAllLevelMpi;
    std::vector< std::vector<BlockInteractions<ParticleGroupClass>>> externalInteractionsLeafLevelMpi;

759
    struct RemoteHandle{
760 761 762 763
        RemoteHandle() : ptrSymb(nullptr), ptrUp(nullptr), ptrDown(nullptr){
            memset(&handleSymb, 0, sizeof(handleSymb));
            memset(&handleUp, 0, sizeof(handleUp));
            memset(&handleDown, 0, sizeof(handleDown));
764 765
        }

766 767 768 769 770 771
        unsigned char * ptrSymb;
        starpu_data_handle_t handleSymb;
        unsigned char * ptrUp;
        starpu_data_handle_t handleUp;
        unsigned char * ptrDown;
        starpu_data_handle_t handleDown;
BRAMAS Berenger's avatar
BRAMAS Berenger committed
772 773

        int intervalSize;
774 775 776 777 778
    };

    std::vector<std::vector<RemoteHandle>> remoteCellGroups;
    std::vector<RemoteHandle> remoteParticleGroupss;

779
    void buildRemoteInteractionsAndHandles(){
780 781
        cleanHandleMpi();

782 783
        // We need to have information about all other blocks
        std::unique_ptr<int[]> nbBlocksPerLevel(new int[tree->getHeight()]);
784
        nbBlocksPerLevel[0] = 0;
785 786 787 788 789 790 791 792 793 794 795 796 797 798 799 800 801 802 803 804 805 806 807 808 809 810 811 812 813 814 815 816 817 818 819 820 821 822 823 824 825 826 827
        for(int idxLevel = 1 ; idxLevel < tree->getHeight() ; ++idxLevel){
            nbBlocksPerLevel[idxLevel] = tree->getNbCellGroupAtLevel(idxLevel);
        }
        // Exchange the number of blocks per proc
        nbBlocksPerLevelAll.resize(tree->getHeight() * comm.processCount());
        FMpi::Assert(MPI_Allgather(nbBlocksPerLevel.get(), tree->getHeight(), MPI_INT,
                                   nbBlocksPerLevelAll.data(), tree->getHeight(), MPI_INT,
                                   comm.getComm()), __LINE__);
        // Compute the number of blocks before mine
        nbBlocksBeforeMinPerLevel.resize(tree->getHeight());
        for(int idxLevel = 1 ; idxLevel < tree->getHeight() ; ++idxLevel){
            nbBlocksBeforeMinPerLevel[idxLevel] = 0;
            for(int idxProc = 0 ; idxProc < comm.processId() ; ++idxProc){
                nbBlocksBeforeMinPerLevel[idxLevel] += nbBlocksPerLevelAll[idxProc*tree->getHeight() + idxLevel];
            }
        }
        // Prepare the block infos
        processesBlockInfos.resize(tree->getHeight());
        std::unique_ptr<int[]> recvBlocksCount(new int[comm.processCount()]);
        std::unique_ptr<int[]> recvBlockDispl(new int[comm.processCount()]);
        // Exchange the block info per level
        for(int idxLevel = 1 ; idxLevel < tree->getHeight() ; ++idxLevel){
            // Count the total number of blocks
            int nbBlocksInLevel = 0;
            recvBlockDispl[0] = 0;
            for(int idxProc = 0 ; idxProc < comm.processCount() ; ++idxProc){
                nbBlocksInLevel += nbBlocksPerLevelAll[idxProc*tree->getHeight() + idxLevel];
                // Count and displacement for the MPI all gatherv
                recvBlocksCount[idxProc] = nbBlocksPerLevelAll[idxProc*tree->getHeight() + idxLevel] * int(sizeof(BlockDescriptor));
                if(idxProc) recvBlockDispl[idxProc] = recvBlockDispl[idxProc-1] + recvBlocksCount[idxProc-1];
            }
            processesBlockInfos[idxLevel].resize(nbBlocksInLevel);
            // Fill my blocks
            std::vector<BlockDescriptor> myBlocksAtLevel;
            myBlocksAtLevel.resize(nbBlocksPerLevel[idxLevel]);
            FAssertLF(tree->getNbCellGroupAtLevel(idxLevel) == int(myBlocksAtLevel.size()));
            FAssertLF(nbBlocksPerLevel[idxLevel] == nbBlocksPerLevelAll[comm.processId()*tree->getHeight() + idxLevel]);

            for(int idxGroup = 0 ; idxGroup < tree->getNbCellGroupAtLevel(idxLevel) ; ++idxGroup){
                CellContainerClass*const currentCells = tree->getCellGroup(idxLevel, idxGroup);
                myBlocksAtLevel[idxGroup].firstIndex = currentCells->getStartingIndex();
                myBlocksAtLevel[idxGroup].lastIndex  = currentCells->getEndingIndex();
                myBlocksAtLevel[idxGroup].owner = comm.processId();
828 829 830
                myBlocksAtLevel[idxGroup].bufferSizeSymb = currentCells->getBufferSizeInByte();
                myBlocksAtLevel[idxGroup].bufferSizeUp   = currentCells->getMultipoleBufferSizeInByte();
                myBlocksAtLevel[idxGroup].bufferSizeDown = currentCells->getLocalBufferSizeInByte();
831 832 833 834 835 836 837 838 839 840 841 842 843

                if(idxLevel == tree->getHeight() - 1){
                    myBlocksAtLevel[idxGroup].leavesBufferSize = tree->getParticleGroup(idxGroup)->getBufferSizeInByte();
                }
                else{
                    myBlocksAtLevel[idxGroup].leavesBufferSize = 0;
                }
            }
            // Exchange with all other
            FMpi::Assert(MPI_Allgatherv(myBlocksAtLevel.data(), int(myBlocksAtLevel.size()*sizeof(BlockDescriptor)), MPI_BYTE,
                                        processesBlockInfos[idxLevel].data(), recvBlocksCount.get(), recvBlockDispl.get(), MPI_BYTE,
                                        comm.getComm()), __LINE__);
        }
844 845 846 847 848 849 850
        // Prepare remate ptr and handles
        remoteCellGroups.resize( tree->getHeight() );
        for(int idxLevel = 1 ; idxLevel < tree->getHeight() ; ++idxLevel){
            remoteCellGroups[idxLevel].resize( processesBlockInfos[idxLevel].size());
        }
        remoteParticleGroupss.resize(processesBlockInfos[tree->getHeight()-1].size());

851 852 853 854 855 856 857 858
        // From now we have the number of blocks for all process
        // we also have the size of the blocks therefor we can
        // create the handles we need
        // We will now detect the relation between our blocks and others
        // During the M2M (which is the same for the L2L)
        // During the M2L and during the P2P
        // I need to insert the task that read my data or that write the data I need.
        // M2L
859
        externalInteractionsAllLevelMpi.clear();
860 861 862 863 864 865 866 867 868 869 870 871 872 873
        externalInteractionsAllLevelMpi.resize(tree->getHeight());
        for(int idxLevel = tree->getHeight()-1 ; idxLevel >= 2 ; --idxLevel){
            // From this level there are no more blocks
            if(tree->getNbCellGroupAtLevel(idxLevel) == 0){
                // We stop here
                break;
            }
            // What are my morton interval at this level
            const MortonIndex myFirstIndex = tree->getCellGroup(idxLevel, 0)->getStartingIndex();
            const MortonIndex myLastIndex = tree->getCellGroup(idxLevel, tree->getNbCellGroupAtLevel(idxLevel)-1)->getEndingIndex();

            externalInteractionsAllLevelMpi[idxLevel].resize(tree->getNbCellGroupAtLevel(idxLevel));

            for(int idxGroup = 0 ; idxGroup < tree->getNbCellGroupAtLevel(idxLevel) ; ++idxGroup){
874
                CellContainerClass* currentCells = tree->getCellGroup(idxLevel, idxGroup);
875 876 877

                std::vector<BlockInteractions<CellContainerClass>>* externalInteractions = &externalInteractionsAllLevelMpi[idxLevel][idxGroup];

BRAMAS Berenger's avatar
BRAMAS Berenger committed
878
#pragma omp task default(none) firstprivate(idxGroup, currentCells, idxLevel, externalInteractions)
879 880 881
                {
                    std::vector<OutOfBlockInteraction> outsideInteractions;

882 883 884 885 886 887 888 889 890 891 892 893 894 895
                    for(int idxCell = 0 ; idxCell < currentCells->getNumberOfCellsInBlock() ; ++idxCell){
                        const MortonIndex mindex = currentCells->getCellMortonIndex(idxCell);

                        MortonIndex interactionsIndexes[189];
                        int interactionsPosition[189];
                        const FTreeCoordinate coord(mindex, idxLevel);
                        int counter = coord.getInteractionNeighbors(idxLevel,interactionsIndexes,interactionsPosition);

                        for(int idxInter = 0 ; idxInter < counter ; ++idxInter){
                            // This interactions need a block owned by someoneelse
                            if(interactionsIndexes[idxInter] < myFirstIndex || myLastIndex <= interactionsIndexes[idxInter]){
                                OutOfBlockInteraction property;
                                property.insideIndex = mindex;
                                property.outIndex    = interactionsIndexes[idxInter];
BRAMAS Berenger's avatar
BRAMAS Berenger committed
896
                                property.relativeOutPosition = interactionsPosition[idxInter];
897
                                property.insideIdxInBlock = idxCell;
BRAMAS Berenger's avatar
BRAMAS Berenger committed
898
                                property.outsideIdxInBlock = -1;
899
                                outsideInteractions.push_back(property);
900 901 902 903 904 905 906 907 908
                            }
                        }
                    }

                    // Manage outofblock interaction
                    FQuickSort<OutOfBlockInteraction, int>::QsSequential(outsideInteractions.data(),int(outsideInteractions.size()));

                    int currentOutInteraction = 0;
                    for(int idxOtherGroup = 0 ; idxOtherGroup < int(processesBlockInfos[idxLevel].size())
BRAMAS Berenger's avatar
BRAMAS Berenger committed
909
                        && currentOutInteraction < int(outsideInteractions.size()) ; ++idxOtherGroup){
910 911 912 913 914 915 916 917 918
                        // Skip my blocks
                        if(idxOtherGroup == nbBlocksBeforeMinPerLevel[idxLevel]){
                            idxOtherGroup += tree->getNbCellGroupAtLevel(idxLevel);
                            if(idxOtherGroup == int(processesBlockInfos[idxLevel].size())){
                                break;
                            }
                            FAssertLF(idxOtherGroup < int(processesBlockInfos[idxLevel].size()));
                        }

BRAMAS Berenger's avatar
BRAMAS Berenger committed
919 920
                        const MortonIndex blockStartIdxOther = processesBlockInfos[idxLevel][idxOtherGroup].firstIndex;
                        const MortonIndex blockEndIdxOther   = processesBlockInfos[idxLevel][idxOtherGroup].lastIndex;
921

BRAMAS Berenger's avatar
BRAMAS Berenger committed
922
                        while(currentOutInteraction < int(outsideInteractions.size()) && outsideInteractions[currentOutInteraction].outIndex < blockStartIdxOther){
923 924 925 926
                            currentOutInteraction += 1;
                        }

                        int lastOutInteraction = currentOutInteraction;
BRAMAS Berenger's avatar
BRAMAS Berenger committed
927
                        while(lastOutInteraction < int(outsideInteractions.size()) && outsideInteractions[lastOutInteraction].outIndex < blockEndIdxOther){
928 929 930 931 932 933
                            lastOutInteraction += 1;
                        }

                        // Create interactions
                        const int nbInteractionsBetweenBlocks = (lastOutInteraction-currentOutInteraction);
                        if(nbInteractionsBetweenBlocks){
934
                            if(remoteCellGroups[idxLevel][idxOtherGroup].ptrSymb == nullptr){
BRAMAS Berenger's avatar
BRAMAS Berenger committed
935
#pragma omp critical(CreateM2LRemotes)
936
                                {
937
                                    if(remoteCellGroups[idxLevel][idxOtherGroup].ptrSymb == nullptr){
BRAMAS Berenger's avatar
BRAMAS Berenger committed
938
                                        const size_t nbBytesInBlockSymb = processesBlockInfos[idxLevel][idxOtherGroup].bufferSizeSymb;
939
                                        unsigned char* memoryBlockSymb = (unsigned char*)FAlignedMemory::AllocateBytes<32>(nbBytesInBlockSymb);
940 941 942
                                        remoteCellGroups[idxLevel][idxOtherGroup].ptrSymb = memoryBlockSymb;
                                        starpu_variable_data_register(&remoteCellGroups[idxLevel][idxOtherGroup].handleSymb, 0,
                                                                      (uintptr_t)remoteCellGroups[idxLevel][idxOtherGroup].ptrSymb, nbBytesInBlockSymb);
BRAMAS Berenger's avatar
BRAMAS Berenger committed
943
                                        const size_t nbBytesInBlockUp = processesBlockInfos[idxLevel][idxOtherGroup].bufferSizeUp;
944
                                        unsigned char* memoryBlockUp = (unsigned char*)FAlignedMemory::AllocateBytes<32>(nbBytesInBlockUp);
945 946 947
                                        remoteCellGroups[idxLevel][idxOtherGroup].ptrUp = memoryBlockUp;
                                        starpu_variable_data_register(&remoteCellGroups[idxLevel][idxOtherGroup].handleUp, 0,
                                                                      (uintptr_t)remoteCellGroups[idxLevel][idxOtherGroup].ptrUp, nbBytesInBlockUp);
948 949
                                    }
                                }
950 951
                            }

952 953
                            externalInteractions->emplace_back();
                            BlockInteractions<CellContainerClass>* interactions = &externalInteractions->back();
954
                            //interactions->otherBlock = remoteCellGroups[idxLevel][idxOtherGroup].ptr;
955 956 957 958 959 960 961 962 963 964 965 966 967 968 969 970 971 972
                            interactions->otherBlockId = idxOtherGroup;
                            interactions->interactions.resize(nbInteractionsBetweenBlocks);
                            std::copy(outsideInteractions.begin() + currentOutInteraction,
                                      outsideInteractions.begin() + lastOutInteraction,
                                      interactions->interactions.begin());
                        }

                        currentOutInteraction = lastOutInteraction;
                    }
                }
            }
        }
        // P2P
        // We create one big vector per block
        {
            const MortonIndex myFirstIndex = tree->getParticleGroup(0)->getStartingIndex();
            const MortonIndex myLastIndex = tree->getParticleGroup(tree->getNbParticleGroup()-1)->getEndingIndex();

973
            externalInteractionsLeafLevelMpi.clear();
974 975 976 977 978 979 980
            externalInteractionsLeafLevelMpi.resize(tree->getNbParticleGroup());
            for(int idxGroup = 0 ; idxGroup < tree->getNbParticleGroup() ; ++idxGroup){
                // Create the vector
                ParticleGroupClass* containers = tree->getParticleGroup(idxGroup);

                std::vector<BlockInteractions<ParticleGroupClass>>* externalInteractions = &externalInteractionsLeafLevelMpi[idxGroup];

BRAMAS Berenger's avatar
BRAMAS Berenger committed
981
#pragma omp task default(none) firstprivate(idxGroup, containers, externalInteractions)
982 983 984
                { // Can be a task(inout:iterCells)
                    std::vector<OutOfBlockInteraction> outsideInteractions;

985
                    for(int idxLeaf = 0 ; idxLeaf < containers->getNumberOfLeavesInBlock() ; ++idxLeaf){
986
                        // ParticleContainerClass particles = containers->template getLeaf<ParticleContainerClass>(mindex);
987
                        const MortonIndex mindex = containers->getLeafMortonIndex(idxLeaf);
988
                        if(containers->exists(mindex)){
989 990 991 992 993 994 995 996 997 998 999
                            MortonIndex interactionsIndexes[26];
                            int interactionsPosition[26];
                            FTreeCoordinate coord(mindex, tree->getHeight()-1);
                            int counter = coord.getNeighborsIndexes(tree->getHeight(),interactionsIndexes,interactionsPosition);

                            for(int idxInter = 0 ; idxInter < counter ; ++idxInter){
                                if(interactionsIndexes[idxInter] < myFirstIndex ||
                                        myLastIndex <= interactionsIndexes[idxInter]){
                                    OutOfBlockInteraction property;
                                    property.insideIndex = mindex;
                                    property.outIndex    = interactionsIndexes[idxInter];
BRAMAS Berenger's avatar
BRAMAS Berenger committed
1000
                                    property.relativeOutPosition = interactionsPosition[idxInter];
BRAMAS Berenger's avatar
BRAMAS Berenger committed
1001
                                    property.outsideIdxInBlock = -1;
BRAMAS Berenger's avatar
BRAMAS Berenger committed
1002
                                    property.insideIdxInBlock = idxLeaf;
1003 1004 1005 1006 1007 1008 1009 1010 1011 1012 1013
                                    outsideInteractions.push_back(property);
                                }
                            }
                        }
                    }

                    // Sort to match external order
                    FQuickSort<OutOfBlockInteraction, int>::QsSequential(outsideInteractions.data(),int(outsideInteractions.size()));

                    int currentOutInteraction = 0;
                    for(int idxOtherGroup = 0 ; idxOtherGroup < int(processesBlockInfos[tree->getHeight()-1].size())
BRAMAS Berenger's avatar
BRAMAS Berenger committed
1014
                        && currentOutInteraction < int(outsideInteractions.size()) ; ++idxOtherGroup){
1015 1016 1017 1018 1019 1020 1021 1022 1023
                        // Skip my blocks
                        if(idxOtherGroup == nbBlocksBeforeMinPerLevel[tree->getHeight()-1]){
                            idxOtherGroup += tree->getNbCellGroupAtLevel(tree->getHeight()-1);
                            if(idxOtherGroup == int(processesBlockInfos[tree->getHeight()-1].size())){
                                break;
                            }
                            FAssertLF(idxOtherGroup < int(processesBlockInfos[tree->getHeight()-1].size()));
                        }

BRAMAS Berenger's avatar
BRAMAS Berenger committed
1024 1025
                        const MortonIndex blockStartIdxOther = processesBlockInfos[tree->getHeight()-1][idxOtherGroup].firstIndex;
                        const MortonIndex blockEndIdxOther   = processesBlockInfos[tree->getHeight()-1][idxOtherGroup].lastIndex;
1026

BRAMAS Berenger's avatar
BRAMAS Berenger committed
1027
                        while(currentOutInteraction < int(outsideInteractions.size()) && outsideInteractions[currentOutInteraction].outIndex < blockStartIdxOther){
1028 1029 1030 1031
                            currentOutInteraction += 1;
                        }

                        int lastOutInteraction = currentOutInteraction;
BRAMAS Berenger's avatar
BRAMAS Berenger committed
1032
                        while(lastOutInteraction < int(outsideInteractions.size()) && outsideInteractions[lastOutInteraction].outIndex < blockEndIdxOther){
1033 1034 1035 1036 1037
                            lastOutInteraction += 1;
                        }

                        const int nbInteractionsBetweenBlocks = (lastOutInteraction-currentOutInteraction);
                        if(nbInteractionsBetweenBlocks){
1038
                            if(remoteParticleGroupss[idxOtherGroup].ptrSymb == nullptr){
BRAMAS Berenger's avatar
BRAMAS Berenger committed
1039
#pragma omp critical(CreateM2LRemotes)
1040
                                {
1041
                                    if(remoteParticleGroupss[idxOtherGroup].ptrSymb == nullptr){
BRAMAS Berenger's avatar
BRAMAS Berenger committed
1042
                                        const size_t nbBytesInBlock = processesBlockInfos[tree->getHeight()-1][idxOtherGroup].leavesBufferSize;
1043
                                        unsigned char* memoryBlock = (unsigned char*)FAlignedMemory::AllocateBytes<32>(nbBytesInBlock);
1044 1045 1046
                                        remoteParticleGroupss[idxOtherGroup].ptrSymb = memoryBlock;
                                        starpu_variable_data_register(&remoteParticleGroupss[idxOtherGroup].handleSymb, 0,
                                                                      (uintptr_t)remoteParticleGroupss[idxOtherGroup].ptrSymb, nbBytesInBlock);
1047 1048
                                    }
                                }
1049 1050
                            }

1051 1052
                            externalInteractions->emplace_back();
                            BlockInteractions<ParticleGroupClass>* interactions = &externalInteractions->back();
1053
                            //interactions->otherBlock = remoteParticleGroupss[idxOtherGroup].ptr;
1054 1055 1056 1057 1058 1059 1060 1061 1062 1063 1064 1065 1066
                            interactions->otherBlockId = idxOtherGroup;
                            interactions->interactions.resize(nbInteractionsBetweenBlocks);
                            std::copy(outsideInteractions.begin() + currentOutInteraction,
                                      outsideInteractions.begin() + lastOutInteraction,
                                      interactions->interactions.begin());
                        }

                        currentOutInteraction = lastOutInteraction;
                    }
                }
            }
        }
    }
1067

1068 1069 1070 1071 1072 1073 1074 1075 1076 1077 1078
    struct MpiDependency{
        int src;
        int dest;
        int level;
        int globalBlockId;
    };

    std::