FGroupTaskStarpuMpiAlgorithm.hpp 118 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

47

48
template <class OctreeClass, class CellContainerClass, class KernelClass, class ParticleGroupClass, class StarPUCpuWrapperClass
49
#ifdef SCALFMM_ENABLE_CUDA_KERNEL
50
    , class StarPUCudaWrapperClass = FStarPUCudaWrapper<KernelClass, FCudaEmptyCellSymb, int, int, FCudaGroupOfCells<FCudaEmptyCellSymb, int, int>,
BRAMAS Berenger's avatar
BRAMAS Berenger committed
51
                                                        FCudaGroupOfParticles<int, 0, 0, int>, FCudaGroupAttachedLeaf<int, 0, 0, int>, FCudaEmptyKernel<int> >
52
#endif
53
#ifdef SCALFMM_ENABLE_OPENCL_KERNEL
54
    , class StarPUOpenClWrapperClass = FStarPUOpenClWrapper<KernelClass, FOpenCLDeviceWrapper<KernelClass>>
55 56
#endif
          >
57
class FGroupTaskStarPUMpiAlgorithm : public FAbstractAlgorithm {
58
protected:
59
    typedef FGroupTaskStarPUMpiAlgorithm<OctreeClass, CellContainerClass, KernelClass, ParticleGroupClass, StarPUCpuWrapperClass
60
#ifdef SCALFMM_ENABLE_CUDA_KERNEL
61
        , StarPUCudaWrapperClass
62
#endif
63
#ifdef SCALFMM_ENABLE_OPENCL_KERNEL
64
    , StarPUOpenClWrapperClass
65 66
#endif
    > ThisClass;
67

68
    int getTag(const int inLevel, const MortonIndex mindex, const int mode) const{
69 70 71
        int shift = 0;
        int height = tree->getHeight();
        while(height) { shift += 1; height >>= 1; }
BRAMAS Berenger's avatar
BRAMAS Berenger committed
72
        return int((((mindex<<shift) + inLevel) << 5) + mode);
73 74
    }

75 76 77 78 79 80 81 82 83
    const FMpi::FComm& comm;

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

84 85 86 87
    struct CellHandles{
        starpu_data_handle_t symb;
        starpu_data_handle_t up;
        starpu_data_handle_t down;
88
        int intervalSize;
89 90 91 92 93
    };

    struct ParticleHandles{
        starpu_data_handle_t symb;
        starpu_data_handle_t down;
94
        int intervalSize;
95 96
    };

97 98 99 100
    std::vector< std::vector< std::vector<BlockInteractions<CellContainerClass>>>> externalInteractionsAllLevel;
    std::vector< std::vector<BlockInteractions<ParticleGroupClass>>> externalInteractionsLeafLevel;

    OctreeClass*const tree;       //< The Tree
101
    KernelClass*const originalCpuKernel;
102

103 104
    std::vector<CellHandles>* cellHandles;
    std::vector<ParticleHandles> particleHandles;
105 106

    starpu_codelet p2m_cl;
BRAMAS Berenger's avatar
BRAMAS Berenger committed
107 108
    starpu_codelet m2m_cl;
    starpu_codelet l2l_cl;
BRAMAS Berenger's avatar
BRAMAS Berenger committed
109
    starpu_codelet l2l_cl_nocommute;
110 111 112 113
    starpu_codelet l2p_cl;

    starpu_codelet m2l_cl_in;
    starpu_codelet m2l_cl_inout;
114
    starpu_codelet m2l_cl_inout_mpi;
115 116 117

    starpu_codelet p2p_cl_in;
    starpu_codelet p2p_cl_inout;
118
    starpu_codelet p2p_cl_inout_mpi;
119

BRAMAS Berenger's avatar
BRAMAS Berenger committed
120 121 122 123 124 125 126

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

127
#ifdef STARPU_USE_CPU
128
    StarPUCpuWrapperClass cpuWrapper;
129
#endif
130
#ifdef SCALFMM_ENABLE_CUDA_KERNEL
131 132
    StarPUCudaWrapperClass cudaWrapper;
#endif
133
#ifdef SCALFMM_ENABLE_OPENCL_KERNEL
134 135
    StarPUOpenClWrapperClass openclWrapper;
#endif
136 137 138

    FStarPUPtrInterface wrappers;
    FStarPUPtrInterface* wrapperptr;
139

140
#ifdef STARPU_SUPPORT_ARBITER
BRAMAS Berenger's avatar
BRAMAS Berenger committed
141
    starpu_arbiter_t arbiterGlobal;
142
#endif
BRAMAS Berenger's avatar
BRAMAS Berenger committed
143 144 145 146 147 148 149 150 151 152 153 154

#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

155
public:
156 157
    FGroupTaskStarPUMpiAlgorithm(const FMpi::FComm& inComm, OctreeClass*const inTree, KernelClass* inKernels)
        :   comm(inComm), tree(inTree), originalCpuKernel(inKernels),
158
          cellHandles(nullptr),
159 160
#ifdef STARPU_USE_CPU
            cpuWrapper(tree->getHeight()),
161
#endif
162
#ifdef SCALFMM_ENABLE_CUDA_KERNEL
163 164
            cudaWrapper(tree->getHeight()),
#endif
165
#ifdef SCALFMM_ENABLE_OPENCL_KERNEL
166
            openclWrapper(tree->getHeight()),
167 168
#endif
            wrapperptr(&wrappers){
169 170 171
        FAssertLF(tree, "tree cannot be null");
        FAssertLF(inKernels, "kernels cannot be null");

172 173
        FAbstractAlgorithm::setNbLevelsInTree(tree->getHeight());

174 175
        struct starpu_conf conf;
        FAssertLF(starpu_conf_init(&conf) == 0);
176
        FStarPUFmmPriorities::Controller().init(&conf, tree->getHeight(), inKernels);
177
        FAssertLF(starpu_init(&conf) == 0);
178
        FAssertLF(starpu_mpi_init ( 0, 0, 0 ) == 0);
179

BRAMAS Berenger's avatar
BRAMAS Berenger committed
180 181
        starpu_malloc_set_align(32);

182 183
        starpu_pthread_mutex_t initMutex;
        starpu_pthread_mutex_init(&initMutex, NULL);
184
#ifdef STARPU_USE_CPU
185 186 187 188 189
        FStarPUUtils::ExecOnWorkers(STARPU_CPU, [&](){
            starpu_pthread_mutex_lock(&initMutex);
            cpuWrapper.initKernel(starpu_worker_get_id(), inKernels);
            starpu_pthread_mutex_unlock(&initMutex);
        });
190
        wrappers.set(FSTARPU_CPU_IDX, &cpuWrapper);
191
#endif
192
#ifdef SCALFMM_ENABLE_CUDA_KERNEL
193 194 195 196 197 198 199
        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
200
#ifdef SCALFMM_ENABLE_OPENCL_KERNEL
201 202 203 204 205 206
        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);
207
#endif
208 209
        starpu_pthread_mutex_destroy(&initMutex);

210 211
        starpu_pause();

212
        cellHandles   = new std::vector<CellHandles>[tree->getHeight()];
213

214
#ifdef STARPU_SUPPORT_ARBITER
BRAMAS Berenger's avatar
BRAMAS Berenger committed
215
        arbiterGlobal = starpu_arbiter_create();
216
#endif
217

BRAMAS Berenger's avatar
BRAMAS Berenger committed
218 219
        initCodelet();
        initCodeletMpi();
BRAMAS Berenger's avatar
BRAMAS Berenger committed
220
        rebuildInteractions();
BRAMAS Berenger's avatar
BRAMAS Berenger committed
221

BRAMAS Berenger's avatar
BRAMAS Berenger committed
222
        FLOG(FLog::Controller << "FGroupTaskStarPUAlgorithm (Max Worker " << starpu_worker_get_count() << ")\n");
BRAMAS Berenger's avatar
BRAMAS Berenger committed
223
#ifdef STARPU_USE_CPU
BRAMAS Berenger's avatar
BRAMAS Berenger committed
224
        FLOG(FLog::Controller << "FGroupTaskStarPUAlgorithm (Max CPU " << starpu_cpu_worker_get_count() << ")\n");
BRAMAS Berenger's avatar
BRAMAS Berenger committed
225
#endif
226
#ifdef SCALFMM_ENABLE_OPENCL_KERNEL
BRAMAS Berenger's avatar
BRAMAS Berenger committed
227 228
        FLOG(FLog::Controller << "FGroupTaskStarPUAlgorithm (Max OpenCL " << starpu_opencl_worker_get_count() << ")\n");
#endif
229
#ifdef SCALFMM_ENABLE_CUDA_KERNEL
BRAMAS Berenger's avatar
BRAMAS Berenger committed
230 231
        FLOG(FLog::Controller << "FGroupTaskStarPUAlgorithm (Max CUDA " << starpu_cuda_worker_get_count() << ")\n");
#endif
BRAMAS Berenger's avatar
BRAMAS Berenger committed
232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262

        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
263 264 265
    }

    ~FGroupTaskStarPUMpiAlgorithm(){
BRAMAS Berenger's avatar
BRAMAS Berenger committed
266 267
        starpu_resume();

268
        cleanHandle();
269
        cleanHandleMpi();
270
        delete[] cellHandles;
271

BRAMAS Berenger's avatar
BRAMAS Berenger committed
272 273 274 275 276 277 278 279 280 281
        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
282
#ifdef SCALFMM_ENABLE_CUDA_KERNEL
BRAMAS Berenger's avatar
BRAMAS Berenger committed
283 284 285 286 287 288 289
        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
290
#ifdef SCALFMM_ENABLE_OPENCL_KERNEL
BRAMAS Berenger's avatar
BRAMAS Berenger committed
291 292 293 294 295 296 297 298 299
        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);

300 301

#ifdef STARPU_SUPPORT_ARBITER
BRAMAS Berenger's avatar
BRAMAS Berenger committed
302
        starpu_arbiter_destroy(arbiterGlobal);
303
#endif
304
        starpu_mpi_shutdown();
305 306 307
        starpu_shutdown();
    }

BRAMAS Berenger's avatar
BRAMAS Berenger committed
308 309
    void rebuildInteractions(){
        setenv("OMP_WAIT_POLICY", "PASSIVE", 1);
310 311 312 313 314 315

        #pragma omp parallel
        #pragma omp single
        buildExternalInteractionVecs();
        buildHandles();

316 317
        #pragma omp parallel
        #pragma omp single
318 319
        buildRemoteInteractionsAndHandles();

BRAMAS Berenger's avatar
BRAMAS Berenger committed
320 321 322 323 324 325 326 327 328 329 330 331 332 333 334
        omp_set_num_threads(1);
    }

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

335
        starpu_resume();
336
        postRecvAllocatedBlocks();
337

338
        if( operationsToProceed & FFmmP2P ) insertParticlesSend();
339

340
        if(operationsToProceed & FFmmP2M && !directOnly) bottomPass();
341

342 343
        if(operationsToProceed & FFmmM2M && !directOnly) upwardPass();
        if(operationsToProceed & FFmmM2L && !directOnly) insertCellsSend();
344

BRAMAS Berenger's avatar
BRAMAS Berenger committed
345 346
        if(operationsToProceed & FFmmM2L && !directOnly) transferPass(FAbstractAlgorithm::upperWorkingLevel, FAbstractAlgorithm::lowerWorkingLevel-1 , true, true);
        if(operationsToProceed & FFmmM2L && !directOnly) transferPass(FAbstractAlgorithm::lowerWorkingLevel-1, FAbstractAlgorithm::lowerWorkingLevel, false, false);
347
        if(operationsToProceed & FFmmM2L && !directOnly) transferPassMpi();
348

349
        if(operationsToProceed & FFmmL2L && !directOnly) downardPass();
350

BRAMAS Berenger's avatar
BRAMAS Berenger committed
351 352
        if(operationsToProceed & FFmmM2L && !directOnly) transferPass(FAbstractAlgorithm::lowerWorkingLevel-1, FAbstractAlgorithm::lowerWorkingLevel, true, true);

353 354
        if( operationsToProceed & FFmmP2P ) directPass();
        if( operationsToProceed & FFmmP2P ) directPassMpi();
BRAMAS Berenger's avatar
BRAMAS Berenger committed
355

BRAMAS Berenger's avatar
BRAMAS Berenger committed
356 357 358 359
        if( operationsToProceed & FFmmL2P && !directOnly) mergePass();
#ifdef STARPU_USE_REDUX
        if( operationsToProceed & FFmmL2P && !directOnly) readParticle();
#endif
360 361 362

        starpu_task_wait_for_all();
        starpu_pause();
BRAMAS Berenger's avatar
BRAMAS Berenger committed
363 364 365 366 367

#ifdef STARPU_USE_CPU
        FTIME_TASKS(cpuWrapper.taskTimeRecorder.end());
        FTIME_TASKS(cpuWrapper.taskTimeRecorder.saveToDisk("/tmp/taskstime-FGroupTaskStarPUAlgorithm.txt"));
#endif
368 369
    }

370

371 372
    void initCodelet(){
        memset(&p2m_cl, 0, sizeof(p2m_cl));
373
#ifdef STARPU_USE_CPU
374
        if(originalCpuKernel->supportP2M(FSTARPU_CPU_IDX)){
375 376 377
            p2m_cl.cpu_funcs[0] = StarPUCpuWrapperClass::bottomPassCallback;
            p2m_cl.where |= STARPU_CPU;
        }
378
#endif
379
#ifdef SCALFMM_ENABLE_CUDA_KERNEL
380
        if(originalCpuKernel->supportP2M(FSTARPU_CUDA_IDX)){
381 382 383 384
            p2m_cl.cuda_funcs[0] = StarPUCudaWrapperClass::bottomPassCallback;
            p2m_cl.where |= STARPU_CUDA;
        }
#endif
385
#ifdef SCALFMM_ENABLE_OPENCL_KERNEL
386
        if(originalCpuKernel->supportP2M(FSTARPU_OPENCL_IDX)){
387 388 389
            p2m_cl.opencl_funcs[0] = StarPUOpenClWrapperClass::bottomPassCallback;
            p2m_cl.where |= STARPU_OPENCL;
        }
390
#endif
391 392 393 394
        p2m_cl.nbuffers = 3;
        p2m_cl.modes[0] = STARPU_R;
        p2m_cl.modes[1] = STARPU_RW;
        p2m_cl.modes[2] = STARPU_R;
395
        p2m_cl.name = "p2m_cl";
396

BRAMAS Berenger's avatar
BRAMAS Berenger committed
397
        memset(&m2m_cl, 0, sizeof(m2m_cl));
398
#ifdef STARPU_USE_CPU
BRAMAS Berenger's avatar
BRAMAS Berenger committed
399 400 401 402
        if(originalCpuKernel->supportM2M(FSTARPU_CPU_IDX)){
            m2m_cl.cpu_funcs[0] = StarPUCpuWrapperClass::upwardPassCallback;
            m2m_cl.where |= STARPU_CPU;
        }
403
#endif
404
#ifdef SCALFMM_ENABLE_CUDA_KERNEL
BRAMAS Berenger's avatar
BRAMAS Berenger committed
405 406 407 408
        if(originalCpuKernel->supportM2M(FSTARPU_CUDA_IDX)){
            m2m_cl.cuda_funcs[0] = StarPUCudaWrapperClass::upwardPassCallback;
            m2m_cl.where |= STARPU_CUDA;
        }
409
#endif
410
#ifdef SCALFMM_ENABLE_OPENCL_KERNEL
BRAMAS Berenger's avatar
BRAMAS Berenger committed
411 412 413 414
        if(originalCpuKernel->supportM2M(FSTARPU_OPENCL_IDX)){
            m2m_cl.opencl_funcs[0] = StarPUOpenClWrapperClass::upwardPassCallback;
            m2m_cl.where |= STARPU_OPENCL;
        }
415
#endif
BRAMAS Berenger's avatar
BRAMAS Berenger committed
416 417 418 419 420 421 422 423 424
        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));
425
#ifdef STARPU_USE_CPU
BRAMAS Berenger's avatar
BRAMAS Berenger committed
426 427 428 429
        if(originalCpuKernel->supportL2L(FSTARPU_CPU_IDX)){
            l2l_cl.cpu_funcs[0] = StarPUCpuWrapperClass::downardPassCallback;
            l2l_cl.where |= STARPU_CPU;
        }
430
#endif
431
#ifdef SCALFMM_ENABLE_CUDA_KERNEL
BRAMAS Berenger's avatar
BRAMAS Berenger committed
432 433 434 435
        if(originalCpuKernel->supportL2L(FSTARPU_CUDA_IDX)){
            l2l_cl.cuda_funcs[0] = StarPUCudaWrapperClass::downardPassCallback;
            l2l_cl.where |= STARPU_CUDA;
        }
436
#endif
437
#ifdef SCALFMM_ENABLE_OPENCL_KERNEL
BRAMAS Berenger's avatar
BRAMAS Berenger committed
438 439 440
        if(originalCpuKernel->supportL2L(FSTARPU_OPENCL_IDX)){
            l2l_cl.opencl_funcs[0] = StarPUOpenClWrapperClass::downardPassCallback;
            l2l_cl.where |= STARPU_OPENCL;
441
        }
BRAMAS Berenger's avatar
BRAMAS Berenger committed
442 443 444 445 446 447 448 449
#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);
450

BRAMAS Berenger's avatar
BRAMAS Berenger committed
451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 472 473 474 475 476 477
        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;

478
        memset(&l2p_cl, 0, sizeof(l2p_cl));
479
#ifdef STARPU_USE_CPU
480
        if(originalCpuKernel->supportL2P(FSTARPU_CPU_IDX)){
481 482 483
            l2p_cl.cpu_funcs[0] = StarPUCpuWrapperClass::mergePassCallback;
            l2p_cl.where |= STARPU_CPU;
        }
484
#endif
485
#ifdef SCALFMM_ENABLE_CUDA_KERNEL
486
        if(originalCpuKernel->supportL2P(FSTARPU_CUDA_IDX)){
487 488 489 490
            l2p_cl.cuda_funcs[0] = StarPUCudaWrapperClass::mergePassCallback;
            l2p_cl.where |= STARPU_CUDA;
        }
#endif
491
#ifdef SCALFMM_ENABLE_OPENCL_KERNEL
492
        if(originalCpuKernel->supportL2P(FSTARPU_OPENCL_IDX)){
493 494 495
            l2p_cl.opencl_funcs[0] = StarPUOpenClWrapperClass::mergePassCallback;
            l2p_cl.where |= STARPU_OPENCL;
        }
496
#endif
497
        l2p_cl.nbuffers = 4;
498
        l2p_cl.modes[0] = STARPU_R;
499 500
        l2p_cl.modes[1] = STARPU_R;
        l2p_cl.modes[2] = STARPU_R;
BRAMAS Berenger's avatar
BRAMAS Berenger committed
501 502 503
#ifdef STARPU_USE_REDUX
        l2p_cl.modes[3] = STARPU_REDUX;
#else
504
        l2p_cl.modes[3] = starpu_data_access_mode(STARPU_RW|STARPU_COMMUTE_IF_SUPPORTED);
BRAMAS Berenger's avatar
BRAMAS Berenger committed
505
#endif
506
        l2p_cl.name = "l2p_cl";
507 508

        memset(&p2p_cl_in, 0, sizeof(p2p_cl_in));
509
#ifdef STARPU_USE_CPU
510
        if(originalCpuKernel->supportP2P(FSTARPU_CPU_IDX)){
511 512 513
            p2p_cl_in.cpu_funcs[0] = StarPUCpuWrapperClass::directInPassCallback;
            p2p_cl_in.where |= STARPU_CPU;
        }
514
#endif
515
#ifdef SCALFMM_ENABLE_CUDA_KERNEL
516
        if(originalCpuKernel->supportP2P(FSTARPU_CUDA_IDX)){
517 518 519 520
            p2p_cl_in.cuda_funcs[0] = StarPUCudaWrapperClass::directInPassCallback;
            p2p_cl_in.where |= STARPU_CUDA;
        }
#endif
521
#ifdef SCALFMM_ENABLE_OPENCL_KERNEL
522
        if(originalCpuKernel->supportP2P(FSTARPU_OPENCL_IDX)){
523 524 525
            p2p_cl_in.opencl_funcs[0] = StarPUOpenClWrapperClass::directInPassCallback;
            p2p_cl_in.where |= STARPU_OPENCL;
        }
526
#endif
527 528
        p2p_cl_in.nbuffers = 2;
        p2p_cl_in.modes[0] = STARPU_R;
BRAMAS Berenger's avatar
BRAMAS Berenger committed
529 530 531
#ifdef STARPU_USE_REDUX
        p2p_cl_in.modes[1] = STARPU_REDUX;
#else
532
        p2p_cl_in.modes[1] = starpu_data_access_mode(STARPU_RW|STARPU_COMMUTE_IF_SUPPORTED);
BRAMAS Berenger's avatar
BRAMAS Berenger committed
533
#endif
534
        p2p_cl_in.name = "p2p_cl_in";
535
        memset(&p2p_cl_inout, 0, sizeof(p2p_cl_inout));
536
#ifdef STARPU_USE_CPU
537
        if(originalCpuKernel->supportP2PExtern(FSTARPU_CPU_IDX)){
538 539 540
            p2p_cl_inout.cpu_funcs[0] = StarPUCpuWrapperClass::directInoutPassCallback;
            p2p_cl_inout.where |= STARPU_CPU;
        }
541
#endif
542
#ifdef SCALFMM_ENABLE_CUDA_KERNEL
543
        if(originalCpuKernel->supportP2PExtern(FSTARPU_CUDA_IDX)){
544 545 546 547
            p2p_cl_inout.cuda_funcs[0] = StarPUCudaWrapperClass::directInoutPassCallback;
            p2p_cl_inout.where |= STARPU_CUDA;
        }
#endif
548
#ifdef SCALFMM_ENABLE_OPENCL_KERNEL
549
        if(originalCpuKernel->supportP2PExtern(FSTARPU_OPENCL_IDX)){
550 551 552
            p2p_cl_inout.opencl_funcs[0] = StarPUOpenClWrapperClass::directInoutPassCallback;
            p2p_cl_inout.where |= STARPU_OPENCL;
        }
553
#endif
554 555
        p2p_cl_inout.nbuffers = 4;
        p2p_cl_inout.modes[0] = STARPU_R;
BRAMAS Berenger's avatar
BRAMAS Berenger committed
556 557 558
#ifdef STARPU_USE_REDUX
        p2p_cl_inout.modes[1] = STARPU_REDUX;
#else
559
        p2p_cl_inout.modes[1] = starpu_data_access_mode(STARPU_RW|STARPU_COMMUTE_IF_SUPPORTED);
BRAMAS Berenger's avatar
BRAMAS Berenger committed
560
#endif
561
        p2p_cl_inout.modes[2] = STARPU_R;
BRAMAS Berenger's avatar
BRAMAS Berenger committed
562 563 564
#ifdef STARPU_USE_REDUX
        p2p_cl_inout.modes[3] = STARPU_REDUX;
#else
565
        p2p_cl_inout.modes[3] = starpu_data_access_mode(STARPU_RW|STARPU_COMMUTE_IF_SUPPORTED);
BRAMAS Berenger's avatar
BRAMAS Berenger committed
566
#endif
567
        p2p_cl_inout.name = "p2p_cl_inout";
568 569

        memset(&m2l_cl_in, 0, sizeof(m2l_cl_in));
570
#ifdef STARPU_USE_CPU
571
        if(originalCpuKernel->supportM2L(FSTARPU_CPU_IDX)){
572 573 574
            m2l_cl_in.cpu_funcs[0] = StarPUCpuWrapperClass::transferInPassCallback;
            m2l_cl_in.where |= STARPU_CPU;
        }
575
#endif
576
#ifdef SCALFMM_ENABLE_CUDA_KERNEL
577
        if(originalCpuKernel->supportM2L(FSTARPU_CUDA_IDX)){
578 579 580 581
            m2l_cl_in.cuda_funcs[0] = StarPUCudaWrapperClass::transferInPassCallback;
            m2l_cl_in.where |= STARPU_CUDA;
        }
#endif
582
#ifdef SCALFMM_ENABLE_OPENCL_KERNEL
583
        if(originalCpuKernel->supportM2L(FSTARPU_OPENCL_IDX)){
584 585 586
            m2l_cl_in.opencl_funcs[0] = StarPUOpenClWrapperClass::transferInPassCallback;
            m2l_cl_in.where |= STARPU_OPENCL;
        }
587
#endif
588 589
        m2l_cl_in.nbuffers = 3;
        m2l_cl_in.modes[0] = STARPU_R;
590
        m2l_cl_in.modes[1] = STARPU_R;
591
        m2l_cl_in.modes[2] = starpu_data_access_mode(STARPU_RW|STARPU_COMMUTE_IF_SUPPORTED);
592
        m2l_cl_in.name = "m2l_cl_in";
BRAMAS Berenger's avatar
BRAMAS Berenger committed
593

594
        memset(&m2l_cl_inout, 0, sizeof(m2l_cl_inout));
595
#ifdef STARPU_USE_CPU
596
        if(originalCpuKernel->supportM2LExtern(FSTARPU_CPU_IDX)){
597 598 599
            m2l_cl_inout.cpu_funcs[0] = StarPUCpuWrapperClass::transferInoutPassCallback;
            m2l_cl_inout.where |= STARPU_CPU;
        }
600
#endif
601
#ifdef SCALFMM_ENABLE_CUDA_KERNEL
602
        if(originalCpuKernel->supportM2LExtern(FSTARPU_CUDA_IDX)){
603 604 605 606
            m2l_cl_inout.cuda_funcs[0] = StarPUCudaWrapperClass::transferInoutPassCallback;
            m2l_cl_inout.where |= STARPU_CUDA;
        }
#endif
607
#ifdef SCALFMM_ENABLE_OPENCL_KERNEL
608
        if(originalCpuKernel->supportM2LExtern(FSTARPU_OPENCL_IDX)){
609 610 611
            m2l_cl_inout.opencl_funcs[0] = StarPUOpenClWrapperClass::transferInoutPassCallback;
            m2l_cl_inout.where |= STARPU_OPENCL;
        }
612
#endif
BRAMAS Berenger's avatar
BRAMAS Berenger committed
613
        m2l_cl_inout.nbuffers = 4;
614
        m2l_cl_inout.modes[0] = STARPU_R;
BRAMAS Berenger's avatar
BRAMAS Berenger committed
615 616
        m2l_cl_inout.modes[1] = starpu_data_access_mode(STARPU_RW|STARPU_COMMUTE_IF_SUPPORTED);
        m2l_cl_inout.modes[2] = STARPU_R;
617
        m2l_cl_inout.modes[3] = STARPU_R;
618
        m2l_cl_inout.name = "m2l_cl_inout";
619 620
    }

BRAMAS Berenger's avatar
BRAMAS Berenger committed
621 622 623 624 625 626 627 628 629 630 631 632 633 634 635 636 637 638 639 640 641 642 643 644 645 646 647 648 649 650 651 652 653 654 655 656 657 658 659 660 661 662 663 664
#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

665 666 667
    /** dealloc in a starpu way all the defined handles */
    void cleanHandle(){
        for(int idxLevel = 0 ; idxLevel < tree->getHeight() ; ++idxLevel){
668 669 670 671
            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);
672
            }
673
            cellHandles[idxLevel].clear();
674 675
        }
        {
676 677 678
            for(int idxHandle = 0 ; idxHandle < int(particleHandles.size()) ; ++idxHandle){
                starpu_data_unregister(particleHandles[idxHandle].symb);
                starpu_data_unregister(particleHandles[idxHandle].down);
679
            }
680
            particleHandles.clear();
681 682 683 684
        }
    }

    ////////////////////////////////////////////////////////////////////////////
685 686

    void initCodeletMpi(){
687
        memset(&p2p_cl_inout_mpi, 0, sizeof(p2p_cl_inout_mpi));
688
#ifdef STARPU_USE_CPU
689
        if(originalCpuKernel->supportP2PMpi(FSTARPU_CPU_IDX)){
690 691 692
            p2p_cl_inout_mpi.where |= STARPU_CPU;
            p2p_cl_inout_mpi.cpu_funcs[0] = StarPUCpuWrapperClass::directInoutPassCallbackMpi;
        }
693
#endif
694
#ifdef SCALFMM_ENABLE_CUDA_KERNEL
695
        if(originalCpuKernel->supportP2PMpi(FSTARPU_CUDA_IDX)){
696 697 698 699
            p2p_cl_inout_mpi.where |= STARPU_CUDA;
            p2p_cl_inout_mpi.cuda_funcs[0] = StarPUCudaWrapperClass::directInoutPassCallbackMpi;
        }
#endif
700
#ifdef SCALFMM_ENABLE_OPENCL_KERNEL
701
        if(originalCpuKernel->supportP2PMpi(FSTARPU_OPENCL_IDX)){
702 703 704
            p2p_cl_inout_mpi.where |= STARPU_OPENCL;
            p2p_cl_inout_mpi.opencl_funcs[0] = StarPUOpenClWrapperClass::directInoutPassCallbackMpi;
        }
705
#endif
706 707
        p2p_cl_inout_mpi.nbuffers = 3;
        p2p_cl_inout_mpi.modes[0] = STARPU_R;
708
        p2p_cl_inout_mpi.modes[1] = starpu_data_access_mode(STARPU_RW|STARPU_COMMUTE_IF_SUPPORTED);
709
        p2p_cl_inout_mpi.modes[2] = STARPU_R;
710
        p2p_cl_inout_mpi.name = "p2p_cl_inout_mpi";
711

712
        memset(&m2l_cl_inout_mpi, 0, sizeof(m2l_cl_inout_mpi));
713
#ifdef STARPU_USE_CPU
714
        if(originalCpuKernel->supportM2LMpi(FSTARPU_CPU_IDX)){
715 716 717
            m2l_cl_inout_mpi.where |= STARPU_CPU;
            m2l_cl_inout_mpi.cpu_funcs[0] = StarPUCpuWrapperClass::transferInoutPassCallbackMpi;
        }
718
#endif
719
#ifdef SCALFMM_ENABLE_CUDA_KERNEL
720
        if(originalCpuKernel->supportM2LMpi(FSTARPU_CUDA_IDX)){
721 722 723 724
            m2l_cl_inout_mpi.where |= STARPU_CUDA;
            m2l_cl_inout_mpi.cuda_funcs[0] = StarPUCudaWrapperClass::transferInoutPassCallbackMpi;
        }
#endif
725
#ifdef SCALFMM_ENABLE_OPENCL_KERNEL
726
        if(originalCpuKernel->supportM2LMpi(FSTARPU_OPENCL_IDX)){
727 728 729
            m2l_cl_inout_mpi.where |= STARPU_OPENCL;
            m2l_cl_inout_mpi.opencl_funcs[0] = StarPUOpenClWrapperClass::transferInoutPassCallbackMpi;
        }
730
#endif
731 732
        m2l_cl_inout_mpi.nbuffers = 4;
        m2l_cl_inout_mpi.modes[0] = STARPU_R;
733
        m2l_cl_inout_mpi.modes[1] = starpu_data_access_mode(STARPU_RW|STARPU_COMMUTE_IF_SUPPORTED);
734 735
        m2l_cl_inout_mpi.modes[2] = STARPU_R;
        m2l_cl_inout_mpi.modes[3] = STARPU_R;
736
        m2l_cl_inout_mpi.name = "m2l_cl_inout_mpi";
737 738
    }

739 740 741 742 743 744
    std::vector<std::pair<MortonIndex,MortonIndex>> processesIntervalPerLevels;
    struct BlockDescriptor{
        MortonIndex firstIndex;
        MortonIndex lastIndex;
        int owner;
        int bufferSize;
745 746 747 748
        size_t bufferSizeSymb;
        size_t bufferSizeUp;
        size_t bufferSizeDown;
        size_t leavesBufferSize;
749 750 751 752 753 754 755 756
    };
    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;

757
    struct RemoteHandle{
758 759 760 761
        RemoteHandle() : ptrSymb(nullptr), ptrUp(nullptr), ptrDown(nullptr){
            memset(&handleSymb, 0, sizeof(handleSymb));
            memset(&handleUp, 0, sizeof(handleUp));
            memset(&handleDown, 0, sizeof(handleDown));
762 763
        }

764 765 766 767 768 769
        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
770 771

        int intervalSize;
772 773 774 775 776
    };

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

777
    void buildRemoteInteractionsAndHandles(){
778 779
        cleanHandleMpi();

780 781
        // We need to have information about all other blocks
        std::unique_ptr<int[]> nbBlocksPerLevel(new int[tree->getHeight()]);
782
        nbBlocksPerLevel[0] = 0;
783 784 785 786 787 788 789 790 791 792 793 794 795 796 797 798 799 800 801 802 803 804 805 806 807 808 809 810 811 812 813 814 815 816 817 818 819 820 821 822 823 824 825
        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();
826 827 828
                myBlocksAtLevel[idxGroup].bufferSizeSymb = currentCells->getBufferSizeInByte();
                myBlocksAtLevel[idxGroup].bufferSizeUp   = currentCells->getMultipoleBufferSizeInByte();
                myBlocksAtLevel[idxGroup].bufferSizeDown = currentCells->getLocalBufferSizeInByte();
829 830 831 832 833 834 835 836 837 838 839 840 841

                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__);
        }
842 843 844 845 846 847 848
        // 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());

849 850 851 852 853 854 855 856
        // 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
857
        externalInteractionsAllLevelMpi.clear();
858 859 860 861 862 863 864 865 866 867 868 869 870 871
        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){
872
                CellContainerClass* currentCells = tree->getCellGroup(idxLevel, idxGroup);
873 874 875 876 877 878 879

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

                #pragma omp task default(none) firstprivate(idxGroup, currentCells, idxLevel, externalInteractions)
                {
                    std::vector<OutOfBlockInteraction> outsideInteractions;

880 881 882 883 884 885 886 887 888 889 890 891 892 893
                    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
894
                                property.relativeOutPosition = interactionsPosition[idxInter];
895 896
                                property.insideIdxInBlock = idxCell;
                                outsideInteractions.push_back(property);
897 898 899 900 901 902 903 904 905 906 907 908 909 910 911 912 913 914 915
                            }
                        }
                    }

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

                    int currentOutInteraction = 0;
                    for(int idxOtherGroup = 0 ; idxOtherGroup < int(processesBlockInfos[idxLevel].size())
                                                && currentOutInteraction < int(outsideInteractions.size()) ; ++idxOtherGroup){
                        // 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
916 917
                        const MortonIndex blockStartIdxOther = processesBlockInfos[idxLevel][idxOtherGroup].firstIndex;
                        const MortonIndex blockEndIdxOther   = processesBlockInfos[idxLevel][idxOtherGroup].lastIndex;
918

BRAMAS Berenger's avatar
BRAMAS Berenger committed
919
                        while(currentOutInteraction < int(outsideInteractions.size()) && outsideInteractions[currentOutInteraction].outIndex < blockStartIdxOther){
920 921 922 923
                            currentOutInteraction += 1;
                        }

                        int lastOutInteraction = currentOutInteraction;
BRAMAS Berenger's avatar
BRAMAS Berenger committed
924
                        while(lastOutInteraction < int(outsideInteractions.size()) && outsideInteractions[lastOutInteraction].outIndex < blockEndIdxOther){
925 926 927 928 929 930
                            lastOutInteraction += 1;
                        }

                        // Create interactions
                        const int nbInteractionsBetweenBlocks = (lastOutInteraction-currentOutInteraction);
                        if(nbInteractionsBetweenBlocks){
931
                            if(remoteCellGroups[idxLevel][idxOtherGroup].ptrSymb == nullptr){
932 933
                                #pragma omp critical(CreateM2LRemotes)
                                {
934
                                    if(remoteCellGroups[idxLevel][idxOtherGroup].ptrSymb == nullptr){
BRAMAS Berenger's avatar
BRAMAS Berenger committed
935
                                        const size_t nbBytesInBlockSymb = processesBlockInfos[idxLevel][idxOtherGroup].bufferSizeSymb;
936
                                        unsigned char* memoryBlockSymb = (unsigned char*)FAlignedMemory::AllocateBytes<32>(nbBytesInBlockSymb);
937 938 939
                                        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
940
                                        const size_t nbBytesInBlockUp = processesBlockInfos[idxLevel][idxOtherGroup].bufferSizeUp;
941
                                        unsigned char* memoryBlockUp = (unsigned char*)FAlignedMemory::AllocateBytes<32>(nbBytesInBlockUp);
942 943 944
                                        remoteCellGroups[idxLevel][idxOtherGroup].ptrUp = memoryBlockUp;
                                        starpu_variable_data_register(&remoteCellGroups[idxLevel][idxOtherGroup].handleUp, 0,
                                                                      (uintptr_t)remoteCellGroups[idxLevel][idxOtherGroup].ptrUp, nbBytesInBlockUp);
945 946
                                    }
                                }
947 948
                            }

949 950
                            externalInteractions->emplace_back();
                            BlockInteractions<CellContainerClass>* interactions = &externalInteractions->back();
951
                            //interactions->otherBlock = remoteCellGroups[idxLevel][idxOtherGroup].ptr;
952 953 954 955 956 957 958 959 960 961 962 963 964 965 966 967 968 969
                            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();

970
            externalInteractionsLeafLevelMpi.clear();
971 972 973 974 975 976 977 978 979 980 981
            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];

                #pragma omp task default(none) firstprivate(idxGroup, containers, externalInteractions)
                { // Can be a task(inout:iterCells)
                    std::vector<OutOfBlockInteraction> outsideInteractions;

982
                    for(int idxLeaf = 0 ; idxLeaf < containers->getNumberOfLeavesInBlock() ; ++idxLeaf){
983
                        // ParticleContainerClass particles = containers->template getLeaf<ParticleContainerClass>(mindex);
984
                        const MortonIndex mindex = containers->getLeafMortonIndex(idxLeaf);
985
                        if(containers->exists(mindex)){
986 987 988 989 990 991 992 993 994 995 996
                            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
997
                                    property.relativeOutPosition = interactionsPosition[idxInter];
998 999 1000 1001 1002 1003 1004 1005 1006 1007 1008 1009 1010 1011 1012 1013 1014 1015 1016 1017 1018
                                    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())
                                                && currentOutInteraction < int(outsideInteractions.size()) ; ++idxOtherGroup){
                        // 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
1019 1020
                        const MortonIndex blockStartIdxOther = processesBlockInfos[tree->getHeight()-1][idxOtherGroup].firstIndex;
                        const MortonIndex blockEndIdxOther   = processesBlockInfos[tree->getHeight()-1][idxOtherGroup].lastIndex;
1021

BRAMAS Berenger's avatar
BRAMAS Berenger committed
1022
                        while(currentOutInteraction < int(outsideInteractions.size()) && outsideInteractions[currentOutInteraction].outIndex < blockStartIdxOther){
1023 1024 1025 1026
                            currentOutInteraction += 1;
                        }

                        int lastOutInteraction = currentOutInteraction;
BRAMAS Berenger's avatar
BRAMAS Berenger committed
1027
                        while(lastOutInteraction < int(outsideInteractions.size()) && outsideInteractions[lastOutInteraction].outIndex < blockEndIdxOther){
1028 1029 1030 1031 1032
                            lastOutInteraction += 1;
                        }

                        const int nbInteractionsBetweenBlocks = (lastOutInteraction-currentOutInteraction);
                        if(nbInteractionsBetweenBlocks){
1033
                            if(remoteParticleGroupss[idxOtherGroup].ptrSymb == nullptr){
1034 1035
                                #pragma omp critical(CreateM2LRemotes)
                                {
1036
                                    if(remoteParticleGroupss[idxOtherGroup].ptrSymb == nullptr){
BRAMAS Berenger's avatar
BRAMAS Berenger committed
1037
                                        const size_t nbBytesInBlock = processesBlockInfos[tree->getHeight()-1][idxOtherGroup].leavesBufferSize;
1038
                                        unsigned char* memoryBlock = (unsigned char*)FAlignedMemory::AllocateBytes<32>(nbBytesInBlock);
1039 1040 1041
                                        remoteParticleGroupss[idxOtherGroup].ptrSymb = memoryBlock;
                                        starpu_variable_data_register(&remoteParticleGroupss[idxOtherGroup].handleSymb, 0,
                                                                      (uintptr_t)remoteParticleGroupss[idxOtherGroup].ptrSymb, nbBytesInBlock);
1042 1043
                                    }
                                }
1044 1045
                            }

1046 1047
                            externalInteractions->emplace_back();
                            BlockInteractions<ParticleGroupClass>* interactions = &externalInteractions->back();
1048
                            //interactions->otherBlock = remoteParticleGroupss[idxOtherGroup].ptr;
1049 1050 1051 1052 1053 1054 1055 1056 1057 1058 1059 1060 1061
                            interactions->otherBlockId = idxOtherGroup;
                            interactions->interactions.resize(nbInteractionsBetweenBlocks);
                            std::copy(outsideInteractions.begin() + currentOutInteraction,
                                      outsideInteractions.begin() + lastOutInteraction,
                                      interactions->interactions.begin());
                        }

                        currentOutInteraction = lastOutInteraction;
                    }
                }
            }
        }
    }
1062

1063 1064 1065 1066 1067 1068 1069 1070 1071 1072 1073
    struct MpiDependency{
        int src;
        int dest;
        int level;
        int globalBlockId;
    };

    std::vector<MpiDependency> toSend;

    void postRecvAllocatedBlocks(){
        std::vector<MpiDependency> toRecv;
1074
        FAssertLF(tree->getHeight() == int(remoteCellGroups.size()));
1075 1076
        const bool directOnly = (tree->getHeight() <= 2);
        if(!directOnly){
1077
            for(int idxLevel = FAbstractAlgorithm::upperWorkingLevel ; idxLevel < FAbstractAlgorithm::lowerWorkingLevel ; ++idxLevel){
1078 1079 1080 1081 1082 1083 1084 1085 1086 1087 1088 1089 1090 1091 1092 1093 1094 1095 1096 1097
                for(int idxHandle = 0 ; idxHandle < int(remoteCellGroups[idxLevel].size()) ; ++idxHandle){
                    if(remoteCellGroups[idxLevel][idxHandle].ptrSymb){
                        FAssertLF(remoteCellGroups[idxLevel][idxHandle].ptrUp);
                        FLOG(FLog::Controller << "[SMpi] " << idxLevel << " Post a recv during M2L for Idx " << processesBlockInfos[idxLevel][idxHandle].firstIndex <<
                             " and dest is " << processesBlockInfos[idxLevel][idxHandle].owner << " tag " << getTag(idxLevel,processesBlockInfos[idxLevel][idxHandle].firstIndex, 0) << "\n");
                        FLOG(FLog::Controller << "[SMpi] " << idxLevel << " Post a recv during M2L for Idx " << processesBlockInfos[idxLevel][idxHandle].firstIndex <<
                             " and dest is " << processesBlockInfos[idxLevel][idxHandle].owner << " tag " << getTag(idxLevel,processesBlockInfos[idxLevel][idxHandle].firstIndex, 1) << "\n");