p2p_distr_mpi.hpp 46.8 KB
Newer Older
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
#ifndef P2P_DISTR_MPI_HPP
#define P2P_DISTR_MPI_HPP

#include <mpi.h>

#include <vector>
#include <memory>
#include <cassert>

#include <type_traits>
#include <omp.h>
#include <algorithm>

#include "scope_timer.hpp"
#include "particles_utils.hpp"
#include "p2p_tree.hpp"

/*
- method to reorder each particle section following the cutoff grid (will permite index too)
- exchange particles (with upper only) and receive particle (from lower only)
- 1 message at a time! so need the offset of each cell of the cutoff grid
- iterate on what has been received with my own particles, fill both rhs
- send back the rhs
- merge rhs
- update particles property
  */

Berenger Bramas's avatar
Debug    
Berenger Bramas committed
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71


template <class partsize_t, class real_number, int size_particle_positions, int size_particle_rhs>
struct ParticleView{
    partsize_t p_index;
    real_number* ptr_particles_positions;
    real_number* ptr_particles_current_rhs;
    partsize_t* ptr_global_idx;
    long int* ptr_cell_idx;

    ParticleView()
        : p_index(-1), ptr_particles_positions(nullptr),
          ptr_particles_current_rhs(nullptr), ptr_global_idx(nullptr),
          ptr_cell_idx(nullptr){}
};

template <class partsize_t, class real_number, int size_particle_positions, int size_particle_rhs>
void swap(ParticleView<partsize_t, real_number, size_particle_positions,size_particle_rhs>& p1,
          ParticleView<partsize_t, real_number, size_particle_positions,size_particle_rhs>& p2){
    if(p1.p_index != -1 && p2.p_index != -1){
        for(int idx_pos = 0 ; idx_pos < size_particle_positions ; ++idx_pos){
            std::swap(p1.ptr_particles_positions[p1.p_index*size_particle_positions+idx_pos],
                      p2.ptr_particles_positions[p2.p_index*size_particle_positions+idx_pos]);
        }
        for(int idx_rhs = 0 ; idx_rhs < size_particle_rhs ; ++idx_rhs){
            std::swap(p1.ptr_particles_current_rhs[p1.p_index*size_particle_rhs+idx_rhs],
                      p2.ptr_particles_current_rhs[p2.p_index*size_particle_rhs+idx_rhs]);
        }
        std::swap(p1.ptr_cell_idx[p1.p_index],p2.ptr_cell_idx[p2.p_index]);
        std::swap(p1.ptr_global_idx[p1.p_index],p2.ptr_global_idx[p2.p_index]);
        std::swap(p1.p_index,p2.p_index);
    }
    else if(p1.p_index != -1){
        p2 = p1;
        p1 = ParticleView<partsize_t, real_number, size_particle_positions,size_particle_rhs>();
    }
    else if(p2.p_index != -1){
        p1 = p2;
        p2 = ParticleView<partsize_t, real_number, size_particle_positions,size_particle_rhs>();
    }
}



72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
template <class partsize_t, class real_number>
class p2p_distr_mpi {
protected:
    static const int MaxNbRhs = 100;

    enum MpiTag{
        TAG_NB_PARTICLES,
        TAG_POSITION_PARTICLES,
        TAG_RESULT_PARTICLES,
    };

    struct NeighborDescriptor{
        partsize_t nbParticlesToExchange;
        int destProc;
        int nbLevelsToExchange;
        bool isRecv;

        std::unique_ptr<real_number[]> toRecvAndMerge;
        std::unique_ptr<real_number[]> toCompute;
        std::unique_ptr<real_number[]> results;
    };

    enum Action{
        NOTHING_TODO,
        RECV_PARTICLES,
        COMPUTE_PARTICLES,
        RELEASE_BUFFER_PARTICLES,
        MERGE_PARTICLES,

        RECV_MOVE_NB_LOW,
        RECV_MOVE_NB_UP,
        RECV_MOVE_LOW,
        RECV_MOVE_UP
    };

    MPI_Comm current_com;

    int my_rank;
    int nb_processes;
    int nb_processes_involved;

    const std::pair<int,int> current_partition_interval;
    const int current_partition_size;
    const std::array<size_t,3> field_grid_dim;

    std::unique_ptr<int[]> partition_interval_size_per_proc;
    std::unique_ptr<int[]> partition_interval_offset_per_proc;

    std::unique_ptr<partsize_t[]> current_offset_particles_for_partition;

    std::vector<std::pair<Action,int>> whatNext;
    std::vector<MPI_Request> mpiRequests;
    std::vector<NeighborDescriptor> neigDescriptors;

    std::array<real_number,3> spatial_box_width;
    std::array<real_number,3> spatial_box_offset;

    const real_number cutoff_radius;
    std::array<long int,3> nb_cell_levels;

public:
    ////////////////////////////////////////////////////////////////////////////

    p2p_distr_mpi(MPI_Comm in_current_com,
                     const std::pair<int,int>& in_current_partitions,
                     const std::array<size_t,3>& in_field_grid_dim,
                     const std::array<real_number,3>& in_spatial_box_width,
                     const std::array<real_number,3>& in_spatial_box_offset,
                     const real_number in_cutoff_radius)
        : current_com(in_current_com),
            my_rank(-1), nb_processes(-1),nb_processes_involved(-1),
            current_partition_interval(in_current_partitions),
            current_partition_size(current_partition_interval.second-current_partition_interval.first),
            field_grid_dim(in_field_grid_dim),
            spatial_box_width(in_spatial_box_width), spatial_box_offset(in_spatial_box_offset),
            cutoff_radius(in_cutoff_radius){

        AssertMpi(MPI_Comm_rank(current_com, &my_rank));
        AssertMpi(MPI_Comm_size(current_com, &nb_processes));

        partition_interval_size_per_proc.reset(new int[nb_processes]);
        AssertMpi( MPI_Allgather( const_cast<int*>(&current_partition_size), 1, MPI_INT,
                                  partition_interval_size_per_proc.get(), 1, MPI_INT,
                                  current_com) );
        assert(partition_interval_size_per_proc[my_rank] == current_partition_size);

        partition_interval_offset_per_proc.reset(new int[nb_processes+1]);
        partition_interval_offset_per_proc[0] = 0;
        for(int idxProc = 0 ; idxProc < nb_processes ; ++idxProc){
            partition_interval_offset_per_proc[idxProc+1] = partition_interval_offset_per_proc[idxProc] + partition_interval_size_per_proc[idxProc];
        }

        current_offset_particles_for_partition.reset(new partsize_t[current_partition_size+1]);

        nb_processes_involved = nb_processes;
        while(nb_processes_involved != 0 && partition_interval_size_per_proc[nb_processes_involved-1] == 0){
            nb_processes_involved -= 1;
        }
        assert(nb_processes_involved != 0);
        for(int idx_proc_involved = 0 ; idx_proc_involved < nb_processes_involved ; ++idx_proc_involved){
            assert(partition_interval_size_per_proc[idx_proc_involved] != 0);
        }

        assert(int(field_grid_dim[IDX_Z]) == partition_interval_offset_per_proc[nb_processes_involved]);

        nb_cell_levels[IDX_X] = spatial_box_width[IDX_X]/cutoff_radius;
        nb_cell_levels[IDX_Y] = spatial_box_width[IDX_Y]/cutoff_radius;
        nb_cell_levels[IDX_Z] = spatial_box_width[IDX_Z]/cutoff_radius;
    }

    virtual ~p2p_distr_mpi(){}

    ////////////////////////////////////////////////////////////////////////////

    long int get_cell_coord_x_from_index(const long int index) const{
        return index % nb_cell_levels[IDX_X];
    }

    long int get_cell_coord_y_from_index(const long int index) const{
        return (index - get_cell_coord_z_from_index(index)*(nb_cell_levels[IDX_X]*nb_cell_levels[IDX_Y]))
                / nb_cell_levels[IDX_X];
    }

    long int get_cell_coord_z_from_index(const long int index) const{
        return index / (nb_cell_levels[IDX_X]*nb_cell_levels[IDX_Y]);
    }

    long int first_cell_level_proc(const int dest_proc) const{
        const real_number field_section_width_z = spatial_box_width[IDX_Z]/real_number(field_grid_dim[IDX_Z]);
        return static_cast<long int>((field_section_width_z*real_number(partition_interval_offset_per_proc[dest_proc]))/cutoff_radius);
    }

    long int last_cell_level_proc(const int dest_proc) const{
        const real_number field_section_width_z = spatial_box_width[IDX_Z]/real_number(field_grid_dim[IDX_Z]);
        return static_cast<long int>((field_section_width_z*real_number(partition_interval_offset_per_proc[dest_proc+1])
                                     - std::numeric_limits<real_number>::epsilon())/cutoff_radius);
    }

    std::array<long int,3> get_cell_coordinate(const real_number pos_x, const real_number pos_y,
                                               const real_number pos_z) const {
        const real_number diff_x = pos_x - spatial_box_offset[IDX_X];
        const real_number diff_y = pos_y - spatial_box_offset[IDX_Y];
        const real_number diff_z = pos_z - spatial_box_offset[IDX_Z];
        std::array<long int,3> coord;
        coord[IDX_X] = static_cast<long int>(diff_x/cutoff_radius);
        coord[IDX_Y] = static_cast<long int>(diff_y/cutoff_radius);
        coord[IDX_Z] = static_cast<long int>(diff_z/cutoff_radius);
        return coord;
    }

    long int get_cell_idx(const real_number pos_x, const real_number pos_y,
                          const real_number pos_z) const {
        std::array<long int,3> coord = get_cell_coordinate(pos_x, pos_y, pos_z);
        return ((coord[IDX_Z]*nb_cell_levels[IDX_Y])+coord[IDX_Y])*nb_cell_levels[IDX_X]+coord[IDX_X];
    }

    real_number compute_distance_r2(const real_number x1, const real_number y1, const real_number z1,
                                    const real_number x2, const real_number y2, const real_number z2) const {
Berenger Bramas's avatar
Debug    
Berenger Bramas committed
230
231
232
233
        real_number diff_x = std::abs(x1-x2);
        while(diff_x > spatial_box_width[IDX_X]/2){
            diff_x = std::abs(diff_x - spatial_box_width[IDX_X]);
        }
234

Berenger Bramas's avatar
Debug    
Berenger Bramas committed
235
236
237
238
        real_number diff_y = std::abs(y1-y2);
        while(diff_y > spatial_box_width[IDX_Y]/2){
            diff_y = std::abs(diff_y - spatial_box_width[IDX_Y]);
        }
239

Berenger Bramas's avatar
Debug    
Berenger Bramas committed
240
241
242
        real_number diff_z = std::abs(z1-z2);
        while(diff_z > spatial_box_width[IDX_Z]/2){
            diff_z = std::abs(diff_z - spatial_box_width[IDX_Z]);
243
        }
Berenger Bramas's avatar
Debug    
Berenger Bramas committed
244
245
246

        return (diff_x*diff_x) + (diff_y*diff_y) + (diff_z*diff_z);
    }
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283

    template <class computer_class, int size_particle_positions, int size_particle_rhs>
    void compute_distr(computer_class& in_computer,
                       const partsize_t current_my_nb_particles_per_partition[],
                       real_number particles_positions[],
                       real_number particles_current_rhs[],
                       partsize_t inout_index_particles[]){
        TIMEZONE("compute_distr");

        // Some processes might not be involved
        if(nb_processes_involved <= my_rank){
            return;
        }

        const long int my_top_z_cell_level = last_cell_level_proc(my_rank);
        const long int my_down_z_cell_level = first_cell_level_proc(my_rank);
        const long int my_nb_cell_levels = 1+my_top_z_cell_level-my_down_z_cell_level;

        current_offset_particles_for_partition[0] = 0;
        partsize_t myTotalNbParticles = 0;
        for(int idxPartition = 0 ; idxPartition < current_partition_size ; ++idxPartition){
            myTotalNbParticles += current_my_nb_particles_per_partition[idxPartition];
            current_offset_particles_for_partition[idxPartition+1] = current_offset_particles_for_partition[idxPartition] + current_my_nb_particles_per_partition[idxPartition];
        }

        // Compute box idx for each particle
        std::unique_ptr<long int[]> particles_coord(new long int[current_offset_particles_for_partition[current_partition_size]]);

        {
            for(int idxPartition = 0 ; idxPartition < current_partition_size ; ++idxPartition){
                #pragma omp parallel for schedule(static)
                for(partsize_t idxPart = current_offset_particles_for_partition[idxPartition] ; idxPart < current_offset_particles_for_partition[idxPartition+1] ; ++idxPart ){
                    particles_coord[idxPart] = get_cell_idx(particles_positions[(idxPart)*size_particle_positions + IDX_X],
                                                                              particles_positions[(idxPart)*size_particle_positions + IDX_Y],
                                                                              particles_positions[(idxPart)*size_particle_positions + IDX_Z]);
                    assert(my_down_z_cell_level <= get_cell_coord_z_from_index(particles_coord[idxPart]));
                    assert(get_cell_coord_z_from_index(particles_coord[idxPart]) <= my_top_z_cell_level);
Berenger Bramas's avatar
Debug    
Berenger Bramas committed
284
285
                    if(inout_index_particles[idxPart] == 58576 || inout_index_particles[idxPart] == 0){// TODO
                        printf("Coord index %ld - %ld (tree index %ld)\n", idxPart, inout_index_particles[idxPart],particles_coord[idxPart]);
286
287
288
                        printf(">> Box index %ld - %ld - %ld\n", get_cell_coord_x_from_index(particles_coord[idxPart]),
                               get_cell_coord_y_from_index(particles_coord[idxPart]),
                               get_cell_coord_z_from_index(particles_coord[idxPart]));
Berenger Bramas's avatar
Debug    
Berenger Bramas committed
289
                        printf("idxPartition %d\n", idxPartition);
290
                    }
291
292
293
                }
            }

Berenger Bramas's avatar
Debug    
Berenger Bramas committed
294
295
            using ParticleView_t = ParticleView<partsize_t, real_number, size_particle_positions,size_particle_rhs>;
            std::vector<ParticleView_t> part_to_sort;
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312

            // Sort each partition in cells
            for(int idxPartition = 0 ; idxPartition < current_partition_size ; ++idxPartition){
                part_to_sort.clear();

                for(partsize_t idxPart = current_offset_particles_for_partition[idxPartition] ; idxPart < current_offset_particles_for_partition[idxPartition+1] ; ++idxPart ){
                    part_to_sort.emplace_back();
                    part_to_sort.back().p_index = idxPart;
                    part_to_sort.back().ptr_particles_positions = particles_positions;
                    part_to_sort.back().ptr_particles_current_rhs = particles_current_rhs;
                    part_to_sort.back().ptr_global_idx = inout_index_particles;
                    part_to_sort.back().ptr_cell_idx = particles_coord.get();
                }

                assert(part_to_sort.size() == (current_offset_particles_for_partition[idxPartition+1]-current_offset_particles_for_partition[idxPartition]));

                std::sort(part_to_sort.begin(), part_to_sort.end(),
Berenger Bramas's avatar
Debug    
Berenger Bramas committed
313
314
315
316
                          [](const ParticleView_t& p1,
                             const ParticleView_t& p2){
                    assert(p1.p_index != -1 && p1.ptr_cell_idx);
                    assert(p2.p_index != -1 && p2.ptr_cell_idx);
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
                    return p1.ptr_cell_idx[p1.p_index] < p2.ptr_cell_idx[p2.p_index];
                });
            }
        }

        // Build the tree
        p2p_tree<std::vector<std::pair<partsize_t,partsize_t>>> my_tree(nb_cell_levels);

        for(int idxPartition = 0 ; idxPartition < current_partition_size ; ++idxPartition){
            long int current_cell_idx = -1;
            partsize_t current_nb_particles_in_cell = 0;
            partsize_t current_cell_offset = 0;

            for(partsize_t idx_part = current_offset_particles_for_partition[idxPartition] ;
                            idx_part != current_offset_particles_for_partition[idxPartition+1]; ++idx_part){
                if(particles_coord[idx_part] != current_cell_idx){
                    if(current_nb_particles_in_cell){
                        my_tree.getCell(current_cell_idx).emplace_back(current_cell_offset,current_nb_particles_in_cell);
                    }
                    current_cell_idx = particles_coord[idx_part];
                    current_nb_particles_in_cell = 1;
                    current_cell_offset = idx_part;
Berenger Bramas's avatar
Debug    
Berenger Bramas committed
339
340
341
342
343
344
345
                    if(inout_index_particles[idx_part] == 58576 || inout_index_particles[idx_part] == 0){// TODO
                        printf("Coord index %ld - %ld (tree index %ld)\n", idx_part, inout_index_particles[idx_part],particles_coord[idx_part]);
                        printf(">> Box index %ld - %ld - %ld\n", get_cell_coord_x_from_index(particles_coord[idx_part]),
                               get_cell_coord_y_from_index(particles_coord[idx_part]),
                               get_cell_coord_z_from_index(particles_coord[idx_part]));
                        printf("current_cell_offset %ld current_nb_particles_in_cell %ld\n", current_cell_offset, current_nb_particles_in_cell);
                    }
346
347
348
349
350
351
352
353
354
355
356
357
358
                }
            }
            if(current_nb_particles_in_cell){
                my_tree.getCell(current_cell_idx).emplace_back(current_cell_offset,current_nb_particles_in_cell);

            }
        }

        printf("[%d] go from cutoff level %ld to %ld\n",
               my_rank, my_down_z_cell_level, my_top_z_cell_level); // TODO remove
        fflush(stdout); // TODO

        // Offset per cell layers
Berenger Bramas's avatar
Debug    
Berenger Bramas committed
359
        long int previous_index = 0;
360
361
362
363
        std::unique_ptr<partsize_t[]> particles_offset_layers(new partsize_t[my_nb_cell_levels+1]());
        for(int idxPartition = 0 ; idxPartition < current_partition_size ; ++idxPartition){
            for(partsize_t idx_part = current_offset_particles_for_partition[idxPartition] ;
                            idx_part != current_offset_particles_for_partition[idxPartition+1]; ++idx_part){
Berenger Bramas's avatar
Debug    
Berenger Bramas committed
364
365
366
367
368
369
                const long int part_box_z_index = get_cell_coord_z_from_index(particles_coord[idx_part]);
                assert(my_down_z_cell_level <= part_box_z_index);
                assert(part_box_z_index <= my_top_z_cell_level);
                particles_offset_layers[part_box_z_index+1-my_down_z_cell_level] += 1;
                assert(previous_index <= part_box_z_index);
                previous_index = part_box_z_index;
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
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
478
479
480
481
482
483
484
485
486
487
488
489
490
491
            }
        }
        for(size_t idx_layer = 0 ; idx_layer < my_nb_cell_levels ; ++idx_layer){
            printf("[%d] nb particles in cutoff level %llu are %ld\n",
                   my_rank, idx_layer, particles_offset_layers[idx_layer+1]); // TODO remove
            fflush(stdout); // TODO
            particles_offset_layers[idx_layer+1] += particles_offset_layers[idx_layer];
        }

        // Reset vectors
        assert(whatNext.size() == 0);
        assert(mpiRequests.size() == 0);
        neigDescriptors.clear();

        // Find process with at least one neighbor
        {
            std::cout << my_rank << " my_top_z_cell_level " << my_top_z_cell_level << std::endl;
            std::cout << my_rank << " my_down_z_cell_level " << my_down_z_cell_level << std::endl;
            std::cout.flush();// TODO

            int dest_proc = (my_rank+1)%nb_processes_involved;
            while(dest_proc != my_rank
                  && (my_top_z_cell_level == first_cell_level_proc(dest_proc)
                      || (my_top_z_cell_level+1)%nb_cell_levels[IDX_Z] == first_cell_level_proc(dest_proc))){
                // Find if we have to send 1 or 2 cell levels
                int nb_levels_to_send = 1;
                if(my_nb_cell_levels > 1 // I have more than one level
                        && (my_top_z_cell_level-1+2)%nb_cell_levels[IDX_Z] <= last_cell_level_proc(dest_proc)){
                    nb_levels_to_send += 1;
                }

                std::cout << my_rank << " dest_proc " << dest_proc << std::endl;
                std::cout << my_rank << " first_cell_level_proc(dest_proc) " << first_cell_level_proc(dest_proc) << std::endl;
                std::cout << my_rank << " last_cell_level_proc(dest_proc) " << last_cell_level_proc(dest_proc) << std::endl;
                std::cout.flush();// TODO

                NeighborDescriptor descriptor;
                descriptor.destProc = dest_proc;
                descriptor.nbLevelsToExchange = nb_levels_to_send;
                descriptor.nbParticlesToExchange = particles_offset_layers[my_nb_cell_levels] - particles_offset_layers[my_nb_cell_levels-nb_levels_to_send];
                descriptor.isRecv = false;

                std::cout << my_rank << "SEND" << std::endl;
                std::cout << "descriptor.destProc " << descriptor.destProc << std::endl;
                std::cout << "descriptor.nbLevelsToExchange " << descriptor.nbLevelsToExchange << std::endl;
                std::cout << "descriptor.nbParticlesToExchange " << descriptor.nbParticlesToExchange << std::endl;
                std::cout << "descriptor.isRecv " << descriptor.isRecv << std::endl;
                std::cout << "neigDescriptors.size() " << neigDescriptors.size() << std::endl;
                std::cout.flush();// TODO

                neigDescriptors.emplace_back(std::move(descriptor));

                dest_proc = (dest_proc+1)%nb_processes_involved;
            }
            std::cout << my_rank << " NO dest_proc " << dest_proc << std::endl;
            std::cout << my_rank << " NO first_cell_level_proc(dest_proc) " << first_cell_level_proc(dest_proc) << std::endl;
            std::cout.flush();// TODO

            int src_proc = (my_rank-1+nb_processes_involved)%nb_processes_involved;
            while(src_proc != my_rank
                  && (last_cell_level_proc(src_proc) == my_down_z_cell_level
                      || (last_cell_level_proc(src_proc)+1)%nb_cell_levels[IDX_Z] == my_down_z_cell_level)){
                // Find if we have to send 1 or 2 cell levels
                int nb_levels_to_recv = 1;
                if(my_nb_cell_levels > 1 // I have more than one level
                        && first_cell_level_proc(src_proc) <= (my_down_z_cell_level-1+2)%nb_cell_levels[IDX_Z]){
                    nb_levels_to_recv += 1;
                }

                std::cout << my_rank << " src_proc " << src_proc << std::endl;
                std::cout << my_rank << " first_cell_level_proc(src_proc) " << first_cell_level_proc(src_proc) << std::endl;
                std::cout.flush();// TODO

                NeighborDescriptor descriptor;
                descriptor.destProc = src_proc;
                descriptor.nbLevelsToExchange = nb_levels_to_recv;
                descriptor.nbParticlesToExchange = -1;
                descriptor.isRecv = true;

                neigDescriptors.emplace_back(std::move(descriptor));

                std::cout << my_rank << "] RECV" << std::endl;
                std::cout << "descriptor.destProc " << descriptor.destProc << std::endl;
                std::cout << "descriptor.nbLevelsToExchange " << descriptor.nbLevelsToExchange << std::endl;
                std::cout << "descriptor.nbParticlesToExchange " << descriptor.nbParticlesToExchange << std::endl;
                std::cout << "descriptor.nbParticlesToExchange " << descriptor.nbParticlesToExchange << std::endl;
                std::cout << "descriptor.isRecv " << descriptor.isRecv << std::endl;
                std::cout << "neigDescriptors.size() " << neigDescriptors.size() << std::endl;
                std::cout.flush();// TODO

                src_proc = (src_proc-1+nb_processes_involved)%nb_processes_involved;
            }
            std::cout << my_rank << " NO src_proc " << src_proc << std::endl;
            std::cout << my_rank << " NO first_cell_level_proc(src_proc) " << first_cell_level_proc(src_proc) << std::endl;
            std::cout.flush();// TODO
        }

        //////////////////////////////////////////////////////////////////////
        /// Exchange the number of particles in each partition
        /// Could involve only here but I do not think it will be a problem
        //////////////////////////////////////////////////////////////////////

        assert(whatNext.size() == 0);
        assert(mpiRequests.size() == 0);


        for(int idxDescr = 0 ; idxDescr < int(neigDescriptors.size()) ; ++idxDescr){
            NeighborDescriptor& descriptor = neigDescriptors[idxDescr];

            if(descriptor.isRecv == false){
                whatNext.emplace_back(std::pair<Action,int>{NOTHING_TODO, -1});
                mpiRequests.emplace_back();
                AssertMpi(MPI_Isend(const_cast<partsize_t*>(&descriptor.nbParticlesToExchange),
                                    1, particles_utils::GetMpiType(partsize_t()),
                                    descriptor.destProc, TAG_NB_PARTICLES,
                                    current_com, &mpiRequests.back()));

                if(descriptor.nbParticlesToExchange){
                    std::cout << my_rank << "] SEND_PARTICLES" << std::endl;
                    std::cout << "descriptor.nbParticlesToExchange " << descriptor.nbParticlesToExchange << std::endl;
                    std::cout << "descriptor.destProc " << descriptor.destProc << std::endl;
                    std::cout << "idxDescr " << idxDescr << std::endl;
Berenger Bramas's avatar
Debug    
Berenger Bramas committed
492
                    std::cout << "send from part " << particles_offset_layers[my_nb_cell_levels-descriptor.nbLevelsToExchange] << std::endl;
493
494
495
496

                    whatNext.emplace_back(std::pair<Action,int>{NOTHING_TODO, -1});
                    mpiRequests.emplace_back();
                    assert(descriptor.nbParticlesToExchange*size_particle_positions < std::numeric_limits<int>::max());
Berenger Bramas's avatar
Debug    
Berenger Bramas committed
497
                    AssertMpi(MPI_Isend(const_cast<real_number*>(&particles_positions[particles_offset_layers[my_nb_cell_levels-descriptor.nbLevelsToExchange]*size_particle_positions]),
498
499
500
501
502
503
                              int(descriptor.nbParticlesToExchange*size_particle_positions), particles_utils::GetMpiType(real_number()),
                              descriptor.destProc, TAG_POSITION_PARTICLES,
                              current_com, &mpiRequests.back()));

                    assert(descriptor.toRecvAndMerge == nullptr);
                    descriptor.toRecvAndMerge.reset(new real_number[descriptor.nbParticlesToExchange*size_particle_rhs]);
Berenger Bramas's avatar
Debug    
Berenger Bramas committed
504
                    whatNext.emplace_back(std::pair<Action,int>{MERGE_PARTICLES, idxDescr});
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
                    mpiRequests.emplace_back();
                    assert(descriptor.nbParticlesToExchange*size_particle_rhs < std::numeric_limits<int>::max());
                    AssertMpi(MPI_Irecv(descriptor.toRecvAndMerge.get(), int(descriptor.nbParticlesToExchange*size_particle_rhs),
                                        particles_utils::GetMpiType(real_number()), descriptor.destProc, TAG_RESULT_PARTICLES,
                                        current_com, &mpiRequests.back()));
                }
            }
            else{
                std::cout << "RECV_PARTICLES " << RECV_PARTICLES << std::endl;
                std::cout << "idxDescr " << idxDescr << std::endl;
                whatNext.emplace_back(std::pair<Action,int>{RECV_PARTICLES, idxDescr});
                mpiRequests.emplace_back();
                AssertMpi(MPI_Irecv(&descriptor.nbParticlesToExchange,
                      1, particles_utils::GetMpiType(partsize_t()), descriptor.destProc, TAG_NB_PARTICLES,
                      current_com, &mpiRequests.back()));
            }
        }

        TIMEZONE_OMP_INIT_PREPARALLEL(omp_get_max_threads())
        #pragma omp parallel default(shared)
        {
            #pragma omp master
            {
                while(mpiRequests.size()){
                    assert(mpiRequests.size() == whatNext.size());

                    int idxDone = int(mpiRequests.size());
                    {
                        TIMEZONE("wait");
                        AssertMpi(MPI_Waitany(int(mpiRequests.size()), mpiRequests.data(), &idxDone, MPI_STATUSES_IGNORE));
                    }
                    const std::pair<Action, int> releasedAction = whatNext[idxDone];
                    std::swap(mpiRequests[idxDone], mpiRequests[mpiRequests.size()-1]);
                    std::swap(whatNext[idxDone], whatNext[mpiRequests.size()-1]);
                    mpiRequests.pop_back();
                    whatNext.pop_back();

                    //////////////////////////////////////////////////////////////////////
                    /// Data to exchange particles
                    //////////////////////////////////////////////////////////////////////
                    if(releasedAction.first == RECV_PARTICLES){
                        NeighborDescriptor& descriptor = neigDescriptors[releasedAction.second];
Berenger Bramas's avatar
Debug    
Berenger Bramas committed
547
                        assert(descriptor.isRecv);
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
572
573
574
                        const int destProc = descriptor.destProc;
                        const partsize_t NbParticlesToReceive = descriptor.nbParticlesToExchange;
                        assert(NbParticlesToReceive != -1);
                        assert(descriptor.toCompute == nullptr);

                        std::cout << my_rank << "] RECV_PARTICLES" << std::endl;
                        std::cout << "descriptor.nbParticlesToExchange " << descriptor.nbParticlesToExchange << std::endl;
                        std::cout << "descriptor.destProc " << descriptor.destProc << std::endl;
                        std::cout << "releasedAction.second " << releasedAction.second << std::endl;

                        if(NbParticlesToReceive){
                            std::cout << "MPI_Irecv " << std::endl;
                            descriptor.toCompute.reset(new real_number[NbParticlesToReceive*size_particle_positions]);
                            whatNext.emplace_back(std::pair<Action,int>{COMPUTE_PARTICLES, releasedAction.second});
                            mpiRequests.emplace_back();
                            assert(NbParticlesToReceive*size_particle_positions < std::numeric_limits<int>::max());
                            AssertMpi(MPI_Irecv(descriptor.toCompute.get(), int(NbParticlesToReceive*size_particle_positions),
                                                particles_utils::GetMpiType(real_number()), destProc, TAG_POSITION_PARTICLES,
                                                current_com, &mpiRequests.back()));
                        }
                    }

                    //////////////////////////////////////////////////////////////////////
                    /// Computation
                    //////////////////////////////////////////////////////////////////////
                    if(releasedAction.first == COMPUTE_PARTICLES){
                        NeighborDescriptor& descriptor = neigDescriptors[releasedAction.second];
Berenger Bramas's avatar
Debug    
Berenger Bramas committed
575
                        assert(descriptor.isRecv);
576
577
578
579
                        const partsize_t NbParticlesToReceive = descriptor.nbParticlesToExchange;

                        assert(descriptor.toCompute != nullptr);
                        descriptor.results.reset(new real_number[NbParticlesToReceive*size_particle_rhs]);
Berenger Bramas's avatar
Debug    
Berenger Bramas committed
580
                        in_computer.template init_result_array<size_particle_rhs>(descriptor.results.get(), NbParticlesToReceive);
581
582
583
584
585
586
587

                        // Compute
                        partsize_t idxPart = 0;
                        while(idxPart != NbParticlesToReceive){
                            const long int current_cell_idx = get_cell_idx(descriptor.toCompute[idxPart*size_particle_positions + IDX_X],
                                                                           descriptor.toCompute[idxPart*size_particle_positions + IDX_Y],
                                                                           descriptor.toCompute[idxPart*size_particle_positions + IDX_Z]);
Berenger Bramas's avatar
Debug    
Berenger Bramas committed
588
                            partsize_t nb_parts_in_cell = 1;
589
590
591
592
593
594
595
596
                            while(idxPart+nb_parts_in_cell != NbParticlesToReceive
                                  && current_cell_idx == get_cell_idx(descriptor.toCompute[(idxPart+nb_parts_in_cell)*size_particle_positions + IDX_X],
                                                                     descriptor.toCompute[(idxPart+nb_parts_in_cell)*size_particle_positions + IDX_Y],
                                                                     descriptor.toCompute[(idxPart+nb_parts_in_cell)*size_particle_positions + IDX_Z])){
                                nb_parts_in_cell += 1;
                            }

                            const std::vector<std::pair<partsize_t,partsize_t>>* neighbors[27];
597
598
                            long int neighbors_indexes[27];
                            const int nbNeighbors = my_tree.getNeighbors(current_cell_idx, neighbors, neighbors_indexes, true);
599

Berenger Bramas's avatar
Debug    
Berenger Bramas committed
600
601
602
603
604
605
606
607
608
609
610
611
612
613
614
615
616
                            for(int idx_test = 0 ; idx_test < nb_parts_in_cell ; ++idx_test){ // TODO
                                if(int(descriptor.toCompute[(idxPart+idx_test)*size_particle_positions + IDX_X]*1000) == int(1.685800e-01*1000)
                                        && int(descriptor.toCompute[(idxPart+idx_test)*size_particle_positions + IDX_Y]*1000) == int(7.524981e-01*1000)
                                        && int(descriptor.toCompute[(idxPart+idx_test)*size_particle_positions + IDX_Z]*1000) == int(9.999596e-01*1000)){
                                    printf("Found a pos %ld\n", idxPart+idx_test);
                                    printf("pos %e %e %e\n",
                                           descriptor.toCompute[(idxPart+idx_test)*size_particle_positions + IDX_X],
                                            descriptor.toCompute[(idxPart+idx_test)*size_particle_positions + IDX_Y],
                                            descriptor.toCompute[(idxPart+idx_test)*size_particle_positions + IDX_Z]);
                                }
                            }
                            printf("Remote part from %ld for %ld at idx %ld\n", idxPart, nb_parts_in_cell, current_cell_idx); // TODO
                            printf("pos of first %e %e %e\n", descriptor.toCompute[idxPart*size_particle_positions + IDX_X],
                                    descriptor.toCompute[idxPart*size_particle_positions + IDX_Y],
                                    descriptor.toCompute[idxPart*size_particle_positions + IDX_Z]); // TODO
                            printf("nbNeighbors %d\n", nbNeighbors); // TODO

617
618
619
620
621
622
623
624
625
626
627
628
                            // with other interval
                            for(size_t idx_neighbor = 0 ; idx_neighbor < nbNeighbors ; ++idx_neighbor){
                                for(size_t idx_2 = 0 ; idx_2 < (*neighbors[idx_neighbor]).size() ; ++idx_2){
                                    for(partsize_t idx_p1 = 0 ; idx_p1 < nb_parts_in_cell ; ++idx_p1){
                                        for(partsize_t idx_p2 = 0 ; idx_p2 < (*neighbors[idx_neighbor])[idx_2].second ; ++idx_p2){
                                            const real_number dist_r2 = compute_distance_r2(descriptor.toCompute[(idxPart+idx_p1)*size_particle_positions + IDX_X],
                                                                                            descriptor.toCompute[(idxPart+idx_p1)*size_particle_positions + IDX_Y],
                                                                                            descriptor.toCompute[(idxPart+idx_p1)*size_particle_positions + IDX_Z],
                                                                                            particles_positions[((*neighbors[idx_neighbor])[idx_2].first+idx_p2)*size_particle_positions + IDX_X],
                                                                                            particles_positions[((*neighbors[idx_neighbor])[idx_2].first+idx_p2)*size_particle_positions + IDX_Y],
                                                                                            particles_positions[((*neighbors[idx_neighbor])[idx_2].first+idx_p2)*size_particle_positions + IDX_Z]);
                                            if(dist_r2 < cutoff_radius*cutoff_radius){
Berenger Bramas's avatar
Debug    
Berenger Bramas committed
629
630
631
632
633
634
635
636
637
638
639
640
641
642
643
644
645
646
647
648
649
650
                                                in_computer.template compute_interaction<size_particle_positions,size_particle_rhs>(
                                                                    &descriptor.toCompute[(idxPart+idx_p1)*size_particle_positions],
                                                                    &descriptor.results[(idxPart+idx_p1)*size_particle_rhs],
                                                                    &particles_positions[((*neighbors[idx_neighbor])[idx_2].first+idx_p2)*size_particle_positions],
                                                                    &particles_current_rhs[((*neighbors[idx_neighbor])[idx_2].first+idx_p2)*size_particle_rhs],
                                                                    dist_r2);
                                            }

                                            if(inout_index_particles[(*neighbors[idx_neighbor])[idx_2].first+idx_p2] == 132){// TODO
                                                printf("test interaction between :\n");
                                                printf("index %ld (%ld) pos %e %e %e\n",
                                                       (idxPart+idx_p1), -1,
                                                       descriptor.toCompute[(idxPart+idx_p1)*size_particle_positions + IDX_X],
                                                       descriptor.toCompute[(idxPart+idx_p1)*size_particle_positions + IDX_Y],
                                                       descriptor.toCompute[(idxPart+idx_p1)*size_particle_positions + IDX_Z]);
                                                printf("index %ld (%ld) pos %e %e %e\n",
                                                       ((*neighbors[idx_neighbor])[idx_2].first+idx_p2),
                                                       inout_index_particles[((*neighbors[idx_neighbor])[idx_2].first+idx_p2)],
                                                       particles_positions[((*neighbors[idx_neighbor])[idx_2].first+idx_p2)*size_particle_positions + IDX_X],
                                                       particles_positions[((*neighbors[idx_neighbor])[idx_2].first+idx_p2)*size_particle_positions + IDX_Y],
                                                       particles_positions[((*neighbors[idx_neighbor])[idx_2].first+idx_p2)*size_particle_positions + IDX_Z]);
                                                printf("Radius = %e (%e)\n", sqrt(dist_r2), dist_r2);
651
652
653
654
655
656
657
658
659
660
661
662
663
664
665
666
667
668
669
670
671
672
673
674
                                            }
                                        }
                                    }
                                }
                            }

                            idxPart += nb_parts_in_cell;
                        }

                        // Send back
                        const int destProc = descriptor.destProc;
                        whatNext.emplace_back(std::pair<Action,int>{RELEASE_BUFFER_PARTICLES, releasedAction.second});
                        mpiRequests.emplace_back();
                        assert(NbParticlesToReceive*size_particle_rhs < std::numeric_limits<int>::max());
                        AssertMpi(MPI_Isend(descriptor.results.get(), int(NbParticlesToReceive*size_particle_rhs),
                                            particles_utils::GetMpiType(real_number()), destProc, TAG_RESULT_PARTICLES,
                                            current_com, &mpiRequests.back()));
                    }
                    //////////////////////////////////////////////////////////////////////
                    /// Computation
                    //////////////////////////////////////////////////////////////////////
                    if(releasedAction.first == RELEASE_BUFFER_PARTICLES){
                        NeighborDescriptor& descriptor = neigDescriptors[releasedAction.second];
                        assert(descriptor.toCompute != nullptr);
Berenger Bramas's avatar
Debug    
Berenger Bramas committed
675
                        assert(descriptor.isRecv);
676
677
678
679
680
                        descriptor.toCompute.release();
                    }
                    //////////////////////////////////////////////////////////////////////
                    /// Merge
                    //////////////////////////////////////////////////////////////////////
Berenger Bramas's avatar
Debug    
Berenger Bramas committed
681
                    if(releasedAction.first == MERGE_PARTICLES){
682
                        NeighborDescriptor& descriptor = neigDescriptors[releasedAction.second];
Berenger Bramas's avatar
Debug    
Berenger Bramas committed
683
                        assert(descriptor.isRecv == false);
684
                        assert(descriptor.toRecvAndMerge != nullptr);
Berenger Bramas's avatar
Debug    
Berenger Bramas committed
685
686
                        in_computer.template reduce_particles_rhs<size_particle_rhs>(&particles_current_rhs[particles_offset_layers[my_nb_cell_levels-descriptor.nbLevelsToExchange]*size_particle_rhs],
                                descriptor.toRecvAndMerge.get(), descriptor.nbParticlesToExchange);
687
688
689
690
691
692
693
694
695
696
697
698
699
700
701
702
703
704
705
706
707
708
709
710
711
712
713
714
715
716
717
718
719
720
721
722
723
724
725
726
727
728
729
730
731
732
733
734
735
736
737
738
739
740
741
742
743
744
745
746
                        descriptor.toRecvAndMerge.release();
                    }
                }
            }
        }

        assert(whatNext.size() == 0);
        assert(mpiRequests.size() == 0);

        // Compute self data
        for(const auto& iter_cell : my_tree){
            const std::vector<std::pair<partsize_t,partsize_t>>& intervals = iter_cell.second;

            for(size_t idx_1 = 0 ; idx_1 < intervals.size() ; ++idx_1){
                // self interval
                for(partsize_t idx_p1 = 0 ; idx_p1 < intervals[idx_1].second ; ++idx_p1){
                    for(partsize_t idx_p2 = idx_p1+1 ; idx_p2 < intervals[idx_1].second ; ++idx_p2){
                        const real_number dist_r2 = compute_distance_r2(particles_positions[(intervals[idx_1].first+idx_p1)*size_particle_positions + IDX_X],
                                                                        particles_positions[(intervals[idx_1].first+idx_p1)*size_particle_positions + IDX_Y],
                                                                        particles_positions[(intervals[idx_1].first+idx_p1)*size_particle_positions + IDX_Z],
                                                                        particles_positions[(intervals[idx_1].first+idx_p2)*size_particle_positions + IDX_X],
                                                                        particles_positions[(intervals[idx_1].first+idx_p2)*size_particle_positions + IDX_Y],
                                                                        particles_positions[(intervals[idx_1].first+idx_p2)*size_particle_positions + IDX_Z]);
                        if(dist_r2 < cutoff_radius*cutoff_radius){
                            in_computer.template compute_interaction<size_particle_positions,size_particle_rhs>(
                                                &particles_positions[(intervals[idx_1].first+idx_p1)*size_particle_positions],
                                                &particles_current_rhs[(intervals[idx_1].first+idx_p1)*size_particle_rhs],
                                                &particles_positions[(intervals[idx_1].first+idx_p2)*size_particle_positions],
                                                &particles_current_rhs[(intervals[idx_1].first+idx_p2)*size_particle_rhs],
                                                dist_r2);
                        }
                    }
                }

                // with other interval
                for(size_t idx_2 = idx_1+1 ; idx_2 < intervals.size() ; ++idx_2){
                    for(partsize_t idx_p1 = 0 ; idx_p1 < intervals[idx_1].second ; ++idx_p1){
                        for(partsize_t idx_p2 = 0 ; idx_p2 < intervals[idx_2].second ; ++idx_p2){
                            const real_number dist_r2 = compute_distance_r2(particles_positions[(intervals[idx_1].first+idx_p1)*size_particle_positions + IDX_X],
                                                                            particles_positions[(intervals[idx_1].first+idx_p1)*size_particle_positions + IDX_Y],
                                                                            particles_positions[(intervals[idx_1].first+idx_p1)*size_particle_positions + IDX_Z],
                                                                            particles_positions[(intervals[idx_2].first+idx_p2)*size_particle_positions + IDX_X],
                                                                            particles_positions[(intervals[idx_2].first+idx_p2)*size_particle_positions + IDX_Y],
                                                                            particles_positions[(intervals[idx_2].first+idx_p2)*size_particle_positions + IDX_Z]);
                            if(dist_r2 < cutoff_radius*cutoff_radius){
                                in_computer.template compute_interaction<size_particle_positions,size_particle_rhs>(
                                                    &particles_positions[(intervals[idx_1].first+idx_p1)*size_particle_positions],
                                                    &particles_current_rhs[(intervals[idx_1].first+idx_p1)*size_particle_rhs],
                                                    &particles_positions[(intervals[idx_2].first+idx_p2)*size_particle_positions],
                                                    &particles_current_rhs[(intervals[idx_2].first+idx_p2)*size_particle_rhs],
                                                    dist_r2);
                            }
                        }
                    }
                }
            }


            const long int currenct_cell_idx = iter_cell.first;
            const std::vector<std::pair<partsize_t,partsize_t>>* neighbors[27];
747
748
749
            long int neighbors_indexes[27];
            const int nbNeighbors = my_tree.getNeighbors(currenct_cell_idx, neighbors, neighbors_indexes, false);

750
751
752
            for(size_t idx_1 = 0 ; idx_1 < intervals.size() ; ++idx_1){
                // with other interval
                for(size_t idx_neighbor = 0 ; idx_neighbor < nbNeighbors ; ++idx_neighbor){
753
754
755
756
757
758
759
760
761
762
763
764
765
766
767
768
769
770
                    if(currenct_cell_idx < neighbors_indexes[idx_neighbor]){
                        for(size_t idx_2 = 0 ; idx_2 < (*neighbors[idx_neighbor]).size() ; ++idx_2){
                            for(partsize_t idx_p1 = 0 ; idx_p1 < intervals[idx_1].second ; ++idx_p1){
                                for(partsize_t idx_p2 = 0 ; idx_p2 < (*neighbors[idx_neighbor])[idx_2].second ; ++idx_p2){
                                    const real_number dist_r2 = compute_distance_r2(particles_positions[(intervals[idx_1].first+idx_p1)*size_particle_positions + IDX_X],
                                                                                    particles_positions[(intervals[idx_1].first+idx_p1)*size_particle_positions + IDX_Y],
                                                                                    particles_positions[(intervals[idx_1].first+idx_p1)*size_particle_positions + IDX_Z],
                                                                                    particles_positions[((*neighbors[idx_neighbor])[idx_2].first+idx_p2)*size_particle_positions + IDX_X],
                                                                                    particles_positions[((*neighbors[idx_neighbor])[idx_2].first+idx_p2)*size_particle_positions + IDX_Y],
                                                                                    particles_positions[((*neighbors[idx_neighbor])[idx_2].first+idx_p2)*size_particle_positions + IDX_Z]);
                                    if(dist_r2 < cutoff_radius*cutoff_radius){
                                        in_computer.template compute_interaction<size_particle_positions,size_particle_rhs>(
                                                            &particles_positions[(intervals[idx_1].first+idx_p1)*size_particle_positions],
                                                            &particles_current_rhs[(intervals[idx_1].first+idx_p1)*size_particle_rhs],
                                                            &particles_positions[((*neighbors[idx_neighbor])[idx_2].first+idx_p2)*size_particle_positions],
                                                            &particles_current_rhs[((*neighbors[idx_neighbor])[idx_2].first+idx_p2)*size_particle_rhs],
                                                            dist_r2);
                                    }
771
772
773
774
775
776
777
778
779
780
781
                                }
                            }
                        }
                    }
                }
            }
        }
    }
};

#endif