forked from fmihpc/vlasiator
-
Notifications
You must be signed in to change notification settings - Fork 1
/
spatial_cell_gpu.cpp
1879 lines (1723 loc) · 82 KB
/
spatial_cell_gpu.cpp
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
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
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
230
231
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
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
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
492
493
494
495
496
497
498
499
500
501
502
503
504
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
547
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
575
576
577
578
579
580
581
582
583
584
585
586
587
588
589
590
591
592
593
594
595
596
597
598
599
600
601
602
603
604
605
606
607
608
609
610
611
612
613
614
615
616
617
618
619
620
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
665
666
667
668
669
670
671
672
673
674
675
676
677
678
679
680
681
682
683
684
685
686
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
747
748
749
750
751
752
753
754
755
756
757
758
759
760
761
762
763
764
765
766
767
768
769
770
771
772
773
774
775
776
777
778
779
780
781
782
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
826
827
828
829
830
831
832
833
834
835
836
837
838
839
840
841
842
843
844
845
846
847
848
849
850
851
852
853
854
855
856
857
858
859
860
861
862
863
864
865
866
867
868
869
870
871
872
873
874
875
876
877
878
879
880
881
882
883
884
885
886
887
888
889
890
891
892
893
894
895
896
897
898
899
900
901
902
903
904
905
906
907
908
909
910
911
912
913
914
915
916
917
918
919
920
921
922
923
924
925
926
927
928
929
930
931
932
933
934
935
936
937
938
939
940
941
942
943
944
945
946
947
948
949
950
951
952
953
954
955
956
957
958
959
960
961
962
963
964
965
966
967
968
969
970
971
972
973
974
975
976
977
978
979
980
981
982
983
984
985
986
987
988
989
990
991
992
993
994
995
996
997
998
999
1000
/*
* This file is part of Vlasiator.
* Copyright 2010-2016 Finnish Meteorological Institute
*
* For details of usage, see the COPYING file and read the "Rules of the Road"
* at http://www.physics.helsinki.fi/vlasiator/
*
* This program is free software; you can redistribute it and/or modify
* it under the terms of the GNU General Public License as published by
* the Free Software Foundation; either version 2 of the License, or
* (at your option) any later version.
*
* This program is distributed in the hope that it will be useful,
* but WITHOUT ANY WARRANTY; without even the implied warranty of
* MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
* GNU General Public License for more details.
*
* You should have received a copy of the GNU General Public License along
* with this program; if not, write to the Free Software Foundation, Inc.,
* 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
*/
#include <unordered_set>
#include "spatial_cell_gpu.hpp"
#include "arch/gpu_base.hpp"
#include "object_wrapper.h"
#include "velocity_mesh_parameters.h"
#ifndef NDEBUG
#define DEBUG_SPATIAL_CELL
#endif
using namespace std;
/** GPU kernel for identifying which blocks have relevant content */
__global__ void __launch_bounds__(WID3,4) update_velocity_block_content_lists_kernel (
vmesh::VelocityMesh *vmesh,
vmesh::VelocityBlockContainer *blockContainer,
split::SplitVector<vmesh::GlobalID>* velocity_block_with_content_list,
split::SplitVector<vmesh::GlobalID>* velocity_block_with_no_content_list,
Realf velocity_block_min_value
) {
const int gpuBlocks = gridDim.x;
const int blocki = blockIdx.x;
const int i = threadIdx.x;
const int j = threadIdx.y;
const int k = threadIdx.z;
const uint ti = k*WID2 + j*WID + i;
__shared__ bool has_content[WID3];
const uint nBlocks = vmesh->size();
for (uint blockLID=blocki; blockLID<nBlocks; blockLID += gpuBlocks) {
vmesh::GlobalID blockGID = vmesh->getGlobalID(blockLID);
#ifdef DEBUG_SPATIAL_CELL
if (blockGID == vmesh->invalidGlobalID()) {
continue;
}
if (blockLID == vmesh->invalidLocalID()) {
continue;
}
#endif
Realf* avgs = blockContainer->getData(blockLID);
has_content[ti] = avgs[ti] >= velocity_block_min_value ? true : false;
__syncthreads();
// Implemented just a simple non-optimized thread OR
for (unsigned int s=WID3/2; s>0; s>>=1) {
if (ti < s) {
has_content[ti] = has_content[ti] || has_content[ti + s];
}
__syncthreads();
}
if (ti==0) {
if (has_content[0]) {
velocity_block_with_content_list->device_push_back(blockGID);
} else {
velocity_block_with_no_content_list->device_push_back(blockGID);
}
}
__syncthreads();
}
}
/** Gpu Kernel to quickly gather blocks and their v-space halo */
__global__ void __launch_bounds__(GPUTHREADS,4) update_blocks_required_halo_kernel (
vmesh::VelocityMesh *vmesh,
Hashinator::Hashmap<vmesh::GlobalID,vmesh::LocalID>* BlocksRequiredMap,
split::SplitVector<vmesh::GlobalID> *velocity_block_with_content_list,
split::SplitVector<vmesh::GlobalID> *BlocksHalo,
const int addWidthV,
// The following 4 vectors are passed just to be able to clear them on-device
split::SplitVector<vmesh::GlobalID>* BlocksRequired,
split::SplitVector<vmesh::GlobalID>* BlocksToRemove,
split::SplitVector<vmesh::GlobalID>* BlocksToAdd,
split::SplitVector<vmesh::GlobalID>* BlocksToMove
) {
const int gpuBlocks = gridDim.x;
const int blocki = blockIdx.x;
const int warpSize = blockDim.x*blockDim.y*blockDim.z;
const vmesh::LocalID ti = threadIdx.z*blockDim.x*blockDim.y + threadIdx.y*blockDim.x + threadIdx.x;
const unsigned long localContentBlocks = velocity_block_with_content_list->size();
vmesh::GlobalID nGIDs[27] = {0};
for (vmesh::LocalID index=blocki*warpSize; index<localContentBlocks; index += gpuBlocks*warpSize) {
if (index+ti < localContentBlocks) {
vmesh::GlobalID GID = velocity_block_with_content_list->at(index+ti);
//BlocksRequiredMap->set_element(GID,GID); // Already added via Hashinator interface
vmesh::LocalID ind0,ind1,ind2;
vmesh->getIndices(GID,ind0,ind1,ind2);
// New algorithm: attempts to minimize thread divergence and use of atomics.
// Gather existence of all 26 neighbours into 32-bit bitmask
vmesh::LocalID exist = 0;
int localCounter=0;
for (int offset_vx=-addWidthV; offset_vx<=addWidthV; offset_vx++) {
for (int offset_vy=-addWidthV; offset_vy<=addWidthV; offset_vy++) {
for (int offset_vz=-addWidthV; offset_vz<=addWidthV; offset_vz++) {
const int nind0 = ind0 + offset_vx;
const int nind1 = ind1 + offset_vy;
const int nind2 = ind2 + offset_vz;
const vmesh::GlobalID nGID
= vmesh->getGlobalID(nind0,nind1,nind2);
if (nGID != vmesh->invalidGlobalID()) {
// NEW:
nGIDs[localCounter] = nGID;
if (BlocksRequiredMap->device_count(nGID) != 0) {
exist = (exist | (1 << localCounter));
}
// OLD:
// if (BlocksRequiredMap->device_count(nGID) == 0) {
// //BlocksHalo->device_push_back(nGID);
// BlocksRequiredMap->set_element(nGID,nGID);
// }
}
localCounter++;
} // for vz
} // for vy
} // for vx
// Quick check: if all halo neighbours already exist, skip to next loop.
if (exist == 134217727) {
continue;
}
/**
Define order of responsibility. Blocks are added by:
first their face neighbours in x
then their face neighbours in y
then their face neighbours in z
then their edge neighbours
then their corner neighbours
Bitmask: (cell 13 is self)
Z
| 8 17 26
| 5 14 23
|2 7 11 16 20 25
| 4 13 22
|1 6 10 15 19 24
| 3 12 21
|0 9 18
L__________________ X
// First the 8 corners: add only if corner cell has no other existing neighours (that we know of)
// Then the 12 blocks which are between two corners
// Add only if the four face neighbours we know of are not there
// Then face neighbours
// Add z-neighbours only if we don't know of any x or y face neighbours
// Add y-neighbours only if we don't know of any x face neighbours
// Add all x-face-neighbours
Start python code:
def val(x,y,z):
# input is distances as -1,0,+1
# returns the index
return (z+1) + (y+1)*3 + (x+1)*9
def coord(inde):
z = (inde % 3)
y = ((inde//3) % 3)
x = ((inde//9) % 3)
return (x-1,y-1,z-1)
def dims2(inde):
(i0,j0,k0) = coord(inde)
listi = []
for i in range(-1,2):
for j in range(-1,2):
for k in range(-1,2):
di = abs(i-i0)
dj = abs(j-j0)
dk = abs(k-k0)
if (di<2) and (dj<2) and (dk<2) and ((di+dj+dk)<3):
listi.append(val(i,j,k))
return listi
def neigh_x(inde):
# looking at x-directional neighbour
# only returns self
listi = [inde]
return listi
def neigh_y(inde):
# looking at y-directional neighbour
# returns the target and any x-directional neighbours of it
(i0,j0,k0) = coord(inde)
listi = []
for i in range(-1,2):
for j in range(-1,2):
for k in range(-1,2):
di = abs(i-i0)
dj = abs(j-j0)
dk = abs(k-k0)
if (di<2) and (dj<1) and (dk<1):
listi.append(val(i,j,k))
return listi
def neigh_z(inde):
# looking at z-directional neighbour
# returns the target and any x-directional or y-directional neighbours of it
(i0,j0,k0) = coord(inde)
listi = []
for i in range(-1,2):
for j in range(-1,2):
for k in range(-1,2):
di = abs(i-i0)
dj = abs(j-j0)
dk = abs(k-k0)
if (di<2) and (dj<2) and (dk<1) and ((di+dj+dk)<2):
listi.append(val(i,j,k))
return listi
def dims1(inde):
(i0,j0,k0) = coord(inde)
listi = []
for i in range(-1,2):
for j in range(-1,2):
for k in range(-1,2):
di = abs(i-i0)
dj = abs(j-j0)
dk = abs(k-k0)
if (di<2) and (dj<2) and (dk<2) and ((di+dj+dk)<2):
listi.append(val(i,j,k))
return listi
def findmask(start,indices):
val = 0
strmoi=""
for x in indices:
val |= (1<<x)
strmoi = strmoi+str(x)+" "
#print(start,indices,val)
print(" // "+str(start)+": "+strmoi+" -> "+str(val))
print(" if ( ( (exist & (vmesh::LocalID)"+str(val)+") == 0) && (nGIDs["+str(start)+"] != 0)) {")
print(" BlocksHalo->device_push_back(nGIDs["+str(start)+"]);")
print(" }")
def findall():
for i in range(-1,2):
for j in range(-1,2):
for k in range(-1,2):
di = abs(i)
dj = abs(j)
dk = abs(k)
if (di+dj+dk==3):
# corner
thiss = val(i,j,k)
findmask(thiss,dims2(thiss))
elif (di+dj+dk==2):
# mid-edge
thiss = val(i,j,k)
findmask(thiss,dims1(thiss))
elif (di+dj+dk==0):
# own, do nothing
pass
elif ((di==1) and (dj+dk==0)):
# x-directional neighbour
thiss = val(i,j,k)
findmask(thiss,neigh_x(thiss))
elif ((dj==1) and (di+dk==0)):
# y-directional neighbour
thiss = val(i,j,k)
findmask(thiss,neigh_y(thiss))
elif ((dk==1) and (di+dj==0)):
# z-directional neighbour
thiss = val(i,j,k)
findmask(thiss,neigh_z(thiss))
else:
print("error passed through!")
**/
// 0: 0 1 3 4 9 10 12 -> 5659
if ( ( (exist & (vmesh::LocalID)5659) == 0) && (nGIDs[0] != 0)) {
BlocksHalo->device_push_back(nGIDs[0]);
}
// 1: 0 1 2 4 10 -> 1047
if ( ( (exist & (vmesh::LocalID)1047) == 0) && (nGIDs[1] != 0)) {
BlocksHalo->device_push_back(nGIDs[1]);
}
// 2: 1 2 4 5 10 11 14 -> 19510
if ( ( (exist & (vmesh::LocalID)19510) == 0) && (nGIDs[2] != 0)) {
BlocksHalo->device_push_back(nGIDs[2]);
}
// 3: 0 3 4 6 12 -> 4185
if ( ( (exist & (vmesh::LocalID)4185) == 0) && (nGIDs[3] != 0)) {
BlocksHalo->device_push_back(nGIDs[3]);
}
// 4: 4 -> 16
if ( ( (exist & (vmesh::LocalID)16) == 0) && (nGIDs[4] != 0)) {
BlocksHalo->device_push_back(nGIDs[4]);
}
// 5: 2 4 5 8 14 -> 16692
if ( ( (exist & (vmesh::LocalID)16692) == 0) && (nGIDs[5] != 0)) {
BlocksHalo->device_push_back(nGIDs[5]);
}
// 6: 3 4 6 7 12 15 16 -> 102616
if ( ( (exist & (vmesh::LocalID)102616) == 0) && (nGIDs[6] != 0)) {
BlocksHalo->device_push_back(nGIDs[6]);
}
// 7: 4 6 7 8 16 -> 66000
if ( ( (exist & (vmesh::LocalID)66000) == 0) && (nGIDs[7] != 0)) {
BlocksHalo->device_push_back(nGIDs[7]);
}
// 8: 4 5 7 8 14 16 17 -> 213424
if ( ( (exist & (vmesh::LocalID)213424) == 0) && (nGIDs[8] != 0)) {
BlocksHalo->device_push_back(nGIDs[8]);
}
// 9: 0 9 10 12 18 -> 267777
if ( ( (exist & (vmesh::LocalID)267777) == 0) && (nGIDs[9] != 0)) {
BlocksHalo->device_push_back(nGIDs[9]);
}
// 10: 1 10 19 -> 525314
if ( ( (exist & (vmesh::LocalID)525314) == 0) && (nGIDs[10] != 0)) {
BlocksHalo->device_push_back(nGIDs[10]);
}
// 11: 2 10 11 14 20 -> 1068036
if ( ( (exist & (vmesh::LocalID)1068036) == 0) && (nGIDs[11] != 0)) {
BlocksHalo->device_push_back(nGIDs[11]);
}
// 12: 3 9 12 15 21 -> 2134536
if ( ( (exist & (vmesh::LocalID)2134536) == 0) && (nGIDs[12] != 0)) {
BlocksHalo->device_push_back(nGIDs[12]);
}
// 14: 5 11 14 17 23 -> 8538144
if ( ( (exist & (vmesh::LocalID)8538144) == 0) && (nGIDs[14] != 0)) {
BlocksHalo->device_push_back(nGIDs[14]);
}
// 15: 6 12 15 16 24 -> 16879680
if ( ( (exist & (vmesh::LocalID)16879680) == 0) && (nGIDs[15] != 0)) {
BlocksHalo->device_push_back(nGIDs[15]);
}
// 16: 7 16 25 -> 33620096
if ( ( (exist & (vmesh::LocalID)33620096) == 0) && (nGIDs[16] != 0)) {
BlocksHalo->device_push_back(nGIDs[16]);
}
// 17: 8 14 16 17 26 -> 67322112
if ( ( (exist & (vmesh::LocalID)67322112) == 0) && (nGIDs[17] != 0)) {
BlocksHalo->device_push_back(nGIDs[17]);
}
// 18: 9 10 12 18 19 21 22 -> 7083520
if ( ( (exist & (vmesh::LocalID)7083520) == 0) && (nGIDs[18] != 0)) {
BlocksHalo->device_push_back(nGIDs[18]);
}
// 19: 10 18 19 20 22 -> 6030336
if ( ( (exist & (vmesh::LocalID)6030336) == 0) && (nGIDs[19] != 0)) {
BlocksHalo->device_push_back(nGIDs[19]);
}
// 20: 10 11 14 19 20 22 23 -> 14175232
if ( ( (exist & (vmesh::LocalID)14175232) == 0) && (nGIDs[20] != 0)) {
BlocksHalo->device_push_back(nGIDs[20]);
}
// 21: 12 18 21 22 24 -> 23334912
if ( ( (exist & (vmesh::LocalID)23334912) == 0) && (nGIDs[21] != 0)) {
BlocksHalo->device_push_back(nGIDs[21]);
}
// 22: 22 -> 4194304
if ( ( (exist & (vmesh::LocalID)4194304) == 0) && (nGIDs[22] != 0)) {
BlocksHalo->device_push_back(nGIDs[22]);
}
// 23: 14 20 22 23 26 -> 80756736
if ( ( (exist & (vmesh::LocalID)80756736) == 0) && (nGIDs[23] != 0)) {
BlocksHalo->device_push_back(nGIDs[23]);
}
// 24: 12 15 16 21 22 24 25 -> 56725504
if ( ( (exist & (vmesh::LocalID)56725504) == 0) && (nGIDs[24] != 0)) {
BlocksHalo->device_push_back(nGIDs[24]);
}
// 25: 16 22 24 25 26 -> 121700352
if ( ( (exist & (vmesh::LocalID)121700352) == 0) && (nGIDs[25] != 0)) {
BlocksHalo->device_push_back(nGIDs[25]);
}
// 26: 14 16 17 22 23 25 26 -> 113459200
if ( ( (exist & (vmesh::LocalID)113459200) == 0) && (nGIDs[26] != 0)) {
BlocksHalo->device_push_back(nGIDs[26]);
}
} // if index
} // for blocks
// One thread also sets the sizes of these vectors to zero
if (blockIdx.x == blockIdx.y == blockIdx.z == threadIdx.x == threadIdx.y == threadIdx.z == 0) {
BlocksRequired->clear();
BlocksToRemove->clear();
BlocksToAdd->clear();
BlocksToMove->clear();
}
}
/** Gpu Kernel to quickly add blocks which have spatial neighbours */
__global__ void __launch_bounds__(GPUTHREADS,4) update_neighbours_have_content_kernel (
vmesh::VelocityMesh *vmesh,
split::SplitVector<vmesh::GlobalID> *BlocksHalo,
Hashinator::Hashmap<vmesh::GlobalID,vmesh::LocalID>* BlocksRequiredMap,
vmesh::GlobalID *neighbor_velocity_block_with_content_list,
const unsigned long neighborContentBlocks
) {
const int gpuBlocks = gridDim.x;
const int blocki = blockIdx.x;
const int warpSize = blockDim.x*blockDim.y*blockDim.z;
const vmesh::LocalID ti = threadIdx.z*blockDim.x*blockDim.y + threadIdx.y*blockDim.x + threadIdx.x;
for (vmesh::LocalID index=blocki*warpSize; index<neighborContentBlocks; index += gpuBlocks*warpSize) {
if (index+ti < neighborContentBlocks) {
vmesh::GlobalID GID = neighbor_velocity_block_with_content_list[index+ti];
if (BlocksRequiredMap->device_count(GID) == 0) {
//BlocksRequiredMap->set_element(GID,GID);
BlocksHalo->device_push_back(GID);
}
}
}
}
/** GPU kernel for selecting only non-existing blocks for addition
This kernel may be non-optimized in itself, but use of it gets rid
of the need of vmesh prefetching back and forth.
*/
__global__ void __launch_bounds__(GPUTHREADS,4) update_blocks_to_add_kernel (
vmesh::VelocityMesh *vmesh,
split::SplitVector<vmesh::GlobalID>* BlocksRequired,
split::SplitVector<vmesh::GlobalID>* BlocksToAdd,
split::SplitVector<vmesh::GlobalID>* BlocksToMove,
const uint nBlocksRequired
) {
const int gpuBlocks = gridDim.x;
const int blocki = blockIdx.x;
const int warpSize = blockDim.x*blockDim.y*blockDim.z;
const uint ti = threadIdx.z*blockDim.x*blockDim.y + threadIdx.y*blockDim.x + threadIdx.x;
for (uint index=blocki*warpSize; index<nBlocksRequired; index += gpuBlocks*warpSize) {
if (index+ti < nBlocksRequired) {
const vmesh::GlobalID GIDreq = BlocksRequired->at(index+ti);
const vmesh::LocalID LIDreq = vmesh->getLocalID(GIDreq);
if ( LIDreq == vmesh->invalidLocalID() ) {
BlocksToAdd->device_push_back(GIDreq);
} else if (LIDreq >= nBlocksRequired) {
BlocksToMove->device_push_back(GIDreq);
}
}
}
}
/** GPU kernel for identifying blocks for deletion.
This kernel may be non-optimized in itself, but use of it gets rid
of the need of vmesh prefetching back and forth.
*/
__global__ void __launch_bounds__(GPUTHREADS,4) update_blocks_to_remove_kernel (
split::SplitVector<vmesh::GlobalID>* velocity_block_with_no_content_list,
Hashinator::Hashmap<vmesh::GlobalID,vmesh::LocalID>* BlocksRequiredMap,
split::SplitVector<vmesh::GlobalID>* BlocksToRemove,
const uint localNoContentBlocks
) {
const int gpuBlocks = gridDim.x;
const int blocki = blockIdx.x;
const int warpSize = blockDim.x*blockDim.y*blockDim.z;
const uint ti = threadIdx.z*blockDim.x*blockDim.y + threadIdx.y*blockDim.x + threadIdx.x;
for (uint index=blocki*warpSize; index<localNoContentBlocks; index += gpuBlocks*warpSize) {
if (index+ti < localNoContentBlocks) {
const vmesh::GlobalID GIDcandidate = velocity_block_with_no_content_list->at(index+ti);
if (BlocksRequiredMap->device_count(GIDcandidate) == 0) {
BlocksToRemove->device_push_back(GIDcandidate);
}
}
}
}
/** GPU kernel for updating blocks based on generated lists */
__global__ void __launch_bounds__(WID3,4) update_velocity_blocks_kernel(
vmesh::VelocityMesh *vmesh,
vmesh::VelocityBlockContainer *blockContainer,
split::SplitVector<vmesh::GlobalID>* BlocksToAdd,
split::SplitVector<vmesh::GlobalID>* BlocksToRemove,
split::SplitVector<vmesh::GlobalID>* BlocksToMove,
vmesh::LocalID nBlocksBeforeAdjust,
vmesh::LocalID nBlocksAfterAdjust,
vmesh::LocalID *VectorIndex,
Realf* gpu_rhoLossAdjust
) {
const int gpuBlocks = gridDim.x;
const int blocki = blockIdx.x;
const int i = threadIdx.x;
const int j = threadIdx.y;
const int k = threadIdx.z;
const uint ti = k*WID2 + j*WID + i;
vmesh::LocalID *addVectorIndex = VectorIndex;
vmesh::LocalID *moveVectorIndex = VectorIndex+1;
const vmesh::LocalID nToAdd = BlocksToAdd->size();
const vmesh::LocalID nToRemove = BlocksToRemove->size();
const vmesh::LocalID nToMove = BlocksToMove->size();
const vmesh::LocalID nToCreate = nToAdd > nToRemove ? (nToAdd-nToRemove) : 0;
Realf local_rhoLoss = 0;
// For tracking mass-loss
__shared__ Realf massloss[WID3];
__shared__ vmesh::LocalID moveIndex;
__shared__ vmesh::LocalID addIndex;
for (vmesh::LocalID m=blocki; m<nToRemove; m += gpuBlocks) {
// Go through all blocks which are to be removed.
// If there is a corresponding block to be added, place that in its stead.
// Otherwise, take the corresponding block from the moved list instead.
const vmesh::GlobalID rmGID = BlocksToRemove->at(m);
const vmesh::LocalID rmLID = vmesh->getLocalID(rmGID);
#ifdef DEBUG_SPATIAL_CELL
if (rmGID == vmesh->invalidGlobalID()) {
continue;
}
if (rmLID == vmesh->invalidLocalID()) {
continue;
}
#endif
// Track mass loss:
Realf* rm_avgs = blockContainer->getData(rmLID);
Real* rm_block_parameters = blockContainer->getParameters(rmLID);
const Real rm_DV3 = rm_block_parameters[BlockParams::DVX]
* rm_block_parameters[BlockParams::DVY]
* rm_block_parameters[BlockParams::DVZ];
// thread-sum for rho
massloss[ti] = rm_avgs[ti]*rm_DV3;
__syncthreads();
// Implemented just a simple non-optimized thread sum
for (unsigned int s=WID3/2; s>0; s>>=1) {
if (ti < s) {
massloss[ti] += massloss[ti + s];
}
__syncthreads();
}
if (ti==0) {
// Bookkeeping only by one thread
local_rhoLoss += massloss[0];
}
__syncthreads();
// Figure out which GID to put here instead
if (rmLID >= nBlocksAfterAdjust) {
// Delete without replacing
if (ti==0) {
vmesh->deleteBlock(rmGID,rmLID);
}
__syncthreads();
continue;
}
if (ti==0) moveIndex = atomicAdd(moveVectorIndex,1);
__syncthreads();
if (moveIndex<nToMove) {
// Move from latter part of vmesh
const vmesh::GlobalID replaceGID = BlocksToMove->at(moveIndex);
const vmesh::LocalID replaceLID = vmesh->getLocalID(replaceGID);
Realf* repl_avgs = blockContainer->getData(replaceLID);
Real* repl_block_parameters = blockContainer->getParameters(replaceLID);
rm_avgs[ti] = repl_avgs[ti];
if (ti < BlockParams::N_VELOCITY_BLOCK_PARAMS) {
rm_block_parameters[ti] = repl_block_parameters[ti];
}
__syncthreads();
if (ti==0) {
// Remove hashmap entry for removed block, add instead created block
vmesh->replaceBlock(rmGID,rmLID,replaceGID);
}
__syncthreads();
#ifdef DEBUG_SPATIAL_CELL
if (vmesh->getGlobalID(rmLID) == vmesh->invalidGlobalID()) {
continue;
}
if (vmesh->getLocalID(replaceGID) == vmesh->invalidLocalID()) {
continue;
}
#endif
continue;
}
if (ti==0) addIndex = atomicAdd(addVectorIndex,1);
__syncthreads();
if (addIndex<nToAdd) {
// New GID
const vmesh::GlobalID addGID = BlocksToAdd->at(addIndex);
#ifdef DEBUG_SPATIAL_CELL
if (addGID == vmesh->invalidGlobalID()) {
__syncthreads();
continue;
}
#endif
rm_avgs[ti] = 0;
if (ti==0) {
// Write in block parameters
vmesh->getCellSize(addGID,&(rm_block_parameters[BlockParams::DVX]));
vmesh->getBlockCoordinates(addGID,&(rm_block_parameters[BlockParams::VXCRD]));
vmesh->replaceBlock(rmGID,rmLID,addGID);
}
__syncthreads();
#ifdef DEBUG_SPATIAL_CELL
if (vmesh->getGlobalID(rmLID) == vmesh->invalidGlobalID()) {
continue;
}
if (vmesh->getLocalID(addGID) == vmesh->invalidLocalID()) {
continue;
}
#endif
continue;
}
#ifdef DEBUG_SPATIAL_CELL
if (ti==0) {
printf("Error! Fall through in update_velocity_blocks_kernel! \n");
}
#endif
__syncthreads();
}
// Now, if we need to expand the size of the vmesh, let's add blocks.
// For thread-safety,this assumes that the localToGlobalMap is already of sufficient size, as should be
// the block_data and block_parameters vectors.
for (vmesh::LocalID m=blocki; m<nToCreate; m += gpuBlocks) {
// Debug check: if we are adding elements, then nToMove should be zero
// We have already used up nToRemove entries from the addition vector.
const vmesh::GlobalID addGID = BlocksToAdd->at(nToRemove+m);
// We need to add the data of addGID to a new LID:
const vmesh::LocalID addLID = nBlocksBeforeAdjust + m;
Realf* add_avgs = blockContainer->getData(addLID);
Real* add_block_parameters = blockContainer->getParameters(addLID);
// Zero out blockdata
add_avgs[ti] = 0;
if (ti==0) {
// Write in block parameters
vmesh->getCellSize(addGID,&(add_block_parameters[BlockParams::DVX]));
vmesh->getBlockCoordinates(addGID,&(add_block_parameters[BlockParams::VXCRD]));
vmesh->placeBlock(addGID,addLID);
}
__syncthreads();
#ifdef DEBUG_SPATIAL_CELL
if (vmesh->getGlobalID(addLID) == vmesh->invalidGlobalID()) {
continue;
}
if (vmesh->getLocalID(addGID) == vmesh->invalidLocalID()) {
continue;
}
#endif
}
// Atomically update accumulated mass loss
if (ti==0) {
Realf old = atomicAdd(gpu_rhoLossAdjust, local_rhoLoss);
}
}
namespace spatial_cell {
int SpatialCell::activePopID = 0;
uint64_t SpatialCell::mpi_transfer_type = 0;
bool SpatialCell::mpiTransferAtSysBoundaries = false;
bool SpatialCell::mpiTransferInAMRTranslation = false;
int SpatialCell::mpiTransferXYZTranslation = 0;
SpatialCell::SpatialCell() {
// Block list and cache always have room for all blocks
this->sysBoundaryLayer=0; // Default value, layer not yet initialized
for (unsigned int i=0; i<WID3; ++i) {
null_block_data[i] = 0.0;
}
// reset spatial cell parameters
for (unsigned int i = 0; i < CellParams::N_SPATIAL_CELL_PARAMS; i++) {
this->parameters[i]=0.0;
}
// reset BVOL derivatives
for (unsigned int i = 0; i < bvolderivatives::N_BVOL_DERIVATIVES; i++) {
this->derivativesBVOL[i]=0;
}
for (unsigned int i = 0; i < MAX_NEIGHBORS_PER_DIM; ++i) {
this->neighbor_number_of_blocks[i] = 0;
this->neighbor_block_data[i] = NULL;
}
//is transferred by default
this->mpiTransferEnabled=true;
// Set correct number of populations
populations.resize(getObjectWrapper().particleSpecies.size());
// Set velocity meshes
for (uint popID=0; popID<populations.size(); ++popID) {
const species::Species& spec = getObjectWrapper().particleSpecies[popID];
populations[popID].vmesh->initialize(spec.velocityMesh);
populations[popID].velocityBlockMinValue = spec.sparseMinValue;
populations[popID].N_blocks = 0;
}
// SplitVectors via pointers for unified memory
velocity_block_with_content_list = new split::SplitVector<vmesh::GlobalID>(1);
velocity_block_with_no_content_list = new split::SplitVector<vmesh::GlobalID>(1);
BlocksHalo = new split::SplitVector<vmesh::GlobalID>(1);
BlocksRequired = new split::SplitVector<vmesh::GlobalID>(1);
BlocksToAdd = new split::SplitVector<vmesh::GlobalID>(1);
BlocksToRemove = new split::SplitVector<vmesh::GlobalID>(1);
BlocksToMove = new split::SplitVector<vmesh::GlobalID>(1);
velocity_block_with_content_list->clear();
velocity_block_with_no_content_list->clear();
BlocksHalo->clear();
BlocksRequired->clear();
BlocksToAdd->clear();
BlocksToRemove->clear();
BlocksToMove->clear();
attachedStream=0;
BlocksRequiredMap = new Hashinator::Hashmap<vmesh::GlobalID,vmesh::LocalID>(7);
gpuMallocHost((void **) &info_vbwcl, sizeof(split::SplitInfo));
gpuMallocHost((void **) &info_vbwncl, sizeof(split::SplitInfo));
gpuMallocHost((void **) &info_toRemove, sizeof(split::SplitInfo));
gpuMallocHost((void **) &info_toAdd, sizeof(split::SplitInfo));
gpuMallocHost((void **) &info_toMove, sizeof(split::SplitInfo));
gpuMallocHost((void **) &info_Required, sizeof(split::SplitInfo));
gpuMallocHost((void **) &info_Halo, sizeof(split::SplitInfo));
gpuMallocHost((void **) &info_brm, sizeof(Hashinator::MapInfo));
}
SpatialCell::~SpatialCell() {
delete velocity_block_with_content_list;
delete velocity_block_with_no_content_list;
delete BlocksHalo;
delete BlocksRequired;
delete BlocksToAdd;
delete BlocksToRemove;
delete BlocksToMove;
delete BlocksRequiredMap;
gpuFreeHost(info_vbwcl);
gpuFreeHost(info_vbwncl);
gpuFreeHost(info_toRemove);
gpuFreeHost(info_toAdd);
gpuFreeHost(info_toMove);
gpuFreeHost(info_Required);
gpuFreeHost(info_Halo);
gpuFreeHost(info_brm);
}
SpatialCell::SpatialCell(const SpatialCell& other) {
velocity_block_with_content_list = new split::SplitVector<vmesh::GlobalID>(*(other.velocity_block_with_content_list));
velocity_block_with_no_content_list = new split::SplitVector<vmesh::GlobalID>(*(other.velocity_block_with_no_content_list));
BlocksHalo = new split::SplitVector<vmesh::GlobalID>(1);
BlocksRequired = new split::SplitVector<vmesh::GlobalID>(1);
BlocksToAdd = new split::SplitVector<vmesh::GlobalID>(1);
BlocksToRemove = new split::SplitVector<vmesh::GlobalID>(1);
BlocksToMove = new split::SplitVector<vmesh::GlobalID>(1);
BlocksHalo->clear();
BlocksRequired->clear();
BlocksToAdd->clear();
BlocksToRemove->clear();
BlocksToMove->clear();
velocity_block_with_content_list->clear();
velocity_block_with_no_content_list->clear();
BlocksRequiredMap = new Hashinator::Hashmap<vmesh::GlobalID,vmesh::LocalID>(7);
// Make space reservation guesses based on popID 0
const uint reserveSize = other.populations[0].vmesh->size()*BLOCK_ALLOCATION_PADDING;
BlocksHalo->reserve(reserveSize,true);
BlocksRequired->reserve(reserveSize,true);
BlocksToAdd->reserve(reserveSize,true);
BlocksToRemove->reserve(reserveSize,true);
BlocksToMove->reserve(reserveSize,true);
velocity_block_with_content_list->reserve(reserveSize,true);
velocity_block_with_no_content_list->reserve(reserveSize,true);
// Member variables
ioLocalCellId = other.ioLocalCellId;
sysBoundaryFlag = other.sysBoundaryFlag;
sysBoundaryLayer = other.sysBoundaryLayer;
sysBoundaryLayerNew = other.sysBoundaryLayerNew;
velocity_block_with_content_list_size = other.velocity_block_with_content_list_size;
initialized = other.initialized;
mpiTransferEnabled = other.mpiTransferEnabled;
for (unsigned int i=0; i<bvolderivatives::N_BVOL_DERIVATIVES; ++i) {
derivativesBVOL[i] = other.derivativesBVOL[i];
}
for (unsigned int i=0; i<CellParams::N_SPATIAL_CELL_PARAMS; ++i) {
parameters[i] = other.parameters[i];
}
for (unsigned int i=0; i<WID3; ++i) {
null_block_data[i] = 0.0;
}
for (unsigned int i=0; i<MAX_NEIGHBORS_PER_DIM; ++i) {
neighbor_block_data[i] = other.neighbor_block_data[i];
neighbor_number_of_blocks[i] = other.neighbor_number_of_blocks[i];
}
if (other.face_neighbor_ranks.size()>0) {
face_neighbor_ranks = std::map<int,std::set<int>>(other.face_neighbor_ranks);
}
if (other.populations.size()>0) {
populations = std::vector<spatial_cell::Population>(other.populations);
}
attachedStream=0;
gpuMallocHost((void **) &info_vbwcl, sizeof(split::SplitInfo));
gpuMallocHost((void **) &info_vbwncl, sizeof(split::SplitInfo));
gpuMallocHost((void **) &info_toRemove, sizeof(split::SplitInfo));
gpuMallocHost((void **) &info_toAdd, sizeof(split::SplitInfo));
gpuMallocHost((void **) &info_toMove, sizeof(split::SplitInfo));
gpuMallocHost((void **) &info_Required, sizeof(split::SplitInfo));
gpuMallocHost((void **) &info_Halo, sizeof(split::SplitInfo));
gpuMallocHost((void **) &info_brm, sizeof(Hashinator::MapInfo));
}
const SpatialCell& SpatialCell::operator=(const SpatialCell& other) {
const uint reserveSize = (other.BlocksRequired)->capacity();
BlocksHalo->clear();
BlocksRequired->clear();
BlocksToAdd->clear();
BlocksToRemove->clear();
BlocksToMove->clear();
velocity_block_with_content_list->clear();
velocity_block_with_no_content_list->clear();
delete BlocksRequiredMap;
BlocksHalo->reserve(reserveSize,true);
BlocksRequired->reserve(reserveSize,true);
BlocksToAdd->reserve(reserveSize,true);
BlocksToRemove->reserve(reserveSize,true);
BlocksToMove->reserve(reserveSize,true);
velocity_block_with_content_list->reserve(reserveSize,true);
velocity_block_with_no_content_list->reserve(reserveSize,true);
const vmesh::LocalID HashmapReqSize = ceil(log2(reserveSize)) +2;
BlocksRequiredMap = new Hashinator::Hashmap<vmesh::GlobalID,vmesh::LocalID>(HashmapReqSize);
// Member variables
ioLocalCellId = other.ioLocalCellId;
sysBoundaryFlag = other.sysBoundaryFlag;
sysBoundaryLayer = other.sysBoundaryLayer;
sysBoundaryLayerNew = other.sysBoundaryLayerNew;
velocity_block_with_content_list_size = other.velocity_block_with_content_list_size;
initialized = other.initialized;
mpiTransferEnabled = other.mpiTransferEnabled;
for (unsigned int i=0; i<bvolderivatives::N_BVOL_DERIVATIVES; ++i) {
derivativesBVOL[i] = other.derivativesBVOL[i];
}
for (unsigned int i=0; i<CellParams::N_SPATIAL_CELL_PARAMS; ++i) {
parameters[i] = other.parameters[i];
}
for (unsigned int i=0; i<WID3; ++i) {
null_block_data[i] = 0.0;
}
for (unsigned int i=0; i<MAX_NEIGHBORS_PER_DIM; ++i) {
neighbor_block_data[i] = other.neighbor_block_data[i];
neighbor_number_of_blocks[i] = other.neighbor_number_of_blocks[i];
}
face_neighbor_ranks = std::map<int,std::set<int>>(other.face_neighbor_ranks);
populations = std::vector<spatial_cell::Population>(other.populations);
attachedStream=0;
return *this;
}
/** Advises unified memory subsystem on preferred location of memory
gpuMemAdviseSetPreferredLocation
gpuMemAdviseUnsetPreferredLocation
gpuMemAdviseSetReadMostly
gpuMemAdviceUnsetReadMostly
gpuMemAdviseSetAccessedBy
gpuMemAdviseUnsetAccessedBy
*/
void SpatialCell::gpu_advise() {
// CHK_ERR( gpuMemAdvise(ptr, count, advise, deviceID) );
// CHK_ERR( gpuMemAdvise(velocity_block_with_content_list, sizeof(velocity_block_with_content_list),gpuMemAdviseSetPreferredLocation, gpu_getDevice()) );
// gpu_getDevice()
int device = gpu_getDevice();
gpuStream_t stream = gpu_getStream();
BlocksHalo->memAdvise(gpuMemAdviseSetPreferredLocation,device,stream);
BlocksRequired->memAdvise(gpuMemAdviseSetPreferredLocation,device,stream);
BlocksToAdd->memAdvise(gpuMemAdviseSetPreferredLocation,device,stream);
BlocksToRemove->memAdvise(gpuMemAdviseSetPreferredLocation,device,stream);
BlocksToMove->memAdvise(gpuMemAdviseSetPreferredLocation,device,stream);
velocity_block_with_content_list->memAdvise(gpuMemAdviseSetPreferredLocation,device,stream);
velocity_block_with_no_content_list->memAdvise(gpuMemAdviseSetPreferredLocation,device,stream);
BlocksRequiredMap->memAdvise(gpuMemAdviseSetPreferredLocation,device,stream);
BlocksHalo->memAdvise(gpuMemAdviseSetAccessedBy,device,stream);
BlocksRequired->memAdvise(gpuMemAdviseSetAccessedBy,device,stream);
BlocksToAdd->memAdvise(gpuMemAdviseSetAccessedBy,device,stream);
BlocksToRemove->memAdvise(gpuMemAdviseSetAccessedBy,device,stream);
BlocksToMove->memAdvise(gpuMemAdviseSetAccessedBy,device,stream);
velocity_block_with_content_list->memAdvise(gpuMemAdviseSetAccessedBy,device,stream);
velocity_block_with_no_content_list->memAdvise(gpuMemAdviseSetAccessedBy,device,stream);
BlocksRequiredMap->memAdvise(gpuMemAdviseSetAccessedBy,device,stream);
// Loop over populations
for (size_t p=0; p<populations.size(); ++p) {
populations[p].blockContainer->gpu_memAdvise(device,stream);
populations[p].vmesh->gpu_memAdvise(device,stream);
}
}
/** Attaches or deattaches unified memory to a GPU stream
When attached, a stream can access this unified memory without
issues.
*/
void SpatialCell::gpu_attachToStream(gpuStream_t stream) {
// Return if attaching is not needed
if (!needAttachedStreams) {
return;
}
// Attach unified memory regions to streams
gpuStream_t newStream;
if (stream==0) {
newStream = gpu_getStream();
} else {
newStream = stream;
}
if (newStream == attachedStream) {
return;
} else {
attachedStream = newStream;
}
CHK_ERR( gpuStreamAttachMemAsync(attachedStream,velocity_block_with_content_list, 0,gpuMemAttachSingle) );
CHK_ERR( gpuStreamAttachMemAsync(attachedStream,velocity_block_with_no_content_list, 0,gpuMemAttachSingle) );
CHK_ERR( gpuStreamAttachMemAsync(attachedStream,BlocksHalo, 0,gpuMemAttachSingle) );
CHK_ERR( gpuStreamAttachMemAsync(attachedStream,BlocksToRemove, 0,gpuMemAttachSingle) );
CHK_ERR( gpuStreamAttachMemAsync(attachedStream,BlocksToAdd, 0,gpuMemAttachSingle) );
CHK_ERR( gpuStreamAttachMemAsync(attachedStream,BlocksToMove, 0,gpuMemAttachSingle) );
CHK_ERR( gpuStreamAttachMemAsync(attachedStream,BlocksRequired, 0,gpuMemAttachSingle) );
CHK_ERR( gpuStreamAttachMemAsync(attachedStream,BlocksRequiredMap, 0,gpuMemAttachSingle) );
// Loop over populations
for (size_t p=0; p<populations.size(); ++p) {
populations[p].blockContainer->gpu_attachToStream(attachedStream);
populations[p].vmesh->gpu_attachToStream(attachedStream);
}
// Also call attach functions on all splitvectors and hashmaps
velocity_block_with_content_list->streamAttach(attachedStream);
velocity_block_with_no_content_list->streamAttach(attachedStream);
BlocksHalo->streamAttach(attachedStream);
BlocksToRemove->streamAttach(attachedStream);
BlocksToAdd->streamAttach(attachedStream);
BlocksToMove->streamAttach(attachedStream);
BlocksRequired->streamAttach(attachedStream);
BlocksRequiredMap->streamAttach(attachedStream);
return;
}
void SpatialCell::gpu_detachFromStream() {
// Return if attaching is not needed
if (!needAttachedStreams) {
return;
}
if (attachedStream == 0) {
// Already detached
return;
}
attachedStream = 0;
CHK_ERR( gpuStreamAttachMemAsync(attachedStream,velocity_block_with_content_list, 0,gpuMemAttachGlobal) );
CHK_ERR( gpuStreamAttachMemAsync(attachedStream,velocity_block_with_no_content_list, 0,gpuMemAttachGlobal) );
CHK_ERR( gpuStreamAttachMemAsync(attachedStream,BlocksHalo, 0,gpuMemAttachGlobal) );
CHK_ERR( gpuStreamAttachMemAsync(attachedStream,BlocksToRemove, 0,gpuMemAttachGlobal) );
CHK_ERR( gpuStreamAttachMemAsync(attachedStream,BlocksToAdd, 0,gpuMemAttachGlobal) );
CHK_ERR( gpuStreamAttachMemAsync(attachedStream,BlocksToMove, 0,gpuMemAttachGlobal) );
CHK_ERR( gpuStreamAttachMemAsync(attachedStream,BlocksRequired, 0,gpuMemAttachGlobal) );
CHK_ERR( gpuStreamAttachMemAsync(attachedStream,BlocksRequiredMap, 0,gpuMemAttachGlobal) );
// Loop over populations
for (size_t p=0; p<populations.size(); ++p) {
populations[p].blockContainer->gpu_detachFromStream();
populations[p].vmesh->gpu_detachFromStream();
}
// Also call detach functions on all splitvectors and hashmaps
velocity_block_with_content_list->streamAttach(0,gpuMemAttachGlobal);
velocity_block_with_no_content_list->streamAttach(0,gpuMemAttachGlobal);
BlocksHalo->streamAttach(0,gpuMemAttachGlobal);
BlocksToRemove->streamAttach(0,gpuMemAttachGlobal);
BlocksToAdd->streamAttach(0,gpuMemAttachGlobal);
BlocksToMove->streamAttach(0,gpuMemAttachGlobal);
BlocksRequired->streamAttach(0,gpuMemAttachGlobal);
BlocksRequiredMap->streamAttach(0,gpuMemAttachGlobal);
return;
}
/** Sends the contents of velocity_block_with_content_list into a device buffer so that it can be accessed
from several streams at once.
*/
void SpatialCell::gpu_uploadContentLists() {
phiprof::start("Upload local content lists");
gpuStream_t stream = gpu_getStream();
velocity_block_with_content_list->copyMetadata(info_vbwcl,stream);
CHK_ERR( gpuStreamSynchronize(stream) );