-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy pathkernel.cu
1525 lines (1368 loc) · 54.2 KB
/
kernel.cu
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
#include "cuda_runtime.h"
#include "device_launch_parameters.h"
#include "cudaKernel.h"
#include "thrust/device_ptr.h"
#include "thrust/remove.h"
#include <stdio.h>
#include <assert.h>
#include <vector>
#include <iostream>
#include "ConstDefine.h"
#define CUDA_CALL(x) { const cudaError_t a = (x); if (a!= cudaSuccess) { printf("\nCUDA Error: %s(err_num=%d)\n", cudaGetErrorString(a), a); cudaDeviceReset(); assert(0);}}
/*
并行计算1个大规模dp
需要提前给定前两次dp的结果，保存在共享内存里
iter: 第几个dp单位；outputIdx：输出结果在全局内存位置；tra1、tra2：两条轨迹，提前被载入共享内存；
*/
//__global__ void DPforward(const int iter, const int* outputIdx,const SPoint *tra1,const SPoint *tra2) {
// SPoint p1 = tra1[threadIdx.x];
// SPoint p2 = tra2[iter - threadIdx.x - 1]; //这样做内存是聚集访问的吗？
// bool subcost;
// if((fabs(p1.x - p2.x) < EPSILON) && (fabs(p1.y - p2.y)<EPSILON)) {
// subcost = 0;
// }
// else
// subcost = 1;
//
//}
/*
SPoint版本
case1：轨迹长度小于512
并行计算n个DP
需要提前给定前两次dp的结果，保存在共享内存里
queryTra[],candidateTra[][]:轨迹
stateTableGPU[][]:对每个candidate的state表
result[]:对于每个candidate的EDR结果
优化方向：
1、轨迹存在share memory里面
2、直接传递轨迹，不使用指针
*/
__global__ void EDRDistance_1(SPoint *queryTra, SPoint **candidateTra, int candidateNum, int queryLength, int *candidateLength, int** stateTableGPU, int *result) {
int blockID = blockIdx.x;
int threadID = threadIdx.x;
if (blockID >= candidateNum) return;
if ((threadID >= candidateLength[blockID]) && (threadID >= queryLength)) return;
const int lenT = candidateLength[blockID];
//int iterNum = queryLength;
//if (lenT > queryLength)
// iterNum = lenT;
const int iterNum = queryLength + lenT - 1;
__shared__ short state[2][MAXTHREAD]; //用于存储前两次的结果
state[0][0] = 0;
state[1][0] = 1;
state[1][1] = 1;
//对两个轨迹排序，保证第一个比第二个短
//首先把轨迹存在共享内存里
__shared__ SPoint queryTraS[MAXTHREAD];
__shared__ SPoint traData[MAXTHREAD];
if (threadID < lenT) {
traData[threadID] = candidateTra[blockID][threadID];
}
if (threadID < queryLength) {
queryTraS[threadID] = queryTra[threadID];
}
const SPoint *tra1, *tra2; //保证tra1比tra2短
int len1, len2;
if (lenT >= queryLength) {
tra1 = queryTraS;
tra2 = traData;
len1 = queryLength;
len2 = lenT;
}
else
{
tra1 = traData;
tra2 = queryTraS;
len1 = lenT;
len2 = queryLength;
}
int myState;
for (int i = 0; i <= iterNum - 1; i++) {//第i轮dp
if (i < len1 - 1) {
if (threadID <= i) {
SPoint p1 = tra1[threadID];
SPoint p2 = tra2[i - threadID]; //这样做内存是聚集访问的吗？
bool subcost;
//if((fabs(p1.x - p2.x) < EPSILON) && (fabs(p1.y - p2.y)<EPSILON)) {
// subcost = 0;
//}
//else
// subcost = 1;
subcost = !((fabs(p1.x - p2.x) < EPSILON) && (fabs(p1.y - p2.y)<EPSILON));
int state_ismatch = state[0][threadID] + subcost;
int state_up = state[1][threadID] + 1;
int state_left = state[1][threadID + 1] + 1;
if (state_ismatch < state_up)
myState = state_ismatch;
else if (state_left < state_up)
myState = state_left;
else
myState = state_up;
//去除if的表达方式，是否可以提升性能？
//myState = (state_ismatch < state_up) * state_ismatch + (state_left < state_up) * state_up + (state_left >= state_up) * state_left;
}
}
else if (i > iterNum - len1) {
if (threadID <= iterNum - i - 1) {
SPoint p1 = tra1[threadID + len1 - (iterNum - i)];
SPoint p2 = tra2[len2 - 1 - threadID]; //这样做内存是聚集访问的吗？
bool subcost;
if ((fabs(p1.x - p2.x) < EPSILON) && (fabs(p1.y - p2.y)<EPSILON)) {
subcost = 0;
}
else
subcost = 1;
int state_ismatch = state[0][threadID + 1] + subcost;
int state_up = state[1][threadID] + 1;
int state_left = state[1][threadID + 1] + 1;
if (state_ismatch < state_up)
myState = state_ismatch;
else if (state_left < state_up)
myState = state_left;
else
myState = state_up;
}
}
else
{
if (threadID < len1) {
SPoint p1 = tra1[threadID];
SPoint p2 = tra2[i - threadID]; //这样做内存是聚集访问的吗？
bool subcost;
if ((fabs(p1.x - p2.x) < EPSILON) && (fabs(p1.y - p2.y)<EPSILON)) {
subcost = 0;
}
else
subcost = 1;
int state_ismatch = state[0][threadID] + subcost;
int state_up = state[1][threadID] + 1;
int state_left = state[1][threadID + 1] + 1;
if (state_ismatch < state_up)
myState = state_ismatch;
else if (state_left < state_up)
myState = state_left;
else
myState = state_up;
}
}
//写myState到share内存,ckecked
int startidx;
//首先将老数据写到全局内存，全写
//startidx是旧的数据应该在全局内存中地址，以i-2计算
//计算应写入全局内存的起始位置
if (i - 2 < len1 - 2) {
startidx = (i - 2 + 2)*(i - 2 + 3) / 2;
if (threadID <= i) {
stateTableGPU[blockID][threadID + startidx] = state[0][threadID];
}
}
else if (i - 2 >= iterNum - len1) {
startidx = (len1 + 1)*(len2 + 1) - (iterNum - (i - 2))*(iterNum - (i - 2) + 1) / 2;
if (threadID <= iterNum - i + 1) {
stateTableGPU[blockID][threadID + startidx] = state[0][threadID];
}
}
else
{
startidx = (len1 + 1)*((i - 2) - (len1 - 2)) + len1*(len1 + 1) / 2;
if (threadID <= len1) {
stateTableGPU[blockID][threadID + startidx] = state[0][threadID];
}
}
//移动新数据到旧数据
state[0][threadID] = state[1][threadID];
//写入新数据
if (i < len1 - 1) {
if (threadID <= i)
state[1][threadID + 1] = myState;
if (threadID == 0) {
state[1][0] = i + 2;
state[1][i + 2] = i + 2;
}
}
else if (i >= iterNum - len1) {
if (threadID <= iterNum - i - 1)
state[1][threadID] = myState;
}
else
{
if (threadID < len1)
state[1][threadID + 1] = myState;
if (threadID == 0) {
state[1][0] = i + 2;
}
}
__syncthreads();
}
//输出结果，最后一次计算一定是由进程0完成的
if (threadID == 0)
result[blockID] = myState;
}
//__global__ void testSharedMemory()
//{
// __shared__ SPoint queryTraS[MAXLENGTH];
// __shared__ SPoint traData[MAXLENGTH];
// __shared__ SPoint traData2[MAXLENGTH];
// SPoint s;
// s.x = 4;
// s.y = 5;
// traData[1535] = s;
// queryTraS[1535] = s;
// traData2[1535] = s;
//}
/*
SPoint版本
同时处理若干个query的EDR，这里以一个EDR计算为单位，每个block计算一个EDR，thread负责一条斜线上state的并行计算。
case1：轨迹长度可长于512，利用循环处理多余512的。
并行计算n个DP
需要提前给定前两次dp的结果，保存在共享内存里
queryTaskNum:总共有几个EDR计算任务
queryTaskInfo[]：每个task对应的qID、candidateID信息，用struct存储
queryTra[],candidateTra[]:轨迹数据，candidateTra保证其内部轨迹不重复
queryTraOffset[],candidateTraOffset[]:每条轨迹的offset，candidateTra保证其内部轨迹不重复
queryLength[],candidateLength[]:每条轨迹的长度（其实offset相减就是长度），其idx和上面的对应
即：candidateLength[id]是第id个candidate Traj的长度
stateTableGPU[][]:对每个candidate的state表
result[]:对于每个candidate的EDR结果
优化方向：
1、轨迹存在share memory里面
2、直接传递轨迹，不使用指针
*/
// EDRDistance_Batch_Handler(validCandTrajNum, taskInfoTableGPU, queryTraGPUBase, queryTraOffsetGPU, candidateOffsetsGPU, queryLengthGPU, candidateTraLengthGPU, resultReturnedGPU, &defaultStream);
__global__ void EDRDistance_Batch(int queryTaskNum, TaskInfoTableForSimilarity* taskInfoTable, SPoint *queryTra, int* queryTraOffset, SPoint** candidateTraOffsets, int* queryLength, int *candidateLength, int *result) {
int blockID = blockIdx.x;
int threadID = threadIdx.x;
if (blockID >= queryTaskNum) return;
__shared__ int thisQueryID;
__shared__ int thisQueryLength;
__shared__ int lenT;
__shared__ int iterNum;
thisQueryID = taskInfoTable[blockID].qID;
if(thisQueryID < 0 ) return; // block保持空闲
__shared__ short state[2][MAXLENGTH + 1];
__shared__ SPoint *queryTraS;
__shared__ SPoint *traData;
__shared__ SPoint *tra1, *tra2;
__shared__ int len1, len2;
if (threadID == 0) {
thisQueryID = taskInfoTable[blockID].qID;
thisQueryLength = queryLength[thisQueryID];
lenT = candidateLength[blockID];
iterNum = thisQueryLength + lenT - 1;
state[0][0] = 0;
// state[0][1] = 1;
state[1][0] = 1;
state[1][1] = 1;
queryTraS = queryTra + queryTraOffset[thisQueryID];
traData = candidateTraOffsets[blockID];
if (lenT >= thisQueryLength) {
tra1 = queryTraS;
tra2 = traData;
len1 = thisQueryLength;
len2 = lenT;
}
else
{
tra1 = traData;
tra2 = queryTraS;
len1 = lenT;
len2 = thisQueryLength;
}
}
__syncthreads(); // 同步保证其他线程空闲
if ((threadID >= lenT) && (threadID >= thisQueryLength)) return;
//__shared__ SPoint queryTraS[MAXLENGTH];
//__shared__ SPoint traData[MAXLENGTH];
//for (int i = 0; i <= lenT - 1;i+=MAXTHREAD)
//{
// if(threadID+i<lenT)
// {
// traData[threadID + i] = SPoint(candidateTraOffsets[blockID][threadID + i]);
// }
//}
//SPoint* queryTraBaseAddr = queryTra + queryTraOffset[thisQueryID];
//for (int i = 0; i <= thisQueryLength - 1;i+=MAXTHREAD)
//{
// if(threadID+i<thisQueryLength)
// {
// queryTraS[threadID + i] = *(queryTraBaseAddr + threadID + i);
// }
//}
int myState[5];// 256*4 = 1024 非 __shared 类型
int nodeID;
SPoint p1;
SPoint p2;
bool subcost;
for (int i = 0; i <= iterNum - 1; i++) { // block 层次循环
if (i < len1 - 1) {
for (int startIdx = 0; startIdx <= i; startIdx += MAXTHREAD) {
nodeID = startIdx + threadID;
if (nodeID <= i) {
p1 = tra1[nodeID]; // fetch from global memory
p2 = tra2[i - nodeID]; // fetch from global memory
// id1 + id2 = i
if ((fabs(p1.x - p2.x) < EPSILON) && (fabs(p1.y - p2.y)<EPSILON)) {
subcost = 0;
}
else
subcost = 1;
//subcost = !((fabs(p1.x - p2.x) < EPSILON) && (fabs(p1.y - p2.y)<EPSILON));
bool c1 = ((state[0][nodeID] + subcost < (state[1][nodeID] + 1)) && (state[0][nodeID] + subcost < (state[1][nodeID + 1] + 1)));
bool c2 = (((state[1][nodeID + 1] + 1) < (state[1][nodeID] + 1)) && (((state[1][nodeID + 1] + 1) < state[0][nodeID] + subcost)));
//去除if的表达方式，是否可以提升性能？
myState[nodeID / MAXTHREAD] = c1 * (state[0][nodeID] + subcost) + c2 * (state[1][nodeID + 1] + 1) + !(c1 || c2) * (state[1][nodeID] + 1);
//if ((state_ismatch < state_up) && (state_ismatch < state_left))
// myState[nodeID/MAXTHREAD] = state_ismatch;
//else if ((state_left < state_up) && ((state_left < state_ismatch)))
// myState[nodeID / MAXTHREAD] = state_left;
//else
// myState[nodeID / MAXTHREAD] = state_up;
////去除if的表达方式，是否可以提升性能？
//myState[nodeID / MAXTHREAD] = (state_ismatch < state_up) && (state_ismatch < state_left) * state_ismatch + ((state_left < state_up) && ((state_left < state_ismatch))) * state_left + !(((state_ismatch < state_up) && (state_ismatch < state_left))||(((state_left < state_up) && ((state_left < state_ismatch))))) * state_up;
}
}
}
else if (i > iterNum - len1) {
for (int startIdx = 0; startIdx <= iterNum - i - 1; startIdx += MAXTHREAD) {
nodeID = startIdx + threadID;
if (nodeID <= iterNum - i - 1) {
// EDR定义理解问题
p1 = tra1[nodeID + len1 - (iterNum - i)]; // fetch from global memory
p2 = tra2[len2 - 1 - nodeID];
if ((fabs(p1.x - p2.x) < EPSILON) && (fabs(p1.y - p2.y)<EPSILON)) {
subcost = 0;
}
else
subcost = 1;
//if (state_ismatch < state_up)
// myState[nodeID / MAXTHREAD] = state_ismatch;
//else if (state_left < state_up)
// myState[nodeID / MAXTHREAD] = state_left;
//else
// myState[nodeID / MAXTHREAD] = state_up;
bool c1 = (((state[0][nodeID + 1] + subcost) < (state[1][nodeID] + 1)) && ((state[0][nodeID + 1] + subcost) < (state[1][nodeID + 1] + 1)));
bool c2 = (((state[1][nodeID + 1] + 1) < (state[1][nodeID] + 1)) && (((state[1][nodeID + 1] + 1) < (state[0][nodeID + 1] + subcost))));
//去除if的表达方式，是否可以提升性能？
myState[nodeID / MAXTHREAD] = c1 * (state[0][nodeID + 1] + subcost) + c2 * (state[1][nodeID + 1] + 1) + !(c1 || c2) * (state[1][nodeID] + 1);
}
}
}
else
{
for (int startIdx = 0; startIdx < len1; startIdx += MAXTHREAD) {
nodeID = startIdx + threadID;
if (nodeID < len1) { // 注意条件判断 保证分配的线程个数正好覆盖计算 否则线程空闲
p1 = tra1[nodeID]; // fetch from global memory
p2 = tra2[i - nodeID]; //这样做内存是聚集访问的吗?
if ((fabs(p1.x - p2.x) < EPSILON) && (fabs(p1.y - p2.y)<EPSILON)) {
subcost = 0;
}
else
subcost = 1;
//int state_ismatch = (state[0][nodeID] + subcost);
//int state_up = (state[1][nodeID] + 1);
//int state_left = (state[1][nodeID + 1] + 1);
//if (state_ismatch < state_up)
// myState[nodeID / MAXTHREAD] = state_ismatch;
//else if (state_left < state_up)
// myState[nodeID / MAXTHREAD] = state_left;
//else
// myState[nodeID / MAXTHREAD] = state_up;
bool c1 = (((state[0][nodeID] + subcost) < (state[1][nodeID] + 1)) && ((state[0][nodeID] + subcost) < (state[1][nodeID + 1] + 1)));
bool c2 = (((state[1][nodeID + 1] + 1) < (state[1][nodeID] + 1)) && (((state[1][nodeID + 1] + 1) < (state[0][nodeID] + subcost))));
//去除if的表达方式，是否可以提升性能？
myState[nodeID / MAXTHREAD] = c1 * (state[0][nodeID] + subcost) + c2 * (state[1][nodeID + 1] + 1) + !(c1 || c2) * (state[1][nodeID] + 1);
}
}
}
// 更新state[0]
//state[1] 给 state[0]
for (int Idx = 0; Idx < MAXLENGTH; Idx += MAXTHREAD)
{
if(threadID + Idx < MAXLENGTH)
state[0][threadID + Idx] = state[1][threadID + Idx];
}
//state[0][threadID] = state[1][threadID];
// 更新state[1]
//写入新数据
if (i < len1 - 1) {
for (int Idx = 0; Idx <= i; Idx += MAXTHREAD) {
if (threadID + Idx <= i)
state[1][Idx + threadID + 1] = myState[Idx / MAXTHREAD];
}
// 增长阶段
if (threadID == 0) {
state[1][0] = i + 2; // 更新头
state[1][i + 2] = i + 2; // 更新尾
}
}
else if (i >= iterNum - len1) {
//if (threadID <= iterNum - i - 1)
// state[1][threadID] = myState;
for (int Idx = 0; Idx <= iterNum - i - 1; Idx += MAXTHREAD) {
if (threadID + Idx <= iterNum - i - 1)
state[1][threadID + Idx] = myState[Idx / MAXTHREAD];
}
// 缩减阶段不用更新
}
else
{
//if (threadID < len1)
// state[1][threadID + 1] = myState;
//if (threadID == 0) {
// state[1][0] = i + 2;
//}
for (int Idx = 0; Idx <= len1; Idx += MAXTHREAD) {
if (threadID + Idx < len1)
state[1][Idx + threadID + 1] = myState[Idx / MAXTHREAD];
}
// 保持阶段
if (threadID == 0) {
state[1][0] = i + 2; // 只需更新头
}
}
__syncthreads(); // 同步一个block的thread
}
// kernel 应该没什么大问题
// for 循环结束 则计算EDR完成
// myState[0]; 就是EDR
if (threadID == 0 && blockID < queryTaskNum)
result[blockID] = myState[0];
// std::cout << "calc EDR success!\n" ;
}
// EDRDistance_Batch_Handler(validCandTrajNum, taskInfoTableGPU, queryTraGPUBase, queryTraOffsetGPU, candidateOffsetsGPU, queryLengthGPU, candidateTraLengthGPU, resultReturnedGPU, &defaultStream);
int EDRDistance_Batch_Handler(int queryTaskNum, TaskInfoTableForSimilarity* taskInfoTable, SPoint *queryTra, int* queryTraOffset, SPoint** candidateTraOffsets, int* queryLength, int *candidateLength, int *result, cudaStream_t *stream)
{
//printf("run kernel now\n");
EDRDistance_Batch <<< queryTaskNum, MAXTHREAD, 0, *stream >> >(queryTaskNum, taskInfoTable, queryTra, queryTraOffset, candidateTraOffsets, queryLength, candidateLength, result);
//CUDA_CALL(cudaDeviceSynchronize());
return 0;
}
__device__ inline int binary_search_intPair(intPair* temp, int left, int right, int val)
{
int mid = (left + right) / 2;
while (left <= right)
{
mid = (left + right) / 2;
if (temp[mid].int_1 == val)
return temp[mid].int_2;
else if (temp[mid].int_1 > val)
{
right = mid - 1;
}
else
left = mid + 1;
}
return 0;
}
__device__ inline int binary_search_intPair_Neighbor(intPair* temp, int left, int right, int val)
{
int mid = (left + right) / 2;
while (left <= right)
{
mid = (left + right) / 2;
if (temp[mid].int_1 == val)
return mid;
else if (temp[mid].int_1 > val)
{
right = mid - 1;
}
else
left = mid + 1;
}
return -1;
}
// -1为没找到
__device__ inline int binary_search_int(int* temp, int left, int right, int val)
{
int mid = (left + right) / 2;
while (left <= right)
{
mid = (left + right) / 2;
if (temp[mid] == val)
return mid;
else if (temp[mid] > val)
{
right = mid - 1;
}
else
left = mid + 1;
}
return -1;
}
__device__ inline int getIdxFromXYGPU(int x, int y)
{
int lenx, leny;
if (x == 0)
lenx = 1;
else
{
lenx = int(log2f(x)) + 1;
}
if (y == 0)
leny = 1;
else
leny = int(log2f(y)) + 1;
int result = 0;
int xbit = 1, ybit = 1;
for (int i = 1; i <= 2 * max(lenx, leny); i++)
{
if ((i & 1) == 1) //奇数
{
result += (x >> (xbit - 1) & 1) * (1 << (i - 1));
xbit = xbit + 1;
}
else //偶数
{
result += (y >> (ybit - 1) & 1) * (1 << (i - 1));
ybit = ybit + 1;
}
}
return result;
}
__device__ inline int findNeighborGPU(int cellNum, int cellID, int * neighborID)
{
int x = 0, y = 0;
for (int bit = 0; bit <= int(log2f(cellNum)) - 1; bit++) {
if (bit % 2 == 0) {
//奇数位
x += ((cellID >> bit)&(1))*(1 << (bit / 2));
}
else {
//偶数位
y += ((cellID >> bit)&(1))*(1 << (bit / 2));
}
}
int cnt = 0;
for (int xx = x - 1; xx <= x + 1; xx++) {
for (int yy = y - 1; yy <= y + 1; yy++) {
if ((xx != x) || (yy != y))
neighborID[cnt++] = getIdxFromXYGPU(xx, yy);
//printf("%d\t", cnt);
}
}
return 0;
}
__device__ inline bool isPositive(short x)
{
return x >= 0;
}
__global__ void Calculate_FD_Sparse(intPair* queryFVGPU, intPair* FVinfo, intPair* FVTable, intPair* SubbedArray, intPair* SubbedArrayOffset, int SubbedArrayJump, int queryCellLength, int startTrajIdx, int checkNum, int cellNum, int trajNumInDB, int nonZeroFVNumInDB, short* FDistance)
{
//第一阶段：并行减法
const int MAX_QUERY_CELLNUMBER = 512;
int blockID = blockIdx.x;
int threadID = threadIdx.x;
int threadIDGlobal = blockDim.x*blockID + threadID;
__shared__ intPair queryCellTraj[MAX_QUERY_CELLNUMBER];
__shared__ intPair dbCellTraj[MAX_QUERY_CELLNUMBER];
//cellchecked记录在query中出现的cell编号，用于在反向减法的时候检查是不是已经减过了。以后可以在归并加中复用此变量。
__shared__ int cellChecked[MAX_QUERY_CELLNUMBER];
for (int i = 0; i <= queryCellLength - 1; i += MAXTHREAD) {
if (threadID + i < queryCellLength)
{
queryCellTraj[threadID + i] = queryFVGPU[threadID + i];
}
}
int dbTrajStartIdx = FVinfo[startTrajIdx + blockID].int_2;
int dbTrajEndIdx;
if (blockID + startTrajIdx == trajNumInDB - 1)
dbTrajEndIdx = nonZeroFVNumInDB - 1;
else
dbTrajEndIdx = FVinfo[startTrajIdx + blockID + 1].int_2 - 1;
for (int i = 0; i <= dbTrajEndIdx - dbTrajStartIdx; i += MAXTHREAD)
{
if (threadID + i <= dbTrajEndIdx - dbTrajStartIdx)
dbCellTraj[threadID + i] = FVTable[dbTrajStartIdx + threadID + i];
}
//1.1:用query减去db
for (int i = 0; i < queryCellLength; i += MAXTHREAD)
{
if (threadID + i < queryCellLength) {
int find = binary_search_intPair(dbCellTraj, 0, dbTrajEndIdx - dbTrajStartIdx, queryCellTraj[threadID + i].int_1);
cellChecked[threadID + i] = queryCellTraj[threadID + i].int_1;
SubbedArray[SubbedArrayJump * blockID + threadID + i].int_1 = queryCellTraj[threadID + i].int_1;
SubbedArray[SubbedArrayJump * blockID + threadID + i].int_2 = queryCellTraj[threadID + i].int_2 - find;
}
if (threadID == 0) {
SubbedArrayOffset[blockID].int_1 = queryCellLength - 1;
SubbedArrayOffset[blockID].int_2 = queryCellLength + dbTrajEndIdx - dbTrajStartIdx;
}
}
//1.2：用db减去query，注意加负号
for (int i = 0; i <= dbTrajEndIdx - dbTrajStartIdx; i += MAXTHREAD)
{
if (threadID + i <= dbTrajEndIdx - dbTrajStartIdx)
{
intPair cellNo = dbCellTraj[threadID + i];
int find = binary_search_int(cellChecked, 0, queryCellLength - 1, cellNo.int_1);
if (find == -1)
{
SubbedArray[SubbedArrayJump * blockID + queryCellLength + threadID + i].int_1 = cellNo.int_1;
SubbedArray[SubbedArrayJump * blockID + queryCellLength + threadID + i].int_2 = -cellNo.int_2;
}
else
SubbedArray[SubbedArrayJump * blockID + queryCellLength + threadID + i].int_1 = -1;
}
}
__syncthreads();
//第二阶段：查找相邻，做减法
//这个阶段改为每个thread处理一个FD
//2.1：合并每个subbedArray
if (threadIDGlobal < checkNum) {
int startMergeIdx = SubbedArrayOffset[threadIDGlobal].int_1 + 1;
int endMergeIdx = SubbedArrayOffset[threadIDGlobal].int_2;
int frontPtr = startMergeIdx;
for (int i = startMergeIdx; i <= endMergeIdx; i++)
{
if (SubbedArray[SubbedArrayJump * threadIDGlobal + i].int_1 != -1)
{
SubbedArray[SubbedArrayJump * threadIDGlobal + frontPtr] = SubbedArray[SubbedArrayJump * threadIDGlobal + i];
frontPtr++;
}
}
SubbedArrayOffset[threadIDGlobal].int_2 = frontPtr - 1;
}
//2.2 查找相邻
int neighborsID[8];
//cell单纯指第几个元素
for (int cell = 0; cell <= SubbedArrayOffset[threadIDGlobal].int_2; cell++)
{
findNeighborGPU(cellNum, cell, neighborsID);
//for (int i = 0; i <= 7; i++)
// neighborsID[i] = 11;
for (int i = 0; i <= 7; i++)
{
int find = binary_search_intPair_Neighbor(&SubbedArray[SubbedArrayJump * threadIDGlobal], 0, SubbedArrayOffset[threadIDGlobal].int_1, neighborsID[i]);
if (find == -1) {
find = binary_search_intPair_Neighbor(&SubbedArray[SubbedArrayJump * threadIDGlobal], SubbedArrayOffset[threadIDGlobal].int_1 + 1, SubbedArrayOffset[threadIDGlobal].int_2, neighborsID[i]);
}
// 如果是-1，说明这个neighbor是0，不用处理
if (find != -1)
{
if (isPositive(SubbedArray[SubbedArrayJump * threadIDGlobal + cell].int_2) != isPositive(SubbedArray[SubbedArrayJump * threadIDGlobal + find].int_2))
{
if (fabsf(SubbedArray[SubbedArrayJump * threadIDGlobal + cell].int_2) > fabsf(SubbedArray[SubbedArrayJump * threadIDGlobal + find].int_2))
{
SubbedArray[SubbedArrayJump * threadIDGlobal + cell].int_2 = SubbedArray[SubbedArrayJump * threadIDGlobal + cell].int_2 + SubbedArray[SubbedArrayJump * threadIDGlobal + find].int_2;
SubbedArray[SubbedArrayJump * threadIDGlobal + find].int_2 = 0;
}
else {
SubbedArray[SubbedArrayJump * threadIDGlobal + find].int_2 = SubbedArray[SubbedArrayJump * threadIDGlobal + find].int_2 + SubbedArray[SubbedArrayJump * threadIDGlobal + cell].int_2;
SubbedArray[SubbedArrayJump * threadIDGlobal + cell].int_2 = 0;
break;
}
}
}
}
}
__syncthreads();
//第三阶段：统计正负个数
//依然是每个block负责一个FD的计算
if (blockID >= checkNum)
return;
int *tempsumPosi = (int*)queryCellTraj;
int *tempsumNega = (int*)dbCellTraj;
tempsumPosi[threadID] = 0;
tempsumNega[threadID] = 0;
for (int i = 0; i <= SubbedArrayOffset[blockID].int_2; i += MAXTHREAD)
{
if (i + threadID <= SubbedArrayOffset[blockID].int_2)
{
tempsumPosi[threadID] += (isPositive(SubbedArray[SubbedArrayJump * blockID + i + threadID].int_2)*SubbedArray[SubbedArrayJump * blockID + i + threadID].int_2);
tempsumNega[threadID] += (-(!isPositive(SubbedArray[SubbedArrayJump * blockID + i + threadID].int_2))*SubbedArray[SubbedArrayJump * blockID + i + threadID].int_2);
}
}
__shared__ int sizeOfTempSum;
if (threadID == 0)
sizeOfTempSum = MAXTHREAD;
__syncthreads();
while ((sizeOfTempSum>1))
{
if (threadID <= (sizeOfTempSum >> 1) - 1)
{
tempsumPosi[threadID] = tempsumPosi[threadID] + tempsumPosi[threadID + (sizeOfTempSum >> 1)];
tempsumNega[threadID] = tempsumNega[threadID] + tempsumNega[threadID + (sizeOfTempSum >> 1)];
}
__syncthreads();
if (threadID == 0)
sizeOfTempSum = (sizeOfTempSum >> 1);
__syncthreads();
}
if (threadID == 0)
FDistance[blockID] = (tempsumPosi[0] > tempsumNega[0]) ? tempsumPosi[0] : tempsumNega[0];
}
//每个block负责一个FD的计算
__global__ void Calculate_FD_NonColumn(short* queryFVGPU, intPair* FVinfo, intPair* FVTable, int startTrajIdx, int checkNum, int cellNum, int trajNumInDB, int nonZeroFVNumInDB, short* FDistance)
{
//第一阶段：并行减法
int blockID = blockIdx.x;
int threadID = threadIdx.x;
int threadIDGlobal = blockDim.x*blockID + threadID;
if (blockID >= checkNum)
return;
__shared__ intPair taskInfo;
if (threadID == 0)
taskInfo = FVinfo[blockID + startTrajIdx];
int nextCnt;
if (blockID + startTrajIdx == trajNumInDB - 1)
nextCnt = nonZeroFVNumInDB;
else
nextCnt = FVinfo[blockID + startTrajIdx + 1].int_2;
__syncthreads();
for (int i = 0; i <= (cellNum - 1); i += MAXTHREAD)
{
int find = binary_search_intPair(FVTable, taskInfo.int_2, (nextCnt - 1), (i + threadID));
//int find = 1;
//int k = cellNum*blockID + (i + threadID);
//queryFVGPU[cellNum*blockID + (i + threadID)] = 2;
queryFVGPU[cellNum*blockID + (i + threadID)] = queryFVGPU[cellNum*blockID + (i + threadID)] - find;
}
//第二阶段：查找相邻，做减法
//这个阶段改为每个thread处理一个FD
int neighborsID[8];
for (int cell = 0; cell <= cellNum - 1; cell++)
{
//只需要一部分线程就行了
if (threadIDGlobal >= checkNum)
break;
if (queryFVGPU[cellNum*threadIDGlobal + cell] != 0)
{
findNeighborGPU(cellNum, cell, neighborsID);
//for (int i = 0; i <= 7; i++)
// neighborsID[i] = 11;
for (int i = 0; i <= 7; i++)
{
if (isPositive(queryFVGPU[cellNum*threadIDGlobal + cell]) != isPositive(queryFVGPU[cellNum*threadIDGlobal + neighborsID[i]])) {
if (fabsf(queryFVGPU[cellNum*threadIDGlobal + cell]) > fabsf(queryFVGPU[cellNum*threadIDGlobal + neighborsID[i]]))
{
queryFVGPU[cellNum*threadIDGlobal + cell] = queryFVGPU[cellNum*threadIDGlobal + cell] + queryFVGPU[cellNum*threadIDGlobal + neighborsID[i]];
queryFVGPU[cellNum*threadIDGlobal + neighborsID[i]] = 0;
}
else
{
queryFVGPU[cellNum*threadIDGlobal + neighborsID[i]] = queryFVGPU[cellNum*threadIDGlobal + neighborsID[i]] + queryFVGPU[cellNum*threadIDGlobal + cell];
queryFVGPU[cellNum*threadIDGlobal + cell] = 0;
break;
}
}
}
}
}
__syncthreads();
//第三阶段：统计正负个数
//依然是每个block负责一个FD的计算
__shared__ int tempsumPosi[MAXTHREAD], tempsumNega[MAXTHREAD];
tempsumPosi[threadID] = 0;
tempsumNega[threadID] = 0;
for (int i = 0; i <= cellNum - 1; i += MAXTHREAD)
{
tempsumPosi[threadID] += (isPositive(queryFVGPU[blockID*cellNum + (i + threadID)])*queryFVGPU[blockID*cellNum + (i + threadID)]);
tempsumNega[threadID] += (-(!isPositive(queryFVGPU[blockID*cellNum + (i + threadID)]))*queryFVGPU[blockID*cellNum + (i + threadID)]);
}
__shared__ int sizeOfTempSum;
if (threadID == 0)
sizeOfTempSum = MAXTHREAD;
__syncthreads();
while ((sizeOfTempSum>1))
{
if (threadID <= (sizeOfTempSum >> 1) - 1)
{
tempsumPosi[threadID] = tempsumPosi[threadID] + tempsumPosi[threadID + (sizeOfTempSum >> 1)];
tempsumNega[threadID] = tempsumNega[threadID] + tempsumNega[threadID + (sizeOfTempSum >> 1)];
}
__syncthreads();
if (threadID == 0)
sizeOfTempSum = (sizeOfTempSum >> 1);
__syncthreads();
}
if (threadID == 0)
FDistance[blockID] = (tempsumPosi[0] > tempsumNega[0]) ? tempsumPosi[0] : tempsumNega[0];
}
//SubbedArrayJump是SubbedArray中每一行有多少个元素，供计算idx用
int Similarity_Pruning_Handler(intPair* queryFVGPU, intPair* FVinfo, intPair* FVTable, intPair* SubbedArray, intPair* SubbedArrayOffset, int SubbedArrayJump, int queryCellLength, int startTrajIdx, int checkNum, int cellNum, int trajNumInDB, int nonZeroFVNumInDB, short* FDistance, cudaStream_t stream)
{
#ifdef NOT_COLUMN_ORIENTED
Calculate_FD_NonColumn << <checkNum, MAXTHREAD, 0, stream >> >(queryFVGPU, FVinfo, FVTable, startTrajIdx, checkNum, cellNum, trajNumInDB, nonZeroFVNumInDB, FDistance);
#else
Calculate_FD_Sparse << <checkNum, MAXTHREAD, 0, stream >> >(queryFVGPU, FVinfo, FVTable, SubbedArray, SubbedArrayOffset, SubbedArrayJump, queryCellLength, startTrajIdx, checkNum, cellNum, trajNumInDB, nonZeroFVNumInDB, FDistance);
#endif
return 0;
}
/*
//先按照能否用一个SM执行一个DP来划分任务，再分别调用两种kernel
//constructing...
可优化：
1、queryTra、queryLength甚至candidateLength可以通过传值的方式直接传递到SM的寄存器，减少全局内存的使用
*/
int handleEDRdistance(SPoint *queryTra, SPoint **candidateTra, int candidateNum, int queryLength, int *candidateLength, int *result) {
MyTimer time1;
time1.start();
int** stateTableGPU = NULL;
//在GPU中为状态表分配内存
int** temp = NULL;
temp = (int**)malloc(sizeof(int*)*candidateNum);
for (int i = 0; i <= candidateNum - 1; i++) {
CUDA_CALL(cudaMalloc((void**)&temp[i], sizeof(int)*(candidateLength[i] + 1)*(queryLength + 1)));
}
CUDA_CALL(cudaMalloc((void***)&stateTableGPU, sizeof(int*)*candidateNum));
CUDA_CALL(cudaMemcpy(stateTableGPU, temp, candidateNum*sizeof(int*), cudaMemcpyHostToDevice));
//为存储的轨迹信息分配内存
SPoint *queryTraGPU = NULL, **candidateTraGPU = NULL;
int *candidateLengthGPU = NULL, *resultGPU = NULL;
CUDA_CALL(cudaMalloc((void**)&queryTraGPU, sizeof(SPoint)*queryLength));
CUDA_CALL(cudaMalloc((void**)&candidateLengthGPU, sizeof(int)*candidateNum));
//CUDA_CALL(cudaMalloc((void**)&resultGPU, sizeof(int)*candidateNum));
SPoint **tempS = (SPoint**)malloc(sizeof(SPoint*)*candidateNum);
for (int i = 0; i <= candidateNum - 1; i++) {
CUDA_CALL(cudaMalloc((void**)&tempS[i], sizeof(SPoint)*candidateLength[i]));
}
CUDA_CALL(cudaMalloc((void***)&candidateTraGPU, sizeof(SPoint*)*candidateNum));
CUDA_CALL(cudaMemcpy(candidateTraGPU, tempS, candidateNum*sizeof(SPoint*), cudaMemcpyHostToDevice));
//
time1.stop();
std::cout << time1.elapse() << std::endl;
time1.start();
//
//最好通过传参数的方法传递轨迹，这就要求轨迹连续存储
//向GPU传递轨迹信息
CUDA_CALL(cudaMemcpy(queryTraGPU, queryTra, queryLength*sizeof(SPoint), cudaMemcpyHostToDevice));
CUDA_CALL(cudaMemcpy(candidateLengthGPU, candidateLength, candidateNum*sizeof(int), cudaMemcpyHostToDevice));
for (int i = 0; i <= candidateNum - 1; i++) {
CUDA_CALL(cudaMemcpy(tempS[i], candidateTra[i], candidateLength[i] * sizeof(SPoint), cudaMemcpyHostToDevice));
}
//for (int i = 0; i <= candidateNum - 1;i++)
// CUDA_CALL(cudaMemcpy(candidateTraGPU[i], candidateTra[i], candidateLength[i]*sizeof(SPoint), cudaMemcpyHostToDevice));
CUDA_CALL(cudaHostAlloc((void**)&result, candidateNum*sizeof(int), cudaHostAllocWriteCombined | cudaHostAllocMapped));
CUDA_CALL(cudaHostGetDevicePointer(&resultGPU, result, 0));
time1.stop();
std::cout << time1.elapse() << std::endl;
time1.start();
//执行kernel
EDRDistance_1 << <candidateNum, MAXTHREAD >> >(queryTraGPU, candidateTraGPU, candidateNum, queryLength, candidateLengthGPU, stateTableGPU, resultGPU);
//取结果
//result = (int*)malloc(candidateNum*sizeof(int));
//CUDA_CALL(cudaMemcpy(result, resultGPU, candidateNum*sizeof(int), cudaMemcpyDeviceToHost));
cudaDeviceSynchronize();
// for (int j = 0; j <= candidateNum - 1;j++)
// std::cout << result[j] << std::endl;
//free GPU!!!!!
time1.stop();
std::cout << time1.elapse() << std::endl;
return 0;
}
inline void __getLastCudaError(const char *errorMessage, const char *file, const int line)
{
cudaError_t err = cudaGetLastError();
if (cudaSuccess != err)
{
fprintf(stderr, "%s(%i) : getLastCudaError() CUDA error : %s : (%d) %s.\n",
file, line, errorMessage, (int)err, cudaGetErrorString(err));
exit(EXIT_FAILURE);
}
}
//using namespace thrust;
//static const int MAXTHREAD = 512; //每个block线程数
cudaError_t addWithCuda(int *c, const int *a, const int *b, unsigned int size);
void CUDAwarmUp() {
CUDA_CALL(cudaSetDeviceFlags(cudaDeviceMapHost));
CUDA_CALL(cudaSetDevice(0));
}
#ifdef _CELL_BASED_STORAGE
int putCellDataSetIntoGPU(Point* pointsPtr, Point*& pointsPtrGPU, int pointNum) {
CUDA_CALL(cudaMalloc((void**)&pointsPtrGPU, pointNum * sizeof(Point))); //分配数据的内存
//debug
//std::cout << pointNum << std::endl;
//debug
CUDA_CALL(cudaMemcpy(pointsPtrGPU, pointsPtr, pointNum * sizeof(Point), cudaMemcpyHostToDevice));//数据拷贝到gpu里
return 0;
}
__global__ void cudaRangeQuery(int* rangeStarts, int* rangeEnds, int candidateCellNum, const Point* pointsPtr, const float xmin, const float ymin, const float xmax, const float ymax, const int *resultOffset, Point* resultPtrCuda) {
int cellNo = blockIdx.x; //candidate里面第几个cell 0,1,2,....
if (cellNo >= candidateCellNum) return;
int tid = threadIdx.x;
if (tid >= 256) return;
int pointNum = rangeEnds[cellNo] - rangeStarts[cellNo] + 1;//block要处理的这个cell有这么多个点
const int offset = rangeStarts[cellNo];
for (int i = tid; i <= pointNum - 1; i += MAXTHREAD) {
float x = pointsPtr[offset + i].x;
float y = pointsPtr[offset + i].y;
uint32_t tid = pointsPtr[offset + i].tID;
uint32_t time = pointsPtr[offset + i].time;
if (x <= xmax &&x >= xmin&&y <= ymax&&y >= ymin) {
resultPtrCuda[resultOffset[cellNo] + i].x = x;
resultPtrCuda[resultOffset[cellNo] + i].y = y;
resultPtrCuda[resultOffset[cellNo] + i].tID = tid;
resultPtrCuda[resultOffset[cellNo] + i].time = time;
}
else
resultPtrCuda[resultOffset[cellNo] + i].tID = -1;
}
}
__global__ void cudaRangeQueryTest(RangeQueryStateTable* stateTable, int stateTableLength, uint8_t* result,
const int maxTrajNum) {
int bID = blockIdx.x;
int tID = threadIdx.x;
__shared__ RangeQueryStateTable sharedStateTable;
// __shared__ uint8_t resultTemp[10000]; //10K
if(tID == 0)
sharedStateTable = (stateTable[bID]); // 意义不大 本来就在类似常量内存中
__syncthreads();//block内thread同步
/*
// 4+4*7=32byte
typedef struct RangeQueryStateTable {// leafnode
// GPU相关
void* ptr; // 指向GPU内存node的指针 连续存储
int candidatePointNum; // 是这个leafnode中的节点数
float xmin;
float ymin;
float xmax;
float ymax;
// cpu相关
int queryID;
int startIdxInAllPoints; //startId in AllPoints数组
}RangeQueryStateTable;*/
int jobID = sharedStateTable.queryID;