-
-
Notifications
You must be signed in to change notification settings - Fork 56.5k
Expand file tree
/
Copy pathgstreamingexecutor.cpp
More file actions
1947 lines (1772 loc) · 76.7 KB
/
gstreamingexecutor.cpp
File metadata and controls
1947 lines (1772 loc) · 76.7 KB
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 OpenCV project.
// It is subject to the license terms in the LICENSE file found in the top-level directory
// of this distribution and at http://opencv.org/license.html.
//
// Copyright (C) 2019-2021 Intel Corporation
#include "precomp.hpp"
#include <memory> // make_shared
#include <ade/util/zip_range.hpp>
#include <opencv2/gapi/opencv_includes.hpp>
#if !defined(GAPI_STANDALONE)
#include <opencv2/gapi/core.hpp> // GCopy -- FIXME - to be removed!
#endif // GAPI_STANDALONE
#include "utils/itt.hpp"
#include "api/gproto_priv.hpp" // ptr(GRunArgP)
#include "compiler/passes/passes.hpp"
#include "backends/common/gbackend.hpp" // createMat
#include "backends/streaming/gstreamingbackend.hpp" // GCopy
#include "compiler/gcompiler.hpp" // for compileIslands
#include <logger.hpp>
#include "executor/gstreamingexecutor.hpp"
#include <opencv2/gapi/streaming/meta.hpp>
#include <opencv2/gapi/streaming/sync.hpp>
#include <opencv2/gapi/util/variant.hpp>
namespace
{
using namespace cv::gimpl::stream;
#if !defined(GAPI_STANDALONE)
class VideoEmitter final: public cv::gimpl::GIslandEmitter {
cv::gapi::wip::IStreamSource::Ptr src;
virtual void halt() override {
src->halt();
}
virtual bool pull(cv::GRunArg &arg) override {
// FIXME: probably we can maintain a pool of (then) pre-allocated
// buffers to avoid runtime allocations.
// Pool size can be determined given the internal queue size.
cv::gapi::wip::Data newData;
if (!src->pull(newData)) {
return false;
}
arg = std::move(static_cast<cv::GRunArg&>(newData));
return true;
}
public:
explicit VideoEmitter(const cv::GRunArg &arg) {
src = cv::util::get<cv::gapi::wip::IStreamSource::Ptr>(arg);
}
};
#endif // GAPI_STANDALONE
class ConstEmitter final: public cv::gimpl::GIslandEmitter {
cv::GRunArg m_arg;
virtual void halt() override {
// Not used here, but in fact can be used.
}
virtual bool pull(cv::GRunArg &arg) override {
arg = const_cast<const cv::GRunArg&>(m_arg); // FIXME: variant workaround
return true;
}
public:
explicit ConstEmitter(const cv::GRunArg &arg) : m_arg(arg) {
}
};
struct DataQueue {
static const char *name() { return "StreamingDataQueue"; }
enum tag { DESYNC }; // Enum of 1 element: purely a syntax sugar
explicit DataQueue(std::size_t capacity) {
// Note: `ptr` is shared<SyncQueue>, while the `q` is a shared<Q>
auto ptr = std::make_shared<cv::gimpl::stream::SyncQueue>();
if (capacity != 0) {
ptr->set_capacity(capacity);
}
q = std::move(ptr);
}
explicit DataQueue(tag t)
: q(new cv::gimpl::stream::DesyncQueue()) {
GAPI_Assert(t == DESYNC);
}
// FIXME: ADE metadata requires types to be copiable
std::shared_ptr<cv::gimpl::stream::Q> q;
};
struct DesyncSpecialCase {
static const char *name() { return "DesyncSpecialCase"; }
};
std::vector<cv::gimpl::stream::Q*> reader_queues( ade::Graph &g,
const ade::NodeHandle &obj)
{
ade::TypedGraph<DataQueue> qgr(g);
std::vector<cv::gimpl::stream::Q*> result;
for (auto &&out_eh : obj->outEdges())
{
result.push_back(qgr.metadata(out_eh).get<DataQueue>().q.get());
}
return result;
}
std::vector<cv::gimpl::stream::Q*> input_queues( ade::Graph &g,
const ade::NodeHandle &obj)
{
ade::TypedGraph<DataQueue> qgr(g);
std::vector<cv::gimpl::stream::Q*> result;
for (auto &&in_eh : obj->inEdges())
{
result.push_back(qgr.metadata(in_eh).contains<DataQueue>()
? qgr.metadata(in_eh).get<DataQueue>().q.get()
: nullptr);
}
return result;
}
void sync_data(cv::GRunArgs &results, cv::GRunArgsP &outputs)
{
for (auto && it : ade::util::zip(ade::util::toRange(outputs),
ade::util::toRange(results)))
{
auto &out_obj = std::get<0>(it);
auto &res_obj = std::get<1>(it);
// FIXME: this conversion should be unified
using T = cv::GRunArgP;
switch (out_obj.index())
{
case T::index_of<cv::Mat*>():
{
auto out_mat_p = cv::util::get<cv::Mat*>(out_obj);
auto view = cv::util::get<cv::RMat>(res_obj).access(cv::RMat::Access::R);
*out_mat_p = cv::gimpl::asMat(view).clone();
} break;
case T::index_of<cv::RMat*>():
*cv::util::get<cv::RMat*>(out_obj) = std::move(cv::util::get<cv::RMat>(res_obj));
break;
case T::index_of<cv::Scalar*>():
*cv::util::get<cv::Scalar*>(out_obj) = std::move(cv::util::get<cv::Scalar>(res_obj));
break;
case T::index_of<cv::detail::VectorRef>():
cv::util::get<cv::detail::VectorRef>(out_obj).mov(cv::util::get<cv::detail::VectorRef>(res_obj));
break;
case T::index_of<cv::detail::OpaqueRef>():
cv::util::get<cv::detail::OpaqueRef>(out_obj).mov(cv::util::get<cv::detail::OpaqueRef>(res_obj));
break;
case T::index_of<cv::MediaFrame*>():
*cv::util::get<cv::MediaFrame*>(out_obj) = std::move(cv::util::get<cv::MediaFrame>(res_obj));
break;
default:
GAPI_Error("This value type is not supported!"); // ...maybe because of STANDALONE mode.
break;
}
}
}
// FIXME: Is there a way to derive function from its GRunArgsP version?
template<class C> using O = cv::util::optional<C>;
void sync_data(cv::gimpl::stream::Result &r, cv::GOptRunArgsP &outputs)
{
namespace own = cv::gapi::own;
for (auto && it : ade::util::zip(ade::util::toRange(outputs),
ade::util::toRange(r.args),
ade::util::toRange(r.flags)))
{
auto &out_obj = std::get<0>(it);
auto &res_obj = std::get<1>(it);
bool available = std::get<2>(it);
using T = cv::GOptRunArgP;
#define HANDLE_CASE(Type) \
case T::index_of<O<Type>*>(): \
if (available) { \
*cv::util::get<O<Type>*>(out_obj) \
= cv::util::make_optional(std::move(cv::util::get<Type>(res_obj))); \
} else { \
cv::util::get<O<Type>*>(out_obj)->reset(); \
}
// FIXME: this conversion should be unified
switch (out_obj.index())
{
HANDLE_CASE(cv::Scalar); break;
HANDLE_CASE(cv::RMat); break;
HANDLE_CASE(cv::MediaFrame); break;
case T::index_of<O<cv::Mat>*>(): {
// Mat: special handling.
auto &mat_opt = *cv::util::get<O<cv::Mat>*>(out_obj);
if (available) {
auto q_map = cv::util::get<cv::RMat>(res_obj).access(cv::RMat::Access::R);
// FIXME: Copy! Maybe we could do some optimization for this case!
// e.g. don't handle RMat for last ilsand in the graph.
// It is not always possible though.
mat_opt = cv::util::make_optional(cv::gimpl::asMat(q_map).clone());
} else {
mat_opt.reset();
}
} break;
case T::index_of<cv::detail::OptionalVectorRef>(): {
// std::vector<>: special handling
auto &vec_opt = cv::util::get<cv::detail::OptionalVectorRef>(out_obj);
if (available) {
vec_opt.mov(cv::util::get<cv::detail::VectorRef>(res_obj));
} else {
vec_opt.reset();
}
} break;
case T::index_of<cv::detail::OptionalOpaqueRef>(): {
// std::vector<>: special handling
auto &opq_opt = cv::util::get<cv::detail::OptionalOpaqueRef>(out_obj);
if (available) {
opq_opt.mov(cv::util::get<cv::detail::OpaqueRef>(res_obj));
} else {
opq_opt.reset();
}
} break;
default:
// ...maybe because of STANDALONE mode.
GAPI_Error("This value type is not supported!");
break;
}
}
#undef HANDLE_CASE
}
// Pops an item from every input queue and combine it to the final
// result. Blocks the current thread. Returns true if the vector has
// been obtained successfully and false if a Stop message has been
// received. Handles Stop x-queue synchronization gracefully.
//
// In fact, the logic behind this method is a little bit more complex.
// The complexity comes from handling the pipeline termination
// messages. This version if GStreamerExecutable is running every
// graph island in its own thread, and threads communicate via bounded
// (limited in size) queues. Threads poll their input queues in the
// infinite loops and pass the data to their Island executables when
// the full input vector (a "stack frame") arrives.
//
// If the input stream is over or stop() is called, "Stop" messages
// are broadcasted in the graph from island to island via queues,
// starting with the emitters (sources). Since queues are bounded,
// thread may block on push() if the queue is full already and is not
// popped for some reason in the reader thread. In order to avoid
// this, once an Island gets Stop on an input, it start reading all
// other input queues until it reaches Stop messages there as well.
// Only then the thread terminates so in theory queues are left
// free'd.
//
// "Stop" messages are sent to the pipeline in these three cases:
// 1. User has called stop(): a "Stop" message is sent to every input
// queue.
// 2. Input video stream has reached its end -- its emitter sends Stop
// to its readers AND asks constant emitters (emitters attached to
// const data -- infinite data generators) to push Stop messages as
// well - in order to maintain a regular Stop procedure as defined
// above.
// 3. "Stop" message coming from a constant emitter after triggering an
// EOS notification -- see (2).
//
// There is a problem with (3). Sometimes it terminates the pipeline
// too early while some frames could still be produced with no issue,
// and our test fails with error like "got 99 frames, expected 100".
// This is how it reproduces:
//
// q1
// [const input] -----------------------> [ ISL2 ] --> [output]
// q0 q2 .->
// [stream input] ---> [ ISL1 ] -------'
//
// Video emitter is pushing frames to q0, and ISL1 is taking every
// frame from this queue and processes it. Meanwhile, q1 is a
// const-input-queue staffed with const data already, ISL2 already
// popped one, and is waiting for data from q2 (of ISL1) to arrive.
//
// When the stream is over, stream emitter pushes the last frame to
// q0, followed by a Stop sign, and _immediately_ notifies const
// emitters to broadcast Stop messages as well. In the above
// configuration, the replicated Stop message via q1 may reach ISL2
// faster than the real Stop message via q2 -- moreover, somewhere in
// q1 or q2 there may be real frames awaiting processing. ISL2 gets
// Stop via q1 and _discards_ any pending data coming from q2 -- so a
// last frame or two may be lost.
//
// A working but not very elegant solution to this problem is to tag
// Stop messages. Stop got via stop() is really a hard stop, while
// broadcasted Stop from a Const input shouldn't initiate the Island
// execution termination. Instead, its associated const data should
// remain somewhere in islands' thread local storage until a real
// "Stop" is received.
//
// Queue reader is the class which encapsulates all this logic and
// provides threads with a managed storage and an easy API to obtain
// data.
class QueueReader
{
bool m_finishing = false; // Set to true once a "soft" stop is received
std::vector<Cmd> m_cmd;
void rewindToStop(std::vector<Q*> &in_queues,
const std::size_t this_id);
public:
cv::gimpl::StreamMsg getInputVector (std::vector<Q*> &in_queues,
cv::GRunArgs &in_constants);
using V = cv::util::variant<cv::GRunArgs, Stop, cv::gimpl::Exception>;
V getResultsVector(std::vector<Q*> &in_queues,
const std::vector<int> &in_mapping,
const std::size_t out_size);
};
void rewindToStop(std::vector<Q*> &in_queues,
const std::size_t this_id)
{
size_t expected_stop_count = std::count_if(in_queues.begin(), in_queues.end(), [] (const Q* ptr) {
return ptr != nullptr;
});
if (expected_stop_count > 0) {
// NB: it requires to subtract own queues id from total waiting queue count
// because it had got stop message before rewind was called
expected_stop_count--;
}
GAPI_LOG_DEBUG(nullptr, "id: " << this_id << ", queues count: " << in_queues.size() <<
", expected stop msg count: " << expected_stop_count);
size_t got_stop_count = 0;
while(got_stop_count < expected_stop_count) {
for (auto &&qit : ade::util::indexed(in_queues)) {
auto id2 = ade::util::index(qit);
auto &q2 = ade::util::value(qit);
if (this_id == id2) continue;
GAPI_LOG_DEBUG(nullptr, "drain next id: " << id2 <<
", stop count (" << got_stop_count << "/" <<
expected_stop_count << ")");
bool got_cmd = true;
while (q2 && got_cmd) {
Cmd cmd;
got_cmd = q2->try_pop(cmd);
if (got_cmd && cv::util::holds_alternative<Stop>(cmd)) {
got_stop_count ++;
GAPI_LOG_DEBUG(nullptr, "got stop from id: " << id2);
break;
}
}
}
}
GAPI_LOG_DEBUG(nullptr, "completed");
}
// This method handles a stop sign got from some input
// island. Reiterate through all _remaining valid_ queues (some of
// them can be set to nullptr already -- see handling in
// getInputVector) and rewind data to every Stop sign per queue.
void QueueReader::rewindToStop(std::vector<Q*> &in_queues,
const std::size_t this_id)
{
::rewindToStop(in_queues, this_id);
}
cv::gimpl::StreamMsg QueueReader::getInputVector(std::vector<Q*> &in_queues,
cv::GRunArgs &in_constants)
{
// NB: Need to release resources from the previous step, to fetch new ones.
// On some systems it might be impossible to allocate new memory
// until the old one is released.
m_cmd.clear();
// NOTE: in order to maintain the GRunArg's underlying object
// lifetime, keep the whole cmd vector (of size == # of inputs)
// in memory.
m_cmd.resize(in_queues.size());
cv::GRunArgs isl_inputs(in_queues.size());
cv::optional<cv::gimpl::Exception> exception;
for (auto &&it : ade::util::indexed(in_queues))
{
auto id = ade::util::index(it);
auto &q = ade::util::value(it);
if (q == nullptr)
{
GAPI_Assert(!in_constants.empty());
// NULL queue means a graph-constant value (like a
// value-initialized scalar)
// It can also hold a constant value received with
// Stop::Kind::CNST message (see above).
isl_inputs[id] = in_constants[id];
continue;
}
q->pop(m_cmd[id]);
switch (m_cmd[id].index())
{
case Cmd::index_of<cv::GRunArg>():
isl_inputs[id] = cv::util::get<cv::GRunArg>(m_cmd[id]);
break;
case Cmd::index_of<Stop>():
{
const auto &stop = cv::util::get<Stop>(m_cmd[id]);
if (stop.kind == Stop::Kind::CNST)
{
// We've got a Stop signal from a const source,
// propagated as a result of real stream reaching its
// end. Sometimes these signals come earlier than
// real EOS Stops so are deprioritized -- just
// remember the Const value here and continue
// processing other queues. Set queue pointer to
// nullptr and update the const_val vector
// appropriately
m_finishing = true;
in_queues[id] = nullptr;
in_constants.resize(in_queues.size());
in_constants[id] = std::move(stop.cdata);
// NEXT time (on a next call to getInputVector()), the
// "q==nullptr" check above will be triggered, but now
// we need to make it manually:
isl_inputs[id] = in_constants[id];
}
else
{
GAPI_Assert(stop.kind == Stop::Kind::HARD);
rewindToStop(in_queues, id);
// After queues are read to the proper indicator,
// indicate end-of-stream
return cv::gimpl::StreamMsg{cv::gimpl::EndOfStream{}};
} // if(Cnst)
break;
}
case Cmd::index_of<cv::gimpl::Exception>():
{
exception =
cv::util::make_optional(cv::util::get<cv::gimpl::Exception>(m_cmd[id]));
break;
}
default:
GAPI_Error("Unsupported cmd type in getInputVector()");
}
} // for(in_queues)
if (exception.has_value()) {
return cv::gimpl::StreamMsg{exception.value()};
}
if (m_finishing)
{
// If the process is about to end (a soft Stop was received
// already) and an island has no other inputs than constant
// inputs, its queues may all become nullptrs. Indicate it as
// "no data".
if (ade::util::all_of(in_queues, [](Q *ptr){return ptr == nullptr;})) {
return cv::gimpl::StreamMsg{cv::gimpl::EndOfStream{}};
}
}
// A regular case - there is data to process
for (auto& arg : isl_inputs) {
if (arg.index() == cv::GRunArg::index_of<cv::Mat>()) {
arg = cv::GRunArg{ cv::make_rmat<cv::gimpl::RMatOnMat>(cv::util::get<cv::Mat>(arg))
, arg.meta
};
}
}
return cv::gimpl::StreamMsg{std::move(isl_inputs)};
}
// This is a special method to obtain a result vector
// for the entire pipeline's outputs.
//
// After introducing desync(), the pipeline output's vector
// can be produced just partially. Also, if a desynchronized
// path has multiple outputs for the pipeline, _these_ outputs
// should still come synchronized to the end user (via pull())
//
//
// This method handles all this.
// It takes a number of input queues, which may or may not be
// equal to the number of pipeline outputs (<=).
// It also takes indexes saying which queue produces which
// output in the resulting pipeline.
//
// `out_results` is always produced with the size of full output
// vector. In the desync case, the number of in_queues will
// be less than this size and some of the items won't be produced.
// In the sync case, there will be a 1-1 mapping.
//
// In the desync case, there _will be_ multiple collector threads
// calling this method, and pushing their whole-pipeline outputs
// (_may be_ partially filled) to the same final output queue.
// The receiver part at the GStreamingExecutor level won't change
// because of that.
QueueReader::V QueueReader::getResultsVector(std::vector<Q*> &in_queues,
const std::vector<int> &in_mapping,
const std::size_t out_size)
{
cv::GRunArgs out_results(out_size);
m_cmd.resize(out_size);
cv::optional<cv::gimpl::Exception> exception;
for (auto &&it : ade::util::indexed(in_queues))
{
auto ii = ade::util::index(it);
auto oi = in_mapping[ii];
auto &q = ade::util::value(it);
q->pop(m_cmd[oi]);
switch (m_cmd[oi].index()) {
case Cmd::index_of<cv::GRunArg>():
out_results[oi] = std::move(cv::util::get<cv::GRunArg>(m_cmd[oi]));
break;
case Cmd::index_of<Stop>():
// In theory, the CNST should never reach here.
// Collector thread never handles the inputs directly
// (collector's input queues are always produced by
// islands in the graph).
rewindToStop(in_queues, ii);
return QueueReader::V(Stop{});
case Cmd::index_of<cv::gimpl::Exception>():
exception =
cv::util::make_optional(cv::util::get<cv::gimpl::Exception>(m_cmd[oi]));
break;
default:
cv::util::throw_error(
std::logic_error("Unexpected cmd kind in getResultsVector"));
} // switch
} // for(in_queues)
if (exception.has_value()) {
return QueueReader::V(exception.value());
}
return QueueReader::V(out_results);
}
// This thread is a plain dump source actor. What it do is just:
// - Check input queue (the only one) for a control command
// - Depending on the state, obtains next data object and pushes it to the
// pipeline.
void emitterActorThread(std::shared_ptr<cv::gimpl::GIslandEmitter> emitter,
Q& in_queue,
std::vector<Q*> out_queues,
std::function<void()> cb_completion)
{
// Wait for the explicit Start command.
// ...or Stop command, this also happens.
Cmd cmd;
in_queue.pop(cmd);
GAPI_Assert( cv::util::holds_alternative<Start>(cmd)
|| cv::util::holds_alternative<Stop>(cmd));
if (cv::util::holds_alternative<Stop>(cmd))
{
for (auto &&oq : out_queues) {
oq->push(cmd);
}
return;
}
GAPI_ITT_STATIC_LOCAL_HANDLE(emitter_hndl, "emitter");
GAPI_ITT_STATIC_LOCAL_HANDLE(emitter_pull_hndl, "emitter_pull");
GAPI_ITT_STATIC_LOCAL_HANDLE(emitter_push_hndl, "emitter_push");
// Now start emitting the data from the source to the pipeline.
while (true)
{
GAPI_ITT_AUTO_TRACE_GUARD(emitter_hndl);
Cmd cancel;
if (in_queue.try_pop(cancel))
{
// if we just popped a cancellation command...
GAPI_Assert(cv::util::holds_alternative<Stop>(cancel));
// Broadcast it to the readers and quit.
for (auto &&oq : out_queues) oq->push(cancel);
return;
}
// Try to obtain next data chunk from the source
cv::GRunArg data;
bool result = false;
try {
result = [&](){
GAPI_ITT_AUTO_TRACE_GUARD(emitter_pull_hndl);
return emitter->pull(data);
}();
} catch (...) {
auto eptr = std::current_exception();
for (auto &&oq : out_queues)
{
oq->push(Cmd{cv::gimpl::Exception{eptr}});
}
// NB: Go to the next iteration.
continue;
}
if (result)
{
GAPI_ITT_AUTO_TRACE_GUARD(emitter_push_hndl);
// // On success, broadcast it to our readers
for (auto &&oq : out_queues)
{
// FIXME: FOR SOME REASON, oq->push(Cmd{data}) doesn't work!!
// empty mats are arrived to the receivers!
// There may be a fatal bug in our variant!
const auto tmp = data;
oq->push(Cmd{tmp});
}
}
else
{
// Otherwise, broadcast STOP message to our readers and quit.
// This usually means end-of-stream, so trigger a callback
for (auto &&oq : out_queues) oq->push(Cmd{Stop{}});
if (cb_completion) cb_completion();
return;
}
}
}
// This thread pulls data from the assigned input queues and makes sure that
// all input args are in sync (timestamps are equal), dropping some inputs if required.
// After getting synchronized inputs from all input queues, the thread pushes them to out queues
void syncActorThread(std::vector<Q*> in_queues,
std::vector<std::vector<Q*>> out_queues) {
using timestamp_t = int64_t;
std::vector<bool> pop_nexts(in_queues.size());
std::vector<Cmd> cmds(in_queues.size());
GAPI_ITT_STATIC_LOCAL_HANDLE(sync_hndl, "sync_actor");
GAPI_ITT_STATIC_LOCAL_HANDLE(sync_pull_1_queue_hndl, "sync_actor_pull_from_1_queue");
GAPI_ITT_STATIC_LOCAL_HANDLE(sync_push_hndl, "sync_actor_push");
while (true) {
GAPI_ITT_AUTO_TRACE_GUARD(sync_hndl);
// pop_nexts indicates which queue still contains earlier timestamps and
// needs to be popped at least one more time.
// For each iteration (frame) we need to pull from each input queue at least once,
// so switch all to true when start processing new frame
for (auto&& p : pop_nexts) {
p = true;
}
timestamp_t max_ts = 0u;
// Iterate through all input queues, pop GRunArg's and compare timestamps.
// Continue pulling from queues whose timestamps are smaller.
// Finish when all timestamps are equal.
do {
for (auto&& it : ade::util::indexed(
ade::util::zip(pop_nexts, in_queues, cmds))) {
auto& val = ade::util::value(it);
auto& pop_next = std::get<0>(val);
if (!pop_next) {
continue;
}
auto& q = std::get<1>(val);
auto& cmd = std::get<2>(val);
{
GAPI_ITT_AUTO_TRACE_GUARD(sync_pull_1_queue_hndl);
q->pop(cmd);
}
if (cv::util::holds_alternative<Stop>(cmd)) {
// We got a stop command from one of the input queues.
// Rewind all input queues till Stop command,
// Push Stop command down the graph, finish the thread
rewindToStop(in_queues, ade::util::index(it));
for (auto &&oqs : out_queues) {
for (auto &&oq : oqs) {
oq->push(Cmd{Stop{}});
}
}
return;
}
// Extract the timestamp
auto& arg = cv::util::get<cv::GRunArg>(cmd);
auto ts = cv::util::any_cast<int64_t>(arg.meta[cv::gapi::streaming::meta_tag::timestamp]);
GAPI_Assert(ts >= 0u);
// TODO: this whole drop logic can be imported via compile args
// to give a user a way to customize it
if (ts < max_ts) {
// Continue popping from this queue
pop_next = true;
} else if (ts == max_ts) {
// Stop popping from this queue
pop_next = false;
} else if (ts > max_ts) {
// We got a timestamp which is greater than timestamps from other queues.
// It means that we need to reiterate through all the queues one more time
// (except the current one)
max_ts = ts;
for (auto&& p : pop_nexts) {
p = true;
}
pop_next = false;
}
}
} while (ade::util::any_of(pop_nexts, [](bool v){ return v; }));
// Finally we got all our inputs synchronized, push them further down the graph
{
GAPI_ITT_AUTO_TRACE_GUARD(sync_push_hndl);
for (auto &&it : ade::util::zip(out_queues, cmds)) {
for (auto &&q : std::get<0>(it)) {
q->push(std::get<1>(it));
}
}
}
}
}
class StreamingInput final: public cv::gimpl::GIslandExecutable::IInput
{
QueueReader &qr;
std::vector<Q*> &in_queues; // FIXME: This can be part of QueueReader
cv::GRunArgs &in_constants; // FIXME: This can be part of QueueReader
cv::optional<cv::gimpl::StreamMsg> last_read_msg;
virtual cv::gimpl::StreamMsg try_get() override
{
// FIXME: This is not very usable at the moment!
return get();
}
public:
explicit StreamingInput(QueueReader &rdr,
std::vector<Q*> &inq,
cv::GRunArgs &inc,
const std::vector<cv::gimpl::RcDesc> &in_descs)
: qr(rdr), in_queues(inq), in_constants(inc)
{
set(in_descs);
}
const cv::gimpl::StreamMsg& read()
{
GAPI_ITT_STATIC_LOCAL_HANDLE(inputs_get_hndl, "StreamingInput::read");
GAPI_ITT_AUTO_TRACE_GUARD(inputs_get_hndl);
last_read_msg =
cv::optional<cv::gimpl::StreamMsg>(
qr.getInputVector(in_queues, in_constants));
return last_read_msg.value();
}
virtual cv::gimpl::StreamMsg get() override
{
GAPI_ITT_STATIC_LOCAL_HANDLE(inputs_get_hndl, "StreamingInput::get");
GAPI_ITT_AUTO_TRACE_GUARD(inputs_get_hndl);
if (!last_read_msg.has_value()) {
(void)read();
}
auto msg = std::move(last_read_msg.value());
last_read_msg = cv::optional<cv::gimpl::StreamMsg>();
return msg;
}
};
class StreamingOutput final: public cv::gimpl::GIslandExecutable::IOutput
{
// These objects form an internal state of the StreamingOutput
struct Posting
{
using V = cv::util::variant<cv::GRunArg,
cv::gimpl::EndOfStream,
cv::gimpl::Exception>;
V data;
bool ready = false;
};
using PostingList = std::list<Posting>;
std::vector<PostingList> m_postings;
std::unordered_map< const void*
, std::pair<int, PostingList::iterator>
> m_postIdx;
std::size_t m_stops_sent = 0u;
// These objects are owned externally
const cv::GMetaArgs &m_metas;
std::vector< std::vector<Q*> > &m_out_queues;
std::shared_ptr<cv::gimpl::GIslandExecutable> m_island;
// NB: StreamingOutput have to be thread-safe.
// Now synchronization approach is quite poor and inefficient.
mutable std::mutex m_mutex;
// Allocate a new data object for output under idx
// Prepare this object for posting
virtual cv::GRunArgP get(int idx) override
{
GAPI_ITT_STATIC_LOCAL_HANDLE(outputs_get_hndl, "StreamingOutput::get (alloc)");
GAPI_ITT_AUTO_TRACE_GUARD(outputs_get_hndl);
std::lock_guard<std::mutex> lock{m_mutex};
using MatType = cv::Mat;
using SclType = cv::Scalar;
// Allocate a new posting first, then bind this GRunArgP to this item
auto iter = m_postings[idx].insert(m_postings[idx].end(), Posting{});
const auto r = desc()[idx];
cv::GRunArg& out_arg = cv::util::get<cv::GRunArg>(iter->data);
cv::GRunArgP ret_val;
switch (r.shape) {
// Allocate a data object based on its shape & meta, and put it into our vectors.
// Yes, first we put a cv::Mat GRunArg, and then specify _THAT_
// pointer as an output parameter - to make sure that after island completes,
// our GRunArg still has the right (up-to-date) value.
// Same applies to other types.
// FIXME: This is absolutely ugly but seem to work perfectly for its purpose.
case cv::GShape::GMAT:
{
auto desc = cv::util::get<cv::GMatDesc>(m_metas[idx]);
if (m_island->allocatesOutputs())
{
out_arg = cv::GRunArg(m_island->allocate(desc));
}
else
{
MatType newMat;
cv::gimpl::createMat(desc, newMat);
auto rmat = cv::make_rmat<cv::gimpl::RMatOnMat>(newMat);
out_arg = cv::GRunArg(std::move(rmat));
}
ret_val = cv::GRunArgP(&cv::util::get<cv::RMat>(out_arg));
}
break;
case cv::GShape::GSCALAR:
{
SclType newScl;
out_arg = cv::GRunArg(std::move(newScl));
ret_val = cv::GRunArgP(&cv::util::get<SclType>(out_arg));
}
break;
case cv::GShape::GARRAY:
{
cv::detail::VectorRef newVec;
cv::util::get<cv::detail::ConstructVec>(r.ctor)(newVec);
out_arg = cv::GRunArg(std::move(newVec));
// VectorRef is implicitly shared so no pointer is taken here
// FIXME: that variant MOVE problem again
const auto &rr = cv::util::get<cv::detail::VectorRef>(out_arg);
ret_val = cv::GRunArgP(rr);
}
break;
case cv::GShape::GOPAQUE:
{
cv::detail::OpaqueRef newOpaque;
cv::util::get<cv::detail::ConstructOpaque>(r.ctor)(newOpaque);
out_arg = cv::GRunArg(std::move(newOpaque));
// OpaqueRef is implicitly shared so no pointer is taken here
// FIXME: that variant MOVE problem again
const auto &rr = cv::util::get<cv::detail::OpaqueRef>(out_arg);
ret_val = cv::GRunArgP(rr);
}
break;
case cv::GShape::GFRAME:
{
cv::MediaFrame frame;
out_arg = cv::GRunArg(std::move(frame));
ret_val = cv::GRunArgP(&cv::util::get<cv::MediaFrame>(out_arg));
}
break;
default:
cv::util::throw_error(std::logic_error("Unsupported GShape"));
}
m_postIdx[cv::gimpl::proto::ptr(ret_val)] = std::make_pair(idx, iter);
return ret_val;
}
virtual void post(cv::GRunArgP&& argp, const std::exception_ptr& exptr) override
{
GAPI_ITT_STATIC_LOCAL_HANDLE(outputs_post_hndl, "StreamingOutput::post");
GAPI_ITT_AUTO_TRACE_GUARD(outputs_post_hndl);
std::lock_guard<std::mutex> lock{m_mutex};
// Mark the output ready for posting. If it is the first in the line,
// actually post it and all its successors which are ready for posting too.
auto it = m_postIdx.find(cv::gimpl::proto::ptr(argp));
GAPI_Assert(it != m_postIdx.end());
const int out_idx = it->second.first;
const auto out_iter = it->second.second;
out_iter->ready = true;
if (exptr) {
out_iter->data = cv::gimpl::Exception{exptr};
}
m_postIdx.erase(it); // Drop the link from the cache anyway
if (out_iter != m_postings[out_idx].begin())
{
return; // There are some pending postings in the beginning, return
}
GAPI_Assert(out_iter == m_postings[out_idx].begin());
auto post_iter = m_postings[out_idx].begin();
while (post_iter != m_postings[out_idx].end() && post_iter->ready == true)
{
Cmd cmd;
switch (post_iter->data.index())
{
case Posting::V::index_of<cv::GRunArg>():
cmd = Cmd{cv::util::get<cv::GRunArg>(post_iter->data)};
break;
case Posting::V::index_of<cv::gimpl::Exception>():
cmd = Cmd{cv::util::get<cv::gimpl::Exception>(post_iter->data)};
break;
case Posting::V::index_of<cv::gimpl::EndOfStream>():
cmd = Cmd{Stop{}};
m_stops_sent++;
break;
default:
GAPI_Error("Unreachable code");
}
for (auto &&q : m_out_queues[out_idx])
{
q->push(cmd);
}
post_iter = m_postings[out_idx].erase(post_iter);
}
}
virtual void post(cv::gimpl::EndOfStream&&) override
{
std::lock_guard<std::mutex> lock{m_mutex};
// If the posting list is empty, just broadcast the stop message.
// If it is not, enqueue the Stop message in the postings list.
for (auto &&it : ade::util::indexed(m_postings))
{
const auto idx = ade::util::index(it);
auto &lst = ade::util::value(it);
if (lst.empty())
{
for (auto &&q : m_out_queues[idx])
{
q->push(Cmd(Stop{}));
}
m_stops_sent++;
}
else
{
Posting p;
p.data = Posting::V{cv::gimpl::EndOfStream{}};
p.ready = true;
lst.push_back(std::move(p)); // FIXME: For some reason {}-ctor didn't work here
}
}
}
void meta(const cv::GRunArgP &out, const cv::GRunArg::Meta &m) override
{
std::lock_guard<std::mutex> lock{m_mutex};
const auto it = m_postIdx.find(cv::gimpl::proto::ptr(out));
GAPI_Assert(it != m_postIdx.end());
const auto out_iter = it->second.second;
cv::util::get<cv::GRunArg>(out_iter->data).meta = m;
}
public:
explicit StreamingOutput(const cv::GMetaArgs &metas,
std::vector< std::vector<Q*> > &out_queues,
const std::vector<cv::gimpl::RcDesc> &out_descs,
std::shared_ptr<cv::gimpl::GIslandExecutable> island)
: m_metas(metas)
, m_out_queues(out_queues)
, m_island(island)
{
set(out_descs);
m_postings.resize(out_descs.size());
}
bool done() const
{
std::lock_guard<std::mutex> lock{m_mutex};
// The streaming actor work is considered DONE for this stream
// when it posted/resent all STOP messages to all its outputs.
return m_stops_sent == desc().size();
}
virtual void post(cv::gimpl::Exception&& error) override