forked from microsoft/onnxruntime
-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy pathinference_session.cc
3129 lines (2684 loc) · 139 KB
/
inference_session.cc
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
// Copyright (c) Microsoft Corporation. All rights reserved.
// Licensed under the MIT License.
#include "core/graph/onnx_protobuf.h"
#include "core/session/inference_session.h"
#include <memory>
#include <sstream>
#include <list>
#include <string>
#include <thread>
#include <queue>
#include "core/common/denormal.h"
#include "core/common/logging/logging.h"
#include "core/common/parse_string.h"
#include "core/common/path_string.h"
#include "core/flatbuffers/flatbuffers_utils.h"
#include "core/flatbuffers/ort_format_version.h"
#include "core/framework/bfc_arena.h"
#include "core/framework/error_code_helper.h"
#include "core/framework/execution_frame.h"
#include "core/framework/feeds_fetches_manager.h"
#include "core/framework/graph_partitioner.h"
#include "core/framework/kernel_def_builder.h"
#include "core/framework/kernel_registry.h"
#include "core/framework/kernel_type_str_resolver.h"
#include "core/framework/kernel_type_str_resolver_utils.h"
#include "core/framework/mldata_type_utils.h"
#include "core/framework/TensorSeq.h"
#include "core/framework/tensorprotoutils.h"
#include "core/framework/tensor_type_and_shape.h"
#include "core/framework/op_kernel_context_internal.h"
#include "core/framework/ort_value_pattern_planner.h"
#include "core/framework/transform_layout_functions.h"
#include "core/framework/utils.h"
#include "core/graph/graph_viewer.h"
#include "core/graph/model.h"
#include "core/optimizer/graph_transformer_utils.h"
#include "core/optimizer/graph_transformer.h"
#include "core/optimizer/layout_transformation/layout_transformation.h"
#include "core/optimizer/insert_cast_transformer.h"
#include "core/optimizer/qdq_transformer/ensure_unique_dq_for_node_unit.h"
#include "core/optimizer/rule_based_graph_transformer.h"
#include "core/optimizer/selectors_actions/selector_action_transformer_apply_contexts.h"
#include "core/optimizer/transformer_memcpy.h"
#include "core/optimizer/transpose_optimization/ort_optimizer_utils.h"
#include "core/platform/Barrier.h"
#include "core/platform/threadpool.h"
#ifdef _WIN32
#include "core/platform/tracing.h"
#include <Windows.h>
#include "core/platform/windows/telemetry.h"
#endif
#include "core/providers/cpu/controlflow/utils.h"
#include "core/providers/cpu/cpu_execution_provider.h"
#ifdef USE_DML // TODO: This is necessary for the workaround in TransformGraph
#include "core/providers/dml/DmlExecutionProvider/src/DmlGraphFusionTransformer.h"
#include "core/providers/dml/DmlExecutionProvider/src/DmlRuntimeGraphFusionTransformer.h"
#include "core/providers/dml/DmlExecutionProvider/src/GraphTransformer.h"
#include "core/providers/dml/dml_session_options_config_keys.h"
#include "core/providers/dml/DmlExecutionProvider/src/ExecutionProvider.h"
#endif
#include "core/session/environment.h"
#include "core/session/user_logging_sink.h"
#include "core/session/IOBinding.h"
#include "core/session/inference_session_utils.h"
#include "core/session/onnxruntime_session_options_config_keys.h"
#include "core/session/onnxruntime_run_options_config_keys.h"
#include "core/util/protobuf_parsing_utils.h"
#include "core/util/thread_utils.h"
#ifdef _WIN32
#include "core/platform/windows/logging/etw_sink.h"
#include "core/common/logging/sinks/composite_sink.h"
#endif
// custom ops are not available in a minimal build unless ORT_MINIMAL_BUILD_CUSTOM_OPS is set
#if !defined(ORT_MINIMAL_BUILD) || defined(ORT_MINIMAL_BUILD_CUSTOM_OPS)
#include "core/framework/customregistry.h"
#include "core/session/custom_ops.h"
#endif
#ifdef ENABLE_TRAINING
#include "core/framework/partial_graph_execution_state.h"
#include "core/framework/stream_execution_context.h"
#include "orttraining/core/optimizer/memory_optimizer/memory_optimizer.h"
#endif
using namespace ONNX_NAMESPACE;
using namespace onnxruntime::common;
namespace onnxruntime {
namespace {
template <typename T>
const T* GetDateFormatString();
template <>
inline const char* GetDateFormatString<char>() {
return "%Y-%m-%d_%H-%M-%S";
}
#ifdef _WIN32
template <>
inline const wchar_t* GetDateFormatString<wchar_t>() {
return L"%Y-%m-%d_%H-%M-%S";
}
#endif
// TODO: use LoggingManager::GetTimestamp and date::operator<<
// (see ostream_sink.cc for an example)
// to simplify this and match the log file timestamp format.
template <typename T>
inline std::basic_string<T> GetCurrentTimeString() {
auto now = std::chrono::system_clock::now();
auto in_time_t = std::chrono::system_clock::to_time_t(now);
std::tm local_tm; // NOLINT
#ifdef _WIN32
ORT_ENFORCE(localtime_s(&local_tm, &in_time_t) == 0);
#else
localtime_r(&in_time_t, &local_tm);
#endif
T time_str[32];
OrtStrftime<T>(time_str, sizeof(time_str), GetDateFormatString<T>(), &local_tm);
return std::basic_string<T>(time_str);
}
#if !defined(ORT_MINIMAL_BUILD)
static bool HasControlflowNodes(const Graph& graph) {
for (const auto& node : graph.Nodes()) {
if (node.ContainsSubgraph()) {
return true;
}
}
return false;
}
static bool HasMemcpyNodes(const Graph& graph) {
for (const auto& node : graph.Nodes()) {
if (node.OpType() == "MemcpyFromHost" || node.OpType() == "MemcpyToHost") {
return true;
}
}
return false;
}
static bool AreAllComputeNodesAssignedToCudaOrJsEp(const Graph& graph) {
bool nodes_on_cpu_and_cuda_and_js_eps_only = true;
for (const auto& node : graph.Nodes()) {
const auto& node_provider = node.GetExecutionProviderType();
// Empty node provider means CPU EP
if (!node_provider.empty() &&
!(node_provider == kCudaExecutionProvider ||
node_provider == kRocmExecutionProvider ||
node_provider == kJsExecutionProvider) &&
node_provider != kCpuExecutionProvider) {
nodes_on_cpu_and_cuda_and_js_eps_only = false;
break;
}
}
// If we see nodes assigned to EPs other than CPU, or CUDA/JS
// (or) if there are Memcpy nodes, then all compute nodes have
// not been parititoned to the CUDA/JS EP.
// We allow CPU EPs to show up in the EP list as long as thre is no Memcpy
// involved as shape subgraphs will be forced onto CPU and these will not have
// Memcpy nodes involved.
return nodes_on_cpu_and_cuda_and_js_eps_only && !HasMemcpyNodes(graph);
}
static bool AreAllNodesInMainGraphAssignedToOneEp(const Graph& graph, ProviderType provider) {
for (const auto& node : graph.Nodes()) {
const auto& node_provider = node.GetExecutionProviderType();
if (node_provider.empty() || node_provider != provider) {
return false;
}
}
return true;
}
static bool HasShapeSubgraphNodes(const Graph& graph) {
bool has_shape_nodes = false;
bool has_cpu_ep_nodes = false;
for (const auto& node : graph.Nodes()) {
if (node.OpType() == "Shape") {
has_shape_nodes = true;
break;
}
}
for (const auto& node : graph.Nodes()) {
const auto& node_provider = node.GetExecutionProviderType();
if (node_provider.empty() || node_provider == kCpuExecutionProvider) {
has_cpu_ep_nodes = true;
break;
}
}
return has_shape_nodes && has_cpu_ep_nodes;
}
Status GetMinimalBuildOptimizationHandling(
std::string_view config_value, bool saving_ort_format,
InferenceSession::MinimalBuildOptimizationHandling& minimal_build_optimization_handling) {
if (config_value == "save") {
if (saving_ort_format) {
minimal_build_optimization_handling =
InferenceSession::MinimalBuildOptimizationHandling::SaveMinimalBuildRuntimeOptimizations;
return Status::OK();
}
return ORT_MAKE_STATUS(ONNXRUNTIME, INVALID_ARGUMENT,
kOrtSessionOptionsConfigMinimalBuildOptimizations,
" value of 'save' is only valid when saving an ORT format model.");
}
if (config_value == "apply") {
minimal_build_optimization_handling =
InferenceSession::MinimalBuildOptimizationHandling::OnlyApplyMinimalBuildOptimizations;
return Status::OK();
}
if (config_value.empty()) {
minimal_build_optimization_handling =
InferenceSession::MinimalBuildOptimizationHandling::ApplyFullBuildOptimizations;
return Status::OK();
}
return ORT_MAKE_STATUS(ONNXRUNTIME, INVALID_ARGUMENT,
"Invalid value for ", kOrtSessionOptionsConfigMinimalBuildOptimizations, ": ", config_value);
};
#endif // !defined(ORT_MINIMAL_BUILD)
} // namespace
std::atomic<uint32_t> InferenceSession::global_session_id_{1};
std::map<uint32_t, InferenceSession*> InferenceSession::active_sessions_;
#ifdef _WIN32
OrtMutex InferenceSession::active_sessions_mutex_; // Protects access to active_sessions_
#endif
static Status FinalizeSessionOptions(const SessionOptions& user_provided_session_options,
const ONNX_NAMESPACE::ModelProto& model_proto,
bool is_model_proto_parsed,
/*out*/ SessionOptions& finalized_session_options) {
#if !defined(ORT_MINIMAL_BUILD)
const logging::Logger& default_logger = logging::LoggingManager::DefaultLogger();
// By now the environment should have initialized. (It is enforced prior to this.)
const Env& env_instance = Env::Default();
bool session_options_from_model = false;
// Get the value held by the environment variable - kOrtLoadConfigFromModelEnvVar
const std::string load_config_from_model_env_var_value =
env_instance.GetEnvironmentVar(inference_session_utils::kOrtLoadConfigFromModelEnvVar);
// Ascertain if the model is to be read for the ORT config from the afore parsed env var
if (!load_config_from_model_env_var_value.empty()) {
// Check if the env var contains an unsupported value
if (load_config_from_model_env_var_value.length() > 1 ||
(load_config_from_model_env_var_value[0] != '0' && load_config_from_model_env_var_value[0] != '1')) {
std::ostringstream oss;
oss << "The only supported values for the environment variable "
<< inference_session_utils::kOrtLoadConfigFromModelEnvVar << " are '0' and '1'. "
<< "The environment variable contained the value: " << load_config_from_model_env_var_value;
return ORT_MAKE_STATUS(ONNXRUNTIME, INVALID_ARGUMENT, oss.str());
}
if (load_config_from_model_env_var_value[0] == '1') {
LOGS(default_logger, INFO) << "Reading the provided model for the ORT config";
session_options_from_model = true;
}
}
// The model is to be read for an ORT config json that may hold some/all session options
if (session_options_from_model) {
SessionOptions constructed_session_options;
// In theory we should not hit this condition unless this internal class' APIs are being called incorrectly.
// This is a good sanity check to enforce that the model has been parsed prior to looking into it for ort config.
ORT_ENFORCE(is_model_proto_parsed, "ModelProto needs to be parsed to check for ORT config within it");
// Use default logger as the session_logger_ hasn't been initialized yet.
inference_session_utils::JsonConfigParser config_parser(default_logger);
auto status = config_parser.ParseOrtConfigJsonInModelProto(model_proto);
if (!status.IsOK()) {
return status;
}
status = config_parser.ParseSessionOptionsFromModelProto(constructed_session_options);
if (!status.IsOK()) {
return status;
}
// use the constructed session options
finalized_session_options = constructed_session_options;
} else {
// use user provided session options instance
finalized_session_options = user_provided_session_options;
}
#else
ORT_UNUSED_PARAMETER(model_proto);
ORT_UNUSED_PARAMETER(is_model_proto_parsed);
finalized_session_options = user_provided_session_options;
#endif // !defined(ORT_MINIMAL_BUILD)
return Status::OK();
}
logging::Severity GetSeverity(const SessionOptions& session_options) {
logging::Severity severity = logging::Severity::kWARNING;
if (session_options.session_log_severity_level == -1) {
severity = logging::LoggingManager::DefaultLogger().GetSeverity();
} else {
ORT_ENFORCE(session_options.session_log_severity_level >= 0 &&
session_options.session_log_severity_level <= static_cast<int>(logging::Severity::kFATAL),
"Invalid session log severity level. Not a valid onnxruntime::logging::Severity value: ",
session_options.session_log_severity_level);
severity = static_cast<logging::Severity>(session_options.session_log_severity_level);
}
return severity;
}
void InferenceSession::SetLoggingManager(const SessionOptions& session_options,
const Environment& session_env) {
logging_manager_ = session_env.GetLoggingManager();
std::unique_ptr<logging::ISink> sink;
if (session_options.user_logging_function) {
sink = std::make_unique<UserLoggingSink>(session_options.user_logging_function,
session_options.user_logging_param);
auto sessionSeverity = GetSeverity(session_options);
auto etwOverrideSeverity = logging::OverrideLevelWithEtw(sessionSeverity);
sink = EnhanceLoggerWithEtw(std::move(sink), sessionSeverity, etwOverrideSeverity);
user_logging_manager_ = std::make_unique<logging::LoggingManager>(std::move(sink),
std::min(sessionSeverity, etwOverrideSeverity),
false,
logging::LoggingManager::InstanceType::Temporal,
&session_options.session_logid);
logging_manager_ = user_logging_manager_.get();
}
}
void InferenceSession::ConstructorCommon(const SessionOptions& session_options,
const Environment& session_env) {
auto status = FinalizeSessionOptions(session_options, model_proto_, is_model_proto_parsed_, session_options_);
ORT_ENFORCE(status.IsOK(), "Could not finalize session options while constructing the inference session. Error Message: ",
status.ErrorMessage());
// a monotonically increasing session id for use in telemetry
session_id_ = global_session_id_.fetch_add(1);
#ifdef _WIN32
std::lock_guard<OrtMutex> lock(active_sessions_mutex_);
active_sessions_[global_session_id_++] = this;
// Register callback for ETW capture state (rundown)
WindowsTelemetry::RegisterInternalCallback(
[this](
LPCGUID SourceId,
ULONG IsEnabled,
UCHAR Level,
ULONGLONG MatchAnyKeyword,
ULONGLONG MatchAllKeyword,
PEVENT_FILTER_DESCRIPTOR FilterData,
PVOID CallbackContext) {
(void)SourceId;
(void)Level;
(void)MatchAnyKeyword;
(void)MatchAllKeyword;
(void)FilterData;
(void)CallbackContext;
// Check if this callback is for capturing state
if ((IsEnabled == EVENT_CONTROL_CODE_CAPTURE_STATE) &&
((MatchAnyKeyword & static_cast<ULONGLONG>(onnxruntime::logging::ORTTraceLoggingKeyword::Session)) != 0)) {
LogAllSessions();
}
});
#endif
SetLoggingManager(session_options, session_env);
// The call to InitLogger depends on the final state of session_options_. Hence it should be invoked
// after the invocation of FinalizeSessionOptions.
InitLogger(logging_manager_); // this sets session_logger_ so that it can be used for logging after this point.
TraceSessionOptions(session_options, false);
#if !defined(ORT_MINIMAL_BUILD)
// Update the number of steps for the graph transformer manager using the "finalized" session options
ORT_ENFORCE(graph_transformer_mgr_.SetSteps(session_options_.max_num_graph_transformation_steps).IsOK());
#endif
bool set_denormal_as_zero =
session_options_.config_options.GetConfigOrDefault(kOrtSessionOptionsConfigSetDenormalAsZero, "0") == "1";
// The only first session option for flush-to-zero and denormal-as-zero is effective to main thread and OpenMP threads.
{
static std::once_flag once;
std::call_once(once, [&] {
SetDenormalAsZero(set_denormal_as_zero);
LOGS(*session_logger_, INFO) << "Flush-to-zero and denormal-as-zero are " << ((set_denormal_as_zero) ? "on" : "off");
});
}
use_per_session_threads_ = session_options.use_per_session_threads;
force_spinning_stop_between_runs_ = session_options_.config_options.GetConfigOrDefault(kOrtSessionOptionsConfigForceSpinningStop, "0") == "1";
if (use_per_session_threads_) {
LOGS(*session_logger_, INFO) << "Creating and using per session threadpools since use_per_session_threads_ is true";
{
if (!external_intra_op_thread_pool_) {
bool allow_intra_op_spinning =
session_options_.config_options.GetConfigOrDefault(kOrtSessionOptionsConfigAllowIntraOpSpinning, "1") == "1";
OrtThreadPoolParams to = session_options_.intra_op_param;
std::basic_stringstream<ORTCHAR_T> ss;
if (to.name) {
ss << to.name << ORT_TSTR("-");
}
ss << ORT_TSTR("session-") << session_id_ << ORT_TSTR("-intra-op");
thread_pool_name_ = ss.str();
to.name = thread_pool_name_.c_str();
to.set_denormal_as_zero = set_denormal_as_zero;
// If the thread pool can use all the processors, then
// we set affinity of each thread to each processor.
to.allow_spinning = allow_intra_op_spinning;
to.dynamic_block_base_ = std::stoi(session_options_.config_options.GetConfigOrDefault(kOrtSessionOptionsConfigDynamicBlockBase, "0"));
LOGS(*session_logger_, INFO) << "Dynamic block base set to " << to.dynamic_block_base_;
// Set custom threading functions
to.custom_create_thread_fn = session_options_.custom_create_thread_fn;
to.custom_thread_creation_options = session_options.custom_thread_creation_options;
to.custom_join_thread_fn = session_options_.custom_join_thread_fn;
if (session_options_.config_options.TryGetConfigEntry(kOrtSessionOptionsConfigIntraOpThreadAffinities, to.affinity_str)) {
ORT_ENFORCE(!to.affinity_str.empty(), "Affinity string must not be empty");
}
to.auto_set_affinity = to.thread_pool_size == 0 &&
session_options_.execution_mode == ExecutionMode::ORT_SEQUENTIAL &&
to.affinity_str.empty();
if (to.custom_create_thread_fn) {
ORT_ENFORCE(to.custom_join_thread_fn, "custom join thread function not set for intra op thread pool");
}
thread_pool_ =
concurrency::CreateThreadPool(&Env::Default(), to, concurrency::ThreadPoolType::INTRA_OP);
}
}
if (session_options_.execution_mode == ExecutionMode::ORT_PARALLEL) {
if (!external_inter_op_thread_pool_) {
bool allow_inter_op_spinning =
session_options_.config_options.GetConfigOrDefault(kOrtSessionOptionsConfigAllowInterOpSpinning, "1") == "1";
OrtThreadPoolParams to = session_options_.inter_op_param;
to.auto_set_affinity = to.thread_pool_size == 0 && session_options_.execution_mode == ExecutionMode::ORT_SEQUENTIAL;
std::basic_stringstream<ORTCHAR_T> ss;
if (to.name) {
ss << to.name << ORT_TSTR("-");
}
ss << ORT_TSTR("session-") << session_id_ << ORT_TSTR("-inter-op");
inter_thread_pool_name_ = ss.str();
to.name = inter_thread_pool_name_.c_str();
to.set_denormal_as_zero = set_denormal_as_zero;
to.allow_spinning = allow_inter_op_spinning;
to.dynamic_block_base_ = std::stoi(session_options_.config_options.GetConfigOrDefault(kOrtSessionOptionsConfigDynamicBlockBase, "0"));
// Set custom threading functions
to.custom_create_thread_fn = session_options_.custom_create_thread_fn;
to.custom_thread_creation_options = session_options.custom_thread_creation_options;
to.custom_join_thread_fn = session_options_.custom_join_thread_fn;
if (to.custom_create_thread_fn) {
ORT_ENFORCE(to.custom_join_thread_fn, "custom join thread function not set for inter op thread pool");
}
inter_op_thread_pool_ =
concurrency::CreateThreadPool(&Env::Default(), to, concurrency::ThreadPoolType::INTER_OP);
if (inter_op_thread_pool_ == nullptr) {
LOGS(*session_logger_, INFO) << "Failed to create the inter-op thread pool for the parallel executor, setting ExecutionMode to SEQUENTIAL";
session_options_.execution_mode = ExecutionMode::ORT_SEQUENTIAL;
}
}
}
} else {
LOGS(*session_logger_, INFO) << "Using global/env threadpools since use_per_session_threads_ is false";
intra_op_thread_pool_from_env_ = session_env.GetIntraOpThreadPool();
inter_op_thread_pool_from_env_ = session_env.GetInterOpThreadPool();
ORT_ENFORCE(session_env.EnvCreatedWithGlobalThreadPools(),
"When the session is not configured to use per session"
" threadpools, the env must be created with the the CreateEnvWithGlobalThreadPools API.");
}
session_profiler_.Initialize(session_logger_);
if (session_options_.enable_profiling) {
StartProfiling(session_options_.profile_file_prefix);
}
telemetry_ = {};
}
void InferenceSession::TraceSessionOptions(const SessionOptions& session_options, bool captureState) {
(void)captureState; // Otherwise Linux build error
LOGS(*session_logger_, INFO) << session_options;
#ifdef _WIN32
TraceLoggingWrite(telemetry_provider_handle,
"SessionOptions",
TraceLoggingKeyword(static_cast<uint64_t>(onnxruntime::logging::ORTTraceLoggingKeyword::Session)),
TraceLoggingLevel(WINEVENT_LEVEL_INFO),
TraceLoggingUInt8(static_cast<UINT8>(session_options.execution_mode), "execution_mode"),
TraceLoggingUInt8(static_cast<UINT8>(session_options.execution_order), "execution_order"),
TraceLoggingBoolean(session_options.enable_profiling, "enable_profiling"),
TraceLoggingString(ORT_TSTR_CONVERT_TO_PRINTABLE_STRING(session_options.optimized_model_filepath).c_str(), "optimized_model_filepath"),
TraceLoggingBoolean(session_options.enable_mem_pattern, "enable_mem_pattern"),
TraceLoggingBoolean(session_options.enable_mem_reuse, "enable_mem_reuse"),
TraceLoggingBoolean(session_options.enable_cpu_mem_arena, "enable_cpu_mem_arena"),
TraceLoggingString(ORT_TSTR_CONVERT_TO_PRINTABLE_STRING(session_options.profile_file_prefix).c_str(), "profile_file_prefix"),
TraceLoggingString(session_options.session_logid.c_str(), "session_logid"),
TraceLoggingInt8(static_cast<INT8>(session_options.session_log_severity_level), "session_log_severity_level"),
TraceLoggingInt8(static_cast<INT8>(session_options.session_log_verbosity_level), "session_log_verbosity_level"),
TraceLoggingUInt32(session_options.max_num_graph_transformation_steps, "max_num_graph_transformation_steps"),
TraceLoggingUInt8(static_cast<UINT8>(session_options.graph_optimization_level), "graph_optimization_level"),
TraceLoggingBoolean(session_options.use_per_session_threads, "use_per_session_threads"),
TraceLoggingBoolean(session_options.thread_pool_allow_spinning, "thread_pool_allow_spinning"),
TraceLoggingBoolean(session_options.use_deterministic_compute, "use_deterministic_compute"),
TraceLoggingBoolean(captureState, "isCaptureState"));
TraceLoggingWrite(
telemetry_provider_handle,
"SessionOptions_IntraOrtThreadPoolParams",
TraceLoggingKeyword(static_cast<uint64_t>(onnxruntime::logging::ORTTraceLoggingKeyword::Session)),
TraceLoggingLevel(WINEVENT_LEVEL_INFO),
TraceLoggingInt32(session_options.intra_op_param.thread_pool_size, "thread_pool_size"),
TraceLoggingBoolean(session_options.intra_op_param.auto_set_affinity, "auto_set_affinity"),
TraceLoggingBoolean(session_options.intra_op_param.allow_spinning, "allow_spinning"),
TraceLoggingInt32(session_options.intra_op_param.dynamic_block_base_, "dynamic_block_base_"),
TraceLoggingUInt32(session_options.intra_op_param.stack_size, "stack_size"),
TraceLoggingString(!session_options.intra_op_param.affinity_str.empty() ? session_options.intra_op_param.affinity_str.c_str() : "", "affinity_str"),
TraceLoggingBoolean(session_options.intra_op_param.set_denormal_as_zero, "set_denormal_as_zero"),
TraceLoggingBoolean(captureState, "isCaptureState"));
for (const auto& config_pair : session_options.config_options.configurations) {
TraceLoggingWrite(
telemetry_provider_handle,
"SessionOptions_ConfigEntry",
TraceLoggingKeyword(static_cast<uint64_t>(onnxruntime::logging::ORTTraceLoggingKeyword::Session)),
TraceLoggingLevel(WINEVENT_LEVEL_INFO),
TraceLoggingString(config_pair.first.c_str(), "Key"),
TraceLoggingString(config_pair.second.c_str(), "Value"),
TraceLoggingBoolean(captureState, "isCaptureState"));
}
#endif
}
InferenceSession::InferenceSession(const SessionOptions& session_options, const Environment& session_env)
:
#if !defined(ORT_MINIMAL_BUILD)
graph_transformer_mgr_(session_options.max_num_graph_transformation_steps),
#endif
environment_(session_env) {
// Initialize assets of this session instance
ConstructorCommon(session_options, session_env);
}
InferenceSession::InferenceSession(const SessionOptions& session_options,
const Environment& session_env,
onnxruntime::concurrency::ThreadPool* external_intra_op_thread_pool,
onnxruntime::concurrency::ThreadPool* external_inter_op_thread_pool)
:
#if !defined(ORT_MINIMAL_BUILD)
graph_transformer_mgr_(session_options.max_num_graph_transformation_steps),
#endif
external_intra_op_thread_pool_(external_intra_op_thread_pool),
external_inter_op_thread_pool_(external_inter_op_thread_pool),
environment_(session_env) {
// Initialize assets of this session instance
ConstructorCommon(session_options, session_env);
}
#if !defined(ORT_MINIMAL_BUILD)
InferenceSession::InferenceSession(const SessionOptions& session_options, const Environment& session_env,
const PathString& model_uri)
: model_location_(model_uri),
graph_transformer_mgr_(session_options.max_num_graph_transformation_steps),
environment_(session_env) {
auto status = Model::Load(model_location_, model_proto_);
ORT_ENFORCE(status.IsOK(), "Given model could not be parsed while creating inference session. Error message: ",
status.ErrorMessage());
is_model_proto_parsed_ = true;
// Finalize session options and initialize assets of this session instance
ConstructorCommon(session_options, session_env);
}
#ifdef _WIN32
InferenceSession::InferenceSession(const SessionOptions& session_options,
const Environment& session_env,
const std::string& model_uri)
: InferenceSession(session_options, session_env, ToPathString(model_uri)) {
}
#endif
InferenceSession::InferenceSession(const SessionOptions& session_options, const Environment& session_env,
std::istream& model_istream)
: graph_transformer_mgr_(session_options.max_num_graph_transformation_steps),
environment_(session_env) {
Status st = Model::Load(model_istream, &model_proto_);
ORT_ENFORCE(st.IsOK(), "Could not parse model successfully while constructing the inference session");
is_model_proto_parsed_ = true;
// Finalize session options and initialize assets of this session instance
ConstructorCommon(session_options, session_env);
}
InferenceSession::InferenceSession(const SessionOptions& session_options, const Environment& session_env,
const void* model_data, int model_data_len)
: graph_transformer_mgr_(session_options.max_num_graph_transformation_steps),
environment_(session_env) {
const bool result = model_proto_.ParseFromArray(model_data, model_data_len);
ORT_ENFORCE(result, "Could not parse model successfully while constructing the inference session");
is_model_proto_parsed_ = true;
// Finalize session options and initialize assets of this session instance
ConstructorCommon(session_options, session_env);
}
#endif // !defined(ORT_MINIMAL_BUILD)
InferenceSession::~InferenceSession() {
if (session_options_.enable_profiling) {
ORT_TRY {
EndProfiling();
}
ORT_CATCH(const std::exception& e) {
// TODO: Currently we have no way to transport this error to the API user
// Maybe this should be refactored, so that profiling must be explicitly
// started and stopped via C-API functions.
// And not like now a session option and therefore profiling must be started
// and stopped implicitly.
ORT_HANDLE_EXCEPTION([&]() {
LOGS(*session_logger_, ERROR) << "Error during EndProfiling(): " << e.what();
});
}
ORT_CATCH(...) {
LOGS(*session_logger_, ERROR) << "Unknown error during EndProfiling()";
}
}
// Unregister the session
#ifdef _WIN32
std::lock_guard<OrtMutex> lock(active_sessions_mutex_);
#endif
active_sessions_.erase(global_session_id_);
#ifdef ONNXRUNTIME_ENABLE_INSTRUMENT
if (session_activity_started_)
TraceLoggingWriteStop(session_activity, "OrtInferenceSessionActivity");
#endif
#if !defined(ORT_MINIMAL_BUILD) && defined(ORT_MEMORY_PROFILE)
GetMemoryProfiler().GenerateMemoryProfile();
#endif
}
common::Status InferenceSession::RegisterExecutionProvider(const std::shared_ptr<IExecutionProvider>& p_exec_provider) {
if (p_exec_provider == nullptr) {
return Status(common::ONNXRUNTIME, common::FAIL, "Received nullptr for exec provider");
}
std::lock_guard<onnxruntime::OrtMutex> l(session_mutex_);
if (is_inited_) {
// adding an EP is pointless as the graph as already been partitioned so no nodes will be assigned to
// the new EP
LOGS(*session_logger_, ERROR) << "Execution providers must be registered before the session is initialized. ";
return common::Status(common::ONNXRUNTIME, common::FAIL,
"Execution providers must be registered before the session is initialized.");
}
const std::string& provider_type = p_exec_provider->Type();
// Some session option values (default or user provided) may not work with some EPs.
// Rather than put the onus on the user to know these, make the appropriate change while logging the change.
if (provider_type == onnxruntime::kDmlExecutionProvider) {
// DML's memory is not byte addressable and hence mem pattern doesn't work.
if (session_options_.enable_mem_pattern) {
LOGS(*session_logger_, INFO)
<< "Having memory pattern enabled is not supported while using the DML Execution Provider. "
<< "So disabling it for this session since it uses the DML Execution Provider.";
session_options_.enable_mem_pattern = false;
}
// Parallel execution mode does not support DML EP
if (session_options_.execution_mode != ExecutionMode::ORT_SEQUENTIAL) {
LOGS(*session_logger_, INFO)
<< "Parallel execution mode does not support the DML Execution Provider. "
<< "So making the execution mode sequential for this session since it uses the DML Execution Provider.";
session_options_.execution_mode = ExecutionMode::ORT_SEQUENTIAL;
}
}
#if !defined(ORT_MINIMAL_BUILD) || defined(ORT_MINIMAL_BUILD_CUSTOM_OPS)
// Register Custom Op if EP requests it
std::vector<OrtCustomOpDomain*> custom_op_domains;
std::vector<OrtCustomOpDomain*> candidate_custom_op_domains;
p_exec_provider->GetCustomOpDomainList(candidate_custom_op_domains);
auto registry_kernels = kernel_registry_manager_.GetKernelRegistriesByProviderType(p_exec_provider->Type());
// Register the custom op domain only if it has not been registered before
if (registry_kernels.empty()) {
custom_op_domains = candidate_custom_op_domains;
} else {
for (auto candidate_custom_op_domain : candidate_custom_op_domains) {
for (auto registry_kernel : registry_kernels) {
const auto& kernel_map = registry_kernel->GetKernelCreateMap();
bool need_register = true;
// If the kernel registry is the ep's custom op registry, we only need to check the first kernel,
// because all kernels in one kernel registry should have the same domain name.
for (auto iter = kernel_map.begin(); iter != kernel_map.end(); iter++) {
if (iter->second.kernel_def->Domain() == candidate_custom_op_domain->domain_) {
need_register = false;
break;
}
}
if (need_register) {
custom_op_domains.push_back(candidate_custom_op_domain);
}
}
}
}
if (!custom_op_domains.empty()) {
if (AddCustomOpDomains(custom_op_domains) != Status::OK()) {
LOGS(*session_logger_, WARNING) << "Can't register custom op domains with ORT for " << provider_type;
}
}
#endif
// if any EPs do not support concurrent calls to Run we add locking around graph execution
if (p_exec_provider->ConcurrentRunSupported() == false) {
is_concurrent_run_supported_ = false;
}
VLOGS(*session_logger_, 1) << "Adding execution provider of type: " << provider_type;
auto p_data_xfr = p_exec_provider->GetDataTransfer();
if (p_data_xfr) {
auto st = data_transfer_mgr_.RegisterDataTransfer(std::move(p_data_xfr));
if (!st.IsOK()) {
return st;
}
}
p_exec_provider->SetLogger(session_logger_);
session_profiler_.AddEpProfilers(p_exec_provider->GetProfiler());
return execution_providers_.Add(provider_type, p_exec_provider);
}
// Custom Op support
#if !defined(ORT_MINIMAL_BUILD) || defined(ORT_MINIMAL_BUILD_CUSTOM_OPS)
common::Status InferenceSession::AddCustomOpDomains(gsl::span<OrtCustomOpDomain* const> op_domains) {
std::shared_ptr<CustomRegistry> custom_registry;
ORT_RETURN_IF_ERROR_SESSIONID_(CreateCustomRegistry(op_domains, custom_registry));
ORT_RETURN_IF_ERROR_SESSIONID_(RegisterCustomRegistry(custom_registry));
return Status::OK();
}
common::Status InferenceSession::RegisterCustomRegistry(std::shared_ptr<CustomRegistry> custom_registry) {
if (custom_registry == nullptr) {
return Status(common::ONNXRUNTIME, common::FAIL, "Received nullptr for custom registry");
}
custom_registries_.push_back(custom_registry);
// Insert session-level customized kernel registry.
kernel_registry_manager_.RegisterKernelRegistry(custom_registry->GetKernelRegistry());
#if !defined(ORT_MINIMAL_BUILD)
custom_schema_registries_.push_back(custom_registry->GetOpschemaRegistry());
#endif
return Status::OK();
}
#endif // !defined(ORT_MINIMAL_BUILD) || defined(ORT_MINIMAL_BUILD_CUSTOM_OPS)
#if !defined(ORT_MINIMAL_BUILD)
common::Status InferenceSession::RegisterGraphTransformer(
std::unique_ptr<onnxruntime::GraphTransformer> p_graph_transformer, TransformerLevel level) {
if (p_graph_transformer == nullptr) {
return Status(common::ONNXRUNTIME, common::FAIL, "Received nullptr for graph transformer");
}
std::lock_guard<onnxruntime::OrtMutex> l(session_mutex_);
if (is_inited_) {
// adding a transformer now is pointless as the graph as already been transformed
LOGS(*session_logger_, ERROR) << "Graph transformers must be registered before the session is initialized.";
return common::Status(common::ONNXRUNTIME, common::FAIL,
"Graph transformers must be registered before the session is initialized.");
}
return graph_transformer_mgr_.Register(std::move(p_graph_transformer), level);
}
common::Status InferenceSession::SaveToOrtFormat(const PathString& filepath) const {
ORT_RETURN_IF_NOT(FLATBUFFERS_LITTLEENDIAN, "ort format only supports little-endian machines");
// Get the byte size of the ModelProto and round it to the next MB and use it as flatbuffers' init_size
// TODO: Investigate whether we should set a max size, and clarify the cost of having a buffer smaller than
// what the total flatbuffers serialized size will be.
constexpr size_t m_bytes = 1024 * 1024;
size_t fbs_buffer_size = std::max(m_bytes, model_->ToProto().ByteSizeLong());
fbs_buffer_size = ((fbs_buffer_size + m_bytes - 1) / m_bytes) * m_bytes;
flatbuffers::FlatBufferBuilder builder(fbs_buffer_size);
auto ort_model_version = builder.CreateString(std::to_string(kOrtModelVersion));
flatbuffers::Offset<fbs::Model> fbs_model;
ORT_RETURN_IF_ERROR(
model_->SaveToOrtFormat(builder, fbs_model));
flatbuffers::Offset<fbs::KernelTypeStrResolver> fbs_kernel_type_str_resolver;
KernelTypeStrResolver kernel_type_str_resolver{};
ORT_RETURN_IF_ERROR(kernel_type_str_resolver.RegisterGraphNodeOpSchemas(model_->MainGraph()));
ORT_RETURN_IF_ERROR(standalone::RegisterCustomOpNodeSchemas(kernel_type_str_resolver, model_->MainGraph()));
for (const auto op_schema : saved_runtime_optimization_produced_node_op_schemas_) {
ORT_RETURN_IF_ERROR(kernel_type_str_resolver.RegisterOpSchema(*op_schema));
}
ORT_RETURN_IF_ERROR(
kernel_type_str_resolver.SaveToOrtFormat(builder, fbs_kernel_type_str_resolver));
fbs::InferenceSessionBuilder sb(builder);
sb.add_ort_version(ort_model_version);
sb.add_model(fbs_model);
sb.add_kernel_type_str_resolver(fbs_kernel_type_str_resolver);
auto session = sb.Finish();
builder.Finish(session, fbs::InferenceSessionIdentifier());
{
std::ofstream file(filepath, std::ios::binary);
uint8_t* buf = builder.GetBufferPointer();
int size = builder.GetSize();
file.write(reinterpret_cast<const char*>(buf), size);
ORT_RETURN_IF_NOT(file, "Failed to save ORT format model to file: ", ToUTF8String(filepath));
}
return Status::OK();
}
common::Status InferenceSession::LoadWithLoader(std::function<common::Status(std::shared_ptr<Model>&)> loader,
const std::string& event_name) {
Status status = Status::OK();
TimePoint tp;
if (session_profiler_.IsEnabled()) {
tp = session_profiler_.Start();
}
ORT_TRY {
std::lock_guard<onnxruntime::OrtMutex> l(session_mutex_);
if (is_model_loaded_) { // already loaded
LOGS(*session_logger_, ERROR) << "This session already contains a loaded model.";
return common::Status(common::ONNXRUNTIME, common::MODEL_LOADED, "This session already contains a loaded model.");
}
std::shared_ptr<onnxruntime::Model> p_tmp_model;
status = loader(p_tmp_model);
ORT_RETURN_IF_ERROR_SESSIONID_(status);
model_ = p_tmp_model;
status = DoPostLoadProcessing(*model_);
ORT_RETURN_IF_ERROR_SESSIONID_(status);
// all steps complete, mark the model as loaded.
is_model_loaded_ = true;
telemetry_.event_name_ = event_name;
}
ORT_CATCH(const std::exception& ex) {
ORT_HANDLE_EXCEPTION([&]() {
status = Status(common::ONNXRUNTIME, common::FAIL, "Exception during loading: " + std::string(ex.what()));
});
}
ORT_CATCH(...) {
LOGS(*session_logger_, ERROR) << "Unknown exception";
status = Status(common::ONNXRUNTIME, common::RUNTIME_EXCEPTION,
"Encountered unknown exception in LoadWithLoader()");
}
if (session_profiler_.IsEnabled()) {
session_profiler_.EndTimeAndRecordEvent(profiling::SESSION_EVENT, event_name, tp);
}
return status;
}
common::Status InferenceSession::LoadOnnxModel(const PathString& model_uri) {
model_location_ = model_uri;
auto loader = [this](std::shared_ptr<onnxruntime::Model>& model) {
#ifdef ENABLE_LANGUAGE_INTEROP_OPS
LoadInterOp(model_location_, interop_domains_, [&](const char* msg) { LOGS(*session_logger_, WARNING) << msg; });
InlinedVector<OrtCustomOpDomain*> domain_ptrs;
domain_ptrs.reserve(interop_domains_.size());
std::copy(std::begin(interop_domains_), std::end(interop_domains_), std::back_inserter(domain_ptrs));
ORT_RETURN_IF_ERROR(AddCustomOpDomains(domain_ptrs));
#endif
const bool strict_shape_type_inference = session_options_.config_options.GetConfigOrDefault(
kOrtSessionOptionsConfigStrictShapeTypeInference, "0") == "1";
return onnxruntime::Model::Load(model_location_, model, HasLocalSchema() ? &custom_schema_registries_ : nullptr,
*session_logger_,
ModelOptions(true, strict_shape_type_inference));
};
common::Status st = LoadWithLoader(loader, "model_loading_uri");
if (!st.IsOK()) {
std::ostringstream oss;
oss << "Load model from " << ToUTF8String(model_uri) << " failed:" << st.ErrorMessage();
return common::Status(st.Category(), st.Code(), oss.str());
}
return Status::OK();
}
#endif // !defined(ORT_MINIMAL_BUILD)
#if !defined(ORT_MINIMAL_BUILD) || defined(ORT_EXTENDED_MINIMAL_BUILD)
common::Status InferenceSession::FilterEnabledOptimizers(InlinedHashSet<std::string>&& optimizers_to_disable) {
optimizers_to_disable_ = std::move(optimizers_to_disable);
return Status::OK();
}
#endif // !defined(ORT_MINIMAL_BUILD) || defined(ORT_EXTENDED_MINIMAL_BUILD)
common::Status InferenceSession::Load(const PathString& model_uri) {
std::string model_type = session_options_.config_options.GetConfigOrDefault(kOrtSessionOptionsConfigLoadModelFormat, "");
bool has_explicit_type = !model_type.empty();
if ((has_explicit_type && model_type == "ORT") ||
(!has_explicit_type && fbs::utils::IsOrtFormatModel(model_uri))) {
return LoadOrtModel(model_uri);
}
#if !defined(ORT_MINIMAL_BUILD)
if (is_model_proto_parsed_) {
return ORT_MAKE_STATUS(ONNXRUNTIME, FAIL,
"ModelProto corresponding to the model to be loaded has already been parsed. "
"Invoke Load().");
}
return LoadOnnxModel(model_uri);
#else
return ORT_MAKE_STATUS(ONNXRUNTIME, INVALID_ARGUMENT, "ONNX format model is not supported in this build.");
#endif
}
#ifdef _WIN32
common::Status InferenceSession::Load(const std::string& model_uri) {
return Load(ToPathString(model_uri));
}
#endif
common::Status InferenceSession::Load(const void* model_data, int model_data_len) {
std::string model_type = session_options_.config_options.GetConfigOrDefault(kOrtSessionOptionsConfigLoadModelFormat, "");
bool has_explicit_type = !model_type.empty();
if ((has_explicit_type && model_type == "ORT") ||
(!has_explicit_type &&
fbs::utils::IsOrtFormatModelBytes(model_data, model_data_len))) {
return LoadOrtModel(model_data, model_data_len);
}
#if !defined(ORT_MINIMAL_BUILD)
if (is_model_proto_parsed_) {
return ORT_MAKE_STATUS(ONNXRUNTIME, FAIL,
"ModelProto corresponding to the model to be loaded has already been parsed. "
"Invoke Load().");
}
auto loader = [this, model_data, model_data_len](std::shared_ptr<onnxruntime::Model>& model) {
ModelProto model_proto;
const bool result = model_proto.ParseFromArray(model_data, model_data_len);
if (!result) {
return Status(common::ONNXRUNTIME, common::INVALID_PROTOBUF,
"Failed to load model because protobuf parsing failed.");
}
#ifdef ENABLE_LANGUAGE_INTEROP_OPS
LoadInterOp(model_proto, interop_domains_, [&](const char* msg) { LOGS(*session_logger_, WARNING) << msg; });
InlinedVector<OrtCustomOpDomain*> domain_ptrs;
domain_ptrs.reserve(interop_domains_.size());
std::copy(std::begin(interop_domains_), std::end(interop_domains_), std::back_inserter(domain_ptrs));
ORT_RETURN_IF_ERROR(AddCustomOpDomains(domain_ptrs));
#endif