forked from microsoft/onnxruntime
-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy pathtest_inference.cc
1434 lines (1206 loc) · 56.9 KB
/
test_inference.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 <memory>
#include <vector>
#include <iostream>
#include <fstream>
#include <sstream>
#include <atomic>
#include <mutex>
#include <algorithm>
#include <gtest/gtest.h>
#include "core/common/common.h"
#include "core/common/make_unique.h"
#include "core/graph/constants.h"
#include "core/session/onnxruntime_c_api.h"
#include "core/session/onnxruntime_cxx_api.h"
#include "core/session/onnxruntime_session_options_config_keys.h"
#include "providers.h"
#include "test_allocator.h"
#include "test_fixture.h"
#include "utils.h"
#include "custom_op_utils.h"
#ifdef _WIN32
#include <Windows.h>
#else
#include <dlfcn.h>
#endif
#ifdef USE_CUDA
#include <cuda_runtime.h>
#endif
// Once we use C++17 this could be replaced with std::size
template <typename T, size_t N>
constexpr size_t countof(T (&)[N]) { return N; }
extern std::unique_ptr<Ort::Env> ort_env;
template <typename OutT>
void RunSession(OrtAllocator* allocator, Ort::Session& session_object,
const std::vector<Input>& inputs,
const char* output_name,
const std::vector<int64_t>& dims_y,
const std::vector<OutT>& values_y,
Ort::Value* output_tensor) {
std::vector<Ort::Value> ort_inputs;
std::vector<const char*> input_names;
for (size_t i = 0; i < inputs.size(); i++) {
input_names.emplace_back(inputs[i].name);
ort_inputs.emplace_back(
Ort::Value::CreateTensor<float>(allocator->Info(allocator), const_cast<float*>(inputs[i].values.data()),
inputs[i].values.size(), inputs[i].dims.data(), inputs[i].dims.size()));
}
std::vector<Ort::Value> ort_outputs;
if (output_tensor)
session_object.Run(Ort::RunOptions{nullptr}, input_names.data(), ort_inputs.data(), ort_inputs.size(),
&output_name, output_tensor, 1);
else {
ort_outputs = session_object.Run(Ort::RunOptions{}, input_names.data(), ort_inputs.data(), ort_inputs.size(),
&output_name, 1);
ASSERT_EQ(ort_outputs.size(), 1u);
output_tensor = &ort_outputs[0];
}
auto type_info = output_tensor->GetTensorTypeAndShapeInfo();
ASSERT_EQ(type_info.GetShape(), dims_y);
size_t total_len = type_info.GetElementCount();
ASSERT_EQ(values_y.size(), total_len);
OutT* f = output_tensor->GetTensorMutableData<OutT>();
for (size_t i = 0; i != total_len; ++i) {
ASSERT_EQ(values_y[i], f[i]);
}
}
template <typename OutT>
static void TestInference(Ort::Env& env, const std::basic_string<ORTCHAR_T>& model_uri,
const std::vector<Input>& inputs,
const char* output_name,
const std::vector<int64_t>& expected_dims_y,
const std::vector<OutT>& expected_values_y,
int provider_type,
OrtCustomOpDomain* custom_op_domain_ptr,
const char* custom_op_library_filename,
void** library_handle = nullptr,
bool test_session_creation_only = false,
void* cuda_compute_stream = nullptr) {
Ort::SessionOptions session_options;
if (provider_type == 1) {
#ifdef USE_CUDA
std::cout << "Running simple inference with cuda provider" << std::endl;
auto cuda_options = CreateDefaultOrtCudaProviderOptionsWithCustomStream(cuda_compute_stream);
session_options.AppendExecutionProvider_CUDA(cuda_options);
#else
ORT_UNUSED_PARAMETER(cuda_compute_stream);
return;
#endif
} else if (provider_type == 2) {
#ifdef USE_DNNL
Ort::ThrowOnError(OrtSessionOptionsAppendExecutionProvider_Dnnl(session_options, 1));
std::cout << "Running simple inference with dnnl provider" << std::endl;
#else
return;
#endif
} else if (provider_type == 3) {
#ifdef USE_NUPHAR
Ort::ThrowOnError(OrtSessionOptionsAppendExecutionProvider_Nuphar(session_options,
/*allow_unaligned_buffers*/ 1, ""));
std::cout << "Running simple inference with nuphar provider" << std::endl;
#else
return;
#endif
} else {
std::cout << "Running simple inference with default provider" << std::endl;
}
if (custom_op_domain_ptr) {
session_options.Add(custom_op_domain_ptr);
}
if (custom_op_library_filename) {
Ort::ThrowOnError(Ort::GetApi().RegisterCustomOpsLibrary(session_options,
custom_op_library_filename, library_handle));
}
// if session creation passes, model loads fine
Ort::Session session(env, model_uri.c_str(), session_options);
// caller wants to test running the model (not just loading the model)
if (!test_session_creation_only) {
// Now run
auto default_allocator = onnxruntime::make_unique<MockedOrtAllocator>();
//without preallocated output tensor
RunSession<OutT>(default_allocator.get(),
session,
inputs,
output_name,
expected_dims_y,
expected_values_y,
nullptr);
//with preallocated output tensor
Ort::Value value_y = Ort::Value::CreateTensor<float>(default_allocator.get(),
expected_dims_y.data(), expected_dims_y.size());
//test it twice
for (int i = 0; i != 2; ++i)
RunSession<OutT>(default_allocator.get(),
session,
inputs,
output_name,
expected_dims_y,
expected_values_y,
&value_y);
}
}
static constexpr PATH_TYPE MODEL_URI = TSTR("testdata/mul_1.onnx");
static constexpr PATH_TYPE SEQUENCE_MODEL_URI = TSTR("testdata/sequence_length.onnx");
static constexpr PATH_TYPE CUSTOM_OP_MODEL_URI = TSTR("testdata/foo_1.onnx");
static constexpr PATH_TYPE CUSTOM_OP_LIBRARY_TEST_MODEL_URI = TSTR("testdata/custom_op_library/custom_op_test.onnx");
static constexpr PATH_TYPE OVERRIDABLE_INITIALIZER_MODEL_URI = TSTR("testdata/overridable_initializer.onnx");
static constexpr PATH_TYPE NAMED_AND_ANON_DIM_PARAM_URI = TSTR("testdata/capi_symbolic_dims.onnx");
static constexpr PATH_TYPE MODEL_WITH_CUSTOM_MODEL_METADATA = TSTR("testdata/model_with_valid_ort_config_json.onnx");
static constexpr PATH_TYPE VARIED_INPUT_CUSTOM_OP_MODEL_URI = TSTR("testdata/VariedInputCustomOp.onnx");
static constexpr PATH_TYPE VARIED_INPUT_CUSTOM_OP_MODEL_URI_2 = TSTR("testdata/foo_3.onnx");
static constexpr PATH_TYPE OPTIONAL_INPUT_OUTPUT_CUSTOM_OP_MODEL_URI = TSTR("testdata/foo_bar_1.onnx");
static constexpr PATH_TYPE OPTIONAL_INPUT_OUTPUT_CUSTOM_OP_MODEL_URI_2 = TSTR("testdata/foo_bar_2.onnx");
static constexpr PATH_TYPE CUSTOM_OP_MODEL_WITH_ATTRIBUTES_URI = TSTR("testdata/foo_bar_3.onnx");
#ifdef ENABLE_LANGUAGE_INTEROP_OPS
static constexpr PATH_TYPE PYOP_FLOAT_MODEL_URI = TSTR("testdata/pyop_1.onnx");
static constexpr PATH_TYPE PYOP_MULTI_MODEL_URI = TSTR("testdata/pyop_2.onnx");
static constexpr PATH_TYPE PYOP_KWARG_MODEL_URI = TSTR("testdata/pyop_3.onnx");
#endif
class CApiTestWithProvider : public testing::Test, public ::testing::WithParamInterface<int> {
};
TEST_P(CApiTestWithProvider, simple) {
// simple inference test
// prepare inputs
std::vector<Input> inputs(1);
Input& input = inputs.back();
input.name = "X";
input.dims = {3, 2};
input.values = {1.0f, 2.0f, 3.0f, 4.0f, 5.0f, 6.0f};
// prepare expected inputs and outputs
std::vector<int64_t> expected_dims_y = {3, 2};
std::vector<float> expected_values_y = {1.0f, 4.0f, 9.0f, 16.0f, 25.0f, 36.0f};
TestInference<float>(*ort_env, MODEL_URI, inputs, "Y", expected_dims_y, expected_values_y, GetParam(),
nullptr, nullptr);
}
TEST(CApiTest, dim_param) {
Ort::SessionOptions session_options;
Ort::Session session(*ort_env, NAMED_AND_ANON_DIM_PARAM_URI, session_options);
auto in0 = session.GetInputTypeInfo(0);
auto in0_ttsi = in0.GetTensorTypeAndShapeInfo();
auto num_input_dims = in0_ttsi.GetDimensionsCount();
ASSERT_GE(num_input_dims, 1u);
// reading 1st dimension only so don't need to malloc int64_t* or const char** values for the Get*Dimensions calls
int64_t dim_value = 0;
const char* dim_param = nullptr;
in0_ttsi.GetDimensions(&dim_value, 1);
in0_ttsi.GetSymbolicDimensions(&dim_param, 1);
ASSERT_EQ(dim_value, -1) << "symbolic dimension should be -1";
ASSERT_EQ(strcmp(dim_param, "n"), 0) << "Expected 'n'. Got: " << dim_param;
auto out0 = session.GetOutputTypeInfo(0);
auto out0_ttsi = out0.GetTensorTypeAndShapeInfo();
auto num_output_dims = out0_ttsi.GetDimensionsCount();
ASSERT_EQ(num_output_dims, 1u);
out0_ttsi.GetDimensions(&dim_value, 1);
out0_ttsi.GetSymbolicDimensions(&dim_param, 1);
ASSERT_EQ(dim_value, -1) << "symbolic dimension should be -1";
ASSERT_EQ(strcmp(dim_param, ""), 0);
}
INSTANTIATE_TEST_SUITE_P(CApiTestWithProviders,
CApiTestWithProvider,
::testing::Values(0, 1, 2, 3, 4));
TEST(CApiTest, custom_op_handler) {
std::cout << "Running custom op inference" << std::endl;
std::vector<Input> inputs(1);
Input& input = inputs[0];
input.name = "X";
input.dims = {3, 2};
input.values = {1.0f, 2.0f, 3.0f, 4.0f, 5.0f, 6.0f};
// prepare expected inputs and outputs
std::vector<int64_t> expected_dims_y = {3, 2};
std::vector<float> expected_values_y = {2.0f, 4.0f, 6.0f, 8.0f, 10.0f, 12.0f};
#ifdef USE_CUDA
cudaStream_t compute_stream = nullptr;
cudaStreamCreateWithFlags(&compute_stream, cudaStreamNonBlocking);
MyCustomOp custom_op{onnxruntime::kCudaExecutionProvider, compute_stream};
#else
MyCustomOp custom_op{onnxruntime::kCpuExecutionProvider, nullptr};
#endif
Ort::CustomOpDomain custom_op_domain("");
custom_op_domain.Add(&custom_op);
#ifdef USE_CUDA
TestInference<float>(*ort_env, CUSTOM_OP_MODEL_URI, inputs, "Y", expected_dims_y, expected_values_y, 1,
custom_op_domain, nullptr, nullptr, false, compute_stream);
cudaStreamDestroy(compute_stream);
#else
TestInference<float>(*ort_env, CUSTOM_OP_MODEL_URI, inputs, "Y", expected_dims_y, expected_values_y, 0,
custom_op_domain, nullptr);
#endif
}
//test custom op which accepts float and double as inputs
TEST(CApiTest, varied_input_custom_op_handler) {
std::vector<Input> inputs(2);
inputs[0].name = "X";
inputs[0].dims = {3};
inputs[0].values = {2.0f, 3.0f, 4.0f};
inputs[1].name = "Y";
inputs[1].dims = {3};
inputs[1].values = {5.0f, 6.0f, 7.0f};
std::vector<int64_t> expected_dims_z = {1};
std::vector<float> expected_values_z = {10.0f};
#ifdef USE_CUDA
cudaStream_t compute_stream = nullptr;
cudaStreamCreateWithFlags(&compute_stream, cudaStreamNonBlocking);
SliceCustomOp slice_custom_op{onnxruntime::kCudaExecutionProvider, compute_stream};
#else
SliceCustomOp slice_custom_op{onnxruntime::kCpuExecutionProvider, nullptr};
#endif
Ort::CustomOpDomain custom_op_domain("abc");
custom_op_domain.Add(&slice_custom_op);
#ifdef USE_CUDA
TestInference<float>(*ort_env, VARIED_INPUT_CUSTOM_OP_MODEL_URI, inputs, "Z",
expected_dims_z, expected_values_z, 1, custom_op_domain, nullptr, nullptr, false, compute_stream);
cudaStreamDestroy(compute_stream);
#else
TestInference<float>(*ort_env, VARIED_INPUT_CUSTOM_OP_MODEL_URI, inputs, "Z",
expected_dims_z, expected_values_z, 0, custom_op_domain, nullptr);
#endif
}
TEST(CApiTest, multiple_varied_input_custom_op_handler) {
#ifdef USE_CUDA
cudaStream_t compute_stream = nullptr;
cudaStreamCreateWithFlags(&compute_stream, cudaStreamNonBlocking);
MyCustomOpMultipleDynamicInputs custom_op{onnxruntime::kCudaExecutionProvider, compute_stream};
#else
MyCustomOpMultipleDynamicInputs custom_op{onnxruntime::kCpuExecutionProvider, nullptr};
#endif
Ort::CustomOpDomain custom_op_domain("");
custom_op_domain.Add(&custom_op);
Ort::SessionOptions session_options;
#ifdef USE_CUDA
auto cuda_options = CreateDefaultOrtCudaProviderOptionsWithCustomStream(compute_stream);
session_options.AppendExecutionProvider_CUDA(cuda_options);
#endif
session_options.Add(custom_op_domain);
Ort::Session session(*ort_env, VARIED_INPUT_CUSTOM_OP_MODEL_URI_2, session_options);
Ort::MemoryInfo info("Cpu", OrtDeviceAllocator, 0, OrtMemTypeDefault);
std::vector<Ort::Value> ort_inputs;
std::vector<const char*> input_names;
// input 0 (float type)
input_names.emplace_back("X");
std::vector<float> input_0_data = {1.f, 2.f, 3.f, 4.f, 5.f, 6.f};
std::vector<int64_t> input_0_dims = {3, 2};
ort_inputs.emplace_back(
Ort::Value::CreateTensor<float>(info, const_cast<float*>(input_0_data.data()),
input_0_data.size(), input_0_dims.data(), input_0_dims.size()));
// input 1 (double type)
input_names.emplace_back("W");
std::vector<double> input_1_data = {2, 3, 4, 5, 6, 7};
std::vector<int64_t> input_1_dims = {3, 2};
ort_inputs.emplace_back(
Ort::Value::CreateTensor<double>(info, const_cast<double*>(input_1_data.data()),
input_1_data.size(), input_1_dims.data(), input_1_dims.size()));
// Run
const char* output_name = "Y";
auto ort_outputs = session.Run(Ort::RunOptions{}, input_names.data(), ort_inputs.data(), ort_inputs.size(),
&output_name, 1);
ASSERT_EQ(ort_outputs.size(), 1u);
// Validate results
std::vector<int64_t> y_dims = {3, 2};
std::vector<float> values_y = {3.f, 5.f, 7.f, 9.f, 11.f, 13.f};
auto type_info = ort_outputs[0].GetTensorTypeAndShapeInfo();
ASSERT_EQ(type_info.GetShape(), y_dims);
size_t total_len = type_info.GetElementCount();
ASSERT_EQ(values_y.size(), total_len);
float* f = ort_outputs[0].GetTensorMutableData<float>();
for (size_t i = 0; i != total_len; ++i) {
ASSERT_EQ(values_y[i], f[i]);
}
#ifdef USE_CUDA
cudaStreamDestroy(compute_stream);
#endif
}
TEST(CApiTest, optional_input_output_custom_op_handler) {
MyCustomOpWithOptionalInput custom_op{onnxruntime::kCpuExecutionProvider};
// `MyCustomOpFooBar` defines a custom op with atmost 3 inputs and the second input is optional.
// In this test, we are going to try and run 2 models - one with the optional input and one without
// the optional input.
Ort::CustomOpDomain custom_op_domain("");
custom_op_domain.Add(&custom_op);
Ort::SessionOptions session_options;
session_options.Add(custom_op_domain);
std::vector<Ort::Value> ort_inputs;
Ort::MemoryInfo info("Cpu", OrtDeviceAllocator, 0, OrtMemTypeDefault);
// input 0
std::vector<float> input_0_data = {1.f};
std::vector<int64_t> input_0_dims = {1};
ort_inputs.emplace_back(
Ort::Value::CreateTensor<float>(info, const_cast<float*>(input_0_data.data()),
input_0_data.size(), input_0_dims.data(), input_0_dims.size()));
// input 1
std::vector<float> input_1_data = {1.f};
std::vector<int64_t> input_1_dims = {1};
ort_inputs.emplace_back(
Ort::Value::CreateTensor<float>(info, const_cast<float*>(input_1_data.data()),
input_1_data.size(), input_1_dims.data(), input_1_dims.size()));
// input 2
std::vector<float> input_2_data = {1.f};
std::vector<int64_t> input_2_dims = {1};
ort_inputs.emplace_back(
Ort::Value::CreateTensor<float>(info, const_cast<float*>(input_2_data.data()),
input_2_data.size(), input_2_dims.data(), input_2_dims.size()));
const char* output_name = "Y";
// Part 1: Model with optional input present
{
std::vector<const char*> input_names = {"X1", "X2", "X3"};
Ort::Session session(*ort_env, OPTIONAL_INPUT_OUTPUT_CUSTOM_OP_MODEL_URI, session_options);
auto ort_outputs = session.Run(Ort::RunOptions{}, input_names.data(), ort_inputs.data(), ort_inputs.size(),
&output_name, 1);
ASSERT_EQ(ort_outputs.size(), 1u);
// Validate results
std::vector<int64_t> y_dims = {1};
std::vector<float> values_y = {3.f};
auto type_info = ort_outputs[0].GetTensorTypeAndShapeInfo();
ASSERT_EQ(type_info.GetShape(), y_dims);
size_t total_len = type_info.GetElementCount();
ASSERT_EQ(values_y.size(), total_len);
float* f = ort_outputs[0].GetTensorMutableData<float>();
for (size_t i = 0; i != total_len; ++i) {
ASSERT_EQ(values_y[i], f[i]);
}
}
// Part 2: Model with optional input absent
{
std::vector<const char*> input_names = {"X1", "X2"};
ort_inputs.erase(ort_inputs.begin() + 2); // remove the last input in the container
Ort::Session session(*ort_env, OPTIONAL_INPUT_OUTPUT_CUSTOM_OP_MODEL_URI_2, session_options);
auto ort_outputs = session.Run(Ort::RunOptions{}, input_names.data(), ort_inputs.data(), ort_inputs.size(),
&output_name, 1);
ASSERT_EQ(ort_outputs.size(), 1u);
// Validate results
std::vector<int64_t> y_dims = {1};
std::vector<float> values_y = {2.f};
auto type_info = ort_outputs[0].GetTensorTypeAndShapeInfo();
ASSERT_EQ(type_info.GetShape(), y_dims);
size_t total_len = type_info.GetElementCount();
ASSERT_EQ(values_y.size(), total_len);
float* f = ort_outputs[0].GetTensorMutableData<float>();
for (size_t i = 0; i != total_len; ++i) {
ASSERT_EQ(values_y[i], f[i]);
}
}
}
TEST(CApiTest, custom_op_with_attributes_handler) {
MyCustomOpWithAttributes custom_op{onnxruntime::kCpuExecutionProvider};
Ort::CustomOpDomain custom_op_domain("");
custom_op_domain.Add(&custom_op);
Ort::SessionOptions session_options;
session_options.Add(custom_op_domain);
Ort::Session session(*ort_env, CUSTOM_OP_MODEL_WITH_ATTRIBUTES_URI, session_options);
Ort::MemoryInfo info("Cpu", OrtDeviceAllocator, 0, OrtMemTypeDefault);
std::vector<Ort::Value> ort_inputs;
std::vector<const char*> input_names;
// input 0 (float type)
input_names.emplace_back("X");
std::vector<float> input_0_data = {1.f};
std::vector<int64_t> input_0_dims = {1};
ort_inputs.emplace_back(
Ort::Value::CreateTensor<float>(info, const_cast<float*>(input_0_data.data()),
input_0_data.size(), input_0_dims.data(), input_0_dims.size()));
// Run
const char* output_name = "Y";
auto ort_outputs = session.Run(Ort::RunOptions{}, input_names.data(), ort_inputs.data(), ort_inputs.size(),
&output_name, 1);
ASSERT_EQ(ort_outputs.size(), 1u);
// Validate results
std::vector<int64_t> y_dims = {1};
std::vector<float> values_y = {15.f};
auto type_info = ort_outputs[0].GetTensorTypeAndShapeInfo();
ASSERT_EQ(type_info.GetShape(), y_dims);
size_t total_len = type_info.GetElementCount();
ASSERT_EQ(values_y.size(), total_len);
float* f = ort_outputs[0].GetTensorMutableData<float>();
for (size_t i = 0; i != total_len; ++i) {
ASSERT_EQ(values_y[i], f[i]);
}
}
// Tests registration of a custom op of the same name for both CPU and CUDA EPs
#ifdef USE_CUDA
TEST(CApiTest, RegisterCustomOpForCPUAndCUDA) {
std::cout << "Tests registration of a custom op of the same name for both CPU and CUDA EPs" << std::endl;
std::vector<Input> inputs(1);
Input& input = inputs[0];
input.name = "X";
input.dims = {3, 2};
input.values = {1.0f, 2.0f, 3.0f, 4.0f, 5.0f, 6.0f};
// prepare expected inputs and outputs
std::vector<int64_t> expected_dims_y = {3, 2};
std::vector<float> expected_values_y = {2.0f, 4.0f, 6.0f, 8.0f, 10.0f, 12.0f};
MyCustomOp custom_op_cpu{onnxruntime::kCpuExecutionProvider, nullptr};
// We are going to test session creation only - hence it is not a problem to use the default stream as the compute stream for the custom op
MyCustomOp custom_op_cuda{onnxruntime::kCudaExecutionProvider, nullptr};
Ort::CustomOpDomain custom_op_domain("");
custom_op_domain.Add(&custom_op_cpu);
custom_op_domain.Add(&custom_op_cuda);
TestInference<float>(*ort_env, CUSTOM_OP_MODEL_URI, inputs, "Y", expected_dims_y,
expected_values_y, 1, custom_op_domain, nullptr, nullptr, true);
}
#endif
//It has memory leak. The OrtCustomOpDomain created in custom_op_library.cc:RegisterCustomOps function was not freed
#if defined(__ANDROID__)
TEST(CApiTest, DISABLED_test_custom_op_library) {
#else
TEST(CApiTest, test_custom_op_library) {
#endif
std::cout << "Running inference using custom op shared library" << std::endl;
std::vector<Input> inputs(2);
inputs[0].name = "input_1";
inputs[0].dims = {3, 5};
inputs[0].values = {1.1f, 2.2f, 3.3f, 4.4f, 5.5f,
6.6f, 7.7f, 8.8f, 9.9f, 10.0f,
11.1f, 12.2f, 13.3f, 14.4f, 15.5f};
inputs[1].name = "input_2";
inputs[1].dims = {3, 5};
inputs[1].values = {15.5f, 14.4f, 13.3f, 12.2f, 11.1f,
10.0f, 9.9f, 8.8f, 7.7f, 6.6f,
5.5f, 4.4f, 3.3f, 2.2f, 1.1f};
// prepare expected inputs and outputs
std::vector<int64_t> expected_dims_y = {3, 5};
std::vector<int32_t> expected_values_y =
{17, 17, 17, 17, 17,
17, 18, 18, 18, 17,
17, 17, 17, 17, 17};
std::string lib_name;
#if defined(_WIN32)
lib_name = "custom_op_library.dll";
#elif defined(__APPLE__)
lib_name = "libcustom_op_library.dylib";
#else
lib_name = "./libcustom_op_library.so";
#endif
void* library_handle = nullptr;
TestInference<int32_t>(*ort_env, CUSTOM_OP_LIBRARY_TEST_MODEL_URI, inputs, "output", expected_dims_y,
expected_values_y, 0, nullptr, lib_name.c_str(), &library_handle);
#ifdef _WIN32
bool success = ::FreeLibrary(reinterpret_cast<HMODULE>(library_handle));
ORT_ENFORCE(success, "Error while closing custom op shared library");
#else
int retval = dlclose(library_handle);
ORT_ENFORCE(retval == 0, "Error while closing custom op shared library");
#endif
}
#if defined(ENABLE_LANGUAGE_INTEROP_OPS)
std::once_flag my_module_flag;
void PrepareModule() {
std::ofstream module("mymodule.py");
module << "class MyKernel:" << std::endl;
module << "\t"
<< "def __init__(self,A,B,C):" << std::endl;
module << "\t\t"
<< "self.a,self.b,self.c = A,B,C" << std::endl;
module << "\t"
<< "def compute(self,x):" << std::endl;
module << "\t\t"
<< "return x*2" << std::endl;
module << "class MyKernel_2:" << std::endl;
module << "\t"
<< "def __init__(self,A,B):" << std::endl;
module << "\t\t"
<< "self.a,self.b = A,B" << std::endl;
module << "\t"
<< "def compute(self,x):" << std::endl;
module << "\t\t"
<< "return x*4" << std::endl;
module << "class MyKernel_3:" << std::endl;
module << "\t"
<< "def __init__(self,A,B):" << std::endl;
module << "\t\t"
<< "self.a,self.b = A,B" << std::endl;
module << "\t"
<< "def compute(self,*kwargs):" << std::endl;
module << "\t\t"
<< "return kwargs[0]*5" << std::endl;
module.close();
}
TEST(CApiTest, test_pyop) {
std::call_once(my_module_flag, PrepareModule);
std::vector<Input> inputs(1);
Input& input = inputs[0];
input.name = "X";
input.dims = {2, 2};
input.values = {1.0f, 2.0f, 3.0f, 4.0f};
std::vector<int64_t> expected_dims_y = {2, 2};
std::vector<float> expected_values_y = {2.0f, 4.0f, 6.0f, 8.0f};
TestInference<float>(*ort_env, PYOP_FLOAT_MODEL_URI, inputs, "Y", expected_dims_y, expected_values_y, 0,
nullptr, nullptr);
}
TEST(CApiTest, test_pyop_multi) {
std::call_once(my_module_flag, PrepareModule);
std::vector<Input> inputs(1);
Input& input = inputs[0];
input.name = "X";
input.dims = {2, 2};
input.values = {1.0f, 2.0f, 3.0f, 4.0f};
std::vector<int64_t> expected_dims_y = {2, 2};
std::vector<float> expected_values_y = {8.0f, 16.0f, 24.0f, 32.0f};
TestInference<float>(*ort_env, PYOP_MULTI_MODEL_URI, inputs, "Z", expected_dims_y, expected_values_y, 0,
nullptr, nullptr);
}
TEST(CApiTest, test_pyop_kwarg) {
std::call_once(my_module_flag, PrepareModule);
std::vector<Input> inputs(1);
Input& input = inputs[0];
input.name = "X";
input.dims = {2, 2};
input.values = {1.0f, 2.0f, 3.0f, 4.0f};
std::vector<int64_t> expected_dims_y = {2, 2};
std::vector<float> expected_values_y = {25.0f, 50.0f, 75.0f, 100.0f};
TestInference<float>(*ort_env, PYOP_KWARG_MODEL_URI, inputs, "Z", expected_dims_y, expected_values_y, 0,
nullptr, nullptr);
}
#endif
#ifdef ORT_RUN_EXTERNAL_ONNX_TESTS
TEST(CApiTest, create_session_without_session_option) {
constexpr PATH_TYPE model_uri = TSTR("../models/opset8/test_squeezenet/model.onnx");
Ort::Session ret(*ort_env, model_uri, Ort::SessionOptions{nullptr});
ASSERT_NE(nullptr, ret);
}
#endif
#ifdef REDUCED_OPS_BUILD
TEST(ReducedOpsBuildTest, test_excluded_ops) {
// In reduced ops build, test a model containing ops not included in required_ops.config cannot be loaded.
// See onnxruntime/test/testdata/reduced_build_test.readme.txt for more details of the setup
constexpr PATH_TYPE model_uri = TSTR("testdata/reduced_build_test.onnx_model_with_excluded_ops");
std::vector<Input> inputs = {{"X", {3}, {-1.0f, 2.0f, -3.0f}}};
std::vector<int64_t> expected_dims_y = {3};
std::vector<float> expected_values_y = {0.1f, 0.1f, 0.1f};
bool failed = false;
try {
//only test model loading, exception expected
TestInference<float>(*ort_env, model_uri, inputs, "Y", expected_dims_y, expected_values_y, 0,
nullptr, nullptr, nullptr, true);
} catch (const Ort::Exception& e) {
failed = e.GetOrtErrorCode() == ORT_NOT_IMPLEMENTED;
}
ASSERT_EQ(failed, true);
}
#endif
TEST(CApiTest, get_allocator_cpu) {
Ort::SessionOptions session_options;
Ort::ThrowOnError(OrtSessionOptionsAppendExecutionProvider_CPU(session_options, 1));
Ort::Session session(*ort_env, NAMED_AND_ANON_DIM_PARAM_URI, session_options);
Ort::MemoryInfo info_cpu = Ort::MemoryInfo::CreateCpu(OrtAllocatorType::OrtArenaAllocator, OrtMemTypeDefault);
Ort::Allocator cpu_allocator(session, info_cpu);
// CPU OrtMemoryInfo does not return OrtArenaAllocator on x86 but rather a device allocator
// which causes MemoryInfo that is used to request the allocator and the actual instance
// of MemoryInfo returned from the allocator exactly match, although they are functionally equivalent.
auto allocator_info = cpu_allocator.GetInfo();
ASSERT_EQ(info_cpu.GetAllocatorName(), allocator_info.GetAllocatorName());
ASSERT_EQ(info_cpu.GetDeviceId(), allocator_info.GetDeviceId());
ASSERT_EQ(info_cpu.GetMemoryType(), allocator_info.GetDeviceId());
void* p = cpu_allocator.Alloc(1024);
ASSERT_NE(p, nullptr);
cpu_allocator.Free(p);
auto mem_allocation = cpu_allocator.GetAllocation(1024);
ASSERT_NE(nullptr, mem_allocation.get());
ASSERT_EQ(1024U, mem_allocation.size());
}
#ifdef USE_CUDA
TEST(CApiTest, get_allocator_cuda) {
Ort::SessionOptions session_options;
Ort::ThrowOnError(OrtSessionOptionsAppendExecutionProvider_CUDA(session_options, 0));
Ort::Session session(*ort_env, NAMED_AND_ANON_DIM_PARAM_URI, session_options);
Ort::MemoryInfo info_cuda("Cuda", OrtAllocatorType::OrtArenaAllocator, 0, OrtMemTypeDefault);
Ort::Allocator cuda_allocator(session, info_cuda);
auto allocator_info = cuda_allocator.GetInfo();
ASSERT_TRUE(info_cuda == allocator_info);
void* p = cuda_allocator.Alloc(1024);
ASSERT_NE(p, nullptr);
cuda_allocator.Free(p);
auto mem_allocation = cuda_allocator.GetAllocation(1024);
ASSERT_NE(nullptr, mem_allocation.get());
ASSERT_EQ(1024U, mem_allocation.size());
}
#endif
TEST(CApiTest, io_binding) {
Ort::SessionOptions session_options;
Ort::ThrowOnError(OrtSessionOptionsAppendExecutionProvider_CPU(session_options, 1));
Ort::Session session(*ort_env, MODEL_URI, session_options);
Ort::MemoryInfo info_cpu = Ort::MemoryInfo::CreateCpu(OrtAllocatorType::OrtArenaAllocator, OrtMemTypeDefault);
const std::array<int64_t, 2> x_shape = {3, 2};
std::array<float, 3 * 2> x_values = {1.0f, 2.0f, 3.0f, 4.0f, 5.0f, 6.0f};
Ort::Value bound_x = Ort::Value::CreateTensor(info_cpu, x_values.data(), x_values.size(),
x_shape.data(), x_shape.size());
const std::array<float, 3 * 2> expected_y = {1.0f, 4.0f, 9.0f, 16.0f, 25.0f, 36.0f};
const std::array<int64_t, 2> y_shape = {3, 2};
std::array<float, 3 * 2> y_values;
Ort::Value bound_y = Ort::Value::CreateTensor(info_cpu, y_values.data(), y_values.size(),
y_shape.data(), y_shape.size());
Ort::IoBinding binding(session);
binding.BindInput("X", bound_x);
binding.BindOutput("Y", bound_y);
session.Run(Ort::RunOptions(), binding);
// Check the values against the bound raw memory
ASSERT_TRUE(std::equal(std::begin(y_values), std::end(y_values), std::begin(expected_y)));
// Now compare values via GetOutputValues
{
std::vector<Ort::Value> output_values = binding.GetOutputValues();
ASSERT_EQ(output_values.size(), 1U);
const Ort::Value& Y_value = output_values[0];
ASSERT_TRUE(Y_value.IsTensor());
Ort::TensorTypeAndShapeInfo type_info = Y_value.GetTensorTypeAndShapeInfo();
ASSERT_EQ(ONNX_TENSOR_ELEMENT_DATA_TYPE_FLOAT, type_info.GetElementType());
auto count = type_info.GetElementCount();
ASSERT_EQ(expected_y.size(), count);
const float* values = Y_value.GetTensorData<float>();
ASSERT_TRUE(std::equal(values, values + count, std::begin(expected_y)));
}
{
std::vector<std::string> output_names = binding.GetOutputNames();
ASSERT_EQ(1U, output_names.size());
ASSERT_EQ(output_names[0].compare("Y"), 0);
}
// Now replace binding of Y with an on device binding instead of pre-allocated memory.
// This is when we can not allocate an OrtValue due to unknown dimensions
{
Ort::MemoryInfo info_cpu_dev("Cpu", OrtAllocatorType::OrtArenaAllocator, 0, OrtMemTypeDefault);
binding.BindOutput("Y", info_cpu_dev);
session.Run(Ort::RunOptions(), binding);
}
// Check the output value allocated based on the device binding.
{
std::vector<Ort::Value> output_values = binding.GetOutputValues();
ASSERT_EQ(output_values.size(), 1U);
const Ort::Value& Y_value = output_values[0];
ASSERT_TRUE(Y_value.IsTensor());
Ort::TensorTypeAndShapeInfo type_info = Y_value.GetTensorTypeAndShapeInfo();
ASSERT_EQ(ONNX_TENSOR_ELEMENT_DATA_TYPE_FLOAT, type_info.GetElementType());
auto count = type_info.GetElementCount();
ASSERT_EQ(expected_y.size(), count);
const float* values = Y_value.GetTensorData<float>();
ASSERT_TRUE(std::equal(values, values + count, std::begin(expected_y)));
}
binding.ClearBoundInputs();
binding.ClearBoundOutputs();
}
#if defined(USE_CUDA) || defined(USE_TENSORRT)
TEST(CApiTest, io_binding_cuda) {
struct CudaMemoryDeleter {
explicit CudaMemoryDeleter(const Ort::Allocator* alloc) {
alloc_ = alloc;
}
void operator()(void* ptr) const {
alloc_->Free(ptr);
}
const Ort::Allocator* alloc_;
};
Ort::SessionOptions session_options;
#ifdef USE_TENSORRT
Ort::ThrowOnError(OrtSessionOptionsAppendExecutionProvider_Tensorrt(session_options, 0));
#else
Ort::ThrowOnError(OrtSessionOptionsAppendExecutionProvider_CUDA(session_options, 0));
#endif
Ort::Session session(*ort_env, MODEL_URI, session_options);
Ort::MemoryInfo info_cuda("Cuda", OrtAllocatorType::OrtArenaAllocator, 0, OrtMemTypeDefault);
Ort::Allocator cuda_allocator(session, info_cuda);
auto allocator_info = cuda_allocator.GetInfo();
ASSERT_TRUE(info_cuda == allocator_info);
const std::array<int64_t, 2> x_shape = {3, 2};
std::array<float, 3 * 2> x_values = {1.0f, 2.0f, 3.0f, 4.0f, 5.0f, 6.0f};
auto input_data = std::unique_ptr<void, CudaMemoryDeleter>(cuda_allocator.Alloc(x_values.size() * sizeof(float)),
CudaMemoryDeleter(&cuda_allocator));
ASSERT_NE(input_data.get(), nullptr);
cudaMemcpy(input_data.get(), x_values.data(), sizeof(float) * x_values.size(), cudaMemcpyHostToDevice);
// Create an OrtValue tensor backed by data on CUDA memory
Ort::Value bound_x = Ort::Value::CreateTensor(info_cuda, reinterpret_cast<float*>(input_data.get()), x_values.size(),
x_shape.data(), x_shape.size());
const std::array<int64_t, 2> expected_y_shape = {3, 2};
const std::array<float, 3 * 2> expected_y = {1.0f, 4.0f, 9.0f, 16.0f, 25.0f, 36.0f};
auto output_data = std::unique_ptr<void, CudaMemoryDeleter>(cuda_allocator.Alloc(expected_y.size() * sizeof(float)),
CudaMemoryDeleter(&cuda_allocator));
ASSERT_NE(output_data.get(), nullptr);
// Create an OrtValue tensor backed by data on CUDA memory
Ort::Value bound_y = Ort::Value::CreateTensor(info_cuda, reinterpret_cast<float*>(output_data.get()),
expected_y.size(), expected_y_shape.data(), expected_y_shape.size());
// Sychronize to make sure the copy on default stream is done since TensorRT isn't using default stream.
cudaStreamSynchronize(nullptr);
Ort::IoBinding binding(session);
binding.BindInput("X", bound_x);
binding.BindOutput("Y", bound_y);
session.Run(Ort::RunOptions(), binding);
// Check the values against the bound raw memory (needs copying from device to host first)
std::array<float, 3 * 2> y_values_0;
cudaMemcpy(y_values_0.data(), output_data.get(), sizeof(float) * y_values_0.size(), cudaMemcpyDeviceToHost);
ASSERT_TRUE(std::equal(std::begin(y_values_0), std::end(y_values_0), std::begin(expected_y)));
// Now compare values via GetOutputValues
{
std::vector<Ort::Value> output_values = binding.GetOutputValues();
ASSERT_EQ(output_values.size(), 1U);
const Ort::Value& Y_value = output_values[0];
ASSERT_TRUE(Y_value.IsTensor());
Ort::TensorTypeAndShapeInfo type_info = Y_value.GetTensorTypeAndShapeInfo();
ASSERT_EQ(ONNX_TENSOR_ELEMENT_DATA_TYPE_FLOAT, type_info.GetElementType());
auto count = type_info.GetElementCount();
ASSERT_EQ(expected_y.size(), count);
const float* values = Y_value.GetTensorData<float>();
std::array<float, 3 * 2> y_values_1;
cudaMemcpy(y_values_1.data(), values, sizeof(float) * y_values_1.size(), cudaMemcpyDeviceToHost);
ASSERT_TRUE(std::equal(std::begin(y_values_1), std::end(y_values_1), std::begin(expected_y)));
}
{
std::vector<std::string> output_names = binding.GetOutputNames();
ASSERT_EQ(1U, output_names.size());
ASSERT_EQ(output_names[0].compare("Y"), 0);
}
// Now replace binding of Y with an on device binding instead of pre-allocated memory.
// This is when we can not allocate an OrtValue due to unknown dimensions
{
binding.BindOutput("Y", info_cuda);
session.Run(Ort::RunOptions(), binding);
}
// Check the output value allocated based on the device binding.
{
std::vector<Ort::Value> output_values = binding.GetOutputValues();
ASSERT_EQ(output_values.size(), 1U);
const Ort::Value& Y_value = output_values[0];
ASSERT_TRUE(Y_value.IsTensor());
Ort::TensorTypeAndShapeInfo type_info = Y_value.GetTensorTypeAndShapeInfo();
ASSERT_EQ(ONNX_TENSOR_ELEMENT_DATA_TYPE_FLOAT, type_info.GetElementType());
auto count = type_info.GetElementCount();
ASSERT_EQ(expected_y.size(), count);
const float* values = Y_value.GetTensorData<float>();
std::array<float, 3 * 2> y_values_2;
cudaMemcpy(y_values_2.data(), values, sizeof(float) * y_values_2.size(), cudaMemcpyDeviceToHost);
ASSERT_TRUE(std::equal(std::begin(y_values_2), std::end(y_values_2), std::begin(expected_y)));
}
// Clean up
binding.ClearBoundInputs();
binding.ClearBoundOutputs();
}
#endif
TEST(CApiTest, create_tensor) {
const char* s[] = {"abc", "kmp"};
int64_t expected_len = 2;
auto default_allocator = onnxruntime::make_unique<MockedOrtAllocator>();
Ort::Value tensor = Ort::Value::CreateTensor(default_allocator.get(), &expected_len, 1,
ONNX_TENSOR_ELEMENT_DATA_TYPE_STRING);
Ort::ThrowOnError(Ort::GetApi().FillStringTensor(tensor, s, expected_len));
auto shape_info = tensor.GetTensorTypeAndShapeInfo();
int64_t len = shape_info.GetElementCount();
ASSERT_EQ(len, expected_len);
std::vector<int64_t> shape_array(len);
size_t data_len = tensor.GetStringTensorDataLength();
std::string result(data_len, '\0');
std::vector<size_t> offsets(len);
tensor.GetStringTensorContent((void*)result.data(), data_len, offsets.data(), offsets.size());
}
TEST(CApiTest, fill_string_tensor) {
const char* s[] = {"abc", "kmp"};
int64_t expected_len = 2;
auto default_allocator = onnxruntime::make_unique<MockedOrtAllocator>();
Ort::Value tensor = Ort::Value::CreateTensor(default_allocator.get(), &expected_len, 1,
ONNX_TENSOR_ELEMENT_DATA_TYPE_STRING);
for (int64_t i = 0; i < expected_len; i++) {
tensor.FillStringTensorElement(s[i], i);
}
auto shape_info = tensor.GetTensorTypeAndShapeInfo();
int64_t len = shape_info.GetElementCount();
ASSERT_EQ(len, expected_len);
}
TEST(CApiTest, get_string_tensor_element) {
const char* s[] = {"abc", "kmp"};
int64_t expected_len = 2;
int64_t element_index = 0;
auto default_allocator = onnxruntime::make_unique<MockedOrtAllocator>();
Ort::Value tensor = Ort::Value::CreateTensor(default_allocator.get(), &expected_len, 1,
ONNX_TENSOR_ELEMENT_DATA_TYPE_STRING);
tensor.FillStringTensor(s, expected_len);
auto expected_string = s[element_index];
size_t expected_string_len = strlen(expected_string);
std::string result(expected_string_len, '\0');
tensor.GetStringTensorElement(expected_string_len, element_index, (void*)result.data());
ASSERT_STREQ(result.c_str(), expected_string);
auto string_len = tensor.GetStringTensorElementLength(element_index);
ASSERT_EQ(expected_string_len, string_len);
}
TEST(CApiTest, create_tensor_with_data) {
float values[] = {3.0f, 1.0f, 2.f, 0.f};
constexpr size_t values_length = sizeof(values) / sizeof(values[0]);
Ort::MemoryInfo info("Cpu", OrtDeviceAllocator, 0, OrtMemTypeDefault);
std::vector<int64_t> dims = {4};
Ort::Value tensor = Ort::Value::CreateTensor<float>(info, values, values_length, dims.data(), dims.size());
const float* new_pointer = tensor.GetTensorData<float>();
ASSERT_EQ(new_pointer, values);
auto type_info = tensor.GetTypeInfo();
auto tensor_info = type_info.GetTensorTypeAndShapeInfo();
ASSERT_NE(tensor_info, nullptr);
ASSERT_EQ(1u, tensor_info.GetDimensionsCount());
}
TEST(CApiTest, create_tensor_with_data_float16) {
// Example with C++. However, what we are feeding underneath is really
// a continuous buffer of uint16_t
// Use 3rd party libraries such as Eigen to convert floats and doubles to float16 types.
Ort::Float16_t values[] = {15360, 16384, 16896, 17408, 17664}; // 1.f, 2.f, 3.f, 4.f, 5.f
constexpr size_t values_length = sizeof(values) / sizeof(values[0]);
std::vector<int64_t> dims = {static_cast<int64_t>(values_length)};
Ort::MemoryInfo info("Cpu", OrtDeviceAllocator, 0, OrtMemTypeDefault);
Ort::Value tensor = Ort::Value::CreateTensor<Ort::Float16_t>(info, values, values_length, dims.data(), dims.size());
const auto* new_pointer = tensor.GetTensorData<Ort::Float16_t>();
ASSERT_EQ(new_pointer, values);
auto type_info = tensor.GetTypeInfo();
auto tensor_info = type_info.GetTensorTypeAndShapeInfo();