Repository navigation
Expand file tree
/
Copy pathpyCore.cpp
More file actions
1949 lines (1737 loc) · 92.7 KB
/
Copy pathpyCore.cpp
File metadata and controls
1949 lines (1737 loc) · 92.7 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
/*
* SPDX-FileCopyrightText: Copyright (c) 1993-2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
* SPDX-License-Identifier: Apache-2.0
*
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
*/
// This contains the core elements of the API, i.e. builder, logger, engine, runtime, context.
#include "ForwardDeclarations.h"
#include "utils.h"
#include <chrono>
#include <cstdint>
#include <iomanip>
#include <pybind11/stl.h>
#include "infer/pyCoreDoc.h"
#include <cuda.h>
#include <cuda_runtime_api.h>
namespace tensorrt
{
using namespace nvinfer1;
// Long lambda functions should go here rather than being inlined into the bindings (1 liners are OK).
namespace lambdas
{
// For IOptimizationProfile
static auto const opt_profile_set_shape
= [](IOptimizationProfile& self, std::string const& inputName, Dims const& min, Dims const& opt, Dims const& max) {
PY_ASSERT_RUNTIME_ERROR(self.setDimensions(inputName.c_str(), OptProfileSelector::kMIN, min),
"Shape provided for min is inconsistent with other shapes.");
PY_ASSERT_RUNTIME_ERROR(self.setDimensions(inputName.c_str(), OptProfileSelector::kOPT, opt),
"Shape provided for opt is inconsistent with other shapes.");
PY_ASSERT_RUNTIME_ERROR(self.setDimensions(inputName.c_str(), OptProfileSelector::kMAX, max),
"Shape provided for max is inconsistent with other shapes.");
};
static auto const opt_profile_get_shape
= [](IOptimizationProfile& self, std::string const& inputName) -> std::vector<Dims> {
std::vector<Dims> shapes{};
Dims minShape = self.getDimensions(inputName.c_str(), OptProfileSelector::kMIN);
if (minShape.nbDims != -1)
{
shapes.emplace_back(minShape);
shapes.emplace_back(self.getDimensions(inputName.c_str(), OptProfileSelector::kOPT));
shapes.emplace_back(self.getDimensions(inputName.c_str(), OptProfileSelector::kMAX));
}
return shapes;
};
static auto const opt_profile_set_shape_input = [](IOptimizationProfile& self, std::string const& inputName,
std::vector<int64_t> const& min, std::vector<int64_t> const& opt,
std::vector<int64_t> const& max) {
PY_ASSERT_RUNTIME_ERROR(self.setShapeValuesV2(inputName.c_str(), OptProfileSelector::kMIN, min.data(), min.size()),
"min input provided for shape tensor is inconsistent with other inputs.");
PY_ASSERT_RUNTIME_ERROR(self.setShapeValuesV2(inputName.c_str(), OptProfileSelector::kOPT, opt.data(), opt.size()),
"opt input provided for shape tensor is inconsistent with other inputs.");
PY_ASSERT_RUNTIME_ERROR(self.setShapeValuesV2(inputName.c_str(), OptProfileSelector::kMAX, max.data(), max.size()),
"max input provided for shape tensor is inconsistent with other inputs.");
};
static auto const opt_profile_get_shape_input
= [](IOptimizationProfile& self, std::string const& inputName) -> std::vector<std::vector<int64_t>> {
std::vector<std::vector<int64_t>> shapes{};
int32_t const shapeSize = self.getNbShapeValues(inputName.c_str());
int64_t const* shapePtr = self.getShapeValuesV2(inputName.c_str(), OptProfileSelector::kMIN);
// In the Python bindings, it is impossible to set only one shape in an optimization profile.
if (shapePtr && shapeSize >= 0)
{
shapes.emplace_back(shapePtr, shapePtr + shapeSize);
shapePtr = self.getShapeValuesV2(inputName.c_str(), OptProfileSelector::kOPT);
PY_ASSERT_RUNTIME_ERROR(shapePtr != nullptr, "Invalid shape for OPT.");
shapes.emplace_back(shapePtr, shapePtr + shapeSize);
shapePtr = self.getShapeValuesV2(inputName.c_str(), OptProfileSelector::kMAX);
PY_ASSERT_RUNTIME_ERROR(shapePtr != nullptr, "Invalid shape for MAX.");
shapes.emplace_back(shapePtr, shapePtr + shapeSize);
}
return shapes;
};
namespace
{
//! Return an optimization profile's CUDA stream as a Python-compatible handle.
size_t optProfileGetProfileStream(IOptimizationProfile const& self)
{
return reinterpret_cast<size_t>(self.getProfileStream());
}
//! Set an optimization profile's CUDA stream from a Python stream handle.
void optProfileSetProfileStream(IOptimizationProfile& self, size_t streamHandle)
{
self.setProfileStream(reinterpret_cast<cudaStream_t>(streamHandle));
}
} // namespace
// For IExecutionContext
static auto const execute_v2 = [](IExecutionContext& self, std::vector<size_t>& bindings) {
return self.executeV2(reinterpret_cast<void**>(bindings.data()));
};
std::vector<char const*> infer_shapes(IExecutionContext& self)
{
int32_t const size{self.getEngine().getNbIOTensors()};
std::vector<char const*> names(size);
int32_t const nbNames = self.inferShapes(names.size(), names.data());
if (nbNames < 0)
{
std::stringstream msg;
msg << "infer_shapes error code: " << nbNames;
py::gil_scoped_acquire gil{};
utils::throwPyError(PyExc_RuntimeError, msg.str().c_str());
}
names.resize(nbNames);
return names;
}
bool execute_async_v3(IExecutionContext& self, size_t streamHandle)
{
return self.enqueueV3(reinterpret_cast<cudaStream_t>(streamHandle));
}
bool set_tensor_address(IExecutionContext& self, char const* tensor_name, size_t memory)
{
return self.setTensorAddress(tensor_name, reinterpret_cast<void*>(memory));
}
size_t get_tensor_address(IExecutionContext& self, char const* tensor_name)
{
return reinterpret_cast<size_t>(self.getTensorAddress(tensor_name));
}
bool set_input_consumed_event(IExecutionContext& self, size_t inputConsumed)
{
return self.setInputConsumedEvent(reinterpret_cast<cudaEvent_t>(inputConsumed));
}
size_t get_input_consumed_event(IExecutionContext& self)
{
return reinterpret_cast<size_t>(self.getInputConsumedEvent());
}
void set_aux_streams(IExecutionContext& self, std::vector<size_t> streamHandle)
{
self.setAuxStreams(reinterpret_cast<cudaStream_t*>(streamHandle.data()), static_cast<int32_t>(streamHandle.size()));
}
template <typename PyIterable>
Dims castDimsFromPyIterable(PyIterable& in)
{
int32_t const maxDims{static_cast<int32_t>(Dims::MAX_DIMS)};
Dims dims{};
dims.nbDims = py::len(in);
PY_ASSERT_RUNTIME_ERROR(
dims.nbDims <= maxDims, "The number of input dims exceeds the maximum allowed number of dimensions");
for (int32_t i = 0; i < dims.nbDims; ++i)
{
dims.d[i] = in[i].template cast<int32_t>();
}
return dims;
}
template <typename PyIterable>
bool setInputShape(IExecutionContext& self, char const* tensorName, PyIterable& in)
{
return self.setInputShape(tensorName, castDimsFromPyIterable<PyIterable>(in));
}
// For IRuntime
static auto const runtime_deserialize_cuda_engine = [](IRuntime& self, py::buffer& serializedEngine) {
py::buffer_info info = serializedEngine.request();
py::gil_scoped_release releaseGil{};
return self.deserializeCudaEngine(info.ptr, info.size * info.itemsize);
};
static auto const runtime_set_dla_workspace_allocation_strategy
= [](IRuntime& self, DLAWorkspaceAllocationStrategy strategy) {
PY_ASSERT_RUNTIME_ERROR(
self.setDLAWorkspaceAllocationStrategy(strategy), "Failed to set the DLA workspace allocation strategy.");
};
static auto const reader_v2_read = [](IStreamReaderV2& self, void* destination, int64_t nbBytes, size_t stream) {
return self.read(destination, nbBytes, reinterpret_cast<cudaStream_t>(stream));
};
// For ICudaEngine
// TODO: Add slicing support?
static auto const engine_getitem = [](ICudaEngine& self, int32_t pyIndex) {
// Support python's negative indexing
int32_t const index = (pyIndex < 0) ? static_cast<int32_t>(self.getNbIOTensors()) + pyIndex : pyIndex;
PY_ASSERT_INDEX_ERROR(index < self.getNbIOTensors());
return self.getIOTensorName(index);
};
std::vector<Dims> get_tensor_profile_shape(ICudaEngine& self, std::string const& tensorName, int32_t profileIndex)
{
std::string const errorMsg{"Could not get profile shape for tensor '" + tensorName
+ "'. Is the tensor name an input and the profile index valid?"};
std::vector<Dims> shapes{};
shapes.emplace_back(
utils::checkDims(self.getProfileShape(tensorName.c_str(), profileIndex, OptProfileSelector::kMIN), errorMsg));
shapes.emplace_back(
utils::checkDims(self.getProfileShape(tensorName.c_str(), profileIndex, OptProfileSelector::kOPT), errorMsg));
shapes.emplace_back(
utils::checkDims(self.getProfileShape(tensorName.c_str(), profileIndex, OptProfileSelector::kMAX), errorMsg));
return shapes;
}
std::vector<std::vector<int64_t>> get_tensor_profile_values(
ICudaEngine& self, int32_t profileIndex, std::string const& tensorName)
{
char const* const name = tensorName.c_str();
bool const isShapeInput{self.isShapeInferenceIO(name) && self.getTensorIOMode(name) == TensorIOMode::kINPUT};
PY_ASSERT_RUNTIME_ERROR(isShapeInput, "Binding index does not correspond to an input shape tensor.");
Dims const shape = self.getTensorShape(name);
PY_ASSERT_RUNTIME_ERROR(shape.nbDims >= 0, "Missing shape for input shape tensor");
auto const shapeSize{utils::volume(shape)};
PY_ASSERT_RUNTIME_ERROR(shapeSize >= 0, "Negative volume for input shape tensor");
std::vector<std::vector<int64_t>> shapes{};
// In the Python bindings, it is impossible to set only one shape in an optimization profile.
int64_t const* shapePtr{self.getProfileTensorValuesV2(name, profileIndex, OptProfileSelector::kMIN)};
if (shapePtr)
{
shapes.emplace_back(shapePtr, shapePtr + shapeSize);
shapePtr = self.getProfileTensorValuesV2(name, profileIndex, OptProfileSelector::kOPT);
shapes.emplace_back(shapePtr, shapePtr + shapeSize);
shapePtr = self.getProfileTensorValuesV2(name, profileIndex, OptProfileSelector::kMAX);
shapes.emplace_back(shapePtr, shapePtr + shapeSize);
}
return shapes;
}
// For IGpuAllocator
void* allocate_async(
IGpuAllocator& self, uint64_t const size, uint64_t const alignment, AllocatorFlags const flags, size_t streamHandle)
{
return self.allocateAsync(size, alignment, flags, reinterpret_cast<cudaStream_t>(streamHandle));
}
bool deallocate_async(IGpuAllocator& self, void* const memory, size_t streamHandle)
{
return self.deallocateAsync(memory, reinterpret_cast<cudaStream_t>(streamHandle));
}
// For IOutputAllocator
void* reallocate_output_async(IOutputAllocator& self, char const* tensorName, void* currentMemory, uint64_t size,
uint64_t alignment, size_t streamHandle)
{
return self.reallocateOutputAsync(
tensorName, currentMemory, size, alignment, reinterpret_cast<cudaStream_t>(streamHandle));
}
#if EXPORT_ALL_BINDINGS
// For IBuilderConfig
static auto const netconfig_get_profile_stream
= [](IBuilderConfig& self) -> size_t { return reinterpret_cast<size_t>(self.getProfileStream()); };
static auto const netconfig_set_profile_stream = [](IBuilderConfig& self, size_t streamHandle) {
self.setProfileStream(reinterpret_cast<cudaStream_t>(streamHandle));
};
static auto const netconfig_create_timing_cache = [](IBuilderConfig& self, py::buffer& serializedTimingCache) {
py::buffer_info info = serializedTimingCache.request();
py::gil_scoped_release releaseGil{};
return self.createTimingCache(info.ptr, info.size * info.itemsize);
};
static auto const get_plugins_to_serialize = [](IBuilderConfig& self) {
std::vector<std::string> paths;
int64_t const nbPlugins = self.getNbPluginsToSerialize();
if (nbPlugins < 0)
{
utils::throwPyError(PyExc_RuntimeError, "Internal error");
}
paths.reserve(nbPlugins);
for (int64_t i = 0; i < nbPlugins; ++i)
{
paths.emplace_back(std::string{self.getPluginToSerialize(i)});
}
return paths;
};
static auto const set_plugins_to_serialize = [](IBuilderConfig& self, std::vector<std::string> const& paths) {
std::vector<char const*> cStrings;
cStrings.reserve(paths.size());
for (auto const& path : paths)
{
cStrings.push_back(path.c_str());
}
self.setPluginsToSerialize(reinterpret_cast<char const* const*>(cStrings.data()), cStrings.size());
};
static auto const get_remote_auto_tuning_config
= [](IBuilderConfig& self) { return std::string{self.getRemoteAutoTuningConfig()}; };
static auto const set_remote_auto_tuning_config
= [](IBuilderConfig& self, std::string const& config) { self.setRemoteAutoTuningConfig(config.c_str()); };
static auto const get_build_route = [](IBuilderConfig& self) { return std::string{self.getBuildRoute()}; };
static auto const set_build_route
= [](IBuilderConfig& self, std::string const& buildRoute) { self.setBuildRoute(buildRoute.c_str()); };
static auto const get_all_build_routes = [](IBuilderConfig& self) { return std::string{self.getAllBuildRoutes()}; };
#endif // EXPORT_ALL_BINDINGS
// For IRefitter
static auto const refitter_get_missing = [](IRefitter& self) {
// First get the number of missing weights.
int32_t const size{self.getMissing(0, nullptr, nullptr)};
// Now that we know how many weights are missing, we can create the buffers appropriately.
std::vector<const char*> layerNames(size);
std::vector<WeightsRole> roles(size);
self.getMissing(size, layerNames.data(), roles.data());
return std::pair<std::vector<const char*>, std::vector<WeightsRole>>{layerNames, roles};
};
static auto const refitter_get_missing_weights = [](IRefitter& self) {
// First get the number of missing weights.
int32_t const size{self.getMissingWeights(0, nullptr)};
// Now that we know how many weights are missing, we can create the buffers appropriately.
std::vector<char const*> names(size);
self.getMissingWeights(size, names.data());
return names;
};
static auto const refitter_get_all = [](IRefitter& self) {
int32_t const size{self.getAll(0, nullptr, nullptr)};
std::vector<char const*> layerNames(size);
std::vector<WeightsRole> roles(size);
self.getAll(size, layerNames.data(), roles.data());
return std::pair<std::vector<const char*>, std::vector<WeightsRole>>{layerNames, roles};
};
static auto const refitter_get_all_weights = [](IRefitter& self) {
int32_t const size{self.getAllWeights(0, nullptr)};
std::vector<char const*> names(size);
self.getAllWeights(size, names.data());
return names;
};
static auto const refitter_refit_cuda_engine_async = [](IRefitter& self, size_t streamHandle) {
return self.refitCudaEngineAsync(reinterpret_cast<cudaStream_t>(streamHandle));
};
static auto const context_set_optimization_profile_async
= [](IExecutionContext& self, int32_t const profileIndex, size_t streamHandle) {
PY_ASSERT_RUNTIME_ERROR(
self.setOptimizationProfileAsync(profileIndex, reinterpret_cast<cudaStream_t>(streamHandle)),
"Error in set optimization profile async.");
return true;
};
void context_set_device_memory(IExecutionContext& self, size_t memory)
{
self.setDeviceMemory(reinterpret_cast<void*>(memory));
}
void context_set_device_memory_v2(IExecutionContext& self, size_t memory, int64_t size)
{
self.setDeviceMemoryV2(reinterpret_cast<void*>(memory), size);
}
void serialization_config_set_flags(ISerializationConfig& self, uint32_t flags)
{
if (!self.setFlags(flags))
{
utils::throwPyError(PyExc_RuntimeError, "Provided serialization flags is incorrect");
}
}
// For IDebugListener, this function is intended to be override by client.
// The bindings here will never be called and is for documentation purpose only.
void docProcessDebugTensor(IDebugListener& self, void const* addr, TensorLocation location, DataType type,
Dims const& shape, char const* name, size_t stream)
{
return;
}
uint64_t getTacticHash(TimingCacheValue const& value)
{
return value.tacticHash;
}
void setTacticHash(TimingCacheValue& value, uint64_t tacticHash)
{
value.tacticHash = tacticHash;
}
float getTimingMSec(TimingCacheValue const& value)
{
return value.timingMSec;
}
void setTimingMSec(TimingCacheValue& value, float timingMSec)
{
value.timingMSec = timingMSec;
}
namespace detail
{
constexpr int64_t kBYTES_PER_KEY = 16;
constexpr int64_t kCHARS_PER_BYTE = 2;
constexpr int64_t kPREFIX_CHARS = 2;
constexpr int64_t kTOTAL_CHARS = kPREFIX_CHARS + kBYTES_PER_KEY * kCHARS_PER_BYTE;
} // namespace detail
TimingCacheKey parseTimingCacheKey(std::string const& text)
{
using namespace detail;
if (text.size() != kTOTAL_CHARS)
{
std::ostringstream msg;
msg << "The text should have exactly " << kTOTAL_CHARS << " characters.";
utils::throwPyError(PyExc_ValueError, msg.str().c_str());
}
int offset = 0;
sscanf(text.c_str(), "0%*[xX]%n", &offset);
PY_ASSERT_VALUE_ERROR(offset == 2, "The text should start with prefix `0x` or `0X`.");
TimingCacheKey key;
for (int64_t i = 0; i < kBYTES_PER_KEY; ++i, offset += kCHARS_PER_BYTE)
{
int64_t numReceived = sscanf(text.c_str() + offset, "%2" SCNx8, &key.data[i]);
PY_ASSERT_VALUE_ERROR(numReceived == 1, "The text has invalid content.");
}
return key;
}
std::string convertTimingCacheKeyToString(TimingCacheKey const& key)
{
using namespace detail;
char buffer[kTOTAL_CHARS + 1] = "0x";
for (int64_t i = 0; i < kBYTES_PER_KEY; ++i)
{
int64_t offset = kPREFIX_CHARS + kCHARS_PER_BYTE * i;
sprintf(buffer + offset, "%02" PRIx8, key.data[i]);
}
return std::string(buffer);
}
std::vector<TimingCacheKey> queryTimingCacheKeys(ITimingCache const& cache)
{
int64_t numKeys = cache.queryKeys(nullptr, 0);
PY_ASSERT_RUNTIME_ERROR(numKeys >= 0, "Failed to get the number of keys in the timing cache");
std::vector<TimingCacheKey> keys(numKeys);
PY_ASSERT_RUNTIME_ERROR(
numKeys == cache.queryKeys(keys.data(), keys.size()), "Failed to get keys from the timing cache");
return keys;
}
} // namespace lambdas
namespace PyGpuAllocatorHelper
{
template <typename TAllocator, typename... Args>
void* allocHelper(TAllocator* allocator, char const* pyFuncName, bool showWarning, Args&&... args) noexcept
{
try
{
py::gil_scoped_acquire gil{};
py::function pyAllocFunc = utils::getOverride(static_cast<TAllocator*>(allocator), pyFuncName, showWarning);
if (!pyAllocFunc)
{
return nullptr;
}
py::object ptr = pyAllocFunc(std::forward<Args>(args)...);
try
{
return reinterpret_cast<void*>(ptr.cast<size_t>());
}
catch (py::cast_error const& e)
{
std::cerr << "[ERROR] Return value of allocate() could not be interpreted as an int" << std::endl;
}
}
catch (std::exception const& e)
{
std::cerr << "[ERROR] Exception caught in allocate(): " << e.what() << std::endl;
}
catch (...)
{
std::cerr << "[ERROR] Exception caught in allocate()" << std::endl;
return nullptr;
}
return nullptr;
}
} // namespace PyGpuAllocatorHelper
class PyGpuAllocator : public IGpuAllocator
{
public:
using IGpuAllocator::IGpuAllocator;
void* allocate(uint64_t size, uint64_t alignment, AllocatorFlags flags) noexcept override
{
return PyGpuAllocatorHelper::allocHelper<IGpuAllocator>(this, "allocate", true, size, alignment, flags);
}
void* reallocate(void* baseAddr, uint64_t alignment, uint64_t newSize) noexcept override
{
return PyGpuAllocatorHelper::allocHelper<IGpuAllocator>(
this, "reallocate", true, reinterpret_cast<size_t>(baseAddr), alignment, newSize);
}
bool deallocate(void* memory) noexcept override
{
try
{
py::gil_scoped_acquire gil{};
py::function pyDeallocate = utils::getOverride(static_cast<IGpuAllocator*>(this), "deallocate");
if (!pyDeallocate)
{
return false;
}
py::object status{};
status = pyDeallocate(reinterpret_cast<size_t>(memory));
return status.cast<bool>();
}
catch (std::exception const& e)
{
std::cerr << "[ERROR] Exception caught in deallocate(): " << e.what() << std::endl;
}
catch (...)
{
std::cerr << "[ERROR] Exception caught in deallocate()" << std::endl;
}
return false;
}
}; // PyGpuAllocator
///////////
class PyGpuAsyncAllocator : public IGpuAsyncAllocator
{
public:
using IGpuAsyncAllocator::IGpuAsyncAllocator;
void* allocateAsync(uint64_t size, uint64_t alignment, AllocatorFlags flags, cudaStream_t stream) noexcept override
{
intptr_t cudaStreamPtr = reinterpret_cast<intptr_t>(stream);
return PyGpuAllocatorHelper::allocHelper<IGpuAsyncAllocator>(
this, "allocate_async", true, size, alignment, flags, cudaStreamPtr);
}
void* reallocate(void* baseAddr, uint64_t alignment, uint64_t newSize) noexcept override
{
return PyGpuAllocatorHelper::allocHelper<IGpuAsyncAllocator>(
this, "reallocate", true, reinterpret_cast<size_t>(baseAddr), alignment, newSize);
}
bool deallocateAsync(void* memory, cudaStream_t stream) noexcept override
{
try
{
py::gil_scoped_acquire gil{};
py::function pyDeallocateAsync
= utils::getOverride(static_cast<IGpuAsyncAllocator*>(this), "deallocate_async");
if (!pyDeallocateAsync)
{
return false;
}
py::object status{};
intptr_t cudaStreamPtr = reinterpret_cast<intptr_t>(stream);
status = pyDeallocateAsync(reinterpret_cast<size_t>(memory), cudaStreamPtr);
return status.cast<bool>();
}
catch (std::exception const& e)
{
std::cerr << "[ERROR] Exception caught in deallocate(): " << e.what() << std::endl;
}
catch (...)
{
std::cerr << "[ERROR] Exception caught in deallocate()" << std::endl;
}
return false;
}
}; // PyGpuAsyncAllocator
/////////////////////////////
class PyOutputAllocator : public IOutputAllocator
{
public:
void* reallocateOutput(
char const* tensorName, void* currentMemory, uint64_t size, uint64_t alignment) noexcept override
{
try
{
py::gil_scoped_acquire gil{};
py::function pyFunc = utils::getOverride(static_cast<IOutputAllocator*>(this), "reallocate_output");
if (!pyFunc)
{
return nullptr;
}
py::object ptr = pyFunc(tensorName, reinterpret_cast<size_t>(currentMemory), size, alignment);
try
{
return reinterpret_cast<void*>(ptr.cast<size_t>());
}
catch (py::cast_error const& e)
{
std::cerr << "[ERROR] Return value of reallocateOutput() could not be interpreted as an int"
<< std::endl;
}
}
catch (std::exception const& e)
{
std::cerr << "[ERROR] Exception caught in reallocateOutput(): " << e.what() << std::endl;
}
catch (...)
{
std::cerr << "[ERROR] Exception caught in reallocateOutput()" << std::endl;
return nullptr;
}
return nullptr;
}
void* reallocateOutputAsync(char const* tensorName, void* currentMemory, uint64_t size, uint64_t alignment,
cudaStream_t stream) noexcept override
{
try
{
py::gil_scoped_acquire gil{};
py::function pyFunc
= utils::getOverride(static_cast<IOutputAllocator*>(this), "reallocate_output_async", false);
if (!pyFunc)
{
//! For legacy implementation, the user might not have implemented this method, so we go for the default
//! method.
return reallocateOutput(tensorName, currentMemory, size, alignment);
}
intptr_t cudaStreamPtr = reinterpret_cast<intptr_t>(stream);
py::object ptr
= pyFunc(tensorName, reinterpret_cast<size_t>(currentMemory), size, alignment, cudaStreamPtr);
try
{
return reinterpret_cast<void*>(ptr.cast<size_t>());
}
catch (py::cast_error const& e)
{
std::cerr << "[ERROR] Return value of reallocateOutputAsync() could not be interpreted as an int"
<< std::endl;
}
}
catch (std::exception const& e)
{
std::cerr << "[ERROR] Exception caught in reallocateOutputAsync(): " << e.what() << std::endl;
}
catch (...)
{
std::cerr << "[ERROR] Exception caught in reallocateOutputAsync()" << std::endl;
return nullptr;
}
return nullptr;
}
void notifyShape(char const* tensorName, Dims const& dims) noexcept override
{
try
{
py::gil_scoped_acquire gil{};
PYBIND11_OVERLOAD_PURE_NAME(void, IOutputAllocator, "notify_shape", notifyShape, tensorName, dims);
}
catch (std::exception const& e)
{
std::cerr << "[ERROR] Exception caught in notifyShape(): " << e.what() << std::endl;
}
catch (...)
{
std::cerr << "[ERROR] Exception caught in notifyShape()" << std::endl;
}
}
};
class PyStreamReaderV2 : public IStreamReaderV2
{
using TFnPointerGetAttribute = CUresult (*)(void*, CUpointer_attribute, CUdeviceptr);
using TFnMemcpyHtoD = CUresult (*)(CUdeviceptr, void const*, size_t);
public:
PyStreamReaderV2()
{
py::gil_scoped_acquire gil{};
mCudaHandle = utils::nvdllOpen(CUDA_LIB_NAME);
if (!mCudaHandle)
{
utils::throwPyError(PyExc_RuntimeError, "[ERROR] Failed to open cuda driver.");
}
mFnPointerGetAttribute
= reinterpret_cast<TFnPointerGetAttribute>(utils::dllGetSym(mCudaHandle, "cuPointerGetAttribute"));
mFnMemcpyHtoD = reinterpret_cast<TFnMemcpyHtoD>(utils::dllGetSym(mCudaHandle, "cuMemcpyHtoD_v2"));
}
~PyStreamReaderV2()
{
try
{
py::gil_scoped_acquire gil{};
utils::dllClose(mCudaHandle);
}
catch (...)
{
std::cerr << "[ERROR] An exception occurred while closing the CUDA driver.";
}
}
int64_t read(void* destination, int64_t nbBytes, cudaStream_t stream) noexcept override
{
try
{
py::gil_scoped_acquire gil{};
if (!mFnPointerGetAttribute || !mFnMemcpyHtoD)
{
utils::throwPyError(PyExc_RuntimeError,
"[ERROR] Read is skipped due to failed to get necessary API entry in cuda driver.");
return 0;
}
py::function pyReadFunc = utils::getOverride(static_cast<IStreamReaderV2*>(this), "read");
if (!pyReadFunc)
{
utils::throwPyError(PyExc_RuntimeError, "[ERROR] Failed to find override read function in python.");
return 0;
}
// Check destination memory location.
uint32_t attributes{};
CUresult ret = mFnPointerGetAttribute(
&attributes, CU_POINTER_ATTRIBUTE_MEMORY_TYPE, reinterpret_cast<CUdeviceptr>(destination));
if (ret == CUDA_ERROR_INVALID_VALUE)
{
attributes = CU_MEMORYTYPE_HOST;
}
else
{
CUDA_CALL_WITH_RET(ret, 0);
}
bool const useH2DCopy = attributes == CU_MEMORYTYPE_DEVICE;
auto copyDestination = static_cast<std::byte*>(destination);
auto cudaStreamPtr = reinterpret_cast<intptr_t>(stream);
// In C++ we can use GDS to reduce host memory usage. For Python, we handle this in the bindings by copying
// data chunk by chunk.
int64_t totalBytesRead{};
while (totalBytesRead < nbBytes)
{
int64_t bytesToRead = nbBytes - totalBytesRead;
py::buffer data = pyReadFunc(bytesToRead, cudaStreamPtr);
py::buffer_info info = data.request();
// User might chunk the memory into pieces to save peak host memory usage.
int64_t bytesRead = std::min(info.size * info.itemsize, bytesToRead);
if (bytesRead == 0)
{
std::cerr
<< "[ERROR] User aborted the operation, the read function in streamReaderV2 returned 0 bytes.";
break;
}
if (useH2DCopy)
{
CUDA_CALL_WITH_RET(mFnMemcpyHtoD(reinterpret_cast<CUdeviceptr>(copyDestination + totalBytesRead),
info.ptr, bytesRead),
totalBytesRead);
}
else
{
std::memcpy(copyDestination + totalBytesRead, info.ptr, bytesRead);
}
totalBytesRead += bytesRead;
}
return totalBytesRead;
}
catch (std::exception const& e)
{
std::cerr << "[ERROR] Exception caught in read(): " << e.what() << std::endl;
}
catch (...)
{
std::cerr << "[ERROR] Exception caught in read()" << std::endl;
}
return 0;
}
bool seek(int64_t offset, SeekPosition where) noexcept override
{
try
{
py::gil_scoped_acquire gil{};
py::function pySeekFunc = utils::getOverride(static_cast<IStreamReaderV2*>(this), "seek");
if (!pySeekFunc)
{
std::cerr << "[ERROR] Failed to find override seek function in python." << std::endl;
return 0;
}
py::bool_ ret = pySeekFunc(offset, where);
return ret;
}
catch (std::exception const& e)
{
std::cerr << "[ERROR] Exception caught in seek(): " << e.what() << std::endl;
}
catch (...)
{
std::cerr << "[ERROR] Exception caught in seek()" << std::endl;
}
return false;
}
private:
void* mCudaHandle{};
TFnPointerGetAttribute mFnPointerGetAttribute{};
TFnMemcpyHtoD mFnMemcpyHtoD{};
};
class PyStreamWriter : public IStreamWriter
{
public:
int64_t write(void const* data, int64_t size) noexcept override
{
try
{
py::gil_scoped_acquire gil{};
py::function pyFunc = utils::getOverride(static_cast<IStreamWriter*>(this), "write");
if (!pyFunc)
{
return 0;
}
auto const pyBytes = py::bytes(static_cast<char const*>(data), size);
py::object bytesWritten = pyFunc(pyBytes);
if (!py::isinstance<py::int_>(bytesWritten))
{
std::cerr << "[ERROR] StreamWriter shall returns the written bytes count in integer." << std::endl;
return 0;
}
return bytesWritten.cast<int64_t>();
}
catch (std::exception const& e)
{
std::cerr << "[ERROR] Exception caught in write(): " << e.what() << std::endl;
}
catch (...)
{
std::cerr << "[ERROR] Exception caught in write()" << std::endl;
}
return 0;
}
};
class PyDebugListener : public IDebugListener
{
public:
bool processDebugTensor(void const* addr, TensorLocation location, DataType type, Dims const& shape,
char const* name, cudaStream_t stream) override
{
try
{
py::gil_scoped_acquire gil{};
py::function pyFunc = utils::getOverride(static_cast<IDebugListener*>(this), "process_debug_tensor");
if (!pyFunc)
{
return false;
}
pyFunc(reinterpret_cast<size_t>(addr), location, type, shape, name, reinterpret_cast<size_t>(stream));
}
catch (std::exception const& e)
{
std::cerr << "[ERROR] Exception caught in processDebugTensor(): " << e.what() << std::endl;
}
catch (...)
{
std::cerr << "[ERROR] Exception caught in processDebugTensor()" << std::endl;
}
return true;
}
};
// NOLINTNEXTLINE(readability-function-cognitive-complexity)
void bindCore(py::module& m)
{
class PyLogger : public ILogger
{
public:
virtual void log(Severity severity, char const* msg) noexcept override
{
try
{
py::gil_scoped_acquire gil{};
PYBIND11_OVERLOAD_PURE_NAME(void, ILogger, "log", log, severity, msg);
}
catch (std::exception const& e)
{
std::cerr << "[ERROR] Exception caught in log(): " << e.what() << std::endl;
}
catch (...)
{
std::cerr << "[ERROR] Exception caught in log()" << std::endl;
}
}
};
py::class_<ILogger, PyLogger> baseLoggerBinding{m, "ILogger", ILoggerDoc::descr, py::module_local()};
py::enum_<ILogger::Severity>(
baseLoggerBinding, "Severity", py::arithmetic(), SeverityDoc::descr, py::module_local())
.value("INTERNAL_ERROR", ILogger::Severity::kINTERNAL_ERROR, SeverityDoc::internal_error)
.value("ERROR", ILogger::Severity::kERROR, SeverityDoc::error)
.value("WARNING", ILogger::Severity::kWARNING, SeverityDoc::warning)
.value("INFO", ILogger::Severity::kINFO, SeverityDoc::info)
.value("VERBOSE", ILogger::Severity::kVERBOSE, SeverityDoc::verbose)
// We export into the outer scope, so we can access with trt.ILogger.X.
.export_values();
baseLoggerBinding.def(py::init<>()).def("log", &ILogger::log, "severity"_a, "msg"_a, ILoggerDoc::log);
class DefaultLogger : public ILogger
{
public:
DefaultLogger(Severity minSeverity = Severity::kWARNING)
: mMinSeverity(minSeverity)
{
}
virtual void log(Severity severity, char const* msg) noexcept override
{
// INFO is the largest value, so this comparison is inverted.
if (severity > mMinSeverity)
return;
// prepend timestamp
std::time_t timestamp = std::time(nullptr);
tm* tm_local = std::localtime(×tamp);
std::cout << "[";
std::cout << std::setw(2) << std::setfill('0') << 1 + tm_local->tm_mon << "/";
std::cout << std::setw(2) << std::setfill('0') << tm_local->tm_mday << "/";
std::cout << std::setw(4) << std::setfill('0') << 1900 + tm_local->tm_year << "-";
std::cout << std::setw(2) << std::setfill('0') << tm_local->tm_hour << ":";
std::cout << std::setw(2) << std::setfill('0') << tm_local->tm_min << ":";
std::cout << std::setw(2) << std::setfill('0') << tm_local->tm_sec << "] ";
std::string loggingPrefix = "[TRT] ";
switch (severity)
{
case Severity::kINTERNAL_ERROR:
{
loggingPrefix += "[F] ";
break;
}
case Severity::kERROR: