Skip to content

Commit 9dad9af

Browse files
kevinch-nvkevinch-nv
authored
[TRT EP] Update build and API usage for TensorRT 10.11 (#24832)
### Description Resolves the following issues starting in TensorRT 10.11: - Version macros changed in `NvInferVersion.h`, update build to look for new macros - Updated deprecated APIs (setShapeValues -> setShapeValuesV2() to support INT64 shape values) ### Motivation and Context - Resolves building TensorRT EP from source with latest 10.11 release. Signed-off-by: Kevin Chen <kevinch@nvidia.com>
1 parent 2bdb57b commit 9dad9af

File tree

2 files changed

+51
-7
lines changed

2 files changed

+51
-7
lines changed

‎cmake/onnxruntime_providers_tensorrt.cmake‎

Lines changed: 21 additions & 6 deletions
Original file line numberDiff line numberDiff line change
@@ -33,12 +33,27 @@
3333
PATH_SUFFIXES include)
3434

3535
file(READ ${TENSORRT_INCLUDE_DIR}/NvInferVersion.h NVINFER_VER_CONTENT)
36-
string(REGEX MATCH "define NV_TENSORRT_MAJOR * +([0-9]+)" NV_TENSORRT_MAJOR "${NVINFER_VER_CONTENT}")
37-
string(REGEX REPLACE "define NV_TENSORRT_MAJOR * +([0-9]+)" "\\1" NV_TENSORRT_MAJOR "${NV_TENSORRT_MAJOR}")
38-
string(REGEX MATCH "define NV_TENSORRT_MINOR * +([0-9]+)" NV_TENSORRT_MINOR "${NVINFER_VER_CONTENT}")
39-
string(REGEX REPLACE "define NV_TENSORRT_MINOR * +([0-9]+)" "\\1" NV_TENSORRT_MINOR "${NV_TENSORRT_MINOR}")
40-
string(REGEX MATCH "define NV_TENSORRT_PATCH * +([0-9]+)" NV_TENSORRT_PATCH "${NVINFER_VER_CONTENT}")
41-
string(REGEX REPLACE "define NV_TENSORRT_PATCH * +([0-9]+)" "\\1" NV_TENSORRT_PATCH "${NV_TENSORRT_PATCH}")
36+
37+
# Starting TRT 10.11, TRT version macros have changed
38+
string(REGEX MATCH "TRT_MAJOR_ENTERPRISE" TRT_VER_CHECK "${NVINFER_VER_CONTENT}")
39+
# Pre TRT 10.11
40+
if("${TRT_VER_CHECK}" STREQUAL "")
41+
string(REGEX MATCH "define NV_TENSORRT_MAJOR * +([0-9]+)" NV_TENSORRT_MAJOR "${NVINFER_VER_CONTENT}")
42+
string(REGEX REPLACE "define NV_TENSORRT_MAJOR * +([0-9]+)" "\\1" NV_TENSORRT_MAJOR "${NV_TENSORRT_MAJOR}")
43+
string(REGEX MATCH "define NV_TENSORRT_MINOR * +([0-9]+)" NV_TENSORRT_MINOR "${NVINFER_VER_CONTENT}")
44+
string(REGEX REPLACE "define NV_TENSORRT_MINOR * +([0-9]+)" "\\1" NV_TENSORRT_MINOR "${NV_TENSORRT_MINOR}")
45+
string(REGEX MATCH "define NV_TENSORRT_PATCH * +([0-9]+)" NV_TENSORRT_PATCH "${NVINFER_VER_CONTENT}")
46+
string(REGEX REPLACE "define NV_TENSORRT_PATCH * +([0-9]+)" "\\1" NV_TENSORRT_PATCH "${NV_TENSORRT_PATCH}")
47+
# TRT 10.11+
48+
else()
49+
string(REGEX MATCH "define TRT_MAJOR_ENTERPRISE * +([0-9]+)" NV_TENSORRT_MAJOR "${NVINFER_VER_CONTENT}")
50+
string(REGEX REPLACE "define TRT_MAJOR_ENTERPRISE * +([0-9]+)" "\\1" NV_TENSORRT_MAJOR "${NV_TENSORRT_MAJOR}")
51+
string(REGEX MATCH "define TRT_MINOR_ENTERPRISE * +([0-9]+)" NV_TENSORRT_MINOR "${NVINFER_VER_CONTENT}")
52+
string(REGEX REPLACE "define TRT_MINOR_ENTERPRISE * +([0-9]+)" "\\1" NV_TENSORRT_MINOR "${NV_TENSORRT_MINOR}")
53+
string(REGEX MATCH "define TRT_PATCH_ENTERPRISE * +([0-9]+)" NV_TENSORRT_PATCH "${NVINFER_VER_CONTENT}")
54+
string(REGEX REPLACE "define TRT_PATCH_ENTERPRISE * +([0-9]+)" "\\1" NV_TENSORRT_PATCH "${NV_TENSORRT_PATCH}")
55+
endif()
56+
4257
math(EXPR NV_TENSORRT_MAJOR_INT "${NV_TENSORRT_MAJOR}")
4358
math(EXPR NV_TENSORRT_MINOR_INT "${NV_TENSORRT_MINOR}")
4459
math(EXPR NV_TENSORRT_PATCH_INT "${NV_TENSORRT_PATCH}")

‎onnxruntime/core/providers/tensorrt/tensorrt_execution_provider.cc‎

Lines changed: 30 additions & 1 deletion
Original file line numberDiff line numberDiff line change
@@ -538,9 +538,18 @@ bool ApplyProfileShapesFromProviderOptions(std::vector<nvinfer1::IOptimizationPr
538538
input_explicit_shape_ranges[input_name][static_cast<int64_t>(j)][i].push_back(opt_value);
539539
}
540540

541+
#if (NV_TENSORRT_MAJOR == 10 && NV_TENSORRT_MINOR > 10) || NV_TENSORRT_MAJOR > 10
542+
std::vector<int64_t> shapes_min_64(shapes_min.begin(), shapes_min.end());
543+
std::vector<int64_t> shapes_opt_64(shapes_opt.begin(), shapes_opt.end());
544+
std::vector<int64_t> shapes_max_64(shapes_max.begin(), shapes_max.end());
545+
trt_profile->setShapeValuesV2(input_name.c_str(), nvinfer1::OptProfileSelector::kMIN, &shapes_min_64[0], shape_size);
546+
trt_profile->setShapeValuesV2(input_name.c_str(), nvinfer1::OptProfileSelector::kMAX, &shapes_opt_64[0], shape_size);
547+
trt_profile->setShapeValuesV2(input_name.c_str(), nvinfer1::OptProfileSelector::kOPT, &shapes_max_64[0], shape_size);
548+
#else
541549
trt_profile->setShapeValues(input_name.c_str(), nvinfer1::OptProfileSelector::kMIN, &shapes_min[0], shape_size);
542550
trt_profile->setShapeValues(input_name.c_str(), nvinfer1::OptProfileSelector::kMAX, &shapes_max[0], shape_size);
543551
trt_profile->setShapeValues(input_name.c_str(), nvinfer1::OptProfileSelector::kOPT, &shapes_opt[0], shape_size);
552+
#endif
544553
}
545554
// Execution tensor
546555
else {
@@ -627,6 +636,17 @@ Status ApplyProfileShapesFromInputTensorValue(std::vector<nvinfer1::IOptimizatio
627636
if (input->isShapeTensor()) {
628637
// shape tensor
629638
int shape_size = nb_dims == 0 ? 1 : static_cast<int>(tensor_shapes[0]);
639+
#if (NV_TENSORRT_MAJOR == 10 && NV_TENSORRT_MINOR > 10) || NV_TENSORRT_MAJOR > 10
640+
std::vector<int64_t> shapes_min(shape_size), shapes_opt(shape_size), shapes_max(shape_size);
641+
for (int j = 0; j < shape_size; j++) {
642+
shapes_min[j] = *(trt_profiles[0]->getShapeValuesV2(input_name.c_str(), nvinfer1::OptProfileSelector::kMIN));
643+
shapes_max[j] = *(trt_profiles[0]->getShapeValuesV2(input_name.c_str(), nvinfer1::OptProfileSelector::kMAX));
644+
shapes_opt[j] = *(trt_profiles[0]->getShapeValuesV2(input_name.c_str(), nvinfer1::OptProfileSelector::kOPT));
645+
}
646+
trt_profile->setShapeValuesV2(input_name.c_str(), nvinfer1::OptProfileSelector::kMIN, &shapes_min[0], shape_size);
647+
trt_profile->setShapeValuesV2(input_name.c_str(), nvinfer1::OptProfileSelector::kMAX, &shapes_max[0], shape_size);
648+
trt_profile->setShapeValuesV2(input_name.c_str(), nvinfer1::OptProfileSelector::kOPT, &shapes_opt[0], shape_size);
649+
#else
630650
std::vector<int32_t> shapes_min(shape_size), shapes_opt(shape_size), shapes_max(shape_size);
631651
for (int j = 0; j < shape_size; j++) {
632652
shapes_min[j] = *(trt_profiles[0]->getShapeValues(input_name.c_str(), nvinfer1::OptProfileSelector::kMIN));
@@ -636,6 +656,7 @@ Status ApplyProfileShapesFromInputTensorValue(std::vector<nvinfer1::IOptimizatio
636656
trt_profile->setShapeValues(input_name.c_str(), nvinfer1::OptProfileSelector::kMIN, &shapes_min[0], shape_size);
637657
trt_profile->setShapeValues(input_name.c_str(), nvinfer1::OptProfileSelector::kMAX, &shapes_max[0], shape_size);
638658
trt_profile->setShapeValues(input_name.c_str(), nvinfer1::OptProfileSelector::kOPT, &shapes_opt[0], shape_size);
659+
#endif
639660
} else {
640661
// execution tensor
641662
nvinfer1::Dims dims_min, dims_opt, dims_max;
@@ -733,10 +754,18 @@ Status ApplyProfileShapesFromInputTensorValue(std::vector<nvinfer1::IOptimizatio
733754
}
734755
*engine_update = true;
735756
}
736-
757+
#if (NV_TENSORRT_MAJOR == 10 && NV_TENSORRT_MINOR > 10) || NV_TENSORRT_MAJOR > 10
758+
std::vector<int64_t> shapes_min_64(shapes_min.begin(), shapes_min.end());
759+
std::vector<int64_t> shapes_opt_64(shapes_opt.begin(), shapes_opt.end());
760+
std::vector<int64_t> shapes_max_64(shapes_max.begin(), shapes_max.end());
761+
trt_profile->setShapeValuesV2(input_name.c_str(), nvinfer1::OptProfileSelector::kMIN, &shapes_min_64[0], shape_size);
762+
trt_profile->setShapeValuesV2(input_name.c_str(), nvinfer1::OptProfileSelector::kMAX, &shapes_opt_64[0], shape_size);
763+
trt_profile->setShapeValuesV2(input_name.c_str(), nvinfer1::OptProfileSelector::kOPT, &shapes_max_64[0], shape_size);
764+
#else
737765
trt_profile->setShapeValues(input_name.c_str(), nvinfer1::OptProfileSelector::kMIN, &shapes_min[0], shape_size);
738766
trt_profile->setShapeValues(input_name.c_str(), nvinfer1::OptProfileSelector::kMAX, &shapes_max[0], shape_size);
739767
trt_profile->setShapeValues(input_name.c_str(), nvinfer1::OptProfileSelector::kOPT, &shapes_opt[0], shape_size);
768+
#endif
740769
} else { // Execution tensor
741770
nvinfer1::Dims dims_min(dims), dims_opt(dims), dims_max(dims);
742771
for (int j = 0, end = nb_dims; j < end; ++j) {

0 commit comments

Comments
 (0)