TensorRT 8.5 GA Release (#879)

Signed-off-by: Kevin Chen <[email protected]> Signed-off-by: Kevin Chen <[email protected]>
onnx · Nov 2, 2022 · 3b008c4 · 3b008c4
1 parent 12eb0d6
commit 3b008c4
Show file tree

Hide file tree

Showing 19 changed files with 1,478 additions and 1,223 deletions.
diff --git a/CMakeLists.txt b/CMakeLists.txt
@@ -28,8 +28,8 @@ add_definitions("-DSOURCE_LENGTH=${SOURCE_LENGTH}")
 # Version information
 #--------------------------------------------------
 set(ONNX2TRT_MAJOR 8)
-set(ONNX2TRT_MINOR 4)
-set(ONNX2TRT_PATCH 3)
+set(ONNX2TRT_MINOR 5)
+set(ONNX2TRT_PATCH 1)
 set(ONNX2TRT_VERSION "${ONNX2TRT_MAJOR}.${ONNX2TRT_MINOR}.${ONNX2TRT_PATCH}" CACHE STRING "ONNX2TRT version")
 
 #--------------------------------------------------
@@ -55,16 +55,10 @@ if (BUILD_ONNXIFI)
   set(ONNXIFI_SOURCES onnx_trt_backend.cpp)
 endif()
 
-# Build executables if BUILD_EXES flag is set
-if (DEFINED BUILD_EXES)
-  set(EXECUTABLE_SOURCES
-    main.cpp
-  )
-  set(API_TESTS_SOURCES
-    getSupportedAPITest.cpp
-    ModelImporter.cpp
-  )
-endif()
+set(API_TESTS_SOURCES
+  getSupportedAPITest.cpp
+  ModelImporter.cpp
+)
 
 if (NOT TARGET protobuf::libprotobuf)
   FIND_PACKAGE(Protobuf REQUIRED)
@@ -135,23 +129,12 @@ if(BUILD_ONNXIFI)
   target_link_libraries(trt_onnxify PUBLIC nvonnxparser_static ${CMAKE_THREAD_LIBS_INIT} ${CMAKE_DL_LIBS})
 endif()
 
-# --------------------------------
-# Converter executable
-# --------------------------------
-if (DEFINED BUILD_EXES)
-  add_executable(onnx2trt ${EXECUTABLE_SOURCES})
-  target_include_directories(onnx2trt PUBLIC ${ONNX_INCLUDE_DIRS})
-  target_link_libraries(onnx2trt PUBLIC ${PROTOBUF_LIB} onnx nvonnxparser_static ${CMAKE_THREAD_LIBS_INIT} ${CMAKE_DL_LIBS}) #${CUDA_LIBRARIES}
-endif()
-
 # --------------------------------
 # API Tests
 # --------------------------------
-if (DEFINED BUILD_EXES)
-  add_executable(getSupportedAPITest ${API_TESTS_SOURCES})
-  target_include_directories(getSupportedAPITest PUBLIC ${ONNX_INCLUDE_DIRS} ${CUDNN_INCLUDE_DIR})
-  target_link_libraries(getSupportedAPITest PUBLIC ${PROTOBUF_LIB} nvonnxparser_static ${CMAKE_THREAD_LIBS_INIT} ${CMAKE_DL_LIBS}) #${CUDA_LIBRARIES}
-endif()
+add_executable(getSupportedAPITest ${API_TESTS_SOURCES})
+target_include_directories(getSupportedAPITest PUBLIC ${ONNX_INCLUDE_DIRS} ${CUDNN_INCLUDE_DIR})
+target_link_libraries(getSupportedAPITest PUBLIC ${PROTOBUF_LIB} nvonnxparser_static ${CMAKE_THREAD_LIBS_INIT} ${CMAKE_DL_LIBS})
 
 # --------------------------------
 # Installation
@@ -166,12 +149,6 @@ install(TARGETS
 install(FILES ${HEADERS}
   DESTINATION include
 )
-if (DEFINED BUILD_EXES)
-  install(TARGETS
-          onnx2trt
-          RUNTIME DESTINATION bin
-  )
-endif()
 
 SET(CPACK_GENERATOR "DEB")
 SET(CPACK_DEBIAN_PACKAGE_MAINTAINER "NVIDIA") #required

diff --git a/ConditionalHelpers.cpp b/ConditionalHelpers.cpp
@@ -58,7 +58,10 @@ Status addConditionalInputLayer(IImporterContext* ctx, nvinfer1::IIfConditional*
         inputsMap[name] = inputLayer;
         const std::string inputLayerName(name);
         ctx->registerLayer(inputLayer, inputLayerName + "_InputLayer");
-        ctx->registerTensor(TensorOrWeights{inputLayer->getOutput(0)}, inputLayerName + "_InputLayer_output");
+        // Note: Since multiple conditionals may use the same external tensor, check unique names for output tensors of
+        // IfConditionalInputLayers to avoid tensor name duplication.
+        ctx->registerTensor(
+            TensorOrWeights{inputLayer->getOutput(0)}, inputLayerName + "_InputLayer_output", /*checkUniqueName*/ true);
     }
     else
     {
@@ -259,11 +262,11 @@ Status getSubgraphTensors(const std::vector<nvinfer1::ILayer*>& newLayers,
     }
 
     using TensorsVec = std::vector<nvinfer1::ITensor*>;
-    auto getOutputs = [&](nvinfer1::ILayer const* l, TensorsVec res) {
+    auto getOutputs = [&](nvinfer1::ILayer const* l, TensorsVec& res) {
         getTensors(l, false, [&](nvinfer1::ITensor* t) { res.emplace_back(t); });
     };
 
-    auto getInputs = [&](nvinfer1::ILayer const* l, TensorsVec res) {
+    auto getInputs = [&](nvinfer1::ILayer const* l, TensorsVec& res) {
         getTensors(l, true, [&](nvinfer1::ITensor* t) { res.emplace_back(t); });
     };
 

diff --git a/ImporterContext.cpp b/ImporterContext.cpp
@@ -1,3 +1,7 @@
+/*
+ * SPDX-License-Identifier: Apache-2.0
+ */
+
 #include "ImporterContext.hpp"
 
 namespace onnx2trt
@@ -25,7 +29,7 @@ void ImporterContext::popBaseNameScope()
     mBaseNameScopeStack.pop_back();
 }
 
-void ImporterContext::registerTensor(TensorOrWeights tensor, std::string const& basename)
+void ImporterContext::registerTensor(TensorOrWeights tensor, std::string const& basename, bool const checkUniqueName)
 {
     // TRT requires unique tensor names.
     std::string const& uniqueName = generateUniqueName(mTensorNames, basename);
@@ -48,7 +52,7 @@ void ImporterContext::registerTensor(TensorOrWeights tensor, std::string const&
                     convertINT64(reinterpret_cast<int64_t*>(weights.values), weights.shape, this), weights.shape};
             }
             // It may be possible for nested subgraphs to have different values for the same initializer.
-            // For multiple name scopes - use unique name to keep track of weights. 
+            // For multiple name scopes - use unique name to keep track of weights.
             if (!mBaseNameScopeStack.empty())
             {
                 tensor.weights().setName(uniqueName.c_str());
@@ -60,13 +64,15 @@ void ImporterContext::registerTensor(TensorOrWeights tensor, std::string const&
         }
     }
 
-    auto const p = this->tensors().emplace(basename, TensorOrWeights{});
+    std::string const& nameToCheck = checkUniqueName ? uniqueName : basename;
+
+    auto const p = this->tensors().emplace(nameToCheck, TensorOrWeights{});
     bool nameIsDuplicate = false;
     if (!mBaseNameScopeStack.empty())
     {
         // Remember original binding so it can be restored when scope is popped.
         auto const q
-            = mBaseNameScopeStack.back().emplace(basename, std::make_pair(p.second, std::move(p.first->second)));
+            = mBaseNameScopeStack.back().emplace(nameToCheck, std::make_pair(p.second, std::move(p.first->second)));
         // Check that scope did not already have a binding for basename.
         nameIsDuplicate = !q.second;
     }
@@ -78,7 +84,7 @@ void ImporterContext::registerTensor(TensorOrWeights tensor, std::string const&
     }
     if (nameIsDuplicate)
     {
-        throw std::runtime_error("ONNX graph has duplicate tensor name: " + basename);
+        throw std::runtime_error("ONNX graph has duplicate tensor name: " + nameToCheck);
     }
     p.first->second = std::move(tensor);
 }
@@ -97,7 +103,7 @@ void ImporterContext::registerLayer(nvinfer1::ILayer* layer, std::string const&
         layer->setName(uniqueName.c_str());
         if (layer->getType() == nvinfer1::LayerType::kCONSTANT)
         {
-            if (basename != uniqueName)
+            if (basename != uniqueName && mConstantLayers.find(uniqueName) != mConstantLayers.end())
             {
                 LOG_ERROR("Constant layer: " << uniqueName << " can be a duplicate of: " << basename);
                 assert(!"Internal error: duplicate constant layers for the same weights");

diff --git a/ImporterContext.hpp b/ImporterContext.hpp
@@ -158,7 +158,8 @@ class ImporterContext final : public IImporterContext
     void popBaseNameScope() override;
 
     // This actually handles weights as well, but is named this way to be consistent with the tensors()
-    void registerTensor(TensorOrWeights tensor, std::string const& basename) override;
+    void registerTensor(
+        TensorOrWeights tensor, std::string const& basename, bool const checkUniqueName = false) override;
 
     void registerLayer(nvinfer1::ILayer* layer, std::string const& basename) override;