TensorRT 10.5 GA Release - 2024-10-01 (#998)

* Staging 10.5 release Signed-off-by: Michal Guzek <mguzek@nvidia.com> * Update minor version in CMakeLists Signed-off-by: Michal Guzek <mguzek@nvidia.com> --------- Signed-off-by: Michal Guzek <mguzek@nvidia.com>
onnx · Oct 8, 2024 · 886aff9 · 886aff9
1 parent 3775e49
commit 886aff9
Show file tree

Hide file tree

Showing 18 changed files with 1,208 additions and 1,012 deletions.
diff --git a/CMakeLists.txt b/CMakeLists.txt
@@ -28,7 +28,7 @@ add_definitions("-DSOURCE_LENGTH=${SOURCE_LENGTH}")
 # Version information
 #--------------------------------------------------
 set(ONNX2TRT_MAJOR 10)
-set(ONNX2TRT_MINOR 4)
+set(ONNX2TRT_MINOR 5)
 set(ONNX2TRT_PATCH 0)
 set(ONNX2TRT_VERSION "${ONNX2TRT_MAJOR}.${ONNX2TRT_MINOR}.${ONNX2TRT_PATCH}" CACHE STRING "ONNX2TRT version")
 

diff --git a/ConditionalHelpers.cpp b/ConditionalHelpers.cpp
@@ -34,19 +34,19 @@ SubgraphPortsMap::const_iterator findLayer(const SubgraphPortsMap& inputs, const
 
 // Add an ConditionalInputLayer between `layer` and its inputs.
 // I.e. input[inIdx] -> layer ==> input[inIdx] -> ConditionalInputLayer -> layer.
-Status addConditionalInputLayer(ImporterContext* ctx, nvinfer1::IIfConditional* conditional, InputsMap& inputsMap,
+void addConditionalInputLayer(ImporterContext* ctx, nvinfer1::IIfConditional* conditional, InputsMap& inputsMap,
     nvinfer1::ILayer& layer, int32_t inIdx)
 {
     auto input = layer.getInput(inIdx);
     if (input == nullptr)
     {
         // Phantom input (an input that is really constant weights).
-        return Status::success();
+        return;
     }
 
     if (layer.getType() == nvinfer1::LayerType::kCONDITIONAL_OUTPUT)
     {
-        return Status::success();
+        return;
     }
 
     auto const name = input->getName();
@@ -70,12 +70,11 @@ Status addConditionalInputLayer(ImporterContext* ctx, nvinfer1::IIfConditional*
     }
     auto ifOutput = N_CHECK(inputLayer->getOutput(0));
     layer.setInput(inIdx, *ifOutput);
-    return Status::success();
 };
 
 // Take a snapshot of the network before and after parsing the subgraph and return a list
 // of newly added network layers.
-Status importSubgraph(ImporterContext* ctx, ::ONNX_NAMESPACE::GraphProto const& subgraph,
+void importSubgraph(ImporterContext* ctx, ::ONNX_NAMESPACE::GraphProto const& subgraph,
     std::vector<nvinfer1::ILayer*>& newLayers, std::vector<TensorOrWeights>& subgraphTensors)
 {
     auto net = ctx->network();
@@ -85,7 +84,7 @@ Status importSubgraph(ImporterContext* ctx, ::ONNX_NAMESPACE::GraphProto const&
     NameScope nameScope(*ctx);
 
     std::vector<Status> errors{};
-    CHECK_STATUS(onnx2trt::parseGraph(ctx, subgraph, errors));
+    onnx2trt::parseGraph(ctx, subgraph, errors);
 
     for (int32_t i = 0; i < subgraph.output_size(); ++i)
     {
@@ -97,12 +96,10 @@ Status importSubgraph(ImporterContext* ctx, ::ONNX_NAMESPACE::GraphProto const&
     {
         newLayers.push_back(net->getLayer(i));
     }
-
-    return Status::success();
 }
 
 // Add an IConditionalInputLayer to `layer`'s inputs, if they don't already exist.
-Status addConditionalInputIfNeeded(ImporterContext* ctx, nvinfer1::IIfConditional* conditional, InputsMap& inputsMap,
+void addConditionalInputIfNeeded(ImporterContext* ctx, nvinfer1::IIfConditional* conditional, InputsMap& inputsMap,
     nvinfer1::ILayer& layer, SubgraphPortsMap subgraphInputsMap)
 {
     // Return all of the layer's inputs that are external to the subgraph that
@@ -125,11 +122,10 @@ Status addConditionalInputIfNeeded(ImporterContext* ctx, nvinfer1::IIfConditiona
         LOG_VERBOSE("Adding Input layer for " << layer.getName());
         addConditionalInputLayer(ctx, conditional, inputsMap, layer, inIdx);
     }
-    return Status::success();
 }
 
 // Add IConditionalInputLayers to `layer`'s inputs.
-Status addIfInputLayers(ImporterContext* ctx, nvinfer1::IIfConditional* conditional, InputsMap& inputsMap,
+void addIfInputLayers(ImporterContext* ctx, nvinfer1::IIfConditional* conditional, InputsMap& inputsMap,
     const std::vector<nvinfer1::ILayer*>& newLayers)
 {
     // Find all of the tensors entering the subgraph.
@@ -143,12 +139,10 @@ Status addIfInputLayers(ImporterContext* ctx, nvinfer1::IIfConditional* conditio
     {
         addConditionalInputIfNeeded(ctx, conditional, inputsMap, *layer, subgraphInputsMap);
     }
-
-    return Status::success();
 }
 
 // Given a subgraph, find all of its external inputs/outputs (tensors entering/exiting the subgraph).
-Status getSubgraphTensors(const std::vector<nvinfer1::ILayer*>& newLayers,
+void getSubgraphTensors(const std::vector<nvinfer1::ILayer*>& newLayers,
     std::unordered_map<nvinfer1::ITensor*, std::set<int32_t>>& externalOutputs, bool extractOutputs,
     const std::vector<std::string>* reportedOutputs = nullptr)
 {
@@ -255,20 +249,19 @@ Status getSubgraphTensors(const std::vector<nvinfer1::ILayer*>& newLayers,
             externalOutputs[tensor].insert(portIndex);
         }
     }
-    return Status::success();
 }
 
-Status getSubgraphOutputs(const std::vector<nvinfer1::ILayer*>& newLayers,
+void getSubgraphOutputs(const std::vector<nvinfer1::ILayer*>& newLayers,
     std::unordered_map<nvinfer1::ITensor*, std::set<int32_t>>& externalOutputs,
     const std::vector<std::string>& reportedOutputs)
 {
-    return getSubgraphTensors(newLayers, externalOutputs, true, &reportedOutputs);
+    getSubgraphTensors(newLayers, externalOutputs, true, &reportedOutputs);
 }
 
-Status getSubgraphInputs(const std::vector<nvinfer1::ILayer*>& newLayers,
+void getSubgraphInputs(const std::vector<nvinfer1::ILayer*>& newLayers,
     std::unordered_map<nvinfer1::ITensor*, std::set<int32_t>>& externalInputs)
 {
-    return getSubgraphTensors(newLayers, externalInputs, false);
+    getSubgraphTensors(newLayers, externalInputs, false);
 }
 
 } // namespace onnx2trt
diff --git a/ConditionalHelpers.hpp b/ConditionalHelpers.hpp
@@ -21,25 +21,25 @@ namespace onnx2trt
 // Given a subgraph, find all of its external inputs (tensors entering the subgraph).
 // The result is returned in `subgraphInputs`, which is a map indexed by ITensor (a tensor entering the subgraph) and
 // with values indicating a set of external input indices.
-Status getSubgraphInputs(std::vector<nvinfer1::ILayer*> const& newLayers,
+void getSubgraphInputs(std::vector<nvinfer1::ILayer*> const& newLayers,
     std::unordered_map<nvinfer1::ITensor*, std::set<int32_t>>& subgraphInputs);
 
 // Given a subgraph, find all of its external outputs (tensors exiting the subgraph).
 // The result is returned in `subgraphInputs`, which is a map indexed by ITensor (a tensor exiting the subgraph) and
 // with values indicating a set of external outputs indices.
-Status getSubgraphOutputs(const std::vector<nvinfer1::ILayer*>& newLayers,
+void getSubgraphOutputs(const std::vector<nvinfer1::ILayer*>& newLayers,
     std::unordered_map<nvinfer1::ITensor*, std::set<int32_t>>& subgraphOutputs,
     const std::vector<std::string>& reportedOutputs);
 
 // Take a snapshot of the network before and after parsing the subgraph and return a list
 // of newly added network layers.
-Status importSubgraph(ImporterContext* ctx, ::ONNX_NAMESPACE::GraphProto const& subgraph,
+void importSubgraph(ImporterContext* ctx, ::ONNX_NAMESPACE::GraphProto const& subgraph,
     std::vector<nvinfer1::ILayer*>& newLayers, std::vector<TensorOrWeights>& subgraphTensors);
 
 using InputsMap = std::unordered_map<std::string, nvinfer1::IIfConditionalInputLayer*>;
 
 // Add IIfConditionalInputLayers to the inputs of the subgraph indicated by `subgraph`.
-onnx2trt::Status addIfInputLayers(ImporterContext* ctx, nvinfer1::IIfConditional* conditional, InputsMap& inputsMap,
+void addIfInputLayers(ImporterContext* ctx, nvinfer1::IIfConditional* conditional, InputsMap& inputsMap,
     const std::vector<nvinfer1::ILayer*>& newLayers);
 
 } // namespace onnx2trt
diff --git a/ImporterContext.hpp b/ImporterContext.hpp
@@ -371,9 +371,9 @@ class ImporterContext
     }
 };
 
-typedef ValueOrStatus<std::vector<TensorOrWeights>> NodeImportResult;
-typedef std::function<NodeImportResult(ImporterContext* ctx, ::ONNX_NAMESPACE::NodeProto const& node,
-    size_t const nodeIdx, std::vector<TensorOrWeights>& inputs)>
+typedef std::vector<TensorOrWeights> NodeOutputs;
+typedef std::function<NodeOutputs(ImporterContext* ctx, ::ONNX_NAMESPACE::NodeProto const& node, size_t const nodeIdx,
+    std::vector<TensorOrWeights>& inputs)>
     NodeImporter;
 
 typedef std::function<void(