ONNX-TensorRT 9.2 Release (#948)

yuanyao-nv · web-flow · commit 306247b30da1 · 2023-12-04T10:42:43.000-08:00
Signed-off-by: Yuan Yao &lt;yuanyao@nvidia.com&gt;
diff --git a/CMakeLists.txt b/CMakeLists.txt
@@ -28,7 +28,7 @@ add_definitions("-DSOURCE_LENGTH=${SOURCE_LENGTH}")
 # Version information
 #--------------------------------------------------
 set(ONNX2TRT_MAJOR 9)
-set(ONNX2TRT_MINOR 1)
+set(ONNX2TRT_MINOR 2)
 set(ONNX2TRT_PATCH 0)
 set(ONNX2TRT_VERSION "${ONNX2TRT_MAJOR}.${ONNX2TRT_MINOR}.${ONNX2TRT_PATCH}" CACHE STRING "ONNX2TRT version")
 
diff --git a/LoopHelpers.cpp b/LoopHelpers.cpp
@@ -13,7 +13,8 @@ nvinfer1::ITensor* addLoopCounter(IImporterContext* ctx, nvinfer1::ILoop* loop,
     nvinfer1::ITensor* initialTensor
         = addConstantScalar(ctx, initial, ::ONNX_NAMESPACE::TensorProto::INT64, nvinfer1::Dims{1, 1})->getOutput(0);
     nvinfer1::ITensor* one
-        = addConstantScalar(ctx, 1, ::ONNX_NAMESPACE::TensorProto::INT64, nvinfer1::Dims{1, 1})->getOutput(0);
+        = addConstantScalar(ctx, static_cast<int64_t>(1), ::ONNX_NAMESPACE::TensorProto::INT64, nvinfer1::Dims{1, 1})
+              ->getOutput(0);
 
     auto counter = loop->addRecurrence(*initialTensor);
     nvinfer1::ITensor* addOne = ctx->network()
diff --git a/ModelImporter.cpp b/ModelImporter.cpp
@@ -208,7 +208,9 @@ Status parseNode(
     }
 
     ASSERT_NODE((node.output().size() <= static_cast<int32_t>(outputs.size())),
-        "Node has more output tensors than TRT expected.", node, nodeIdx, ErrorCode::kINVALID_GRAPH);
+        "Node has more output tensors than TRT expected, expected output size is "
+            << outputs.size() << ", actual output size is " << node.output().size() << ".",
+        node, nodeIdx, ErrorCode::kINVALID_GRAPH);
 
     // Set output names and register outputs with the context.
     std::ostringstream ssOutputs{};
diff --git a/README.md b/README.md
@@ -4,7 +4,7 @@
 
 Parses ONNX models for execution with [TensorRT](https://developer.nvidia.com/tensorrt).
 
-See also the [TensorRT documentation](https://docs.nvidia.com/deeplearning/sdk/#inference).
+See also the [TensorRT documentation](https://docs.nvidia.com/deeplearning/tensorrt/).
 
 For the list of recent changes, see the [changelog](docs/Changelog.md).
 
@@ -48,8 +48,8 @@ Current supported ONNX operators are found in the [operator support matrix](docs
 ### Dependencies
 
  - [Protobuf >= 3.0.x](https://github.com/google/protobuf/releases)
- - [TensorRT 9.1](https://developer.nvidia.com/tensorrt)
- - [TensorRT 9.1 open source libaries] (https://github.com/NVIDIA/TensorRT/)
+ - [TensorRT 9.2](https://developer.nvidia.com/tensorrt)
+ - [TensorRT 9.2 open source libaries] (https://github.com/NVIDIA/TensorRT/)
 
 ### Building
 
@@ -99,7 +99,7 @@ Refer to the link or run `polygraphy run -h` for more information on CLI options
 
 Python bindings for the ONNX-TensorRT parser are packaged in the shipped `.whl` files.
 
-TensorRT 9.1 supports ONNX release 1.14.0. Install it with:
+TensorRT 9.2 supports ONNX release 1.14.0. Install it with:
 
     python3 -m pip install onnx==1.14.0
 
diff --git a/Status.hpp b/Status.hpp
@@ -38,13 +38,14 @@
 #define MAKE_NODE_ERROR(desc, code, node, index)                                                                       \
     onnx2trt::Status((code), (desc), __FILENAME__, __LINE__, __func__, (index), (node.name()), (node.op_type()))
 
-#define ASSERT_NODE(condition, desc, node, index, error_code)                                                          \
+#define ASSERT_NODE(condition, msg, node, index, error_code)                                                           \
     do                                                                                                                 \
     {                                                                                                                  \
         if (!(condition))                                                                                              \
         {                                                                                                              \
-            std::string error = std::string("Assertion failed:  " #condition) + ". " + desc;                           \
-            return MAKE_NODE_ERROR((error), (error_code), node, index);                                                \
+            std::stringstream error;                                                                                   \
+            error << "Assertion failed: " << #condition << ": " << msg;                                                \
+            return MAKE_NODE_ERROR((error.str()), (error_code), node, index);                                          \
         }                                                                                                              \
     } while (0)
 
diff --git a/builtin_op_importers.cpp b/builtin_op_importers.cpp
diff --git a/builtin_op_static_checkers.cpp b/builtin_op_static_checkers.cpp
@@ -350,6 +350,8 @@ DEFINE_OP_CHECKER(GRU)
     }
 }
 
+DEFINE_OP_EMPTY_CHECKER(Hardmax)
+
 DEFINE_OP_EMPTY_CHECKER(HardSigmoid)
 
 DEFINE_OP_EMPTY_CHECKER(Identity)
@@ -830,11 +832,6 @@ DEFINE_OP_CHECKER(Det)
     STATIC_CHECK(false, ErrorCode::kUNSUPPORTED_NODE, node, errors, nodeIndex);
 }
 
-DEFINE_OP_CHECKER(Hardmax)
-{
-    STATIC_CHECK(false, ErrorCode::kUNSUPPORTED_NODE, node, errors, nodeIndex);
-}
-
 DEFINE_OP_CHECKER(ImageDecoder)
 {
     STATIC_CHECK(false, ErrorCode::kUNSUPPORTED_NODE, node, errors, nodeIndex);
diff --git a/docs/Changelog.md b/docs/Changelog.md
@@ -2,6 +2,12 @@
 
 # ONNX-TensorRT Changelog
 
+# TensorRT 9.2 GA Release - 2023-12-4
+For more details, see the 9.2 GA release notes for the fixes since 9.1 GA.
+
+- Added support for `Hardmax`
+- Fixed type inference for few operators to use native ONNX types
+
 # TensorRT 9.1 GA Release - 2023-10-18
 For more details, see the 9.1 GA release notes for the fixes since 9.0 GA.
 
diff --git a/docs/operators.md b/docs/operators.md
@@ -2,7 +2,7 @@
 
 # Supported ONNX Operators
 
-TensorRT 9.1 supports operators up to Opset 19. Latest information of ONNX operators can be found [here](https://github.com/onnx/onnx/blob/master/docs/Operators.md)
+TensorRT 9.2 supports operators up to Opset 19. Latest information of ONNX operators can be found [here](https://github.com/onnx/onnx/blob/master/docs/Operators.md)
 
 TensorRT supports the following ONNX data types: DOUBLE, FLOAT32, FLOAT16, BFLOAT16, INT32, INT64, FP8, INT8, UINT8, and BOOL
 
@@ -80,7 +80,7 @@ See below for the support matrix of ONNX operators in ONNX-TensorRT.
 | HannWindow                | N          |
 | HardSwish                 | Y          | FP32, FP16, BF16, FP8, INT8 |
 | HardSigmoid               | Y          | FP32, FP16, BF16, FP8, INT8 |
-| Hardmax                   | N          |
+| Hardmax                   | Y          | FP32, FP16, BF16 | `axis` dimension of input must be a build-time constant
 | Identity                  | Y          | FP32, FP16, BF16, INT32, INT64, FP8, INT8, BOOL |
 | If                        | Y          | FP32, FP16, BF16, INT32, INT64, BOOL | Output tensors of the two conditional branches must have broadcastable shapes, and must have different names
 | ImageScaler               | Y          | FP32, FP16, BF16|
@@ -116,7 +116,7 @@ See below for the support matrix of ONNX operators in ONNX-TensorRT.
 | NonMaxSuppression         | Y          | FP32, FP16 |
 | NonZero                   | Y          | FP32, FP16
 | Not                       | Y          | BOOL |
-| OneHot                    | Y          | FP32, FP16, BF16, INT32, INT64, BOOL |
+| OneHot                    | Y          | FP32, FP16, BF16, INT32, INT64, BOOL | `depth` must be a build-time constant
 | Optional                  | N          |
 | OptionalGetElement        | N          |
 | OptionalHasElement        | N          |
diff --git a/onnx2trt_utils.cpp b/onnx2trt_utils.cpp
diff --git a/onnx2trt_utils.hpp b/onnx2trt_utils.hpp
diff --git a/onnx_tensorrt/__init__.py b/onnx_tensorrt/__init__.py

Original file line number	Diff line number	Diff line change
`@@ -208,7 +208,9 @@ Status parseNode(`
`208`	`208`	`}`
`209`	`209`
`210`	`210`	`ASSERT_NODE((node.output().size() <= static_cast<int32_t>(outputs.size())),`
`211`		`- "Node has more output tensors than TRT expected.", node, nodeIdx, ErrorCode::kINVALID_GRAPH);`
	`211`	`+ "Node has more output tensors than TRT expected, expected output size is "`
	`212`	`+ << outputs.size() << ", actual output size is " << node.output().size() << ".",`
	`213`	`+ node, nodeIdx, ErrorCode::kINVALID_GRAPH);`
`212`	`214`
`213`	`215`	`// Set output names and register outputs with the context.`
`214`	`216`	`std::ostringstream ssOutputs{};`
Original file line number	Diff line number	Diff line change
`@@ -350,6 +350,8 @@ DEFINE_OP_CHECKER(GRU)`
`350`	`350`	`}`
`351`	`351`	`}`
`352`	`352`
	`353`	`+DEFINE_OP_EMPTY_CHECKER(Hardmax)`
	`354`	`+`
`353`	`355`	`DEFINE_OP_EMPTY_CHECKER(HardSigmoid)`
`354`	`356`
`355`	`357`	`DEFINE_OP_EMPTY_CHECKER(Identity)`
`@@ -830,11 +832,6 @@ DEFINE_OP_CHECKER(Det)`
`830`	`832`	`STATIC_CHECK(false, ErrorCode::kUNSUPPORTED_NODE, node, errors, nodeIndex);`
`831`	`833`	`}`
`832`	`834`
`833`		`-DEFINE_OP_CHECKER(Hardmax)`
`834`		`-{`
`835`		`- STATIC_CHECK(false, ErrorCode::kUNSUPPORTED_NODE, node, errors, nodeIndex);`
`836`		`-}`
`837`		`-`
`838`	`835`	`DEFINE_OP_CHECKER(ImageDecoder)`
`839`	`836`	`{`
`840`	`837`	`STATIC_CHECK(false, ErrorCode::kUNSUPPORTED_NODE, node, errors, nodeIndex);`