Plugin class for user-implemented layers. More...
#include <NvInferRuntimePlugin.h>
Public Member Functions | |
| virtual nvinfer1::DataType | getOutputDataType (int32_t index, nvinfer1::DataType const *inputTypes, int32_t nbInputs) const noexcept=0 |
| Return the DataType of the plugin output at the requested index. More... | |
| virtual TRT_DEPRECATED bool | isOutputBroadcastAcrossBatch (int32_t outputIndex, bool const *inputIsBroadcasted, int32_t nbInputs) const noexcept=0 |
| Return true if the output tensor is broadcast across a batch. More... | |
| virtual TRT_DEPRECATED bool | canBroadcastInputAcrossBatch (int32_t inputIndex) const noexcept=0 |
| Return true if the plugin can use an input tensor that is broadcast across batch without replication. More... | |
| virtual void | configurePlugin (Dims const *inputDims, int32_t nbInputs, Dims const *outputDims, int32_t nbOutputs, DataType const *inputTypes, DataType const *outputTypes, bool const *inputIsBroadcast, bool const *outputIsBroadcast, PluginFormat floatFormat, int32_t maxBatchSize) noexcept=0 |
| Configure the layer with input and output data types. More... | |
| IPluginV2Ext ()=default | |
| ~IPluginV2Ext () override=default | |
| virtual void | attachToContext (cudnnContext *, cublasContext *, IGpuAllocator *) noexcept |
| Attach the plugin object to an execution context and grant the plugin the access to some context resources. More... | |
| virtual void | detachFromContext () noexcept |
| Detach the plugin object from its execution context. More... | |
| IPluginV2Ext * | clone () const noexcept override=0 |
| Clone the plugin object. This copies over internal plugin parameters as well and returns a new plugin object with these parameters. If the source plugin is pre-configured with configurePlugin(), the returned object must also be pre-configured. The returned object must allow attachToContext() with a new execution context. Cloned plugin objects can share the same per-engine immutable resource (e.g. weights) with the source object (e.g. via ref-counting) to avoid duplication. More... | |
 Public Member Functions inherited from nvinfer1::IPluginV2 | |
| virtual AsciiChar const * | getPluginType () const noexcept=0 |
| Return the plugin type. Should match the plugin name returned by the corresponding plugin creator. More... | |
| virtual AsciiChar const * | getPluginVersion () const noexcept=0 |
| Return the plugin version. Should match the plugin version returned by the corresponding plugin creator. More... | |
| virtual int32_t | getNbOutputs () const noexcept=0 |
| Get the number of outputs from the layer. More... | |
| virtual Dims | getOutputDimensions (int32_t index, Dims const *inputs, int32_t nbInputDims) noexcept=0 |
| Get the dimension of an output tensor. More... | |
| virtual bool | supportsFormat (DataType type, PluginFormat format) const noexcept=0 |
| Check format support. More... | |
| virtual int32_t | initialize () noexcept=0 |
| Initialize the layer for execution. This is called when the engine is created. More... | |
| virtual void | terminate () noexcept=0 |
| Release resources acquired during plugin layer initialization. This is called when the engine is destroyed. More... | |
| virtual size_t | getWorkspaceSize (int32_t maxBatchSize) const noexcept=0 |
| Find the workspace size required by the layer. More... | |
| virtual int32_t | enqueue (int32_t batchSize, void const *const *inputs, void *const *outputs, void *workspace, cudaStream_t stream) noexcept=0 |
| Execute the layer. More... | |
| virtual size_t | getSerializationSize () const noexcept=0 |
| Find the size of the serialization buffer required to store the plugin configuration in a binary file. More... | |
| virtual void | serialize (void *buffer) const noexcept=0 |
| Serialize the layer. More... | |
| virtual void | destroy () noexcept=0 |
| Destroy the plugin object. This will be called when the network, builder or engine is destroyed. More... | |
| virtual void | setPluginNamespace (AsciiChar const *pluginNamespace) noexcept=0 |
| Set the namespace that this plugin object belongs to. Ideally, all plugin objects from the same plugin library must have the same namespace. More... | |
| virtual AsciiChar const * | getPluginNamespace () const noexcept=0 |
| Return the namespace of the plugin object. More... | |
Protected Member Functions | |
| int32_t | getTensorRTVersion () const noexcept override |
| Return the API version with which this plugin was built. The upper byte reserved by TensorRT and is used to differentiate this from IPluginV2. More... | |
| void | configureWithFormat (Dims const *, int32_t, Dims const *, int32_t, DataType, PluginFormat, int32_t) noexcept override |
| Derived classes must not implement this. In a C++11 API it would be override final. More... | |
Detailed Description
Plugin class for user-implemented layers.
Plugins are a mechanism for applications to implement custom layers. This interface provides additional capabilities to the IPluginV2 interface by supporting different output data types and broadcast across batches.
- See also
- IPluginV2
- Deprecated:
- Deprecated in TensorRT 8.5. Implement IPluginV2DynamicExt or IPluginV2IOExt depending on your requirement.
Constructor & Destructor Documentation
◆ IPluginV2Ext()
|
default |
◆ ~IPluginV2Ext()
|
overridedefault |
Member Function Documentation
◆ attachToContext()
|
inlinevirtualnoexcept |
Attach the plugin object to an execution context and grant the plugin the access to some context resources.
- Parameters
-
cudnn The cuDNN context handle of the execution context. Will be a valid cuDNN context handle, or nullptr if TacticSource::kCUDNN is disabled. cublas The cuBLAS context handle of the execution context. Will be a valid cuBLAS context handle, or nullptr if TacticSource::kCUBLAS is disabled. allocator The allocator used by the execution context
This function is called automatically for each plugin when a new execution context is created. If the context was created without resources, this method is not called until the resources are assigned. It is also called if new resources are assigned to the context.
If the plugin needs per-context resource, it can be allocated here. The plugin can also get context-owned cuDNN and cuBLAS context here.
- Note
- The TacticSource::kCUDNN and TacticSource::kCUBLAS flag is disabled by default. The allocator pointer is unique to each building or execution context instance having overlapping lifetimes. It can be used as a key to manage resources across plugin instances sharing the same context. Plugins attached to different contexts will have different handles as their execution will not overlap.
- See also
- TacticSources
- getPluginCudnnHandle(void* executionContextIdentifier)
- getPluginCublasHandle(void* excecutionContextIdentifier)
- Note
- In the automotive safety context, the cuDNN and cuBLAS parameters will be nullptr because cuDNN and cuBLAS are not used by the safe runtime.
Usage considerations
- Allowed context for the API call
- Thread-safe: Yes, this method is required to be thread-safe and may be called from multiple threads when building networks on multiple devices sharing the same plugin.
◆ canBroadcastInputAcrossBatch()
|
pure virtualnoexcept |
Return true if the plugin can use an input tensor that is broadcast across batch without replication.
- Parameters
-
inputIndex Index of input that could be broadcast. Will be in the valid range between 0 and nbInputs - 1 where nbInputs is the maximum number of input tensors supported by this plugin.
- Returns
- true if the index is in the valid range and the plugin is able to broadcast a single copy of this input tensor across the batch. False otherwise.
For each input whose tensor is semantically broadcast across a batch, TensorRT calls this method before calling configurePlugin. If canBroadcastInputAcrossBatch returns true, TensorRT will not replicate the input tensor; i.e., there will be a single copy that the plugin must share across the batch. If it returns false, TensorRT will replicate the input tensor so that it appears like a non-broadcasted tensor.
This method is called only for inputs that can be broadcast.
Usage considerations
Allowed context for the API call
- Thread-safe: Yes, this method is required to be thread-safe and may be called from multiple threads when building networks on multiple devices sharing the same plugin.
- Deprecated:
- Deprecated in TensorRT 10.0. Implicit batch support is removed in TensorRT 10.0.
◆ clone()
|
overridepure virtualnoexcept |
Clone the plugin object. This copies over internal plugin parameters as well and returns a new plugin object with these parameters. If the source plugin is pre-configured with configurePlugin(), the returned object must also be pre-configured. The returned object must allow attachToContext() with a new execution context. Cloned plugin objects can share the same per-engine immutable resource (e.g. weights) with the source object (e.g. via ref-counting) to avoid duplication.
- Returns
- A pointer to a cloned plugin object if cloning was successful, otherwise nullptr.
Usage considerations
- Allowed context for the API call
- Thread-safe: Yes, this method is required to be thread-safe and may be called from multiple threads when building networks on multiple devices sharing the same plugin.
Implements nvinfer1::IPluginV2.
Implemented in nvinfer1::IPluginV2DynamicExt.
◆ configurePlugin()
|
pure virtualnoexcept |
Configure the layer with input and output data types.
This function is called by the builder prior to initialize(). It provides an opportunity for the layer to make algorithm choices on the basis of its weights, dimensions, data types and maximum batch size.
- Parameters
-
inputDims The input tensor dimensions. Will be an array of length nbInputs. nbInputs The number of inputs. Will be a non-negative integer. outputDims The output tensor dimensions. Will be an array of length nbOutputs. nbOutputs The number of outputs. Will be a positive integer. inputTypes The data types selected for the plugin inputs. Will be an array of length nbInputs. outputTypes The data types selected for the plugin outputs. Will be an array of length nbOutputs. inputIsBroadcast True for each input that the plugin must broadcast across the batch. Will be an array of length nbInputs. outputIsBroadcast True for each output that TensorRT will broadcast across the batch. Will be an array of length nbOutputs. floatFormat The format selected for the engine for the floating point inputs/outputs. maxBatchSize The maximum batch size. Will be a positive integer.
The dimensions passed here do not include the outermost batch size (i.e. for 2-D image networks, they will be 3-dimensional CHW dimensions). When inputIsBroadcast or outputIsBroadcast is true, the outermost batch size for that input or output must be treated as if it is one. Index 'i' of inputIsBroadcast is true only if the input is semantically broadcast across the batch and calling canBroadcastInputAcrossBatch with argument 'i' returns true. Index 'i' of outputIsBroadcast is true only if calling isOutputBroadcastAcrossBatch with argument 'i' returns true.
- Warning
- for the floatFormat field, the values PluginFormat::kCHW4, PluginFormat::kCHW16, and PluginFormat::kCHW32 will not be passed in, this is to keep backward compatibility with TensorRT 5.x series. Use PluginV2IOExt or PluginV2DynamicExt for other PluginFormats.
Usage considerations
- Allowed context for the API call
- Thread-safe: Yes, this method is required to be thread-safe and may be called from multiple threads when building networks on multiple devices sharing the same plugin. However, TensorRT will not call this method from two threads simultaneously on a given clone of a plugin.
◆ configureWithFormat()
|
inlineoverrideprotectedvirtualnoexcept |
Derived classes must not implement this. In a C++11 API it would be override final.
IPluginV2Ext::configureWithFormat() is a NOP operation for all classes derived from IPluginV2Ext. These classes call configurePlugin() instead.
Implements nvinfer1::IPluginV2.
◆ detachFromContext()
|
inlinevirtualnoexcept |
Detach the plugin object from its execution context.
This function is called automatically for each plugin when an execution context is destroyed or the context resources are unassigned from the context.
If the plugin owns per-context resource, it can be released here.
Usage considerations
- Allowed context for the API call
- Thread-safe: Yes, this method is required to be thread-safe and may be called from multiple threads when building networks on multiple devices sharing the same plugin.
◆ getOutputDataType()
|
pure virtualnoexcept |
Return the DataType of the plugin output at the requested index.
- Parameters
-
index The output tensor index in the valid range between 0 and getNbOutputs()-1. inputTypes The data types of the input tensors, stored in an array of length nbInputs. nbInputs The number of input tensors. Will be a non-negative integer.
- Returns
- The data type of the output tensor with the provided index if the input tensors have the data types provided in inputTypes, provided the output tensor index is in the valid range. DataType::kFLOAT must be returned if the index is not in the valid range.
The default behavior must be to return the type of the first input, or DataType::kFLOAT if the layer has no inputs. The returned data type must have a format that is supported by the plugin.
- See also
- supportsFormat()
- Warning
- DataType:kBOOL and DataType::kUINT8 are not supported.
Usage considerations
- Allowed context for the API call
- Thread-safe: Yes, this method is required to be thread-safe and may be called from multiple threads when building networks on multiple devices sharing the same plugin.
◆ getTensorRTVersion()
|
inlineoverrideprotectedvirtualnoexcept |
Return the API version with which this plugin was built. The upper byte reserved by TensorRT and is used to differentiate this from IPluginV2.
- Returns
- In the lower three bytes, the TensorRT version in the format (major * 100 + minor) * 100 + patch. In the upper byte, the value 1.
Do not override this method as it is used by the TensorRT library to maintain backwards-compatibility with plugins.
Usage considerations
- Allowed context for the API call
- Thread-safe: Yes, the implementation provided here is safe to call from any thread.
Reimplemented from nvinfer1::IPluginV2.
Reimplemented in nvinfer1::IPluginV2IOExt.
◆ isOutputBroadcastAcrossBatch()
|
pure virtualnoexcept |
Return true if the output tensor is broadcast across a batch.
- Parameters
-
outputIndex The index of the output tensor, which will be in the valid range between 0 and nbOutputs()-1. inputIsBroadcasted A boolean array of length nbInputs. The i-th element will be true if and only if the tensor for the ith input is broadcast across a batch. nbInputs The number of inputs. Will be a non-negative integer.
The values in inputIsBroadcasted refer to broadcasting at the semantic level, i.e. are unaffected by whether method canBroadcastInputAcrossBatch requests physical replication of the values.
Usage considerations
Allowed context for the API call
- Thread-safe: Yes, this method is required to be thread-safe and may be called from multiple threads when building networks on multiple devices sharing the same plugin.
- Deprecated:
- Deprecated in TensorRT 10.0. Implicit batch support is removed in TensorRT 10.0.
The documentation for this class was generated from the following file:
