NvInferRuntime.h

Go to the documentation of this file.

/*
 * Copyright 1993-2021 NVIDIA Corporation.  All rights reserved.
 *
 * NOTICE TO LICENSEE:
 *
 * This source code and/or documentation ("Licensed Deliverables") are
 * subject to NVIDIA intellectual property rights under U.S. and
 * international Copyright laws.
 *
 * These Licensed Deliverables contained herein is PROPRIETARY and
 * CONFIDENTIAL to NVIDIA and is being provided under the terms and
 * conditions of a form of NVIDIA software license agreement by and
 * between NVIDIA and Licensee ("License Agreement") or electronically
 * accepted by Licensee.  Notwithstanding any terms or conditions to
 * the contrary in the License Agreement, reproduction or disclosure
 * of the Licensed Deliverables to any third party without the express
 * written consent of NVIDIA is prohibited.
 *
 * NOTWITHSTANDING ANY TERMS OR CONDITIONS TO THE CONTRARY IN THE
 * LICENSE AGREEMENT, NVIDIA MAKES NO REPRESENTATION ABOUT THE
 * SUITABILITY OF THESE LICENSED DELIVERABLES FOR ANY PURPOSE.  IT IS
 * PROVIDED "AS IS" WITHOUT EXPRESS OR IMPLIED WARRANTY OF ANY KIND.
 * NVIDIA DISCLAIMS ALL WARRANTIES WITH REGARD TO THESE LICENSED
 * DELIVERABLES, INCLUDING ALL IMPLIED WARRANTIES OF MERCHANTABILITY,
 * NONINFRINGEMENT, AND FITNESS FOR A PARTICULAR PURPOSE.
 * NOTWITHSTANDING ANY TERMS OR CONDITIONS TO THE CONTRARY IN THE
 * LICENSE AGREEMENT, IN NO EVENT SHALL NVIDIA BE LIABLE FOR ANY
 * SPECIAL, INDIRECT, INCIDENTAL, OR CONSEQUENTIAL DAMAGES, OR ANY
 * DAMAGES WHATSOEVER RESULTING FROM LOSS OF USE, DATA OR PROFITS,
 * WHETHER IN AN ACTION OF CONTRACT, NEGLIGENCE OR OTHER TORTIOUS
 * ACTION, ARISING OUT OF OR IN CONNECTION WITH THE USE OR PERFORMANCE
 * OF THESE LICENSED DELIVERABLES.
 *
 * U.S. Government End Users.  These Licensed Deliverables are a
 * "commercial item" as that term is defined at 48 C.F.R. 2.101 (OCT
 * 1995), consisting of "commercial computer software" and "commercial
 * computer software documentation" as such terms are used in 48
 * C.F.R. 12.212 (SEPT 1995) and is provided to the U.S. Government
 * only as a commercial end item.  Consistent with 48 C.F.R.12.212 and
 * 48 C.F.R. 227.7202-1 through 227.7202-4 (JUNE 1995), all
 * U.S. Government End Users acquire the Licensed Deliverables with
 * only those rights set forth herein.
 *
 * Any use of the Licensed Deliverables in individual and commercial
 * software must include, in the user documentation and internal
 * comments to the code, the above Disclaimer and U.S. Government End
 * Users Notice.
 */
 
#ifndef NV_INFER_RUNTIME_H
#define NV_INFER_RUNTIME_H
 
 
#include "NvInferImpl.h"
#include "NvInferRuntimeCommon.h"
 
namespace nvinfer1
{
 
class IExecutionContext; 
class ICudaEngine;       
class IPluginFactory;    
 
 
class INoCopy
{
protected:
    INoCopy() = default;
    virtual ~INoCopy() = default;
    INoCopy(const INoCopy& other) = delete;
    INoCopy& operator=(const INoCopy& other) = delete;
    INoCopy(INoCopy&& other) = delete;
    INoCopy& operator=(INoCopy&& other) = delete;
};
 
enum class EngineCapability : int32_t
{
    kDEFAULT = 0,
 
    kSAFE_GPU = 1,
 
    kSAFE_DLA = 2,
};
 
namespace impl
{
template <>
struct EnumMaxImpl<EngineCapability>
{
    static constexpr int32_t kVALUE = 3;
};
} // namespace impl
 
class Weights
{
public:
    DataType type;      
    const void* values; 
    int64_t count;      
};
 
class IHostMemory : public INoCopy
{
public:
    virtual ~IHostMemory() noexcept = default;
 
    void* data() const noexcept
    {
        return mImpl->data();
    }
 
    std::size_t size() const noexcept
    {
        return mImpl->size();
    }
 
    DataType type() const noexcept
    {
        return mImpl->type();
    }
    TRT_DEPRECATED void destroy() noexcept
    {
        delete this;
    }
 
protected:
    apiv::VHostMemory* mImpl;
};
 
enum class DimensionOperation : int32_t
{
    kSUM = 0,       
    kPROD = 1,      
    kMAX = 2,       
    kMIN = 3,       
    kSUB = 4,       
    kEQUAL = 5,     
    kLESS = 6,      
    kFLOOR_DIV = 7, 
    kCEIL_DIV = 8   
};
 
template <>
constexpr inline int32_t EnumMax<DimensionOperation>() noexcept
{
    return 9;
}
 
enum class TensorLocation : int32_t
{
    kDEVICE = 0, 
    kHOST = 1,   
};
 
namespace impl
{
template <>
struct EnumMaxImpl<TensorLocation>
{
    static constexpr int32_t kVALUE = 2;
};
} // namespace impl
 
class IDimensionExpr : public INoCopy
{
public:
    bool isConstant() const noexcept
    {
        return mImpl->isConstant();
    }
 
    int32_t getConstantValue() const noexcept
    {
        return mImpl->getConstantValue();
    }
 
protected:
    apiv::VDimensionExpr* mImpl;
    virtual ~IDimensionExpr() noexcept = default;
};
 
class IExprBuilder : public INoCopy
{
public:
    const IDimensionExpr* constant(int32_t value) noexcept
    {
        return mImpl->constant(value);
    }
 
    const IDimensionExpr* operation(
        DimensionOperation op, const IDimensionExpr& first, const IDimensionExpr& second) noexcept
    {
        return mImpl->operation(op, first, second);
    }
 
protected:
    apiv::VExprBuilder* mImpl;
    virtual ~IExprBuilder() noexcept = default;
};
 
class DimsExprs
{
public:
    int32_t nbDims;                          
    const IDimensionExpr* d[Dims::MAX_DIMS]; 
};
 
struct DynamicPluginTensorDesc
{
    PluginTensorDesc desc;
 
    Dims min;
 
    Dims max;
};
 
class IPluginV2DynamicExt : public nvinfer1::IPluginV2Ext
{
public:
    IPluginV2DynamicExt* clone() const noexcept override = 0;
 
    virtual DimsExprs getOutputDimensions(
        int32_t outputIndex, const DimsExprs* inputs, int32_t nbInputs, IExprBuilder& exprBuilder) noexcept
        = 0;
 
    static constexpr int32_t kFORMAT_COMBINATION_LIMIT = 100;
 
    virtual bool supportsFormatCombination(
        int32_t pos, const PluginTensorDesc* inOut, int32_t nbInputs, int32_t nbOutputs) noexcept
        = 0;
 
    virtual void configurePlugin(const DynamicPluginTensorDesc* in, int32_t nbInputs,
        const DynamicPluginTensorDesc* out, int32_t nbOutputs) noexcept
        = 0;
 
    virtual size_t getWorkspaceSize(const PluginTensorDesc* inputs, int32_t nbInputs, const PluginTensorDesc* outputs,
        int32_t nbOutputs) const noexcept
        = 0;
 
    virtual int32_t enqueue(const PluginTensorDesc* inputDesc, const PluginTensorDesc* outputDesc,
        const void* const* inputs, void* const* outputs, void* workspace, cudaStream_t stream) noexcept
        = 0;
 
protected:
    int32_t getTensorRTVersion() const noexcept override
    {
        return (static_cast<int32_t>(PluginVersion::kV2_DYNAMICEXT) << 24 | (NV_TENSORRT_VERSION & 0xFFFFFF));
    }
 
    virtual ~IPluginV2DynamicExt() noexcept {}
 
private:
    // Following are obsolete base class methods, and must not be implemented or used.
 
    void configurePlugin(Dims const*, int32_t, Dims const*, int32_t, DataType const*, DataType const*, bool const*,
        bool const*, PluginFormat, int32_t) noexcept override final
    {
    }
 
    bool supportsFormat(DataType, PluginFormat) const noexcept override final
    {
        return false;
    }
 
    Dims getOutputDimensions(int32_t, Dims const*, int32_t) noexcept override final
    {
        return Dims{-1, {}};
    }
 
    bool isOutputBroadcastAcrossBatch(int32_t, bool const*, int32_t) const noexcept override final
    {
        return false;
    }
 
    bool canBroadcastInputAcrossBatch(int32_t) const noexcept override final
    {
        return true;
    }
 
    size_t getWorkspaceSize(int32_t) const noexcept override final
    {
        return 0;
    }
 
    int32_t enqueue(int32_t, const void* const*, void* const*, void*, cudaStream_t) noexcept override final
    {
        return 1;
    }
};
 
class IProfiler
{
public:
    virtual void reportLayerTime(const char* layerName, float ms) noexcept = 0;
 
    virtual ~IProfiler() noexcept {}
};
 
enum class WeightsRole : int32_t
{
    kKERNEL = 0,   
    kBIAS = 1,     
    kSHIFT = 2,    
    kSCALE = 3,    
    kCONSTANT = 4, 
    kANY = 5,      
};
 
template <>
constexpr inline int32_t EnumMax<WeightsRole>() noexcept
{
    return 5;
}
 
enum class DeviceType : int32_t
{
    kGPU, 
    kDLA, 
};
 
template <>
constexpr inline int32_t EnumMax<DeviceType>() noexcept
{
    return 2;
}
 
class IRuntime : public INoCopy
{
public:
    virtual ~IRuntime() noexcept = default;
 
    TRT_DEPRECATED nvinfer1::ICudaEngine* deserializeCudaEngine(
        const void* blob, std::size_t size, IPluginFactory* pluginFactory) noexcept
    {
        return mImpl->deserializeCudaEngine(blob, size, nullptr);
    }
 
    void setDLACore(int32_t dlaCore) noexcept
    {
        mImpl->setDLACore(dlaCore);
    }
 
    int32_t getDLACore() const noexcept
    {
        return mImpl->getDLACore();
    }
 
    int32_t getNbDLACores() const noexcept
    {
        return mImpl->getNbDLACores();
    }
 
    TRT_DEPRECATED void destroy() noexcept
    {
        delete this;
    }
 
    void setGpuAllocator(IGpuAllocator* allocator) noexcept
    {
        mImpl->setGpuAllocator(allocator);
    }
 
    //
    void setErrorRecorder(IErrorRecorder* recorder) noexcept
    {
        mImpl->setErrorRecorder(recorder);
    }
 
    IErrorRecorder* getErrorRecorder() const noexcept
    {
        return mImpl->getErrorRecorder();
    }
 
    nvinfer1::ICudaEngine* deserializeCudaEngine(const void* blob, std::size_t size) noexcept
    {
        return mImpl->deserializeCudaEngine(blob, size, nullptr);
    }
 
protected:
    apiv::VRuntime* mImpl;
};
 
class IRefitter : public INoCopy
{
public:
    virtual ~IRefitter() noexcept = default;
 
    bool setWeights(const char* layerName, WeightsRole role, Weights weights) noexcept
    {
        return mImpl->setWeights(layerName, role, weights);
    }
 
    bool refitCudaEngine() noexcept
    {
        return mImpl->refitCudaEngine();
    }
 
    int32_t getMissing(int32_t size, const char** layerNames, WeightsRole* roles) noexcept
    {
        return mImpl->getMissing(size, layerNames, roles);
    }
 
    int32_t getAll(int32_t size, const char** layerNames, WeightsRole* roles) noexcept
    {
        return mImpl->getAll(size, layerNames, roles);
    }
 
    TRT_DEPRECATED void destroy() noexcept
    {
        delete this;
    }
 
    bool setDynamicRange(const char* tensorName, float min, float max) noexcept
    {
        return mImpl->setDynamicRange(tensorName, min, max);
    }
 
    float getDynamicRangeMin(const char* tensorName) const noexcept
    {
        return mImpl->getDynamicRangeMin(tensorName);
    }
 
    float getDynamicRangeMax(const char* tensorName) const noexcept
    {
        return mImpl->getDynamicRangeMax(tensorName);
    }
 
    int32_t getTensorsWithDynamicRange(int32_t size, const char** tensorNames) const noexcept
    {
        return mImpl->getTensorsWithDynamicRange(size, tensorNames);
    }
 
    //
    void setErrorRecorder(IErrorRecorder* recorder) noexcept
    {
        mImpl->setErrorRecorder(recorder);
    }
 
    IErrorRecorder* getErrorRecorder() const noexcept
    {
        return mImpl->getErrorRecorder();
    }
 
    bool setNamedWeights(const char* name, Weights weights) noexcept
    {
        return mImpl->setNamedWeights(name, weights);
    }
 
    int32_t getMissingWeights(int32_t size, const char** weightsNames) noexcept
    {
        return mImpl->getMissingWeights(size, weightsNames);
    }
 
    int32_t getAllWeights(int32_t size, const char** weightsNames) noexcept
    {
        return mImpl->getAllWeights(size, weightsNames);
    }
 
protected:
    apiv::VRefitter* mImpl;
};
 
enum class OptProfileSelector : int32_t
{
    kMIN = 0, 
    kOPT = 1, 
    kMAX = 2  
};
 
template <>
constexpr inline int32_t EnumMax<OptProfileSelector>() noexcept
{
    return 3;
}
 
class IOptimizationProfile : public INoCopy
{
public:
    bool setDimensions(const char* inputName, OptProfileSelector select, Dims dims) noexcept
    {
        return mImpl->setDimensions(inputName, select, dims);
    }
 
    Dims getDimensions(const char* inputName, OptProfileSelector select) const noexcept
    {
        return mImpl->getDimensions(inputName, select);
    }
 
    bool setShapeValues(
        const char* inputName, OptProfileSelector select, const int32_t* values, int32_t nbValues) noexcept
    {
        return mImpl->setShapeValues(inputName, select, values, nbValues);
    }
 
    int32_t getNbShapeValues(const char* inputName) const noexcept
    {
        return mImpl->getNbShapeValues(inputName);
    }
 
    int32_t const* getShapeValues(const char* inputName, OptProfileSelector select) const noexcept
    {
        return mImpl->getShapeValues(inputName, select);
    }
 
    bool setExtraMemoryTarget(float target) noexcept
    {
        return mImpl->setExtraMemoryTarget(target);
    }
 
    float getExtraMemoryTarget() const noexcept
    {
        return mImpl->getExtraMemoryTarget();
    }
 
    bool isValid() const noexcept
    {
        return mImpl->isValid();
    }
 
protected:
    apiv::VOptimizationProfile* mImpl;
    virtual ~IOptimizationProfile() noexcept = default;
};
 
enum class TacticSource : int32_t
{
    kCUBLAS = 0,    
    kCUBLAS_LT = 1, 
    kCUDNN = 2      
};
 
template <>
constexpr inline int32_t EnumMax<TacticSource>() noexcept
{
    return 3;
} 
 
using TacticSources = uint32_t;
 
class ICudaEngine : public INoCopy
{
public:
    virtual ~ICudaEngine() noexcept = default;
 
    int32_t getNbBindings() const noexcept
    {
        return mImpl->getNbBindings();
    }
 
    int32_t getBindingIndex(const char* name) const noexcept
    {
        return mImpl->getBindingIndex(name);
    }
 
    const char* getBindingName(int32_t bindingIndex) const noexcept
    {
        return mImpl->getBindingName(bindingIndex);
    }
 
    bool bindingIsInput(int32_t bindingIndex) const noexcept
    {
        return mImpl->bindingIsInput(bindingIndex);
    }
 
    Dims getBindingDimensions(int32_t bindingIndex) const noexcept
    {
        return mImpl->getBindingDimensions(bindingIndex);
    }
 
    DataType getBindingDataType(int32_t bindingIndex) const noexcept
    {
        return mImpl->getBindingDataType(bindingIndex);
    }
 
    int32_t getMaxBatchSize() const noexcept
    {
        return mImpl->getMaxBatchSize();
    }
 
    int32_t getNbLayers() const noexcept
    {
        return mImpl->getNbLayers();
    }
 
    IHostMemory* serialize() const noexcept
    {
        return mImpl->serialize();
    }
 
    IExecutionContext* createExecutionContext() noexcept
    {
        return mImpl->createExecutionContext();
    }
 
    TRT_DEPRECATED void destroy() noexcept
    {
        delete this;
    }
 
    TensorLocation getLocation(int32_t bindingIndex) const noexcept
    {
        return mImpl->getLocation(bindingIndex);
    }
 
    IExecutionContext* createExecutionContextWithoutDeviceMemory() noexcept
    {
        return mImpl->createExecutionContextWithoutDeviceMemory();
    }
 
    size_t getDeviceMemorySize() const noexcept
    {
        return mImpl->getDeviceMemorySize();
    }
 
    bool isRefittable() const noexcept
    {
        return mImpl->isRefittable();
    }
 
    int32_t getBindingBytesPerComponent(int32_t bindingIndex) const noexcept
    {
        return mImpl->getBindingBytesPerComponent(bindingIndex);
    }
 
    int32_t getBindingComponentsPerElement(int32_t bindingIndex) const noexcept
    {
        return mImpl->getBindingComponentsPerElement(bindingIndex);
    }
 
    TensorFormat getBindingFormat(int32_t bindingIndex) const noexcept
    {
        return mImpl->getBindingFormat(bindingIndex);
    }
 
    const char* getBindingFormatDesc(int32_t bindingIndex) const noexcept
    {
        return mImpl->getBindingFormatDesc(bindingIndex);
    }
 
    int32_t getBindingVectorizedDim(int32_t bindingIndex) const noexcept
    {
        return mImpl->getBindingVectorizedDim(bindingIndex);
    }
 
    const char* getName() const noexcept
    {
        return mImpl->getName();
    }
 
    int32_t getNbOptimizationProfiles() const noexcept
    {
        return mImpl->getNbOptimizationProfiles();
    }
 
    Dims getProfileDimensions(int32_t bindingIndex, int32_t profileIndex, OptProfileSelector select) const noexcept
    {
        return mImpl->getProfileDimensions(bindingIndex, profileIndex, select);
    }
 
    const int32_t* getProfileShapeValues(int32_t profileIndex, int32_t inputIndex, OptProfileSelector select) const
        noexcept
    {
        return mImpl->getProfileShapeValues(profileIndex, inputIndex, select);
    }
 
    bool isShapeBinding(int32_t bindingIndex) const noexcept
    {
        return mImpl->isShapeBinding(bindingIndex);
    }
 
    bool isExecutionBinding(int32_t bindingIndex) const noexcept
    {
        return mImpl->isExecutionBinding(bindingIndex);
    }
 
    EngineCapability getEngineCapability() const noexcept
    {
        return mImpl->getEngineCapability();
    }
 
    //
    void setErrorRecorder(IErrorRecorder* recorder) noexcept
    {
        return mImpl->setErrorRecorder(recorder);
    }
 
    IErrorRecorder* getErrorRecorder() const noexcept
    {
        return mImpl->getErrorRecorder();
    }
 
    bool hasImplicitBatchDimension() const noexcept
    {
        return mImpl->hasImplicitBatchDimension();
    }
 
    TacticSources getTacticSources() const noexcept
    {
        return mImpl->getTacticSources();
    }
 
protected:
    apiv::VCudaEngine* mImpl;
};
 
class IExecutionContext : public INoCopy
{
public:
    virtual ~IExecutionContext() noexcept = default;
 
    bool execute(int32_t batchSize, void* const* bindings) noexcept
    {
        return mImpl->execute(batchSize, bindings);
    }
 
    bool enqueue(int32_t batchSize, void* const* bindings, cudaStream_t stream, cudaEvent_t* inputConsumed) noexcept
    {
        return mImpl->enqueue(batchSize, bindings, stream, inputConsumed);
    }
 
    void setDebugSync(bool sync) noexcept
    {
        mImpl->setDebugSync(sync);
    }
 
    bool getDebugSync() const noexcept
    {
        return mImpl->getDebugSync();
    }
 
    void setProfiler(IProfiler* profiler) noexcept
    {
        mImpl->setProfiler(profiler);
    }
 
    IProfiler* getProfiler() const noexcept
    {
        return mImpl->getProfiler();
    }
 
    const ICudaEngine& getEngine() const noexcept
    {
        return mImpl->getEngine();
    }
 
    TRT_DEPRECATED void destroy() noexcept
    {
        delete this;
    }
 
    void setName(const char* name) noexcept
    {
        mImpl->setName(name);
    }
 
    const char* getName() const noexcept
    {
        return mImpl->getName();
    }
 
 
    TRT_DEPRECATED
    void setDeviceMemory(void* memory) noexcept
    {
        mImpl->setDeviceMemory(memory);
    }
 
    Dims getStrides(int32_t bindingIndex) const noexcept
    {
        return mImpl->getStrides(bindingIndex);
    }
 
public:
    TRT_DEPRECATED
    bool setOptimizationProfile(int32_t profileIndex) noexcept
    {
        return mImpl->setOptimizationProfile(profileIndex);
    }
 
    int32_t getOptimizationProfile() const noexcept
    {
        return mImpl->getOptimizationProfile();
    }
 
    bool setBindingDimensions(int32_t bindingIndex, Dims dimensions) noexcept
    {
        return mImpl->setBindingDimensions(bindingIndex, dimensions);
    }
 
    Dims getBindingDimensions(int32_t bindingIndex) const noexcept
    {
        return mImpl->getBindingDimensions(bindingIndex);
    }
 
    bool setInputShapeBinding(int32_t bindingIndex, int32_t const* data) noexcept
    {
        return mImpl->setInputShapeBinding(bindingIndex, data);
    }
 
    bool getShapeBinding(int32_t bindingIndex, int32_t* data) const noexcept
    {
        return mImpl->getShapeBinding(bindingIndex, data);
    }
 
    bool allInputDimensionsSpecified() const noexcept
    {
        return mImpl->allInputDimensionsSpecified();
    }
 
    bool allInputShapesSpecified() const noexcept
 
    {
        return mImpl->allInputShapesSpecified();
    }
 
    //
    void setErrorRecorder(IErrorRecorder* recorder) noexcept
    {
        mImpl->setErrorRecorder(recorder);
    }
 
    IErrorRecorder* getErrorRecorder() const noexcept
    {
        return mImpl->getErrorRecorder();
    }
 
    bool executeV2(void* const* bindings) noexcept
    {
        return mImpl->executeV2(bindings);
    }
 
    bool enqueueV2(void* const* bindings, cudaStream_t stream, cudaEvent_t* inputConsumed) noexcept
    {
        return mImpl->enqueueV2(bindings, stream, inputConsumed);
    }
 
    bool setOptimizationProfileAsync(int32_t profileIndex, cudaStream_t stream) noexcept
    {
        return mImpl->setOptimizationProfileAsync(profileIndex, stream);
    }
 
protected:
    apiv::VExecutionContext* mImpl;
}; // class IExecutionContext
} // namespace nvinfer1
 
extern "C" TENSORRTAPI void* createInferRuntime_INTERNAL(void* logger, int32_t version) noexcept;
 
extern "C" TENSORRTAPI void* createInferRefitter_INTERNAL(void* engine, void* logger, int32_t version) noexcept;
 
extern "C" TENSORRTAPI nvinfer1::IPluginRegistry* getPluginRegistry() noexcept;
 
namespace nvinfer1
{
namespace // unnamed namespace avoids linkage surprises when linking objects built with different versions of this
          // header.
{
inline IRuntime* createInferRuntime(ILogger& logger) noexcept
{
    return static_cast<IRuntime*>(createInferRuntime_INTERNAL(&logger, NV_TENSORRT_VERSION));
}
 
inline IRefitter* createInferRefitter(ICudaEngine& engine, ILogger& logger) noexcept
{
    return static_cast<IRefitter*>(createInferRefitter_INTERNAL(&engine, &logger, NV_TENSORRT_VERSION));
}
 
} // namespace
 
template <typename T>
class PluginRegistrar
{
public:
    PluginRegistrar()
    {
        getPluginRegistry()->registerCreator(instance, "");
    }
 
private:
    T instance{};
};
 
} // namespace nvinfer1
 
#define REGISTER_TENSORRT_PLUGIN(name)                                                                                 \
    static nvinfer1::PluginRegistrar<name> pluginRegistrar##name {}
#endif // NV_INFER_RUNTIME_H