NvInferRuntime.h

Go to the documentation of this file.

/*
 * Copyright 1993-2020 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 "NvInferRuntimeCommon.h"
 
namespace nvinfer1
{
 
class IExecutionContext; 
class ICudaEngine; 
class IPluginFactory; 
 
enum class EngineCapability : int32_t
{
    kDEFAULT = 0,  
    kSAFE_GPU = 1, 
    kSAFE_DLA = 2, 
};
 
template <>
constexpr inline int32_t EnumMax<EngineCapability>()
{
    return 3;
}
 
class Weights
{
public:
    DataType type;      
    const void* values; 
    int64_t count;      
};
 
class IHostMemory
{
public:
    virtual void* data() const noexcept = 0;       
    virtual std::size_t size() const noexcept = 0; 
    virtual DataType type() const noexcept = 0;    
    virtual void destroy() noexcept = 0;           
protected:
    virtual ~IHostMemory() {}
};
 
class IPlugin
{
public:
    virtual int32_t getNbOutputs() const TRTNOEXCEPT = 0;
 
    virtual Dims getOutputDimensions(int32_t index, const Dims* inputs, int32_t nbInputDims) TRTNOEXCEPT = 0;
 
    virtual void configure(const Dims* inputDims, int32_t nbInputs, const Dims* outputDims, int32_t nbOutputs,
        int32_t maxBatchSize) TRTNOEXCEPT = 0;
 
    virtual int32_t initialize() TRTNOEXCEPT = 0;
 
    virtual void terminate() TRTNOEXCEPT = 0;
 
    virtual size_t getWorkspaceSize(int32_t maxBatchSize) const TRTNOEXCEPT = 0;
 
    virtual int32_t enqueue(int32_t batchSize, const void* const* inputs, void** outputs, void* workspace,
        cudaStream_t stream) TRTNOEXCEPT = 0;
 
    virtual size_t getSerializationSize() TRTNOEXCEPT = 0;
 
    virtual void serialize(void* buffer) TRTNOEXCEPT = 0;
 
    virtual ~IPlugin() {}
};
 
class IPluginExt : public IPlugin
{
public:
    virtual int32_t getTensorRTVersion() const TRTNOEXCEPT
    {
        return NV_TENSORRT_VERSION;
    }
 
    virtual bool supportsFormat(DataType type, PluginFormat format) const TRTNOEXCEPT = 0;
 
    virtual void configureWithFormat(const Dims* inputDims, int32_t nbInputs, const Dims* outputDims, int32_t nbOutputs,
        DataType type, PluginFormat format, int32_t maxBatchSize) TRTNOEXCEPT = 0;
 
    virtual ~IPluginExt() {}
 
protected:
    void configure(const Dims* /*inputDims*/, int32_t /*nbInputs*/, const Dims* /*outputDims*/, int32_t /*nbOutputs*/,
        int32_t /*maxBatchSize*/) _TENSORRT_FINAL TRTNOEXCEPT
    {
    }
};
 
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>()
{
    return 9;
}
 
class IDimensionExpr
{
public:
    virtual bool isConstant() const = 0;
 
    virtual int32_t getConstantValue() const = 0;
 
protected:
    virtual ~IDimensionExpr() {}
};
 
class IExprBuilder
{
public:
    virtual const IDimensionExpr* constant(int32_t value) = 0;
 
    virtual const IDimensionExpr* operation(DimensionOperation op, const IDimensionExpr& first, const IDimensionExpr& second) = 0;
 
protected:
    virtual ~IExprBuilder() {}
};
 
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 _TENSORRT_OVERRIDE TRTNOEXCEPT = 0;
 
    virtual DimsExprs getOutputDimensions(
        int32_t outputIndex, const DimsExprs* inputs, int32_t nbInputs, IExprBuilder& exprBuilder)
        = 0;
 
    static constexpr int32_t kFORMAT_COMBINATION_LIMIT = 100;
 
    virtual bool supportsFormatCombination(
        int32_t pos, const PluginTensorDesc* inOut, int32_t nbInputs, int32_t nbOutputs) TRTNOEXCEPT = 0;
 
    virtual void configurePlugin(const DynamicPluginTensorDesc* in, int32_t nbInputs,
        const DynamicPluginTensorDesc* out, int32_t nbOutputs) TRTNOEXCEPT = 0;
 
    virtual size_t getWorkspaceSize(const PluginTensorDesc* inputs, int32_t nbInputs, const PluginTensorDesc* outputs,
        int32_t nbOutputs) const TRTNOEXCEPT = 0;
 
    virtual int32_t enqueue(const PluginTensorDesc* inputDesc, const PluginTensorDesc* outputDesc,
        const void* const* inputs, void* const* outputs, void* workspace, cudaStream_t stream) TRTNOEXCEPT = 0;
 
protected:
    int32_t getTensorRTVersion() const _TENSORRT_OVERRIDE TRTNOEXCEPT
    {
        return (static_cast<int32_t>(PluginVersion::kV2_DYNAMICEXT) << 24 | (NV_TENSORRT_VERSION & 0xFFFFFF));
    }
 
    virtual ~IPluginV2DynamicExt() {}
 
    // Rest of the methods below are obsolete inherited methods, and marked final when using a C++11 compiler.
    // Derived classes should not override them.
 
    TRT_DEPRECATED
    Dims getOutputDimensions(
        int32_t /*index*/, const Dims* /*inputs*/, int32_t /*nbInputDims*/) _TENSORRT_FINAL TRTNOEXCEPT
    {
        return Dims{-1, {}, {}};
    }
 
    TRT_DEPRECATED
    bool isOutputBroadcastAcrossBatch(int32_t /*outputIndex*/, const bool* /*inputIsBroadcasted*/,
        int32_t /*nbInputs*/) const _TENSORRT_FINAL TRTNOEXCEPT
    {
        return false;
    }
 
    TRT_DEPRECATED
    bool canBroadcastInputAcrossBatch(int32_t /*inputIndex*/) const _TENSORRT_FINAL TRTNOEXCEPT
    {
        return true;
    }
 
    TRT_DEPRECATED
    bool supportsFormat(DataType /*type*/, PluginFormat /*format*/) const _TENSORRT_FINAL TRTNOEXCEPT
    {
        return false;
    }
 
    TRT_DEPRECATED
    void configurePlugin(const Dims* /*inputDims*/, int32_t /*nbInputs*/, const Dims* /*outputDims*/,
        int32_t /*nbOutputs*/, const DataType* /*inputTypes*/, const DataType* /*outputTypes*/,
        const bool* /*inputIsBroadcast*/, const bool* /*outputIsBroadcast*/, PluginFormat /*floatFormat*/,
        int32_t /*maxBatchSize*/) _TENSORRT_FINAL TRTNOEXCEPT
    {
    }
 
    TRT_DEPRECATED
    size_t getWorkspaceSize(int32_t /*maxBatchSize*/) const _TENSORRT_FINAL TRTNOEXCEPT
    {
        return 0;
    }
 
    TRT_DEPRECATED
    int32_t enqueue(int32_t /*batchSize*/, const void* const* /*inputs*/, void** /*outputs*/, void* /*workspace*/,
        cudaStream_t /*stream*/) _TENSORRT_FINAL TRTNOEXCEPT
    {
        return 1;
    }
};
 
class IProfiler
{
public:
    virtual void reportLayerTime(const char* layerName, float ms) TRTNOEXCEPT = 0;
 
    virtual ~IProfiler() {}
};
 
enum class WeightsRole : int32_t
{
    kKERNEL = 0,   
    kBIAS = 1,     
    kSHIFT = 2,    
    kSCALE = 3,    
    kCONSTANT = 4, 
};
 
template <>
constexpr inline int32_t EnumMax<WeightsRole>()
{
    return 5;
}
 
enum class DeviceType : int32_t
{
    kGPU, 
    kDLA, 
};
 
template <>
constexpr inline int32_t EnumMax<DeviceType>()
{
    return 2;
}
 
class IRuntime
{
public:
    virtual nvinfer1::ICudaEngine* deserializeCudaEngine(const void* blob, std::size_t size, IPluginFactory* pluginFactory) noexcept = 0;
 
    virtual void setDLACore(int32_t dlaCore) noexcept = 0;
 
    virtual int32_t getDLACore() const noexcept = 0;
 
    virtual int32_t getNbDLACores() const noexcept = 0;
 
    virtual void destroy() noexcept = 0;
 
protected:
    virtual ~IRuntime() {}
 
public:
    virtual void setGpuAllocator(IGpuAllocator* allocator) noexcept = 0;
 
    //
    virtual void setErrorRecorder(IErrorRecorder* recorder) noexcept = 0;
 
    virtual IErrorRecorder* getErrorRecorder() const noexcept = 0;
 
    nvinfer1::ICudaEngine* deserializeCudaEngine(const void* blob, std::size_t size) noexcept
    {
        return deserializeCudaEngine(blob, size, nullptr);
    }
};
 
class IRefitter
{
public:
    virtual bool setWeights(const char* layerName, WeightsRole role, Weights weights) TRTNOEXCEPT = 0;
 
    virtual bool refitCudaEngine() TRTNOEXCEPT = 0;
 
    virtual int32_t getMissing(int32_t size, const char** layerNames, WeightsRole* roles) TRTNOEXCEPT = 0;
 
    virtual int32_t getAll(int32_t size, const char** layerNames, WeightsRole* roles) TRTNOEXCEPT = 0;
 
    virtual void destroy() TRTNOEXCEPT = 0;
 
protected:
    virtual ~IRefitter() {}
 
public:
    virtual bool setDynamicRange(const char* tensorName, float min, float max) TRTNOEXCEPT = 0;
 
    virtual float getDynamicRangeMin(const char* tensorName) const TRTNOEXCEPT = 0;
 
    virtual float getDynamicRangeMax(const char* tensorName) const TRTNOEXCEPT = 0;
 
    virtual int32_t getTensorsWithDynamicRange(int32_t size, const char** tensorNames) const TRTNOEXCEPT = 0;
 
    //
    virtual void setErrorRecorder(IErrorRecorder* recorder) TRTNOEXCEPT = 0;
 
    virtual IErrorRecorder* getErrorRecorder() const TRTNOEXCEPT = 0;
};
 
class IPluginFactory
{
public:
    virtual IPlugin* createPlugin(const char* layerName, const void* serialData, size_t serialLength) TRTNOEXCEPT = 0;
 
    virtual ~IPluginFactory() {}
};
 
enum class OptProfileSelector : int32_t
{
    kMIN = 0, 
    kOPT = 1, 
    kMAX = 2  
};
 
template <>
constexpr inline int32_t EnumMax<OptProfileSelector>()
{
    return 3;
}
 
class IOptimizationProfile
{
public:
    virtual bool setDimensions(const char* inputName, OptProfileSelector select, Dims dims) noexcept = 0;
 
    virtual Dims getDimensions(const char* inputName, OptProfileSelector select) const noexcept = 0;
 
    virtual bool setShapeValues(
        const char* inputName, OptProfileSelector select, const int32_t* values, int32_t nbValues) noexcept
        = 0;
 
    virtual int32_t getNbShapeValues(const char* inputName) const noexcept = 0;
 
    virtual const int32_t* getShapeValues(const char* inputName, OptProfileSelector select) const noexcept = 0;
 
    virtual bool setExtraMemoryTarget(float target) noexcept = 0;
 
    virtual float getExtraMemoryTarget() const noexcept = 0;
 
    virtual bool isValid() const noexcept = 0;
 
protected:
    ~IOptimizationProfile() noexcept = default;
};
 
class ICudaEngine
{
public:
    virtual int32_t getNbBindings() const noexcept = 0;
 
    virtual int32_t getBindingIndex(const char* name) const noexcept = 0;
 
    virtual const char* getBindingName(int32_t bindingIndex) const noexcept = 0;
 
    virtual bool bindingIsInput(int32_t bindingIndex) const noexcept = 0;
 
    virtual Dims getBindingDimensions(int32_t bindingIndex) const noexcept = 0;
 
    virtual DataType getBindingDataType(int32_t bindingIndex) const noexcept = 0;
 
    virtual int32_t getMaxBatchSize() const noexcept = 0;
 
    virtual int32_t getNbLayers() const noexcept = 0;
 
    TRT_DEPRECATED
    virtual std::size_t getWorkspaceSize() const noexcept = 0;
 
    virtual IHostMemory* serialize() const noexcept = 0;
 
    virtual IExecutionContext* createExecutionContext() noexcept = 0;
 
    virtual void destroy() noexcept = 0;
 
    virtual TensorLocation getLocation(int32_t bindingIndex) const noexcept = 0;
 
protected:
    virtual ~ICudaEngine() {}
 
public:
    virtual IExecutionContext* createExecutionContextWithoutDeviceMemory() noexcept = 0;
 
    virtual size_t getDeviceMemorySize() const noexcept = 0;
 
    virtual bool isRefittable() const noexcept = 0;
 
    virtual int32_t getBindingBytesPerComponent(int32_t bindingIndex) const noexcept = 0;
 
    virtual int32_t getBindingComponentsPerElement(int32_t bindingIndex) const noexcept = 0;
 
    virtual TensorFormat getBindingFormat(int32_t bindingIndex) const noexcept = 0;
 
    virtual const char* getBindingFormatDesc(int32_t bindingIndex) const noexcept = 0;
 
    virtual int32_t getBindingVectorizedDim(int32_t bindingIndex) const noexcept = 0;
 
    virtual const char* getName() const noexcept = 0;
 
    virtual int32_t getNbOptimizationProfiles() const noexcept = 0;
 
    virtual Dims getProfileDimensions(int32_t bindingIndex, int32_t profileIndex, OptProfileSelector select) const
        noexcept
        = 0;
 
    virtual const int32_t* getProfileShapeValues(
        int32_t profileIndex, int32_t inputIndex, OptProfileSelector select) const noexcept
        = 0;
 
    virtual bool isShapeBinding(int32_t bindingIndex) const noexcept = 0;
 
    virtual bool isExecutionBinding(int32_t bindingIndex) const noexcept = 0;
 
    virtual EngineCapability getEngineCapability() const noexcept = 0;
 
    //
    virtual void setErrorRecorder(IErrorRecorder* recorder) noexcept = 0;
 
    virtual IErrorRecorder* getErrorRecorder() const noexcept = 0;
 
    virtual bool hasImplicitBatchDimension() const TRTNOEXCEPT = 0;
};
 
class IExecutionContext
{
public:
    virtual bool execute(int32_t batchSize, void** bindings) noexcept = 0;
 
    virtual bool enqueue(int32_t batchSize, void** bindings, cudaStream_t stream, cudaEvent_t* inputConsumed) noexcept
        = 0;
 
    virtual void setDebugSync(bool sync) noexcept = 0;
 
    virtual bool getDebugSync() const noexcept = 0;
 
    virtual void setProfiler(IProfiler*) noexcept = 0;
 
    virtual IProfiler* getProfiler() const noexcept = 0;
 
    virtual const ICudaEngine& getEngine() const noexcept = 0;
 
    virtual void destroy() noexcept = 0;
 
protected:
    virtual ~IExecutionContext() noexcept {}
 
public:
    virtual void setName(const char* name) noexcept = 0;
 
    virtual const char* getName() const noexcept = 0;
 
    virtual void setDeviceMemory(void* memory) noexcept = 0;
 
    virtual Dims getStrides(int32_t bindingIndex) const noexcept = 0;
 
public:
    TRT_DEPRECATED
    virtual bool setOptimizationProfile(int32_t profileIndex) noexcept = 0;
 
    virtual int32_t getOptimizationProfile() const noexcept = 0;
 
    virtual bool setBindingDimensions(int32_t bindingIndex, Dims dimensions) noexcept = 0;
 
    virtual Dims getBindingDimensions(int32_t bindingIndex) const noexcept = 0;
 
    virtual bool setInputShapeBinding(int32_t bindingIndex, const int32_t* data) noexcept = 0;
 
    virtual bool getShapeBinding(int32_t bindingIndex, int32_t* data) const noexcept = 0;
 
    virtual bool allInputDimensionsSpecified() const noexcept = 0;
 
    virtual bool allInputShapesSpecified() const noexcept = 0;
 
    //
    virtual void setErrorRecorder(IErrorRecorder* recorder) noexcept = 0;
 
    virtual IErrorRecorder* getErrorRecorder() const noexcept = 0;
 
    virtual bool executeV2(void** bindings) noexcept = 0;
 
    virtual bool enqueueV2(void** bindings, cudaStream_t stream, cudaEvent_t* inputConsumed) noexcept = 0;
 
    virtual bool setOptimizationProfileAsync(int32_t profileIndex, cudaStream_t stream) noexcept = 0;
}; // class IExecutionContext
} // namespace nvinfer1
 
extern "C" TENSORRTAPI void* createInferRuntime_INTERNAL(void* logger, int32_t version);
 
extern "C" TENSORRTAPI void* createInferRefitter_INTERNAL(void* engine, void* logger, int32_t version);
 
namespace nvinfer1
{
namespace // unnamed namespace avoids linkage surprises when linking objects built with different versions of this header.
{
inline IRuntime* createInferRuntime(ILogger& logger)
{
    return static_cast<IRuntime*>(createInferRuntime_INTERNAL(&logger, NV_TENSORRT_VERSION));
}
 
inline IRefitter* createInferRefitter(ICudaEngine& engine, ILogger& logger)
{
    return static_cast<IRefitter*>(createInferRefitter_INTERNAL(&engine, &logger, NV_TENSORRT_VERSION));
}
}
}
 
#endif // NV_INFER_RUNTIME_H