tensorNet.h

Go to the documentation of this file.

/*
 * Copyright (c) 2017, NVIDIA CORPORATION. All rights reserved.
 *
 * Permission is hereby granted, free of charge, to any person obtaining a
 * copy of this software and associated documentation files (the "Software"),
 * to deal in the Software without restriction, including without limitation
 * the rights to use, copy, modify, merge, publish, distribute, sublicense,
 * and/or sell copies of the Software, and to permit persons to whom the
 * Software is furnished to do so, subject to the following conditions:
 *
 * The above copyright notice and this permission notice shall be included in
 * all copies or substantial portions of the Software.
 *
 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
 * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
 * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
 * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER
 * DEALINGS IN THE SOFTWARE.
 */
 
#ifndef __TENSOR_NET_H__
#define __TENSOR_NET_H__

// forward declaration of IInt8Calibrator
namespace nvinfer1 { class IInt8Calibrator; }

// includes
#include <NvInfer.h>

#include <jetson-utils/cudaUtility.h>
#include <jetson-utils/timespec.h>

#include <vector>
#include <sstream>
#include <math.h>


#if NV_TENSORRT_MAJOR > 1
typedef nvinfer1::DimsCHW Dims3;

#define DIMS_C(x) x.d[0]
#define DIMS_H(x) x.d[1]
#define DIMS_W(x) x.d[2]

#else
typedef nvinfer1::Dims3 Dims3; 

#define DIMS_C(x) x.c
#define DIMS_H(x) x.h
#define DIMS_W(x) x.w

#ifndef NV_TENSORRT_MAJOR
#define NV_TENSORRT_MAJOR 1
#define NV_TENSORRT_MINOR 0
#endif
#endif


#define DEFAULT_MAX_BATCH_SIZE  1

#define LOG_TRT "[TRT]   "


enum precisionType
{
        TYPE_DISABLED = 0,      
        TYPE_FASTEST,           
        TYPE_FP32,              
        TYPE_FP16,              
        TYPE_INT8,              
        NUM_PRECISIONS          
};

const char* precisionTypeToStr( precisionType type );

precisionType precisionTypeFromStr( const char* str );

enum deviceType
{
        DEVICE_GPU = 0,                 
        DEVICE_DLA,                             
        DEVICE_DLA_0 = DEVICE_DLA,      
        DEVICE_DLA_1,                           
        NUM_DEVICES                             
};

const char* deviceTypeToStr( deviceType type );

deviceType deviceTypeFromStr( const char* str );

enum modelType
{
        MODEL_CUSTOM = 0,       
        MODEL_CAFFE,            
        MODEL_ONNX,             
        MODEL_UFF                       
};

const char* modelTypeToStr( modelType type );

modelType modelTypeFromStr( const char* str );

enum profilerQuery
{
        PROFILER_PREPROCESS = 0,
        PROFILER_NETWORK,
        PROFILER_POSTPROCESS,
        PROFILER_VISUALIZE,
        PROFILER_TOTAL,
};

const char* profilerQueryToStr( profilerQuery query );

enum profilerDevice
{
        PROFILER_CPU = 0,       
        PROFILER_CUDA,          
};


class tensorNet
{
public:
        virtual ~tensorNet();
        
        bool LoadNetwork( const char* prototxt, const char* model, const char* mean=NULL,
                                   const char* input_blob="data", const char* output_blob="prob",
                                   uint32_t maxBatchSize=DEFAULT_MAX_BATCH_SIZE, precisionType precision=TYPE_FASTEST,
                                   deviceType device=DEVICE_GPU, bool allowGPUFallback=true,
                                   nvinfer1::IInt8Calibrator* calibrator=NULL, cudaStream_t stream=NULL );

        bool LoadNetwork( const char* prototxt, const char* model, const char* mean,
                                   const char* input_blob, const std::vector<std::string>& output_blobs,
                                   uint32_t maxBatchSize=DEFAULT_MAX_BATCH_SIZE, precisionType precision=TYPE_FASTEST,
                                   deviceType device=DEVICE_GPU, bool allowGPUFallback=true,
                                   nvinfer1::IInt8Calibrator* calibrator=NULL, cudaStream_t stream=NULL );

        bool LoadNetwork( const char* prototxt, const char* model, const char* mean,
                                   const char* input_blob, const Dims3& input_dims, 
                                   const std::vector<std::string>& output_blobs,
                                   uint32_t maxBatchSize=DEFAULT_MAX_BATCH_SIZE, 
                                   precisionType precision=TYPE_FASTEST,
                                   deviceType device=DEVICE_GPU, bool allowGPUFallback=true,
                                   nvinfer1::IInt8Calibrator* calibrator=NULL, cudaStream_t stream=NULL );

        void EnableLayerProfiler();

        void EnableDebug();

        inline bool AllowGPUFallback() const                            { return mAllowGPUFallback; }

        inline deviceType GetDevice() const                             { return mDevice; }

        inline precisionType GetPrecision() const                       { return mPrecision; }

        inline bool IsPrecision( precisionType type ) const     { return (mPrecision == type); }

        static precisionType FindFastestPrecision( deviceType device=DEVICE_GPU, bool allowInt8=true );

        static std::vector<precisionType> DetectNativePrecisions( deviceType device=DEVICE_GPU );
        
        static bool DetectNativePrecision( const std::vector<precisionType>& nativeTypes, precisionType type );

        static bool DetectNativePrecision( precisionType precision, deviceType device=DEVICE_GPU );

        inline cudaStream_t GetStream() const                           { return mStream; }

        cudaStream_t CreateStream( bool nonBlocking=true );

        void SetStream( cudaStream_t stream );

        inline const char* GetPrototxtPath() const                      { return mPrototxtPath.c_str(); }

        inline const char* GetModelPath() const                         { return mModelPath.c_str(); }

        inline modelType GetModelType() const                           { return mModelType; }

        inline bool IsModelType( modelType type ) const         { return (mModelType == type); }

        inline float GetNetworkTime()                                                                           { return GetProfilerTime(PROFILER_NETWORK, PROFILER_CUDA); }
        
        inline float2 GetProfilerTime( profilerQuery query )                                     { PROFILER_QUERY(query); return mProfilerTimes[query]; }
        
        inline float GetProfilerTime( profilerQuery query, profilerDevice device ) { PROFILER_QUERY(query); return (device == PROFILER_CPU) ? mProfilerTimes[query].x : mProfilerTimes[query].y; }
        
        inline void PrintProfilerTimes()
        {
                printf("\n");
                printf(LOG_TRT "----------------------------------------------\n");
                printf(LOG_TRT "Timing Report %s\n", GetModelPath());
                printf(LOG_TRT "----------------------------------------------\n");

                for( uint32_t n=0; n <= PROFILER_TOTAL; n++ )
                {
                        const profilerQuery query = (profilerQuery)n;

                        if( PROFILER_QUERY(query) )
                                printf(LOG_TRT "%-12s  CPU %8.5fms  CUDA %8.5fms\n", profilerQueryToStr(query), mProfilerTimes[n].x, mProfilerTimes[n].y);
                }

                printf(LOG_TRT "----------------------------------------------\n\n");

                static bool first_run=true;

                if( first_run )
                {
                        printf(LOG_TRT "note -- when processing a single image, run 'sudo jetson_clocks' before\n"
                                  "                to disable DVFS for more accurate profiling/timing measurements\n\n");
                        
                        first_run = false;
                }
        }

protected:

        tensorNet();
                          
        bool ProfileModel( const std::string& deployFile, const std::string& modelFile,
                                        const char* input, const Dims3& inputDims,
                                    const std::vector<std::string>& outputs, uint32_t maxBatchSize, 
                                    precisionType precision, deviceType device, bool allowGPUFallback,
                                    nvinfer1::IInt8Calibrator* calibrator, std::ostream& modelStream);

        class Logger : public nvinfer1::ILogger                 
        {
                void log( Severity severity, const char* msg ) override
                {
                        if( severity != Severity::kINFO /*|| mEnableDebug*/ )
                                printf(LOG_TRT "%s\n", msg);
                }
        } gLogger;

        class Profiler : public nvinfer1::IProfiler
        {
        public:
                Profiler() : timingAccumulator(0.0f)    { }
                
                virtual void reportLayerTime(const char* layerName, float ms)
                {
                        printf(LOG_TRT "layer %s - %f ms\n", layerName, ms);
                        timingAccumulator += ms;
                }
                
                float timingAccumulator;
                
        } gProfiler;

        inline void PROFILER_BEGIN( profilerQuery query )               
        { 
                const uint32_t evt = query*2; 
                const uint32_t flag = (1 << query);

                CUDA(cudaEventRecord(mEventsGPU[evt], mStream)); 
                timestamp(&mEventsCPU[evt]); 

                mProfilerQueriesUsed |= flag;
                mProfilerQueriesDone &= ~flag;
        }

        inline void PROFILER_END( profilerQuery query )         
        { 
                const uint32_t evt = query*2+1; 

                CUDA(cudaEventRecord(mEventsGPU[evt])); 
                timestamp(&mEventsCPU[evt]); 
                timespec cpuTime; 
                timeDiff(mEventsCPU[evt-1], mEventsCPU[evt], &cpuTime);
                mProfilerTimes[query].x = timeFloat(cpuTime);

                if( mEnableProfiler && query == PROFILER_NETWORK ) 
                { 
                        printf(LOG_TRT "layer network time - %f ms\n", gProfiler.timingAccumulator); 
                        gProfiler.timingAccumulator = 0.0f; 
                        printf(LOG_TRT "note -- when processing a single image, run 'sudo jetson_clocks' before\n"
                                  "                to disable DVFS for more accurate profiling/timing measurements\n"); 
                }
        }
        
        inline bool PROFILER_QUERY( profilerQuery query )
        {
                const uint32_t flag = (1 << query);

                if( query == PROFILER_TOTAL )
                {
                        mProfilerTimes[PROFILER_TOTAL].x = 0.0f;
                        mProfilerTimes[PROFILER_TOTAL].y = 0.0f;

                        for( uint32_t n=0; n < PROFILER_TOTAL; n++ )
                        {
                                if( PROFILER_QUERY((profilerQuery)n) )
                                {
                                        mProfilerTimes[PROFILER_TOTAL].x += mProfilerTimes[n].x;
                                        mProfilerTimes[PROFILER_TOTAL].y += mProfilerTimes[n].y;
                                }
                        }

                        return true;
                }
                else if( mProfilerQueriesUsed & flag )
                {
                        if( !(mProfilerQueriesDone & flag) )
                        {
                                const uint32_t evt = query*2;
                                float cuda_time = 0.0f;
                                CUDA(cudaEventElapsedTime(&cuda_time, mEventsGPU[evt], mEventsGPU[evt+1]));
                                mProfilerTimes[query].y = cuda_time;
                                mProfilerQueriesDone |= flag;
                                //mProfilerQueriesUsed &= ~flag;
                        }

                        return true;
                }

                return false;
        }

protected:

        /* Member Variables */
        std::string mPrototxtPath;
        std::string mModelPath;
        std::string mMeanPath;
        std::string mInputBlobName;
        std::string mCacheEnginePath;
        std::string mCacheCalibrationPath;

        deviceType    mDevice;
        precisionType mPrecision;
        modelType     mModelType;
        cudaStream_t  mStream;
        cudaEvent_t   mEventsGPU[PROFILER_TOTAL * 2];
        timespec      mEventsCPU[PROFILER_TOTAL * 2];

        nvinfer1::IRuntime* mInfer;
        nvinfer1::ICudaEngine* mEngine;
        nvinfer1::IExecutionContext* mContext;
        
        uint32_t mWidth;
        uint32_t mHeight;
        uint32_t mInputSize;
        float*   mInputCPU;
        float*   mInputCUDA;
        float2   mProfilerTimes[PROFILER_TOTAL + 1];
        uint32_t mProfilerQueriesUsed;
        uint32_t mProfilerQueriesDone;
        uint32_t mMaxBatchSize;
        bool        mEnableProfiler;
        bool     mEnableDebug;
        bool        mAllowGPUFallback;

        Dims3 mInputDims;
        
        struct outputLayer
        {
                std::string name;
                Dims3 dims;
                uint32_t size;
                float* CPU;
                float* CUDA;
        };
        
        std::vector<outputLayer> mOutputs;
};

#endif