diff --git a/README.md b/README.md index 6178e40..bb2fd53 100644 --- a/README.md +++ b/README.md @@ -53,6 +53,7 @@ Following models are implemented. |Name | Description | |-|-| +|[mlp](./mlp) | the very basic model for starters, properly documented | |[lenet](./lenet) | the simplest, as a "hello world" of this project | |[alexnet](./alexnet)| easy to implement, all layers are supported in tensorrt | |[googlenet](./googlenet)| GoogLeNet (Inception v1) | diff --git a/mlp/CMakeLists.txt b/mlp/CMakeLists.txt new file mode 100644 index 0000000..0dc9c98 --- /dev/null +++ b/mlp/CMakeLists.txt @@ -0,0 +1,24 @@ +cmake_minimum_required(VERSION 3.14) # change the version, if asked by compiler +project(mlp) + +set(CMAKE_CXX_STANDARD 14) + +# include and link dirs of tensorrt, you need adapt them if yours are different +include_directories(/usr/include/x86_64-linux-gnu/) +link_directories(/usr/lib/x86_64-linux-gnu/) + +# include and link dirs of cuda for inference +include_directories(/usr/local/cuda/include) +link_directories(/usr/local/cuda/lib64) + +# create link for executable files +add_executable(mlp mlp.cpp) + +# perform linking with nvinfer libraries +target_link_libraries(mlp nvinfer) + +# link with cuda libraries for Inference +target_link_libraries(mlp cudart) + +add_definitions(-O2 -pthread) + diff --git a/mlp/README.md b/mlp/README.md new file mode 100644 index 0000000..5bfc408 --- /dev/null +++ b/mlp/README.md @@ -0,0 +1,57 @@ +# MLP + +MLP is the most basic net in this tensorrtx project for starters. You can learn the basic procedures of building +TensorRT app from the provided APIs. The process of building a TensorRT engine explained in the chart below. + +![TensorRT Image](https://user-images.githubusercontent.com/33795294/148565279-795b12da-5243-4e7e-881b-263eb7658683.jpg) + +## Helper Files + +`logging.h` : A logger file for using NVIDIA TRT API (mostly same for all models) + +`mlp.wts` : Converted weight file (simple file, you can open and check it) + +## TensorRT C++ API + +``` +// 1. generate mlp.wts from https://github.com/wang-xinyu/pytorchx/tree/master/mlp -- or use the given .wts file + +// 2. put mlp.wts into tensorrtx/mlp (if using the generated weights) + +// 3. build and run + + cd tensorrtx/mlp + + mkdir build + + cd build + + cmake .. + + make + + sudo ./mlp -s // serialize model to plan file i.e. 'mlp.engine' + + sudo ./mlp -d // deserialize plan file and run inference +``` + +## TensorRT Python API + +``` +# 1. Generate mlp.wts from https://github.com/wang-xinyu/pytorchx/tree/master/mlp -- or use the given .wts file + +# 2. Put mlp.wts into tensorrtx/mlp (if using the generated weights) + +# 3. Install Python dependencies (tensorrt/pycuda/numpy) + +# 4. Run + + cd tensorrtx/mlp + + python mlp.py -s # serialize model to plan file, i.e. 'mlp.engine' + + python mlp.py -d # deserialize plan file and run inference +``` + +## Note +It also supports the latest CUDA-11.4 and TensorRT-8.2.x diff --git a/mlp/logging.h b/mlp/logging.h new file mode 100644 index 0000000..0edb75f --- /dev/null +++ b/mlp/logging.h @@ -0,0 +1,503 @@ +/* + * Copyright (c) 2019, NVIDIA CORPORATION. All rights reserved. + * + * Licensed under the Apache License, Version 2.0 (the "License"); + * you may not use this file except in compliance with the License. + * You may obtain a copy of the License at + * + * http://www.apache.org/licenses/LICENSE-2.0 + * + * Unless required by applicable law or agreed to in writing, software + * distributed under the License is distributed on an "AS IS" BASIS, + * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + * See the License for the specific language governing permissions and + * limitations under the License. + */ + +#ifndef TENSORRT_LOGGING_H +#define TENSORRT_LOGGING_H + +#include "NvInferRuntimeCommon.h" +#include +#include +#include +#include +#include +#include +#include + +using Severity = nvinfer1::ILogger::Severity; + +class LogStreamConsumerBuffer : public std::stringbuf +{ +public: + LogStreamConsumerBuffer(std::ostream& stream, const std::string& prefix, bool shouldLog) + : mOutput(stream) + , mPrefix(prefix) + , mShouldLog(shouldLog) + { + } + + LogStreamConsumerBuffer(LogStreamConsumerBuffer&& other) + : mOutput(other.mOutput) + { + } + + ~LogStreamConsumerBuffer() + { + // std::streambuf::pbase() gives a pointer to the beginning of the buffered part of the output sequence + // std::streambuf::pptr() gives a pointer to the current position of the output sequence + // if the pointer to the beginning is not equal to the pointer to the current position, + // call putOutput() to log the output to the stream + if (pbase() != pptr()) + { + putOutput(); + } + } + + // synchronizes the stream buffer and returns 0 on success + // synchronizing the stream buffer consists of inserting the buffer contents into the stream, + // resetting the buffer and flushing the stream + virtual int sync() + { + putOutput(); + return 0; + } + + void putOutput() + { + if (mShouldLog) + { + // prepend timestamp + std::time_t timestamp = std::time(nullptr); + tm* tm_local = std::localtime(×tamp); + std::cout << "["; + std::cout << std::setw(2) << std::setfill('0') << 1 + tm_local->tm_mon << "/"; + std::cout << std::setw(2) << std::setfill('0') << tm_local->tm_mday << "/"; + std::cout << std::setw(4) << std::setfill('0') << 1900 + tm_local->tm_year << "-"; + std::cout << std::setw(2) << std::setfill('0') << tm_local->tm_hour << ":"; + std::cout << std::setw(2) << std::setfill('0') << tm_local->tm_min << ":"; + std::cout << std::setw(2) << std::setfill('0') << tm_local->tm_sec << "] "; + // std::stringbuf::str() gets the string contents of the buffer + // insert the buffer contents pre-appended by the appropriate prefix into the stream + mOutput << mPrefix << str(); + // set the buffer to empty + str(""); + // flush the stream + mOutput.flush(); + } + } + + void setShouldLog(bool shouldLog) + { + mShouldLog = shouldLog; + } + +private: + std::ostream& mOutput; + std::string mPrefix; + bool mShouldLog; +}; + +//! +//! \class LogStreamConsumerBase +//! \brief Convenience object used to initialize LogStreamConsumerBuffer before std::ostream in LogStreamConsumer +//! +class LogStreamConsumerBase +{ +public: + LogStreamConsumerBase(std::ostream& stream, const std::string& prefix, bool shouldLog) + : mBuffer(stream, prefix, shouldLog) + { + } + +protected: + LogStreamConsumerBuffer mBuffer; +}; + +//! +//! \class LogStreamConsumer +//! \brief Convenience object used to facilitate use of C++ stream syntax when logging messages. +//! Order of base classes is LogStreamConsumerBase and then std::ostream. +//! This is because the LogStreamConsumerBase class is used to initialize the LogStreamConsumerBuffer member field +//! in LogStreamConsumer and then the address of the buffer is passed to std::ostream. +//! This is necessary to prevent the address of an uninitialized buffer from being passed to std::ostream. +//! Please do not change the order of the parent classes. +//! +class LogStreamConsumer : protected LogStreamConsumerBase, public std::ostream +{ +public: + //! \brief Creates a LogStreamConsumer which logs messages with level severity. + //! Reportable severity determines if the messages are severe enough to be logged. + LogStreamConsumer(Severity reportableSeverity, Severity severity) + : LogStreamConsumerBase(severityOstream(severity), severityPrefix(severity), severity <= reportableSeverity) + , std::ostream(&mBuffer) // links the stream buffer with the stream + , mShouldLog(severity <= reportableSeverity) + , mSeverity(severity) + { + } + + LogStreamConsumer(LogStreamConsumer&& other) + : LogStreamConsumerBase(severityOstream(other.mSeverity), severityPrefix(other.mSeverity), other.mShouldLog) + , std::ostream(&mBuffer) // links the stream buffer with the stream + , mShouldLog(other.mShouldLog) + , mSeverity(other.mSeverity) + { + } + + void setReportableSeverity(Severity reportableSeverity) + { + mShouldLog = mSeverity <= reportableSeverity; + mBuffer.setShouldLog(mShouldLog); + } + +private: + static std::ostream& severityOstream(Severity severity) + { + return severity >= Severity::kINFO ? std::cout : std::cerr; + } + + static std::string severityPrefix(Severity severity) + { + switch (severity) + { + case Severity::kINTERNAL_ERROR: return "[F] "; + case Severity::kERROR: return "[E] "; + case Severity::kWARNING: return "[W] "; + case Severity::kINFO: return "[I] "; + case Severity::kVERBOSE: return "[V] "; + default: assert(0); return ""; + } + } + + bool mShouldLog; + Severity mSeverity; +}; + +//! \class Logger +//! +//! \brief Class which manages logging of TensorRT tools and samples +//! +//! \details This class provides a common interface for TensorRT tools and samples to log information to the console, +//! and supports logging two types of messages: +//! +//! - Debugging messages with an associated severity (info, warning, error, or internal error/fatal) +//! - Test pass/fail messages +//! +//! The advantage of having all samples use this class for logging as opposed to emitting directly to stdout/stderr is +//! that the logic for controlling the verbosity and formatting of sample output is centralized in one location. +//! +//! In the future, this class could be extended to support dumping test results to a file in some standard format +//! (for example, JUnit XML), and providing additional metadata (e.g. timing the duration of a test run). +//! +//! TODO: For backwards compatibility with existing samples, this class inherits directly from the nvinfer1::ILogger +//! interface, which is problematic since there isn't a clean separation between messages coming from the TensorRT +//! library and messages coming from the sample. +//! +//! In the future (once all samples are updated to use Logger::getTRTLogger() to access the ILogger) we can refactor the +//! class to eliminate the inheritance and instead make the nvinfer1::ILogger implementation a member of the Logger +//! object. + +class Logger : public nvinfer1::ILogger +{ +public: + Logger(Severity severity = Severity::kWARNING) + : mReportableSeverity(severity) + { + } + + //! + //! \enum TestResult + //! \brief Represents the state of a given test + //! + enum class TestResult + { + kRUNNING, //!< The test is running + kPASSED, //!< The test passed + kFAILED, //!< The test failed + kWAIVED //!< The test was waived + }; + + //! + //! \brief Forward-compatible method for retrieving the nvinfer::ILogger associated with this Logger + //! \return The nvinfer1::ILogger associated with this Logger + //! + //! TODO Once all samples are updated to use this method to register the logger with TensorRT, + //! we can eliminate the inheritance of Logger from ILogger + //! + nvinfer1::ILogger& getTRTLogger() + { + return *this; + } + + //! + //! \brief Implementation of the nvinfer1::ILogger::log() virtual method + //! + //! Note samples should not be calling this function directly; it will eventually go away once we eliminate the + //! inheritance from nvinfer1::ILogger + //! + void log(Severity severity, const char* msg) noexcept override + { + LogStreamConsumer(mReportableSeverity, severity) << "[TRT] " << std::string(msg) << std::endl; + } + + //! + //! \brief Method for controlling the verbosity of logging output + //! + //! \param severity The logger will only emit messages that have severity of this level or higher. + //! + void setReportableSeverity(Severity severity) + { + mReportableSeverity = severity; + } + + //! + //! \brief Opaque handle that holds logging information for a particular test + //! + //! This object is an opaque handle to information used by the Logger to print test results. + //! The sample must call Logger::defineTest() in order to obtain a TestAtom that can be used + //! with Logger::reportTest{Start,End}(). + //! + class TestAtom + { + public: + TestAtom(TestAtom&&) = default; + + private: + friend class Logger; + + TestAtom(bool started, const std::string& name, const std::string& cmdline) + : mStarted(started) + , mName(name) + , mCmdline(cmdline) + { + } + + bool mStarted; + std::string mName; + std::string mCmdline; + }; + + //! + //! \brief Define a test for logging + //! + //! \param[in] name The name of the test. This should be a string starting with + //! "TensorRT" and containing dot-separated strings containing + //! the characters [A-Za-z0-9_]. + //! For example, "TensorRT.sample_googlenet" + //! \param[in] cmdline The command line used to reproduce the test + // + //! \return a TestAtom that can be used in Logger::reportTest{Start,End}(). + //! + static TestAtom defineTest(const std::string& name, const std::string& cmdline) + { + return TestAtom(false, name, cmdline); + } + + //! + //! \brief A convenience overloaded version of defineTest() that accepts an array of command-line arguments + //! as input + //! + //! \param[in] name The name of the test + //! \param[in] argc The number of command-line arguments + //! \param[in] argv The array of command-line arguments (given as C strings) + //! + //! \return a TestAtom that can be used in Logger::reportTest{Start,End}(). + static TestAtom defineTest(const std::string& name, int argc, char const* const* argv) + { + auto cmdline = genCmdlineString(argc, argv); + return defineTest(name, cmdline); + } + + //! + //! \brief Report that a test has started. + //! + //! \pre reportTestStart() has not been called yet for the given testAtom + //! + //! \param[in] testAtom The handle to the test that has started + //! + static void reportTestStart(TestAtom& testAtom) + { + reportTestResult(testAtom, TestResult::kRUNNING); + assert(!testAtom.mStarted); + testAtom.mStarted = true; + } + + //! + //! \brief Report that a test has ended. + //! + //! \pre reportTestStart() has been called for the given testAtom + //! + //! \param[in] testAtom The handle to the test that has ended + //! \param[in] result The result of the test. Should be one of TestResult::kPASSED, + //! TestResult::kFAILED, TestResult::kWAIVED + //! + static void reportTestEnd(const TestAtom& testAtom, TestResult result) + { + assert(result != TestResult::kRUNNING); + assert(testAtom.mStarted); + reportTestResult(testAtom, result); + } + + static int reportPass(const TestAtom& testAtom) + { + reportTestEnd(testAtom, TestResult::kPASSED); + return EXIT_SUCCESS; + } + + static int reportFail(const TestAtom& testAtom) + { + reportTestEnd(testAtom, TestResult::kFAILED); + return EXIT_FAILURE; + } + + static int reportWaive(const TestAtom& testAtom) + { + reportTestEnd(testAtom, TestResult::kWAIVED); + return EXIT_SUCCESS; + } + + static int reportTest(const TestAtom& testAtom, bool pass) + { + return pass ? reportPass(testAtom) : reportFail(testAtom); + } + + Severity getReportableSeverity() const + { + return mReportableSeverity; + } + +private: + //! + //! \brief returns an appropriate string for prefixing a log message with the given severity + //! + static const char* severityPrefix(Severity severity) + { + switch (severity) + { + case Severity::kINTERNAL_ERROR: return "[F] "; + case Severity::kERROR: return "[E] "; + case Severity::kWARNING: return "[W] "; + case Severity::kINFO: return "[I] "; + case Severity::kVERBOSE: return "[V] "; + default: assert(0); return ""; + } + } + + //! + //! \brief returns an appropriate string for prefixing a test result message with the given result + //! + static const char* testResultString(TestResult result) + { + switch (result) + { + case TestResult::kRUNNING: return "RUNNING"; + case TestResult::kPASSED: return "PASSED"; + case TestResult::kFAILED: return "FAILED"; + case TestResult::kWAIVED: return "WAIVED"; + default: assert(0); return ""; + } + } + + //! + //! \brief returns an appropriate output stream (cout or cerr) to use with the given severity + //! + static std::ostream& severityOstream(Severity severity) + { + return severity >= Severity::kINFO ? std::cout : std::cerr; + } + + //! + //! \brief method that implements logging test results + //! + static void reportTestResult(const TestAtom& testAtom, TestResult result) + { + severityOstream(Severity::kINFO) << "&&&& " << testResultString(result) << " " << testAtom.mName << " # " + << testAtom.mCmdline << std::endl; + } + + //! + //! \brief generate a command line string from the given (argc, argv) values + //! + static std::string genCmdlineString(int argc, char const* const* argv) + { + std::stringstream ss; + for (int i = 0; i < argc; i++) + { + if (i > 0) + ss << " "; + ss << argv[i]; + } + return ss.str(); + } + + Severity mReportableSeverity; +}; + +namespace +{ + +//! +//! \brief produces a LogStreamConsumer object that can be used to log messages of severity kVERBOSE +//! +//! Example usage: +//! +//! LOG_VERBOSE(logger) << "hello world" << std::endl; +//! +inline LogStreamConsumer LOG_VERBOSE(const Logger& logger) +{ + return LogStreamConsumer(logger.getReportableSeverity(), Severity::kVERBOSE); +} + +//! +//! \brief produces a LogStreamConsumer object that can be used to log messages of severity kINFO +//! +//! Example usage: +//! +//! LOG_INFO(logger) << "hello world" << std::endl; +//! +inline LogStreamConsumer LOG_INFO(const Logger& logger) +{ + return LogStreamConsumer(logger.getReportableSeverity(), Severity::kINFO); +} + +//! +//! \brief produces a LogStreamConsumer object that can be used to log messages of severity kWARNING +//! +//! Example usage: +//! +//! LOG_WARN(logger) << "hello world" << std::endl; +//! +inline LogStreamConsumer LOG_WARN(const Logger& logger) +{ + return LogStreamConsumer(logger.getReportableSeverity(), Severity::kWARNING); +} + +//! +//! \brief produces a LogStreamConsumer object that can be used to log messages of severity kERROR +//! +//! Example usage: +//! +//! LOG_ERROR(logger) << "hello world" << std::endl; +//! +inline LogStreamConsumer LOG_ERROR(const Logger& logger) +{ + return LogStreamConsumer(logger.getReportableSeverity(), Severity::kERROR); +} + +//! +//! \brief produces a LogStreamConsumer object that can be used to log messages of severity kINTERNAL_ERROR +// ("fatal" severity) +//! +//! Example usage: +//! +//! LOG_FATAL(logger) << "hello world" << std::endl; +//! +inline LogStreamConsumer LOG_FATAL(const Logger& logger) +{ + return LogStreamConsumer(logger.getReportableSeverity(), Severity::kINTERNAL_ERROR); +} + +} // anonymous namespace + +#endif // TENSORRT_LOGGING_H diff --git a/mlp/mlp.cpp b/mlp/mlp.cpp new file mode 100644 index 0000000..e2217b4 --- /dev/null +++ b/mlp/mlp.cpp @@ -0,0 +1,323 @@ +#include "NvInfer.h" // TensorRT library +#include "iostream" // Standard input/output library +#include "logging.h" // logging file -- by NVIDIA +#include // for weight maps +#include // for file-handling +#include // for timing the execution + +// provided by nvidia for using TensorRT APIs +using namespace nvinfer1; + +// Logger from TRT API +static Logger gLogger; + +const int INPUT_SIZE = 1; +const int OUTPUT_SIZE = 1; + +/** //////////////////////////// +// DEPLOYMENT RELATED ///////// +////////////////////////////*/ +std::map loadWeights(const std::string file) { + /** + * Parse the .wts file and store weights in dict format. + * + * @param file path to .wts file + * @return weight_map: dictionary containing weights and their values + */ + + std::cout << "[INFO]: Loading weights..." << file << std::endl; + std::map weightMap; + + // Open Weight file + std::ifstream input(file); + assert(input.is_open() && "[ERROR]: Unable to load weight file..."); + + // Read number of weights + int32_t count; + input >> count; + assert(count > 0 && "Invalid weight map file."); + + // Loop through number of line, actually the number of weights & biases + while (count--) { + // TensorRT weights + Weights wt{DataType::kFLOAT, nullptr, 0}; + uint32_t size; + // Read name and type of weights + std::string w_name; + input >> w_name >> std::dec >> size; + wt.type = DataType::kFLOAT; + + uint32_t *val = reinterpret_cast(malloc(sizeof(val) * size)); + for (uint32_t x = 0, y = size; x < y; ++x) { + // Change hex values to uint32 (for higher values) + input >> std::hex >> val[x]; + } + wt.values = val; + wt.count = size; + + // Add weight values against its name (key) + weightMap[w_name] = wt; + } + return weightMap; +} + +ICudaEngine *createMLPEngine(unsigned int maxBatchSize, IBuilder *builder, IBuilderConfig *config, DataType dt) { + /** + * Create Multi-Layer Perceptron using the TRT Builder and Configurations + * + * @param maxBatchSize: batch size for built TRT model + * @param builder: to build engine and networks + * @param config: configuration related to Hardware + * @param dt: datatype for model layers + * @return engine: TRT model + */ + + std::cout << "[INFO]: Creating MLP using TensorRT..." << std::endl; + + // Load Weights from relevant file + std::map weightMap = loadWeights("../mlp.wts"); + + // Create an empty network + INetworkDefinition *network = builder->createNetworkV2(0U); + + // Create an input with proper *name + ITensor *data = network->addInput("data", DataType::kFLOAT, Dims3{1, 1, 1}); + assert(data); + + // Add layer for MLP + IFullyConnectedLayer *fc1 = network->addFullyConnected(*data, 1, + weightMap["linear.weight"], + weightMap["linear.bias"]); + assert(fc1); + + // set output with *name + fc1->getOutput(0)->setName("out"); + + // mark the output + network->markOutput(*fc1->getOutput(0)); + + // Set configurations + builder->setMaxBatchSize(1); + // Set workspace size + config->setMaxWorkspaceSize(1 << 20); + + // Build CUDA Engine using network and configurations + ICudaEngine *engine = builder->buildEngineWithConfig(*network, *config); + assert(engine != nullptr); + + // Don't need the network any more + // free captured memory + network->destroy(); + + // Release host memory + for (auto &mem: weightMap) { + free((void *) (mem.second.values)); + } + + return engine; +} + +void APIToModel(unsigned int maxBatchSize, IHostMemory **modelStream) { + /** + * Create engine using TensorRT APIs + * + * @param maxBatchSize: for the deployed model configs + * @param modelStream: shared memory to store serialized model + */ + + // Create builder with the help of logger + IBuilder *builder = createInferBuilder(gLogger); + + // Create hardware configs + IBuilderConfig *config = builder->createBuilderConfig(); + + // Build an engine + ICudaEngine *engine = createMLPEngine(maxBatchSize, builder, config, DataType::kFLOAT); + assert(engine != nullptr); + + // serialize the engine into binary stream + (*modelStream) = engine->serialize(); + + // free up the memory + engine->destroy(); + builder->destroy(); +} + +void performSerialization() { + /** + * Serialization Function + */ + // Shared memory object + IHostMemory *modelStream{nullptr}; + + // Write model into stream + APIToModel(1, &modelStream); + assert(modelStream != nullptr); + + + std::cout << "[INFO]: Writing engine into binary..." << std::endl; + + // Open the file and write the contents there in binary format + std::ofstream p("../mlp.engine", std::ios::binary); + if (!p) { + std::cerr << "could not open plan output file" << std::endl; + return; + } + p.write(reinterpret_cast(modelStream->data()), modelStream->size()); + + // Release the memory + modelStream->destroy(); + + std::cout << "[INFO]: Successfully created TensorRT engine..." << std::endl; + std::cout << "\n\tRun inference using `./mlp -d`" << std::endl; + +} + +/** //////////////////////////// +// INFERENCE RELATED ////////// +////////////////////////////*/ +void doInference(IExecutionContext &context, float *input, float *output, int batchSize) { + /** + * Perform inference using the CUDA context + * + * @param context: context created by engine + * @param input: input from the host + * @param output: output to save on host + * @param batchSize: batch size for TRT model + */ + + // Get engine from the context + const ICudaEngine &engine = context.getEngine(); + + // Pointers to input and output device buffers to pass to engine. + // Engine requires exactly IEngine::getNbBindings() number of buffers. + assert(engine.getNbBindings() == 2); + void *buffers[2]; + + // In order to bind the buffers, we need to know the names of the input and output tensors. + // Note that indices are guaranteed to be less than IEngine::getNbBindings() + const int inputIndex = engine.getBindingIndex("data"); + const int outputIndex = engine.getBindingIndex("out"); + + // Create GPU buffers on device -- allocate memory for input and output + cudaMalloc(&buffers[inputIndex], batchSize * INPUT_SIZE * sizeof(float)); + cudaMalloc(&buffers[outputIndex], batchSize * OUTPUT_SIZE * sizeof(float)); + + // create CUDA stream for simultaneous CUDA operations + cudaStream_t stream; + cudaStreamCreate(&stream); + + // copy input from host (CPU) to device (GPU) in stream + cudaMemcpyAsync(buffers[inputIndex], input, batchSize * INPUT_SIZE * sizeof(float), cudaMemcpyHostToDevice, stream); + + // execute inference using context provided by engine + context.enqueue(batchSize, buffers, stream, nullptr); + + // copy output back from device (GPU) to host (CPU) + cudaMemcpyAsync(output, buffers[outputIndex], batchSize * OUTPUT_SIZE * sizeof(float), cudaMemcpyDeviceToHost, + stream); + + // synchronize the stream to prevent issues + // (block CUDA and wait for CUDA operations to be completed) + cudaStreamSynchronize(stream); + + // Release stream and buffers (memory) + cudaStreamDestroy(stream); + cudaFree(buffers[inputIndex]); + cudaFree(buffers[outputIndex]); +} + +void performInference() { + /** + * Get inference using the pre-trained model + */ + + // stream to write model + char *trtModelStream{nullptr}; + size_t size{0}; + + // read model from the engine file + std::ifstream file("../mlp.engine", std::ios::binary); + if (file.good()) { + file.seekg(0, file.end); + size = file.tellg(); + file.seekg(0, file.beg); + trtModelStream = new char[size]; + assert(trtModelStream); + file.read(trtModelStream, size); + file.close(); + } + + // create a runtime (required for deserialization of model) with NVIDIA's logger + IRuntime *runtime = createInferRuntime(gLogger); + assert(runtime != nullptr); + + // deserialize engine for using the char-stream + ICudaEngine *engine = runtime->deserializeCudaEngine(trtModelStream, size, nullptr); + assert(engine != nullptr); + + // create execution context -- required for inference executions + IExecutionContext *context = engine->createExecutionContext(); + assert(context != nullptr); + + float out[1]; // array for output + float data[1]; // array for input + for (float &i: data) + i = 12.0; // put any value for input + + // time the execution + auto start = std::chrono::system_clock::now(); + + // do inference using the parameters + doInference(*context, data, out, 1); + + // time the execution + auto end = std::chrono::system_clock::now(); + std::cout << "\n[INFO]: Time taken by execution: " + << std::chrono::duration_cast(end - start).count() << "ms" << std::endl; + + + // free the captured space + context->destroy(); + engine->destroy(); + runtime->destroy(); + + std::cout << "\nInput:\t" << data[0]; + std::cout << "\nOutput:\t"; + for (float i: out) { + std::cout << i; + } + std::cout << std::endl; +} + +int checkArgs(int argc, char **argv) { + /** + * Parse command line arguments + * + * @param argc: argument count + * @param argv: arguments vector + * @return int: a flag to perform operation + */ + + if (argc != 2) { + std::cerr << "[ERROR]: Arguments not right!" << std::endl; + std::cerr << "./mlp -s // serialize model to plan file" << std::endl; + std::cerr << "./mlp -d // deserialize plan file and run inference" << std::endl; + return -1; + } + if (std::string(argv[1]) == "-s") { + return 1; + } else if (std::string(argv[1]) == "-d") { + return 2; + } + return -1; +} + +int main(int argc, char **argv) { + int args = checkArgs(argc, argv); + if (args == 1) + performSerialization(); + else if (args == 2) + performInference(); + return 0; +} diff --git a/mlp/mlp.py b/mlp/mlp.py new file mode 100644 index 0000000..aad38ea --- /dev/null +++ b/mlp/mlp.py @@ -0,0 +1,248 @@ +import argparse +import os +import numpy as np +import struct + +# required for the model creation +import tensorrt as trt + +# required for the inference using TRT engine +import pycuda.autoinit +import pycuda.driver as cuda + +# Sizes of input and output for TensorRT model +INPUT_SIZE = 1 +OUTPUT_SIZE = 1 + +# path of .wts (weight file) and .engine (model file) +WEIGHT_PATH = "./mlp.wts" +ENGINE_PATH = "./mlp.engine" + +# input and output names are must for the TRT model +INPUT_BLOB_NAME = 'data' +OUTPUT_BLOB_NAME = 'out' + +# A logger provided by NVIDIA-TRT +gLogger = trt.Logger(trt.Logger.INFO) + + +################################ +# DEPLOYMENT RELATED ########### +################################ +def load_weights(file_path): + """ + Parse the .wts file and store weights in dict format + :param file_path: + :return weight_map: dictionary containing weights and their values + """ + print(f"[INFO]: Loading weights: {file_path}") + assert os.path.exists(file_path), '[ERROR]: Unable to load weight file.' + + weight_map = {} + with open(file_path, "r") as f: + lines = [line.strip() for line in f] + + # count for total # of weights + count = int(lines[0]) + assert count == len(lines) - 1 + + # Loop through counts and get the exact num of values against weights + for i in range(1, count + 1): + splits = lines[i].split(" ") + name = splits[0] + cur_count = int(splits[1]) + + # len of splits must be greater than current weight counts + assert cur_count + 2 == len(splits) + + # loop through all weights and unpack from the hexadecimal values + values = [] + for j in range(2, len(splits)): + # hex string to bytes to float + values.append(struct.unpack(">f", bytes.fromhex(splits[j]))) + + # store in format of { 'weight.name': [weights_val0, weight_val1, ..] } + weight_map[name] = np.array(values, dtype=np.float32) + + return weight_map + + +def create_mlp_engine(max_batch_size, builder, config, dt): + """ + Create Multi-Layer Perceptron using the TRT Builder and Configurations + :param max_batch_size: batch size for built TRT model + :param builder: to build engine and networks + :param config: configuration related to Hardware + :param dt: datatype for model layers + :return engine: TRT model + """ + print("[INFO]: Creating MLP using TensorRT...") + # load weight maps from the file + weight_map = load_weights(WEIGHT_PATH) + + # build an empty network using builder + network = builder.create_network() + + # add an input to network using the *input-name + data = network.add_input(INPUT_BLOB_NAME, dt, (1, 1, INPUT_SIZE)) + assert data + + # add the layer with output-size (number of outputs) + linear = network.add_fully_connected(input=data, + num_outputs=OUTPUT_SIZE, + kernel=weight_map['linear.weight'], + bias=weight_map['linear.bias']) + assert linear + + # set the name for output layer + linear.get_output(0).name = OUTPUT_BLOB_NAME + + # mark this layer as final output layer + network.mark_output(linear.get_output(0)) + + # set the batch size of current builder + builder.max_batch_size = max_batch_size + + # create the engine with model and hardware configs + engine = builder.build_engine(network, config) + + # free captured memory + del network + del weight_map + + # return engine + return engine + + +def api_to_model(max_batch_size): + """ + Create engine using TensorRT APIs + :param max_batch_size: for the deployed model configs + :return: + """ + # Create Builder with logger provided by TRT + builder = trt.Builder(gLogger) + + # Create configurations from Engine Builder + config = builder.create_builder_config() + + # Create MLP Engine + engine = create_mlp_engine(max_batch_size, builder, config, trt.float32) + assert engine + + # Write the engine into binary file + print("[INFO]: Writing engine into binary...") + with open(ENGINE_PATH, "wb") as f: + # write serialized model in file + f.write(engine.serialize()) + + # free the memory + del engine + del builder + + +################################ +# INFERENCE RELATED ############ +################################ +def perform_inference(input_val): + """ + Get inference using the pre-trained model + :param input_val: a number as an input + :return: + """ + + def do_inference(inf_context, inf_host_in, inf_host_out): + """ + Perform inference using the CUDA context + :param inf_context: context created by engine + :param inf_host_in: input from the host + :param inf_host_out: output to save on host + :return: + """ + + inference_engine = inf_context.engine + # Input and output bindings are required for inference + assert inference_engine.num_bindings == 2 + + # allocate memory in GPU using CUDA bindings + device_in = cuda.mem_alloc(inf_host_in.nbytes) + device_out = cuda.mem_alloc(inf_host_out.nbytes) + + # create bindings for input and output + bindings = [int(device_in), int(device_out)] + + # create CUDA stream for simultaneous CUDA operations + stream = cuda.Stream() + + # copy input from host (CPU) to device (GPU) in stream + cuda.memcpy_htod_async(device_in, inf_host_in, stream) + + # execute inference using context provided by engine + inf_context.execute_async(bindings=bindings, stream_handle=stream.handle) + + # copy output back from device (GPU) to host (CPU) + cuda.memcpy_dtoh_async(inf_host_out, device_out, stream) + + # synchronize the stream to prevent issues + # (block CUDA and wait for CUDA operations to be completed) + stream.synchronize() + + # create a runtime (required for deserialization of model) with NVIDIA's logger + runtime = trt.Runtime(gLogger) + assert runtime + + # read and deserialize engine for inference + with open(ENGINE_PATH, "rb") as f: + engine = runtime.deserialize_cuda_engine(f.read()) + assert engine + + # create execution context -- required for inference executions + context = engine.create_execution_context() + assert context + + # create input as array + data = np.array([input_val], dtype=np.float32) + + # capture free memory for input in GPU + host_in = cuda.pagelocked_empty((INPUT_SIZE), dtype=np.float32) + + # copy input-array from CPU to Flatten array in GPU + np.copyto(host_in, data.ravel()) + + # capture free memory for output in GPU + host_out = cuda.pagelocked_empty(OUTPUT_SIZE, dtype=np.float32) + + # do inference using required parameters + do_inference(context, host_in, host_out) + + print(f'\n[INFO]: Predictions using pre-trained model..\n\tInput:\t{input_val}\n\tOutput:\t{host_out[0]:.4f}') + + +def get_args(): + """ + Parse command line arguments + :return arguments: parsed arguments + """ + arg_parser = argparse.ArgumentParser() + arg_parser.add_argument('-s', action='store_true') + arg_parser.add_argument('-d', action='store_true') + arguments = vars(arg_parser.parse_args()) + # check for the arguments + if not (arguments['s'] ^ arguments['d']): + print("[ERROR]: Arguments not right!\n") + print("\tpython mlp.py -s # serialize model to engine file") + print("\tpython mlp.py -d # deserialize engine file and run inference") + exit() + + return arguments + + +if __name__ == "__main__": + args = get_args() + if args['s']: + api_to_model(max_batch_size=1) + print("[INFO]: Successfully created TensorRT engine...") + print("\n\tRun inference using `python mlp.py -d`\n") + else: + perform_inference(input_val=4.0) + diff --git a/mlp/mlp.wts b/mlp/mlp.wts new file mode 100644 index 0000000..01ef0db --- /dev/null +++ b/mlp/mlp.wts @@ -0,0 +1,3 @@ +2 +linear.weight 1 3fff7e32 +linear.bias 1 3c138a5a