# 限定使用的最小cmake版本
cmake_minimum_required(VERSION 3.2 FATAL_ERROR)
# 項(xiàng)目名稱(chēng)：FCPlugin 編程語(yǔ)言：CXX和C（如果不指定LANGUAGES即為CXX和C）
project(FCPlugin LANGUAGES CXX C)# 設(shè)置一個(gè)宏set_ifndef，當(dāng)變量var沒(méi)找到時(shí)，將其設(shè)定為val
macro(set_ifndef var val)if(NOT ${var})set(${var} ${val})endif()message(STATUS "Configurable variable ${var} set to ${${var}}")
endmacro()# -------- 配置 --------
# 設(shè)置模塊的名稱(chēng)為fcplugin，這個(gè)名稱(chēng)必須與pyFullyConnected.cpp中的名稱(chēng)一致
set_ifndef(PY_MODULE_NAME fcplugin)
# 設(shè)置C++標(biāo)準(zhǔn)為C++11
set(CMAKE_CXX_STANDARD 11)
# pybind11默認(rèn)支持C++14，我們使用C++11標(biāo)準(zhǔn)
set(PYBIND11_CPP_STANDARD -std=c++11)
# $ENV{HOME}代表環(huán)境變量HOME，調(diào)用宏set_ifndef把它的下屬文件夾pybind11賦值給變量PYBIND11_DIR
set_ifndef(PYBIND11_DIR $ENV{HOME}/pybind11/)# 下面這些set_ifndef也是一些變量值的設(shè)置
set_ifndef(CUDA_VERSION 10.0)
set_ifndef(CUDA_ROOT /usr/local/cuda-${CUDA_VERSION})
set_ifndef(CUDNN_ROOT ${CUDA_ROOT})
set_ifndef(PYTHON_ROOT /usr/include)
set_ifndef(TRT_LIB_DIR /usr/lib/x86_64-linux-gnu)
set_ifndef(TRT_INC_DIR /usr/include/x86_64-linux-gnu)# 輸出提示信息：以下變量的值如果不顯式提供，則從已得到的變量中派生得到
message("\nThe following variables are derived from the values of the previous variables unless provided explicitly:\n")# 查找包含cuda_runtime_api.h的路徑，將該路徑賦值給變量_CUDA_INC_DIR
# HINTS ${CUDA_ROOT} 指定${CUDA_ROOT}為額外的搜索路徑
# PATH_SUFFIXES include 指定額外要搜索的子目錄include
find_path(_CUDA_INC_DIR cuda_runtime_api.h HINTS ${CUDA_ROOT} PATH_SUFFIXES include)
set_ifndef(CUDA_INC_DIR ${_CUDA_INC_DIR})find_library(_CUDA_LIB cudart HINTS ${CUDA_ROOT} PATH_SUFFIXES lib lib64)
set_ifndef(CUDA_LIB ${_CUDA_LIB})find_library(_CUBLAS_LIB cublas HINTS ${CUDA_ROOT} PATH_SUFFIXES lib lib64)
set_ifndef(CUBLAS_LIB ${_CUBLAS_LIB})find_path(_CUDNN_INC_DIR cudnn.h HINTS ${CUDNN_ROOT} PATH_SUFFIXES include x86_64-linux-gnu)
set_ifndef(CUDNN_INC_DIR ${_CUDNN_INC_DIR})find_library(_CUDNN_LIB cudnn HINTS ${CUDNN_ROOT} PATH_SUFFIXES lib lib64 x86_64-linux-gnu)
set_ifndef(CUDNN_LIB ${_CUDNN_LIB})find_library(_TRT_INC_DIR NvInfer.h HINTS ${TRT_INC_DIR} PATH_SUFFIXES include x86_64-linux-gnu)
set_ifndef(TRT_INC_DIR ${_TRT_INC_DIR})find_library(_NVINFER_LIB nvinfer HINTS ${TRT_LIB_DIR} PATH_SUFFIXES lib lib64 x86_64-linux-gnu)
set_ifndef(NVINFER_LIB ${_NVINFER_LIB})find_library(_NVPARSERS_LIB nvparsers HINTS ${TRT_LIB_DIR} PATH_SUFFIXES lib lib64 x86_64-linux-gnu)
set_ifndef(NVPARSERS_LIB ${_NVPARSERS_LIB})find_library(_NVINFER_PLUGIN_LIB nvinfer_plugin HINTS ${TRT_LIB_DIR} PATH_SUFFIXES lib lib64 x86_64-linux-gnu)
set_ifndef(NVINFER_PLUGIN_LIB ${_NVINFER_PLUGIN_LIB})find_path(_PYTHON2_INC_DIR Python.h HINTS ${PYTHON_ROOT} PATH_SUFFIXES python2.7)
set_ifndef(PYTHON2_INC_DIR ${_PYTHON2_INC_DIR})find_path(_PYTHON3_INC_DIR Python.h HINTS ${PYTHON_ROOT} PATH_SUFFIXES python3.7 python3.6 python3.5 python3.4)
set_ifndef(PYTHON3_INC_DIR ${_PYTHON3_INC_DIR})# -------- 構(gòu)建 --------# 添加include文件夾
include_directories(${TRT_INC_DIR} ${CUDA_INC_DIR} ${CUDNN_INC_DIR} ${PYBIND11_DIR}/include/)# 添加子目錄，使我們可以檢索pybind11_add_module
add_subdirectory(${PYBIND11_DIR} ${CMAKE_BINARY_DIR}/pybind11)# GLOB會(huì)遍歷指定目錄下的文件，將符合的組成一個(gè)列表，賦值給變量
# GLOB_RECURSE會(huì)遍歷${CMAKE_SOURCE_DIR}/plugin/目錄和其子目錄下的所有.cpp文件，將他們組成一個(gè)列表，賦值給變量SOURCE_FILES
file(GLOB_RECURSE SOURCE_FILES ${CMAKE_SOURCE_DIR}/plugin/*.cpp)# Bindings library. The module name MUST MATCH the module name specified in the .cpp
if(PYTHON3_INC_DIR AND NOT (${PYTHON3_INC_DIR} STREQUAL "None"))pybind11_add_module(${PY_MODULE_NAME} SHARED THIN_LTO ${SOURCE_FILES})target_include_directories(${PY_MODULE_NAME} BEFORE PUBLIC ${PYTHON3_INC_DIR})target_link_libraries(${PY_MODULE_NAME} PRIVATE ${CUDNN_LIB} ${CUDA_LIB} ${CUBLAS_LIB} ${NVINFER_LIB} ${NVPARSERS_LIB} ${NVINFER_PLUGIN_LIB})
endif()if(PYTHON2_INC_DIR AND NOT (${PYTHON2_INC_DIR} STREQUAL "None"))# Suffix the cmake target name with a 2 to differentiate from the Python 3 bindings target.pybind11_add_module(${PY_MODULE_NAME}2 SHARED THIN_LTO ${SOURCE_FILES})target_include_directories(${PY_MODULE_NAME}2 BEFORE PUBLIC ${PYTHON2_INC_DIR})target_link_libraries(${PY_MODULE_NAME}2 PRIVATE ${CUDNN_LIB} ${CUDA_LIB} ${CUBLAS_LIB} ${NVINFER_LIB} ${NVPARSERS_LIB} ${NVINFER_PLUGIN_LIB})# Rename to remove the .cpython-35... extension.set_target_properties(${PY_MODULE_NAME}2 PROPERTIES OUTPUT_NAME ${PY_MODULE_NAME} SUFFIX ".so")# Python 2 requires an empty __init__ file to be able to import.file(WRITE ${CMAKE_BINARY_DIR}/__init__.py "")
endif()
FullyConnected.h
#ifndef _FULLY_CONNECTED_H_
#define _FULLY_CONNECTED_H_#include <cassert>
#include <cstring>
#include <cuda_runtime_api.h>
#include <cudnn.h>
#include <cublas_v2.h>
#include <stdexcept>#include "NvInfer.h"
#include "NvCaffeParser.h"#define CHECK(status) { if (status != 0) throw std::runtime_error(__FILE__ + __LINE__ + std::string{"CUDA Error: "} + std::to_string(status)); }// Helpers to move data to/from the GPU.
nvinfer1::Weights copyToDevice(const void* hostData, int count)
{void* deviceData;CHECK(cudaMalloc(&deviceData, count * sizeof(float)));CHECK(cudaMemcpy(deviceData, hostData, count * sizeof(float), cudaMemcpyHostToDevice));return nvinfer1::Weights{nvinfer1::DataType::kFLOAT, deviceData, count};
}int copyFromDevice(char* hostBuffer, nvinfer1::Weights deviceWeights)
{*reinterpret_cast<int*>(hostBuffer) = deviceWeights.count;CHECK(cudaMemcpy(hostBuffer + sizeof(int), deviceWeights.values, deviceWeights.count * sizeof(float), cudaMemcpyDeviceToHost));return sizeof(int) + deviceWeights.count * sizeof(float);
}class FCPlugin: public nvinfer1::IPluginExt
{
public:// In this simple case we're going to infer the number of output channels from the bias weights.// The knowledge that the kernel weights are weights[0] and the bias weights are weights[1] was// divined from the caffe innardsFCPlugin(const nvinfer1::Weights* weights, int nbWeights){assert(nbWeights == 2);mKernelWeights = copyToDevice(weights[0].values, weights[0].count);mBiasWeights = copyToDevice(weights[1].values, weights[1].count);}// Create the plugin at runtime from a byte stream.FCPlugin(const void* data, size_t length){const char* d = reinterpret_cast<const char*>(data);const char* check = d;// Deserialize kernel.const int kernelCount = reinterpret_cast<const int*>(d)[0];mKernelWeights = copyToDevice(d + sizeof(int), kernelCount);d += sizeof(int) + mKernelWeights.count * sizeof(float);// Deserialize bias.const int biasCount = reinterpret_cast<const int*>(d)[0];mBiasWeights = copyToDevice(d + sizeof(int), biasCount);d += sizeof(int) + mBiasWeights.count * sizeof(float);// Check that the sizes are what we expected.assert(d == check + length);}virtual int getNbOutputs() const override { return 1; }virtual nvinfer1::Dims getOutputDimensions(int index, const nvinfer1::Dims* inputs, int nbInputDims) override{assert(index == 0 && nbInputDims == 1 && inputs[0].nbDims == 3);return nvinfer1::DimsCHW{static_cast<int>(mBiasWeights.count), 1, 1};}virtual int initialize() override{CHECK(cudnnCreate(&mCudnn));CHECK(cublasCreate(&mCublas));// Create cudnn tensor descriptors for bias addition.CHECK(cudnnCreateTensorDescriptor(&mSrcDescriptor));CHECK(cudnnCreateTensorDescriptor(&mDstDescriptor));return 0;}virtual void terminate() override{CHECK(cudnnDestroyTensorDescriptor(mSrcDescriptor));CHECK(cudnnDestroyTensorDescriptor(mDstDescriptor));CHECK(cublasDestroy(mCublas));CHECK(cudnnDestroy(mCudnn));}// This plugin requires no workspace memory during build time.virtual size_t getWorkspaceSize(int maxBatchSize) const override { return 0; }virtual int enqueue(int batchSize, const void* const* inputs, void** outputs, void* workspace, cudaStream_t stream) override{int nbOutputChannels = mBiasWeights.count;int nbInputChannels = mKernelWeights.count / nbOutputChannels;constexpr float kONE = 1.0f, kZERO = 0.0f;// Do matrix multiplication.cublasSetStream(mCublas, stream);cudnnSetStream(mCudnn, stream);CHECK(cublasSgemm(mCublas, CUBLAS_OP_T, CUBLAS_OP_N, nbOutputChannels, batchSize, nbInputChannels, &kONE,reinterpret_cast<const float*>(mKernelWeights.values), nbInputChannels,reinterpret_cast<const float*>(inputs[0]), nbInputChannels, &kZERO,reinterpret_cast<float*>(outputs[0]), nbOutputChannels));// Add bias.CHECK(cudnnSetTensor4dDescriptor(mSrcDescriptor, CUDNN_TENSOR_NCHW, CUDNN_DATA_FLOAT, 1, nbOutputChannels, 1, 1));CHECK(cudnnSetTensor4dDescriptor(mDstDescriptor, CUDNN_TENSOR_NCHW, CUDNN_DATA_FLOAT, batchSize, nbOutputChannels, 1, 1));CHECK(cudnnAddTensor(mCudnn, &kONE, mSrcDescriptor, mBiasWeights.values, &kONE, mDstDescriptor, outputs[0]));return 0;}// For this sample, we'll only support float32 with NCHW.virtual bool supportsFormat(nvinfer1::DataType type, nvinfer1::PluginFormat format) const override{return (type == nvinfer1::DataType::kFLOAT && format == nvinfer1::PluginFormat::kNCHW);}void configureWithFormat(const nvinfer1::Dims* inputDims, int nbInputs, const nvinfer1::Dims* outputDims, int nbOutputs, nvinfer1::DataType type, nvinfer1::PluginFormat format, int maxBatchSize){assert(nbInputs == 1 && inputDims[0].d[1] == 1 && inputDims[0].d[2] == 1);assert(nbOutputs == 1 && outputDims[0].d[1] == 1 && outputDims[0].d[2] == 1);assert(mKernelWeights.count == inputDims[0].d[0] * inputDims[0].d[1] * inputDims[0].d[2] * mBiasWeights.count);}virtual size_t getSerializationSize() override{return sizeof(int) * 2 + mKernelWeights.count * sizeof(float) + mBiasWeights.count * sizeof(float);}virtual void serialize(void* buffer) override{char* d = reinterpret_cast<char*>(buffer);const char* check = d;d += copyFromDevice(d, mKernelWeights);d += copyFromDevice(d, mBiasWeights);assert(d == check + getSerializationSize());}// Free buffers.virtual ~FCPlugin(){cudaFree(const_cast<void*>(mKernelWeights.values));mKernelWeights.values = nullptr;cudaFree(const_cast<void*>(mBiasWeights.values));mBiasWeights.values = nullptr;}private:cudnnHandle_t mCudnn;cublasHandle_t mCublas;nvinfer1::Weights mKernelWeights{nvinfer1::DataType::kFLOAT, nullptr}, mBiasWeights{nvinfer1::DataType::kFLOAT, nullptr};cudnnTensorDescriptor_t mSrcDescriptor, mDstDescriptor;
};class FCPluginFactory : public nvcaffeparser1::IPluginFactoryExt, public nvinfer1::IPluginFactory
{
public:bool isPlugin(const char* name) override { return isPluginExt(name); }bool isPluginExt(const char* name) override { return !strcmp(name, "ip2"); }// Create a plugin using provided weights.virtual nvinfer1::IPlugin* createPlugin(const char* layerName, const nvinfer1::Weights* weights, int nbWeights) override{assert(isPluginExt(layerName) && nbWeights == 2);assert(mPlugin == nullptr);// This plugin will need to be manually destroyed after parsing the network, by calling destroyPlugin.mPlugin = new FCPlugin{weights, nbWeights};return mPlugin;}// Create a plugin from serialized data.virtual nvinfer1::IPlugin* createPlugin(const char* layerName, const void* serialData, size_t serialLength) override{assert(isPlugin(layerName));// This will be automatically destroyed when the engine is destroyed.return new FCPlugin{serialData, serialLength};}// User application destroys plugin when it is safe to do so.// Should be done after consumers of plugin (like ICudaEngine) are destroyed.void destroyPlugin() { delete mPlugin; }FCPlugin* mPlugin{ nullptr };
};#endif //_FULLY_CONNECTED_H
pyFullyConnected.cpp
#include "FullyConnected.h"
#include "NvInfer.h"
#include "NvCaffeParser.h"
#include <pybind11/pybind11.h>PYBIND11_MODULE(fcplugin, m)
{namespace py = pybind11;// This allows us to use the bindings exposed by the tensorrt module.py::module::import("tensorrt");// Note that we only need to bind the constructors manually. Since all other methods override IPlugin functionality, they will be automatically available in the python bindings.// The `std::unique_ptr<FCPlugin, py::nodelete>` specifies that Python is not responsible for destroying the object. This is required because the destructor is private.py::class_<FCPlugin, nvinfer1::IPluginExt, std::unique_ptr<FCPlugin, py::nodelete>>(m, "FCPlugin")// Bind the normal constructor as well as the one which deserializes the plugin.def(py::init<const nvinfer1::Weights*, int>()).def(py::init<const void*, size_t>());// Our custom plugin factory derives from both nvcaffeparser1::IPluginFactoryExt and nvinfer1::IPluginFactorypy::class_<FCPluginFactory, nvcaffeparser1::IPluginFactoryExt, nvinfer1::IPluginFactory>(m, "FCPluginFactory")// Bind the default constructor..def(py::init<>())// The destroy_plugin function does not override either of the base classes, so we must bind it explicitly..def("destroy_plugin", &FCPluginFactory::destroyPlugin);
}
sample.py
# This sample uses a Caffe model along with a custom plugin to create a TensorRT engine.
from random import randint
from PIL import Image
import numpy as np
import tempfileimport pycuda.driver as cuda
import pycuda.autoinitimport tensorrt as trttry:from build import fcplugin
except ImportError as err:raise ImportError("""ERROR: Failed to import module ({})
Please build the FullyConnected sample plugin.
For more information, see the included README.md
Note that Python 2 requires the presence of `__init__.py` in the build folder""".format(err))# Allows us to import from common.
import sys, os
sys.path.insert(1, os.path.join(sys.path[0], ".."))
import common# You can set the logger severity higher to suppress messages (or lower to display more messages).
TRT_LOGGER = trt.Logger(trt.Logger.WARNING)# Define some global constants about the model.
class ModelData(object):INPUT_NAME = "input"INPUT_SHAPE = (1, 28, 28)OUTPUT_NAME = "prob"OUTPUT_SHAPE = (10, )DTYPE = trt.float32# Uses a parser to retrieve mean data from a binary_proto.
def retrieve_mean(mean_proto):with trt.CaffeParser() as parser:return parser.parse_binary_proto(mean_proto)# Create the parser's plugin factory. The factory is global because it has
# to be destroyed after the engine is destroyed.
fc_factory = fcplugin.FCPluginFactory()# For more information on TRT basics, refer to the introductory parser samples.
def build_engine(deploy_file, model_file):with trt.Builder(TRT_LOGGER) as builder, builder.create_network() as network, trt.CaffeParser() as parser:builder.max_workspace_size = common.GiB(1)# Set the parser's plugin factory. Note that we bind the factory to a reference so# that we can destroy it later. (parser.plugin_factory_ext is a write-only attribute)parser.plugin_factory_ext = fc_factory# Parse the model and build the engine.model_tensors = parser.parse(deploy=deploy_file, model=model_file, network=network, dtype=ModelData.DTYPE)network.mark_output(model_tensors.find(ModelData.OUTPUT_NAME))return builder.build_cuda_engine(network)# Tries to load an engine from the provided engine_path, or builds and saves an engine to the engine_path.
def get_engine(deploy_file, model_file, engine_path):try:with open(engine_path, "rb") as f, trt.Runtime(TRT_LOGGER) as runtime:# Note that we have to provide the plugin factory when deserializing an engine built with an IPlugin or IPluginExt.return runtime.deserialize_cuda_engine(f.read(), fc_factory)except:# Fallback to building an engine if the engine cannot be loaded for any reason.engine = build_engine(deploy_file, model_file)with open(engine_path, "wb") as f:f.write(engine.serialize())return engine# Loads a test case into the provided pagelocked_buffer.
def load_normalized_test_case(data_path, mean):case_num = randint(0, 9)test_case_path = os.path.join(data_path, str(case_num) + ".pgm")# Flatten the image into a 1D array, and normalize.img = np.array(Image.open(test_case_path)).ravel() - meanreturn img, case_numdef main():# Get data files for the model.data_path, [deploy_file, model_file, mean_proto] = common.find_sample_data(description="Runs an MNIST network using a Caffe model file", subfolder="mnist", find_files=["mnist.prototxt", "mnist.caffemodel", "mnist_mean.binaryproto"])# Cache the engine in a temporary directory.engine_path = os.path.join(tempfile.gettempdir(), "mnist.engine")with get_engine(deploy_file, model_file, engine_path) as engine, engine.create_execution_context() as context:# Build an engine, allocate buffers and create a stream.# For more information on buffer allocation, refer to the introductory samples.inputs, outputs, bindings, stream = common.allocate_buffers(engine)mean = retrieve_mean(mean_proto)# For more information on performing inference, refer to the introductory samples.inputs[0].host, case_num = load_normalized_test_case(data_path, mean)# The common.do_inference function will return a list of outputs - we only have one in this case.[output] = common.do_inference(context, bindings=bindings, inputs=inputs, outputs=outputs, stream=stream)pred = np.argmax(output)print("Test Case: " + str(case_num))print("Prediction: " + str(pred))# After the engine is destroyed, we destroy the plugin. This function is exposed through the binding code in plugin/pyFullyConnected.cpp.fc_factory.destroy_plugin()if __name__ == "__main__":main()
