一、配置简介

windows10 + GTX 1050Ti 4GB显存

cuda 11.5

TensorRT-8.6.1.6

visual studio 2019

.onnx模型，input为float32的{1,1,N,3}；output为float32的{N,7}

二、代码示例

#include <vector>
#include <iostream>
#include <fstream>
#include "utils.h"

/*
* 用于记录TensorRT的相关操作日志
*/
class Logger : public nvinfer1::ILogger {
public:
	void log(Severity severity, const char* msg) noexcept override {
		using namespace std;
		string s;
		bool printMsg = true;
		switch (severity) {
		case Severity::kINTERNAL_ERROR:
			s = "INTERNAL_ERROR";
			break;
		case Severity::kERROR:
			s = "ERROR";
			break;
		case Severity::kWARNING:
			s = "WARNING";
			break;
		case Severity::kINFO:
			s = "INFO";
			printMsg = m_printVerbose;
			break;
		case Severity::kVERBOSE:
			s = "VERBOSE";
			printMsg = m_printVerbose;
			break;
		}
		if (printMsg)
			std::cout << s << ": " << msg << endl;
	}
public:
	void setPrintVerbose(bool printVerbose) {
		this->m_printVerbose = printVerbose;
	};

private:
	bool m_printVerbose = true;
};
Logger gLogger;
const wchar_t* trtModelName = L"model.trt";

int main() {
	/*
	* 1.读取序列化的trt模型
	*/
	std::ifstream trtModelFile(trtModelName, std::ios_base::in | std::ios_base::binary);
	if (!trtModelFile)
		throw "TRT Model Path Error!";
	trtModelFile.seekg(0, std::ios::end);
	int m_size = (int)trtModelFile.tellg();
	trtModelFile.seekg(0, std::ios::beg);

	char* p_modelBuff = new char[m_size];

	trtModelFile.read(p_modelBuff, m_size);
	trtModelFile.close();

	/*
	* 2.构造推理环境
	*/
	nvinfer1::IRuntime* runtime = nvinfer1::createInferRuntime(gLogger);				// Create an instance of a safe::IRuntime class. 创建运行实例
	nvinfer1::ICudaEngine* engine = runtime->deserializeCudaEngine((void*)p_modelBuff, m_size, NULL);	// Deserialize an engine from a byte array. 反序列化引擎
	
	if (engine == NULL)
		throw "Deserialize model failed!\n";

	nvinfer1::IExecutionContext* p_context = engine->createExecutionContext();			// Create an execution context. 创建执行语境
	// 创建cuda流
	cudaStream_t stream;
	cudaStreamCreate(&stream);
	
	/*
	* 3.创建输入输出
	*/
	pcl::PointCloud<pcl::PointXYZ>::Ptr cloud(new pcl::PointCloud<pcl::PointXYZ>);
	std::string pointcloudPath = "pointcloud.xyz";
	ReadCloudFromTxt(pointcloudPath, cloud);
	NormalizePoints(cloud);
	int size = cloud->points.size();

	float* host_input = new float[size * 3];
	float* host_output = new float[size * 7];

	for (int i = 0; i < size; ++i) {				// 展平x y z x y z…
		host_input[i * 3] = cloud->points[i].x;
		host_input[i * 3 + 1] = cloud->points[i].y;
		host_input[i * 3 + 2] = cloud->points[i].z;
	}

	//for (int i = 0; i < engine->getNbBindings(); i++){
	//	nvinfer1::Dims dims = engine->getBindingDimensions(i);
	//	printf("index %d, dims: (");
	//	for (int d = 0; d < dims.nbDims; d++)
	//	{
	//		if (d < dims.nbDims - 1)
	//			printf("%d,", dims.d[d]);
	//		else
	//			printf("%d", dims.d[d]);
	//	}
	//	printf(")\n");
	//}

	void* buffers[2]{ 0 };
	cudaMalloc(&buffers[0], size * 3 * sizeof(float));
	cudaMalloc(&buffers[1], size * 7 * sizeof(float));      //输出结果的大小
	cudaMemcpyAsync(buffers[0], host_input, size * 3 * sizeof(float), cudaMemcpyHostToDevice);

	/*
	* 4.动态输入维度固定
	*/
	nvinfer1::Dims dims4;
	dims4.nbDims = 4;		// 定义维度
	dims4.d[0] = 1;
	dims4.d[1] = 1;
	dims4.d[2] = size;
	dims4.d[3] = 3;

	p_context->setBindingDimensions(0, dims4);			// Set the dynamic dimensions of an input binding. 动态维度需要在推理时固定！！！切记！！！

	/*
	* 5.执行推理
	*/
	p_context->enqueueV2(buffers, (cudaStream_t)stream, nullptr);
	cudaStreamSynchronize(stream);

	/*
	* 6.数据后处理
	*/
	cudaMemcpyAsync(host_output, buffers[1], size * 7 * sizeof(float), cudaMemcpyDeviceToHost);

	int* label = new int[size];
	for (int i = 0; i < size; i++) {
		label[i] = 0; // 初始化数组元素
	}

	GetFinalLabel(host_output, label, size);

	// 导出可视化
	std::fstream fs;
	fs.open("result_.txt", std::ios::out);
	if (!fs)
		return -1;
	for (size_t i = 0; i < size; i++){	
		fs << cloud->points[i].x << " " << cloud->points[i].y << " " << cloud->points[i].z << " " << label[i] << std::endl;
	}
	fs.close();

	delete[] p_modelBuff;
	delete[] label;
	delete[] host_input;
	delete[] host_output;

	std::cout << "Hello world" << std::endl;
	return 0;
}

注意事项：

1.代码中ReadCloudFromTxt和NormalizePoints函数分别是读取点云文件与对点云做归一化操作；非重点，不同模型需要不同的数据。

2.第4步中，对动态输入做固定是使用动态trt模型的必要步骤，这是区别于静态输入的根本所在。