pointnet C++推理部署--tensorrt框架

classification

如上图所示，由于直接export出的onnx文件有两个输出节点，不方便处理，所以编写脚本删除不需要的输出节点193：

import onnx


onnx_model = onnx.load("cls.onnx")
graph = onnx_model.graph
 
inputs = graph.input
for input in inputs:
    print('input',input.name)
    
outputs = graph.output
for output in outputs:
    print('output',output.name)

graph.output.remove(outputs[1])
onnx.save(onnx_model, 'cls_modified.onnx')

C++推理代码：

#include <iostream>
#include <fstream>
#include <vector>
#include <algorithm>
#include <cuda_runtime.h>
#include <NvInfer.h>
#include <NvInferRuntime.h>
#include <NvOnnxParser.h>


const int point_num = 1024;


void pc_normalize(std::vector<float>& points)
{
	float mean_x = 0, mean_y = 0, mean_z = 0;
	for (size_t i = 0; i < point_num; ++i)
	{
		mean_x += points[3 * i];
		mean_y += points[3 * i + 1];
		mean_z += points[3 * i + 2];
	}
	mean_x /= point_num;
	mean_y /= point_num;
	mean_z /= point_num;

	for (size_t i = 0; i < point_num; ++i)
	{
		points[3 * i] -= mean_x;
		points[3 * i + 1] -= mean_y;
		points[3 * i + 2] -= mean_z;
	}

	float m = 0;
	for (size_t i = 0; i < point_num; ++i)
	{
		if (sqrt(pow(points[3 * i], 2) + pow(points[3 * i + 1], 2) + pow(points[3 * i + 2], 2)) > m)
			m = sqrt(pow(points[3 * i], 2) + pow(points[3 * i + 1], 2) + pow(points[3 * i + 2], 2));
	}

	for (size_t i = 0; i < point_num; ++i)
	{
		points[3 * i] /= m;
		points[3 * i + 1] /= m;
		points[3 * i + 2] /= m;
	}
}

class TRTLogger : public nvinfer1::ILogger 
{
public:
	virtual void log(Severity severity, nvinfer1::AsciiChar const* msg) noexcept override
	{
		if (severity <= Severity::kINFO) 
			printf(msg);
	}
} logger;

std::vector<unsigned char> load_file(const std::string& file) 
{
	std::ifstream in(file, std::ios::in | std::ios::binary);
	if (!in.is_open())
		return {};

	in.seekg(0, std::ios::end);
	size_t length = in.tellg();

	std::vector<uint8_t> data;
	if (length > 0) 
	{
		in.seekg(0, std::ios::beg);
		data.resize(length);
		in.read((char*)& data[0], length);
	}
	in.close();
	return data;
}

void classfier(std::vector<float> & points)
{
	TRTLogger logger;
	nvinfer1::ICudaEngine* engine;

//#define BUILD_ENGINE

#ifdef  BUILD_ENGINE
	nvinfer1::IBuilder* builder = nvinfer1::createInferBuilder(logger);
	nvinfer1::IBuilderConfig* config = builder->createBuilderConfig();
	nvinfer1::INetworkDefinition* network = builder->createNetworkV2(1);

	nvonnxparser::IParser* parser = nvonnxparser::createParser(*network, logger);
	if (!parser->parseFromFile("cls_modified.onnx", 1))
	{
		printf("Failed to parser onnx\n");
		return;
	}

	int maxBatchSize = 1;
	config->setMaxWorkspaceSize(1 << 32);

	engine = builder->buildEngineWithConfig(*network, *config);
	if (engine == nullptr) {
		printf("Build engine failed.\n");
		return;
	}

	nvinfer1::IHostMemory* model_data = engine->serialize();
	FILE* f = fopen("cls.engine", "wb");
	fwrite(model_data->data(), 1, model_data->size(), f);
	fclose(f);

	model_data->destroy();
	parser->destroy();
	engine->destroy();
	network->destroy();
	config->destroy();
	builder->destroy();
#endif  

	auto engine_data = load_file("cls.engine");
	nvinfer1::IRuntime* runtime = nvinfer1::createInferRuntime(logger);
	engine = runtime->deserializeCudaEngine(engine_data.data(), engine_data.size());
	if (engine == nullptr)
	{
		printf("Deserialize cuda engine failed.\n");
		runtime->destroy();
		return;
	}

	nvinfer1::IExecutionContext* execution_context = engine->createExecutionContext();
	cudaStream_t stream = nullptr;
	cudaStreamCreate(&stream);

	float* input_data_host = nullptr;
	const size_t input_numel = 1 * 3 * point_num;
	cudaMallocHost(&input_data_host, input_numel * sizeof(float));
	for (size_t i = 0; i < 3; i++)
	{
		for (size_t j = 0; j < point_num; j++)
		{
			input_data_host[point_num * i + j] = points[3 * j + i];
		}
	}

	float* input_data_device = nullptr;
	float output_data_host[10];
	float* output_data_device = nullptr;
	cudaMalloc(&input_data_device, input_numel * sizeof(float));
	cudaMalloc(&output_data_device, sizeof(output_data_host));
	cudaMemcpyAsync(input_data_device, input_data_host, input_numel * sizeof(float), cudaMemcpyHostToDevice, stream);
	float* bindings[] = { input_data_device, output_data_device };

	bool success = execution_context->enqueueV2((void**)bindings, stream, nullptr);
	cudaMemcpyAsync(output_data_host, output_data_device, sizeof(output_data_host), cudaMemcpyDeviceToHost, stream);
	cudaStreamSynchronize(stream);

	int predict_label = std::max_element(output_data_host, output_data_host + 10) - output_data_host;
	std::cout << "\npredict_label: " << predict_label << std::endl;

	cudaStreamDestroy(stream);
	execution_context->destroy();
	engine->destroy();
	runtime->destroy();
}


int main()
{
	std::vector<float> points;
	std::ifstream infile;
	float x, y, z, nx, ny, nz;
	char ch;
	infile.open("bed_0610.txt");
	for (size_t i = 0; i < point_num; i++)
	{
		infile >> x >> ch >> y >> ch >> z >> ch >> nx >> ch >> ny >> ch >> nz;
		points.push_back(x);
		points.push_back(y);
		points.push_back(z);
	}
	infile.close();

	pc_normalize(points);

	classfier(points);

	return 0;
}

其中推理引擎的构建也可以直接使用tensorrt的bin目录下的trtexec.exe。

LZ也实现了cuda版本的前处理代码，但似乎效率比cpu前处理还低。可能是数据量不够大吧（才10^3数量级），而且目前LZ的cuda水平也只是入门阶段...

#include <iostream>
#include <fstream>
#include <vector>
#include <algorithm>
#include <cuda_runtime.h>
#include <cuda_runtime_api.h>
#include <NvInfer.h>
#include <NvInferRuntime.h>
#include <NvOnnxParser.h>


const int point_num = 1024;
const int thread_num = 1024;
const int block_num = 1;

__global__ void array_sum(float* data, float* val, int N)
{
	__shared__ double share_dTemp[thread_num];
	const int nStep = gridDim.x * blockDim.x;
	const int tid = blockIdx.x * blockDim.x + threadIdx.x;
	double dTempSum = 0.0;
	for (int i = tid; i < N; i += nStep)
	{
		dTempSum += data[i];
	}
	share_dTemp[threadIdx.x] = dTempSum;
	__syncthreads();

	for (int i = blockDim.x / 2; i != 0; i /= 2)
	{
		if (threadIdx.x < i)
		{
			share_dTemp[threadIdx.x] += share_dTemp[threadIdx.x + i];
		}
		__syncthreads();
	}

	if (0 == threadIdx.x)
	{
		atomicAdd(val, share_dTemp[0]);
	}
}

__global__ void array_sub(float* data, float val, int N)
{
	const int tid = blockIdx.x * blockDim.x + threadIdx.x;
	const int nStep = blockDim.x * gridDim.x;

	for (int i = tid; i < N; i += nStep)
	{
		data[i] = data[i] - val;
	}
}

__global__ void array_L2(float* in, float* out, int N)
{
	const int tid = blockIdx.x * blockDim.x + threadIdx.x;
	const int nStep = blockDim.x * gridDim.x;

	for (int i = tid; i < N; i += nStep)
	{
		out[i] = sqrt(pow(in[i], 2) + pow(in[i + N], 2) + pow(in[i + 2 * N], 2));
	}
}

__global__ void array_max(float* mem, int numbers) 
{
	int tid = threadIdx.x;
	int idof = blockIdx.x * blockDim.x;
	int idx = tid + idof;
	extern __shared__ float tep[];
	if (idx >= numbers) return;
	tep[tid] = mem[idx];
	unsigned int bi = 0;
	for (int s = 1; s < blockDim.x; s = (s << 1))
	{
		unsigned int kid = tid << (bi + 1);
		if ((kid + s) >= blockDim.x || (idof + kid + s) >= numbers) break;
		tep[kid] = tep[kid] > tep[kid + s] ? tep[kid] : tep[kid + s];
		++bi;
		__syncthreads();
	}
	if (tid == 0) 
	{
		mem[blockIdx.x] = tep[0];
	}
}

__global__ void array_div(float* data, float val, int N)
{
	const int tid = blockIdx.x * blockDim.x + threadIdx.x;
	const int nStep = blockDim.x * gridDim.x;

	for (int i = tid; i < N; i += nStep)
	{
		data[i] = data[i] / val;
	}
}

void pc_normalize_gpu(float* points)
{
	float *mean_x = NULL,  *mean_y = NULL,  *mean_z = NULL;
	cudaMalloc((void**)& mean_x, sizeof(float));
	cudaMalloc((void**)& mean_y, sizeof(float));
	cudaMalloc((void**)& mean_z, sizeof(float));
	array_sum << <thread_num, block_num >> > (points + 0 * point_num, mean_x, point_num);
	array_sum << <thread_num, block_num >> > (points + 1 * point_num, mean_y, point_num);
	array_sum << <thread_num, block_num >> > (points + 2 * point_num, mean_z, point_num);

	float mx, my, mz;
	cudaMemcpy(&mx, mean_x, sizeof(float), cudaMemcpyDeviceToHost);
	cudaMemcpy(&my, mean_y, sizeof(float), cudaMemcpyDeviceToHost);
	cudaMemcpy(&mz, mean_z, sizeof(float), cudaMemcpyDeviceToHost);

	array_sub << <thread_num, block_num >> > (points + 0 * point_num, mx / point_num, point_num);
	array_sub << <thread_num, block_num >> > (points + 1 * point_num, my / point_num, point_num);
	array_sub << <thread_num, block_num >> > (points + 2 * point_num, mz / point_num, point_num);
	//float* pts = (float*)malloc(sizeof(float) * point_num);
	//cudaMemcpy(pts, points, sizeof(float) * point_num, cudaMemcpyDeviceToHost);
	//for (size_t i = 0; i < point_num; i++)
	//{
	//	std::cout << pts[i] << std::endl;
	//}

	float* L2 = NULL;
	cudaMalloc((void**)& L2, sizeof(float) * point_num);
	array_L2 << <thread_num, block_num >> > (points, L2, point_num);
	//float* l2 = (float*)malloc(sizeof(float) * point_num);
	//cudaMemcpy(l2, L2, sizeof(float) * point_num, cudaMemcpyDeviceToHost);
	//for (size_t i = 0; i < point_num; i++)
	//{
	//	std::cout << l2[i] << std::endl;
	//}

	int tmp_num = point_num;
	int share_size = sizeof(float) * thread_num;
	int block_num = (tmp_num + thread_num - 1) / thread_num;
	do {
		array_max << <block_num, thread_num, share_size >> > (L2, thread_num);
		tmp_num = block_num;
		block_num = (tmp_num + thread_num - 1) / thread_num;
	} while (tmp_num > 1);

	float max;
	cudaMemcpy(&max, L2, sizeof(float), cudaMemcpyDeviceToHost);
	//std::cout << max << std::endl;

	array_div << <thread_num, block_num >> > (points + 0 * point_num, max, point_num);
	array_div << <thread_num, block_num >> > (points + 1 * point_num, max, point_num);
	array_div << <thread_num, block_num >> > (points + 2 * point_num, max, point_num);


}

class TRTLogger : public nvinfer1::ILogger 
{
public:
	virtual void log(Severity severity, nvinfer1::AsciiChar const* msg) noexcept override
	{
		if (severity <= Severity::kINFO) 
			printf(msg);
	}
} logger;

std::vector<unsigned char> load_file(const std::string& file) 
{
	std::ifstream in(file, std::ios::in | std::ios::binary);
	if (!in.is_open())
		return {};

	in.seekg(0, std::ios::end);
	size_t length = in.tellg();

	std::vector<uint8_t> data;
	if (length > 0) 
	{
		in.seekg(0, std::ios::beg);
		data.resize(length);
		in.read((char*)& data[0], length);
	}
	in.close();
	return data;
}

void classfier(std::vector<float> & points)
{
	TRTLogger logger;
	nvinfer1::ICudaEngine* engine;

//#define BUILD_ENGINE

#ifdef  BUILD_ENGINE
	nvinfer1::IBuilder* builder = nvinfer1::createInferBuilder(logger);
	nvinfer1::IBuilderConfig* config = builder->createBuilderConfig();
	nvinfer1::INetworkDefinition* network = builder->createNetworkV2(1);

	nvonnxparser::IParser* parser = nvonnxparser::createParser(*network, logger);
	if (!parser->parseFromFile("cls_modified.onnx", 1))
	{
		printf("Failed to parser onnx\n");
		return;
	}

	int maxBatchSize = 1;
	config->setMaxWorkspaceSize(1 << 32);

	engine = builder->buildEngineWithConfig(*network, *config);
	if (engine == nullptr) {
		printf("Build engine failed.\n");
		return;
	}

	nvinfer1::IHostMemory* model_data = engine->serialize();
	FILE* f = fopen("cls.engine", "wb");
	fwrite(model_data->data(), 1, model_data->size(), f);
	fclose(f);

	model_data->destroy();
	parser->destroy();
	engine->destroy();
	network->destroy();
	config->destroy();
	builder->destroy();
#endif  

	auto engine_data = load_file("cls.engine");
	nvinfer1::IRuntime* runtime = nvinfer1::createInferRuntime(logger);
	engine = runtime->deserializeCudaEngine(engine_data.data(), engine_data.size());
	if (engine == nullptr)
	{
		printf("Deserialize cuda engine failed.\n");
		runtime->destroy();
		return;
	}

	nvinfer1::IExecutionContext* execution_context = engine->createExecutionContext();
	cudaStream_t stream = nullptr;
	cudaStreamCreate(&stream);

	float* input_data_host = nullptr;
	const size_t input_numel = 1 * 3 * point_num;
	cudaMallocHost(&input_data_host, input_numel * sizeof(float));
	for (size_t i = 0; i < 3; i++)
	{
		for (size_t j = 0; j < point_num; j++)
		{
			input_data_host[point_num * i + j] = points[3 * j + i];
		}
	}

	float* input_data_device = nullptr;
	float output_data_host[10];
	float* output_data_device = nullptr;
	cudaMalloc(&input_data_device, input_numel * sizeof(float));
	cudaMalloc(&output_data_device, sizeof(output_data_host));
	cudaMemcpyAsync(input_data_device, input_data_host, input_numel * sizeof(float), cudaMemcpyHostToDevice, stream);
	pc_normalize_gpu(input_data_device);
	float* bindings[] = { input_data_device, output_data_device };

	bool success = execution_context->enqueueV2((void**)bindings, stream, nullptr);
	cudaMemcpyAsync(output_data_host, output_data_device, sizeof(output_data_host), cudaMemcpyDeviceToHost, stream);
	cudaStreamSynchronize(stream);

	int predict_label = std::max_element(output_data_host, output_data_host + 10) - output_data_host;
	std::cout << "\npredict_label: " << predict_label << std::endl;

	cudaStreamDestroy(stream);
	execution_context->destroy();
	engine->destroy();
	runtime->destroy();
}


int main()
{
	std::vector<float> points;
	std::ifstream infile;
	float x, y, z, nx, ny, nz;
	char ch;
	infile.open("sofa_0020.txt");
	for (size_t i = 0; i < point_num; i++)
	{
		infile >> x >> ch >> y >> ch >> z >> ch >> nx >> ch >> ny >> ch >> nz;
		points.push_back(x);
		points.push_back(y);
		points.push_back(z);
	}
	infile.close();

	classfier(points);

	return 0;
}