目录
- 前言
- 1. Tensor封装
- 总结
前言
杜老师推出的 tensorRT从零起步高性能部署 课程，之前有看过一遍，但是没有做笔记，很多东西也忘了。这次重新撸一遍，顺便记记笔记。
本次课程学习 tensorRT 高级-tensor封装，索引计算，内存标记及自动复制
课程大纲可看下面的思维导图
1. Tensor封装
这节我们学习 tensor 的封装，张量是 CNN 中常见的基本单元，尤其是计算偏移量的工作需要封装，其次是内存的复制、分配需要引用 memory 进行封装，避免使用时面对指针不好管控
Tensor 封装主要考虑以下几点：
1. Tensor 的封装是针对输入与输出的，使得对输入或输出的操作更加的便捷
2. Tensor 内部的内存管理则是对 MixMemory 进行了包装
3. Tensor 封装所考虑的，是便于访问，因此有 offset 函数，实现索引的计算
在看代码之前，我们先把握下 tensor 封装的四个重点：
1. 内存的管理，可以使用 MixMemory 解决
2. 内存的复用，依然可以用 MixMemory 解决
3. 内存的 copy，比如说 cpu → \rightarrow → gpu，gpu → \rightarrow → cpu
- 解决方案（从 caffe 上学到的思路）
- a. 定义内存的状态，表示内存当前最新的内容在哪里（GPU/CPU/Init）
- b. 懒分配原则，当你需要使用时，才会考虑分配内存
- c. 获取内存地址，即表示：想拿到最新的数据，比如说 tensor.cpu 表示我想拿到最新的数据，并且把它放到 cpu 上
4. 索引的计算，比如说，我有 5d tensor(B，D，C，H，W)，此时我要获取 B = 1，D = 3，C = 0，H = 5，W = 9 的位置元素属于非常基础且非常频繁的一个能力
了解完重点之后，我们再来看代码：
trt-tensor.hpp
#ifndef TRT_TENSOR_HPP
#define TRT_TENSOR_HPP
#include <string>
#include <memory>
#include <vector>
#include <map>
#include <opencv2/opencv.hpp>
#include "mix-memory.hpp"
struct CUstream_st;
typedef CUstream_st CUStreamRaw;
typedef CUStreamRaw* CUStream;
namespace TRT{
enum class DataHead : int{
Init = 0,
Device = 1,
Host = 2
};
enum class DataType : int {
Float = 0,
Float16 = 1,
Int32 = 2,
UInt8 = 3
};
int data_type_size(DataType dt);
const char* data_head_string(DataHead dh);
const char* data_type_string(DataType dt);
class Tensor {
public:
Tensor(const Tensor& other) = delete;
Tensor& operator = (const Tensor& other) = delete;
explicit Tensor(DataType dtype = DataType::Float, std::shared_ptr<MixMemory> data = nullptr, int device_id = CURRENT_DEVICE_ID);
explicit Tensor(int n, int c, int h, int w, DataType dtype = DataType::Float, std::shared_ptr<MixMemory> data = nullptr, int device_id = CURRENT_DEVICE_ID);
explicit Tensor(int ndims, const int* dims, DataType dtype = DataType::Float, std::shared_ptr<MixMemory> data = nullptr, int device_id = CURRENT_DEVICE_ID);
explicit Tensor(const std::vector<int>& dims, DataType dtype = DataType::Float, std::shared_ptr<MixMemory> data = nullptr, int device_id = CURRENT_DEVICE_ID);
virtual ~Tensor();
int numel() const;
inline int ndims() const{return shape_.size();}
inline int size(int index) const{return shape_[index];}
inline int shape(int index) const{return shape_[index];}
inline int batch() const{return shape_[0];}
inline int channel() const{return shape_[1];}
inline int height() const{return shape_[2];}
inline int width() const{return shape_[3];}
inline DataType type() const { return dtype_; }
inline const std::vector<int>& dims() const { return shape_; }
inline const std::vector<size_t>& strides() const {return strides_;}
inline int bytes() const { return bytes_; }
inline int bytes(int start_axis) const { return count(start_axis) * element_size(); }
inline int element_size() const { return data_type_size(dtype_); }
inline DataHead head() const { return head_; }
std::shared_ptr<Tensor> clone() const;
Tensor& release();
Tensor& set_to(float value);
bool empty() const;
template<typename ... _Args>
int offset(int index, _Args ... index_args) const{
const int index_array[] = {index, index_args...};
return offset_array(sizeof...(index_args) + 1, index_array);
}
int offset_array(const std::vector<int>& index) const;
int offset_array(size_t size, const int* index_array) const;
template<typename ... _Args>
Tensor& resize(int dim_size, _Args ... dim_size_args){
const int dim_size_array[] = {dim_size, dim_size_args...};
return resize(sizeof...(dim_size_args) + 1, dim_size_array);
}
Tensor& resize(int ndims, const int* dims);
Tensor& resize(const std::vector<int>& dims);
Tensor& resize_single_dim(int idim, int size);
int count(int start_axis = 0) const;
int device() const{return device_id_;}
Tensor& to_gpu(bool copy=true);
Tensor& to_cpu(bool copy=true);
inline void* cpu() const { ((Tensor*)this)->to_cpu(); return data_->cpu(); }
inline void* gpu() const { ((Tensor*)this)->to_gpu(); return data_->gpu(); }
template<typename DType> inline const DType* cpu() const { return (DType*)cpu(); }
template<typename DType> inline DType* cpu() { return (DType*)cpu(); }
template<typename DType, typename ... _Args>
inline DType* cpu(int i, _Args&& ... args) { return cpu<DType>() + offset(i, args...); }
template<typename DType> inline const DType* gpu() const { return (DType*)gpu(); }
template<typename DType> inline DType* gpu() { return (DType*)gpu(); }
template<typename DType, typename ... _Args>
inline DType* gpu(int i, _Args&& ... args) { return gpu<DType>() + offset(i, args...); }
template<typename DType, typename ... _Args>
inline DType& at(int i, _Args&& ... args) { return *(cpu<DType>() + offset(i, args...)); }
std::shared_ptr<MixMemory> get_data() const {return data_;}
std::shared_ptr<MixMemory> get_workspace() const {return workspace_;}
Tensor& set_workspace(std::shared_ptr<MixMemory> workspace) {workspace_ = workspace; return *this;}
bool is_stream_owner() const {return stream_owner_;}
CUStream get_stream() const{return stream_;}
Tensor& set_stream(CUStream stream, bool owner=false){stream_ = stream; stream_owner_ = owner; return *this;}
Tensor& set_mat (int n, const cv::Mat& image);
Tensor& set_norm_mat(int n, const cv::Mat& image, float mean[3], float std[3]);
cv::Mat at_mat(int n = 0, int c = 0) { return cv::Mat(height(), width(), CV_32F, cpu<float>(n, c)); }
Tensor& synchronize();
const char* shape_string() const{return shape_string_;}
const char* descriptor() const;
Tensor& copy_from_gpu(size_t offset, const void* src, size_t num_element, int device_id = CURRENT_DEVICE_ID);
Tensor& copy_from_cpu(size_t offset, const void* src, size_t num_element);
void reference_data(const std::vector<int>& shape, void* cpu_data, size_t cpu_size, void* gpu_data, size_t gpu_size, DataType dtype);
/**
# 以下代码是python中加载Tensor
import numpy as np
def load_tensor(file):
with open(file, "rb") as f:
binary_data = f.read()
magic_number, ndims, dtype = np.frombuffer(binary_data, np.uint32, count=3, offset=0)
assert magic_number == 0xFCCFE2E2, f"{file} not a tensor file."
dims = np.frombuffer(binary_data, np.uint32, count=ndims, offset=3 * 4)
if dtype == 0:
np_dtype = np.float32
elif dtype == 1:
np_dtype = np.float16
else:
assert False, f"Unsupport dtype = {dtype}, can not convert to numpy dtype"
return np.frombuffer(binary_data, np_dtype, offset=(ndims + 3) * 4).reshape(*dims)
**/
bool save_to_file(const std::string& file) const;
bool load_from_file(const std::string& file);
private:
Tensor& compute_shape_string();
Tensor& adajust_memory_by_update_dims_or_type();
void setup_data(std::shared_ptr<MixMemory> data);
private:
std::vector<int> shape_;
std::vector<size_t> strides_;
size_t bytes_ = 0;
DataHead head_ = DataHead::Init;
DataType dtype_ = DataType::Float;
CUStream stream_ = nullptr;
bool stream_owner_ = false;
int device_id_ = 0;
char shape_string_[100];
char descriptor_string_[100];
std::shared_ptr<MixMemory> data_;
std::shared_ptr<MixMemory> workspace_;
};
}; // namespace TRT
#endif // TRT_TENSOR_HPP
trt-tensor.cpp
#include "trt-tensor.hpp"
#include <algorithm>
#include <cuda_runtime.h>
#include "cuda-tools.hpp"
#include "simple-logger.hpp"
using namespace cv;
using namespace std;
namespace TRT{
int data_type_size(DataType dt){
switch (dt) {
case DataType::Float: return sizeof(float);
case DataType::Int32: return sizeof(int);
case DataType::UInt8: return sizeof(uint8_t);
default: {
INFOE("Not support dtype: %d", dt);
return -1;
}
}
}
inline static int get_device(int device_id){
if(device_id != CURRENT_DEVICE_ID){
CUDATools::check_device_id(device_id);
return device_id;
}
checkRuntime(cudaGetDevice(&device_id));
return device_id;
}
const char* data_head_string(DataHead dh){
switch(dh){
case DataHead::Init: return "Init";
case DataHead::Device: return "Device";
case DataHead::Host: return "Host";
default: return "Unknow";
}
}
const char* data_type_string(DataType dt){
switch(dt){
case DataType::Float: return "Float32";
case DataType::Float16: return "Float16";
case DataType::Int32: return "Int32";
case DataType::UInt8: return "UInt8";
default: return "Unknow";
}
}
Tensor::Tensor(int n, int c, int h, int w, DataType dtype, shared_ptr<MixMemory> data, int device_id) {
this->dtype_ = dtype;
this->device_id_ = get_device(device_id);
descriptor_string_[0] = 0;
setup_data(data);
resize(n, c, h, w);
}
Tensor::Tensor(const std::vector<int>& dims, DataType dtype, shared_ptr<MixMemory> data, int device_id){
this->dtype_ = dtype;
this->device_id_ = get_device(device_id);
descriptor_string_[0] = 0;
setup_data(data);
resize(dims);
}
Tensor::Tensor(int ndims, const int* dims, DataType dtype, shared_ptr<MixMemory> data, int device_id) {
this->dtype_ = dtype;
this->device_id_ = get_device(device_id);
descriptor_string_[0] = 0;
setup_data(data);
resize(ndims, dims);
}
Tensor::Tensor(DataType dtype, shared_ptr<MixMemory> data, int device_id){
shape_string_[0] = 0;
descriptor_string_[0] = 0;
this->device_id_ = get_device(device_id);
dtype_ = dtype;
setup_data(data);
}
Tensor::~Tensor() {
release();
}
const char* Tensor::descriptor() const{
char* descriptor_ptr = (char*)descriptor_string_;
int device_id = device();
snprintf(descriptor_ptr, sizeof(descriptor_string_),
"Tensor:%p, %s, %s, CUDA:%d",
data_.get(),
data_type_string(dtype_),
shape_string_,
device_id
);
return descriptor_ptr;
}
Tensor& Tensor::compute_shape_string(){
// clean string
shape_string_[0] = 0;
char* buffer = shape_string_;
size_t buffer_size = sizeof(shape_string_);
for(int i = 0; i < shape_.size(); ++i){
int size = 0;
if(i < shape_.size() - 1)
size = snprintf(buffer, buffer_size, "%d x ", shape_[i]);
else
size = snprintf(buffer, buffer_size, "%d", shape_[i]);
buffer += size;
buffer_size -= size;
}
return *this;
}
void Tensor::reference_data(const vector<int>& shape, void* cpu_data, size_t cpu_size, void* gpu_data, size_t gpu_size, DataType dtype){
dtype_ = dtype;
data_->reference_data(cpu_data, cpu_size, gpu_data, gpu_size);
setup_data(data_);
resize(shape);
}
void Tensor::setup_data(shared_ptr<MixMemory> data){
data_ = data;
if(data_ == nullptr){
data_ = make_shared<MixMemory>(device_id_);
}else{
device_id_ = data_->device_id();
}
head_ = DataHead::Init;
if(data_->cpu()){
head_ = DataHead::Host;
}
if(data_->gpu()){
head_ = DataHead::Device;
}
}
shared_ptr<Tensor> Tensor::clone() const{
auto new_tensor = make_shared<Tensor>(shape_, dtype_);
if(head_ == DataHead::Init)
return new_tensor;
if(head_ == DataHead::Host){
memcpy(new_tensor->cpu(), this->cpu(), this->bytes_);
}else if(head_ == DataHead::Device){
CUDATools::AutoDevice auto_device_exchange(device());
checkRuntime(cudaMemcpyAsync(new_tensor->gpu(), this->gpu(), bytes_, cudaMemcpyDeviceToDevice, stream_));
}
return new_tensor;
}
Tensor& Tensor::copy_from_gpu(size_t offset, const void* src, size_t num_element, int device_id){
if(head_ == DataHead::Init)
to_gpu(false);
size_t offset_location = offset * element_size();
if(offset_location >= bytes_){
INFOE("Offset location[%lld] >= bytes_[%lld], out of range", offset_location, bytes_);
return *this;
}
size_t copyed_bytes = num_element * element_size();
size_t remain_bytes = bytes_ - offset_location;
if(copyed_bytes > remain_bytes){
INFOE("Copyed bytes[%lld] > remain bytes[%lld], out of range", copyed_bytes, remain_bytes);
return *this;
}
if(head_ == DataHead::Device){
int current_device_id = get_device(device_id);
int gpu_device_id = device();
if(current_device_id != gpu_device_id){
checkRuntime(cudaMemcpyPeerAsync(gpu<unsigned char>() + offset_location, gpu_device_id, src, current_device_id, copyed_bytes, stream_));
//checkRuntime(cudaMemcpyAsync(gpu<unsigned char>() + offset_location, src, copyed_bytes, cudaMemcpyDeviceToDevice, stream_));
}
else{
checkRuntime(cudaMemcpyAsync(gpu<unsigned char>() + offset_location, src, copyed_bytes, cudaMemcpyDeviceToDevice, stream_));
}
}else if(head_ == DataHead::Host){
CUDATools::AutoDevice auto_device_exchange(this->device());
checkRuntime(cudaMemcpyAsync(cpu<unsigned char>() + offset_location, src, copyed_bytes, cudaMemcpyDeviceToHost, stream_));
}else{
INFOE("Unsupport head type %d", head_);
}
return *this;
}
Tensor& Tensor::copy_from_cpu(size_t offset, const void* src, size_t num_element){
if(head_ == DataHead::Init)
to_cpu(false);
size_t offset_location = offset * element_size();
if(offset_location >= bytes_){
INFOE("Offset location[%lld] >= bytes_[%lld], out of range", offset_location, bytes_);
return *this;
}
size_t copyed_bytes = num_element * element_size();
size_t remain_bytes = bytes_ - offset_location;
if(copyed_bytes > remain_bytes){
INFOE("Copyed bytes[%lld] > remain bytes[%lld], out of range", copyed_bytes, remain_bytes);
return *this;
}
if(head_ == DataHead::Device){
CUDATools::AutoDevice auto_device_exchange(this->device());
checkRuntime(cudaMemcpyAsync((char*)data_->gpu() + offset_location, src, copyed_bytes, cudaMemcpyHostToDevice, stream_));
}else if(head_ == DataHead::Host){
//checkRuntime(cudaMemcpyAsync((char*)data_->cpu() + offset_location, src, copyed_bytes, cudaMemcpyHostToHost, stream_));
memcpy((char*)data_->cpu() + offset_location, src, copyed_bytes);
}else{
INFOE("Unsupport head type %d", head_);
}
return *this;
}
Tensor& Tensor::release() {
data_->release_all();
shape_.clear();
bytes_ = 0;
head_ = DataHead::Init;
if(stream_owner_ && stream_ != nullptr){
CUDATools::AutoDevice auto_device_exchange(this->device());
checkRuntime(cudaStreamDestroy(stream_));
}
stream_owner_ = false;
stream_ = nullptr;
return *this;
}
bool Tensor::empty() const{
return data_->cpu() == nullptr && data_->gpu() == nullptr;
}
int Tensor::count(int start_axis) const {
if(start_axis >= 0 && start_axis < shape_.size()){
int size = 1;
for (int i = start_axis; i < shape_.size(); ++i)
size *= shape_[i];
return size;
}else{
return 0;
}
}
Tensor& Tensor::resize(const std::vector<int>& dims) {
return resize(dims.size(), dims.data());
}
int Tensor::numel() const{
int value = shape_.empty() ? 0 : 1;
for(int i = 0; i < shape_.size(); ++i){
value *= shape_[i];
}
return value;
}
Tensor& Tensor::resize_single_dim(int idim, int size){
assert(idim >= 0 && idim < shape_.size());
auto new_shape = shape_;
new_shape[idim] = size;
return resize(new_shape);
}
Tensor& Tensor::resize(int ndims, const int* dims) {
vector<int> setup_dims(ndims);
for(int i = 0; i < ndims; ++i){
int dim = dims[i];
if(dim == -1){
assert(ndims == shape_.size());
dim = shape_[i];
}
setup_dims[i] = dim;
}
this->shape_ = setup_dims;
// strides = element_size
this->strides_.resize(setup_dims.size());
size_t prev_size = element_size();
size_t prev_shape = 1;
for(int i = (int)strides_.size() - 1; i >= 0; --i){
if(i + 1 < strides_.size()){
prev_size = strides_[i+1];
prev_shape = shape_[i+1];
}
strides_[i] = prev_size * prev_shape;
}
this->adajust_memory_by_update_dims_or_type();
this->compute_shape_string();
return *this;
}
Tensor& Tensor::adajust_memory_by_update_dims_or_type(){
int needed_size = this->numel() * element_size();
if(needed_size > this->bytes_){
head_ = DataHead::Init;
}
this->bytes_ = needed_size;
return *this;
}
Tensor& Tensor::synchronize(){
CUDATools::AutoDevice auto_device_exchange(this->device());
checkRuntime(cudaStreamSynchronize(stream_));
return *this;
}
Tensor& Tensor::to_gpu(bool copy) {
if (head_ == DataHead::Device)
return *this;
head_ = DataHead::Device;
data_->gpu(bytes_);
if (copy && data_->cpu() != nullptr) {
CUDATools::AutoDevice auto_device_exchange(this->device());
checkRuntime(cudaMemcpyAsync(data_->gpu(), data_->cpu(), bytes_, cudaMemcpyHostToDevice, stream_));
}
return *this;
}
Tensor& Tensor::to_cpu(bool copy) {
if (head_ == DataHead::Host)
return *this;
head_ = DataHead::Host;
data_->cpu(bytes_);
if (copy && data_->gpu() != nullptr) {
CUDATools::AutoDevice auto_device_exchange(this->device());
checkRuntime(cudaMemcpyAsync(data_->cpu(), data_->gpu(), bytes_, cudaMemcpyDeviceToHost, stream_));
checkRuntime(cudaStreamSynchronize(stream_));
}
return *this;
}
template<typename _T>
static inline void memset_any_type(_T* ptr, size_t count, _T value){
for (size_t i = 0; i < count; ++i)
*ptr++ = value;
}
Tensor& Tensor::set_to(float value) {
int c = count();
if (dtype_ == DataType::Float) {
memset_any_type(cpu<float>(), c, value);
}
else if(dtype_ == DataType::Int32) {
memset_any_type(cpu<int>(), c, (int)value);
}
else if(dtype_ == DataType::UInt8) {
memset_any_type(cpu<uint8_t>(), c, (uint8_t)value);
}
else{
INFOE("Unsupport type: %d", dtype_);
}
return *this;
}
int Tensor::offset_array(size_t size, const int* index_array) const{
assert(size <= shape_.size());
int value = 0;
for(int i = 0; i < shape_.size(); ++i){
if(i < size)
value += index_array[i];
if(i + 1 < shape_.size())
value *= shape_[i+1];
}
return value;
}
int Tensor::offset_array(const std::vector<int>& index_array) const{
return offset_array(index_array.size(), index_array.data());
}
Tensor& Tensor::set_norm_mat(int n, const cv::Mat& image, float mean[3], float std[3]) {
assert(image.channels() == 3 && !image.empty() && type() == DataType::Float);
assert(ndims() == 4 && n < shape_[0]);
to_cpu(false);
int width = shape_[3];
int height = shape_[2];
float scale = 1 / 255.0;
cv::Mat inputframe = image;
if(inputframe.size() != cv::Size(width, height))
cv::resize(inputframe, inputframe, cv::Size(width, height));
if(CV_MAT_DEPTH(inputframe.type()) != CV_32F){
inputframe.convertTo(inputframe, CV_32F, scale);
}
cv::Mat ms[3];
for (int c = 0; c < 3; ++c)
ms[c] = cv::Mat(height, width, CV_32F, cpu<float>(n, c));
split(inputframe, ms);
assert((void*)ms[0].data == (void*)cpu<float>(n));
for (int c = 0; c < 3; ++c)
ms[c] = (ms[c] - mean[c]) / std[c];
return *this;
}
Tensor& Tensor::set_mat(int n, const cv::Mat& _image) {
cv::Mat image = _image;
assert(!image.empty() && CV_MAT_DEPTH(image.type()) == CV_32F && type() == DataType::Float);
assert(shape_.size() == 4 && n < shape_[0] && image.channels() == shape_[1]);
to_cpu(false);
int width = shape_[3];
int height = shape_[2];
if (image.size() != cv::Size(width, height))
cv::resize(image, image, cv::Size(width, height));
if (image.channels() == 1) {
memcpy(cpu<float>(n), image.data, width * height * sizeof(float));
return *this;
}
vector<cv::Mat> ms(image.channels());
for (int i = 0; i < ms.size(); ++i)
ms[i] = cv::Mat(height, width, CV_32F, cpu<float>(n, i));
cv::split(image, &ms[0]);
assert((void*)ms[0].data == (void*)cpu<float>(n));
return *this;
}
bool Tensor::save_to_file(const std::string& file) const{
if(empty()) return false;
FILE* f = fopen(file.c_str(), "wb");
if(f == nullptr) return false;
int ndims = this->ndims();
unsigned int head[3] = {0xFCCFE2E2, ndims, static_cast<unsigned int>(dtype_)};
fwrite(head, 1, sizeof(head), f);
fwrite(shape_.data(), 1, sizeof(shape_[0]) * shape_.size(), f);
fwrite(cpu(), 1, bytes_, f);
fclose(f);
return true;
}
bool Tensor::load_from_file(const std::string& file){
FILE* f = fopen(file.c_str(), "rb");
if(f == nullptr){
INFOE("Open %s failed.", file.c_str());
return false;
}
unsigned int head[3] = {0};
fread(head, 1, sizeof(head), f);
if(head[0] != 0xFCCFE2E2){
fclose(f);
INFOE("Invalid tensor file %s, magic number mismatch", file.c_str());
return false;
}
int ndims = head[1];
auto dtype = (TRT::DataType)head[2];
vector<int> dims(ndims);
fread(dims.data(), 1, ndims * sizeof(dims[0]), f);
this->dtype_ = dtype;
this->resize(dims);
fread(this->cpu(), 1, bytes_, f);
fclose(f);
return true;
}
};
头文件中首先定义了两个枚举类，分别是 DataHead 和 DataType，用于表示数据目前的状态（初始化、CPU、GPU）和数据类型，然后定义了一个 Tensor 类，用于获取和设置张量的属性，如维度、形状、类型等
Tensor 存在多个构造函数，比如接受 n,c,h,w 作为 shape，MixMemory 作为 data，关于 tensor 内存的管理和复用我们是通过上节课封装的 MixMemory 来实现的，因此我们重点来看 tensor 内存的拷贝和索引的计算
tensor 内存的拷贝是通过 to_cpu 和 to_gpu 函数实现的，在 to_cpu 函数中，它首先会检查当前内存数据的状态，如果当前的数据已经在 CPU 上了，它会立即返回，避免再次拷贝；如果数据不在 CPU 上，那就说明在 GPU 上，此时我们会将 head_ 设置为 CPU，通过调用 data_->cpu(byes_) 分配一块 CPU 内存，调用的是 MixMemory 的 cpu 方法，像之前说的一样，如果当前分配的 cpu 内存大小已经满足了要求，则它会直接拿之前的内存，实现内存的复用。最后通过 CUDA 的 cudaMemcpyAsync 函数从 GPU 异步复制数据到 CPU，同时还引入了流同步，确保复制完成
tensor 索引的计算是通过 offset 函数来完成的，该函数接收一个变参，然后将它转换成一个数组送到 offset_array 中进行索引计算，遵循我们之前讲解的左乘右加原则，它直接返回的是偏移量，代表 tensor 中的某个元素在内存中的位置
我们接下来看下 main.cpp 中的不同
// tensor的建立并不会立即分配内存，而是在第一次需要使用的时候进行分配
TRT::Tensor input_data({input_batch, input_channel, input_height, input_width}, TRT::DataType::Float);
// 为input关联stream，使得在同一个pipeline中执行复制操作
input_data.set_stream(stream);
首先是 input_data 的构建，我们可以利用封装好的 Tensor 来实现
// 利用opencv mat的内存地址引用，实现input与mat的关联，然后利用split函数一次性完成mat到input的复制
cv::Mat channel_based[3];
for(int i = 0; i < 3; ++i)
// 注意这里 2 - i是实现bgr -> rgb的方式
// 这里cpu提供的参数0是表示batch的索引是0，第二个参数表示通道的索引，因此获取的是0, 2-i通道的地址
// 而tensor最大的好处就是帮忙计算索引，否则手动计算就得写很多代码
channel_based[i] = cv::Mat(input_height, input_width, CV_32F, input_data.cpu<float>(0, 2-i));
cv::split(image, channel_based);
其次是预处理部分，我们采用了 opencv 中的 split 方法将三个通道分割开，这样的性能更好，这也是从 caffe 中学习到了
// 如果不写，input_data.gpu获取gpu地址时会自动进行复制
// 目的就是把内存复制变为隐式进行
input_data.to_gpu();
...
float* bindings[] = {input_data.gpu<float>(), output_data.gpu<float>()};
还有一点要注意的是，我们不需要显性的去做内存复制，在需要的时候会隐式的完成，比如 input_data.gpu 获取 gpu 的数据时，会将 cpu 的数据自动复制到 gpu 上
以上就是 tensor 封装的分析，封装好的 tensor 不用显性的去调用 cuda 的 API，接口更加高级，性能更好。更多细节还是需要多去看
这是一个相对复杂的版本的 tensor 封装，其实也可以只考虑之前提过的四条来封装 tensor，也能解决绝大部分问题
总结
本次课程学习了 tensor 的封装，对 tensor 的封装重点考虑四个方面：内存的管理、内存的复用、内存的拷贝以及索引的计算，其中前面两个可以用 MixMemory 来解决，内存的拷贝通过定义内存的状态，懒分配原则实现的，而索引的计算则是通过之前说的左乘右加完成的。tensor 的封装使得输入和输出的操作更加的便捷，索引的计算也更加的方便。
![【汇编语言】使用DS和[address]实现字的传送](https://img-blog.csdnimg.cn/5c1755fcc93e48fdba7f452d01814922.png)
