#ifndef CAFFE_SYNCEDMEM_HPP_
#define CAFFE_SYNCEDMEM_HPP_
#include <cstdlib>
#include "caffe/common.hpp"
#include "caffe/util/math_functions.hpp"
namespace caffe {
// Theoretically, CaffeMallocHost and CaffeFreeHost should simply call the
// cudaMallocHost and cudaFree functions in order to create pinned memory.
// However, those codes rely on the existence of a cuda GPU (I don't know
// why that is a must since allocating memory should not be accessing the
// GPU resource, but it just creates an error as of Cuda 5.0) and will cause
// problem when running on a machine without GPU. Thus, we simply define
// these two functions for safety and possible future change if the problem
// of calling cuda functions disappears in a future version.
//
// In practice, although we are creating unpinned memory here, as long as we
// are constantly accessing them the memory pages almost always stays in
// the physical memory (assuming we have large enough memory installed), and
// does not seem to create a memory bottleneck here.
//SynceMem:做内存同步的操作。
//两个全局的内联函数。
//机器支持GPU并且安装了cuda,通过cudaMallocHost分配的hostmemory将会被pinned(内存不会被paged out),
//注:内存有页作为基本的管理单元,分配的内存可以常住在内存空间中对提高效率有帮助,空间不会被别的进程抢占。
//内存越大,能背分配的pinned内存自然也越大。对于单一的GPU提升不显著,但对多个GPU可以显著提高稳定性。
inline void CaffeMallocHost(void** ptr, size_t size) {
*ptr = malloc(size);
CHECK(*ptr) << "host allocation of size " << size << " failed";
}
inline void CaffeFreeHost(void* ptr) {
free(ptr);
}
/**
* @brief Manages memory allocation and synchronization between the host (CPU)
* and device (GPU).
*
* TODO(dox): more thorough description.
*/
class SyncedMemory { //SyncedMemory类,首先是构造函数和析构函数。
public:
SyncedMemory() //参数构造函数:负责初始化
: cpu_ptr_(NULL), gpu_ptr_(NULL), size_(0), head_(UNINITIALIZED),
own_cpu_data_(false) {}
explicit SyncedMemory(size_t size) //带explicit关键字:单个参数构造函数,explicit禁止单参数构造函数的隐式转换
: cpu_ptr_(NULL), gpu_ptr_(NULL), size_(size), head_(UNINITIALIZED),
own_cpu_data_(false) {}
~SyncedMemory(); //其在析构时,也调用CaffeFreeHost
const void* cpu_data(); //获得cpu上data的地址
void set_cpu_data(void* data); //将cpu的data指针指向一个新的区域由data指针传入,并且将原来申请的内存释放。
const void* gpu_data(); //获得gpu数据地址//获得set gpu的数据地址
void* mutable_cpu_data(); //返回cpu上data的指针
void* mutable_gpu_data(); //返回gpu上data指针
enum SyncedHead { UNINITIALIZED, HEAD_AT_CPU, HEAD_AT_GPU, SYNCED }; //SyncedHead是个枚举类型,用来设定head_状态,
SyncedHead head() { return head_; } //head()函数即返回相应的数据状态,
size_t size() { return size_; } //size()函数返回数据大小
private:
void to_cpu();
void to_gpu();
void* cpu_ptr_;
void* gpu_ptr_;
size_t size_;
SyncedHead head_;
bool own_cpu_data_;
//head未被初始化的状态:首先需要分配内存,根据cpu和gpu的情况而定,再将数据从cpu或gpu拷贝到另一处。
//之后函数会重新标记Head状态,数据是否在cpu或者在gpu中。
//cpu_ptr和gpu_ptr分别是cpu和gpu中的数据指针
DISABLE_COPY_AND_ASSIGN(SyncedMemory); //禁止该类的拷贝与赋值
}; // class SyncedMemory
} // namespace caffe
#endif // CAFFE_SYNCEDMEM_HPP_
#define CAFFE_SYNCEDMEM_HPP_
#include <cstdlib>
#include "caffe/common.hpp"
#include "caffe/util/math_functions.hpp"
namespace caffe {
// Theoretically, CaffeMallocHost and CaffeFreeHost should simply call the
// cudaMallocHost and cudaFree functions in order to create pinned memory.
// However, those codes rely on the existence of a cuda GPU (I don't know
// why that is a must since allocating memory should not be accessing the
// GPU resource, but it just creates an error as of Cuda 5.0) and will cause
// problem when running on a machine without GPU. Thus, we simply define
// these two functions for safety and possible future change if the problem
// of calling cuda functions disappears in a future version.
//
// In practice, although we are creating unpinned memory here, as long as we
// are constantly accessing them the memory pages almost always stays in
// the physical memory (assuming we have large enough memory installed), and
// does not seem to create a memory bottleneck here.
//SynceMem:做内存同步的操作。
//两个全局的内联函数。
//机器支持GPU并且安装了cuda,通过cudaMallocHost分配的hostmemory将会被pinned(内存不会被paged out),
//注:内存有页作为基本的管理单元,分配的内存可以常住在内存空间中对提高效率有帮助,空间不会被别的进程抢占。
//内存越大,能背分配的pinned内存自然也越大。对于单一的GPU提升不显著,但对多个GPU可以显著提高稳定性。
inline void CaffeMallocHost(void** ptr, size_t size) {
*ptr = malloc(size);
CHECK(*ptr) << "host allocation of size " << size << " failed";
}
inline void CaffeFreeHost(void* ptr) {
free(ptr);
}
/**
* @brief Manages memory allocation and synchronization between the host (CPU)
* and device (GPU).
*
* TODO(dox): more thorough description.
*/
class SyncedMemory { //SyncedMemory类,首先是构造函数和析构函数。
public:
SyncedMemory() //参数构造函数:负责初始化
: cpu_ptr_(NULL), gpu_ptr_(NULL), size_(0), head_(UNINITIALIZED),
own_cpu_data_(false) {}
explicit SyncedMemory(size_t size) //带explicit关键字:单个参数构造函数,explicit禁止单参数构造函数的隐式转换
: cpu_ptr_(NULL), gpu_ptr_(NULL), size_(size), head_(UNINITIALIZED),
own_cpu_data_(false) {}
~SyncedMemory(); //其在析构时,也调用CaffeFreeHost
const void* cpu_data(); //获得cpu上data的地址
void set_cpu_data(void* data); //将cpu的data指针指向一个新的区域由data指针传入,并且将原来申请的内存释放。
const void* gpu_data(); //获得gpu数据地址//获得set gpu的数据地址
void* mutable_cpu_data(); //返回cpu上data的指针
void* mutable_gpu_data(); //返回gpu上data指针
enum SyncedHead { UNINITIALIZED, HEAD_AT_CPU, HEAD_AT_GPU, SYNCED }; //SyncedHead是个枚举类型,用来设定head_状态,
SyncedHead head() { return head_; } //head()函数即返回相应的数据状态,
size_t size() { return size_; } //size()函数返回数据大小
private:
void to_cpu();
void to_gpu();
void* cpu_ptr_;
void* gpu_ptr_;
size_t size_;
SyncedHead head_;
bool own_cpu_data_;
//head未被初始化的状态:首先需要分配内存,根据cpu和gpu的情况而定,再将数据从cpu或gpu拷贝到另一处。
//之后函数会重新标记Head状态,数据是否在cpu或者在gpu中。
//cpu_ptr和gpu_ptr分别是cpu和gpu中的数据指针
DISABLE_COPY_AND_ASSIGN(SyncedMemory); //禁止该类的拷贝与赋值
}; // class SyncedMemory
} // namespace caffe
#endif // CAFFE_SYNCEDMEM_HPP_