【caffe】syncedmem.cpp/syncedmem.hpp代码解析

该模块主要用于同步CPU端和GPU端内存数据

#ifndef CAFFE_SYNCEDMEM_HPP_
#define CAFFE_SYNCEDMEM_HPP_

#include <cstdlib>

#include "caffe/common.hpp"

namespace caffe {

// If CUDA is available and in GPU mode, host memory will be allocated pinned,
// using cudaMallocHost. It avoids dynamic pinning for transfers (DMA).
// The improvement in performance seems negligible in the single GPU case,
// but might be more significant for parallel training. Most importantly,
// it improved stability for large models on many GPUs.
/*
使用的内存分为Pageable(交换页)和Pinned(Page-locked)两种模式
CaffeMallocHost和CaffeFreeHost都是为主机分配内存，而不是为显卡分配内存。
当启用GPU模式时，优先使用cuda函数cudaMallocHost、cudaFreeHost为主机分配和释放内存
而不用malloc、free，因为cudaMallocHost、cudaFreeHost效率更高
*/
inline void CaffeMallocHost(void** ptr, size_t size, bool* use_cuda) {
#ifndef CPU_ONLY
  if (Caffe::mode() == Caffe::GPU) {
    CUDA_CHECK(cudaMallocHost(ptr, size));
    *use_cuda = true;
    return;
  }
#endif
  *ptr = malloc(size);
  *use_cuda = false;
  CHECK(*ptr) << "host allocation of size " << size << " failed";
}

inline void CaffeFreeHost(void* ptr, bool use_cuda) {
#ifndef CPU_ONLY
  if (use_cuda) {
    CUDA_CHECK(cudaFreeHost(ptr));
    return;
  }
#endif
  free(ptr);
}


/**
 * @brief Manages memory allocation and synchronization between the host (CPU)
 *        and device (GPU).
 *
 * TODO(dox): more thorough description.
 */
class SyncedMemory {
 public:
  SyncedMemory()
      : cpu_ptr_(NULL), gpu_ptr_(NULL), size_(0), head_(UNINITIALIZED),
        own_cpu_data_(false), cpu_malloc_use_cuda_(false), own_gpu_data_(false),
        gpu_device_(-1) {}
  explicit SyncedMemory(size_t size)
      : cpu_ptr_(NULL), gpu_ptr_(NULL), size_(size), head_(UNINITIALIZED),
        own_cpu_data_(false), cpu_malloc_use_cuda_(false), own_gpu_data_(false),
        gpu_device_(-1) {}
  ~SyncedMemory();
  const void* cpu_data();
  void set_cpu_data(void* data);
  const void* gpu_data();
  void set_gpu_data(void* data);
  void* mutable_cpu_data();
  void* mutable_gpu_data();
  enum SyncedHead { UNINITIALIZED, HEAD_AT_CPU, HEAD_AT_GPU, SYNCED };
  SyncedHead head() { return head_; }
  size_t size() { return size_; }

#ifndef CPU_ONLY
  void async_gpu_push(const cudaStream_t& stream);
#endif

 private:
  void to_cpu();
  void to_gpu();
  void* cpu_ptr_;
  void* gpu_ptr_;
  size_t size_;
  SyncedHead head_;
  bool own_cpu_data_;
  bool cpu_malloc_use_cuda_;
  bool own_gpu_data_;
  int gpu_device_;

  DISABLE_COPY_AND_ASSIGN(SyncedMemory);
};  // class SyncedMemory

}  // namespace caffe

#endif  // CAFFE_SYNCEDMEM_HPP_

-----------------------------------------------------------

#include "caffe/common.hpp"
#include "caffe/syncedmem.hpp"
#include "caffe/util/math_functions.hpp"

namespace caffe {

SyncedMemory::~SyncedMemory() {
  /* 释放cpu端内存 */
  if (cpu_ptr_ && own_cpu_data_) {
    CaffeFreeHost(cpu_ptr_, cpu_malloc_use_cuda_);
  }

  /* 释放gpu端内存 */
#ifndef CPU_ONLY
  if (gpu_ptr_ && own_gpu_data_) {
    int initial_device;
    cudaGetDevice(&initial_device);
    if (gpu_device_ != -1) {
      CUDA_CHECK(cudaSetDevice(gpu_device_));
    }
    CUDA_CHECK(cudaFree(gpu_ptr_));
    cudaSetDevice(initial_device);
  }
#endif  // CPU_ONLY
}

/*
1：当head_=UNINITIALIZED时
则开辟cpu端内存，并将内存初始化为0
head_修改为HEAD_AT_CPU;
当head_=HEAD_AT_GPU时，则将GPU上的数据同步给cpu端
head_修改为SYNCED
当head_=HEAD_AT_CPU或SYNCED时，则不处理
own_cpu_data_表示拥有cpu端数据标志位
2：当head_=HEAD_AT_GPU时:
将gpu端内存同步到cpu端内存
主要同步函数为caffe_gpu_memcpy，在math_functions.cu中实现，代码为：
void caffe_gpu_memcpy(const size_t N, const void* X, void* Y) 
{
	if (X != Y) 
	{
		// 将X指向的数据拷贝给Y //
		CUDA_CHECK(cudaMemcpy(Y, X, N, cudaMemcpyDefault));  // NOLINT(caffe/alt_fn)
	}
}
可知，底层调用了cuda函数cudaMemcpy将gpu_ptr_指向的数据同步给cpu_ptr_

因此to_cpu函数的作用是将gpu端内存数据同步给cpu端内存数据
*/
inline void SyncedMemory::to_cpu() {
  switch (head_) {
  case UNINITIALIZED:
    CaffeMallocHost(&cpu_ptr_, size_, &cpu_malloc_use_cuda_);
    caffe_memset(size_, 0, cpu_ptr_);
    head_ = HEAD_AT_CPU;
    own_cpu_data_ = true;
    break;
  case HEAD_AT_GPU:
#ifndef CPU_ONLY
    if (cpu_ptr_ == NULL) {
      CaffeMallocHost(&cpu_ptr_, size_, &cpu_malloc_use_cuda_);
      own_cpu_data_ = true;
    }
    caffe_gpu_memcpy(size_, gpu_ptr_, cpu_ptr_);
    head_ = SYNCED;
#else
    NO_GPU;
#endif
    break;
  case HEAD_AT_CPU:
  case SYNCED:
    break;
  }
}

/* 将cpu端内存数据同步给gpu端内存 */
inline void SyncedMemory::to_gpu() {
#ifndef CPU_ONLY
  switch (head_) {
  case UNINITIALIZED:
    CUDA_CHECK(cudaGetDevice(&gpu_device_));
    CUDA_CHECK(cudaMalloc(&gpu_ptr_, size_));
    caffe_gpu_memset(size_, 0, gpu_ptr_);
    head_ = HEAD_AT_GPU;
    own_gpu_data_ = true;
    break;
  case HEAD_AT_CPU:
    if (gpu_ptr_ == NULL) {
      CUDA_CHECK(cudaGetDevice(&gpu_device_));
      CUDA_CHECK(cudaMalloc(&gpu_ptr_, size_));
      own_gpu_data_ = true;
    }
    caffe_gpu_memcpy(size_, cpu_ptr_, gpu_ptr_);
    head_ = SYNCED;
    break;
  case HEAD_AT_GPU:
  case SYNCED:
    break;
  }
#else
  NO_GPU;
#endif
}
/* 将GPU数据同步至CPU */
const void* SyncedMemory::cpu_data() {
  to_cpu();
  return (const void*)cpu_ptr_;
}
/* 更新cpu端数据，释放cpu端内存 */
void SyncedMemory::set_cpu_data(void* data) {
  CHECK(data);
  if (own_cpu_data_) {
    CaffeFreeHost(cpu_ptr_, cpu_malloc_use_cuda_);
  }
  cpu_ptr_ = data;
  head_ = HEAD_AT_CPU;
  own_cpu_data_ = false;
}
/* 将CPU数据同步至GPU */
const void* SyncedMemory::gpu_data() {
#ifndef CPU_ONLY
  to_gpu();
  return (const void*)gpu_ptr_;
#else
  NO_GPU;
  return NULL;
#endif
}
/* 更新gpu端数据，释放gpu端内存 */
void SyncedMemory::set_gpu_data(void* data) {
#ifndef CPU_ONLY
  CHECK(data);
  if (own_gpu_data_) {
    int initial_device;
    cudaGetDevice(&initial_device);
    if (gpu_device_ != -1) {
      CUDA_CHECK(cudaSetDevice(gpu_device_));
    }
    CUDA_CHECK(cudaFree(gpu_ptr_));
    cudaSetDevice(initial_device);
  }
  gpu_ptr_ = data;
  head_ = HEAD_AT_GPU;
  own_gpu_data_ = false;
#else
  NO_GPU;
#endif
}
/* 将GPU数据同步至CPU后，head_指向HEAD_AT_CPU，返回cpu端内存指针 */
void* SyncedMemory::mutable_cpu_data() {
  to_cpu();
  head_ = HEAD_AT_CPU;
  return cpu_ptr_;
}
/* 将CPU数据同步至GPU后，head_指向HEAD_AT_GPU，返回gpu端内存指针 */
void* SyncedMemory::mutable_gpu_data() {
#ifndef CPU_ONLY
  to_gpu();
  head_ = HEAD_AT_GPU;
  return gpu_ptr_;
#else
  NO_GPU;
  return NULL;
#endif
}

/* 基于cuda函数cudaMemcpyAsync，将cpu_ptr_数据异步的传给gpu_ptr_ */
#ifndef CPU_ONLY
void SyncedMemory::async_gpu_push(const cudaStream_t& stream) {
  CHECK(head_ == HEAD_AT_CPU);
  if (gpu_ptr_ == NULL) {
    CUDA_CHECK(cudaGetDevice(&gpu_device_));
    CUDA_CHECK(cudaMalloc(&gpu_ptr_, size_));
    own_gpu_data_ = true;
  }
  const cudaMemcpyKind put = cudaMemcpyHostToDevice;
  CUDA_CHECK(cudaMemcpyAsync(gpu_ptr_, cpu_ptr_, size_, put, stream));
  // Assume caller will synchronize on the stream before use
  head_ = SYNCED;
}
#endif

}  // namespace caffe