笔者在caffe上做各种实验有一段时间了，但一直都只是在修改配置文件或者存在某些新的idea却难以实现的地步，很多时候实现一些idea需要深入到底层去修改或者添加一下新的layer等，这样也就要求对caffe的底层源码有一个较为深层的理解，这个系列的博客将记录和分享在学习caffe源码中的体会和过程，作为一个EE转CS的编程菜鸟，如有错误，希望读者指正。
caffe大致上可以分为Blob, Layer, Net, Solver这四个大的模块。

Solver: An interface for classes that perform     optimization on Nets
Net: Connects Layers together into a directed acyclic graph (DAG) specified by a NetParameter
Layer: An interface for the units of computation which can be composed into a Net
Blob: A wrapper around SyncedMemory holders serving as the basic computational unit through which Layers, Nets, and Solvers interact

Blob作为caffe数据流通的一个基本数据结构，层与层之间的数据是通过blob来实现的，具有cpu和gpu之间同步的能力，可以看作一个四维数组(num, channels, heights, width)。主要代码位于/home/xxx/caffe/include/blob.hpp。

*1. 主要变量*

shared_ptr<SyncedMemory> data_;
shared_ptr<SyncedMemory> diff_;
shared_ptr<SyncedMemory> shape_data_;
vector<int> shape_;
int count_;
int capacity_;

Blob作为一个基本的数据结构，内部变量并不复杂，首先是data指针，采用的是boost库的一个智能指针shared_ptr。这一部分主要用来申请内存存储data用于正向传播。diff_用来存储误差，shape_data 和 shape_都是用来存储Blob的形状，只是一个是老版本，一个是新版本，count表示Blob中的元素个数，也就是 num * channel * height * width，capacity表示当前元素的个数。
caffe基于blob进行存储和交换数据，为了便于优化，Blob提供了统一的内存接口存储某种类型的数据，可以根据cpu到gpu的同步需要，屏蔽cpu/gpu在计算上的开销。

*2. 主要函数*

template <typename Dtype>
class Blob {
 public:
  Blob()
       : data_(), diff_(), count_(0), capacity_(0) {}

  /// @brief Deprecated; use <code>Blob(const vector<int>& shape)</code>.
  explicit Blob(const int num, const int channels, const int height,
      const int width);
  explicit Blob(const vector<int>& shape);

  /// @brief Deprecated; use <code>Reshape(const vector<int>& shape)</code>.
  void Reshape(const int num, const int channels, const int height,
      const int width);

Blob作为一个基础的数据结构，其中构造函数开辟了一个内存空间来存储数据，Reshape()函数在layer中的reshape和forward中来adjust dimension，同时会改变Blob的大小，此时内存将会被重新分配。
blob中的index是可以从负坐标开始读的，这一点类似于python，对于blob中的四个基本变量，num channel height width，可以直接通过shape(0), shape(1), shape(2), shape(3)。

计算offset

inline int offset(const int n, const int c = , const int h = , const int w = )
inline int offset(const vector<int>& indices)

offset计算的方式也支持两种方式，直接指定n,c,h,w或者放到一个vector中进行计算，偏差是根据对应的n,c,h,w，返回的offset是

从一个blob中copy数据 ，通过开关控制是否copy_diff,如果是False则copy data。reshape控制是否需要reshape。

inline Dtype data_at(const int n, const int c, const int h, const int w)
inline Dtype diff_at(const int n, const int c, const int h, const int w)
inline Dtype data_at(const vector<int>& index)
inline Dtype diff_at(const vector<int>& index)
inline const shared_ptr<SyncedMemory>& data()
inline const shared_ptr<SyncedMemory>& diff()

这一部分函数主要通过给定的位置访问数据，根据位置计算与数据起始的偏差offset，在通过cpu_data*指针获得地址。

const Dtype* cpu_data() const;
void set_cpu_data(Dtype* data);
const int* gpu_shape() const;
const Dtype* gpu_data() const;
const Dtype* cpu_diff() const;
const Dtype* gpu_diff() const;
Dtype* mutable_cpu_data();
Dtype* mutable_gpu_data();
Dtype* mutable_cpu_diff();
Dtype* mutable_gpu_diff();

data就是存储前向传递的信息的数据，diff指的是神经网络在反向传播时候的梯度。

void FromProto(const BlobProto& proto, bool reshape = true);
void ToProto(BlobProto* proto, bool write_diff = false) const;

这两个函数主要是将数据序列化，存储到BlobProto，这里说到Proto是谷歌的一个数据序列化的存储格式，可以实现语言、平台无关、可扩展的序列化结构数据格式。Caffe里面数据的存储都采用这一结构.
总结一下Blob.hpp的一些函数：
Reshape()可以改变一个Blob的大小；
ReshapeLike()为data和diff重新分配了一块空间，大小和另一个Blob一样。
Num_axes()返回的是Blob的大小
Count()计算得到 num * channels * height * width
Offset() 可得到输入Blob数据的(n,c,h,w)的偏移量位置
CopyFrom()从source拷贝数据，copy_diff来作为标志区分拷贝的是data还是diff。
FromProto()从proto读数据进俩，其实就是反序列化
ToProto()把blob数据保存到proto中
ShareData() / ShareDiff() 从other的blob复制data和diff的值

#ifndef CAFFE_BLOB_HPP_
#define CAFFE_BLOB_HPP_

#include <algorithm>
#include <string>
#include <vector>

#include "caffe/common.hpp"  //单例化caffe类，并且封装了boost和cuda随机数生成的函数，提供了统一接口
#include "caffe/proto/caffe.pb.h"
#include "caffe/syncedmem.hpp"

/*
主要是分配内存和释放内存。class yncedMemory定义了内存分配管理和CPU与GPU之间同步的函数。
 Blob会使用SyncedMem自动决定什么时候去copy data以提高运行效率，通常情况是仅当gnu或cpu修改后有copy操作。
 */

const int kMaxBlobAxes = ; //在头文件中为它添加 extern 声明,以使其能被多个文件共享

namespace caffe {

/**
 * @brief A wrapper around SyncedMemory holders serving as the basic
 *        computational unit through which Layer%s, Net%s, and Solver%s
 *        interact.
 *
 * TODO(dox): more thorough description.
 */
template <typename Dtype>   //模板类，虚拟类型Dtype
class Blob {
 public:
  Blob() //构造函数：初始化列表 {空函数体}
       : data_(), diff_(), count_(), capacity_() {}


  //当构造函数被声明 explicit 时,编译器将不使用它作为转换操作符。
  /// @brief Deprecated; use <code>Blob(const vector<int>& shape)</code>.
  explicit Blob(const int num, const int channels, const int height,
      const int width);   //可以通过设置数据维度（N,C,H,W）初始化
  //const 传递过来的参数在函数内不可以改变(无意义，因为本身就是形参)


  //const引用参数在函数内为常量不可变
  explicit Blob(const vector<int>& shape); //也可以通过传入vector<int>直接传入维数

  /// @brief Deprecated; use <code>Reshape(const vector<int>& shape)</code>.
  void Reshape(const int num, const int channels, const int height,
      const int width);


  /**
   * @brief Change the dimensions of the blob, allocating new memory if
   *        necessary.
   *
   * This function can be called both to create an initial allocation
   * of memory, and to adjust the dimensions of a top blob during Layer::Reshape
   * or Layer::Forward. When changing the size of blob, memory will only be
   * reallocated if sufficient memory does not already exist, and excess memory
   * will never be freed.
   *
   * Note that reshaping an input blob and immediately calling Net::Backward is
   *
   * an error; either Net::Forward or Net::Reshape need to be called to
   * propagate the new input shape to higher layers.
   */

  void Reshape(const vector<int>& shape);
  void Reshape(const BlobShape& shape);
  void ReshapeLike(const Blob& other);


  //内联函数 通过内联函数，编译器不需要跳转到内存其他地址去执行函数调用，也不需要保留函数调用时的现场数据。
  // const 成员函数，任何不会修改数据成员的函数都应该声明为const 类型。

  // 输出blob的形状
  inline string shape_string() const { //
    ostringstream stream;
    for (int i = ; i < shape_.size(); ++i) {
      stream << shape_[i] << " ";
    }
    stream << "(" << count_ << ")";
    return stream.str();
  }

  inline const vector<int>& shape() const { return shape_; }
  /**
   * @brief Returns the dimension of the index-th axis (or the negative index-th
   *        axis from the end, if index is negative).
   *
   * @param index the axis index, which may be negative as it will be
   *        "canonicalized" using CanonicalAxisIndex.
   *        Dies on out of range index.
   */
  inline int shape(int index) const {  //根据索引返回维数，对于维数(N,C,H,W),shape(0)返回N,shape(-1)返回W。
    return shape_[CanonicalAxisIndex(index)];
  }

  inline int num_axes() const { return shape_.size(); }    //返回Blob维度数，对于维数(N,C,H,W)，返回4

  inline int count() const { return count_; }               //返回Blob维度数，对于维数(N,C,H,W)，返回N×C×H×W

  /**
   * @brief Compute the volume of a slice; i.e., the product of dimensions
   *        among a range of axes.
   *
   * @param start_axis The first axis to include in the slice.
   *
   * @param end_axis The first axis to exclude from the slice.
   */
  //对于维数(N,C,H,W)，count(0, 3)返回N×C×H
  inline int count(int start_axis, int end_axis) const {
    CHECK_LE(start_axis, end_axis);
    CHECK_GE(start_axis, );
    CHECK_GE(end_axis, );
    CHECK_LE(start_axis, num_axes());
    CHECK_LE(end_axis, num_axes());
    int count = ;
    for (int i = start_axis; i < end_axis; ++i) {
      count *= shape(i);
    }
    return count;
  }


  /**
   * @brief Compute the volume of a slice spanning from a particular first
   *        axis to the final axis.
   *
   * @param start_axis The first axis to include in the slice.
   */
  //对于维数(N,C,H,W)，count(1)返回C×H×W
  inline int count(int start_axis) const {
    return count(start_axis, num_axes());
  }


  /**
   * @brief Returns the 'canonical' version of a (usually) user-specified axis,
   *        allowing for negative indexing (e.g., -1 for the last axis).
   *
   * @param axis_index the axis index.
   *        If 0 <= index < num_axes(), return index.
   *        If -num_axes <= index <= -1, return (num_axes() - (-index)),
   *        e.g., the last axis index (num_axes() - 1) if index == -1,
   *        the second to last if index == -2, etc.
   *        Dies on out of range index.
   */
  inline int CanonicalAxisIndex(int axis_index) const {
    CHECK_GE(axis_index, -num_axes())
        << "axis " << axis_index << " out of range for " << num_axes()
        << "-D Blob with shape " << shape_string();
    CHECK_LT(axis_index, num_axes())
        << "axis " << axis_index << " out of range for " << num_axes()
        << "-D Blob with shape " << shape_string();
    if (axis_index < ) {
      return axis_index + num_axes();
    }
    return axis_index;
  }



  /// @brief Deprecated legacy shape accessor num: use shape(0) instead.
  inline int num() const { return LegacyShape(); }
  /// @brief Deprecated legacy shape accessor channels: use shape(1) instead.
  inline int channels() const { return LegacyShape(); }
  /// @brief Deprecated legacy shape accessor height: use shape(2) instead.
  inline int height() const { return LegacyShape(); }
  /// @brief Deprecated legacy shape accessor width: use shape(3) instead.
  inline int width() const { return LegacyShape(); }
  inline int LegacyShape(int index) const {
    CHECK_LE(num_axes(), )
        << "Cannot use legacy accessors on Blobs with > 4 axes.";
    CHECK_LT(index, );
    CHECK_GE(index, -);
    if (index >= num_axes() || index < -num_axes()) {
      // Axis is out of range, but still in [0, 3] (or [-4, -1] for reverse
      // indexing) -- this special case simulates the one-padding used to fill
      // extraneous axes of legacy blobs.
      return ;
    }
    return shape(index);
  }



  inline int offset(const int n, const int c = , const int h = ,
      const int w = ) const {    //计算物理偏移量，(n,c,h,w)的偏移量为((n∗C+c)∗H+h)∗W+w
    CHECK_GE(n, );
    CHECK_LE(n, num());
    CHECK_GE(channels(), );
    CHECK_LE(c, channels());
    CHECK_GE(height(), );
    CHECK_LE(h, height());
    CHECK_GE(width(), );
    CHECK_LE(w, width());
    return ((n * channels() + c) * height() + h) * width() + w;
  }

  inline int offset(const vector<int>& indices) const {
    CHECK_LE(indices.size(), num_axes());
    int offset = ;
    for (int i = ; i < num_axes(); ++i) {
      offset *= shape(i);
      if (indices.size() > i) {
        CHECK_GE(indices[i], );
        CHECK_LT(indices[i], shape(i));
        offset += indices[i];
      }
    }
    return offset;
  }


  /**
   * @brief Copy from a source Blob.
   *
   * @param source the Blob to copy from
   * @param copy_diff: if false, copy the data; if true, copy the diff
   * @param reshape: if false, require this Blob to be pre-shaped to the shape
   *        of other (and die otherwise); if true, Reshape this Blob to other's
   *        shape if necessary
   */
  void CopyFrom(const Blob<Dtype>& source, bool copy_diff = false,   //从source拷贝数据， copy_diff来作为标志区分是拷贝data还是diff。
      bool reshape = false);

  inline Dtype data_at(const int n, const int c, const int h,
      const int w) const {
    return cpu_data()[offset(n, c, h, w)];
  }

  inline Dtype diff_at(const int n, const int c, const int h,
      const int w) const {
    return cpu_diff()[offset(n, c, h, w)];
  }

  inline Dtype data_at(const vector<int>& index) const {
    return cpu_data()[offset(index)];
  }

  inline Dtype diff_at(const vector<int>& index) const {
    return cpu_diff()[offset(index)];
  }

  inline const shared_ptr<SyncedMemory>& data() const {
    CHECK(data_);
    return data_;
  }

  inline const shared_ptr<SyncedMemory>& diff() const {
    CHECK(diff_);
    return diff_;
  }\


  /*
    // 假定数据在 CPU 上进行初始化，我们有一个 blob
    const Dtype* foo;
    Dtype* bar;
    foo = blob.gpu_data(); // 数据从 CPU 复制到 GPU
    foo = blob.cpu_data(); // 没有数据复制，两者都有最新的内容
    bar = blob.mutable_gpu_data(); // 没有数据复制
    // ... 一些操作 ...
    bar = blob.mutable_gpu_data(); // 仍在 GPU，没有数据复制
    foo = blob.cpu_data(); // 由于 GPU 修改了数值，数据从 GPU 复制到 CPU
    foo = blob.gpu_data(); //没有数据复制，两者都有最新的内容
    bar = blob.mutable_cpu_data(); // 依旧没有数据复制
    bar = blob.mutable_gpu_data(); //数据从 CPU 复制到 GPU
    bar = blob.mutable_cpu_data(); //数据从 GPU 复制到 CPU

   */

  const Dtype* cpu_data() const;  //数据访问，const方式只读，不允许改写数据
  void set_cpu_data(Dtype* data);
  const int* gpu_shape() const;
  const Dtype* gpu_data() const;
  const Dtype* cpu_diff() const;
  const Dtype* gpu_diff() const;
  Dtype* mutable_cpu_data();      //mutable方式可改写数据（对diff_的访问也是类似的）
  Dtype* mutable_gpu_data();
  Dtype* mutable_cpu_diff();
  Dtype* mutable_gpu_diff();
  void Update();

  void FromProto(const BlobProto& proto, bool reshape = true); //从proto读数据进来，其实就是反序列化
  void ToProto(BlobProto* proto, bool write_diff = false) const; //blob数据保存到proto中

  /// @brief Compute the sum of absolute values (L1 norm) of the data.
  Dtype asum_data() const;
  /// @brief Compute the sum of absolute values (L1 norm) of the diff.
  Dtype asum_diff() const;
  /// @brief Compute the sum of squares (L2 norm squared) of the data.
  Dtype sumsq_data() const;
  /// @brief Compute the sum of squares (L2 norm squared) of the diff.
  Dtype sumsq_diff() const;

  /// @brief Scale the blob data by a constant factor.
  void scale_data(Dtype scale_factor);
  /// @brief Scale the blob diff by a constant factor.
  void scale_diff(Dtype scale_factor);

  /**
   * @brief Set the data_ shared_ptr to point to the SyncedMemory holding the
   *        data_ of Blob other -- useful in Layer%s which simply perform a copy
   *        in their Forward pass.
   *
   * This deallocates the SyncedMemory holding this Blob's data_, as
   * shared_ptr calls its destructor when reset with the "=" operator.
   */
  void ShareData(const Blob& other); //Blob& other 赋值给data_

  /**
   * @brief Set the diff_ shared_ptr to point to the SyncedMemory holding the
   *        diff_ of Blob other -- useful in Layer%s which simply perform a copy
   *        in their Forward pass.
   *
   * This deallocates the SyncedMemory holding this Blob's diff_, as
   * shared_ptr calls its destructor when reset with the "=" operator.
   */
  void ShareDiff(const Blob& other); //Blob& other 赋值给diff_

  bool ShapeEquals(const BlobProto& other);

 protected:
  shared_ptr<SyncedMemory> data_; //存储前向传递数据
  shared_ptr<SyncedMemory> diff_; //存储反向传递梯度
  shared_ptr<SyncedMemory> shape_data_;
  vector<int> shape_;  //参数维度
  int count_; //Blob存储的元素个数（shape_所有元素乘积）
  int capacity_;//当前Blob的元素个数（控制动态分配）

  DISABLE_COPY_AND_ASSIGN(Blob);
};  // class Blob

}  // namespace caffe

#endif  // CAFFE_BLOB_HPP_