caffe cuda 程序分析

caffe 用cuda 来进行加速， 看完caffe CPU的程序，再去看 CUDA程序， 变得很容易了

在layer文件中， cuda 程序和CPU程序类似，只是调用math_function.hpp函数不一样，一个调用cpu 一个调用gpu，下面以absval_layer.cu为例：

#include <vector>

#include "caffe/layer.hpp"
#include "caffe/util/math_functions.hpp"
#include "caffe/vision_layers.hpp"

namespace caffe {

template <typename Dtype>
void AbsValLayer<Dtype>::Forward_gpu(
    const vector<Blob<Dtype>*>& bottom, const vector<Blob<Dtype>*>& top) {
  const int count = top[0]->count();
  Dtype* top_data = top[0]->mutable_gpu_data();
  // 与cpu程序区别，这里调用GPU 函数
<span style="color:#ff6666;">  caffe_gpu_abs(count, bottom[0]->gpu_data(), top_data);</span>
}

template <typename Dtype>
void AbsValLayer<Dtype>::Backward_gpu(const vector<Blob<Dtype>*>& top,
    const vector<bool>& propagate_down, const vector<Blob<Dtype>*>& bottom) {
  const int count = top[0]->count();
  const Dtype* top_diff = top[0]->gpu_diff();
  if (propagate_down[0]) {
    const Dtype* bottom_data = bottom[0]->gpu_data();
    Dtype* bottom_diff = bottom[0]->mutable_gpu_diff();

<pre name="code" class="cpp"><span style="color:#ff0000;"> // 与cpu程序区别，这里调用GPU 函数</span>

caffe_gpu_sign(count, bottom_data, bottom_diff); caffe_gpu_mul(count, bottom_diff, top_diff, bottom_diff); }}INSTANTIATE_LAYER_GPU_FUNCS(AbsValLayer);} // namespace caffe

从上边可以看出，重点在 math_function上， 看那边cuda 程序如何封装的即可。

layer中 还有一种方式

#include <algorithm>
#include <vector>

#include "caffe/layer.hpp"
#include "caffe/vision_layers.hpp"

// if x > 0  Computes y = x + log(1 + exp(-x))  ;  otherwise y =log(1 + exp(-x))

namespace caffe {

const float kBNLL_THRESHOLD = 50.;

先封装一个函数，用于前向
<span style="background-color: rgb(255, 204, 51);">template <typename Dtype>
__global__ void BNLLForward(const int n, const Dtype* in, Dtype* out) {
  CUDA_KERNEL_LOOP(index, n) {
    // if x > 0  Computes y = x + log(1 + exp(-x))  ;  otherwise y =log(1 + exp(-x))
    out[index] = in[index] > 0 ?
        in[index] + log(1. + exp(-in[index])) :
        log(1. + exp(in[index]));
  }
}</span>

template <typename Dtype>
void BNLLLayer<Dtype>::Forward_gpu(const vector<Blob<Dtype>*>& bottom,
    const vector<Blob<Dtype>*>& top) {
  const Dtype* bottom_data = bottom[0]->gpu_data();
  Dtype* top_data = top[0]->mutable_gpu_data();
  const int count = bottom[0]->count();

  // 这里与cpu 不同 
  // NOLINT_NEXT_LINE(whitespace/operators)
  BNLLForward<Dtype><<<CAFFE_GET_BLOCKS(count), CAFFE_CUDA_NUM_THREADS>>>(
      count, bottom_data, top_data);
  CUDA_POST_KERNEL_CHECK;
}

<span style="background-color: rgb(255, 204, 51);">template <typename Dtype>
__global__ void BNLLBackward(const int n, const Dtype* in_diff,
    const Dtype* in_data, Dtype* out_diff) {
  CUDA_KERNEL_LOOP(index, n) {
    // if x < 50 ,x = x, else x = 50;
	// if x > 0 ,y' = exp(x) / (1 + exp(x));
    Dtype expval = exp(min(in_data[index], Dtype(kBNLL_THRESHOLD)));
    out_diff[index] = in_diff[index] * expval / (expval + 1.);
  }
}</span>

template <typename Dtype>
void BNLLLayer<Dtype>::Backward_gpu(const vector<Blob<Dtype>*>& top,
    const vector<bool>& propagate_down,
    const vector<Blob<Dtype>*>& bottom) {
  if (propagate_down[0]) {
    const Dtype* bottom_data = bottom[0]->gpu_data();
    const Dtype* top_diff = top[0]->gpu_diff();
    Dtype* bottom_diff = bottom[0]->mutable_gpu_diff();
    const int count = bottom[0]->count();

   <span style="background-color: rgb(102, 204, 204);"> //　与cpu区别
    // NOLINT_NEXT_LINE(whitespace/operators)
    BNLLBackward<Dtype><<<CAFFE_GET_BLOCKS(count), CAFFE_CUDA_NUM_THREADS>>>(
        count, top_diff, bottom_data, bottom_diff);</span>
    CUDA_POST_KERNEL_CHECK;
  }
}

INSTANTIATE_LAYER_GPU_FUNCS(BNLLLayer);


}  // namespace caffe