通过整理LeNet、AlexNet、VGG16、googLeNet、ResNet、MLP统计出的常用算子(不包括ReLU)，表格是对比。

Cpu版

Gpu版

  for (int i = 0; i < count; ++i) {

    int c = (i / dim) % channels / div_factor;

    top_data[i] = std::max(bottom_data[i], Dtype(0))

        + slope_data[c] * std::min(bottom_data[i], Dtype(0));

  }

Kernel代码

  for (int i = blockIdx.x * blockDim.x + threadIdx.x; \

       i < (n);       i += blockDim.x * gridDim.x)

{

    int c = (index / dim) % channels / div_factor;

    out[index] = in[index] > 0 ? in[index] : in[index] * slope_data[c];

}

 

Im2col

Cpu版：

Gpu版本

channel_size = height * width;

for (int channel = channels; channel--; data_im += channel_size) {

  for (int kernel_row = 0; kernel_row < kernel_h; kernel_row++) {

    for (int kernel_col = 0; kernel_col < kernel_w; kernel_col++) {

      int input_row = -pad_h + kernel_row * dilation_h;

      for (int output_rows = output_h; output_rows; output_rows--) {

        if (!is_a_ge_zero_and_a_lt_b(input_row, height)) {

          for (int output_cols = output_w; output_cols; output_cols--) {

            *(data_col++) = 0;

          }

        } else {

          int input_col = -pad_w + kernel_col * dilation_w;

          for (int output_col = output_w; output_col; output_col--) {

            if (is_a_ge_zero_and_a_lt_b(input_col, width)) {

              *(data_col++) = data_im[input_row * width + input_col];

            } else {

              *(data_col++) = 0;

            }

            input_col += stride_w;

          }

        }

        input_row += stride_h;

      }

    }

kernel   

for (int i = blockIdx.x * blockDim.x + threadIdx.x; \

       i < (n);       i += blockDim.x * gridDim.x)

 {

    const int h_index = i/ width_col;

    const int h_col = h_index % height_col;

    const int w_col = i % width_col;

    const int c_im = h_index / height_col;

    const int c_col = c_im * kernel_h * kernel_w;

    const int h_offset = h_col * stride_h - pad_h;

    const int w_offset = w_col * stride_w - pad_w;

    Dtype* data_col_ptr = data_col;

    data_col_ptr += (c_col * height_col + h_col) * width_col + w_col;

    const Dtype* data_im_ptr = data_im;

    data_im_ptr += (c_im * height + h_offset) * width + w_offset;

    for (int i = 0; i < kernel_h; ++i) {

      for (int j = 0; j < kernel_w; ++j) {

        int h_im = h_offset + i * dilation_h;

        int w_im = w_offset + j * dilation_w;

        *data_col_ptr =

            (h_im >= 0 && w_im >= 0 && h_im < height && w_im < width) ?

            data_im_ptr[i * dilation_h * width + j * dilation_w] : 0;

        data_col_ptr += height_col * width_col;

      }

    }

  }

Host代码：

  int num_kernels = channels * height_col * width_col;

  im2col_gpu_kernel<Dtype><<<(num_kernels+511)/512, 512>>>( … …)；

 

Pool算子

Cpu版:AVE版本的,MAX类似,缺少Stochastic

Gpu版本

for (int n = 0; n < bottom[0]->num(); ++n) {

  for (int c = 0; c < channels_; ++c) {

    for (int ph = 0; ph < pooled_height_; ++ph) {

      for (int pw = 0; pw < pooled_width_; ++pw) {

        int hstart = ph * stride_h_ - pad_h_;

        int wstart = pw * stride_w_ - pad_w_;

        int hend = min(hstart + kernel_h_, height_);

        int wend = min(wstart + kernel_w_, width_);

        hstart = max(hstart, 0);

        wstart = max(wstart, 0);

        const int pool_index = ph * pooled_width_ + pw;

        for (int h = hstart; h < hend; ++h) {

          for (int w = wstart; w < wend; ++w) {   //AVE

            top_data[ph * pooled_width_ + pw] +=

                    bottom_data[h * width_ + w];

          }

        }

      }

    }

// compute offset

bottom_data += bottom[0]->offset(0, 1);

top_data += top[0]->offset(0, 1);

  }

}

const int w = index % width;

const int h = (index / width) % height;

const int c = (index / width / height) % channels;

const int n = index / width / height / channels;

const int phstart =

         (h + pad_h < kernel_h) ? 0 : (h + pad_h - kernel_h) / stride_h + 1;

const int phend = min((h + pad_h) / stride_h + 1, pooled_height);

const int pwstart =

         (w + pad_w < kernel_w) ? 0 : (w + pad_w - kernel_w) / stride_w + 1;

const int pwend = min((w + pad_w) / stride_w + 1, pooled_width);

Dtype gradient = 0;

const int offset = (n * channels + c) * pooled_height * pooled_width;

const Dtype* const top_diff_slice = top_diff + offset;

if (mask) {

const int* const mask_slice = mask + offset;

  for (int ph = phstart; ph < phend; ++ph) {

   for (int pw = pwstart; pw < pwend; ++pw) {

      if (mask_slice[ph * pooled_width + pw] == h * width + w) {

         gradient += top_diff_slice[ph * pooled_width + pw];

      }

    }

}

}

bottom_diff[index] = gradient;

 

FC算子(InnerProduct)

Cpu版:

Gpu版本：和CPU版本一致

caffe_cpu_gemm<Dtype>(CblasNoTrans, transpose_ ? CblasNoTrans : CblasTrans,M_, N_, K_, (Dtype)1.,bottom_data, weight, (Dtype)0., top_data);

if (bias_term_) {

    caffe_cpu_gemm<Dtype>(CblasNoTrans, CblasNoTrans, M_, N_, 1, (Dtype)1.,bias_multiplier_.cpu_data(),this->blobs_[1]->cpu_data(), (Dtype)1., top_data);

}

 

Dropout算子

Cpu版:

Gpu版本

// Create random numbers

caffe_rng_bernoulli(count, 1. - threshold_, mask);

for (int i = 0; i < count; ++i) {

  top_data[i] = bottom_data[i] * mask[i] * scale_;

}

  

Softmax算子

Cpu版:

Gpu版本

for (int i = 0; i < outer_num_; ++i) {

  // initialize scale_data to the first plane

  caffe_copy(inner_num_, bottom_data + i * dim, scale_data);

  for (int j = 0; j < channels; j++) {

    for (int k = 0; k < inner_num_; k++) {

      scale_data[k] = std::max(scale_data[k],

        bottom_data[i * dim + j * inner_num_ + k]);

    }

  }

  // subtraction

 caffe_cpu_gemm<Dtype>(CblasNoTrans,CblasNoTrans,channels, inner_num_,

        1, -1., sum_multiplier_.cpu_data(), scale_data, 1., top_data);

  // exponentiation

  caffe_exp<Dtype>(dim, top_data, top_data);

  // sum after exp

  caffe_cpu_gemv<Dtype>(CblasTrans, channels, inner_num_, 1.,

      top_data, sum_multiplier_.cpu_data(), 0., scale_data);

  // division

  for (int j = 0; j < channels; j++) {

    caffe_div(inner_num_, top_data, scale_data, top_data);

    top_data += inner_num_;

  }

}

kernel  以kernel_channel_sum为例

int n = index / spatial_dim;

int s = index % spatial_dim;

Dtype sum = 0;

for (int c = 0; c < channels; ++c) {

  sum += data[(n * channels + c) * spatial_dim + s];

}

channel_sum[index] = sum;

Host代码：

kernel_channel_max<Dtype><<<CAFFE_GET_BLOCKS(outer_num_ * inner_num_),CAFFE_CUDA_NUM_THREADS>>>(outer_num_, channels, inner_num_, top_data,scale_data);

kernel_channel_subtract<Dtype><<<CAFFE_GET_BLOCKS(count),CAFFE_CUDA_NUM_THREADS>>>(count, outer_num_, channels, inner_num_,scale_data, top_data);

kernel_exp<Dtype><<<CAFFE_GET_BLOCKS(count), CAFFE_CUDA_NUM_THREADS>>>(count, top_data, top_data);

kernel_channel_sum<Dtype><<<CAFFE_GET_BLOCKS(outer_num_ * inner_num_),CAFFE_CUDA_NUM_THREADS>>>(outer_num_, channels, inner_num_, top_data,scale_data);

kernel_channel_div<Dtype><<<CAFFE_GET_BLOCKS(count),CAFFE_CUDA_NUM_THREADS>>>(count, outer_num_, channels, inner_num_,scale_data, top_data);

 

Sigmoid算子

Cpu版:

Gpu版本

for (int i = 0; i < count; ++i) {

    top_data[i] = sigmoid(bottom_data[i]);

}

Sigmoid：

inline Dtype sigmoid(Dtype x) {

  return 0.5 * tanh(0.5 * x) + 0.5;

}

out[index] = 0.5 * tanh(0.5 * in[index]) + 0.5;

LRN算子

Cpu版:

Gpu版本：和Cpu一样

split_layer_->Forward(bottom, split_top_vec_);

square_layer_->Forward(square_bottom_vec_, square_top_vec_);//PowerLayer

pool_layer_->Forward(square_top_vec_, pool_top_vec_);

power_layer_->Forward(pool_top_vec_, power_top_vec_);

product_layer_->Forward(product_bottom_vec_, top);//EltwiseLayer

 

Split算子

Cpu版:

Gpu版本：和Cpu一样

for (int i = 0; i < top.size(); ++i) {

    top[i]->ShareData(*bottom[0]);

}

ShareData：

void Blob<Dtype>::ShareData(const Blob& other) {

  CHECK_EQ(count_, other.count());

  data_ = other.data();

}

 

Eltwise算子

Cpu版:MAX版本的

Gpu版本

// Initialize

    mask = max_idx_.mutable_cpu_data();

    caffe_set(count, -1, mask);

    caffe_set(count, Dtype(-FLT_MAX), top_data);

    // bottom 0 & 1

    bottom_data_a = bottom[0]->cpu_data();

    bottom_data_b = bottom[1]->cpu_data();

    for (int idx = 0; idx < count; ++idx) {

      if (bottom_data_a[idx] > bottom_data_b[idx]) {

        top_data[idx] = bottom_data_a[idx];  // maxval

        mask[idx] = 0;  // maxid

      } else {

        top_data[idx] = bottom_data_b[idx];  // maxval

        mask[idx] = 1;  // maxid

      }

    }

    // bottom 2++

    for (int blob_idx = 2; blob_idx < bottom.size(); ++blob_idx) {

      bottom_data_b = bottom[blob_idx]->cpu_data();

      for (int idx = 0; idx < count; ++idx) {

        if (bottom_data_b[idx] > top_data[idx]) {

          top_data[idx] = bottom_data_b[idx];  // maxval

          mask[idx] = blob_idx;  // maxid

        }

      }

    }

kernel   

Dtype maxval = -FLT_MAX;

int maxidx = -1;

if (bottom_data_a[index] > bottom_data_b[index]) {

  // only update for very first bottom_data blob (blob_idx == 0)

  if (blob_idx == 0) {

    maxval = bottom_data_a[index];

    top_data[index] = maxval;

    maxidx = blob_idx;

    mask[index] = maxidx;

  }

} else {

  maxval = bottom_data_b[index];

  top_data[index] = maxval;

  maxidx = blob_idx + 1;

  mask[index] = maxidx;

}

Host代码：

mask = max_idx_.mutable_gpu_data();

MaxForward<Dtype><<<CAFFE_GET_BLOCKS(count), CAFFE_CUDA_NUM_THREADS>>>(count,bottom[0]->gpu_data(), bottom[1]->gpu_data(), 0, top_data, mask);

for (int i = 2; i < bottom.size(); ++i) {

MaxForward<Dtype><<<CAFFE_GET_BLOCKS(count), CAFFE_CUDA_NUM_THREADS>>>(count, top_data, bottom[i]->gpu_data(), i-1, top_data, mask);

}

SigmoidCrossEntropyLoss算子

Cpu版:

Gpu版本

for (int i = 0; i < bottom[0]->count(); ++i) {

  const int target_value = static_cast<int>(target[i]);

  if (has_ignore_label_ && target_value == ignore_label_) {

    continue;

  }

  loss -= input_data[i] * (target[i] - (input_data[i] >= 0)) -

      log(1 + exp(input_data[i] - 2 * input_data[i] * (input_data[i] >= 0)));

  ++valid_count;

}

//return max(1.0,normalizer_)

normalizer_ = get_normalizer(normalization_, valid_count);

top[0]->mutable_cpu_data()[0] = loss / normalizer_;

kernel

const int target_value = static_cast<int>(target[i]);

if (has_ignore_label_ && target_value == ignore_label_) {

  loss[i] = 0;

  counts[i] = 0;

} else {

  loss[i] = input_data[i] * (target[i] - (input_data[i] >= 0)) -

      log(1 + exp(input_data[i] - 2 * input_data[i] *

      (input_data[i] >= 0)));

  counts[i] = 1;

}

Host代码：

SigmoidCrossEntropyLossForwardGPU<Dtype><<<CAFFE_GET_BLOCKS(count),CAFFE_CUDA_NUM_THREADS>>>(count, input_data, target, loss_data,

has_ignore_label_, ignore_label_, count_data);

// Only launch another CUDA kernel if we actually need the valid count.

if (normalization_ == LossParameter_NormalizationMode_VALID &&

    has_ignore_label_) {

  caffe_gpu_asum(count, count_data, &valid_count);

} else {

  valid_count = count;

}

Scale算子

Cpu版:

Gpu版本

for (int n = 0; n < outer_dim_; ++n) {

    for (int d = 0; d < scale_dim_; ++d) {

      const Dtype factor = scale_data[d];

      caffe_cpu_scale(inner_dim_, factor, bottom_data, top_data);

      bottom_data += inner_dim_;

      top_data += inner_dim_;

    }

}

caffe_cpu_scale：

for (int i = 0; i < N; ++i) {

  Y[i] += alpha;

}

kernel

const int scale_index = (index / inner_dim) % scale_dim;

out[index] = in[index] * scale[scale_index];

BatchNorm算子

Cpu版:代码太长待分析

Gpu版本

 

实现逻辑与CPU版本一致，只不过所有的函数调用都成了caffe_gpu_gemm这种