Caffe常用算子GPU和CPU对比

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。

Prelu

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