src/model/layer/cudnn_dropout.cc

/*
 * Licensed to the Apache Software Foundation (ASF) under one
 * or more contributor license agreements.  See the NOTICE file
 * distributed with this work for additional information
 * regarding copyright ownership.  The ASF licenses this file
 * to you under the Apache License, Version 2.0 (the
 * "License"); you may not use this file except in compliance
 * with the License.  You may obtain a copy of the License at
 *
 *     http://www.apache.org/licenses/LICENSE-2.0
 *
 * Unless required by applicable law or agreed to in writing, software
 * distributed under the License is distributed on an "AS IS" BASIS,
 * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
 * See the License for the specific language governing permissions and
 * limitations under the License.
 */
#include "./cudnn_dropout.h"
#ifdef USE_CUDNN
#include <cudnn.h>
// cudnn dropout is added in cudnn 5
#if CUDNN_MAJOR >= 5

#include <chrono>

#include "./cudnn_utils.h"
#include "singa/utils/logging.h"

namespace singa {
RegisterLayerClass(cudnn_dropout, CudnnDropout);
CudnnDropout::~CudnnDropout() {
  if (drop_desc_ != nullptr)
    CUDNN_CHECK(cudnnDestroyDropoutDescriptor(drop_desc_));
  if (x_desc_ != nullptr) CUDNN_CHECK(cudnnDestroyTensorDescriptor(x_desc_));
  if (y_desc_ != nullptr) CUDNN_CHECK(cudnnDestroyTensorDescriptor(y_desc_));
}

void CudnnDropout::InitCudnn(int size, DataType dtype,
                             std::shared_ptr<Device> dev, Context* ctx) {
  CHECK(!has_init_cudnn_);
  CUDNN_CHECK(cudnnCreateTensorDescriptor(&x_desc_));
  CUDNN_CHECK(cudnnCreateTensorDescriptor(&y_desc_));
  CUDNN_CHECK(cudnnCreateDropoutDescriptor(&drop_desc_));

  CUDNN_CHECK(cudnnSetTensor4dDescriptor(
      x_desc_, CUDNN_TENSOR_NCHW, GetCudnnDataType(dtype), 1, 1, 1, size));
  CUDNN_CHECK(cudnnSetTensor4dDescriptor(
      y_desc_, CUDNN_TENSOR_NCHW, GetCudnnDataType(dtype), 1, 1, 1, size));

  cudnnDropoutGetStatesSize(ctx->cudnn_handle, &state_size_);
  state_ = Tensor(Shape{state_size_}, dev, kChar);
  cudnnDropoutGetReserveSpaceSize(x_desc_, &reserve_size_);
  mask_ = Tensor(Shape{reserve_size_}, dev, kChar);
  // TODO(wangwei) update for async running,
  // where reserve_size_ may not available
  CHECK_EQ(reserve_size_, mask_.MemSize());

  // TODO(wangwei) get seed from ctx or user config?
  auto seed = std::chrono::system_clock::now().time_since_epoch().count();
  cudnnSetDropoutDescriptor(drop_desc_, ctx->cudnn_handle, 1 - dropout_ratio_,
                            state_.block()->mutable_data(), state_size_, seed);
  has_init_cudnn_ = true;
}

const Tensor CudnnDropout::Forward(int flag, const Tensor& input) {
  if (flag & kTrain) {
    auto size = input.Size();
    DataType dtype = input.data_type();
    auto dev = input.device();
    if (!has_init_cudnn_) {
      input.device()->Exec([size, dtype, this, dev](Context* ctx) {
          this->InitCudnn(size, dtype, dev, ctx);
          }, {}, {this->state_.block()}, "InitCudnn");
    } else {
      int n, c, h, w, s;
      cudnnDataType_t type;
      CUDNN_CHECK(cudnnGetTensor4dDescriptor(x_desc_, &type,
            &n, &c, &h, &w, &s, &s, &s, &s));
      if (size != static_cast<size_t>(w))
        input.device()->Exec([size, dtype, this, dev](Context* ctx) {
            this->InitCudnn(size, dtype, dev, ctx);
            }, {}, {this->state_.block()}, "InitCudnn");
    }
    Tensor output;
    output.ResetLike(input);
    output.device()->Exec([input, output, this](Context* ctx) {
      Block* inblock = input.block(), * outblock = output.block(),
             * mblock = mask_.block();
      cudnnDropoutForward(ctx->cudnn_handle, this->drop_desc_, this->x_desc_,
                          inblock->data(), this->y_desc_,
                          outblock->mutable_data(), mblock->mutable_data(),
                          this->reserve_size_);
    }, {input.block()}, {output.block(), mask_.block()}, "cudnnDropoutForward");
    return output;
  } else {
    return input;
  }
}

const std::pair<Tensor, vector<Tensor>> CudnnDropout::Backward(
    int flag, const Tensor& grad) {
  vector<Tensor> param_grad;
  Tensor dx;
  if (flag & kTrain) {
    dx.ResetLike(grad);
    dx.device()->Exec([dx, grad, this](Context* ctx) {
      Block* dyblock = grad.block(), * dxblock = dx.block(),
             * mblock = this->mask_.block();
      cudnnDropoutBackward(ctx->cudnn_handle, this->drop_desc_, this->y_desc_,
                           dyblock->data(), this->x_desc_,
                           dxblock->mutable_data(), mblock->mutable_data(),
                           this->reserve_size_);
    }, {grad.block(), mask_.block()}, {dx.block()}, "cudnnDropoutBackward");
  } else {
    LOG(ERROR) << "Do not call backward for evaluation phase";
  }
  return std::make_pair(dx, param_grad);
}
void CudnnDropout::ToDevice(std::shared_ptr<Device> device) {
  Dropout::ToDevice(device);
  state_.ToDevice(device);
}
}  // namespace singa
#endif  // CUDNN_MAJOR>=5
#endif  // USE_CUDNN