src/model/layer/cudnn_softmax.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 "singa/singa_config.h"
#include "./cudnn_softmax.h"
#ifdef USE_CUDNN
#include <cudnn.h>
#include "./cudnn_utils.h"
#include "singa/utils/logging.h"
namespace singa {

RegisterLayerClass(cudnn_softmax, CudnnSoftmax);
CudnnSoftmax::~CudnnSoftmax() {
  if (desc_ != nullptr) CUDNN_CHECK(cudnnDestroyTensorDescriptor(desc_));
}

void CudnnSoftmax::Setup(const Shape& in_sample, const LayerConf &conf) {
  Softmax::Setup(in_sample, conf);
  SoftmaxConf sft_conf = conf.softmax_conf();
  std::string algorithm = sft_conf.algorithm();
  CHECK(algorithm == "accurate" || algorithm == "fast" || algorithm == "log")
    << "CudnnSoftmax only supports three algorithm preferences: "
    << "accurate, fast and log.";
  if (algorithm == "accurate")
    algorithm_ = CUDNN_SOFTMAX_ACCURATE;
  else if (algorithm == "fast")
    algorithm_ = CUDNN_SOFTMAX_FAST;
  else algorithm_ = CUDNN_SOFTMAX_LOG;
}

void CudnnSoftmax::InitCudnn(Shape shape, DataType dtype) {
  if (!has_init_cudnn_)
    CUDNN_CHECK(cudnnCreateTensorDescriptor(&desc_));

  if (shape.size() == 1u)
    CUDNN_CHECK(cudnnSetTensor4dDescriptor( desc_,
      CUDNN_TENSOR_NCHW, GetCudnnDataType(dtype), 1, shape[0], 1, 1));
  else
    CUDNN_CHECK(cudnnSetTensor4dDescriptor( desc_, CUDNN_TENSOR_NCHW,
      GetCudnnDataType(dtype), shape[0], shape[1], 1, 1));
  has_init_cudnn_ = true;
}

const Tensor CudnnSoftmax::Forward(int flag, const Tensor& input) {
  CHECK(buf_.empty());
  auto shape = input.shape();
  CHECK_LE(shape.size(), 2u)
    << "Tensor shape should range from 1 to 2D;"
    << "otherwise, add flatten layer to transform";
  DataType dtype = input.data_type();
  if (!has_init_cudnn_) {
    InitCudnn(shape, dtype);
  } else {
    int n, c, h, w, s;
    cudnnDataType_t type;
    CUDNN_CHECK(cudnnGetTensor4dDescriptor(desc_, &type, &n, &c, &h, &w,
          &s, &s, &s, &s));
    if ((shape.size() == 1u && shape[0] != static_cast<size_t>(c)) ||
        (shape.size() == 2u &&
         (shape[0] != static_cast<size_t>(n)
          || shape[1] != static_cast<size_t>(c))))
      InitCudnn(shape, dtype);
  }

  Tensor output;
  output.ResetLike(input);
  output.device()->Exec([input, output, this](Context* ctx) {
    Block* inblock = input.block(), * outblock = output.block();
    float alpha = 1.0f, beta = 0.0f;
    cudnnSoftmaxForward(ctx->cudnn_handle, this->algorithm_,
                        CUDNN_SOFTMAX_MODE_INSTANCE,
                        &alpha, this->desc_, inblock->data(), &beta,
                        this->desc_, outblock->mutable_data());
  }, {input.block()}, {output.block()});
  if (flag & kTrain) buf_.push(output);
  return output;
}

const std::pair<Tensor, vector<Tensor>> CudnnSoftmax::Backward(
    int flag, const Tensor& grad) {
  vector<Tensor> param_grad;
  CHECK(!buf_.empty());
  Tensor dx, output = buf_.top();
  buf_.pop();
  dx.ResetLike(grad);
  dx.device()->Exec([dx, grad, output, this](Context* ctx) {
    Block* dyblock = grad.block(), * dxblock = dx.block(),
           * yblock = output.block();
    float alpha = 1.0f, beta = 0.0f;
    cudnnSoftmaxBackward(ctx->cudnn_handle, this->algorithm_,
                         CUDNN_SOFTMAX_MODE_INSTANCE,
                         &alpha, this->desc_, yblock->data(), this->desc_,
                         dyblock->data(), &beta, this->desc_,
                         dxblock->mutable_data());
  }, {grad.block(), output.block()}, {dx.block()});
  return std::make_pair(dx, param_grad);
}
}  // namespace singa
#endif  // USE_CUDNN