/* Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserve.

Licensed 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 "paddle/fluid/framework/tensor.h"
#include "paddle/fluid/operators/lrn_op.h"
#include "paddle/fluid/platform/mkldnn_helper.h"

namespace paddle {
namespace operators {

using paddle::framework::Tensor;
using paddle::platform::MKLDNNDeviceContext;

namespace {
mkldnn::algorithm LRNAlgorithm(const paddle::framework::ExecutionContext& ctx) {
  mkldnn::algorithm algorithm = mkldnn::lrn_across_channels;

  std::string algorithm_str = ctx.Attr<std::string>("algorithm");
  if (algorithm_str == "WITHIN_CHANNEL") {
    algorithm = mkldnn::lrn_within_channel;
  }
  return algorithm;
}
}  // namespace

template <typename T>
class LRNMKLDNNOpKernel : public paddle::framework::OpKernel<T> {
 public:
  void Compute(const paddle::framework::ExecutionContext& ctx) const override {
    PADDLE_ENFORCE(std::is_same<T, float>::value,
                   "MKLDNN LRN must use float data.");
    PADDLE_ENFORCE(paddle::platform::is_cpu_place(ctx.GetPlace()),
                   "MKLDNN LRN must use CPUPlace.");

    auto& dev_ctx = ctx.template device_context<MKLDNNDeviceContext>();
    const auto& mkldnn_engine = dev_ctx.GetEngine();

    auto x = ctx.Input<Tensor>("X");
    auto out = ctx.Output<Tensor>("Out");
    auto mid = ctx.Output<Tensor>("MidOut");

    auto input_data = x->data<T>();
    auto output_data = out->mutable_data<T>(ctx.GetPlace());
    mid->mutable_data<T>(ctx.GetPlace());

    const std::string key = ctx.op().Output("Out");
    const std::string key_src_memory = key + "@lrn_src_memory";
    const std::string key_pd = key + "@lrn_pd";
    const std::string key_workspace_memory = key + "@lrn_workspace_memory";

    const int n = ctx.Attr<int>("n");
    const float alpha = ctx.Attr<float>("alpha");
    const float beta = ctx.Attr<float>("beta");
    const float k = ctx.Attr<float>("k");

    auto algorithm = LRNAlgorithm(ctx);

    auto e_mid = framework::EigenTensor<T, 4>::From(*mid);
    e_mid = e_mid.constant(k);

    auto dims = paddle::framework::vectorize2int(x->dims());

    auto src_md = paddle::platform::MKLDNNMemDesc(
        dims, mkldnn::memory::data_type::f32, mkldnn::memory::format::nchw);

    auto dst_md = paddle::platform::MKLDNNMemDesc(
        dims, mkldnn::memory::data_type::f32, mkldnn::memory::format::nchw);

    auto forward_desc = mkldnn::lrn_forward::desc{
        mkldnn::prop_kind::forward, algorithm, src_md, n, alpha, beta, k};

    auto forward_pd = std::make_shared<mkldnn::lrn_forward::primitive_desc>(
        forward_desc, mkldnn_engine);

    dev_ctx.SetBlob(key_pd, forward_pd);

    auto src_memory_pd = mkldnn::memory::primitive_desc{src_md, mkldnn_engine};
    auto src_memory = std::make_shared<mkldnn::memory>(
        src_memory_pd, static_cast<void*>(const_cast<float*>(input_data)));

    dev_ctx.SetBlob(key_src_memory, src_memory);
    auto dst_memory = mkldnn::memory{{dst_md, mkldnn_engine},
                                     static_cast<void*>(output_data)};

    auto workspace_md = forward_pd->workspace_primitive_desc();
    auto workspace_memory = std::make_shared<mkldnn::memory>(workspace_md);

    dev_ctx.SetBlob(key_workspace_memory, workspace_memory);

    auto forward_op = mkldnn::lrn_forward{*forward_pd, *src_memory,
                                          *workspace_memory, dst_memory};

    std::vector<mkldnn::primitive> pipeline = {forward_op};
    mkldnn::stream(mkldnn::stream::kind::eager).submit(pipeline).wait();
  }
};

template <typename T>
class LRNMKLDNNGradOpKernel : public paddle::framework::OpKernel<T> {
 public:
  void Compute(const paddle::framework::ExecutionContext& ctx) const override {
    PADDLE_ENFORCE(std::is_same<T, float>::value,
                   "MKLDNN LRN must use float data.");
    PADDLE_ENFORCE(paddle::platform::is_cpu_place(ctx.GetPlace()),
                   "MKLDNN LRN must use CPUPlace.");

    auto x = ctx.Input<Tensor>("X");

    auto out_grad = ctx.Input<Tensor>(framework::GradVarName("Out"));
    auto x_grad = ctx.Output<Tensor>(framework::GradVarName("X"));

    const std::string key = ctx.op().Input("Out");
    const std::string key_src_memory = key + "@lrn_src_memory";
    const std::string key_pd = key + "@lrn_pd";
    const std::string key_workspace_memory = key + "@lrn_workspace_memory";

    const int n = ctx.Attr<int>("n");
    const float alpha = ctx.Attr<float>("alpha");
    const float beta = ctx.Attr<float>("beta");
    const float k = ctx.Attr<float>("k");

    auto& dev_ctx = ctx.template device_context<MKLDNNDeviceContext>();
    const auto& mkldnn_engine = dev_ctx.GetEngine();

    auto x_grad_data = x_grad->mutable_data<T>(ctx.GetPlace());
    auto out_grad_data = out_grad->data<T>();

    auto dims = paddle::framework::vectorize2int(x->dims());

    auto src_md = paddle::platform::MKLDNNMemDesc(
        dims, mkldnn::memory::data_type::f32, mkldnn::memory::format::nchw);

    auto diff_src_md = paddle::platform::MKLDNNMemDesc(
        dims, mkldnn::memory::data_type::f32, mkldnn::memory::format::nchw);

    auto diff_dst_md = paddle::platform::MKLDNNMemDesc(
        dims, mkldnn::memory::data_type::f32, mkldnn::memory::format::nchw);

    auto diff_dst_memory =
        mkldnn::memory{{diff_dst_md, mkldnn_engine},
                       static_cast<void*>(const_cast<float*>(out_grad_data))};

    auto diff_src_memory = mkldnn::memory{{diff_src_md, mkldnn_engine},
                                          static_cast<void*>(x_grad_data)};

    auto algorithm = LRNAlgorithm(ctx);

    auto backward_desc = mkldnn::lrn_backward::desc{
        algorithm, src_md, diff_src_md, n, alpha, beta, k};

    auto forward_pd = dev_ctx.GetBlob(key_pd);

    auto backward_pd = mkldnn::lrn_backward::primitive_desc{
        backward_desc, mkldnn_engine,
        *static_cast<mkldnn::lrn_forward::primitive_desc*>(forward_pd.get())};

    std::shared_ptr<void> workspace_memory =
        dev_ctx.GetBlob(key_workspace_memory);

    auto src_memory = dev_ctx.GetBlob(key_src_memory);
    auto backward_op = mkldnn::lrn_backward{
        backward_pd, *static_cast<mkldnn::memory*>(src_memory.get()),
        diff_dst_memory, *static_cast<mkldnn::memory*>(workspace_memory.get()),
        diff_src_memory};

    std::vector<mkldnn::primitive> pipeline = {backward_op};
    mkldnn::stream(mkldnn::stream::kind::eager).submit(pipeline).wait();
  }
};
}  // namespace operators
}  // namespace paddle

namespace ops = paddle::operators;

REGISTER_OP_KERNEL(lrn, MKLDNN, paddle::platform::CPUPlace,
                   ops::LRNMKLDNNOpKernel<float>);
REGISTER_OP_KERNEL(lrn_grad, MKLDNN, paddle::platform::CPUPlace,
                   ops::LRNMKLDNNGradOpKernel<float>);