sequence_concat_op.h

/* Copyright (c) 2016 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. */

#pragma once
#include "paddle/framework/op_registry.h"
#include "paddle/operators/strided_memcpy.h"

namespace paddle {
namespace operators {

using Tensor = framework::Tensor;
using LoDTensor = framework::LoDTensor;
using LoD = framework::LoD;

// Concat LoD, the initialized LoD of Output is lod(x0),
// if axis is not 0, the LoD(Out) will be the same as Inputs, if axis is 0:
// Case1:
//  There is one level, the Output LoD will be modified:
//  LoD(x0) = {{0,2,4}}
//  LoD(x1) = {{0,1,5}}
//  LoD(Out) = {{0,3,9}}
// Case2:
//  There is two level, and concat level is 1,
//  the Output LoD will be modified as followed:
//  LoD(x0) = {{0,2,4}, {0,1,2,3,4}}
//  LoD(x1) = {{0,3,5}, {0,1,3,4,5}}
//  LoD(Out) = {{0,5,9}, {0,1,2,4,5,6,7,8,9}}
template <typename T>
LoD concatLoD(const std::vector<const T*> ins, const size_t axis,
              const size_t level) {
  auto out_lod = ins[0]->lod();
  const size_t n = ins.size();
  if (axis == 0UL) {
    if (level == 0) {
      for (size_t i = 1; i < n; ++i) {
        for (size_t j = 0; j < ins[i]->lod()[0].size(); ++j) {
          out_lod[0][j] += ins[i]->lod()[0][j];
        }
      }
    } else if (level == 1) {
      PADDLE_ENFORCE_EQ(ins[0]->NumLevels(), 2UL,
                        "If the level is 1, all of the inputs "
                        "should be the the nested sequence.");
      for (size_t i = 1; i < n; ++i) {
        for (size_t j = 0; j < ins[i]->lod()[0].size(); ++j) {
          out_lod[0].push_back(ins[i]->lod()[0][j]);
        }
        for (size_t j = 0; j < ins[i]->lod()[1].size(); ++j) {
          out_lod[1][j] += ins[i]->lod()[1][j];
        }
      }
    }
  }
  return out_lod;
}

template <typename Place, typename T>
class SequenceConcatOpKernel : public framework::OpKernel<T> {
 public:
  void Compute(const framework::ExecutionContext& ctx) const override {
    auto ins = ctx.MultiInput<LoDTensor>("X");
    auto* out = ctx.Output<LoDTensor>("Out");
    const size_t axis = static_cast<size_t>(ctx.Attr<int>("axis"));
    const size_t level = static_cast<size_t>(ctx.Attr<int>("level"));
    const size_t n = ins.size();

    for (size_t i = 1; i < n; ++i) {
      PADDLE_ENFORCE_EQ(ins[0]->NumLevels(), ins[i]->NumLevels(),
                        "The level number of all the input LoDTensors "
                        "should be the same.");
      PADDLE_ENFORCE_EQ(ins[0]->dims().size(), ins[i]->dims().size(),
                        "The dimensions size of all the input LoDTensors "
                        "should be the same.");

      const size_t dims_size = ins[i]->dims().size();
      for (size_t j = 0; j < dims_size; ++j) {
        if (j == axis) continue;
        PADDLE_ENFORCE_EQ(ins[0]->dims()[j], ins[i]->dims()[j],
                          "The dimensions of all the input LoDTensors "
                          "except for the specify axis should be "
                          "matched exactly.");
      }
    }

    out->mutable_data<T>(ctx.GetPlace());
    auto out_lod = concatLoD<LoDTensor>(ins, axis, level);
    out->set_lod(out_lod);

    auto out_lod_level = out_lod[level];
    for (size_t i = 0; i < out_lod_level.size() - 1; ++i) {
      Tensor out_t = out->Slice<T>(static_cast<int>(out_lod_level[i]),
                                   static_cast<int>(out_lod_level[i + 1]));
      auto out_stride = framework::stride(out_t.dims());
      size_t offset = 0;

      for (size_t j = 0; j < n; ++j) {
        auto in_lod_level = ins[j]->lod()[level];
        auto in_stride = framework::stride(ins[j]->dims());
        Tensor in_t = ins[j]->Slice<T>(static_cast<int>(in_lod_level[i]),
                                       static_cast<int>(in_lod_level[i + 1]));
        size_t axis_dim = in_t.dims()[axis];
        StridedMemcpy<T>(ctx.device_context(), in_t.data<T>(), in_stride,
                         in_t.dims(), out_stride, out_t.data<T>() + offset);
        offset += axis_dim * in_stride[axis];
      }
    }
  }
};

template <typename Place, typename T>
class SequenceConcatGradOpKernel : public framework::OpKernel<T> {
 public:
  void Compute(const framework::ExecutionContext& ctx) const override {
    auto ins = ctx.MultiInput<framework::LoDTensor>("X");
    auto* out_grad =
        ctx.Input<framework::LoDTensor>(framework::GradVarName("Out"));
    auto x_grads =
        ctx.MultiOutput<framework::LoDTensor>(framework::GradVarName("X"));
    size_t axis = static_cast<size_t>(ctx.Attr<int>("axis"));
    size_t level = static_cast<size_t>(ctx.Attr<int>("level"));
    const size_t n = x_grads.size();

    // Set Grad(X) LoD as X
    for (size_t i = 0; i < n; i++) {
      x_grads[i]->set_lod(ins[i]->lod());
      x_grads[i]->mutable_data<T>(ctx.GetPlace());
    }

    auto out_lod = concatLoD<LoDTensor>(ins, axis, level);
    auto out_lod_level = out_lod[level];

    for (size_t i = 0; i < out_lod_level.size() - 1; ++i) {
      Tensor out_grad_t =
          out_grad->Slice<T>(static_cast<int>(out_lod_level[i]),
                             static_cast<int>(out_lod_level[i + 1]));
      auto out_grad_stride = framework::stride(out_grad_t.dims());
      size_t offset = 0;

      for (size_t j = 0; j < n; ++j) {
        auto x_grad_lod_level = x_grads[j]->lod()[level];
        auto x_grad_stride = framework::stride(x_grads[j]->dims());
        Tensor x_grad_t =
            x_grads[j]->Slice<T>(static_cast<int>(x_grad_lod_level[i]),
                                 static_cast<int>(x_grad_lod_level[i + 1]));
        size_t axis_dim = x_grad_t.dims()[axis];
        StridedMemcpy<T>(ctx.device_context(), out_grad_t.data<T>() + offset,
                         out_grad_stride, out_grad_t.dims(), x_grad_stride,
                         x_grad_t.data<T>());
        offset += axis_dim * out_grad_stride[axis];
      }
    }
  }
};

}  // namespace operators
}  // namespace paddle