gradient_accumulator.cc

// Copyright (c) 2019 PaddlePaddle Authors. All Rights Reserved.
//
// 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/imperative/gradient_accumulator.h"

#include <algorithm>
#include <memory>
#include <utility>

#include "paddle/fluid/framework/convert_utils.h"
#include "paddle/fluid/framework/lod_tensor.h"
#include "paddle/fluid/framework/selected_rows_utils.h"
#include "paddle/fluid/imperative/layer.h"
#include "paddle/fluid/operators/math/selected_rows_functor.h"
#include "paddle/fluid/platform/bfloat16.h"
#include "paddle/fluid/platform/complex.h"
#include "paddle/fluid/platform/device_context.h"
#include "paddle/fluid/platform/float16.h"
#include "paddle/fluid/platform/profiler.h"
#include "paddle/phi/kernels/funcs/blas/blas.h"
#include "paddle/phi/kernels/funcs/math_function.h"
#ifdef PADDLE_WITH_XPU
#include "xpu/refactor/math.h"
#endif
#ifdef PADDLE_WITH_ASCEND_CL
#include "paddle/fluid/platform/device/npu/npu_op_runner.h"
#endif
#ifdef PADDLE_WITH_MLU
#include "paddle/fluid/operators/mlu/mlu_baseop.h"
#endif
#ifdef PADDLE_WITH_CUSTOM_DEVICE
#include "paddle/phi/backends/device_manager.h"
#endif

namespace paddle {
namespace imperative {

static void MoveOrCopyVar(framework::Variable* dst,
                          framework::Variable* src,
                          bool force_copy) {
  if (!force_copy) {
    VLOG(6) << "Just Move Variable when sum gradients within this graph";
    *dst = std::move(*src);
    return;
  }

  VLOG(6) << "Copy occurs when sum gradients within this graph";
  if (src->IsType<framework::LoDTensor>()) {
    auto& src_tensor = src->Get<framework::LoDTensor>();
    if (!dst->IsType<framework::LoDTensor>()) {
      dst->Clear();
    }
    auto* dst_tensor = dst->GetMutable<framework::LoDTensor>();
    framework::TensorCopy(src_tensor, src_tensor.place(), dst_tensor);
    dst_tensor->set_lod(src_tensor.lod());
  } else if (src->IsType<phi::SelectedRows>()) {
    auto& src_selected_rows = src->Get<phi::SelectedRows>();
    if (!dst->IsType<phi::SelectedRows>()) {
      dst->Clear();
    }
    auto* dst_selected_rows = dst->GetMutable<phi::SelectedRows>();
    framework::TensorCopy(src_selected_rows.value(),
                          src_selected_rows.value().place(),
                          dst_selected_rows->mutable_value());
    dst_selected_rows->set_rows(src_selected_rows.rows());
    dst_selected_rows->set_height(src_selected_rows.height());
  } else {
    PADDLE_THROW(platform::errors::PermissionDenied(
        "Only support LoDTensor and SelectedRows for sum gradient"));
  }
}

template <typename T>
class TensorAddFunctor
    : public std::unary_function<const platform::Place&, void> {
 public:
  TensorAddFunctor(int64_t numel, const T* x, T* y)
      : numel_(numel), x_(x), y_(y) {}

  void operator()(const platform::CPUPlace& place) const {
    phi::CPUContext* ctx = dynamic_cast<phi::CPUContext*>(
        platform::DeviceContextPool::Instance().Get(place));
    auto blas = phi::funcs::GetBlas<phi::CPUContext, T>(*ctx);
    blas.AXPY(numel_, 1., x_, y_);
  }

#ifdef PADDLE_WITH_XPU
  void operator()(const platform::XPUPlace& place) const {
    using XPUType = typename XPUTypeTrait<T>::Type;
    platform::XPUDeviceContext* ctx = dynamic_cast<platform::XPUDeviceContext*>(
        platform::DeviceContextPool::Instance().Get(place));
    int r = xpu::add<XPUType>(ctx->x_context(),
                              reinterpret_cast<const XPUType*>(x_),
                              reinterpret_cast<const XPUType*>(y_),
                              reinterpret_cast<XPUType*>(y_),
                              static_cast<int>(numel_));
    PADDLE_ENFORCE_EQ(
        r,
        XPU_SUCCESS,
        platform::errors::External(
            "XPU add kernel return wrong value[%d %s]", r, XPUAPIErrorMsg[r]));
  }
#else
  void operator()(const platform::XPUPlace& place) const {
    PADDLE_THROW(platform::errors::PermissionDenied(
        "Gradient accumulation on place (%s) "
        "is not supported in imperative mode",
        place));
  }
#endif

#if defined(PADDLE_WITH_CUDA) || defined(PADDLE_WITH_HIP)
  void operator()(const platform::CUDAPlace& place) const {
    platform::CUDADeviceContext* ctx =
        dynamic_cast<platform::CUDADeviceContext*>(
            platform::DeviceContextPool::Instance().Get(place));
    auto blas = phi::funcs::GetBlas<platform::CUDADeviceContext, T>(*ctx);
    blas.AXPY(numel_, 1., x_, y_);
  }
#else
  void operator()(const platform::CUDAPlace& place) const {
    PADDLE_THROW(platform::errors::PermissionDenied(
        "Gradient accumulation on place (%s) "
        "is not supported in imperative mode",
        place));
  }
#endif

#ifdef PADDLE_WITH_MLU
  void operator()(const platform::MLUPlace& place) const {
    // TODO(fwg): SUPPORT it
    PADDLE_THROW(platform::errors::PermissionDenied(
        "Gradient accumulation on place (%s) "
        "is not supported in imperative mode",
        place));
  }
#else
  void operator()(const platform::MLUPlace& place) const {
    PADDLE_THROW(platform::errors::PermissionDenied(
        "Gradient accumulation on place (%s) "
        "is not supported in imperative mode",
        place));
  }
#endif

#ifdef PADDLE_WITH_ASCEND_CL
  void operator()(const platform::NPUPlace& place) const {
    // TODO(zhiqiu): SUPPORT it
    PADDLE_THROW(platform::errors::PermissionDenied(
        "Gradient accumulation on place (%s) "
        "is not supported in imperative mode",
        place));
  }
#else
  void operator()(const platform::NPUPlace& place) const {
    PADDLE_THROW(platform::errors::PermissionDenied(
        "Gradient accumulation on place (%s) "
        "is not supported in imperative mode",
        place));
  }
#endif

  void operator()(const platform::NPUPinnedPlace& place) const {
    PADDLE_THROW(platform::errors::PermissionDenied(
        "Gradient accumulation on place (%s) "
        "is not supported in imperative mode",
        place));
  }
  // there is NO blas in CUDAPinnedPlace
  void operator()(const platform::CUDAPinnedPlace& place) const {
    PADDLE_THROW(platform::errors::PermissionDenied(
        "Gradient accumulation on place (%s) "
        "is not supported in imperative mode",
        place));
  }
  // there is NO support in IPUPlace
  void operator()(const platform::IPUPlace& place) const {
    PADDLE_THROW(platform::errors::PermissionDenied(
        "Gradient accumulation on place (%s) "
        "is not supported in imperative mode",
        place));
  }
  void operator()(const platform::CustomPlace& place) const {
#ifdef PADDLE_WITH_CUSTOM_DEVICE
    platform::CustomDeviceContext* ctx =
        dynamic_cast<platform::CustomDeviceContext*>(
            platform::DeviceContextPool::Instance().Get(place));
    phi::stream::Stream stream(place, ctx->stream());
    auto device = phi::DeviceManager::GetDeviceWithPlace(place);
    device->BlasAXPBY<T>(stream, static_cast<size_t>(numel_), 1., x_, 1., y_);
#else
    PADDLE_THROW(platform::errors::PermissionDenied(
        "Gradient accumulation on place (%s) "
        "is not supported in imperative mode",
        place));
#endif
  }

 private:
  int64_t numel_;
  const T* x_;
  mutable T* y_;
};

#ifdef PADDLE_WITH_XPU
template <typename T>
void XPUTensorAddFunctor(const platform::Place& place,
                         const framework::Tensor& src,
                         framework::Tensor* dst) {
  using XPUType = typename XPUTypeTrait<T>::Type;
  platform::XPUDeviceContext* ctx = dynamic_cast<platform::XPUDeviceContext*>(
      platform::DeviceContextPool::Instance().Get(place));
  const XPUType* x = reinterpret_cast<const XPUType*>(src.data<T>());
  XPUType* y = reinterpret_cast<XPUType*>(dst->mutable_data<T>(place));
  int r = xpu::add<XPUType>(
      ctx->x_context(), x, y, y, static_cast<int>(src.numel()));
  PADDLE_ENFORCE_EQ(
      r,
      XPU_SUCCESS,
      platform::errors::External(
          "XPU add kernel return wrong value[%d %s]", r, XPUAPIErrorMsg[r]));
}
#endif

template <typename DeviceContext, typename T>
void TensorAddImpl(const framework::Tensor& src,
                   framework::Tensor* dst,
                   const platform::Place& place) {
  platform::DeviceContextPool& pool = platform::DeviceContextPool::Instance();
  paddle::platform::DeviceContext* ctx = pool.Get(place);
  auto dev_ctx = dynamic_cast<DeviceContext*>(ctx);
  phi::funcs::ElementwiseAddTo<DeviceContext, T> func;
  func(dev_ctx, src, dst);
}

template <typename TType>
TType* GetInnerMutableTensor(framework::Variable* dst) {
  auto* dst_tensor = dst->GetMutable<TType>();
  return dst_tensor;
}

template <typename TType>
TType* GetInnerMutableTensor(paddle::experimental::Tensor* dst) {
  auto* dst_tensor = static_cast<TType*>(dst->impl().get());
  return dst_tensor;
}

template <typename TType>
const TType& GetInnerTensor(const framework::Variable& src) {
  return src.Get<TType>();
}

template <typename TType>
TType& GetInnerTensor(const paddle::experimental::Tensor& src) {
  PADDLE_ENFORCE_EQ(
      src.initialized(),
      true,
      platform::errors::Fatal("We only add tensor with value if a tensor is "
                              "NOT INITILIZED, it should just move instead of "
                              "calling this method."));
  auto* src_tensor = static_cast<TType*>(src.impl().get());
  return *src_tensor;
}

template <typename TType>
TType* GetEmptyInnerTensor(paddle::experimental::Tensor* dst) {
  PADDLE_ENFORCE_EQ(
      dst->defined(),
      false,
      platform::errors::Fatal(
          "The underlying Tensor implementation should be nullptr"));
  dst->set_impl(std::make_shared<TType>());
  auto* dst_tensor = static_cast<TType*>(dst->impl().get());
  return dst_tensor;
}

template <typename TType>
TType* GetEmptyInnerTensor(paddle::imperative::VariableWrapper* dst) {
  auto* dst_tensor = dst->MutableVar()->GetMutable<TType>();
  return dst_tensor;
}

template <typename VarType>
void TensorAdd(const VarType& src, VarType* dst) {
  phi::DenseTensor* dst_tensor = GetInnerMutableTensor<phi::DenseTensor>(dst);
  const phi::DenseTensor& src_tensor = GetInnerTensor<phi::DenseTensor>(src);

  auto numel = src_tensor.numel();

  // FIXME(minqiyang): loss_grad op will pass a zero grad of label
  // ugly fix for it
  if (numel == 0) {
    return;
  }

  PADDLE_ENFORCE_EQ(
      dst_tensor->numel(),
      numel,
      platform::errors::PreconditionNotMet(
          "The number of elements of source tensor and destination tensor "
          "should be equal, but got the number of elements of source tensor is "
          "%zu and the number of elements of destination tensor is %zu.",
          numel,
          dst_tensor->numel()));

  auto data_type = framework::TransToProtoVarType(src_tensor.dtype());
  auto place = src_tensor.place();

  PADDLE_ENFORCE_EQ(framework::TransToProtoVarType(dst_tensor->dtype()),
                    data_type,
                    platform::errors::PreconditionNotMet(
                        "The data type of source tensor and destination tensor "
                        "should be equal, Otherwise, the calculation results "
                        "will be incorrect."));

  // if src and dst are in different place, copy dst to src's place
  if (dst_tensor->place() != place) {
    paddle::framework::TensorCopySync(*dst_tensor, place, dst_tensor);
  }
#define PADDLE_TENSOR_ADD(cpp_type)                                  \
  if (data_type == framework::DataTypeTrait<cpp_type>::DataType()) { \
    TensorAddFunctor<cpp_type> func(                                 \
        numel,                                                       \
        src_tensor.data<cpp_type>(),                                 \
        dst_tensor->mutable_data<cpp_type>(place));                  \
    platform::VisitPlace(place, func);                               \
    return;                                                          \
  }

#ifdef PADDLE_WITH_ASCEND_CL
  if (platform::is_npu_place(place)) {
    platform::DeviceContextPool& pool = platform::DeviceContextPool::Instance();
    platform::DeviceContext* ctx = pool.Get(place);
    auto dev_ctx = dynamic_cast<platform::NPUDeviceContext*>(ctx);
    if (data_type == framework::DataTypeTrait<float>::DataType()) {
      dst_tensor->mutable_data<float>(place);
    } else if (data_type == framework::DataTypeTrait<double>::DataType()) {
      dst_tensor->mutable_data<double>(place);
    } else if (data_type ==
               framework::DataTypeTrait<platform::float16>::DataType()) {
      dst_tensor->mutable_data<platform::float16>(place);
    } else {
      PADDLE_THROW(platform::errors::Unimplemented(
          "Gradient accumulation of data type (%s) on place (%s) is not "
          "supported in imperative mode",
          framework::DataTypeToString(data_type),
          place));
    }
    const auto& runner = operators::NpuOpRunner(
        "Add", {*dst_tensor, src_tensor}, {*dst_tensor}, {});
    runner.Run(dev_ctx->stream());
    return;
  }
#endif

#ifdef PADDLE_WITH_XPU
  if (platform::is_xpu_place(place)) {
    if (data_type == framework::DataTypeTrait<float>::DataType()) {
      XPUTensorAddFunctor<float>(place, src_tensor, dst_tensor);
    } else if (data_type ==
               framework::DataTypeTrait<platform::float16>::DataType()) {
      XPUTensorAddFunctor<platform::float16>(place, src_tensor, dst_tensor);
    } else {
      PADDLE_THROW(platform::errors::Unimplemented(
          "Gradient accumulation of data type (%s) on place (%s) is not "
          "supported in imperative mode",
          framework::DataTypeToString(data_type),
          place));
    }
    return;
  }
#endif

#ifdef PADDLE_WITH_MLU
  if (platform::is_mlu_place(place)) {
    platform::DeviceContextPool& pool = platform::DeviceContextPool::Instance();
    platform::DeviceContext* ctx = pool.Get(place);
    auto dev_ctx = dynamic_cast<platform::MLUDeviceContext*>(ctx);
    if (data_type == framework::DataTypeTrait<float>::DataType()) {
      dst_tensor->mutable_data<float>(place);
    } else if (data_type ==
               framework::DataTypeTrait<platform::float16>::DataType()) {
      dst_tensor->mutable_data<platform::float16>(place);
    } else {
      PADDLE_THROW(platform::errors::Unimplemented(
          "Gradient accumulation of data type (%s) on place (%s) is not "
          "supported in imperative mode",
          framework::DataTypeToString(data_type),
          place));
    }
    static const float alpha = 1.f;
    static const float beta = 1.f;
    operators::MLUCnnlTensorDesc src_tensor_desc(src_tensor);
    operators::MLUCnnlTensorDesc dst_tensor_desc(*dst_tensor);
    PADDLE_ENFORCE_MLU_SUCCESS(
        cnnlAssignAdd(dev_ctx->cnnl_handle(),
                      static_cast<const void*>(&alpha),
                      src_tensor_desc.get(),
                      operators::GetBasePtr(&src_tensor),
                      nullptr,
                      0,
                      static_cast<const void*>(&beta),
                      dst_tensor_desc.get(),
                      operators::GetBasePtr(dst_tensor)));
    return;
  }
#endif

  PADDLE_TENSOR_ADD(float);

#ifndef PADDLE_WITH_XPU
  // NOTE(phlrain): xpu only support float
  PADDLE_TENSOR_ADD(double);
  // NOTE(chenweihang): only support complex grad tensor accumulated,
  // support selected rows if needed in the future
  PADDLE_TENSOR_ADD(platform::complex<float>);
  PADDLE_TENSOR_ADD(platform::complex<double>);
#endif

#undef PADDLE_TENSOR_ADD

  if (data_type == framework::proto::VarType::FP16) {
    if (platform::is_gpu_place(place)) {
#if defined(PADDLE_WITH_CUDA) || defined(PADDLE_WITH_HIP)
      return TensorAddImpl<platform::CUDADeviceContext, platform::float16>(
          src_tensor, dst_tensor, place);
#else
      PADDLE_THROW(platform::errors::Unimplemented(
          "Gradient accumulation of data type (%s) on place (%s) is not "
          "supported in imperative mode",
          framework::DataTypeToString(data_type),
          place));
#endif
    } else if (platform::is_cpu_place(place)) {
      return TensorAddImpl<phi::CPUContext, platform::float16>(
          src_tensor, dst_tensor, place);
    }
  }
  if (data_type == framework::proto::VarType::BF16) {
    if (platform::is_gpu_place(place)) {
#if defined(PADDLE_WITH_CUDA) || defined(PADDLE_WITH_HIP)
      return TensorAddImpl<platform::CUDADeviceContext, platform::bfloat16>(
          src_tensor, dst_tensor, place);
#else
      PADDLE_THROW(platform::errors::Unimplemented(
          "Gradient accumulation of data type (%s) on place (%s) is not "
          "supported in imperative mode",
          framework::DataTypeToString(data_type),
          place));
#endif
    } else if (platform::is_cpu_place(place)) {
      return TensorAddImpl<phi::CPUContext, platform::bfloat16>(
          src_tensor, dst_tensor, place);
    }
  }
  PADDLE_THROW(platform::errors::Unimplemented(
      "Gradient accumulation of data type (%s) on place (%s) is not "
      "supported in imperative mode",
      framework::DataTypeToString(data_type),
      place));
}

template void TensorAdd<framework::Variable>(const framework::Variable& src,
                                             framework::Variable* dst);
template void TensorAdd<paddle::experimental::Tensor>(
    const paddle::experimental::Tensor& src, paddle::experimental::Tensor* dst);

template <typename VarType>
void SelectedRowsAddToTensor(const VarType& src, VarType* dst) {
  phi::DenseTensor* dst_tensor = GetInnerMutableTensor<phi::DenseTensor>(dst);
  const phi::SelectedRows& src_selected_rows =
      GetInnerTensor<phi::SelectedRows>(src);
  auto place = dst_tensor->place();
  auto data_type =
      framework::TransToProtoVarType(src_selected_rows.value().dtype());
  platform::DeviceContextPool& pool = platform::DeviceContextPool::Instance();

#define PADDLE_SELECTED_ROWS_ADD_TO_TENSOR(dev_ctx_type, cpp_type)           \
  if (data_type == framework::DataTypeTrait<cpp_type>::DataType()) {         \
    paddle::platform::DeviceContext* dev_ctx = pool.Get(place);              \
    paddle::operators::math::SelectedRowsAddToTensor<dev_ctx_type, cpp_type> \
        functor;                                                             \
    functor(*(dynamic_cast<dev_ctx_type*>(dev_ctx)),                         \
            src_selected_rows,                                               \
            dst_tensor);                                                     \
    return;                                                                  \
  }

#if defined(PADDLE_WITH_CUDA) || defined(PADDLE_WITH_HIP)
  if (paddle::platform::is_gpu_place(place)) {
    PADDLE_SELECTED_ROWS_ADD_TO_TENSOR(platform::CUDADeviceContext, float);
    PADDLE_SELECTED_ROWS_ADD_TO_TENSOR(platform::CUDADeviceContext, double);
  } else {
#endif
    PADDLE_SELECTED_ROWS_ADD_TO_TENSOR(phi::CPUContext, float);
    PADDLE_SELECTED_ROWS_ADD_TO_TENSOR(phi::CPUContext, double);
#if defined(PADDLE_WITH_CUDA) || defined(PADDLE_WITH_HIP)
  }
#endif

#undef PADDLE_SELECTED_ROWS_ADD_TO_TENSOR

  PADDLE_THROW(platform::errors::InvalidArgument(
      "Not supported data type %s for SelectedRowsAddToTensor",
      framework::DataTypeToString(data_type)));
}

template void SelectedRowsAddToTensor(const framework::Variable& src,
                                      framework::Variable* dst);
template void SelectedRowsAddToTensor(const paddle::experimental::Tensor& src,
                                      paddle::experimental::Tensor* dst);

template <typename VarType>
void SelectedRowsAddTensor(const VarType& src_selected_rows_var,
                           const VarType& src_tensor_var,
                           VarType* dst_tensor_var) {
  const phi::SelectedRows& src_selected_rows =
      GetInnerTensor<phi::SelectedRows>(src_selected_rows_var);
  const phi::DenseTensor& src_tensor =
      GetInnerTensor<phi::DenseTensor>(src_tensor_var);
  const auto& place = src_tensor.place();
  auto data_type = framework::TransToProtoVarType(src_tensor.dtype());
  auto* dev_ctx = platform::DeviceContextPool::Instance().Get(place);

  phi::DenseTensor* dst_tensor =
      GetInnerMutableTensor<phi::DenseTensor>(dst_tensor_var);
  dst_tensor->Resize(src_tensor.dims());
  dst_tensor->mutable_data(place, src_tensor.dtype());

#define PADDLE_SELECTED_ROWS_ADD_TENSOR(dev_ctx_type, cpp_type)            \
  if (data_type == framework::DataTypeTrait<cpp_type>::DataType()) {       \
    paddle::operators::math::SelectedRowsAddTensor<dev_ctx_type, cpp_type> \
        functor;                                                           \
    functor(*(dynamic_cast<dev_ctx_type*>(dev_ctx)),                       \
            src_selected_rows,                                             \
            src_tensor,                                                    \
            dst_tensor);                                                   \
    return;                                                                \
  }

#if defined(PADDLE_WITH_CUDA) || defined(PADDLE_WITH_HIP)
  if (platform::is_gpu_place(place)) {
    PADDLE_SELECTED_ROWS_ADD_TENSOR(platform::CUDADeviceContext, float);
    PADDLE_SELECTED_ROWS_ADD_TENSOR(platform::CUDADeviceContext, double);
  } else {
#endif
    PADDLE_SELECTED_ROWS_ADD_TENSOR(phi::CPUContext, float);
    PADDLE_SELECTED_ROWS_ADD_TENSOR(phi::CPUContext, double);
#if defined(PADDLE_WITH_CUDA) || defined(PADDLE_WITH_HIP)
  }
#endif

  PADDLE_THROW(platform::errors::InvalidArgument(
      "Not supported data type %s for SelectedRowsAddToTensor",
      framework::DataTypeToString(data_type)));

#undef PADDLE_SELECTED_ROWS_ADD_TENSOR
}

template void SelectedRowsAddTensor(
    const framework::Variable& src_selected_rows_var,
    const framework::Variable& src_tensor_var,
    framework::Variable* dst_tensor_var);
template void SelectedRowsAddTensor(
    const paddle::experimental::Tensor& src_selected_rows_var,
    const paddle::experimental::Tensor& src_tensor_var,
    paddle::experimental::Tensor* dst_tensor_var);

// Note(chenweihang): when two selected rows need to be added,
//   adding one to another is not equal to merging two selected rows
//   to one then add it to a empty selected rows, the after is correct
template <typename ReturnVarType, typename VarType>
std::shared_ptr<ReturnVarType> SelectedRowsMerge(const VarType& src1,
                                                 const VarType& src2) {
  const phi::SelectedRows& src_selected_rows1 =
      GetInnerTensor<phi::SelectedRows>(src1);
  const phi::SelectedRows& src_selected_rows2 =
      GetInnerTensor<phi::SelectedRows>(src2);

  auto place = src_selected_rows1.value().place();
  auto data_type =
      framework::TransToProtoVarType(src_selected_rows1.value().dtype());
  platform::DeviceContextPool& pool = platform::DeviceContextPool::Instance();

  std::vector<const phi::SelectedRows*> src_selected_rows;
  src_selected_rows.emplace_back(&src_selected_rows1);
  src_selected_rows.emplace_back(&src_selected_rows2);

  auto dst_var = std::make_shared<ReturnVarType>("Temp");
  phi::SelectedRows* dst_selected_rows =
      GetEmptyInnerTensor<phi::SelectedRows>(dst_var.get());

#define PADDLE_SELECTED_ROWS_ADD(dev_ctx_type, cpp_type)               \
  if (data_type == framework::DataTypeTrait<cpp_type>::DataType()) {   \
    paddle::platform::DeviceContext* dev_ctx = pool.Get(place);        \
    paddle::operators::math::scatter::MergeAdd<dev_ctx_type, cpp_type> \
        merge_add;                                                     \
    merge_add(*(dynamic_cast<dev_ctx_type*>(dev_ctx)),                 \
              src_selected_rows,                                       \
              dst_selected_rows);                                      \
    return dst_var;                                                    \
  }

#if defined(PADDLE_WITH_CUDA) || defined(PADDLE_WITH_HIP)
  if (paddle::platform::is_gpu_place(place)) {
    PADDLE_SELECTED_ROWS_ADD(platform::CUDADeviceContext, float);
    PADDLE_SELECTED_ROWS_ADD(platform::CUDADeviceContext, double);
  } else {
#endif
    PADDLE_SELECTED_ROWS_ADD(phi::CPUContext, float);
    PADDLE_SELECTED_ROWS_ADD(phi::CPUContext, double);
#if defined(PADDLE_WITH_CUDA) || defined(PADDLE_WITH_HIP)
  }
#endif

#undef PADDLE_SELECTED_ROWS_ADD
  PADDLE_THROW(platform::errors::InvalidArgument(
      "Not supported data type %s for SelectedRowsMerge",
      framework::DataTypeToString(data_type)));
}

template std::shared_ptr<paddle::experimental::Tensor> SelectedRowsMerge(
    const paddle::experimental::Tensor& src1,
    const paddle::experimental::Tensor& src2);
template std::shared_ptr<paddle::imperative::VariableWrapper> SelectedRowsMerge(
    const framework::Variable& src1, const framework::Variable& src2);

void VariableWrapperAdd(std::shared_ptr<VariableWrapper> var,
                        VariableWrapper* dst_var,
                        bool unchange_input) {
  auto& src = var->Var();
  auto* dst = dst_var->MutableVar();
  if (dst->IsType<framework::LoDTensor>()) {
    if (src.IsType<framework::LoDTensor>()) {
      TensorAdd<framework::Variable>(src, dst);
    } else if (src.IsType<phi::SelectedRows>()) {
      SelectedRowsAddToTensor(src, dst);
    } else {
      PADDLE_THROW(platform::errors::InvalidArgument(
          "Unexpected branch, output variable type is %s",
          framework::ToTypeName(dst->Type())));
    }
  } else {
    if (src.IsType<framework::LoDTensor>()) {
      if (unchange_input) {
        framework::Variable new_dst;
        SelectedRowsAddTensor(*dst, src, &new_dst);
        *dst = std::move(new_dst);
      } else {
        auto* src_mutable = var->MutableVar();
        SelectedRowsAddToTensor(*dst, src_mutable);
        *dst = std::move(*(var->MutableVar()));
      }
    } else if (src.IsType<phi::SelectedRows>()) {
      auto temp = SelectedRowsMerge<VariableWrapper>(src, *dst);
      *dst = std::move(*(temp->MutableVar()));
    } else {
      PADDLE_THROW(platform::errors::InvalidArgument(
          "Unexpected branch, output variable type is %s",
          framework::ToTypeName(dst->Type())));
    }
  }
}

static platform::Place GetPlaceOfVar(
    const std::shared_ptr<VariableWrapper>& var) {
  platform::Place place;
  if (var->Var().IsType<framework::LoDTensor>()) {
    place = var->Var().Get<framework::LoDTensor>().place();
  } else if (var->Var().IsType<phi::SelectedRows>()) {
    place = var->Var().Get<phi::SelectedRows>().place();
  } else {
    PADDLE_THROW(platform::errors::InvalidArgument(
        "only support LoDTensor and SelectedRows in dygraph"));
  }
  return place;
}

void GradientAccumulator::AccumulateGrad() {
  /**
   * If the leaf gradient has been calculated done, the inner_var_
   * should be added to the var_.
   */
  if (!var_->IsLeafGrad() || !SumGradCompleted() || !HasInnerVar()) {
    return;
  }
  PADDLE_ENFORCE_EQ(HasInnerVar(),
                    true,
                    platform::errors::InvalidArgument(
                        "Leaf tensor should have inner var to store results of "
                        "this auto-grad"));
  PADDLE_ENFORCE_EQ(inner_var_->Var().IsInitialized(),
                    true,
                    platform::errors::InvalidArgument(
                        "Interior var of Leaf tensor should be initialized."));
  auto* src = inner_var_->MutableVar();
  auto* dst = var_->MutableVar();
  if (!var_->IsEmpty()) {
    VLOG(6) << "Leaf Var(" << var_->Name()
            << ")'s Gradient has been initizlized, will accumulate on "
               "previous gradient.";
    if (dst->IsType<framework::LoDTensor>()) {
      if (src->IsType<framework::LoDTensor>()) {
        TensorAdd<framework::Variable>(*src, dst);
      } else if (src->IsType<phi::SelectedRows>()) {
        SelectedRowsAddToTensor(*src, dst);
      }
    } else if (dst->IsType<phi::SelectedRows>()) {
      if (src->IsType<framework::LoDTensor>()) {
        SelectedRowsAddToTensor(*dst, src);
        *dst = std::move(*src);
      } else if (src->IsType<phi::SelectedRows>()) {
        auto temp = SelectedRowsMerge<VariableWrapper>(*src, *dst);
        *dst = std::move(*(temp->MutableVar()));
      }
    } else {
      PADDLE_THROW(platform::errors::PermissionDenied(
          "Only support LoDTensor and SelectedRows for gradient var"));
    }
  } else {
    VLOG(6)
        << "Leaf Var(" << var_->Name()
        << ")'s Gradient has not been initialized, not accumulate. Just move";
    *(dst) = std::move(*src);
    var_->SetType(inner_var_->Type());
    var_->SetDataType(inner_var_->DataType());
    var_->SetIsEmpty(false);
  }
  inner_var_.reset();
}

void GradientAccumulator::CallGradientHooks() {
  PADDLE_ENFORCE_EQ(var_->IsLeafGrad(),
                    true,
                    platform::errors::Unavailable(
                        "Only leaf gradient Tensor can deal with by gradient "
                        "hook in gradient accumulator."));
  PADDLE_ENFORCE_EQ(
      SumGradCompleted(),
      true,
      platform::errors::PreconditionNotMet(
          "Only can call gradient hooks after sum gradient completed."));
  PADDLE_ENFORCE_EQ(
      HasInnerVar(),
      true,
      platform::errors::PreconditionNotMet(
          "Leaf Tensor's inner var is nullptr when call gradient hook."));
  PADDLE_ENFORCE_EQ(
      inner_var_->Var().IsInitialized(),
      true,
      platform::errors::PreconditionNotMet("Leaf Tensor's inner var "
                                           "is not initialized when "
                                           "call gradient hook."));
  if (var_->HasVariableWrapperHook()) {
    VLOG(3) << "Call " << var_->GetVariableWrapperHooks().size()
            << " hooks of leaf gradient accumulator's inner var `"
            << var_->Name() << "`.";
    auto tmp_var = inner_var_;
    VLOG(3) << "Input var " << var_->Name() << "'s hook size - "
            << var_->GetVariableWrapperHooks().size();
    for (const auto& hook_pair : var_->GetVariableWrapperHooks()) {
      tmp_var = (*hook_pair.second)(tmp_var);
      CheckVar(inner_var_, tmp_var);
    }
    inner_var_ = tmp_var;
  }
}

void GradientAccumulator::CallReduceHooks() {
  PADDLE_ENFORCE_EQ(
      var_->IsLeafGrad(),
      true,
      platform::errors::Unavailable("Only leaf gradient Tensor can deal with "
                                    "by reduce hook in gradient accumulator."));
  PADDLE_ENFORCE_EQ(SumGradCompleted(),
                    true,
                    platform::errors::PreconditionNotMet(
                        "Only can call reduce hooks after the gradient "
                        "summation is completed in current batch."));
  PADDLE_ENFORCE_EQ(HasInnerVar(),
                    false,
                    platform::errors::PreconditionNotMet(
                        "Only can call reduce hooks after the "
                        "gradient accumulation is completed in "
                        "current batch or across batchs."));
  if (var_->HasVoidHook()) {
    for (const auto& hook : var_->GetVoidHooks()) {
      VLOG(3) << "call gradient accumulator backward hooks.";
      (*hook)();
    }
  }
}

void EagerGradientAccumulator::SumGrad(std::shared_ptr<VariableWrapper> var,
                                       size_t trace_id,
                                       bool unchange_input) {
  /**
   * If var has grad node, it indicates that this var would be an input
   * of a grad op. Therefore, it should not be changed.
   */
  if (var->HasGradNode()) {
    unchange_input = true;
  }

  auto* dst_var = Var();
  platform::Place place = GetPlaceOfVar(var);
  if (!dst_var->OverridedStopGradient()) {
    if (CurCnt() == 0) {
      MoveOrCopyVar(dst_var->MutableVar(), var->MutableVar(), unchange_input);
    } else {
      VLOG(6) << "Sum Gradient for: " << dst_var->Name()
              << " within this graph.";
      VariableWrapperAdd(var, dst_var, unchange_input);
    }
  } else {
    if (!dst_var->Var().IsInitialized() ||
        !dst_var->Var().Get<framework::LoDTensor>().IsInitialized()) {
      VLOG(6) << "Set StopGradient Grad: " << dst_var->Name() << " as zero ";
      auto* dev_ctx = platform::DeviceContextPool::Instance().Get(place);
      if (!dst_var->Var().IsInitialized()) {
        auto* tensor =
            dst_var->MutableVar()->GetMutable<framework::LoDTensor>();
        VLOG(6) << "Dims of " << dst_var->Name() << " is set as: "
                << var->Var().Get<framework::LoDTensor>().dims();
        tensor->Resize(var->Var().Get<framework::LoDTensor>().dims());
        tensor->mutable_data(place,
                             framework::TransToPhiDataType(var->DataType()));
        phi::funcs::set_constant(*dev_ctx, tensor, 0.0);
      } else {
        auto* tensor =
            dst_var->MutableVar()->GetMutable<framework::LoDTensor>();
        tensor->mutable_data(place,
                             framework::TransToPhiDataType(var->DataType()));
        phi::funcs::set_constant(*dev_ctx, tensor, 0.0);
      }
    }
  }

  // Type may be changed after OP run, such as VarTypeInference
  // so synchronous VariableWrapper with Variable.
  if (dst_var->Var().IsType<framework::LoDTensor>()) {
    dst_var->SetType(framework::proto::VarType::LOD_TENSOR);
  } else if (dst_var->Var().IsType<phi::SelectedRows>()) {
    dst_var->SetType(framework::proto::VarType::SELECTED_ROWS);
  }

  // Increase curent count
  IncreaseCurCnt();
}

void SortedGradientAccumulator::SumGrad(std::shared_ptr<VariableWrapper> var,
                                        size_t trace_id,
                                        bool unchange_input) {
  auto* dst_var = Var();
  platform::Place place = GetPlaceOfVar(var);
  if (!dst_var->OverridedStopGradient()) {
    if (ref_cnt_ == 1) {
      MoveOrCopyVar(dst_var->MutableVar(),
                    var->MutableVar(),
                    unchange_input || var->HasGradNode());
    } else {
      if (tmp_grad_vars_.empty()) {
        tmp_grad_vars_.reserve(ref_cnt_);
      }

      tmp_grad_vars_.emplace_back(std::move(var), trace_id, unchange_input);

      if (tmp_grad_vars_.size() != ref_cnt_) {
        return;
      }

      VLOG(6) << "Sum Gradient for: " << dst_var->Name()
              << " within this graph.";
      std::sort(tmp_grad_vars_.begin(),
                tmp_grad_vars_.end(),
                [](const SavedVarInfo& info1, const SavedVarInfo& info2) {
                  return info1.trace_id > info2.trace_id;
                });

      for (auto& var_info : tmp_grad_vars_) {
        if (var_info.var->HasGradNode()) {
          var_info.unchange_input = true;
        }
      }

#if defined(PADDLE_WITH_CUDA) || defined(PADDLE_WITH_HIP)
      if (paddle::platform::is_gpu_place(place)) {
        // sum selected rows firstly
        for (auto& var_info : tmp_grad_vars_) {
          if (!var_info.var->Var().IsType<phi::SelectedRows>()) {
            continue;
          }

          if (CurCnt() == 0) {
            MoveOrCopyVar(dst_var->MutableVar(),
                          var_info.var->MutableVar(),
                          var_info.unchange_input);
          } else {
            VariableWrapperAdd(var_info.var, dst_var, var_info.unchange_input);
          }

          var_info.var = nullptr;
          // Increase count
          IncreaseCurCnt();
        }

        for (auto& var_info : tmp_grad_vars_) {
          if (!var_info.var) {
            continue;
          }

          PADDLE_ENFORCE_EQ(var_info.var->Var().IsType<framework::LoDTensor>(),
                            true,
                            platform::errors::PermissionDenied(
                                "Gradient var must be LoDTensor"));
          if (CurCnt() == 0) {
            MoveOrCopyVar(dst_var->MutableVar(),
                          var_info.var->MutableVar(),
                          var_info.unchange_input);
          } else {
            VariableWrapperAdd(var_info.var, dst_var, var_info.unchange_input);
          }

          var_info.var = nullptr;
          // Increase count
          IncreaseCurCnt();
        }
      } else {
#endif
        for (auto& var_info : tmp_grad_vars_) {
          if (!var_info.var) {
            continue;
          }
          PADDLE_ENFORCE_EQ(
              var_info.var->Var().IsType<framework::LoDTensor>() ||
                  var_info.var->Var().IsType<phi::SelectedRows>(),
              true,
              platform::errors::PermissionDenied("The type of Gradient "
                                                 "var must be LoDTensor "
                                                 "or SelectedRows"));
          if (CurCnt() == 0) {
            MoveOrCopyVar(dst_var->MutableVar(),
                          var_info.var->MutableVar(),
                          var_info.unchange_input);
          } else {
            VariableWrapperAdd(var_info.var, dst_var, var_info.unchange_input);
          }
          var_info.var = nullptr;
          // Increase count
          IncreaseCurCnt();
        }
#if defined(PADDLE_WITH_CUDA) || defined(PADDLE_WITH_HIP)
      }
#endif
      tmp_grad_vars_.clear();
    }
  } else {
    if (!dst_var->Var().IsInitialized() ||
        !dst_var->Var().Get<framework::LoDTensor>().IsInitialized()) {
      VLOG(6) << "Set StopGradient Grad: " << var->Name() << " as zero";
      auto* dev_ctx = platform::DeviceContextPool::Instance().Get(place);
      if (!dst_var->Var().IsInitialized()) {
        auto* tensor =
            dst_var->MutableVar()->GetMutable<framework::LoDTensor>();
        VLOG(6) << "Dims of " << dst_var->Name() << " is set as: "
                << var->Var().Get<framework::LoDTensor>().dims();
        tensor->Resize(var->Var().Get<framework::LoDTensor>().dims());
        tensor->mutable_data(place,
                             framework::TransToPhiDataType(var->DataType()));
        phi::funcs::set_constant(*dev_ctx, tensor, 0.0);
      } else {
        auto* tensor =
            dst_var->MutableVar()->GetMutable<framework::LoDTensor>();
        tensor->mutable_data(place,
                             framework::TransToPhiDataType(var->DataType()));
        phi::funcs::set_constant(*dev_ctx, tensor, 0.0);
      }
    }
    // looks like tmp_grad_vars will not have any member but just in case
    tmp_grad_vars_.clear();
  }

  if (dst_var->Var().IsType<framework::LoDTensor>()) {
    dst_var->SetType(framework::proto::VarType::LOD_TENSOR);
  } else if (dst_var->Var().IsType<phi::SelectedRows>()) {
    dst_var->SetType(framework::proto::VarType::SELECTED_ROWS);
  }
}

}  // namespace imperative
}  // namespace paddle