gradient_accumulator.cc 14.2 KB
Newer Older
J
Jiabin Yang 已提交
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18
// 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>
19
#include "paddle/fluid/framework/framework.pb.h"
J
Jiabin Yang 已提交
20 21 22 23 24
#include "paddle/fluid/framework/lod_tensor.h"
#include "paddle/fluid/framework/selected_rows.h"
#include "paddle/fluid/imperative/layer.h"
#include "paddle/fluid/operators/math/blas.h"
#include "paddle/fluid/operators/math/math_function.h"
25
#include "paddle/fluid/operators/math/selected_rows_functor.h"
J
Jiabin Yang 已提交
26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88
#include "paddle/fluid/platform/device_context.h"
#include "paddle/fluid/platform/profiler.h"

namespace paddle {
namespace imperative {

template <typename T>
class TensorAddFunctor : public boost::static_visitor<> {
 public:
  TensorAddFunctor(int64_t numel, const T* x, T* y)
      : numel_(numel), x_(x), y_(y) {}

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

#ifdef PADDLE_WITH_CUDA
  void operator()(const platform::CUDAPlace& place) {
    platform::CUDADeviceContext* ctx =
        dynamic_cast<platform::CUDADeviceContext*>(
            platform::DeviceContextPool::Instance().Get(place));
    auto blas = operators::math::GetBlas<platform::CUDADeviceContext, T>(*ctx);
    blas.AXPY(numel_, 1., x_, y_);
  }
#else
  void operator()(const platform::CUDAPlace& place) {
    PADDLE_THROW("Do NOT support gradient merge in place %s", place);
  }
#endif

  // there is NO blas in CUDAPinnedPlace
  void operator()(const platform::CUDAPinnedPlace& place) {
    PADDLE_THROW("Do NOT support gradient merge in place %s", place);
  }

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

void TensorAdd(const framework::Variable& src, framework::Variable* dst) {
  auto* dst_tensor = dst->GetMutable<framework::LoDTensor>();
  auto& src_tensor = src.Get<framework::LoDTensor>();

  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, true,
                    "dst_numel %d vs. src_numel %d", dst_tensor->numel(),
                    numel);

  auto data_type = src_tensor.type();
  auto place = src_tensor.place();

89
#define PADDLE_TENSOR_ADD(cpp_type)                                  \
J
Jiabin Yang 已提交
90 91 92 93 94 95 96 97
  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));                  \
    boost::apply_visitor(func, place);                               \
    return;                                                          \
  }

98 99
  PADDLE_TENSOR_ADD(float);
  PADDLE_TENSOR_ADD(double);
J
Jiabin Yang 已提交
100

101
#undef PADDLE_TENSOR_ADD
J
Jiabin Yang 已提交
102 103 104 105 106

  PADDLE_THROW("Not supported data type %s for AddTo",
               framework::DataTypeToString(data_type));
}

107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233
void SelectedRowsAddToTensor(const framework::Variable& src,
                             framework::Variable* dst) {
  auto* dst_tensor = dst->GetMutable<framework::LoDTensor>();
  auto& src_selected_rows = src.Get<framework::SelectedRows>();
  auto place = dst_tensor->place();
  auto data_type = src_selected_rows.value().type();
  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;                                                                  \
  }

#ifdef PADDLE_WITH_CUDA
  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(platform::CPUDeviceContext, float);
    PADDLE_SELECTED_ROWS_ADD_TO_TENSOR(platform::CPUDeviceContext, double);
#ifdef PADDLE_WITH_CUDA
  }
#endif

#undef PADDLE_SELECTED_ROWS_ADD_TO_TENSOR

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

// 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
std::shared_ptr<VarBase> SelectedRowsMerge(const framework::Variable& src1,
                                           const framework::Variable& src2) {
  auto& src_selected_rows1 = src1.Get<framework::SelectedRows>();
  auto& src_selected_rows2 = src2.Get<framework::SelectedRows>();
  auto place = src_selected_rows1.value().place();
  auto data_type = src_selected_rows1.value().type();
  platform::DeviceContextPool& pool = platform::DeviceContextPool::Instance();

  std::vector<const framework::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<VarBase>(false, "Temp");
  auto* dst_selected_rows =
      dst_var->MutableVar()->GetMutable<framework::SelectedRows>();

#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;                                                       \
  }

#ifdef PADDLE_WITH_CUDA
  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(platform::CPUDeviceContext, float);
    PADDLE_SELECTED_ROWS_ADD(platform::CPUDeviceContext, double);
#ifdef PADDLE_WITH_CUDA
  }
#endif

#undef PADDLE_SELECTED_ROWS_ADD

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

void VarBaseAdd(std::shared_ptr<VarBase> var, VarBase* var_) {
  auto& src = var->Var();
  auto* dst = var_->MutableVar();
  if (dst->IsType<framework::LoDTensor>()) {
    if (src.IsType<framework::LoDTensor>()) {
      TensorAdd(src, dst);
    } else if (src.IsType<framework::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>()) {
      auto* src_mutable = var->MutableVar();
      SelectedRowsAddToTensor(*dst, src_mutable);
      *dst = std::move(*(var->MutableVar()));
      var_->SetType(framework::proto::VarType::LOD_TENSOR);
    } else if (src.IsType<framework::SelectedRows>()) {
      std::shared_ptr<VarBase> temp = SelectedRowsMerge(src, *dst);
      *dst = std::move(*(temp->MutableVar()));
    } else {
      PADDLE_THROW(platform::errors::InvalidArgument(
          "Unexpected branch, output variable type is %s",
          framework::ToTypeName(dst->Type())));
    }
  }
}

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

J
Jiabin Yang 已提交
234 235 236
void EagerGradientAccumulator::Add(std::shared_ptr<VarBase> var,
                                   size_t trace_id) {
  auto* dst_var = var_->MutableVar();
237
  platform::Place place = GetPlaceOfVarBase(var);
238 239 240
  if (!var_->OverridedStopGradient()) {
    VLOG(3) << "Sum Gradient for: " << var_->Name();
    if (cur_cnt_ == 0) {
241 242 243
      if (var->Var().IsType<framework::SelectedRows>()) {
        var_->SetType(framework::proto::VarType::SELECTED_ROWS);
      }
244 245
      *dst_var = std::move(*(var->MutableVar()));
    } else {
246
      VarBaseAdd(var, var_);
247
    }
J
Jiabin Yang 已提交
248
  } else {
249 250
    if (!var_->Var().IsInitialized() ||
        !var_->Var().Get<framework::LoDTensor>().IsInitialized()) {
251 252
      VLOG(6) << "Set StopGradient Grad: " << var_->Name() << " as zero ";

253
      auto* dev_ctx = platform::DeviceContextPool::Instance().Get(place);
254 255 256 257 258 259 260 261 262 263 264 265
      if (!var_->Var().IsInitialized()) {
        auto* tensor = var_->MutableVar()->GetMutable<framework::LoDTensor>();
        VLOG(6) << "Dims of " << var_->Name() << " is set as: "
                << var->Var().Get<framework::LoDTensor>().dims();
        tensor->Resize(var->Var().Get<framework::LoDTensor>().dims());
        tensor->mutable_data(place, var->DataType());
        operators::math::set_constant(*dev_ctx, tensor, 0.0);
      } else {
        auto* tensor = var_->MutableVar()->GetMutable<framework::LoDTensor>();
        tensor->mutable_data(place, var->DataType());
        operators::math::set_constant(*dev_ctx, tensor, 0.0);
      }
266
    }
J
Jiabin Yang 已提交
267 268 269 270 271 272 273
  }
  ++cur_cnt_;
}

void SortedGradientAccumulator::Add(std::shared_ptr<VarBase> var,
                                    size_t trace_id) {
  auto* dst_var = var_->MutableVar();
274
  platform::Place place = GetPlaceOfVarBase(var);
275 276
  if (!var_->OverridedStopGradient()) {
    if (ref_cnt_ == 1) {
277 278 279 280 281 282
      if (var->Var().IsType<framework::SelectedRows>()) {
        var_->SetType(framework::proto::VarType::SELECTED_ROWS);
        *dst_var = std::move(*(var->MutableVar()));
      } else {
        *dst_var = std::move(*(var->MutableVar()));
      }
283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299
    } else {
      if (tmp_grad_vars_.empty()) {
        tmp_grad_vars_.reserve(ref_cnt_);
      }

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

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

      std::sort(tmp_grad_vars_.begin(), tmp_grad_vars_.end(),
                [](const std::pair<std::shared_ptr<VarBase>, size_t>& p1,
                   const std::pair<std::shared_ptr<VarBase>, size_t>& p2) {
                  return p1.second > p2.second;
                });

300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338
#ifdef PADDLE_WITH_CUDA
      if (paddle::platform::is_gpu_place(place)) {
        bool dst_varbase_is_initialized = false;
        // accumulate selected rows firstly
        for (size_t i = 0; i < tmp_grad_vars_.size(); ++i) {
          if (tmp_grad_vars_[i]
                  .first->Var()
                  .IsType<framework::SelectedRows>()) {
            if (!dst_varbase_is_initialized) {
              dst_varbase_is_initialized = true;
              var_->SetType(framework::proto::VarType::SELECTED_ROWS);
              *dst_var = std::move(*(tmp_grad_vars_[i].first->MutableVar()));
            } else {
              VarBaseAdd(tmp_grad_vars_[i].first, var_);
            }
          }
        }
        // accumulate lod tensor
        for (size_t i = 0; i < tmp_grad_vars_.size(); ++i) {
          if (!dst_varbase_is_initialized) {
            dst_varbase_is_initialized = true;
            *dst_var = std::move(*(tmp_grad_vars_[0].first->MutableVar()));
          }
          if (tmp_grad_vars_[i].first->Var().IsType<framework::LoDTensor>()) {
            VarBaseAdd(tmp_grad_vars_[i].first, var_);
          }
        }
      } else {
#endif
        if (tmp_grad_vars_[0].first->Var().IsType<framework::SelectedRows>()) {
          var_->SetType(framework::proto::VarType::SELECTED_ROWS);
          *dst_var = std::move(*(tmp_grad_vars_[0].first->MutableVar()));
        } else {
          *dst_var = std::move(*(tmp_grad_vars_[0].first->MutableVar()));
        }
        for (size_t i = 1; i < tmp_grad_vars_.size(); ++i) {
          VarBaseAdd(tmp_grad_vars_[i].first, var_);
        }
#ifdef PADDLE_WITH_CUDA
339
      }
340
#endif
341
      tmp_grad_vars_.clear();
J
Jiabin Yang 已提交
342
    }
343 344 345 346 347
  } else {
    if (!var_->Var().IsInitialized() ||
        !var_->Var().Get<framework::LoDTensor>().IsInitialized()) {
      VLOG(6) << "Set StopGradient Grad: " << var->Name() << " as zero";
      auto* dev_ctx = platform::DeviceContextPool::Instance().Get(place);
348 349 350 351 352 353 354 355 356 357 358 359
      if (!var_->Var().IsInitialized()) {
        auto* tensor = var_->MutableVar()->GetMutable<framework::LoDTensor>();
        VLOG(6) << "Dims of " << var_->Name() << " is set as: "
                << var->Var().Get<framework::LoDTensor>().dims();
        tensor->Resize(var->Var().Get<framework::LoDTensor>().dims());
        tensor->mutable_data(place, var->DataType());
        operators::math::set_constant(*dev_ctx, tensor, 0.0);
      } else {
        auto* tensor = var_->MutableVar()->GetMutable<framework::LoDTensor>();
        tensor->mutable_data(place, var->DataType());
        operators::math::set_constant(*dev_ctx, tensor, 0.0);
      }
J
Jiabin Yang 已提交
360
    }
361
    // looks like tmp_grad_vars will not have any member but just in case
J
Jiabin Yang 已提交
362 363 364 365 366 367
    tmp_grad_vars_.clear();
  }
}

}  // namespace imperative
}  // namespace paddle