grad_node_info.cc 16.2 KB
Newer Older
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15
// Copyright (c) 2021 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/eager/grad_node_info.h"
16
#include "paddle/fluid/eager/accumulation/accumulation_node.h"
17
#include "paddle/fluid/eager/autograd_meta.h"
18 19
#include "paddle/fluid/eager/utils.h"

20 21
#include "paddle/phi/common/data_type.h"
#include "paddle/phi/core/dense_tensor.h"
22
#include "paddle/phi/core/sparse_coo_tensor.h"
23

24 25 26
#include "paddle/fluid/framework/convert_utils.h"
#include "paddle/fluid/framework/data_type.h"
#include "paddle/fluid/framework/data_type_transform.h"
27
#include "paddle/fluid/framework/var_type.h"
28

29 30 31 32 33 34
#include "paddle/fluid/platform/enforce.h"
#include "paddle/fluid/platform/errors.h"

#include "glog/logging.h"

/**
35
 * Implementation of GradNodeBase, Edge and GradTensorHolder.
36 37 38 39
**/
namespace egr {

GradNodeBase::GradNodeBase(size_t bwd_in_slot_num, size_t bwd_out_slot_num) {
J
Jiabin Yang 已提交
40
  VLOG(6) << "Construct GradNodeBase";
41 42 43 44 45
  bwd_in_meta_.resize(bwd_in_slot_num);
  bwd_out_meta_.resize(bwd_out_slot_num);
  adj_edges_.resize(bwd_out_slot_num);
}

46 47 48 49 50 51 52
void GradNodeBase::AddEdges(std::vector<AutogradMeta*>* metas, size_t slot_id) {
  PADDLE_ENFORCE_LT(
      slot_id, adj_edges_.size(),
      paddle::platform::errors::InvalidArgument(
          "Given slot id is out of range of adj_edges outter size, "
          "adj_edges is designed to has the same size of grad "
          "inputs's slot num."));
53 54 55

  for (size_t i = 0; i < metas->size(); i++) {
    const auto& meta = (*metas)[i];
56 57 58
    // adj_edges has as same rank as fwd inputs, and record it's output rank
    // from
    // its pre-ops
59
    if (meta && !meta->StopGradient()) {
60
      auto node = meta->GetMutableGradNode();
61
      if (!node || !node.get()) {
62
        meta->SetGradNode(std::make_shared<egr::GradNodeAccumulation>(meta));
63
      }
64 65 66 67
      VLOG(6) << "Add Edges for slot: " << slot_id << ", the Edge is from "
              << this->name() << " (addr: " << this << ") "
              << " to " << meta->GetMutableGradNode()->name()
              << " (addr: " << meta->GetMutableGradNode().get() << ")";
68 69 70

      adj_edges_[slot_id].emplace_back(meta->GetMutableGradNode(),
                                       meta->OutRankInfo());
J
Jiabin Yang 已提交
71 72
    } else {
      adj_edges_[slot_id].emplace_back();
73
    }
74 75 76
  }
}

77
void GradNodeBase::AddEdges(AutogradMeta* meta, size_t slot_id) {
78 79 80 81 82 83
  PADDLE_ENFORCE_LT(
      slot_id, adj_edges_.size(),
      paddle::platform::errors::InvalidArgument(
          "Given slot id is out of range of adj_edges outter size, "
          "adj_edges is designed to has the same size of grad "
          "inputs's slot num."));
84

85
  if (meta && !meta->StopGradient()) {
86
    auto node = meta->GetMutableGradNode();
87
    if (!node || !node.get()) {
88
      meta->SetGradNode(std::make_shared<egr::GradNodeAccumulation>(meta));
89
    }
90
    VLOG(6) << "Add Edges for slot: " << slot_id << ", the Edge is from "
91 92 93
            << this->name() << " (addr: " << this << ") "
            << " to " << meta->GetMutableGradNode()->name()
            << " (addr: " << meta->GetMutableGradNode().get() << ")";
94 95 96

    adj_edges_[slot_id].emplace_back(meta->GetMutableGradNode(),
                                     meta->OutRankInfo());
J
Jiabin Yang 已提交
97 98
  } else {
    adj_edges_[slot_id].emplace_back();
99
  }
100 101
}

102
const std::vector<std::vector<GradSlotMeta>>& GradNodeBase::InputMeta() const {
103 104 105
  return bwd_in_meta_;
}

106
const std::vector<std::vector<GradSlotMeta>>& GradNodeBase::OutputMeta() const {
107 108 109
  return bwd_out_meta_;
}

110
void GradNodeBase::SetGradInMeta(const paddle::experimental::Tensor& fwd_out,
111
                                 size_t slot_rank) {
112
  VLOG(6) << "Set GradSlotMeta for Grad Inputs";
113
  auto* fwd_out_meta = egr::EagerUtils::nullable_autograd_meta(fwd_out);
114 115 116 117 118 119
  PADDLE_ENFORCE_LE(
      slot_rank, (bwd_in_meta_.size() - 1),
      paddle::platform::errors::InvalidArgument(
          "Slot Rank should less equal than bwd_in_meta_ size, since "
          "bwd_in_meta_ is designed to hold as same num as backward "
          "inputs."));
120 121 122 123 124 125 126 127
  auto& metas = bwd_in_meta_.at(slot_rank);
  if (metas.size() == 0) {
    metas.resize(1);
  }

  auto& meta = metas[0];
  meta.SetStopGradient(fwd_out_meta->StopGradient());

128
  if (!fwd_out.initialized()) {
129 130 131 132 133
    VLOG(6)
        << "Skip Configuring GradSlotMeta for uninitialized GradInput Tensor";
    return;
  }

134
  phi::DenseTensor* dense_tensor = nullptr;
135 136 137
  // Record TensorMeta
  if (phi::DenseTensor::classof(fwd_out.impl().get())) {
    // Only Copy Meta
138 139 140 141 142
    dense_tensor = static_cast<phi::DenseTensor*>(fwd_out.impl().get());
  } else if (phi::SparseCooTensor::classof(fwd_out.impl().get())) {
    phi::SparseCooTensor* coo_tensor =
        static_cast<phi::SparseCooTensor*>(fwd_out.impl().get());
    dense_tensor = coo_tensor->mutable_non_zero_elements();
143 144 145
  } else {
    VLOG(6) << "Unable to initialize the DenseTensorMeta of GradSlotMeta with "
               "non-DenseTensor argument.";
146
  }
147 148 149 150 151 152 153
  PADDLE_ENFORCE_NE(
      dense_tensor->meta().dtype, phi::DataType::UNDEFINED,
      paddle::platform::errors::Fatal(
          "Attempting to copy DenseTensorMeta with phi::DataType::UNDEFINED,"
          "which is illegal."));

  meta.SetTensorMeta(dense_tensor->meta());
C
Chen Weihang 已提交
154
  meta.SetPlace(fwd_out.place());
155 156 157 158 159

  if (paddle::framework::IsComplexType(
          paddle::framework::TransToProtoVarType(dense_tensor->type()))) {
    need_complex_to_real_ = true;
  }
160 161
}

162 163 164
void GradNodeBase::SetGradInMeta(
    const std::vector<paddle::experimental::Tensor>& fwd_out,
    size_t slot_rank) {
165
  VLOG(6) << "Set GradSlotMeta for Grad Inputs";
166
  size_t slot_size = fwd_out.size();
167 168 169 170 171 172
  PADDLE_ENFORCE_LE(
      slot_rank, (bwd_in_meta_.size() - 1),
      paddle::platform::errors::InvalidArgument(
          "Slot Rank should less equal than bwd_in_meta_ size, since "
          "bwd_in_meta_ is designed to hold as same num as backward "
          "inputs."));
173
  auto& metas = bwd_in_meta_.at(slot_rank);
174
  // Init stop gradient vector before use to avoid push back
175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194
  if (metas.size() < slot_size) {
    VLOG(7) << "Init bwd_in_meta_ with slot rank: " << slot_rank;
    metas.resize(slot_size);
  }
  for (size_t i = 0; i < slot_size; i++) {
    auto& meta = metas[i];
    const auto& fwd_out_tensor = fwd_out[i];
    auto* fwd_out_meta =
        egr::EagerUtils::nullable_autograd_meta(fwd_out_tensor);
    PADDLE_ENFORCE_NOT_NULL(fwd_out_meta,
                            paddle::platform::errors::PreconditionNotMet(
                                "Bwd_in_meta should only be called while "
                                "autograd_meta is not null. If you got this "
                                "error, it indicates bugs in framework."));
    if (fwd_out_meta->StopGradient()) {
      // Set Stop Gradient only when its true or non-initialized autograd_meta,
      // since all default value is false.
      meta.SetStopGradient(fwd_out_meta->StopGradient());
    }

195
    if (!fwd_out_tensor.initialized()) {
196 197 198 199 200
      VLOG(6)
          << "Skip Configuring GradSlotMeta for uninitialized GradInput Tensor";
      return;
    }

201 202 203 204 205 206 207 208 209 210 211 212
    // Record TensorMeta
    if (phi::DenseTensor::classof(fwd_out_tensor.impl().get())) {
      // Only Copy Meta
      phi::DenseTensor* dense_tensor =
          static_cast<phi::DenseTensor*>(fwd_out_tensor.impl().get());

      PADDLE_ENFORCE_NE(
          dense_tensor->meta().dtype, phi::DataType::UNDEFINED,
          paddle::platform::errors::Fatal("Attempting to copy DenseTensorMeta "
                                          "with phi::DataType::UNDEFINED,"
                                          "which is illegal."));
      meta.SetTensorMeta(dense_tensor->meta());
C
Chen Weihang 已提交
213
      meta.SetPlace(fwd_out_tensor.place());
214

215 216 217 218 219 220 221 222 223
      if (paddle::framework::IsComplexType(
              paddle::framework::TransToProtoVarType(dense_tensor->type()))) {
        need_complex_to_real_ = true;
      }
    } else {
      VLOG(6) << "Unable to initialize the DenseTensorMeta of GradSlotMeta "
                 "with non-DenseTensor argument.";
    }
  }
224 225
}

226
void GradNodeBase::SetGradOutMeta(const paddle::experimental::Tensor& fwd_in,
227
                                  size_t slot_rank) {
228
  auto* fwd_in_meta = egr::EagerUtils::nullable_autograd_meta(fwd_in);
229
  PADDLE_ENFORCE_LE(
230
      (slot_rank + 1), bwd_out_meta_.size(),
231 232 233 234
      paddle::platform::errors::InvalidArgument(
          "Slot Rank should less equal than bwd_out_meta_ size, "
          "since bwd_out_meta_ is designed to hold as same num as "
          "backward outputs."));
235
  auto& metas = bwd_out_meta_.at(slot_rank);
236
  // Init stop gradient vector before use to avoid push back
237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258
  if (metas.size() == 0) {
    metas.resize(1);
  }
  auto& meta = metas[0];
  if (fwd_in_meta) {
    meta.SetStopGradient(fwd_in_meta->StopGradient());
  } else {
    meta.SetStopGradient(true);
  }

  // Record TensorMeta
  if (fwd_in.impl() && fwd_in.impl().get()) {
    if (phi::DenseTensor::classof(fwd_in.impl().get())) {
      // Only Copy Meta
      phi::DenseTensor* dense_tensor =
          static_cast<phi::DenseTensor*>(fwd_in.impl().get());
      PADDLE_ENFORCE_NE(
          dense_tensor->meta().dtype, phi::DataType::UNDEFINED,
          paddle::platform::errors::Fatal("Attempting to copy DenseTensorMeta "
                                          "with phi::DataType::UNDEFINED,"
                                          "which is illegal."));
      meta.SetTensorMeta(dense_tensor->meta());
C
Chen Weihang 已提交
259
      meta.SetPlace(fwd_in.place());
260
    }
261 262 263
  } else {
    VLOG(6) << "Unable to initialize the DenseTensorMeta of GradSlotMeta with "
               "non-DenseTensor argument.";
264 265 266
  }
}

267 268 269
void GradNodeBase::SetGradOutMeta(
    const std::vector<paddle::experimental::Tensor>& fwd_in, size_t slot_rank) {
  size_t slot_size = fwd_in.size();
270
  PADDLE_ENFORCE_LE(
271
      slot_rank, (bwd_out_meta_.size() - 1),
272 273 274 275
      paddle::platform::errors::InvalidArgument(
          "Slot Rank should less equal than bwd_out_meta_ size, "
          "since bwd_out_meta_ is designed to hold as same num as "
          "backward outputs."));
276
  auto& metas = bwd_out_meta_.at(slot_rank);
277
  // Init stop gradient vector before use to avoid push back
278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303
  if (metas.size() < slot_size) {
    metas.resize(slot_size);
  }
  for (size_t i = 0; i < slot_size; i++) {
    const auto& fwd_in_tensor = fwd_in[i];
    auto& meta = metas[i];
    auto* fwd_in_meta = egr::EagerUtils::nullable_autograd_meta(fwd_in_tensor);
    if (fwd_in_meta) {
      // Set Stop Gradient only when its true or non-initialized autograd_meta,
      // since all default value is false.
      meta.SetStopGradient(fwd_in_meta->StopGradient());
    }

    // Record TensorMeta
    if (fwd_in_tensor.impl() && fwd_in_tensor.impl().get()) {
      if (phi::DenseTensor::classof(fwd_in_tensor.impl().get())) {
        // Only Copy Meta
        phi::DenseTensor* dense_tensor =
            static_cast<phi::DenseTensor*>(fwd_in_tensor.impl().get());

        PADDLE_ENFORCE_NE(dense_tensor->meta().dtype, phi::DataType::UNDEFINED,
                          paddle::platform::errors::Fatal(
                              "Attempting to copy DenseTensorMeta with "
                              "phi::DataType::UNDEFINED,"
                              "which is illegal."));
        meta.SetTensorMeta(dense_tensor->meta());
C
Chen Weihang 已提交
304
        meta.SetPlace(fwd_in_tensor.place());
305 306 307 308 309
      }
    } else {
      VLOG(6) << "Unable to initialize the DenseTensorMeta of GradSlotMeta "
                 "with non-DenseTensor argument.";
    }
310
  }
311 312 313 314 315 316 317 318 319
}

void GradNodeBase::SetDefaultGradInOutMeta() {
  PADDLE_ENFORCE((bwd_out_meta_.size() == 1) && (bwd_in_meta_.size() == 1),
                 paddle::platform::errors::PreconditionNotMet(
                     "We can only support 1 input and 1 output in default grad "
                     "meta setter, other size of inputs and outputs should "
                     "create with Setter and Getters"));
  // Default stop_gradient is false and slot id is 0, slot size is 1;
320 321
  bwd_out_meta_[0].resize(1);
  bwd_in_meta_[0].resize(1);
322 323
}

324 325 326 327 328
int64_t GradNodeBase::RegisterGradientHook(
    size_t slot_id, size_t rank, std::shared_ptr<egr::TensorHook>&& hook) {
  gradient_hooks_.emplace(next_hook_id_,
                          std::make_tuple(slot_id, rank, std::move(hook)));
  return next_hook_id_++;
329 330
}

331 332 333 334
const std::vector<std::vector<Edge>>& GradNodeBase::GetEdges() const {
  return adj_edges_;
}

335 336 337 338
std::vector<std::vector<Edge>>& GradNodeBase::GetMutableEdges() {
  return adj_edges_;
}

339 340 341 342
std::vector<std::vector<paddle::experimental::Tensor>>
GradNodeBase::ApplyGradientHooks(
    const std::vector<std::vector<paddle::experimental::Tensor>>& tensors) {
  std::vector<std::vector<paddle::experimental::Tensor>> outs(tensors.size());
343 344 345 346 347
  for (auto& hook_pair : gradient_hooks_) {
    size_t slot_id = std::get<0>(hook_pair.second);
    size_t rank = std::get<1>(hook_pair.second);

    auto hook = std::get<2>(hook_pair.second);
348 349 350 351 352 353 354 355 356 357 358 359

    PADDLE_ENFORCE(slot_id < tensors.size(),
                   paddle::platform::errors::Fatal(
                       "Slot_id from registered hook should be smaller than "
                       "slot size of grad_tensors"));

    PADDLE_ENFORCE(rank < tensors[slot_id].size(),
                   paddle::platform::errors::Fatal(
                       "rank of slot %d from registered hook should be smaller "
                       "than rank size of grad_tensors",
                       slot_id));

360
    std::vector<paddle::experimental::Tensor>& slot_out = outs[slot_id];
361
    slot_out.resize(tensors[slot_id].size());
362
    paddle::experimental::Tensor& out = slot_out[rank];
363
    if (!out.defined() || !out.initialized()) {
364
      out = (*hook)(tensors[slot_id][rank]);
365
    } else {
366
      // If more than one hook is registered, the input to the next hook func
367
      // should be the output of the previous hook
368
      out = (*hook)(out);
369 370 371 372 373 374 375 376 377 378 379 380
    }
  }

  for (size_t i = 0; i < outs.size(); i++) {
    if (outs[i].empty() && (!tensors[i].empty())) {
      outs[i].resize(tensors[i].size());
    }
    // TODO(Jiabin): Optimize this if we only add hook slot by slot
    for (size_t j = 0; j < outs[i].size(); j++) {
      if (!outs[i][j].defined() || !outs[i][j].initialized()) {
        outs[i][j] = tensors[i][j];
      }
381
      CheckTensor(tensors[i][j], outs[i][j]);
382 383 384 385 386 387
    }
  }

  return outs;
}

388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428
void GradNodeBase::HandleComplexGradToRealGrad(
    std::vector<std::vector<paddle::experimental::Tensor>>* out_grads) {
  for (size_t slot_id = 0; slot_id < out_grads->size(); slot_id++) {
    const std::vector<paddle::experimental::Tensor>& slot_out_grads =
        (*out_grads)[slot_id];
    for (size_t rank_id = 0; rank_id < slot_out_grads.size(); rank_id++) {
      const GradSlotMeta& slot_meta = bwd_out_meta_[slot_id][rank_id];

      PADDLE_ENFORCE(
          slot_meta.HasTensorMeta() > 0,
          paddle::platform::errors::Fatal(
              "We require TensorMeta in GradInputMeta() to obtain forward data "
              "types."
              "However, no TensorMeta is detected in bwd_out_meta_."));

      auto fwd_data_type = paddle::framework::TransToProtoVarType(
          slot_meta.GetTensorMeta().dtype);
      const paddle::experimental::Tensor& grad = slot_out_grads[rank_id];

      if (paddle::framework::IsComplexType(fwd_data_type)) continue;

      // Only Handle Complex To Real for DenseTensor for now
      if (phi::DenseTensor::classof(grad.impl().get())) {
        phi::DenseTensor* grad_dense_tensor =
            static_cast<phi::DenseTensor*>(grad.impl().get());

        auto curr_data_type =
            paddle::framework::TransToProtoVarType(grad_dense_tensor->type());
        if (!paddle::framework::IsComplexType(curr_data_type)) continue;

        // Convert Complex GradOut to Real
        auto out = std::make_shared<phi::DenseTensor>();
        paddle::framework::TransComplexToReal(fwd_data_type, curr_data_type,
                                              *grad_dense_tensor, out.get());

        (*out_grads)[slot_id][rank_id].set_impl(out);
      }
    }
  }
}

429
}  // namespace egr