parallel_executor.cc 22.7 KB
Newer Older
Y
Yang Yang 已提交
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15
/* Copyright (c) 2016 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/framework/parallel_executor.h"
Y
Yu Yang 已提交
16 17
#include "ThreadPool.h"
#include "executor.h"
Y
Yu Yang 已提交
18
#include "lod_tensor.h"
Y
Yu Yang 已提交
19
#include "lod_tensor_array.h"
Y
Yu Yang 已提交
20
#include "op_registry.h"
Y
Yu Yang 已提交
21
#include "paddle/fluid/operators/math/concat.h"
Y
Yang Yang 已提交
22 23

namespace paddle {
Y
Yu Yang 已提交
24 25
namespace framework {

Y
Yu Yang 已提交
26 27 28 29 30 31
#ifdef PADDLE_WITH_CUDA

// FIXME: CHECK the return value of x;
#define NCCL_INVOKE(x) x
#endif

Y
Yu Yang 已提交
32 33
struct OpHandle;

Y
Yu Yang 已提交
34 35 36 37 38
struct VarHandleBase {
  virtual ~VarHandleBase() {}
  virtual std::string DebugString() const = 0;

  OpHandle *generated_op_;
Y
Yu Yang 已提交
39
  std::unordered_set<OpHandle *> pending_ops_;
Y
Yu Yang 已提交
40 41 42 43 44 45 46 47 48
};

struct VarHandle : public VarHandleBase {
  std::string DebugString() const override {
    std::stringstream ss;
    ss << name_ << ":" << place_;
    return ss.str();
  }

Y
Yu Yang 已提交
49 50
  // version field currently is not used, however, just store the version to
  // debug easily.
Y
Yu Yang 已提交
51 52 53
  size_t version_;
  std::string name_;
  platform::Place place_;
Y
Yu Yang 已提交
54
};
Y
Yu Yang 已提交
55

Y
Yu Yang 已提交
56
struct DependencyVarHandle : public VarHandleBase {
Y
Yu Yang 已提交
57
  std::string DebugString() const override { return "Dependency Variable"; }
Y
Yu Yang 已提交
58 59 60
};

struct OpHandle {
Y
Yu Yang 已提交
61 62 63 64 65
  std::vector<VarHandleBase *> inputs_;
  std::vector<VarHandleBase *> outputs_;
  std::unordered_map<platform::Place, platform::DeviceContext *,
                     platform::PlaceHash>
      dev_ctx_;
Y
Yu Yang 已提交
66 67 68 69 70

  std::string DebugString() {
    std::stringstream ss;
    ss << "(";
    for (auto *var : inputs_) {
Y
Yu Yang 已提交
71
      ss << var->DebugString() << ", ";
Y
Yu Yang 已提交
72 73 74
    }
    ss << ") --> (";
    for (auto *var : outputs_) {
Y
Yu Yang 已提交
75
      ss << var->DebugString() << ", ";
Y
Yu Yang 已提交
76 77 78 79 80 81
    }
    ss << ")\n";
    return ss.str();
  }

  virtual ~OpHandle() {}
Y
Yu Yang 已提交
82

Y
Yu Yang 已提交
83
  virtual void Run() { PADDLE_THROW("Not implemented"); }
Y
Yu Yang 已提交
84
  virtual void Wait(platform::DeviceContext *waited_dev) {}
Y
Yu Yang 已提交
85 86 87 88
};

struct ComputationOpHandle : public OpHandle {
  std::unique_ptr<OperatorBase> op_;
Y
Yu Yang 已提交
89 90
  Scope *scope_;
  platform::Place place_;
Y
Yu Yang 已提交
91

Y
Yu Yang 已提交
92 93
  explicit ComputationOpHandle(const OpDesc &op_desc, Scope *scope,
                               platform::Place place)
Y
Yu Yang 已提交
94
      : op_(framework::OpRegistry::CreateOp(op_desc)),
Y
Yu Yang 已提交
95
        scope_(scope),
Y
Yu Yang 已提交
96 97 98 99 100 101 102
        place_(place) {}

  void Run() override {
    // Wait other op if necessary
    auto *cur_ctx = dev_ctx_[place_];
    for (auto *in : inputs_) {
      if (in->generated_op_ && in->generated_op_->dev_ctx_[place_] != cur_ctx) {
Y
Yu Yang 已提交
103
        in->generated_op_->Wait(cur_ctx);
Y
Yu Yang 已提交
104 105 106 107 108
      }
    }

    op_->Run(*scope_, place_);
  }
Y
Yu Yang 已提交
109 110 111 112

  void Wait(platform::DeviceContext *waited_dev) override {
    this->dev_ctx_.at(place_)->Wait();
  }
Y
Yu Yang 已提交
113 114
};

Y
Yu Yang 已提交
115 116 117 118 119 120 121 122 123 124 125 126 127
struct ScaleLossGradOpHandle : public OpHandle {
  float coeff_;
  Scope *scope_;
  platform::Place place_;

  explicit ScaleLossGradOpHandle(size_t num_dev, Scope *scope,
                                 platform::Place place)
      : coeff_(static_cast<float>(1.0 / num_dev)),
        scope_(scope),
        place_(place) {}

  void Run() override {
    std::string var_name = static_cast<VarHandle *>(this->outputs_[0])->name_;
Y
Yu Yang 已提交
128

Y
Yu Yang 已提交
129 130 131 132 133 134 135 136 137 138 139 140 141 142
    float *tmp = scope_->FindVar(var_name)
                     ->GetMutable<framework::LoDTensor>()
                     ->mutable_data<float>(make_ddim({1}), place_);

    if (platform::is_cpu_place(place_)) {
      *tmp = coeff_;
    } else {
      memory::Copy(
          boost::get<platform::CUDAPlace>(place_), tmp, platform::CPUPlace(),
          &coeff_, sizeof(float),
          static_cast<platform::CUDADeviceContext *>(this->dev_ctx_[place_])
              ->stream());
    }
  }
Y
Yu Yang 已提交
143 144 145 146

  void Wait(platform::DeviceContext *waited_dev) override {
    this->dev_ctx_.at(place_)->Wait();
  }
Y
Yu Yang 已提交
147 148
};

Y
Yu Yang 已提交
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
struct FetchedData {
 public:
  std::vector<framework::LoDTensor> tensors_;

  explicit FetchedData(size_t num_fetched) { tensors_.resize(num_fetched); }
};

struct FetchOpHandle : public OpHandle {
  std::shared_ptr<FetchedData> data_;
  size_t offset_;
  std::vector<Scope *> *local_scopes_;
  std::vector<LoDTensor> tensors_;

  ~FetchOpHandle() {
    for (auto *input_var : inputs_) {
      input_var->pending_ops_.erase(this);
    }
    for (auto &pair : dev_ctx_) {
      pair.second->Wait();
    }

    // Lazily merge tensors. Will faster code.
    MergeTensors();
  }

  void Run() override {
Y
Debug  
Yu Yang 已提交
175 176 177 178
    for (auto *input : inputs_) {
      input->generated_op_->Wait(nullptr);
    }

Y
Yu Yang 已提交
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
    tensors_.resize(inputs_.size());
    auto *var = static_cast<VarHandle *>(inputs_[0]);
    auto &var_name = var->name_;
    platform::CPUPlace cpu;
    auto &scopes = *local_scopes_;

    for (size_t i = 0; i < scopes.size(); ++i) {
      auto &scope = scopes[i];
      auto &t = scope->FindVar(var_name)->Get<framework::LoDTensor>();
      if (platform::is_gpu_place(var->place_)) {
        TensorCopy(t, cpu, *dev_ctx_[t.place()], &tensors_[i]);
      } else {
        tensors_[i].ShareDataWith(t);
        tensors_[i].set_lod(t.lod());
      }
    }
  }

  void Wait(platform::DeviceContext *waited_dev) override {
    PADDLE_THROW("Nobody should wait FetchOp. Unexpceted Error");
  }

 private:
  void MergeTensors() const {
    std::vector<const LoDTensor *> tensors_ptr;
    for (auto &t : tensors_) {
      tensors_ptr.emplace_back(&t);
    }
    data_->tensors_[offset_].MergeLoDTensor(tensors_ptr, platform::CPUPlace());
  }
};

Y
Yu Yang 已提交
211 212
class ParallelExecutorPrivate {
 public:
Y
Yu Yang 已提交
213 214 215
  explicit ParallelExecutorPrivate(size_t num_threads = 12)
      : pool_(num_threads) {}

Y
Stash  
Yu Yang 已提交
216 217
  std::vector<platform::Place> places_;

Y
Yu Yang 已提交
218
  std::vector<Scope *> local_scopes_;
Y
Yu Yang 已提交
219
  Scope *global_scope_;
Y
Yu Yang 已提交
220

Y
Yu Yang 已提交
221 222 223 224 225 226 227 228 229 230 231 232 233 234 235
#ifdef PADDLE_WITH_CUDA
  struct NCCLContext {
    std::unique_ptr<platform::CUDADeviceContext> ctx_;
    ncclComm_t comm;

    explicit NCCLContext(int dev_id) {
      ctx_.reset(new platform::CUDADeviceContext(platform::CUDAPlace(dev_id)));
    }

    cudaStream_t stream() const { return ctx_->stream(); }

    int device_id() const {
      return boost::get<platform::CUDAPlace>(ctx_->GetPlace()).device;
    }

Y
Update  
Yu Yang 已提交
236 237
    static void InitNCCLContext(std::unordered_map<int, NCCLContext> &contexts,
                                const std::vector<platform::Place> &places) {
Y
Yu Yang 已提交
238 239 240 241 242
      std::vector<ncclComm_t> comms;
      std::vector<int> devs;
      comms.resize(contexts.size());
      devs.reserve(contexts.size());

Y
Update  
Yu Yang 已提交
243 244
      for (auto &p : places) {
        devs.push_back(boost::get<platform::CUDAPlace>(p).device);
Y
Yu Yang 已提交
245 246 247 248 249 250
      }

      NCCL_INVOKE(platform::dynload::ncclCommInitAll(
          &comms[0], static_cast<int>(contexts.size()), &devs[0]));

      int i = 0;
Y
Update  
Yu Yang 已提交
251 252
      for (auto &dev_id : devs) {
        contexts.at(dev_id).comm = comms[i++];
Y
Yu Yang 已提交
253 254 255 256
      }
    }
  };

Y
Update  
Yu Yang 已提交
257
  std::unordered_map<int, NCCLContext> communication_streams_;
Y
Yu Yang 已提交
258 259 260 261 262 263 264 265

  NCCLContext &GetNCCLCtx(platform::Place p) {
    int dev_id = boost::get<platform::CUDAPlace>(p).device;
    return communication_streams_.at(dev_id);
  }

#endif

Y
Yu Yang 已提交
266 267 268 269 270 271 272 273 274 275 276 277 278
  platform::DeviceContext *CommunicationDevCtx(const platform::Place &place) {
    if (platform::is_cpu_place(place) || local_scopes_.size() == 1) {
      return const_cast<platform::DeviceContext *>(
          platform::DeviceContextPool::Instance().Get(place));
    } else {
#ifdef PADDLE_WITH_CUDA
      return GetNCCLCtx(place).ctx_.get();
#else
      PADDLE_THROW("Not compiled with CUDA")
#endif
    }
  }

Y
Yu Yang 已提交
279 280 281 282 283 284
  platform::Place main_place_;

  std::unordered_map<platform::Place,
                     std::unordered_map<std::string, std::map<int, VarHandle>>,
                     platform::PlaceHash>
      vars_;
Y
Yu Yang 已提交
285 286
  std::unordered_set<std::unique_ptr<VarHandleBase>> dep_vars_;

Y
Yu Yang 已提交
287
  std::vector<std::unique_ptr<OpHandle>> ops_;
Y
Yu Yang 已提交
288

Y
Yu Yang 已提交
289
  // Use a simpler thread pool, might be faster.
Y
Yu Yang 已提交
290
  ThreadPool pool_;
Y
Yu Yang 已提交
291 292

  std::unique_ptr<platform::EnforceNotMet> exception_;
Y
Yu Yang 已提交
293 294 295 296
};

// TODO(yy): Move this function somewhere
ncclDataType_t ToNCCLDataType(std::type_index type) {
Y
Stash  
Yu Yang 已提交
297 298 299 300 301 302 303 304 305
  if (type == typeid(float)) {  // NOLINT
    return ncclFloat;
  } else if (type == typeid(double)) {  // NOLINT
    return ncclDouble;
  } else if (type == typeid(int)) {  // NOLINT
    return ncclInt;
  } else {
    PADDLE_THROW("Not supported");
  }
Y
Yu Yang 已提交
306 307
}

Y
Yu Yang 已提交
308 309 310 311 312 313 314 315 316 317 318 319 320 321 322
struct NCCLAllReduceOpHandle : public OpHandle {
  ParallelExecutorPrivate *member_;

  explicit NCCLAllReduceOpHandle(ParallelExecutorPrivate *member)
      : member_(member) {}

  void Run() override {
    if (this->inputs_.size() == 1) {
      return;  // No need to all reduce when GPU count = 1;
    } else {
      auto &var_name = static_cast<VarHandle *>(this->inputs_[0])->name_;

      int dtype = -1;
      size_t numel = 0;

Y
Update  
Yu Yang 已提交
323 324
      platform::dynload::ncclGroupStart();

Y
Yu Yang 已提交
325 326 327
      for (size_t i = 0; i < member_->local_scopes_.size(); ++i) {
        auto &p = member_->places_[i];
        auto *s = member_->local_scopes_[i];
Y
Yu Yang 已提交
328 329 330 331 332 333 334 335 336 337 338 339 340 341
        int dev_id = boost::get<platform::CUDAPlace>(p).device;

        auto &lod_tensor = s->FindVar(var_name)->Get<framework::LoDTensor>();
        void *buffer = const_cast<void *>(lod_tensor.data<void>());
        if (dtype == -1) {
          dtype = ToNCCLDataType(lod_tensor.type());
        }

        if (numel == 0) {
          numel = static_cast<size_t>(lod_tensor.numel());
        }

        auto &nccl_ctx = member_->communication_streams_.at(dev_id);

Y
Update  
Yu Yang 已提交
342 343 344
        platform::dynload::ncclAllReduce(
            buffer, buffer, numel, static_cast<ncclDataType_t>(dtype), ncclSum,
            nccl_ctx.comm, nccl_ctx.stream());
Y
Yu Yang 已提交
345 346
      }

Y
Update  
Yu Yang 已提交
347
      platform::dynload::ncclGroupEnd();
Y
Yu Yang 已提交
348 349
    }
  }
Y
Yu Yang 已提交
350 351

  void Wait(platform::DeviceContext *waited_dev) override {
Y
Debug  
Yu Yang 已提交
352 353 354
    for (auto &pair : member_->communication_streams_) {
      pair.second.ctx_->Wait();
    }
Y
Yu Yang 已提交
355
  }
Y
Yu Yang 已提交
356 357
};

Y
Yu Yang 已提交
358 359 360 361 362 363
ParallelExecutor::ParallelExecutor(
    const std::vector<platform::Place> &places,
    const std::unordered_set<std::string> &params,
    const ProgramDesc &startup_program, const ProgramDesc &main_program,
    const std::string &loss_var_name, Scope *scope)
    : member_(new ParallelExecutorPrivate()) {
Y
Stash  
Yu Yang 已提交
364
  member_->places_ = places;
Y
Yu Yang 已提交
365
  member_->global_scope_ = scope;
Y
Yu Yang 已提交
366 367 368 369
  // Step 1. RunStartupProgram and Bcast the params to devs.
  Executor exe(places[0]);
  exe.Run(startup_program, scope, 0);
  // Create local scopes
Y
Yu Yang 已提交
370 371
  for (size_t i = 0; i < member_->places_.size(); ++i) {
    member_->local_scopes_.push_back(&scope->NewScope());
Y
Yu Yang 已提交
372 373 374 375
  }
  member_->main_place_ = places[0];

  // Bcast Parameters to all GPUs
Y
Yu Yang 已提交
376
  BuildNCCLCommunicator();
Y
Yu Yang 已提交
377 378 379
  if (platform::is_gpu_place(member_->main_place_) &&
      member_->local_scopes_.size() != 1) {  // Is CUDA
    BCastParamsToGPUs(startup_program);
Y
Yu Yang 已提交
380 381 382 383 384 385
  }
  // Startup Program has been run. All local scopes has correct parameters.

  // Step 2. Convert main_program to SSA form and dependency graph. Also, insert
  // ncclOp
  ConstructDependencyGraph(params, main_program, loss_var_name);
Y
Yu Yang 已提交
386 387

  // Step 3. Create vars in each scope;
Y
Yu Yang 已提交
388
  for (auto *scope : member_->local_scopes_) {
Y
Yu Yang 已提交
389 390 391 392 393 394 395 396
    for (auto *var : main_program.Block(0).AllVars()) {
      if (scope->FindVar(var->Name()) != nullptr) {
        continue;
      }

      InitializeVariable(scope->Var(var->Name()), var->GetType());
    }
  }
Y
Yu Yang 已提交
397 398 399 400 401
}

void ParallelExecutor::ConstructDependencyGraph(
    const std::unordered_set<std::string> &params,
    const ProgramDesc &main_program, const std::string &loss_var_name) const {
Y
Yu Yang 已提交
402
  std::unordered_set<std::string> grads;
Y
Yu Yang 已提交
403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418
  for (auto &each_param : params) {
    grads.insert(each_param + "@GRAD");
  }

  bool is_forwarding = true;
  for (auto *op : main_program.Block(0).AllOps()) {
    bool change_forward = false;

    if (!is_forwarding) {
      // FIXME(yy): Do not hard code like this
      if (op->OutputArgumentNames().size() == 1 &&
          op->OutputArgumentNames()[0] == loss_var_name + "@GRAD") {
        continue;  // Drop fill 1. for backward coeff;
      }
    }

Y
Yu Yang 已提交
419 420 421 422 423
    for (size_t i = 0; i < member_->places_.size(); ++i) {
      auto &p = member_->places_[i];
      auto *s = member_->local_scopes_[i];

      member_->ops_.emplace_back(new ComputationOpHandle(*op, s, p));
Y
Yu Yang 已提交
424
      auto *op_handle = member_->ops_.back().get();
Y
Yu Yang 已提交
425 426
      op_handle->dev_ctx_[p] = const_cast<platform::DeviceContext *>(
          platform::DeviceContextPool::Instance().Get(p));
Y
Yu Yang 已提交
427 428 429 430

      auto var_names = op->InputArgumentNames();

      for (auto &each_var_name : var_names) {
Y
Yu Yang 已提交
431
        VarHandle *var = GetVarHandle(each_var_name, p);
Y
Yu Yang 已提交
432
        op_handle->inputs_.emplace_back(var);
Y
Yu Yang 已提交
433
        var->pending_ops_.emplace(op_handle);
Y
Yu Yang 已提交
434 435 436 437
      }
      var_names = op->OutputArgumentNames();

      for (auto &each_var_name : var_names) {
Y
Yu Yang 已提交
438
        GenerateVar(op_handle, each_var_name, p);
Y
Yu Yang 已提交
439 440 441 442 443
      }

      if (is_forwarding) {
        if (var_names.size() == 1 && var_names[0] == loss_var_name) {
          // Insert ScaleCost OpHandle
Y
Yu Yang 已提交
444
          member_->ops_.emplace_back(new ScaleLossGradOpHandle(
Y
Yu Yang 已提交
445
              this->member_->local_scopes_.size(), s, p));
Y
Yu Yang 已提交
446
          op_handle = member_->ops_.back().get();
Y
Yu Yang 已提交
447

Y
Yu Yang 已提交
448
          op_handle->dev_ctx_[p] = member_->CommunicationDevCtx(p);
Y
Yu Yang 已提交
449

Y
Yu Yang 已提交
450 451 452 453 454 455
          // FIXME: Currently ScaleLossGradOp only use device_count as scale
          // factor. So it does not depend on any other operators.
          // VarHandle *loss = GetVarHandle(loss_var_name, place);
          // loss->pending_ops_.emplace_back(op_handle);
          // op_handle->inputs_.emplace_back(loss);

Y
Yu Yang 已提交
456
          GenerateVar(op_handle, loss_var_name + "@GRAD", p);
Y
Yu Yang 已提交
457 458 459 460 461 462 463 464 465 466 467 468 469 470
          change_forward = true;
        }
      }
    }

    if (change_forward) {
      is_forwarding = false;
    }

    if (!is_forwarding) {
      auto var_names = op->OutputArgumentNames();
      for (auto &og : var_names) {
        if (grads.count(og) != 0) {  // is param grad
          // Insert NCCL AllReduce Op
Y
Yu Yang 已提交
471
          member_->ops_.emplace_back(new NCCLAllReduceOpHandle(member_));
Y
Yu Yang 已提交
472 473
          auto *op_handle = member_->ops_.back().get();

Y
Yu Yang 已提交
474 475 476
          for (size_t i = 0; i < member_->places_.size(); ++i) {
            auto &p = member_->places_[i];
            auto &vars = member_->vars_[p][og];
Y
Yu Yang 已提交
477 478 479 480 481 482

            if (vars.empty()) {  // This device has no data. continue.
              continue;
            }
            auto *prev_grad = &vars[vars.size() - 1];
            op_handle->inputs_.emplace_back(prev_grad);
Y
Yu Yang 已提交
483
            prev_grad->pending_ops_.emplace(op_handle);
Y
Yu Yang 已提交
484
            auto &var = vars[vars.size()];
Y
Yu Yang 已提交
485
            var.place_ = p;
Y
Yu Yang 已提交
486 487 488 489
            var.generated_op_ = op_handle;
            var.name_ = og;
            var.version_ = vars.size() - 1;
            op_handle->outputs_.emplace_back(&var);
Y
Yu Yang 已提交
490

Y
Yu Yang 已提交
491
            op_handle->dev_ctx_[p] = member_->CommunicationDevCtx(p);
Y
Yu Yang 已提交
492 493 494 495 496
          }
        }
      }
    }
  }
Y
Yu Yang 已提交
497

Y
Yu Yang 已提交
498 499 500
  /*
    Dependency graph has been constructed. However, there are still data
    harzaeds need to be handled.
Y
Yu Yang 已提交
501
   */
Y
Yu Yang 已提交
502 503
  PolishGraphToSupportDataHarzaeds();
}
Y
Yu Yang 已提交
504

Y
Yu Yang 已提交
505 506 507 508 509 510 511 512
/**
 * We only handle write after read(WAR), since it should not have a write
 * after write in program. If there are write after write operators, we need
 * prune them.
 *
 * https://en.wikipedia.org/wiki/Hazard_(computer_architecture)#Write_after_read_(WAR)
 */
void ParallelExecutor::PolishGraphToSupportDataHarzaeds() const {
Y
Yu Yang 已提交
513 514 515 516 517 518 519 520 521 522 523
  for (auto &place_pair : member_->vars_) {
    for (auto &name_pair : place_pair.second) {
      if (name_pair.second.size() <= 1) {
        return;
      }
      auto it_new = name_pair.second.rbegin();
      auto it_old = name_pair.second.rbegin();
      ++it_old;
      for (; it_old != name_pair.second.rend(); it_new = it_old, ++it_old) {
        auto *write_op = it_new->second.generated_op_;
        auto &read_ops = it_old->second.pending_ops_;
Y
Yu Yang 已提交
524 525 526 527 528 529
        auto *ex_write_op = it_old->second.generated_op_;

        if (ex_write_op == nullptr) {  // Nobody write this var.
          continue;
        }

Y
Yu Yang 已提交
530 531
        for (auto *read_op : read_ops) {
          // Manually add a dependency var from read_op to write_op;
Y
Yu Yang 已提交
532 533 534 535
          if (read_op == write_op) {
            // Read Write is the same op.
            continue;
          }
Y
Yu Yang 已提交
536 537

          auto *dep_var = new DependencyVarHandle();
Y
Yu Yang 已提交
538

Y
Yu Yang 已提交
539 540 541
          dep_var->generated_op_ = read_op;
          read_op->outputs_.emplace_back(dep_var);

Y
Yu Yang 已提交
542
          dep_var->pending_ops_.emplace(write_op);
Y
Yu Yang 已提交
543 544 545 546 547 548
          write_op->inputs_.emplace_back(dep_var);
          member_->dep_vars_.emplace(dep_var);
        }
      }
    }
  }
Y
Yu Yang 已提交
549 550 551 552 553 554 555 556 557 558 559 560 561 562 563 564 565 566 567 568 569 570 571 572 573 574 575 576 577 578 579 580 581 582 583
}

void ParallelExecutor::GenerateVar(OpHandle *op_handle,
                                   const std::string &each_var_name,
                                   const platform::Place &place) const {
  auto &vars = member_->vars_[place][each_var_name];
  size_t version = vars.size();
  auto &var = vars[version];
  var.version_ = version;
  var.generated_op_ = op_handle;
  var.name_ = each_var_name;
  var.place_ = place;
  op_handle->outputs_.emplace_back(&var);
}

VarHandle *ParallelExecutor::GetVarHandle(const std::string &each_var_name,
                                          const platform::Place &place) const {
  auto &var_holders = member_->vars_[place];
  auto &var_holder = var_holders[each_var_name];
  VarHandle *var = nullptr;
  if (var_holder.empty()) {
    auto &init_var = var_holder[0];
    init_var.place_ = place;
    init_var.name_ = each_var_name;
    init_var.generated_op_ = nullptr;
    init_var.version_ = 0;
    var = &init_var;
  } else {
    var = &var_holder.rbegin()->second;
  }
  return var;
}

void ParallelExecutor::BCastParamsToGPUs(
    const ProgramDesc &startup_program) const {
Y
Yu Yang 已提交
584
#ifdef PADDLE_WITH_CUDA
Y
Yu Yang 已提交
585
  auto *main_scope = member_->local_scopes_[0];
Y
Yu Yang 已提交
586

Y
Yu Yang 已提交
587 588 589 590 591 592 593 594
  for (auto *var_desc : startup_program.Block(0).AllVars()) {
    if (var_desc->GetType() == proto::VarType::LOD_TENSOR) {
      auto &main_tensor =
          main_scope->FindVar(var_desc->Name())->Get<LoDTensor>();
      ncclDataType_t data_type = ToNCCLDataType(main_tensor.type());
      auto &dims = main_tensor.dims();
      size_t numel = main_tensor.numel();

Y
Stash  
Yu Yang 已提交
595
      platform::dynload::ncclGroupStart();
Y
Yu Yang 已提交
596

Y
Update  
Yu Yang 已提交
597 598 599 600 601 602
      for (size_t i = 0; i < member_->places_.size(); ++i) {
        auto place = member_->places_[i];
        void *buffer;
        if (i == 0) {
          buffer = const_cast<void *>(main_tensor.data<void>());
        } else {
Y
Yu Yang 已提交
603
          auto local_scope = member_->local_scopes_[i];
Y
Update  
Yu Yang 已提交
604 605 606 607 608
          auto *t = local_scope->Var(var_desc->Name())->GetMutable<LoDTensor>();
          t->Resize(dims);
          buffer = t->mutable_data(place, main_tensor.type());
        }

Y
Stash  
Yu Yang 已提交
609
        auto &nccl_ctx = member_->GetNCCLCtx(place);
Y
Update  
Yu Yang 已提交
610
        platform::dynload::ncclBcast(buffer, numel, data_type, 0, nccl_ctx.comm,
Y
Stash  
Yu Yang 已提交
611
                                     nccl_ctx.stream());
Y
Yu Yang 已提交
612
      }
Y
Stash  
Yu Yang 已提交
613 614 615
      platform::dynload::ncclGroupEnd();
    }
  }
Y
Yu Yang 已提交
616 617 618 619
#else
  PADDLE_THROW("Not compiled with CUDA");
#endif
}
Y
Yu Yang 已提交
620

Y
Yu Yang 已提交
621 622
void ParallelExecutor::BuildNCCLCommunicator() const {
#ifdef PADDLE_WITH_CUDA
Y
Yu Yang 已提交
623
  for (auto &place : member_->places_) {
Y
Yu Yang 已提交
624
    int dev_id = boost::get<platform::CUDAPlace>(place).device;
Y
Yu Yang 已提交
625

Y
Yu Yang 已提交
626 627
    member_->communication_streams_.emplace(
        dev_id, ParallelExecutorPrivate::NCCLContext(dev_id));
Y
Yu Yang 已提交
628
  }
Y
Yu Yang 已提交
629 630

  ParallelExecutorPrivate::NCCLContext::InitNCCLContext(
Y
Update  
Yu Yang 已提交
631
      member_->communication_streams_, member_->places_);
Y
Yu Yang 已提交
632
#endif
Y
Yu Yang 已提交
633 634
}

Y
Yu Yang 已提交
635 636 637
void ParallelExecutor::Run(const std::vector<std::string> &fetch_tensors,
                           const std::string &fetched_var_name) {
  auto fetched_data = std::make_shared<FetchedData>(fetch_tensors.size());
Y
Yu Yang 已提交
638
  // Version --> VarHandle
Y
Yu Yang 已提交
639
  member_->exception_.reset();
Y
Yu Yang 已提交
640
  std::unordered_map<VarHandleBase *, std::atomic<bool>> pending_vars;
Y
Yu Yang 已提交
641 642 643 644 645
  std::unordered_map<OpHandle *, size_t> pending_ops;

  for (auto &place_pair : member_->vars_) {
    for (auto &name_pair : place_pair.second) {
      for (auto &version_pair : name_pair.second) {
Y
Yu Yang 已提交
646 647 648
        pending_vars[&version_pair.second].store(
            version_pair.second.generated_op_ == nullptr,
            std::memory_order_relaxed);
Y
Yu Yang 已提交
649 650 651 652
      }
    }
  }

Y
Yu Yang 已提交
653
  for (auto &var : member_->dep_vars_) {
Y
Yu Yang 已提交
654 655
    pending_vars[var.get()].store(var->generated_op_ == nullptr,
                                  std::memory_order_relaxed);
Y
Yu Yang 已提交
656 657
  }

Y
Yu Yang 已提交
658 659
  std::vector<OpHandle *> to_run;

Y
Yu Yang 已提交
660
  for (auto &op : member_->ops_) {
Y
Yu Yang 已提交
661 662 663 664 665 666 667
    if (op->inputs_.empty()) {  // Special case, Op has no input.
      to_run.emplace_back(op.get());
    } else {
      pending_ops.insert({op.get(), op->inputs_.size()});
    }
  }

Y
Yu Yang 已提交
668 669 670 671 672 673 674 675 676 677 678 679 680 681 682 683 684 685 686 687 688 689 690 691 692 693 694 695 696 697 698 699
  std::unordered_map<std::string, std::vector<VarHandleBase *>> fetched_vars;

  for (auto &fetch_var_name : fetch_tensors) {
    for (auto &pair : member_->vars_) {
      auto it = pair.second.find(fetch_var_name);
      if (it != pair.second.end()) {
        fetched_vars[fetch_var_name].push_back(&it->second.rbegin()->second);
      }
    }
  }

  std::vector<FetchOpHandle> fetch_ops;

  for (size_t i = 0; i < fetch_tensors.size(); ++i) {
    auto &var_name = fetch_tensors[i];
    auto &vars = fetched_vars[var_name];
    fetch_ops.emplace_back();
    FetchOpHandle *op = &fetch_ops.back();
    op->data_ = fetched_data;
    op->offset_ = i;
    op->local_scopes_ = &member_->local_scopes_;
    for (auto &p : member_->places_) {
      op->dev_ctx_[p] = this->member_->GetNCCLCtx(p).ctx_.get();
    }

    for (auto *var : vars) {
      var->pending_ops_.emplace(op);
      op->inputs_.emplace_back(var);
    }
    pending_ops.insert({op, op->inputs_.size()});
  }

Y
Yu Yang 已提交
700 701
  for (auto *op : to_run) {
    RunOp(pending_vars, op);
Y
Yu Yang 已提交
702 703
  }

Y
Yu Yang 已提交
704
  while (!pending_ops.empty()) {
Y
Yu Yang 已提交
705
    VarHandleBase *ready_var = nullptr;
Y
Yu Yang 已提交
706
    for (auto &pair : pending_vars) {
Y
Yu Yang 已提交
707
      if (pair.second.load(std::memory_order_acquire)) {
Y
Yu Yang 已提交
708
        ready_var = pair.first;
Y
Yu Yang 已提交
709 710
      }
    }
Y
Yu Yang 已提交
711
    if (ready_var == nullptr) {
Y
Yu Yang 已提交
712 713 714 715 716 717
      // FIXME use conditional var instead of busy wait.

      if (member_->exception_) {
        throw * member_->exception_;
      }

Y
Yu Yang 已提交
718
      continue;
Y
Yu Yang 已提交
719
    }
Y
Yu Yang 已提交
720
    pending_vars.erase(ready_var);
Y
Yu Yang 已提交
721
    to_run.clear();
Y
Yu Yang 已提交
722 723 724 725 726
    for (auto *op : ready_var->pending_ops_) {
      auto &deps = pending_ops[op];
      --deps;
      if (deps == 0) {
        to_run.emplace_back(op);
Y
Yu Yang 已提交
727 728 729 730
      }
    }
    for (auto *op : to_run) {
      pending_ops.erase(op);
Y
Yu Yang 已提交
731
      RunOp(pending_vars, op);
Y
Yu Yang 已提交
732 733
    }
  }
Y
Yu Yang 已提交
734 735 736
  fetch_ops.clear();
  *member_->global_scope_->Var(fetched_var_name)->GetMutable<LoDTensorArray>() =
      fetched_data->tensors_;
Y
Yu Yang 已提交
737
}
Y
Yu Yang 已提交
738 739

void ParallelExecutor::RunOp(
Y
Yu Yang 已提交
740
    std::unordered_map<VarHandleBase *, std::atomic<bool>> &pending_vars,
Y
Yu Yang 已提交
741
    OpHandle *op) const {
Y
Debug  
Yu Yang 已提交
742 743
  std::vector<std::atomic<bool> *> *ready_buffer =
      new std::vector<std::atomic<bool> *>();
Y
Yu Yang 已提交
744
  for (auto *var : op->outputs_) {
Y
Debug  
Yu Yang 已提交
745
    ready_buffer->emplace_back(&pending_vars[var]);
Y
Yu Yang 已提交
746 747 748 749
  }

  auto op_run = [ready_buffer, op, this] {
    try {
Y
Yu Yang 已提交
750
      VLOG(10) << op->DebugString() << " " << this;
Y
Yu Yang 已提交
751
      op->Run();
Y
Yu Yang 已提交
752
      VLOG(10) << "Done " << this;
Y
Debug  
Yu Yang 已提交
753
      for (auto *ready : *ready_buffer) {
Y
Yu Yang 已提交
754
        ready->store(true, std::memory_order_release);
Y
Yu Yang 已提交
755
      }
Y
Debug  
Yu Yang 已提交
756
      delete ready_buffer;
Y
Yu Yang 已提交
757 758 759 760 761 762
    } catch (platform::EnforceNotMet ex) {
      member_->exception_.reset(new platform::EnforceNotMet(ex));
    } catch (...) {
      LOG(FATAL) << "Unknown exception catched";
    }
  };
Y
Debug  
Yu Yang 已提交
763
  VLOG(3) << "Enqueue";
Y
Yu Yang 已提交
764
  member_->pool_.enqueue(op_run);
Y
Debug  
Yu Yang 已提交
765
  VLOG(3) << "Done";
Y
Yu Yang 已提交
766
}
Y
Yu Yang 已提交
767
}  // namespace framework
Y
Yang Yang 已提交
768
}  // namespace paddle