multi_devices_graph_pass.cc 34.1 KB
Newer Older
Y
Yu Yang 已提交
1 2 3 4 5 6 7 8 9 10 11 12 13
//   Copyright (c) 2018 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.
C
chengduoZH 已提交
14
#include <algorithm>
Y
Yancey1989 已提交
15
#include <fstream>
C
chengduoZH 已提交
16
#include <string>
C
chengduoZH 已提交
17
#include <utility>
C
chengduoZH 已提交
18 19
#include <vector>

20
#include "paddle/fluid/framework/details/all_reduce_op_handle.h"
C
chengduoZH 已提交
21
#include "paddle/fluid/framework/details/broadcast_op_handle.h"
Y
Yu Yang 已提交
22
#include "paddle/fluid/framework/details/computation_op_handle.h"
23
#include "paddle/fluid/framework/details/data_balance_op_handle.h"
24
#include "paddle/fluid/framework/details/fused_broadcast_op_handle.h"
X
Xin Pan 已提交
25
#include "paddle/fluid/framework/details/multi_devices_graph_pass.h"
C
chengduoZH 已提交
26
#include "paddle/fluid/framework/details/reduce_op_handle.h"
Y
Yancey1989 已提交
27
#include "paddle/fluid/framework/details/rpc_op_handle.h"
Y
Yu Yang 已提交
28
#include "paddle/fluid/framework/details/scale_loss_grad_op_handle.h"
X
better  
Xin Pan 已提交
29
#include "paddle/fluid/framework/ir/graph_helper.h"
X
Xin Pan 已提交
30
#include "paddle/fluid/framework/ir/node.h"
Y
Fix bug  
yuyang18 已提交
31
#include "paddle/fluid/framework/op_info.h"
Y
Yu Yang 已提交
32
#include "paddle/fluid/framework/scope.h"
Y
Yu Yang 已提交
33

Y
Yu Yang 已提交
34 35 36
namespace paddle {
namespace framework {
namespace details {
X
Xin Pan 已提交
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 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122
namespace {
void PolishGraphToSupportDataHazards(ir::Graph *graph) {
  for (auto &var_map : graph->Get<GraphVars>(kGraphVars)) {
    for (auto &name_pair : var_map) {
      if (name_pair.second.size() <= 1) {
        continue;
      }
      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) {
        OpHandleBase *write_op = (*it_new)->GeneratedOp();
        const auto &read_ops = (*it_old)->PendingOps();

        for (auto *read_op : read_ops) {
          // Manually add a dependency var from read_op to write_op;
          if (read_op == write_op) {
            // Read Write is the same op.
            continue;
          }
          bool has_dep = false;
          for (auto *r_out : read_op->Outputs()) {
            for (auto *w_in : write_op->Inputs()) {
              if (r_out->Node() == w_in->Node()) {
                has_dep = true;
                break;
              }
            }
          }
          if (has_dep) continue;

          auto *dep_var = new DummyVarHandle(graph->CreateControlDepVar());
          read_op->AddOutput(dep_var);
          write_op->AddInput(dep_var);
          graph->Get<GraphDepVars>(kGraphDepVars).emplace(dep_var);
        }
      }
    }
  }
}

VarHandle *CreateOrGetLatestVarHandle(ir::Graph *graph, ir::Node *node,
                                      const platform::Place &place,
                                      size_t place_offset) {
  auto &var_holders = graph->Get<GraphVars>(kGraphVars)[place_offset];
  auto &var_holder = var_holders[node->Name()];
  VarHandle *var = nullptr;
  if (var_holder.empty()) {
    if (node->Var()) {
      var = new VarHandle(graph->CreateVarNode(node->Var()), 0, place_offset,
                          node->Name(), place);
    } else {
      var = new VarHandle(
          graph->CreateEmptyNode(node->Name(), ir::Node::Type::kVariable), 0,
          place_offset, node->Name(), place);
    }
    var_holder.emplace_back(var);
  } else {
    var = var_holder.rbegin()->get();
  }
  return var;
}

void CreateOpOutput(ir::Graph *graph, OpHandleBase *op_handle,
                    ir::Node *new_node, const platform::Place &place,
                    size_t place_offset) {
  auto &vars =
      graph->Get<GraphVars>(kGraphVars)[place_offset][new_node->Name()];
  size_t version = vars.size();
  auto var =
      new VarHandle(new_node, version, place_offset, new_node->Name(), place);
  vars.emplace_back(var);
  op_handle->AddOutput(var);
}

void AddOutputToLeafOps(ir::Graph *graph) {
  for (auto &op : graph->Get<GraphOps>(kGraphOps)) {
    if (!op->Outputs().empty()) {
      continue;
    }
    auto *dummy_leaf = new DummyVarHandle(graph->CreateControlDepVar());
    graph->Get<GraphDepVars>(kGraphDepVars).emplace(dummy_leaf);
    op->AddOutput(dummy_leaf);
  }
}
}  // namespace
Y
Yu Yang 已提交
123

X
Xin Pan 已提交
124 125 126 127 128 129
static const char kLossVarName[] = "loss_var_name";
static const char kPlaces[] = "places";
static const char kParams[] = "params";
static const char kLocalScopes[] = "local_scopes";
static const char kStrategy[] = "strategy";

X
Xin Pan 已提交
130
void MultiDevSSAGraphBuilder::Init() const {
X
clean  
Xin Pan 已提交
131 132 133
  all_vars_.clear();
  balance_vars_.clear();

X
Xin Pan 已提交
134 135 136 137
  loss_var_name_ = Get<const std::string>(kLossVarName);
  places_ = Get<const std::vector<platform::Place>>(kPlaces);
  local_scopes_ = Get<const std::vector<Scope *>>(kLocalScopes);
  strategy_ = Get<const BuildStrategy>(kStrategy);
Y
Yu Yang 已提交
138
#ifdef PADDLE_WITH_CUDA
X
Xin Pan 已提交
139
  nccl_ctxs_ = &Get<platform::NCCLContextMap>("nccl_ctxs");
Y
Yu Yang 已提交
140
#endif
X
Xin Pan 已提交
141

X
Xin Pan 已提交
142
  for (auto &p : Get<const std::unordered_set<std::string>>(kParams)) {
Y
Yu Yang 已提交
143 144
    grad_names_.insert(GradVarName(p));
  }
Y
Yancey1989 已提交
145
  balance_vars_.resize(places_.size(), 0);
Y
yuyang18 已提交
146 147 148 149 150
  if (strategy_.enable_data_balance_ && places_.size() == 1) {
    LOG(WARNING) << "It is no need to enable data balance when there is only "
                    "one place. enable_data_balance is set to False.";
    strategy_.enable_data_balance_ = false;
  }
Y
Yu Yang 已提交
151 152
}

X
Xin Pan 已提交
153 154
void MultiDevSSAGraphBuilder::CreateOpHandleIOs(ir::Graph *result,
                                                ir::Node *node,
Y
Yu Yang 已提交
155 156
                                                size_t place_id) const {
  auto p = places_[place_id];
X
Xin Pan 已提交
157
  auto *op_handle = result->Get<GraphOps>(kGraphOps).back().get();
X
Xin Pan 已提交
158 159
  op_handle->SetDeviceContext(p,
                              platform::DeviceContextPool::Instance().Get(p));
T
wip  
typhoonzero 已提交
160

161 162
  for (ir::Node *input : node->inputs) {
    VarHandle *var = CreateOrGetLatestVarHandle(result, input, p, place_id);
T
wip  
typhoonzero 已提交
163 164 165
    op_handle->AddInput(var);
  }

166
  for (ir::Node *output : node->outputs) {
X
polish  
Xin Pan 已提交
167 168 169 170 171 172 173 174
    ir::Node *new_node = nullptr;
    if (output->Var()) {
      new_node = result->CreateVarNode(output->Var());
    } else {
      new_node =
          result->CreateEmptyNode(output->Name(), ir::Node::Type::kVariable);
    }
    CreateOpOutput(result, op_handle, new_node, p, place_id);
T
wip  
typhoonzero 已提交
175 176
  }
}
Y
fix pe  
Yancey1989 已提交
177 178

std::vector<std::string> MultiDevSSAGraphBuilder::FindDistTrainSendVars(
X
Xin Pan 已提交
179
    const std::vector<ir::Node *> &nodes) const {
Y
fix pe  
Yancey1989 已提交
180
  std::vector<std::string> send_vars;
Y
Yancey1989 已提交
181 182
  // since parameters are all in block 0,
  // it's enough to only scan send ops in block 0
183 184
  for (auto &node : nodes) {
    OpDesc *op = node->Op();
Y
Yancey1989 已提交
185 186
    // TODO(Yancey1989): use a graceful method to find send op,
    // instead of the the hard code string
187
    if (op->Type() == "send") {
Y
fix pe  
Yancey1989 已提交
188 189 190 191 192 193 194 195 196 197
      auto op_vars = op->InputArgumentNames();
      send_vars.reserve(send_vars.size() +
                        std::distance(op_vars.begin(), op_vars.end()));
      send_vars.insert(send_vars.end(), op_vars.begin(), op_vars.end());
    }
  }
  return send_vars;
}

std::vector<std::string> MultiDevSSAGraphBuilder::FindDistTrainRecvVars(
X
Xin Pan 已提交
198
    const std::vector<ir::Node *> &nodes) const {
Y
fix pe  
Yancey1989 已提交
199
  std::vector<std::string> recv_vars;
200 201
  for (auto &node : nodes) {
    OpDesc *op = node->Op();
Y
Yancey1989 已提交
202 203 204
    // TODO(Yancey1989): use a graceful method to find recv op,
    // instead of the hard code string
    if (op->Type() == "recv") {
Y
fix pe  
Yancey1989 已提交
205 206 207 208 209 210 211 212 213
      auto op_vars = op->OutputArgumentNames();
      recv_vars.reserve(recv_vars.size() +
                        std::distance(op_vars.begin(), op_vars.end()));
      recv_vars.insert(recv_vars.end(), op_vars.begin(), op_vars.end());
    }
  }
  return recv_vars;
}

Y
Yancey1989 已提交
214 215 216 217
size_t MultiDevSSAGraphBuilder::GetAppropriateDeviceID(
    const std::vector<std::string> &var_names) const {
  int64_t numel_sum = 0;
  for (auto var_name : var_names) {
X
Xin Pan 已提交
218
    if (all_vars_.find(var_name) == all_vars_.end()) continue;
Y
Yancey1989 已提交
219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234
    auto var_desc = all_vars_.at(var_name);
    PADDLE_ENFORCE_NOT_NULL(var_desc);
    auto dim = framework::make_ddim(var_desc->GetShape());
    int64_t numel = framework::product(dim);
    PADDLE_ENFORCE_GT(numel, 0);
    numel_sum += numel;
  }

  auto smallest =
      std::min_element(std::begin(balance_vars_), std::end(balance_vars_));
  size_t dev_id =
      static_cast<size_t>(std::distance(std::begin(balance_vars_), smallest));
  balance_vars_[dev_id] += numel_sum;
  return dev_id;
}

X
better  
Xin Pan 已提交
235 236 237 238 239
// Topology sort the graph nodes from inputs to outputs.
// Since SSAGraphBuilder depends on forward/backward nodes to assign devices
// to parameter/gradients before optimizer ops, topo sort is insufficient. (
// some optimizer ops might not depend on any nodes), we manually move all
// optimizer nodes after last backward nodes.
X
Xin Pan 已提交
240 241 242
// However, the assumption by SSAGraphBuilder should be relaxed in the future.
std::vector<ir::Node *> SortOpsAndDelayOptimizeOp(const ir::Graph &graph) {
  std::vector<ir::Node *> ret = ir::TopologySortOperations(graph);
X
better  
Xin Pan 已提交
243 244 245 246 247
  size_t last_backward = 0;
  for (size_t i = 0; i < ret.size(); ++i) {
    if (boost::get<int>(
            ret[i]->Op()->GetAttr(OpProtoAndCheckerMaker::OpRoleAttrName())) ==
        static_cast<int>(OpRole::kBackward)) {
X
Xin Pan 已提交
248
      last_backward = i;
X
better  
Xin Pan 已提交
249 250 251
    }
  }

X
Xin Pan 已提交
252 253 254 255
  std::vector<ir::Node *> optimize_ops;
  std::vector<ir::Node *> sorted_ret;
  for (size_t i = 0; i < ret.size(); ++i) {
    if (i < last_backward) {
X
Xin Pan 已提交
256 257 258
      if (static_cast<bool>(boost::get<int>(ret[i]->Op()->GetAttr(
                                OpProtoAndCheckerMaker::OpRoleAttrName())) &
                            static_cast<int>(OpRole::kOptimize))) {
X
Xin Pan 已提交
259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276
        optimize_ops.push_back(ret[i]);
      } else {
        sorted_ret.push_back(ret[i]);
      }
    } else if (i == last_backward) {
      sorted_ret.push_back(ret[i]);
      // Verify that no operations before optimize ops depends on optimize ops.
      std::unordered_set<ir::Node *> optimize_set(optimize_ops.begin(),
                                                  optimize_ops.end());
      for (ir::Node *n : sorted_ret) {
        for (ir::Node *in : n->inputs) {
          for (ir::Node *pre_n : in->inputs) {
            PADDLE_ENFORCE(optimize_set.find(pre_n) == optimize_set.end(),
                           "optimize operations cannot be depended by forward "
                           "or backward node %s -> %s",
                           pre_n->Name(), n->Name());
          }
        }
X
Xin Pan 已提交
277
      }
X
Xin Pan 已提交
278 279 280 281
      sorted_ret.insert(sorted_ret.end(), optimize_ops.begin(),
                        optimize_ops.end());
    } else {
      sorted_ret.push_back(ret[i]);
X
Xin Pan 已提交
282 283
    }
  }
X
better  
Xin Pan 已提交
284 285 286
  return sorted_ret;
}

X
Xin Pan 已提交
287
std::unique_ptr<ir::Graph> MultiDevSSAGraphBuilder::ApplyImpl(
X
Xin Pan 已提交
288
    std::unique_ptr<ir::Graph> graph) const {
X
Xin Pan 已提交
289
  Init();
X
Xin Pan 已提交
290
  // Give the topology sort order and rebuild the graph structure.
X
better  
Xin Pan 已提交
291
  std::vector<ir::Node *> sorted_ops = SortOpsAndDelayOptimizeOp(*graph);
X
Xin Pan 已提交
292 293
  auto nodes = graph->ReleaseNodes();
  ir::Graph &result = *graph;
294 295

  for (auto &node : nodes) {
X
Xin Pan 已提交
296
    if (node->IsVar() && node->Var()) {
X
Xin Pan 已提交
297
      all_vars_.emplace(node->Name(), node->Var());
298
    }
C
fix ci  
chengduoZH 已提交
299
  }
C
chengduoZH 已提交
300
  std::unordered_set<std::string> og_has_been_broadcast;
Y
Yu Yang 已提交
301 302

  // We cannot invoke resize. It is a bug of GCC 4.8
X
Xin Pan 已提交
303 304 305 306
  result.Set(kGraphVars, new GraphVars(places_.size()));
  result.Set(kGraphDepVars, new GraphDepVars);
  result.Set(kGraphOps, new GraphOps);
  result.Set(kShardedVarDevice, new ShardedVarDevice);
307

Y
fix pe  
Yancey1989 已提交
308
  // find send/recv vars so that we can place the distributed training
309
  // related op in the place 0
X
Xin Pan 已提交
310 311
  auto send_vars = FindDistTrainSendVars(sorted_ops);
  auto recv_vars = FindDistTrainRecvVars(sorted_ops);
T
typhoonzero 已提交
312

C
chengduoZH 已提交
313 314 315
  std::vector<std::unordered_set<std::string>> bcast_var_name_set;
  bcast_var_name_set.resize(places_.size());

C
chengduoZH 已提交
316
  size_t cur_device_id = 0;
Y
Yu Yang 已提交
317
  bool is_forwarding = true;
Y
Yancey1989 已提交
318
  bool is_dist_train = false;
319

X
better  
Xin Pan 已提交
320
  for (ir::Node *node : sorted_ops) {
Y
Yancey1989 已提交
321
    if (boost::get<int>(
322
            node->Op()->GetAttr(OpProtoAndCheckerMaker::OpRoleAttrName())) ==
Y
Yancey1989 已提交
323
        static_cast<int>(OpRole::kRPC)) {
Y
Yancey1989 已提交
324 325 326 327 328 329 330 331 332 333 334 335 336
      int op_dev_id = CreateRPCOp(&result, node);
      PADDLE_ENFORCE(op_dev_id != -1,
                     "Can not schedule the RPC operator to the right place.");
      if (node->Op()->Type() == "recv") {
        auto recv_vars_attr =
            boost::get<std::vector<std::string>>(node->Op()->GetNullableAttr(
                OpProtoAndCheckerMaker::OpRoleVarAttrName()));
        PADDLE_ENFORCE(recv_vars_attr.size() == 2UL);  // [parameter, gradient]
        if (recv_vars_attr[0].find(".block") == std::string::npos) {
          bcast_var_name_set[op_dev_id].emplace(recv_vars_attr[0]);
        }
      }
      is_dist_train = true;
337 338 339
    } else if (boost::get<int>(node->Op()->GetAttr(
                   OpProtoAndCheckerMaker::OpRoleAttrName())) ==
               static_cast<int>(OpRole::kDist)) {
Y
Yancey1989 已提交
340 341 342 343 344
      int op_dev_id = CreateDistTrainOp(&result, node);
      if (node->Op()->Type() == "concat") {
        auto origin_param_name = node->Op()->OutputArgumentNames()[0];
        bcast_var_name_set[op_dev_id].emplace(origin_param_name);
      }
X
Xin Pan 已提交
345
    } else if (IsScaleLossOp(node)) {
Y
Yu Yang 已提交
346
      // user can customize loss@grad if not use_default_grad_scale_
Y
yuyang18 已提交
347 348
      if (strategy_.gradient_scale_ !=
          BuildStrategy::GradientScaleStrategy::kCustomized) {
X
Xin Pan 已提交
349
        // TODO(paddle-dev): Why is there no input for this op_handle?
350
        auto loss_grad_name = node->Op()->OutputArgumentNames()[0];
351
        CreateScaleLossGradOp(&result, loss_grad_name, node->outputs[0]);
Y
Yu Yang 已提交
352
      }
353 354 355 356
      // This assumes the backward generating code will ensure IsScaleLossOp
      // is true only for the op that scale the final scalar loss.
      // It also assumes backward op will always follow the forward op in
      // the block.
Y
Yu Yang 已提交
357
      is_forwarding = false;
Y
Yu Yang 已提交
358
    } else {
X
Xin Pan 已提交
359
      int op_dev_id = GetOpDeviceID(result, node);
C
chengduo 已提交
360
      if (op_dev_id != -1) {  // This op only runs on one specific device.
X
Xin Pan 已提交
361
        CreateComputationalOp(&result, node, op_dev_id);
362
        for (ir::Node *n : node->outputs) {
X
Xin Pan 已提交
363
          graph->Get<ShardedVarDevice>(kShardedVarDevice)
X
Xin Pan 已提交
364
              .emplace(n->Name(), op_dev_id);
C
chengduoZH 已提交
365
        }
C
chengduo 已提交
366 367 368
      } else {
        // This op runs on all devices, and its output may have parameter's
        // gradients.
X
Xin Pan 已提交
369
        // TODO(paddle-dev): Why is so special about "read" op?
370 371
        if (node->Op()->Type() == "read" && strategy_.enable_data_balance_) {
          node->Op()->SetAttr("throw_eof_exp", false);
X
Xin Pan 已提交
372
          CreateComputationalOps(&result, node, places_.size());
373
          const auto &data_var_names = node->Op()->Output("Out");
374
          InsertDataBalanceOp(&result, data_var_names);
F
fengjiayi 已提交
375
        } else {
X
Xin Pan 已提交
376
          CreateComputationalOps(&result, node, places_.size());
377 378
        }

C
chengduo 已提交
379 380 381
        if (!is_forwarding && places_.size() > 1) {
          // Currently, we assume that once gradient is generated, it can be
          // broadcast, and each gradient is only broadcast once.
382
          if (static_cast<bool>(boost::get<int>(node->Op()->GetAttr(
C
chengduo 已提交
383 384 385
                                    OpProtoAndCheckerMaker::OpRoleAttrName())) &
                                static_cast<int>(OpRole::kBackward))) {
            try {
386 387
              auto backward_vars = boost::get<std::vector<std::string>>(
                  node->Op()->GetNullableAttr(
C
chengduo 已提交
388
                      OpProtoAndCheckerMaker::OpRoleVarAttrName()));
Y
yuyang18 已提交
389

C
chengduo 已提交
390
              PADDLE_ENFORCE_EQ(backward_vars.size() % 2, 0);
Y
yuyang18 已提交
391

C
chengduo 已提交
392 393 394 395
              for (size_t i = 0; i < backward_vars.size(); i += 2) {
                auto &p_name = backward_vars[i];
                auto &g_name = backward_vars[i + 1];
                VLOG(10) << "Bcast " << g_name << " for parameter " << p_name;
Y
yuyang18 已提交
396

C
chengduo 已提交
397 398 399 400
                switch (strategy_.reduce_) {
                  case BuildStrategy::ReduceStrategy::kReduce:
                    cur_device_id = GetAppropriateDeviceID({g_name});
                    CreateReduceOp(&result, g_name, cur_device_id);
X
Xin Pan 已提交
401
                    graph->Get<ShardedVarDevice>(kShardedVarDevice)
X
Xin Pan 已提交
402
                        .emplace(g_name, cur_device_id);
Y
Yancey1989 已提交
403 404 405
                    if (!is_dist_train) {
                      bcast_var_name_set[cur_device_id].emplace(p_name);
                    }
C
chengduo 已提交
406 407 408 409 410 411 412 413 414 415 416 417 418
                    break;
                  case BuildStrategy::ReduceStrategy::kAllReduce:
                    if (IsSparseGradient(g_name)) {
                      CreateReduceOp(&result, g_name, 0);
                      CreateBroadcastOp(&result, g_name, 0);
                    } else {
                      InsertAllReduceOp(&result, g_name);
                    }
                    break;
                  default:
                    LOG(FATAL) << "Unknown reduce strategy ";
                    break;
                }
Y
yuyang18 已提交
419
              }
C
chengduo 已提交
420
            } catch (boost::bad_get e) {
C
chengduoZH 已提交
421
            }
Y
Yu Yang 已提交
422 423 424 425 426
          }
        }
      }
    }
  }
427 428 429 430 431
  bool use_gpu = false;
#ifdef PADDLE_WITH_CUDA
  use_gpu = nccl_ctxs_ != nullptr;
#endif

Y
Yancey1989 已提交
432 433 434 435 436
  // Insert broadcast operators principle:
  // 1. Broadcast optimized parameters in Reduce strategy;
  // 2. No need broadcast optimized parameters in AllReduce strategy because of
  //    the optimization sub-graph would be run on every GPU;
  // 3. Allways broadcast received parameters in Distribute Training.
Y
Yancey1989 已提交
437 438 439
  if ((use_gpu &&
       strategy_.reduce_ == BuildStrategy::ReduceStrategy::kReduce) ||
      is_dist_train) {
440 441 442 443 444 445 446 447
    if (strategy_.fuse_broadcast_op_) {
      CreateFusedBroadcastOp(&result, bcast_var_name_set);
    } else {
      for (size_t dev_id = 0; dev_id < bcast_var_name_set.size(); ++dev_id) {
        auto &to_bcast_set = bcast_var_name_set[dev_id];
        for (auto &bcast_name : to_bcast_set) {
          CreateBroadcastOp(&result, bcast_name, dev_id);
        }
448
      }
C
chengduoZH 已提交
449 450
    }
  }
Y
Yu Yang 已提交
451
  /*
X
Xin Pan 已提交
452 453 454
  Dependency graph has been constructed. However, there are still data
  hazards need to be handled.
 */
Y
Yu Yang 已提交
455
  PolishGraphToSupportDataHazards(&result);
Y
Yu Yang 已提交
456

Y
Yu Yang 已提交
457 458 459 460
  /*
   * Only variables should be the leaves of graph.
   */
  AddOutputToLeafOps(&result);
X
Xin Pan 已提交
461
  PADDLE_ENFORCE(!ir::HasCircle(result));
Q
qiaolongfei 已提交
462
  return graph;
Y
Yu Yang 已提交
463 464
}

Y
Yancey1989 已提交
465 466 467
bool MultiDevSSAGraphBuilder::IsSparseGradient(const std::string &og) const {
  PADDLE_ENFORCE(all_vars_.count(og) != 0);
  if (all_vars_.at(og)->GetType() == proto::VarType::SELECTED_ROWS) {
C
fix ci  
chengduoZH 已提交
468 469 470
    return true;
  }
  return false;
471 472
}

473 474 475 476 477 478 479 480 481 482 483 484 485
void MultiDevSSAGraphBuilder::SetCommunicationContext(
    OpHandleBase *op_handle, const platform::Place &p) const {
#ifdef PADDLE_WITH_CUDA
  if (nccl_ctxs_ == nullptr) {
    op_handle->SetDeviceContext(p,
                                platform::DeviceContextPool::Instance().Get(p));
  }
#else
  op_handle->SetDeviceContext(p,
                              platform::DeviceContextPool::Instance().Get(p));
#endif
}

X
Xin Pan 已提交
486
void MultiDevSSAGraphBuilder::CreateBroadcastOp(ir::Graph *result,
C
chengduoZH 已提交
487
                                                const std::string &p_name,
C
chengduoZH 已提交
488
                                                size_t src_dev_id) const {
C
chengduoZH 已提交
489
#ifdef PADDLE_WITH_CUDA
X
polish  
Xin Pan 已提交
490 491 492
  auto *op_handle = new BroadcastOpHandle(
      result->CreateEmptyNode("broadcast", ir::Node::Type::kOperation),
      local_scopes_, places_, nccl_ctxs_);
C
chengduoZH 已提交
493
#else
X
polish  
Xin Pan 已提交
494 495 496
  auto *op_handle = new BroadcastOpHandle(
      result->CreateEmptyNode("broadcast", ir::Node::Type::kOperation),
      local_scopes_, places_);
C
chengduoZH 已提交
497
#endif
X
Xin Pan 已提交
498
  result->Get<GraphOps>(kGraphOps).emplace_back(op_handle);
X
Xin Pan 已提交
499

X
Xin Pan 已提交
500
  auto *in =
X
Xin Pan 已提交
501
      result->Get<GraphVars>(kGraphVars).at(src_dev_id).at(p_name).back().get();
C
chengduoZH 已提交
502 503 504 505
  op_handle->AddInput(in);

  for (size_t i = 0; i < places_.size(); ++i) {
    auto &p = places_[i];
C
chengduoZH 已提交
506
    SetCommunicationContext(op_handle, p);
X
Xin Pan 已提交
507
    auto &vars = result->Get<GraphVars>(kGraphVars).at(i).at(p_name);
X
polish  
Xin Pan 已提交
508 509 510
    auto *out_var = new VarHandle(
        result->CreateEmptyNode(p_name, ir::Node::Type::kVariable), vars.size(),
        i, p_name, p);
C
chengduoZH 已提交
511 512 513 514 515
    vars.emplace_back(out_var);
    op_handle->AddOutput(out_var);
  }
}

516 517 518 519 520 521 522 523 524 525 526 527 528 529 530 531 532 533 534 535 536 537 538 539 540 541 542 543 544 545 546 547 548 549 550 551 552 553
void MultiDevSSAGraphBuilder::CreateFusedBroadcastOp(
    ir::Graph *result,
    const std::vector<std::unordered_set<std::string>> &bcast_varnames) const {
#ifdef PADDLE_WITH_CUDA
  auto *op_handle = new FusedBroadcastOpHandle(
      result->CreateEmptyNode("fused_broadcast", ir::Node::Type::kOperation),
      local_scopes_, places_, nccl_ctxs_);
#else
  auto *op_handle = new FusedBroadcastOpHandle(
      result->CreateEmptyNode("fused_broadcast", ir::Node::Type::kOperation),
      local_scopes_, places_);
#endif
  result->Get<GraphOps>(kGraphOps).emplace_back(op_handle);

  for (size_t i = 0; i < places_.size(); ++i) {
    auto &p = places_[i];
    SetCommunicationContext(op_handle, p);
  }

  for (size_t dev_id = 0; dev_id < bcast_varnames.size(); ++dev_id) {
    for (auto &p_name : bcast_varnames[dev_id]) {
      auto *in =
          result->Get<GraphVars>(kGraphVars).at(dev_id).at(p_name).back().get();
      op_handle->AddInput(in);
      for (size_t out_dev_id = 0; out_dev_id < places_.size(); ++out_dev_id) {
        auto &p = places_[out_dev_id];
        auto &vars =
            result->Get<GraphVars>(kGraphVars).at(out_dev_id).at(p_name);
        auto *out_var = new VarHandle(
            result->CreateEmptyNode(p_name, ir::Node::Type::kVariable),
            vars.size(), out_dev_id, p_name, p);
        vars.emplace_back(out_var);
        op_handle->AddOutput(out_var);
      }
    }
  }
}

X
Xin Pan 已提交
554
void MultiDevSSAGraphBuilder::CreateComputationalOp(ir::Graph *result,
555
                                                    ir::Node *node,
C
chengduoZH 已提交
556
                                                    int dev_id) const {
X
Xin Pan 已提交
557
  result->Get<GraphOps>(kGraphOps).emplace_back(
X
Xin Pan 已提交
558
      new ComputationOpHandle(result->CreateOpNode(node->Op()),
559 560
                              local_scopes_[dev_id], places_[dev_id]));
  CreateOpHandleIOs(result, node, dev_id);
C
chengduoZH 已提交
561 562
}

X
Xin Pan 已提交
563
void MultiDevSSAGraphBuilder::InsertAllReduceOp(ir::Graph *result,
C
chengduoZH 已提交
564
                                                const std::string &og) const {
Y
Yu Yang 已提交
565
#ifdef PADDLE_WITH_CUDA
X
Xin Pan 已提交
566
  result->Get<GraphOps>(kGraphOps).emplace_back(new AllReduceOpHandle(
X
polish  
Xin Pan 已提交
567 568
      result->CreateEmptyNode("allreduce", ir::Node::Type::kOperation),
      local_scopes_, places_, nccl_ctxs_));
C
chengduoZH 已提交
569
#else
X
Xin Pan 已提交
570
  result->Get<GraphOps>(kGraphOps).emplace_back(new AllReduceOpHandle(
X
polish  
Xin Pan 已提交
571 572
      result->CreateEmptyNode("allreduce", ir::Node::Type::kOperation),
      local_scopes_, places_));
C
chengduoZH 已提交
573
#endif
X
Xin Pan 已提交
574
  auto *op_handle = result->Get<GraphOps>(kGraphOps).back().get();
Y
Yu Yang 已提交
575 576 577

  for (size_t i = 0; i < places_.size(); ++i) {
    auto &p = places_[i];
C
chengduoZH 已提交
578
    SetCommunicationContext(op_handle, p);
X
Xin Pan 已提交
579
    auto &vars = result->Get<GraphVars>(kGraphVars)[i][og];
Y
Yu Yang 已提交
580 581
    PADDLE_ENFORCE(!vars.empty());
    auto &prev_grad = vars.back();
Y
Yu Yang 已提交
582 583
    op_handle->AddInput(prev_grad.get());

X
Xin Pan 已提交
584
    auto var =
X
polish  
Xin Pan 已提交
585 586
        new VarHandle(result->CreateEmptyNode(og, ir::Node::Type::kVariable),
                      vars.size(), i, og, p);
Y
Yu Yang 已提交
587 588 589 590 591
    vars.emplace_back(var);
    op_handle->AddOutput(var);
  }
}

592
void MultiDevSSAGraphBuilder::InsertDataBalanceOp(
X
Xin Pan 已提交
593
    ir::Graph *result, const std::vector<std::string> &datas) const {
F
fengjiayi 已提交
594
#ifdef PADDLE_WITH_CUDA
X
Xin Pan 已提交
595
  result->Get<GraphOps>(kGraphOps).emplace_back(new DataBalanceOpHandle(
X
polish  
Xin Pan 已提交
596 597
      result->CreateEmptyNode("data_balance", ir::Node::Type::kOperation),
      local_scopes_, places_, nccl_ctxs_));
F
fengjiayi 已提交
598
#else
X
Xin Pan 已提交
599
  result->Get<GraphOps>(kGraphOps).emplace_back(new DataBalanceOpHandle(
X
polish  
Xin Pan 已提交
600 601
      result->CreateEmptyNode("data_balance", ir::Node::Type::kOperation),
      local_scopes_, places_));
F
fengjiayi 已提交
602
#endif
X
Xin Pan 已提交
603
  auto *op_handle = result->Get<GraphOps>(kGraphOps).back().get();
604 605 606 607
  for (size_t i = 0; i < places_.size(); ++i) {
    auto &p = places_[i];
    SetCommunicationContext(op_handle, p);
    for (const std::string &d_name : datas) {
X
Xin Pan 已提交
608
      auto &vars = result->Get<GraphVars>(kGraphVars)[i][d_name];
609 610
      PADDLE_ENFORCE(!vars.empty());
      op_handle->AddInput(vars.back().get());
X
polish  
Xin Pan 已提交
611 612 613
      auto var = new VarHandle(
          result->CreateEmptyNode(d_name, ir::Node::Type::kVariable),
          vars.size(), i, d_name, p);
614 615 616 617 618 619
      vars.emplace_back(var);
      op_handle->AddOutput(var);
    }
  }
}

X
Xin Pan 已提交
620 621
int MultiDevSSAGraphBuilder::GetOpDeviceID(const ir::Graph &graph,
                                           ir::Node *node) const {
Y
yuyang18 已提交
622
  if (strategy_.reduce_ != BuildStrategy::ReduceStrategy::kReduce) {
C
chengduoZH 已提交
623 624
    return -1;
  }
625
  int op_role = boost::get<int>(
626
      node->Op()->GetAttr(framework::OpProtoAndCheckerMaker::OpRoleAttrName()));
627 628
  if (op_role != static_cast<int>(framework::OpRole::kOptimize)) {
    return -1;
C
chengduoZH 已提交
629
  }
630
  auto param_grad = boost::get<std::vector<std::string>>(
X
Xin Pan 已提交
631
      node->Op()->GetAttr(OpProtoAndCheckerMaker::OpRoleVarAttrName()));
632 633

  PADDLE_ENFORCE_EQ(param_grad.size(), 2U);
X
Xin Pan 已提交
634
  int dev_id = GetVarDeviceID(graph, param_grad[1]);
X
Xin Pan 已提交
635 636
  PADDLE_ENFORCE_NE(dev_id, -1, "dev_id should not be -1.[%s, %s, %s]",
                    node->Op()->Type(), param_grad[0], param_grad[1]);
637
  return dev_id;
638 639
}

X
Xin Pan 已提交
640 641
int MultiDevSSAGraphBuilder::GetVarDeviceID(const ir::Graph &graph,
                                            const std::string &varname) const {
X
Xin Pan 已提交
642
  auto &sharded_var_device = graph.Get<ShardedVarDevice>(kShardedVarDevice);
X
Xin Pan 已提交
643 644
  auto got = sharded_var_device.find(varname);
  return got == sharded_var_device.end() ? -1 : got->second;
C
chengduoZH 已提交
645 646
}

647
void MultiDevSSAGraphBuilder::CreateScaleLossGradOp(
648 649
    ir::Graph *result, const std::string &loss_grad_name,
    ir::Node *out_var_node) const {
Y
Yu Yang 已提交
650
  for (size_t i = 0; i < places_.size(); ++i) {
Y
yuyang18 已提交
651 652
    // Insert ScaleCost OpHandle
    auto *dev_ctx = platform::DeviceContextPool::Instance().Get(places_[i]);
X
Xin Pan 已提交
653
    auto *op_handle = new ScaleLossGradOpHandle(
X
polish  
Xin Pan 已提交
654
        result->CreateEmptyNode("scale_loss_grad", ir::Node::Type::kOperation),
Y
yuyang18 已提交
655
        local_scopes_.size(), local_scopes_[i], places_[i], dev_ctx);
X
Xin Pan 已提交
656
    result->Get<GraphOps>(kGraphOps).emplace_back(op_handle);
Y
Yu Yang 已提交
657 658 659 660 661 662 663

    // 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);

664 665
    CreateOpOutput(result, op_handle,
                   result->CreateVarNode(out_var_node->Var()), places_[i], i);
Y
Yu Yang 已提交
666 667 668
  }
}

X
Xin Pan 已提交
669
void MultiDevSSAGraphBuilder::CreateComputationalOps(ir::Graph *result,
670
                                                     ir::Node *node,
T
typhoonzero 已提交
671 672
                                                     size_t num_places) const {
  for (size_t scope_idx = 0; scope_idx < num_places; ++scope_idx) {
Y
Yu Yang 已提交
673 674
    auto p = places_[scope_idx];
    auto s = local_scopes_[scope_idx];
X
Xin Pan 已提交
675
    result->Get<GraphOps>(kGraphOps).emplace_back(
X
Xin Pan 已提交
676
        new ComputationOpHandle(result->CreateOpNode(node->Op()), s, p));
677
    CreateOpHandleIOs(result, node, scope_idx);
Y
Yu Yang 已提交
678 679 680
  }
}

X
Xin Pan 已提交
681
VarHandle *MultiDevSSAGraphBuilder::CreateReduceOp(ir::Graph *result,
C
chengduoZH 已提交
682 683
                                                   const std::string &og,
                                                   int dst_dev_id) const {
C
chengduoZH 已提交
684
#ifdef PADDLE_WITH_CUDA
X
Xin Pan 已提交
685
  result->Get<GraphOps>(kGraphOps).emplace_back(new ReduceOpHandle(
X
polish  
Xin Pan 已提交
686 687
      result->CreateEmptyNode("reduce", ir::Node::Type::kOperation),
      local_scopes_, places_, nccl_ctxs_));
C
chengduoZH 已提交
688
#else
X
Xin Pan 已提交
689
  result->Get<GraphOps>(kGraphOps).emplace_back(new ReduceOpHandle(
X
polish  
Xin Pan 已提交
690 691
      result->CreateEmptyNode("reduce", ir::Node::Type::kOperation),
      local_scopes_, places_));
C
chengduoZH 已提交
692
#endif
X
Xin Pan 已提交
693
  auto *op_handle = result->Get<GraphOps>(kGraphOps).back().get();
C
chengduoZH 已提交
694 695 696

  for (size_t i = 0; i < places_.size(); ++i) {
    auto &p = places_[i];
C
chengduoZH 已提交
697
    SetCommunicationContext(op_handle, p);
X
Xin Pan 已提交
698
    auto &vars = result->Get<GraphVars>(kGraphVars)[i][og];
C
chengduoZH 已提交
699 700 701 702
    PADDLE_ENFORCE(!vars.empty());
    auto &prev_grad = vars.back();
    op_handle->AddInput(prev_grad.get());
  }
X
Xin Pan 已提交
703
  auto &vars = result->Get<GraphVars>(kGraphVars)[dst_dev_id][og];
X
polish  
Xin Pan 已提交
704 705 706
  auto var =
      new VarHandle(result->CreateEmptyNode(og, ir::Node::Type::kVariable),
                    vars.size(), dst_dev_id, og, places_[dst_dev_id]);
C
chengduoZH 已提交
707 708 709 710 711
  vars.emplace_back(var);
  op_handle->AddOutput(var);
  return var;
}

Y
Yancey1989 已提交
712 713
int MultiDevSSAGraphBuilder::CreateDistTrainOp(ir::Graph *result,
                                               ir::Node *node) const {
Y
Yancey1989 已提交
714
  int op_dev_id = -1;
715 716 717
  std::vector<std::string> input_var_names;
  std::vector<std::string> output_var_names;
  for (ir::Node *input : node->inputs) {
X
Xin Pan 已提交
718
    input_var_names.push_back(input->Name());
719 720
  }
  for (ir::Node *output : node->outputs) {
X
Xin Pan 已提交
721
    output_var_names.push_back(output->Name());
722 723 724
  }

  if (node->Op()->Type() == "split_byref" ||
725 726
      node->Op()->Type() == "split_selected_rows" ||
      node->Op()->Type() == "split_ids") {
X
Xin Pan 已提交
727
    // TODO(paddle-dev): getting the first var is not safe.
X
Xin Pan 已提交
728
    op_dev_id = GetVarDeviceID(*result, input_var_names[0]);
Y
Yancey1989 已提交
729
    if (strategy_.reduce_ == BuildStrategy::ReduceStrategy::kAllReduce) {
730 731
      op_dev_id = GetAppropriateDeviceID(input_var_names);
      for (auto &varname : input_var_names) {
X
Xin Pan 已提交
732
        result->Get<ShardedVarDevice>(kShardedVarDevice)
X
Xin Pan 已提交
733
            .emplace(varname, op_dev_id);
Y
Yancey1989 已提交
734 735
      }
    }
736
    for (auto &varname : output_var_names) {
X
Xin Pan 已提交
737
      result->Get<ShardedVarDevice>(kShardedVarDevice)
X
Xin Pan 已提交
738
          .emplace(varname, op_dev_id);
Y
Yancey1989 已提交
739
    }
740
  } else if (node->Op()->Type() == "concat") {
X
Xin Pan 已提交
741
    op_dev_id = GetVarDeviceID(*result, input_var_names[0]);
742
    for (auto &varname : output_var_names) {
X
Xin Pan 已提交
743
      result->Get<ShardedVarDevice>(kShardedVarDevice)
X
Xin Pan 已提交
744
          .emplace(varname, op_dev_id);
Y
yi.wu 已提交
745
    }
Y
Yancey1989 已提交
746
  } else {
747
    LOG(ERROR) << "got unexpected dist op: " << node->Op()->Type();
W
Wu Yi 已提交
748
    PADDLE_THROW(
Y
Yancey1989 已提交
749 750 751 752 753
        "the distribute training related op should be in [split_byref, "
        "concat].");
  }

  PADDLE_ENFORCE(op_dev_id != -1,
754 755
                 "can not find right place for distributed op: %s",
                 node->Op()->Type());
Y
Yancey1989 已提交
756

757
  CreateComputationalOp(result, node, op_dev_id);
Y
Yancey1989 已提交
758
  return op_dev_id;
W
Wu Yi 已提交
759 760 761 762 763 764 765 766 767 768 769 770 771 772
}

void SetOpInputsAllPlaces(ir::Graph *result, ir::Node *node, int num_places) {
  auto *op_handle = result->Get<GraphOps>(kGraphOps).back().get();
  for (ir::Node *input : node->inputs) {
    VarHandle *var = nullptr;
    for (int place_offset = 0; place_offset < num_places; ++place_offset) {
      auto &var_holders = result->Get<GraphVars>(kGraphVars)[place_offset];
      auto &var_holder = var_holders[input->Name()];
      if (!var_holder.empty()) {
        var = var_holder.rbegin()->get();
        op_handle->AddInput(var);
      }
    }
Y
Yancey1989 已提交
773 774 775
  }
}

776
// Create RPC related op handles that connects its in ops and out ops.
Y
Yancey1989 已提交
777 778
int MultiDevSSAGraphBuilder::CreateRPCOp(ir::Graph *result,
                                         ir::Node *node) const {
Y
Yancey1989 已提交
779
  int op_dev_id = -1;
780
  if (node->Op()->Type() == "send") {
X
Xin Pan 已提交
781
    // TODO(paddle-dev): getting the first var is not safe.
X
Xin Pan 已提交
782
    op_dev_id = GetVarDeviceID(*result, node->inputs[0]->Name());
X
Xin Pan 已提交
783 784
    PADDLE_ENFORCE(!ir::IsControlDepVar(*node->inputs[0]),
                   "This hack no longer holds, please fix.");
Y
Yancey1989 已提交
785 786 787
    // the variable name which contains .block means it was splited by
    // split_byref op
    if (strategy_.reduce_ == BuildStrategy::ReduceStrategy::kAllReduce &&
X
Xin Pan 已提交
788
        node->inputs[0]->Name().find(".block") == std::string::npos) {
789 790
      std::vector<std::string> input_var_names;
      for (ir::Node *n : node->inputs) {
X
Xin Pan 已提交
791
        input_var_names.push_back(n->Name());
792
      }
W
Wu Yi 已提交
793 794 795 796 797 798
      auto send_param_grad = boost::get<std::vector<std::string>>(
          node->Op()->GetAttr(OpProtoAndCheckerMaker::OpRoleVarAttrName()));
      PADDLE_ENFORCE_EQ(send_param_grad.size(), 2U);
      op_dev_id = GetAppropriateDeviceID({send_param_grad[1]});
      VLOG(10) << "send grad " << input_var_names[0] << " origin "
               << send_param_grad[1] << " place: " << op_dev_id;
799
      for (auto &varname : input_var_names) {
X
Xin Pan 已提交
800
        result->Get<ShardedVarDevice>(kShardedVarDevice)
X
Xin Pan 已提交
801
            .emplace(varname, op_dev_id);
Y
Yancey1989 已提交
802
      }
W
Wu Yi 已提交
803 804
      result->Get<ShardedVarDevice>(kShardedVarDevice)
          .emplace(send_param_grad[1], op_dev_id);
Y
Yancey1989 已提交
805
    }
806 807 808
  } else if (node->Op()->Type() == "recv") {
    std::vector<std::string> output_var_names;
    for (ir::Node *n : node->outputs) {
X
Xin Pan 已提交
809
      output_var_names.push_back(n->Name());
810
    }
W
Wu Yi 已提交
811 812 813 814 815 816 817 818 819 820
    auto recv_param_grad = boost::get<std::vector<std::string>>(
        node->Op()->GetAttr(OpProtoAndCheckerMaker::OpRoleVarAttrName()));
    if (recv_param_grad.size() == 2U) {
      op_dev_id = GetVarDeviceID(*result, recv_param_grad[1]);
      VLOG(10) << "recv param " << recv_param_grad[0]
               << " get grad place: " << recv_param_grad[1]
               << " place: " << op_dev_id;
    } else {
      op_dev_id = GetAppropriateDeviceID(output_var_names);
    }
821
    for (auto &varname : output_var_names) {
X
Xin Pan 已提交
822
      result->Get<ShardedVarDevice>(kShardedVarDevice)
X
Xin Pan 已提交
823
          .emplace(varname, op_dev_id);
Y
Yancey1989 已提交
824 825
    }
  } else {
W
Wu Yi 已提交
826
    // send_barrier, fetch_barrier will run on place 0;
Y
Yancey1989 已提交
827 828 829 830
    op_dev_id = 0;
  }

  PADDLE_ENFORCE(op_dev_id != -1, "can not find the right place for rpc op: %s",
831
                 node->Op()->Type());
X
Xin Pan 已提交
832
  result->Get<GraphOps>(kGraphOps).emplace_back(new RPCOpHandle(
833 834
      result->CreateOpNode(node->Op()), *node->Op(), local_scopes_[op_dev_id],
      node->Op()->Type(), places_[op_dev_id]));
Y
fix pe  
Yancey1989 已提交
835

W
Wu Yi 已提交
836 837
  if (node->Op()->Type() == "send") {
    CreateOpHandleIOs(result, node, op_dev_id);
Y
Yancey1989 已提交
838
  } else {
W
Wu Yi 已提交
839 840 841 842 843 844
    // send_barrier, recv, fetch_barrier's inputs are deps var, get them from
    // all places
    auto p = places_[op_dev_id];
    auto *op_handle = result->Get<GraphOps>(kGraphOps).back().get();
    op_handle->SetDeviceContext(p,
                                platform::DeviceContextPool::Instance().Get(p));
Y
Yancey1989 已提交
845

W
Wu Yi 已提交
846 847 848 849 850 851 852 853 854 855 856 857 858 859 860 861 862 863
    SetOpInputsAllPlaces(result, node, places_.size());
    for (ir::Node *output : node->outputs) {
      int outvar_dev_id = op_dev_id;
      if (node->Op()->Type() == "fetch_barrier") {
        outvar_dev_id = GetVarDeviceID(*result, output->Name());
        PADDLE_ENFORCE_NE(outvar_dev_id, -1);
      }
      p = places_[outvar_dev_id];
      ir::Node *new_node = nullptr;
      if (output->Var()) {
        new_node = result->CreateVarNode(output->Var());
      } else {
        new_node =
            result->CreateEmptyNode(output->Name(), ir::Node::Type::kVariable);
      }
      CreateOpOutput(result, op_handle, new_node, p, outvar_dev_id);
    }
  }
Y
Yancey1989 已提交
864
  return op_dev_id;
Y
Yu Yang 已提交
865 866
}

867
bool MultiDevSSAGraphBuilder::IsScaleLossOp(ir::Node *node) const {
Y
yuyang18 已提交
868
  return boost::get<int>(
869
             node->Op()->GetAttr(OpProtoAndCheckerMaker::OpRoleAttrName())) ==
Y
Fix bug  
yuyang18 已提交
870 871 872
             (static_cast<int>(OpRole::kBackward) |
              static_cast<int>(OpRole::kLoss)) &&
         !loss_var_name_.empty();  // If loss_var is empty. This is test mode
Y
Yu Yang 已提交
873
}
Y
Yu Yang 已提交
874 875 876
}  // namespace details
}  // namespace framework
}  // namespace paddle
X
Xin Pan 已提交
877

X
Xin Pan 已提交
878
REGISTER_PASS(multi_devices_pass,
X
Xin Pan 已提交
879 880 881 882 883 884
              paddle::framework::details::MultiDevSSAGraphBuilder)
    .RequirePassAttr(paddle::framework::details::kLossVarName)
    .RequirePassAttr(paddle::framework::details::kPlaces)
    .RequirePassAttr(paddle::framework::details::kParams)
    .RequirePassAttr(paddle::framework::details::kLocalScopes)
    .RequirePassAttr(paddle::framework::details::kStrategy);