executor.cc 4.1 KB
Newer Older
Q
qijun 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 Reserve.

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/framework/executor.h"
Y
Yang Yang 已提交
16

Y
Yang Yang 已提交
17
#include <algorithm>
Y
Yang Yang 已提交
18
#include <iostream>
Q
qijun 已提交
19
#include <memory>
Y
Yang Yang 已提交
20
#include <set>
Y
Yang Yang 已提交
21
#include <vector>
Y
Yang Yang 已提交
22

Q
QI JUN 已提交
23
#include "paddle/framework/feed_fetch_type.h"
Y
Yang Yang 已提交
24
#include "paddle/framework/lod_tensor.h"
Q
qijun 已提交
25
#include "paddle/framework/op_registry.h"
Q
qijun 已提交
26
#include "paddle/framework/scope.h"
Q
qijun 已提交
27 28 29 30

namespace paddle {
namespace framework {

Y
Yang Yang 已提交
31 32 33
const std::string kFeedOpType = "feed";
const std::string kFetchOpType = "fetch";

Y
Yu Yang 已提交
34
Executor::Executor(const std::vector<platform::Place>& places) : own_(true) {
Y
Yang Yang 已提交
35
  PADDLE_ENFORCE_GT(places.size(), 0);
Q
qijun 已提交
36
  device_contexts_.resize(places.size());
Q
qijun 已提交
37
  for (size_t i = 0; i < places.size(); i++) {
Q
qijun 已提交
38
    if (platform::is_cpu_place(places[i])) {
Q
qijun 已提交
39 40 41
      device_contexts_[i] = new platform::CPUDeviceContext(
          boost::get<platform::CPUPlace>(places[i]));
    } else if (platform::is_gpu_place(places[i])) {
Q
qijun 已提交
42
#ifdef PADDLE_WITH_CUDA
Q
qijun 已提交
43 44
      device_contexts_[i] = new platform::CUDADeviceContext(
          boost::get<platform::GPUPlace>(places[i]));
Q
qijun 已提交
45
#else
Q
qijun 已提交
46 47 48
      PADDLE_THROW(
          "'GPUPlace' is not supported, Please re-compile with WITH_GPU "
          "option");
Q
qijun 已提交
49 50
#endif
    }
Q
qijun 已提交
51 52 53
  }
}

Q
qijun 已提交
54
Executor::~Executor() {
Y
Yu Yang 已提交
55 56 57 58
  if (own_) {
    for (auto& device_context : device_contexts_) {
      delete device_context;
    }
Q
qijun 已提交
59 60 61
  }
}

Q
QI JUN 已提交
62 63 64 65 66 67 68 69 70
static void CreateTensor(Variable* var, VarDesc::VarType var_type) {
  if (var_type == VarDesc::LOD_TENSOR) {
    var->GetMutable<LoDTensor>();
  } else if (var_type == VarDesc::SELECTED_ROWS) {
    var->GetMutable<SelectedRows>();
  } else if (var_type == VarDesc::FEED_MINIBATCH) {
    var->GetMutable<FeedFetchList>();
  } else if (var_type == VarDesc::FETCH_LIST) {
    var->GetMutable<FeedFetchList>();
Y
Yu Yang 已提交
71 72
  } else if (var_type == VarDesc::STEP_SCOPES) {
    var->GetMutable<std::vector<framework::Scope>>();
Q
QI JUN 已提交
73 74
  } else {
    PADDLE_THROW(
Y
Yu Yang 已提交
75 76 77
        "Variable type %d is not in "
        "[LoDTensor, SelectedRows, FEED_MINIBATCH, FETCH_LIST]",
        var_type);
Q
QI JUN 已提交
78 79 80
  }
}

Y
Yu Yang 已提交
81 82
void Executor::Run(const ProgramDescBind& pdesc, Scope* scope, int block_id,
                   bool create_local_scope) {
Y
Yang Yang 已提交
83
  // TODO(tonyyang-svail):
Y
Yang Yang 已提交
84
  //    - only runs on the first device (i.e. no interdevice communication)
Y
Yang Yang 已提交
85
  //    - will change to use multiple blocks for RNN op and Cond Op
86
  PADDLE_ENFORCE_LT(static_cast<size_t>(block_id), pdesc.Size());
87
  auto& block = pdesc.Block(block_id);
Y
Yang Yang 已提交
88
  auto& device = device_contexts_[0];
Y
Yang Yang 已提交
89

Y
Yu Yang 已提交
90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108
  Scope* local_scope = scope;
  if (create_local_scope) {
    local_scope = &scope->NewScope();
    for (auto& var : block.AllVars()) {
      if (var->Persistable()) {
        auto* ptr = scope->Var(var->Name());
        CreateTensor(ptr, var->GetType());
        VLOG(3) << "Create Variable " << var->Name()
                << " global, which pointer is " << ptr;
      } else {
        auto* ptr = local_scope->Var(var->Name());
        CreateTensor(ptr, var->GetType());
        VLOG(3) << "Create Variable " << var->Name()
                << " locally, which pointer is " << ptr;
      }
    }
  } else {
    for (auto& var : block.AllVars()) {
      auto* ptr = local_scope->Var(var->Name());
109
      CreateTensor(ptr, var->GetType());
Y
Yu Yang 已提交
110 111
      VLOG(3) << "Create variable " << var->Name() << ", which pointer is "
              << ptr;
Y
Yang Yang 已提交
112
    }
Y
Yang Yang 已提交
113
  }
Y
Yang Yang 已提交
114

115 116
  for (auto& op_desc : block.AllOps()) {
    auto op = paddle::framework::OpRegistry::CreateOp(*op_desc);
Y
Yu Yang 已提交
117 118 119 120
    op->Run(*local_scope, *device);
  }
  if (create_local_scope) {
    scope->DeleteScope(local_scope);
Q
qijun 已提交
121
  }
Q
qijun 已提交
122 123
}

Y
Yu Yang 已提交
124 125 126
Executor::Executor(const platform::DeviceContext& device)
    : device_contexts_({&device}), own_(false) {}

Q
qijun 已提交
127 128
}  // namespace framework
}  // namespace paddle