init.cc 4.9 KB
Newer Older
1
/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved.
2

3 4 5
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
6

7
    http://www.apache.org/licenses/LICENSE-2.0
8

9 10 11 12 13
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. */
Y
Yang Yu 已提交
14
#include <string.h>  // for strdup
15
#include <algorithm>
D
dzhwinter 已提交
16
#include <stdexcept>
17
#include <string>
18
#include <vector>
19

Y
Yi Wang 已提交
20 21 22 23
#include "paddle/fluid/framework/init.h"
#include "paddle/fluid/framework/operator.h"
#include "paddle/fluid/platform/device_context.h"
#include "paddle/fluid/platform/place.h"
24
#include "paddle/fluid/string/piece.h"
25 26 27 28

namespace paddle {
namespace framework {

29 30 31
DEFINE_string(devices, "", "The devices to be used.");
DEFINE_bool(init_p2p, true, "Whether to init p2p.");

32
std::once_flag gflags_init_flag;
X
Xin Pan 已提交
33
std::once_flag p2p_init_flag;
34

35 36
using paddle::platform::DeviceContextPool;

37
void Init(std::vector<std::string> argv) {
W
wanghaoshuang 已提交
38
  InitGflags(argv);
39 40 41 42 43 44 45 46 47 48
  // init devices
  std::vector<int> devices;
  std::string token;
  std::istringstream tokenStream(FLAGS_devices);
  while (std::getline(tokenStream, token, ',')) {
    devices.push_back(std::stoi(token));
  }
  InitDevices(FLAGS_init_p2p, devices);
}

49
void InitGflags(std::vector<std::string> argv) {
50
  std::call_once(gflags_init_flag, [&]() {
W
wanghaoshuang 已提交
51
    argv.push_back("dummy");
52 53 54 55 56 57 58 59 60 61 62 63 64
    int argc = argv.size();
    char **arr = new char *[argv.size()];
    std::string line;
    for (size_t i = 0; i < argv.size(); i++) {
      arr[i] = &argv[i][0];
      line += argv[i];
      line += ' ';
    }
    google::ParseCommandLineFlags(&argc, &arr, true);
    VLOG(1) << "Init commandline: " << line;
  });
}

X
Xin Pan 已提交
65
void InitP2P(int count) {
X
fix  
Xin Pan 已提交
66
#ifdef PADDLE_WITH_CUDA
X
Xin Pan 已提交
67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82
  std::call_once(p2p_init_flag, [&]() {
    for (int i = 0; i < count; ++i) {
      for (int j = 0; j < count; ++j) {
        if (i == j) continue;
        int can_acess = -1;
        PADDLE_ENFORCE(cudaDeviceCanAccessPeer(&can_acess, i, j),
                       "Failed to test P2P access.");
        if (can_acess != 1) {
          LOG(WARNING) << "Cannot enable P2P access from " << i << " to " << j;
        } else {
          cudaSetDevice(i);
          cudaDeviceEnablePeerAccess(j, 0);
        }
      }
    }
  });
X
fix  
Xin Pan 已提交
83
#endif
X
Xin Pan 已提交
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
void InitP2P(std::vector<int> devices) {
#ifdef PADDLE_WITH_CUDA
  std::call_once(p2p_init_flag, [&]() {
    int count = devices.size();
    for (int i = 0; i < count; ++i) {
      for (int j = 0; j < count; ++j) {
        if (devices[i] == devices[j]) continue;
        int can_acess = -1;
        PADDLE_ENFORCE(
            cudaDeviceCanAccessPeer(&can_acess, devices[i], devices[j]),
            "Failed to test P2P access.");
        if (can_acess != 1) {
          LOG(WARNING) << "Cannot enable P2P access from " << devices[i]
                       << " to " << devices[j];
        } else {
          cudaSetDevice(devices[i]);
          cudaDeviceEnablePeerAccess(devices[j], 0);
        }
      }
    }
  });
#endif
}

X
Xin Pan 已提交
110
void InitDevices(bool init_p2p) {
111
  /*Init all available devices by default */
112

113
  std::vector<platform::Place> places;
114
  places.emplace_back(platform::CPUPlace());
D
dzhwinter 已提交
115
  int count = 0;
116

117
#ifdef PADDLE_WITH_CUDA
D
dzhwinter 已提交
118 119 120 121
  try {
    count = platform::GetCUDADeviceCount();
  } catch (const std::exception &exp) {
    LOG(WARNING) << "Compiled with WITH_GPU, but no GPU found in runtime.";
122
  }
123
#else
124
  LOG(WARNING)
D
dzhwinter 已提交
125
      << "'CUDA' is not supported, Please re-compile with WITH_GPU option";
126 127
#endif

D
dzhwinter 已提交
128 129 130
  for (int i = 0; i < count; ++i) {
    places.emplace_back(platform::CUDAPlace(i));
  }
X
Xin Pan 已提交
131 132 133
  if (init_p2p) {
    InitP2P(count);
  }
134
  platform::DeviceContextPool::Init(places);
135 136
}

137 138 139 140 141 142 143 144 145 146 147 148 149 150 151
void InitDevices(bool init_p2p, const std::vector<int> devices) {
  std::vector<platform::Place> places;
  int count = 0;
#ifdef PADDLE_WITH_CUDA
  try {
    count = platform::GetCUDADeviceCount();
  } catch (const std::exception &exp) {
    LOG(WARNING) << "Compiled with WITH_GPU, but no GPU found in runtime.";
  }
#else
  LOG(WARNING)
      << "'CUDA' is not supported, Please re-compile with WITH_GPU option";
#endif

  for (size_t i = 0; i < devices.size(); ++i) {
W
wanghaoshuang 已提交
152
    if (devices[i] >= count || devices[i] < 0) {
153 154 155 156 157 158 159 160 161 162 163 164
      LOG(WARNING) << "Invalid devices id.";
      continue;
    }
    places.emplace_back(platform::CUDAPlace(devices[i]));
  }
  if (init_p2p) {
    InitP2P(devices);
  }
  places.emplace_back(platform::CPUPlace());
  platform::DeviceContextPool::Init(places);
}

Y
Yang Yu 已提交
165
void InitGLOG(const std::string &prog_name) {
Y
Yang Yu 已提交
166 167 168
  // glog will not hold the ARGV[0] inside.
  // Use strdup to alloc a new string.
  google::InitGoogleLogging(strdup(prog_name.c_str()));
Y
Yang Yu 已提交
169
  google::InstallFailureSignalHandler();
Y
Yang Yu 已提交
170 171
}

172 173
}  // namespace framework
}  // namespace paddle