init.cc 4.3 KB
Newer Older
1
/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved.
D
dzhwinter 已提交
2

L
Luo Tao 已提交
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
D
dzhwinter 已提交
6

L
Luo Tao 已提交
7
    http://www.apache.org/licenses/LICENSE-2.0
D
dzhwinter 已提交
8

L
Luo Tao 已提交
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
D
dzhwinter 已提交
15
#include <algorithm>
D
dzhwinter 已提交
16
#include <stdexcept>
D
dzhwinter 已提交
17 18
#include <string>

Y
Yi Wang 已提交
19
#include "paddle/fluid/framework/operator.h"
T
tensor-tang 已提交
20
#include "paddle/fluid/platform/cpu_helper.h"
T
tensor-tang 已提交
21
#include "paddle/fluid/platform/cpu_info.h"
Y
Yu Yang 已提交
22
#include "paddle/fluid/platform/cuda_device_guard.h"
Y
Yi Wang 已提交
23
#include "paddle/fluid/platform/device_context.h"
24
#include "paddle/fluid/platform/init.h"
Y
Yi Wang 已提交
25
#include "paddle/fluid/platform/place.h"
26
#include "paddle/fluid/string/piece.h"
D
dzhwinter 已提交
27

T
tensor-tang 已提交
28 29
DEFINE_int32(paddle_num_threads, 1,
             "Number of threads for each paddle instance.");
T
tensor-tang 已提交
30

D
dzhwinter 已提交
31 32 33 34
namespace paddle {
namespace framework {

std::once_flag gflags_init_flag;
X
Xin Pan 已提交
35
std::once_flag p2p_init_flag;
D
dzhwinter 已提交
36

37
void InitGflags(std::vector<std::string> argv) {
D
dzhwinter 已提交
38
  std::call_once(gflags_init_flag, [&]() {
W
wanghaoshuang 已提交
39
    argv.insert(argv.begin(), "dummy");
D
dzhwinter 已提交
40 41 42 43 44 45 46 47 48 49 50 51 52
    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;
  });
}

53 54 55 56 57 58 59 60 61 62 63 64 65 66 67
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 {
Y
Yu Yang 已提交
68
          platform::CUDADeviceGuard guard(devices[i]);
69 70 71 72 73 74 75 76
          cudaDeviceEnablePeerAccess(devices[j], 0);
        }
      }
    }
  });
#endif
}

X
Xin Pan 已提交
77
void InitDevices(bool init_p2p) {
78
  /*Init all available devices by default */
79
  std::vector<int> devices;
80
#ifdef PADDLE_WITH_CUDA
D
dzhwinter 已提交
81
  try {
82 83 84 85
    int count = platform::GetCUDADeviceCount();
    for (int i = 0; i < count; ++i) {
      devices.push_back(i);
    }
D
dzhwinter 已提交
86 87
  } catch (const std::exception &exp) {
    LOG(WARNING) << "Compiled with WITH_GPU, but no GPU found in runtime.";
88
  }
D
dzhwinter 已提交
89
#endif
90
  InitDevices(init_p2p, devices);
D
dzhwinter 已提交
91 92
}

93 94 95 96 97 98 99 100 101 102 103 104
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.";
  }
#endif

  for (size_t i = 0; i < devices.size(); ++i) {
W
wanghaoshuang 已提交
105
    if (devices[i] >= count || devices[i] < 0) {
106 107 108 109 110 111 112 113 114 115
      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);
116
#ifndef PADDLE_WITH_MKLDNN
T
tensor-tang 已提交
117
  platform::SetNumThreads(FLAGS_paddle_num_threads);
118
#endif
T
tensor-tang 已提交
119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134

  if (platform::jit::MayIUse(platform::jit::avx512_common)) {
#ifndef __AVX512F__
    LOG(WARNING) << "AVX512F is available, Please re-compile on local machine";
#endif
  }
  if (platform::jit::MayIUse(platform::jit::avx2)) {
#ifndef __AVX2__
    LOG(WARNING) << "AVX2 is available, Please re-compile on local machine";
#endif
  }
  if (platform::jit::MayIUse(platform::jit::avx)) {
#ifndef __AVX__
    LOG(WARNING) << "AVX is available, Please re-compile on local machine";
#endif
  }
135 136
}

Y
Yang Yu 已提交
137
void InitGLOG(const std::string &prog_name) {
Y
Yang Yu 已提交
138 139 140
  // 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
141
  google::InstallFailureSignalHandler();
Y
Yang Yu 已提交
142 143
}

D
dzhwinter 已提交
144 145
}  // namespace framework
}  // namespace paddle