init.cc 5.1 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"
22
#include "paddle/fluid/string/split.h"
S
sneaxiy 已提交
23
#ifdef PADDLE_WITH_CUDA
Y
Yu Yang 已提交
24
#include "paddle/fluid/platform/cuda_device_guard.h"
S
sneaxiy 已提交
25
#endif
Y
Yi Wang 已提交
26
#include "paddle/fluid/platform/device_context.h"
27
#include "paddle/fluid/platform/init.h"
Y
Yi Wang 已提交
28
#include "paddle/fluid/platform/place.h"
29
#include "paddle/fluid/string/piece.h"
D
dzhwinter 已提交
30

T
tensor-tang 已提交
31 32
DEFINE_int32(paddle_num_threads, 1,
             "Number of threads for each paddle instance.");
T
tensor-tang 已提交
33

D
dzhwinter 已提交
34 35 36 37
namespace paddle {
namespace framework {

std::once_flag gflags_init_flag;
X
Xin Pan 已提交
38
std::once_flag p2p_init_flag;
D
dzhwinter 已提交
39

40
void InitGflags(std::vector<std::string> argv) {
D
dzhwinter 已提交
41
  std::call_once(gflags_init_flag, [&]() {
C
chengduo 已提交
42
    FLAGS_logtostderr = true;
W
wanghaoshuang 已提交
43
    argv.insert(argv.begin(), "dummy");
D
dzhwinter 已提交
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);
M
minqiyang 已提交
53
    VLOG(1) << "Init commandline: " << line;
D
dzhwinter 已提交
54 55 56
  });
}

57 58 59 60 61 62 63 64 65 66 67 68 69 70 71
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
72
          platform::CUDADeviceGuard guard(devices[i]);
73 74 75 76 77 78 79 80
          cudaDeviceEnablePeerAccess(devices[j], 0);
        }
      }
    }
  });
#endif
}

X
Xin Pan 已提交
81
void InitDevices(bool init_p2p) {
82
  /*Init all available devices by default */
83
  std::vector<int> devices;
84
#ifdef PADDLE_WITH_CUDA
D
dzhwinter 已提交
85
  try {
86 87
    // use user specified GPUs in single-node multi-process mode.
    devices = platform::GetSelectedDevices();
D
dzhwinter 已提交
88 89
  } catch (const std::exception &exp) {
    LOG(WARNING) << "Compiled with WITH_GPU, but no GPU found in runtime.";
90
  }
D
dzhwinter 已提交
91
#endif
92
  InitDevices(init_p2p, devices);
D
dzhwinter 已提交
93 94
}

95 96 97 98
void InitDevices(bool init_p2p, const std::vector<int> devices) {
  std::vector<platform::Place> places;

  for (size_t i = 0; i < devices.size(); ++i) {
99 100 101
    // In multi process multi gpu mode, we may have gpuid = 7
    // but count = 1.
    if (devices[i] < 0) {
102 103 104
      LOG(WARNING) << "Invalid devices id.";
      continue;
    }
105

106 107 108 109 110 111 112
    places.emplace_back(platform::CUDAPlace(devices[i]));
  }
  if (init_p2p) {
    InitP2P(devices);
  }
  places.emplace_back(platform::CPUPlace());
  platform::DeviceContextPool::Init(places);
113
  platform::DeviceTemporaryAllocator::Init();
114
#ifndef PADDLE_WITH_MKLDNN
T
tensor-tang 已提交
115
  platform::SetNumThreads(FLAGS_paddle_num_threads);
116
#endif
T
tensor-tang 已提交
117

T
tensor-tang 已提交
118
#if !defined(_WIN32) && !defined(__APPLE__) && !defined(__OSX__)
T
tensor-tang 已提交
119
  if (platform::MayIUse(platform::avx)) {
T
tensor-tang 已提交
120 121 122 123
#ifndef __AVX__
    LOG(WARNING) << "AVX is available, Please re-compile on local machine";
#endif
  }
124 125 126 127 128 129 130 131 132 133

// Throw some informations when CPU instructions mismatch.
#define AVX_GUIDE(compiletime, runtime)                                     \
  LOG(FATAL)                                                                \
      << "This version is compiled on higher instruction(" #compiletime     \
         ") system, you may encounter illegal instruction error running on" \
         " your local CPU machine. Please reinstall the " #runtime          \
         " version or compile from source code."

#ifdef __AVX512F__
T
tensor-tang 已提交
134 135
  if (!platform::MayIUse(platform::avx512f)) {
    if (platform::MayIUse(platform::avx2)) {
136
      AVX_GUIDE(AVX512, AVX2);
T
tensor-tang 已提交
137
    } else if (platform::MayIUse(platform::avx)) {
138 139 140 141 142
      AVX_GUIDE(AVX512, AVX);
    } else {
      AVX_GUIDE(AVX512, NonAVX);
    }
  }
T
tensor-tang 已提交
143
#endif
144 145

#ifdef __AVX2__
T
tensor-tang 已提交
146 147
  if (!platform::MayIUse(platform::avx2)) {
    if (platform::MayIUse(platform::avx)) {
148 149 150 151
      AVX_GUIDE(AVX2, AVX);
    } else {
      AVX_GUIDE(AVX2, NonAVX);
    }
T
tensor-tang 已提交
152 153
  }
#endif
154 155

#ifdef __AVX__
T
tensor-tang 已提交
156
  if (!platform::MayIUse(platform::avx)) {
157
    AVX_GUIDE(AVX, NonAVX);
T
tensor-tang 已提交
158
  }
159 160
#endif
#undef AVX_GUIDE
T
tensor-tang 已提交
161 162

#endif
163 164
}

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()));
P
peizhilin 已提交
169
#ifndef _WIN32
Y
Yang Yu 已提交
170
  google::InstallFailureSignalHandler();
P
peizhilin 已提交
171
#endif
Y
Yang Yu 已提交
172 173
}

D
dzhwinter 已提交
174 175
}  // namespace framework
}  // namespace paddle