init.cc 13.5 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. */
14
#include <csignal>
C
chengduo 已提交
15
#include <fstream>
D
dzhwinter 已提交
16 17
#include <string>

T
tensor-tang 已提交
18
#include "paddle/fluid/platform/cpu_helper.h"
T
tensor-tang 已提交
19
#include "paddle/fluid/platform/cpu_info.h"
20 21
#include "paddle/fluid/platform/npu_info.h"
#include "paddle/fluid/string/split.h"
22
#if defined(PADDLE_WITH_CUDA) || defined(PADDLE_WITH_HIP)
Y
Yu Yang 已提交
23
#include "paddle/fluid/platform/cuda_device_guard.h"
24 25
#endif
#ifdef PADDLE_WITH_CUDA
26
#include "paddle/fluid/platform/dynload/cupti.h"
S
sneaxiy 已提交
27
#endif
Y
Yi Wang 已提交
28
#include "paddle/fluid/platform/device_context.h"
29
#include "paddle/fluid/platform/init.h"
Y
Yi Wang 已提交
30
#include "paddle/fluid/platform/place.h"
31

32
#ifdef PADDLE_WITH_XPU
Q
QingshuChen 已提交
33 34
#include "paddle/fluid/platform/xpu/xpu_header.h"
#include "paddle/fluid/platform/xpu/xpu_info.h"
35 36
#endif

37 38 39
#ifdef WITH_WIN_DUMP_DBG
#include <stdio.h>
#include <time.h>
40 41 42
#ifndef NOMINMAX
#define NOMINMAX  // msvc max/min macro conflict with std::min/max
#endif
43
#include <windows.h>
44

45 46 47
#include "DbgHelp.h"
#endif

48
DECLARE_int32(paddle_num_threads);
Z
Zeng Jinle 已提交
49 50 51 52
PADDLE_DEFINE_EXPORTED_int32(
    multiple_of_cupti_buffer_size, 1,
    "Multiple of the CUPTI device buffer size. If the timestamps have "
    "been dropped when you are profiling, try increasing this value.");
T
tensor-tang 已提交
53

54 55 56 57
namespace paddle {
namespace platform {

void ParseCommandLineFlags(int argc, char **argv, bool remove) {
58
  ::GFLAGS_NAMESPACE::ParseCommandLineFlags(&argc, &argv, remove);
59 60 61 62 63
}

}  // namespace platform
}  // namespace paddle

D
dzhwinter 已提交
64 65 66
namespace paddle {
namespace framework {

67 68 69 70
#ifdef _WIN32
#define strdup _strdup
#endif

D
dzhwinter 已提交
71
std::once_flag gflags_init_flag;
72
std::once_flag glog_init_flag;
73
std::once_flag npu_init_flag;
D
dzhwinter 已提交
74

75
bool InitGflags(std::vector<std::string> args) {
76
  bool successed = false;
D
dzhwinter 已提交
77
  std::call_once(gflags_init_flag, [&]() {
C
chengduo 已提交
78
    FLAGS_logtostderr = true;
L
Leo Chen 已提交
79 80 81 82 83
    // NOTE(zhiqiu): dummy is needed, since the function
    // ParseNewCommandLineFlags in gflags.cc starts processing
    // commandline strings from idx 1.
    // The reason is, it assumes that the first one (idx 0) is
    // the filename of executable file.
84 85
    args.insert(args.begin(), "dummy");
    std::vector<char *> argv;
D
dzhwinter 已提交
86
    std::string line;
87 88 89 90
    int argc = args.size();
    for (auto &arg : args) {
      argv.push_back(const_cast<char *>(arg.data()));
      line += arg;
D
dzhwinter 已提交
91 92
      line += ' ';
    }
L
Leo Chen 已提交
93 94
    VLOG(1) << "Before Parse: argc is " << argc
            << ", Init commandline: " << line;
95 96

    char **arr = argv.data();
97
    ::GFLAGS_NAMESPACE::ParseCommandLineFlags(&argc, &arr, true);
98
    successed = true;
99 100

    VLOG(1) << "After Parse: argc is " << argc;
D
dzhwinter 已提交
101
  });
102
  return successed;
D
dzhwinter 已提交
103 104
}

105
#ifdef PADDLE_WITH_CUDA
106 107 108 109
void InitCupti() {
#ifdef PADDLE_WITH_CUPTI
  if (FLAGS_multiple_of_cupti_buffer_size == 1) return;
  size_t attrValue = 0, attrValueSize = sizeof(size_t);
G
GaoWei8 已提交
110 111 112 113 114 115 116 117 118 119 120 121
#define MULTIPLY_ATTR_VALUE(attr)                                            \
  {                                                                          \
    PADDLE_ENFORCE_EQ(                                                       \
        !platform::dynload::cuptiActivityGetAttribute(attr, &attrValueSize,  \
                                                      &attrValue),           \
        true, platform::errors::Unavailable("Get cupti attribute failed.")); \
    attrValue *= FLAGS_multiple_of_cupti_buffer_size;                        \
    LOG(WARNING) << "Set " #attr " " << attrValue << " byte";                \
    PADDLE_ENFORCE_EQ(                                                       \
        !platform::dynload::cuptiActivitySetAttribute(attr, &attrValueSize,  \
                                                      &attrValue),           \
        true, platform::errors::Unavailable("Set cupti attribute failed.")); \
122 123 124 125 126 127 128 129 130
  }
  MULTIPLY_ATTR_VALUE(CUPTI_ACTIVITY_ATTR_DEVICE_BUFFER_SIZE);
  MULTIPLY_ATTR_VALUE(CUPTI_ACTIVITY_ATTR_DEVICE_BUFFER_SIZE_CDP);
#if CUDA_VERSION >= 9000
  MULTIPLY_ATTR_VALUE(CUPTI_ACTIVITY_ATTR_PROFILING_SEMAPHORE_POOL_SIZE);
#endif
#undef MULTIPLY_ATTR_VALUE
#endif
}
131
#endif
132

133
void InitDevices() {
134 135 136
// CUPTI attribute should be set before any CUDA context is created (see CUPTI
// documentation about CUpti_ActivityAttribute).
#ifdef PADDLE_WITH_CUDA
137
  InitCupti();
138
#endif
139
  /*Init all available devices by default */
140
  std::vector<int> devices;
141
#if defined(PADDLE_WITH_CUDA) || defined(PADDLE_WITH_HIP)
D
dzhwinter 已提交
142
  try {
143 144
    // use user specified GPUs in single-node multi-process mode.
    devices = platform::GetSelectedDevices();
D
dzhwinter 已提交
145 146
  } catch (const std::exception &exp) {
    LOG(WARNING) << "Compiled with WITH_GPU, but no GPU found in runtime.";
147
  }
148 149 150 151 152 153 154 155
#endif
#ifdef PADDLE_WITH_XPU
  try {
    // use user specified XPUs in single-node multi-process mode.
    devices = platform::GetXPUSelectedDevices();
  } catch (const std::exception &exp) {
    LOG(WARNING) << "Compiled with WITH_XPU, but no XPU found in runtime.";
  }
156 157 158 159 160 161 162 163 164 165 166
#endif
#ifdef PADDLE_WITH_ASCEND_CL
  // NOTE(zhiqiu): use singleton to explicitly init and finalize ACL
  platform::AclInstance::Instance();  // NOLINT
  try {
    // use user specified XPUs in single-node multi-process mode.
    devices = platform::GetSelectedNPUDevices();
  } catch (const std::exception &exp) {
    LOG(WARNING)
        << "Compiled with PADDLE_WITH_ASCEND_CL, but no NPU found in runtime.";
  }
D
dzhwinter 已提交
167
#endif
168
  InitDevices(devices);
D
dzhwinter 已提交
169 170
}

171
void InitDevices(const std::vector<int> devices) {
172 173 174
  std::vector<platform::Place> places;

  for (size_t i = 0; i < devices.size(); ++i) {
175 176 177
    // In multi process multi gpu mode, we may have gpuid = 7
    // but count = 1.
    if (devices[i] < 0) {
178 179 180
      LOG(WARNING) << "Invalid devices id.";
      continue;
    }
181

182
#if defined(PADDLE_WITH_CUDA) || defined(PADDLE_WITH_HIP)
183
    places.emplace_back(platform::CUDAPlace(devices[i]));
184 185 186
#endif
#ifdef PADDLE_WITH_XPU
    places.emplace_back(platform::XPUPlace(devices[i]));
187 188 189
#endif
#ifdef PADDLE_WITH_ASCEND_CL
    places.emplace_back(platform::NPUPlace(devices[i]));
190
#endif
191 192
  }
  places.emplace_back(platform::CPUPlace());
193
#if defined(PADDLE_WITH_CUDA) || defined(PADDLE_WITH_HIP)
194 195
  places.emplace_back(platform::CUDAPinnedPlace());
#endif
196
  platform::DeviceContextPool::Init(places);
Q
qingqing01 已提交
197

198
#ifndef PADDLE_WITH_MKLDNN
T
tensor-tang 已提交
199
  platform::SetNumThreads(FLAGS_paddle_num_threads);
200
#endif
T
tensor-tang 已提交
201

T
tensor-tang 已提交
202
#if !defined(_WIN32) && !defined(__APPLE__) && !defined(__OSX__)
T
tensor-tang 已提交
203
  if (platform::MayIUse(platform::avx)) {
T
tensor-tang 已提交
204 205 206 207
#ifndef __AVX__
    LOG(WARNING) << "AVX is available, Please re-compile on local machine";
#endif
  }
208 209

// Throw some informations when CPU instructions mismatch.
210 211 212 213 214 215
#define AVX_GUIDE(compiletime, runtime)                                  \
  PADDLE_THROW(platform::errors::Unavailable(                            \
      "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."))
216 217

#ifdef __AVX512F__
T
tensor-tang 已提交
218 219
  if (!platform::MayIUse(platform::avx512f)) {
    if (platform::MayIUse(platform::avx2)) {
220
      AVX_GUIDE(AVX512, AVX2);
T
tensor-tang 已提交
221
    } else if (platform::MayIUse(platform::avx)) {
222 223 224 225 226
      AVX_GUIDE(AVX512, AVX);
    } else {
      AVX_GUIDE(AVX512, NonAVX);
    }
  }
T
tensor-tang 已提交
227
#endif
228 229

#ifdef __AVX2__
T
tensor-tang 已提交
230 231
  if (!platform::MayIUse(platform::avx2)) {
    if (platform::MayIUse(platform::avx)) {
232 233 234 235
      AVX_GUIDE(AVX2, AVX);
    } else {
      AVX_GUIDE(AVX2, NonAVX);
    }
T
tensor-tang 已提交
236 237
  }
#endif
238 239

#ifdef __AVX__
T
tensor-tang 已提交
240
  if (!platform::MayIUse(platform::avx)) {
241
    AVX_GUIDE(AVX, NonAVX);
T
tensor-tang 已提交
242
  }
243 244
#endif
#undef AVX_GUIDE
T
tensor-tang 已提交
245 246

#endif
247 248
}

C
chengduo 已提交
249
#ifndef _WIN32
250 251 252
// Description Quoted from
// https://pubs.opengroup.org/onlinepubs/9699919799/basedefs/signal.h.html
const struct {
253
  int signal_number;
254 255 256
  const char *name;
  const char *error_string;
} SignalErrorStrings[] = {
257 258 259 260 261 262
    {SIGSEGV, "SIGSEGV", "Segmentation fault"},
    {SIGILL, "SIGILL", "Illegal instruction"},
    {SIGFPE, "SIGFPE", "Erroneous arithmetic operation"},
    {SIGABRT, "SIGABRT", "Process abort signal"},
    {SIGBUS, "SIGBUS", "Access to an undefined portion of a memory object"},
    {SIGTERM, "SIGTERM", "Termination signal"},
263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281
};

bool StartsWith(const char *str, const char *prefix) {
  size_t len_prefix = strlen(prefix);
  size_t len_str = strlen(str);
  return len_str < len_prefix ? false : memcmp(prefix, str, len_prefix) == 0;
}

const char *ParseSignalErrorString(const std::string &str) {
  for (size_t i = 0;
       i < (sizeof(SignalErrorStrings) / sizeof(*(SignalErrorStrings))); ++i) {
    if (std::string::npos != str.find(SignalErrorStrings[i].name)) {
      return SignalErrorStrings[i].error_string;
    }
  }
  return "Unknown signal";
}

// Handle SIGSEGV, SIGILL, SIGFPE, SIGABRT, SIGBUS, and SIGTERM.
282
void SignalHandle(const char *data, int size) {
C
chengduo 已提交
283
  try {
284 285
    // NOTE1: The glog FailureSignalHandler dumped messages
    //   are deal with line by line
286
    auto signal_msg_dunmer_ptr = SignalMessageDumper::Instance().Get();
287 288 289
    // NOTE2: we only deal with the time info ane signal info,
    //   the stack trace will generated by paddle self
    if (StartsWith(data, "*** Aborted at")) {
290
      *signal_msg_dunmer_ptr << "\n  [TimeInfo: " << std::string(data, size - 1)
291
                             << "]\n";
292 293 294 295 296
    } else if (StartsWith(data, "***")) {
      std::string signal_info(data, size - 1);
      std::string useless_substr("; stack trace:");
      size_t start_pos = signal_info.rfind(useless_substr);
      signal_info.replace(start_pos, useless_substr.length(), "");
297
      *signal_msg_dunmer_ptr << "  [SignalInfo: " << signal_info << "]\n";
298 299 300

      // NOTE3: Final singal error message print.
      // Here does not throw an exception,
301
      // otherwise it will casue "terminate called recursively"
302
      std::ostringstream sout;
303 304 305 306 307 308 309 310 311 312 313
      sout << "\n\n--------------------------------------\n";
      sout << "C++ Traceback (most recent call last):";
      sout << "\n--------------------------------------\n";
      auto traceback = platform::GetCurrentTraceBackString(/*for_signal=*/true);
      if (traceback.empty()) {
        sout
            << "No stack trace in paddle, may be caused by external reasons.\n";
      } else {
        sout << traceback;
      }

314 315 316 317 318 319 320
      sout << "\n----------------------\nError Message "
              "Summary:\n----------------------\n";
      sout << platform::errors::Fatal(
                  "`%s` is detected by the operating system.",
                  ParseSignalErrorString(signal_info))
                  .to_string();
      std::cout << sout.str() << (*signal_msg_dunmer_ptr).str() << std::endl;
321
    }
C
chengduo 已提交
322
  } catch (...) {
323 324 325
    // Since the program has already triggered a system error,
    // no further processing is required here, glog FailureSignalHandler
    // will Kill program by the default signal handler
C
chengduo 已提交
326 327
  }
}
328 329 330 331 332 333 334 335 336 337 338 339 340
#endif  // _WIN32

void DisableSignalHandler() {
#ifndef _WIN32
  for (size_t i = 0;
       i < (sizeof(SignalErrorStrings) / sizeof(*(SignalErrorStrings))); ++i) {
    int signal_number = SignalErrorStrings[i].signal_number;
    struct sigaction sig_action;
    memset(&sig_action, 0, sizeof(sig_action));
    sigemptyset(&sig_action.sa_mask);
    sig_action.sa_handler = SIG_DFL;
    sigaction(signal_number, &sig_action, NULL);
  }
C
chengduo 已提交
341
#endif
342
}
C
chengduo 已提交
343

344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382
#ifdef WITH_WIN_DUMP_DBG
typedef BOOL(WINAPI *MINIDUMP_WRITE_DUMP)(
    IN HANDLE hProcess, IN DWORD ProcessId, IN HANDLE hFile,
    IN MINIDUMP_TYPE DumpType,
    IN CONST PMINIDUMP_EXCEPTION_INFORMATION ExceptionParam,
    OPTIONAL IN PMINIDUMP_USER_STREAM_INFORMATION UserStreamParam,
    OPTIONAL IN PMINIDUMP_CALLBACK_INFORMATION CallbackParam OPTIONAL);
void CreateDumpFile(LPCSTR lpstrDumpFilePathName,
                    EXCEPTION_POINTERS *pException) {
  HANDLE hDumpFile = CreateFile(lpstrDumpFilePathName, GENERIC_WRITE, 0, NULL,
                                CREATE_ALWAYS, FILE_ATTRIBUTE_NORMAL, NULL);
  MINIDUMP_EXCEPTION_INFORMATION dumpInfo;
  dumpInfo.ExceptionPointers = pException;
  dumpInfo.ThreadId = GetCurrentThreadId();
  dumpInfo.ClientPointers = TRUE;
  MINIDUMP_WRITE_DUMP MiniDumpWriteDump_;
  HMODULE hDbgHelp = LoadLibrary("DBGHELP.DLL");
  MiniDumpWriteDump_ =
      (MINIDUMP_WRITE_DUMP)GetProcAddress(hDbgHelp, "MiniDumpWriteDump");
  MiniDumpWriteDump_(GetCurrentProcess(), GetCurrentProcessId(), hDumpFile,
                     MiniDumpWithPrivateReadWriteMemory, &dumpInfo, NULL, NULL);
  CloseHandle(hDumpFile);
}

LONG ApplicationCrashHandler(EXCEPTION_POINTERS *pException) {
  time_t time_seconds = time(0);
  struct tm now_time;
  localtime_s(&now_time, &time_seconds);

  char buf[1024];
  sprintf_s(buf, "C:\\Paddle%04d%02d%02d-%02d%02d%02d.dmp",
            1900 + now_time.tm_year, 1 + now_time.tm_mon, now_time.tm_mday,
            now_time.tm_hour, now_time.tm_min, now_time.tm_sec);

  CreateDumpFile(buf, pException);
  return EXCEPTION_EXECUTE_HANDLER;
}
#endif

Y
Yang Yu 已提交
383
void InitGLOG(const std::string &prog_name) {
384
  std::call_once(glog_init_flag, [&]() {
385 386 387 388 389 390
// glog will not hold the ARGV[0] inside.
// Use strdup to alloc a new string.
#ifdef WITH_WIN_DUMP_DBG
    SetUnhandledExceptionFilter(
        (LPTOP_LEVEL_EXCEPTION_FILTER)ApplicationCrashHandler);
#endif
391
    google::InitGoogleLogging(strdup(prog_name.c_str()));
C
chengduo 已提交
392
#ifndef _WIN32
393 394
    google::InstallFailureSignalHandler();
    google::InstallFailureWriter(&SignalHandle);
C
chengduo 已提交
395
#endif
396
  });
Y
Yang Yu 已提交
397 398
}

D
dzhwinter 已提交
399 400
}  // namespace framework
}  // namespace paddle