/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved.

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 <string.h>  // for strdup

#include <algorithm>
#include <fstream>
#include <iostream>
#include <memory>
#include <set>
#include <stdexcept>
#include <string>

#include "paddle/fluid/framework/operator.h"
#include "paddle/fluid/platform/cpu_helper.h"
#include "paddle/fluid/platform/cpu_info.h"
#include "paddle/fluid/string/split.h"
#ifdef PADDLE_WITH_CUDA
#include "paddle/fluid/platform/cuda_device_guard.h"
#include "paddle/fluid/platform/dynload/cupti.h"
#endif
#include "paddle/fluid/platform/device_context.h"
#include "paddle/fluid/platform/init.h"
#include "paddle/fluid/platform/place.h"
#include "paddle/fluid/string/piece.h"

#ifdef PADDLE_WITH_XPU
#include "paddle/fluid/platform/xpu_header.h"
#include "paddle/fluid/platform/xpu_info.h"
#endif

DECLARE_int32(paddle_num_threads);
DEFINE_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.");

namespace paddle {
namespace platform {

void ParseCommandLineFlags(int argc, char **argv, bool remove) {
  google::ParseCommandLineFlags(&argc, &argv, remove);
}

}  // namespace platform
}  // namespace paddle

namespace paddle {
namespace framework {

#ifdef _WIN32
#define strdup _strdup
#endif

std::once_flag gflags_init_flag;
std::once_flag glog_init_flag;
std::once_flag p2p_init_flag;

bool InitGflags(std::vector<std::string> args) {
  bool successed = false;
  std::call_once(gflags_init_flag, [&]() {
    FLAGS_logtostderr = true;
    // 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.
    args.insert(args.begin(), "dummy");
    std::vector<char *> argv;
    std::string line;
    int argc = args.size();
    for (auto &arg : args) {
      argv.push_back(const_cast<char *>(arg.data()));
      line += arg;
      line += ' ';
    }
    VLOG(1) << "Before Parse: argc is " << argc
            << ", Init commandline: " << line;

    char **arr = argv.data();
    google::ParseCommandLineFlags(&argc, &arr, true);
    successed = true;

    VLOG(1) << "After Parse: argc is " << argc;
  });
  return successed;
}

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_CUDA_SUCCESS(
            cudaDeviceCanAccessPeer(&can_acess, devices[i], devices[j]));
        if (can_acess != 1) {
          VLOG(2) << "Cannot enable P2P access from " << devices[i] << " to "
                  << devices[j];
        } else {
          platform::CUDADeviceGuard guard(devices[i]);
          cudaDeviceEnablePeerAccess(devices[j], 0);
        }
      }
    }
  });
#endif
}

void InitCupti() {
#ifdef PADDLE_WITH_CUPTI
  if (FLAGS_multiple_of_cupti_buffer_size == 1) return;
  size_t attrValue = 0, attrValueSize = sizeof(size_t);
#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.")); \
  }
  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
}

void InitDevices(bool init_p2p) {
  // CUPTI attribute should be set before any CUDA context is created (see CUPTI
  // documentation about CUpti_ActivityAttribute).
  InitCupti();
  /*Init all available devices by default */
  std::vector<int> devices;
#ifdef PADDLE_WITH_CUDA
  try {
    // use user specified GPUs in single-node multi-process mode.
    devices = platform::GetSelectedDevices();
  } catch (const std::exception &exp) {
    LOG(WARNING) << "Compiled with WITH_GPU, but no GPU found in runtime.";
  }
#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.";
  }
#endif
  InitDevices(init_p2p, devices);
}

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

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

#ifdef PADDLE_WITH_CUDA
    places.emplace_back(platform::CUDAPlace(devices[i]));
#endif
#ifdef PADDLE_WITH_XPU
    places.emplace_back(platform::XPUPlace(devices[i]));
#endif
  }
  if (init_p2p) {
    InitP2P(devices);
  }
  places.emplace_back(platform::CPUPlace());
#ifdef PADDLE_WITH_CUDA
  places.emplace_back(platform::CUDAPinnedPlace());
#endif
  platform::DeviceContextPool::Init(places);

#ifndef PADDLE_WITH_MKLDNN
  platform::SetNumThreads(FLAGS_paddle_num_threads);
#endif

#if !defined(_WIN32) && !defined(__APPLE__) && !defined(__OSX__)
  if (platform::MayIUse(platform::avx)) {
#ifndef __AVX__
    LOG(WARNING) << "AVX is available, Please re-compile on local machine";
#endif
  }

// Throw some informations when CPU instructions mismatch.
#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."))

#ifdef __AVX512F__
  if (!platform::MayIUse(platform::avx512f)) {
    if (platform::MayIUse(platform::avx2)) {
      AVX_GUIDE(AVX512, AVX2);
    } else if (platform::MayIUse(platform::avx)) {
      AVX_GUIDE(AVX512, AVX);
    } else {
      AVX_GUIDE(AVX512, NonAVX);
    }
  }
#endif

#ifdef __AVX2__
  if (!platform::MayIUse(platform::avx2)) {
    if (platform::MayIUse(platform::avx)) {
      AVX_GUIDE(AVX2, AVX);
    } else {
      AVX_GUIDE(AVX2, NonAVX);
    }
  }
#endif

#ifdef __AVX__
  if (!platform::MayIUse(platform::avx)) {
    AVX_GUIDE(AVX, NonAVX);
  }
#endif
#undef AVX_GUIDE

#endif
}

#ifndef _WIN32
// Description Quoted from
// https://pubs.opengroup.org/onlinepubs/9699919799/basedefs/signal.h.html
const struct {
  const char *name;
  const char *error_string;
} SignalErrorStrings[] = {
    {"SIGSEGV", "Segmentation fault"},
    {"SIGILL", "Illegal instruction"},
    {"SIGFPE", "Erroneous arithmetic operation"},
    {"SIGABRT", "Process abort signal"},
    {"SIGBUS", "Access to an undefined portion of a memory object"},
    {"SIGTERM", "Termination signal"},
};

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.
void SignalHandle(const char *data, int size) {
  try {
    // NOTE1: The glog FailureSignalHandler dumped messages
    //   are deal with line by line
    auto signal_msg_dunmer_ptr = SignalMessageDumper::Instance().Get();
    // NOTE2: we only deal with the time info ane signal info,
    //   the stack trace will generated by paddle self
    if (StartsWith(data, "*** Aborted at")) {
      *signal_msg_dunmer_ptr << "  [TimeInfo: " << std::string(data, size - 1)
                             << "]\n";
    } 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(), "");
      *signal_msg_dunmer_ptr << "  [SignalInfo: " << signal_info << "]\n";
      // NOTE3: Here does not throw an exception,
      // otherwise it will casue "terminate called recursively"
      auto exp = platform::EnforceNotMet(
          platform::errors::Fatal(
              "A serious error (%s) is detected by the operating system.",
              ParseSignalErrorString(signal_info)),
          __FILE__, __LINE__);
      std::cout << exp.what() << (*signal_msg_dunmer_ptr).str() << std::endl;
    }
  } catch (...) {
    // 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
  }
}
#endif

void InitGLOG(const std::string &prog_name) {
  std::call_once(glog_init_flag, [&]() {
    // glog will not hold the ARGV[0] inside.
    // Use strdup to alloc a new string.
    google::InitGoogleLogging(strdup(prog_name.c_str()));
#ifndef _WIN32
    google::InstallFailureSignalHandler();
    google::InstallFailureWriter(&SignalHandle);
#endif
  });
}

}  // namespace framework
}  // namespace paddle