threading_backend.cc 8.72 KB
Newer Older
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19
/*
 * Licensed to the Apache Software Foundation (ASF) under one
 * or more contributor license agreements.  See the NOTICE file
 * distributed with this work for additional information
 * regarding copyright ownership.  The ASF licenses this file
 * to you 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.
 */

20 21 22 23 24 25 26
/*!
 * \file threading_backend.cc
 * \brief Native threading backend
 */
#include <tvm/runtime/threading_backend.h>
#include <dmlc/logging.h>
#include <thread>
27
#include <algorithm>
28 29
#if defined(__linux__) || defined(__ANDROID__)
#include <fstream>
30
#include <sstream>
31 32
#else
#endif
33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51
#if defined(__linux__)
#include <sched.h>
#endif

namespace tvm {
namespace runtime {
namespace threading {

class ThreadGroup::Impl {
 public:
  Impl(int num_workers,
       std::function<void(int)> worker_callback,
       bool exclude_worker0)
      : num_workers_(num_workers) {
    CHECK_GE(num_workers, 1)
      << "Requested a non-positive number of worker threads.";
    for (int i = exclude_worker0; i < num_workers_; ++i) {
      threads_.emplace_back([worker_callback, i] { worker_callback(i); });
    }
52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75
    InitSortedOrder();
  }
  ~Impl() { Join(); }

  void Join() {
    for (auto& t : threads_) {
      if (t.joinable()) t.join();
    }
  }

  int Configure(AffinityMode mode, int nthreads, bool exclude_worker0) {
    int num_workers_used = 0;
    if (mode == kLittle) {
      num_workers_used = little_count_;
    } else if (mode == kBig) {
      num_workers_used = big_count_;
    } else {
      // use default
      num_workers_used = threading::MaxConcurrency();
    }
    // if a specific number was given, use that
    if (nthreads) {
      num_workers_used = nthreads;
    }
76 77 78 79 80 81
    // if MaxConcurrency restricted the number of workers (e.g., due to
    // hyperthreading), respect the restriction. On CPUs with N logical cores
    // and N/2 physical cores this will set affinity to the first N/2 logical
    // ones.
    num_workers_used = std::min(num_workers_, num_workers_used);

82 83
    const char *val = getenv("TVM_BIND_THREADS");
    if (val == nullptr || atoi(val) == 1) {
84 85
      // Do not set affinity if there are more workers than found cores
      if (sorted_order_.size() >= static_cast<unsigned int>(num_workers_)) {
86
          SetAffinity(exclude_worker0, mode == kLittle);
87 88 89 90 91 92
      } else {
        LOG(WARNING)
          << "The thread affinity cannot be set when the number of workers"
          << "is larger than the number of available cores in the system.";
      }
    }
93
    return num_workers_used;
94 95 96 97
  }

 private:
  // bind worker threads to disjoint cores
98 99
  // if worker 0 is offloaded to master, i.e. exclude_worker0 is true,
  // the master thread is bound to core 0.
100
  void SetAffinity(bool exclude_worker0, bool reverse = false) {
101 102 103 104 105 106 107 108 109 110 111 112 113 114 115
#if defined(__ANDROID__)
#ifndef CPU_SET
#define CPU_SETSIZE 1024
#define __NCPUBITS (8 * sizeof (uint64_t))
    typedef struct {
      uint64_t __bits[CPU_SETSIZE / __NCPUBITS];
    } cpu_set_t;

#define CPU_SET(cpu, cpusetp) \
    ((cpusetp)->__bits[(cpu)/__NCPUBITS] |= (1UL << ((cpu) % __NCPUBITS)))
#define CPU_ZERO(cpusetp) \
    memset((cpusetp), 0, sizeof(cpu_set_t))
#endif
#endif
#if defined(__linux__) || defined(__ANDROID__)
116 117
    CHECK_GE(sorted_order_.size(), num_workers_);

118
    for (unsigned i = 0; i < threads_.size(); ++i) {
119 120 121 122 123 124
      unsigned core_id;
      if (reverse) {
        core_id = sorted_order_[sorted_order_.size() - (i + exclude_worker0) - 1];
      } else {
        core_id = sorted_order_[i + exclude_worker0];
      }
125 126
      cpu_set_t cpuset;
      CPU_ZERO(&cpuset);
127
      CPU_SET(core_id, &cpuset);
128 129 130 131 132 133 134
#if defined(__ANDROID__)
      sched_setaffinity(threads_[i].native_handle(), sizeof(cpu_set_t), &cpuset);
#else
      pthread_setaffinity_np(threads_[i].native_handle(),
          sizeof(cpu_set_t), &cpuset);
#endif
    }
135
    if (exclude_worker0) {  // master thread run task
136 137 138 139 140
      // Master thread will have free migration on needed cores.
      // Typically, the OS will schedule the master thread to run at core 0,
      // which is idle, when other workers are running.
      // See the comment inside SetMasterThreadFullCpuAffinity function to get more detail.
      SetMasterThreadFullCpuAffinity(reverse);
141 142
    }
#endif
143 144
  }

145
  void SetMasterThreadFullCpuAffinity(bool reverse) {
146 147 148
#if defined(__linux__) || defined(__ANDROID__)
    cpu_set_t cpuset;
    CPU_ZERO(&cpuset);
149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173
    // For example, we have 2xA72 + 4xA53 (id is 0 - 5, 4, 5 is A72 big core)
    // And we use config_threadpool API to set we will only use 4xA53.
    // The sorted_order will be [4, 5, 0, 1, 2, 3].
    // When to call this API, we have spawn threads on little cores for other workers
    // in SetAffinity function. And for tvm master thread, it should also run on little cores,
    // not big cores (4, 5).

    // Note: this works well on x86 too. Because x86 doesn't have BIG.LITTLE,
    // our implementation will use kBig mode by default and will let master thread
    // run on intended cores.
    if (reverse) {
      for (int i = 0; i < little_count_; ++i) {
        CPU_SET(sorted_order_[sorted_order_.size() - i - 1], &cpuset);
      }
    } else {
      int big_count = big_count_;
      // Imagine our x86 has cores 0 - 7
      // physical cores are 0 - 3, logical cores are 4 - 7, big_count_ is 8
      // we wish we run on physical cores, not logical cores to avoid contention issue.
#if defined(_M_X64) || defined(__x86_64__)
      big_count /= 2;  // ignore hyper-threading
#endif
      for (int i = 0; i < big_count; ++i) {
        CPU_SET(sorted_order_[i], &cpuset);
      }
174 175 176 177 178 179 180 181 182
    }
#if defined(__ANDROID__)
    sched_setaffinity(pthread_self(), sizeof(cpu_set_t), &cpuset);
#else
    pthread_setaffinity_np(pthread_self(), sizeof(cpu_set_t), &cpuset);
#endif
#endif
  }

183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223
  void InitSortedOrder() {
    unsigned int threads = std::thread::hardware_concurrency();
    std::vector<std::pair <unsigned int, int64_t> > max_freqs;

    for (unsigned int i = 0; i < threads; ++i) {
      int64_t cur_freq = 0;
      #if defined(__linux__) || defined(__ANDROID__)
        std::ostringstream filepath;
        filepath << "/sys/devices/system/cpu/cpu"  << i << "/cpufreq/cpuinfo_max_freq";
        std::ifstream ifs(filepath.str());
        if (!ifs.fail()) {
          if (!(ifs >> cur_freq)) {
            cur_freq = -1;
          }
          ifs.close();
        }
      #endif
      max_freqs.push_back(std::make_pair(i, cur_freq));
    }

    auto fcmpbyfreq = [] (const std::pair<unsigned int, int64_t> &a,
                          const std::pair<unsigned int, int64_t> &b) {
        return a.second == b.second ? a.first < b.first : a.second > b.second;
    };
    std::sort(max_freqs.begin(), max_freqs.end(), fcmpbyfreq);
    int64_t big_freq = max_freqs.begin()->second;
    int64_t little_freq = max_freqs.rbegin()->second;
    for (auto it = max_freqs.begin(); it != max_freqs.end(); it++) {
      sorted_order_.push_back(it->first);
      if (big_freq == it->second) {
        big_count_++;
      }
      if (big_freq != little_freq && little_freq == it->second) {
        little_count_++;
      }
    }
    if (big_count_ + little_count_ != static_cast<int>(sorted_order_.size())) {
      LOG(WARNING) << "more than two frequencies detected!";
    }
  }

224 225
  int num_workers_;
  std::vector<std::thread> threads_;
226 227 228
  std::vector<unsigned int> sorted_order_;
  int big_count_ = 0;
  int little_count_ = 0;
229 230 231 232 233 234 235 236 237
};

ThreadGroup::ThreadGroup(int num_workers,
                         std::function<void(int)> worker_callback,
                         bool exclude_worker0)
  : impl_(new ThreadGroup::Impl(num_workers, worker_callback, exclude_worker0)) {}
ThreadGroup::~ThreadGroup() { delete impl_; }
void ThreadGroup::Join() { impl_->Join(); }

238 239 240 241
int ThreadGroup::Configure(AffinityMode mode, int nthreads, bool exclude_worker0) {
  return impl_->Configure(mode, nthreads, exclude_worker0);
}

242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262
void Yield() {
  std::this_thread::yield();
}

int MaxConcurrency() {
  int max_concurrency = 1;
  const char *val = getenv("TVM_NUM_THREADS");
  if (val == nullptr) {
    val = getenv("OMP_NUM_THREADS");
  }
  if (val != nullptr) {
    max_concurrency = atoi(val);
  } else {
    max_concurrency = std::thread::hardware_concurrency();
#if defined(_M_X64) || defined(__x86_64__)
    max_concurrency /= 2;  // ignore hyper-threading
#endif
  }
  return std::max(max_concurrency, 1);
}

263

264 265 266
}  // namespace threading
}  // namespace runtime
}  // namespace tvm