deeplearning4j/deeplearning4j

/* ******************************************************************************
 *
 *
 * This program and the accompanying materials are made available under the
 * terms of the Apache License, Version 2.0 which is available at
 * https://www.apache.org/licenses/LICENSE-2.0.
 *
 *  See the NOTICE file distributed with this work for additional
 *  information regarding copyright ownership.
 * 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.
 *
 * SPDX-License-Identifier: Apache-2.0
 ******************************************************************************/

//
// @author raver119@gmail.com
//
#include <exceptions/cuda_exception.h>
#include <execution/AffinityManager.h>
#include <execution/LaunchContext.h>
#include <helpers/logger.h>

thread_local int globalThreadToDevice = -1;

namespace sd {
std::mutex AffinityManager::_currentMutex;
std::mutex AffinityManager::_numberMutex;
int AffinityManager::_numberOfDevices = -1;

int AffinityManager::currentDeviceId() {
  // if there's no affinity set - set it now
  if (globalThreadToDevice < 0) {
    // this block must be thread-local
    _currentMutex.lock();

    globalThreadToDevice = _lastDevice++;

    // we need to check if we've got deviceId >= number of actual devices, and reset to zero otherwise
    if (globalThreadToDevice >= numberOfDevices()) {
      globalThreadToDevice = 0;
      _lastDevice = numberOfDevices() > 1 ? 1 : 0;
    }

    _currentMutex.unlock();

    setCurrentNativeDevice(globalThreadToDevice);
  }

  // if we already know affinity - just return it
  if (globalThreadToDevice >= 0) return globalThreadToDevice;

  int dev = 0;
  auto res = cudaGetDevice(&dev);

  if (res != 0) throw cuda_exception::build("cudaGetDevice failed", res);

  return dev;
}

int AffinityManager::currentNativeDeviceId() {
  int dev = 0;
  auto res = cudaGetDevice(&dev);

  if (res != 0) throw cuda_exception::build("cudaGetDevice failed", res);

  return dev;
}

int AffinityManager::numberOfDevices() {
  _numberMutex.lock();
  // we want to cache number of devices
  if (_numberOfDevices <= 0) {
    int dev = 0;
    auto res = cudaGetDeviceCount(&dev);

    if (res != 0) throw cuda_exception::build("cudaGetDeviceCount failed", res);

    _numberOfDevices = dev;
  }
  _numberMutex.unlock();

  return _numberOfDevices;
}

void AffinityManager::setCurrentNativeDevice(int deviceId) {
  auto res = cudaSetDevice(deviceId);
  if (res != 0) throw cuda_exception::build("setCurrentDevice failed", res);
}

void AffinityManager::setCurrentDevice(int deviceId) {
  auto previousDeviceId = globalThreadToDevice;
  if (previousDeviceId >= 0 && LaunchContext::isInitialized()) {
    auto res = cudaStreamSynchronize(*LaunchContext::defaultContext()->getCudaStream());
    if (res != 0) throw cuda_exception::build("setCurrentDevice -> sync failed", res);

    res = cudaStreamSynchronize(*LaunchContext::defaultContext()->getCudaSpecialStream());
    if (res != 0) throw cuda_exception::build("setCurrentDevice -> specialSync failed", res);

    if (deviceId != previousDeviceId) {
      // discard existing stuff
      LaunchContext::releaseBuffers();
    }
  }

  if (deviceId != previousDeviceId) {
    auto res = cudaSetDevice(deviceId);
    if (res != 0) throw cuda_exception::build("cudaSetDevice failed", res);
  }

  // update thread-device affinity
  globalThreadToDevice = deviceId;
}

std::atomic<int> AffinityManager::_lastDevice;  // = std::atomic<int>(initialV);
}  // namespace sd