nccl_helper.h 2.4 KB
Newer Older
Y
Yu Yang 已提交
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16
//   Copyright (c) 2018 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.

#pragma once

Y
Yu Yang 已提交
17
#include <thread>
Y
Yu Yang 已提交
18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36
#include <typeindex>
#include "paddle/fluid/platform/dynload/nccl.h"
#include "paddle/fluid/platform/enforce.h"

namespace paddle {
namespace platform {

inline ncclDataType_t ToNCCLDataType(std::type_index type) {
  if (type == typeid(float)) {  // NOLINT
    return ncclFloat;
  } else if (type == typeid(double)) {  // NOLINT
    return ncclDouble;
  } else if (type == typeid(int)) {  // NOLINT
    return ncclInt;
  } else {
    PADDLE_THROW("Not supported");
  }
}

Y
Yu Yang 已提交
37 38 39 40 41 42 43 44 45 46 47 48 49
class NCCLGroupGuard {
 public:
  inline NCCLGroupGuard() {
    mutex().lock();
    PADDLE_ENFORCE(dynload::ncclGroupStart());
  }

  inline ~NCCLGroupGuard() {
    PADDLE_ENFORCE(dynload::ncclGroupEnd());
    mutex().unlock();
  }

 private:
Y
Yu Yang 已提交
50
  static std::mutex &mutex() {
Y
Yu Yang 已提交
51 52 53 54 55
    static std::mutex mtx;
    return mtx;
  }
};

Y
Yu Yang 已提交
56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89
struct NCCLContext {
  std::unique_ptr<CUDADeviceContext> ctx_;
  ncclComm_t comm_;

  explicit NCCLContext(int dev_id)
      : ctx_(new CUDADeviceContext(CUDAPlace(dev_id))) {}

  cudaStream_t stream() const { return ctx_->stream(); }

  int device_id() const {
    return boost::get<platform::CUDAPlace>(ctx_->GetPlace()).device;
  }

  static void InitNCCLContext(std::unordered_map<int, NCCLContext> &contexts,
                              const std::vector<platform::Place> &places) {
    std::vector<ncclComm_t> comms;
    std::vector<int> devs;
    comms.resize(contexts.size());
    devs.reserve(contexts.size());

    for (auto &p : places) {
      devs.push_back(boost::get<platform::CUDAPlace>(p).device);
    }

    PADDLE_ENFORCE(platform::dynload::ncclCommInitAll(
        &comms[0], static_cast<int>(contexts.size()), &devs[0]));

    int i = 0;
    for (auto &dev_id : devs) {
      contexts.at(dev_id).comm_ = comms[i++];
    }
  }
};

Y
Yu Yang 已提交
90 91
}  // namespace platform
}  // namespace paddle