collective_helper.cc 5.5 KB
Newer Older
1 2 3 4 5 6 7 8 9 10 11 12 13 14
//   Copyright (c) 2019 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.

15
#if defined(PADDLE_WITH_NCCL)
16
#include "paddle/fluid/platform/collective_helper.h"
17
#include <utility>
18 19 20 21 22 23 24 25 26 27 28 29 30 31 32

namespace paddle {
namespace platform {

class NCCLCommImpl : public NCCLComm {
 public:
  void set_ring_id(int ring_id) { ring_id_ = ring_id; }
  int ring_id() const override { return ring_id_; }

  void set_nranks(int nranks) { nranks_ = nranks; }
  int nranks() const override { return nranks_; }

  void set_rank(int rank) { rank_ = rank; }
  int rank() const override { return rank_; }

33
  int device_id() const override {
34
    return BOOST_GET_CONST(CUDAPlace, dev_ctx_->GetPlace()).device;
35
  }
36

37 38
  void set_comm(ncclComm_t comm) { comm_ = comm; }
  ncclComm_t comm() const override { return comm_; }
39 40 41

  cudaStream_t stream() const override { return dev_ctx_->stream(); }

42 43 44
  void set_dev_ctx(std::unique_ptr<CUDADeviceContext>&& dev_ctx) {
    dev_ctx_ = std::move(dev_ctx);
  }
45
  CUDADeviceContext* dev_context() const override { return dev_ctx_.get(); }
46

47 48 49 50
 private:
  int ring_id_;
  int nranks_;
  int rank_;
51
  ncclComm_t comm_;
52
  std::unique_ptr<CUDADeviceContext> dev_ctx_;
53 54 55 56
};

NCCLComm* NCCLCommContext::CreateNCCLComm(ncclUniqueId* nccl_id, int nranks,
                                          int rank, int dev_id, int ring_id) {
G
GaoWei8 已提交
57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75
  PADDLE_ENFORCE_NOT_NULL(nccl_id,
                          platform::errors::InvalidArgument(
                              "The nccl unique id should not be null."));
  PADDLE_ENFORCE_GT(
      nranks, 1,
      platform::errors::InvalidArgument(
          "Expected nranks > 1. But received nranks is %d.", nranks));
  PADDLE_ENFORCE_GE(rank, 0,
                    platform::errors::InvalidArgument(
                        "Expected rank >= 0. But received rank is %d.", rank));
  PADDLE_ENFORCE_LT(
      rank, nranks,
      platform::errors::InvalidArgument(
          "Expected rank < nranks. But received rank is %d, nranks is %d.",
          rank, nranks));
  PADDLE_ENFORCE_GE(
      dev_id, 0,
      platform::errors::InvalidArgument(
          "Expected dev_id >= 0. But received dev_id is %d.", dev_id));
76 77

  ncclComm_t comm = nullptr;
78 79
  PADDLE_ENFORCE_CUDA_SUCCESS(cudaSetDevice(dev_id));
  PADDLE_ENFORCE_CUDA_SUCCESS(
80 81
      platform::dynload::ncclCommInitRank(&comm, nranks, *nccl_id, rank));

82
  auto* comm_wrapper = AssignNCCLComm(comm, nranks, rank, dev_id, ring_id);
83

84
  VLOG(1) << "nccl communicator of rank " << rank << " in ring " << ring_id
85
          << " has been created on device " << dev_id;
86

87 88 89 90
  std::call_once(once_flag_, []() {
    std::atexit([]() { NCCLCommContext::Instance().ReleaseNCCLComms(); });
  });

91
  return comm_wrapper;
92 93 94 95
}

void NCCLCommContext::CreateAllNCCLComms(const std::vector<int>& dev_ids,
                                         int ring_id) {
G
GaoWei8 已提交
96 97 98 99 100
  PADDLE_ENFORCE_GT(
      dev_ids.size(), 0,
      platform::errors::InvalidArgument("Expected the size of dev_ids > 0. But "
                                        "received the size of dev_ids is %d.",
                                        dev_ids.size()));
101 102 103 104 105 106

  const int kDevices = dev_ids.size();
  ncclComm_t comms[kDevices];
  PADDLE_ENFORCE_CUDA_SUCCESS(platform::dynload::ncclCommInitAll(
      comms, dev_ids.size(), dev_ids.data()));

G
GaoWei8 已提交
107 108 109 110 111
  PADDLE_ENFORCE_EQ(comm_map_.count(ring_id), 0,
                    platform::errors::InvalidArgument(
                        "Expected comm_map_.count(ring_id) = 0. But received "
                        "comm_map_.count(ring_id) is %d.",
                        comm_map_.count(ring_id)));
112
  for (size_t i = 0; i < dev_ids.size(); ++i) {
113 114 115
    AssignNCCLComm(comms[i], dev_ids.size(), i, dev_ids[i], ring_id);
    VLOG(1) << "nccl communicator of rank " << i << " in ring " << ring_id
            << " has been created on device " << dev_ids[i];
116 117 118 119 120
  }

  std::call_once(once_flag_, []() {
    std::atexit([]() { NCCLCommContext::Instance().ReleaseNCCLComms(); });
  });
121 122
}

123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153
NCCLComm* NCCLCommContext::AssignNCCLComm(ncclComm_t comm, int nranks, int rank,
                                          int dev_id, int ring_id) {
  std::unique_ptr<CUDADeviceContext> dev_ctx(
      new CUDADeviceContext(CUDAPlace(dev_id)));

  NCCLCommImpl* c = new NCCLCommImpl;
  c->set_ring_id(ring_id);
  c->set_nranks(nranks);
  c->set_rank(rank);
  c->set_comm(comm);
  c->set_dev_ctx(std::move(dev_ctx));

  comm_map_mutex_.lock();
  if (comm_map_.count(ring_id) == 0) {
    comm_map_.emplace(ring_id, std::map<int, std::unique_ptr<NCCLComm>>());
  }
  auto& dev2comm = comm_map_[ring_id];

  dev2comm.emplace(dev_id, std::unique_ptr<NCCLComm>(c));
  comm_map_mutex_.unlock();

  if (ring_id == 0) {
    auto* dev_ctx = static_cast<platform::CUDADeviceContext*>(
        platform::DeviceContextPool::Instance().Get(
            platform::CUDAPlace(dev_id)));
    dev_ctx->set_nccl_comm(comm);
  }

  return comm_map_[ring_id][dev_id].get();
}

154
void NCCLCommContext::ReleaseNCCLComms() {
155 156 157 158
  for (auto& p : comm_map_) {
    for (auto& q : p.second) {
      q.second.reset();
    }
159 160 161 162 163 164 165
  }
}

}  // namespace platform
}  // namespace paddle

#endif