src/channel.cu


1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51

/*************************************************************************
 * Copyright (c) 2015-2018, NVIDIA CORPORATION. All rights reserved.
 *
 * See LICENSE.txt for license information
 ************************************************************************/

#include "channel.h"
#include "param.h"

NCCL_PARAM(Buffsize, "BUFFSIZE", DEFAULT_BUFFER_SIZE_BYTES);

ncclResult_t initChannel(struct ncclComm* comm, int channelid) {
  struct ncclChannel* channel = comm->channels+channelid;
  channel->id = channelid;

  // Setup intermediate buffering
  channel->buffSize = ncclParamBuffsize();

  // Ring index to user rank table.
  NCCLCHECK(ncclCudaCalloc(&channel->ring.devUserRanks, comm->nRanks));
  NCCLCHECK(ncclCalloc(&channel->ring.userRanks, comm->nRanks));

  // Communication structures with peers.
  NCCLCHECK(ncclCudaCalloc(&channel->devPeers, comm->nRanks));
  NCCLCHECK(ncclCalloc(&channel->peers, comm->nRanks));
  for (size_t i=0; i<comm->nRanks; ++i) {
    channel->peers[i].send.comm = comm;
    channel->peers[i].recv.comm = comm;
  }

  // Per-channel operation list.
  NCCLCHECK(ncclCudaHostAlloc((void**)&channel->collectives, (void**)&channel->devCollectives, sizeof(struct ncclColl)*NCCL_MAX_OPS));
  return ncclSuccess;
}

ncclResult_t freeChannel(struct ncclChannel* channel, int nRanks) {
  // Operation list
  NCCLCHECK(ncclCudaHostFree(channel->collectives));

  // Free Ring index to rank tables
  free(channel->ring.userRanks);
  CUDACHECK(cudaFree(channel->ring.devUserRanks));

  // Free transport proxy resources
  for (int r=0; r<nRanks; r++) {
    struct ncclPeer* peer = channel->peers+r;
    if (peer->send.transportResources) NCCLCHECK(peer->send.transportComm->free(peer->send.transportResources));
    if (peer->recv.transportResources) NCCLCHECK(peer->recv.transportComm->free(peer->recv.transportResources));
  }
  return ncclSuccess;
}