/**
* \brief NCCL implementation of \ref gpucomm_new.
*/
static int comm_new(gpucomm **comm_ptr, gpucontext *ctx,
gpucommCliqueId comm_id, int ndev, int rank) {
gpucomm *comm;
ncclResult_t err;
ASSERT_CTX(ctx);
GA_CHECK(setup_lib(ctx->err));
comm = calloc(1, sizeof(*comm)); // Allocate memory
if (comm == NULL) {
*comm_ptr = NULL; // Set to NULL if failed
return error_sys(ctx->err, "calloc");
}
comm->ctx = (cuda_context *)ctx; // convert to underlying cuda context
// So that context would not be destroyed before communicator
comm->ctx->refcnt++;
cuda_enter(comm->ctx); // Use device
err = ncclCommInitRank(&comm->c, ndev, *((ncclUniqueId *)&comm_id), rank);
cuda_exit(comm->ctx);
TAG_COMM(comm);
if (err != ncclSuccess) {
*comm_ptr = NULL; // Set to NULL if failed
comm_clear(comm);
return error_nccl(ctx->err, "ncclCommInitRank", err);
}
*comm_ptr = comm;
return GA_NO_ERROR;
}
NCCL<Dtype>::NCCL(shared_ptr<Solver<Dtype> > solver, const string& uid)
: GPUParams<Dtype>(solver, getDevice()),
solver_(solver), barrier_() {
this->Configure(solver.get());
Caffe::set_multiprocess(true);
ncclUniqueId nccl_uid;
memcpy(&nccl_uid, &uid[0], NCCL_UNIQUE_ID_BYTES); // NOLINT(caffe/alt_fn)
NCCL_CHECK(ncclCommInitRank(&comm_,
Caffe::solver_count(),
nccl_uid,
Caffe::solver_rank()));
Init();
}