gpucomm = collectives.GpuComm(_local_id,_local_size,_local_rank)
return gpucomm
if __name__ == '__main__':
comm = get_internode_comm()
rank=comm.rank
device='cuda'+str(rank)
size=comm.size
from test_exchanger import init_device, clean_device
_,ctx,arr,shared_x,shared_xx = init_device(device=device)
gpucomm = get_intranode_comm(rank,size, ctx)
if rank==0: print 'original array %s' % arr
# prepare nccl32 exchanger
from exchanger_strategy import Exch_nccl32
exch = Exch_nccl32(intercomm=comm, intracomm=gpucomm, avg=False)
exch.prepare(ctx, [shared_x])
exch.exchange()
import sys
sys.path.append('../../lib/base/')
device = sys.argv[1]
from mpi4py import MPI
comm = MPI.COMM_WORLD
rank = comm.rank
size = comm.size
# device='gpu'+str(rank)
from test_exchanger import init_device, clean_device
drv, ctx, arr, shared_x, shared_xx = init_device(device=device)
if rank == 0: print 'original array %s' % arr
# prepare copper exchanger
from exchanger_strategy import Exch_copper
exch = Exch_copper(comm, avg=False)
exch.prepare(ctx, drv, [shared_x])
exch.exchange()
if rank == 0: print 'copper summation: %s' % shared_x.get_value()
# prepare ar exchanger
from exchanger_strategy import Exch_allreduce
exch = Exch_allreduce(comm, avg=False)