import time
import numpy as np
t = time.time()
from mpi4py import MPI
import kokkos.gpu.core as kokkos
comm = MPI.COMM_WORLD
size = comm.Get_size()
rank = comm.Get_rank()
kokkos.start(rank)
t = time.time() - t
print("Kokkos init time: ", t, flush=True)
n_local = 1000000000
N = size * n_local
itemsize = MPI.DOUBLE.Get_size()
if rank == 0:
nbytes = N * itemsize
else:
nbytes = 0
t = time.time()
#Create shared block
win = MPI.Win.Allocate_shared(nbytes, itemsize, comm=comm)
#Create numpy array in shared memory
buf, itemsize = win.Shared_query(0)
assert itemsize == MPI.DOUBLE.Get_size()
import time
if __name__ == '__main__':
#time.sleep(20) #To give me enough time to attach strace
m = 2
def thing(i):
n = 100000
array = np.arange(1, n + 1, dtype='float64')
result = kokkos.reduction(array, i)
return result
for i in range(m):
multiload_contexts[i].load_stub_library("cuda")
multiload_contexts[i].load_stub_library("cudart")
with multiload_contexts[i]:
a = 1
import numpy as np
import kokkos.gpu.core as kokkos
kokkos.start(i)
for i in range(m):
with multiload_contexts[i]:
ctx = thing(i)
print(i, ctx)
for i in range(m):
with multiload_contexts[i]:
b = 1
kokkos.end()
parser.add_argument('-trials',
metavar='trials',
type=int,
help="Number of warmup runs")
args = parser.parse_args()
if __name__ == '__main__':
m = args.m
n_local = 10000**2
N = m * n_local
t = time.time()
import numpy as np
import kokkos.gpu.core as kokkos
kokkos.start(m)
t = time.time() - t
print("Initilize time: ", t, flush=True)
t = time.time()
array = np.arange(1, N + 1, dtype='float64')
result = np.zeros(m, dtype='float64')
t = time.time() - t
print("Initilize array time: ", t, flush=True)
def reduction(array, i):
p = kokkos.dev_copy(array, (int)(np.sqrt(len(array))), i)
result = kokkos.reduction(p, (int)(np.sqrt(len(array))), i)
times = []
import numpy as np import kokkos.gpu.core as kokkos "Test of Kokkos Wrapper" n = 1000000 a = np.arange(1, n + 1, dtype='float64') kokkos.start(0) result = kokkos.reduction(a) print(result) kokkos.end()