def testExscan(self):
size = self.COMM.Get_size()
rank = self.COMM.Get_rank()
for array, typecode in arrayimpl.loop():
for op in (MPI.SUM, MPI.PROD, MPI.MAX, MPI.MIN):
if skip_op(typecode, op): continue
buf = array(range(size), typecode)
StartWaitFree(
self.COMM.Exscan_init(MPI.IN_PLACE,
buf.as_mpi(),
op)
)
if rank == 1:
for i, value in enumerate(buf):
self.assertEqual(value, i)
elif rank > 1:
max_val = maxvalue(buf)
for i, value in enumerate(buf):
if op == MPI.SUM:
if (i * rank) < max_val:
self.assertAlmostEqual(value, i * rank)
elif op == MPI.PROD:
if (i ** rank) < max_val:
self.assertAlmostEqual(value, i ** rank)
elif op == MPI.MAX:
self.assertEqual(value, i)
elif op == MPI.MIN:
self.assertEqual(value, i)
def testReduce(self):
size = self.COMM.Get_size()
rank = self.COMM.Get_rank()
for array, typecode in arrayimpl.loop():
for op in (MPI.SUM, MPI.PROD, MPI.MAX, MPI.MIN):
if skip_op(typecode, op): continue
for root in range(size):
count = size
if rank == root:
buf = array(range(size), typecode)
sbuf = MPI.IN_PLACE
rbuf = buf.as_mpi()
else:
buf = array(range(size), typecode)
buf2 = array(range(size), typecode)
sbuf = buf.as_mpi()
rbuf = buf2.as_mpi()
StartWaitFree(
self.COMM.Reduce_init(sbuf, rbuf, op, root)
)
if rank == root:
max_val = maxvalue(buf)
for i, value in enumerate(buf):
if op == MPI.SUM:
if (i * size) < max_val:
self.assertAlmostEqual(value, i*size)
elif op == MPI.PROD:
if (i ** size) < max_val:
self.assertAlmostEqual(value, i**size)
elif op == MPI.MAX:
self.assertEqual(value, i)
elif op == MPI.MIN:
self.assertEqual(value, i)
def testGatherv3(self):
size = self.COMM.Get_size()
rank = self.COMM.Get_rank()
for array, typecode in arrayimpl.loop():
for root in range(size):
for count in range(size + 1):
#
sbuf = array(root, typecode, count).as_raw()
rbuf = array(-1, typecode, count * size).as_raw()
sendbuf = sbuf
recvbuf = [rbuf, count]
if rank != root: recvbuf = None
StartWaitFree(
self.COMM.Gatherv_init(sendbuf, recvbuf, root))
if recvbuf is not None:
for v in rbuf:
self.assertEqual(v, root)
#
sbuf = array(root, typecode, count).as_raw()
if rank == root:
rbuf = array(-1, typecode, count * size).as_raw()
else:
rbuf = None
StartWaitFree(self.COMM.Gatherv_init(sbuf, rbuf, root))
if rank == root:
for v in rbuf:
self.assertEqual(v, root)
def testAccumulate(self):
group = self.WIN.Get_group()
size = group.Get_size()
group.Free()
for array, typecode in arrayimpl.loop():
with arrayimpl.test(self):
if unittest.is_mpi_gpu('openmpi', array): continue
if unittest.is_mpi_gpu('mvapich2', array): continue
if typecode in 'FDG': continue
for count in range(self.COUNT_MIN, 10):
for rank in range(size):
with self.subTest(rank=rank, count=count):
ones = array([1]*count, typecode)
sbuf = array(range(count), typecode)
rbuf = array(-1, typecode, count+1)
for op in (
MPI.SUM, MPI.PROD,
MPI.MAX, MPI.MIN,
MPI.REPLACE,
):
self.WIN.Lock(rank)
self.WIN.Put(ones.as_mpi(), rank)
self.WIN.Flush(rank)
r = self.WIN.Raccumulate(sbuf.as_mpi(),
rank, op=op)
r.Wait()
self.WIN.Flush(rank)
r = self.WIN.Rget(rbuf.as_mpi_c(count), rank)
r.Wait()
self.WIN.Unlock(rank)
#
for i in range(count):
self.assertEqual(sbuf[i], i)
self.assertEqual(rbuf[i], op(1, i))
self.assertEqual(rbuf[-1], -1)
def testScatterv3(self):
size = self.COMM.Get_size()
rank = self.COMM.Get_rank()
for array, typecode in arrayimpl.loop():
for root in range(size):
for count in range(size + 1):
#
sbuf = array(root, typecode, count * size).as_raw()
rbuf = array(-1, typecode, count).as_raw()
sendbuf = [sbuf, count]
recvbuf = rbuf
if rank != root: sendbuf = None
StartWaitFree(
self.COMM.Scatterv_init(sendbuf, recvbuf, root))
for v in rbuf:
self.assertEqual(v, root)
#
if rank == root:
sbuf = array(root, typecode, count * size).as_raw()
else:
sbuf = None
rbuf = array(-1, typecode, count).as_raw()
StartWaitFree(self.COMM.Scatterv_init(sbuf, rbuf, root))
for v in rbuf:
self.assertEqual(v, root)
def testReduceScatterBlock(self):
size = self.COMM.Get_size()
rank = self.COMM.Get_rank()
for array, typecode in arrayimpl.loop():
# segfault as of OpenMPI 4.1.1
if unittest.is_mpi_gpu('openmpi', array): continue
for op in (MPI.SUM, MPI.MAX, MPI.MIN, MPI.PROD):
if skip_op(typecode, op): continue
for rcnt in range(1, size+1):
sbuf = array([rank]*rcnt*size, typecode)
rbuf = array(-1, typecode, rcnt)
if op == MPI.PROD:
sbuf = array([rank+1]*rcnt*size, typecode)
self.COMM.Ireduce_scatter_block(sbuf.as_mpi(),
rbuf.as_mpi(),
op).Wait()
max_val = maxvalue(rbuf)
v_sum = (size*(size-1))/2
v_prod = 1
for i in range(1,size+1): v_prod *= i
v_max = size-1
v_min = 0
for i, value in enumerate(rbuf):
if op == MPI.SUM:
if v_sum <= max_val:
self.assertAlmostEqual(value, v_sum)
elif op == MPI.PROD:
if v_prod <= max_val:
self.assertAlmostEqual(value, v_prod)
elif op == MPI.MAX:
self.assertEqual(value, v_max)
elif op == MPI.MIN:
self.assertEqual(value, v_min)
def testNeighborAlltoall(self):
for comm in create_topo_comms(self.COMM):
rsize, ssize = get_neighbors_count(comm)
for array, typecode in arrayimpl.loop():
for v in range(3):
sbuf = array(v, typecode, (ssize, 3))
rbuf = array(-1, typecode, (rsize, 3))
comm.Neighbor_alltoall(sbuf.as_mpi(), rbuf.as_mpi_c(3))
for value in rbuf.flat:
self.assertEqual(value, v)
sbuf = array(v, typecode, (ssize, 3))
rbuf = array(-1, typecode, (rsize, 3))
comm.Neighbor_alltoall(sbuf.as_mpi(), rbuf.as_mpi())
for value in rbuf.flat:
self.assertEqual(value, v)
sbuf = array(v, typecode, (ssize, 3))
rbuf = array(-1, typecode, (rsize, 3))
comm.Neighbor_alltoallv(sbuf.as_mpi_c(3), rbuf.as_mpi_c(3))
for value in rbuf.flat:
self.assertEqual(value, v)
sbuf = array(v, typecode, (ssize, 3))
rbuf = array(-1, typecode, (rsize, 3))
comm.Ineighbor_alltoall(sbuf.as_mpi(),
rbuf.as_mpi()).Wait()
for value in rbuf.flat:
self.assertEqual(value, v)
sbuf = array(v, typecode, (ssize, 3))
rbuf = array(-1, typecode, (rsize, 3))
comm.Ineighbor_alltoallv(sbuf.as_mpi_c(3),
rbuf.as_mpi_c(3)).Wait()
for value in rbuf.flat:
self.assertEqual(value, v)
comm.Free()
def testGather(self):
size = self.COMM.Get_size()
rank = self.COMM.Get_rank()
for array, typecode in arrayimpl.loop():
for root in range(size):
count = root+3
if rank == root:
sbuf = MPI.IN_PLACE
buf = array(-1, typecode, (size, count))
#buf.flat[(rank*count):((rank+1)*count)] = \
# array(root, typecode, count)
s, e = rank*count, (rank+1)*count
for i in range(s, e): buf.flat[i] = root
rbuf = buf.as_mpi()
else:
buf = array(root, typecode, count)
sbuf = buf.as_mpi()
rbuf = None
self.COMM.Igather(sbuf, rbuf, root=root).Wait()
for value in buf.flat:
self.assertEqual(value, root)
if rank == root:
sbuf = None
self.COMM.Igather(sbuf, rbuf, root=root).Wait()
for value in buf.flat:
self.assertEqual(value, root)
def testNeighborAlltoallwBottom(self):
size = self.COMM.Get_size()
for comm in create_topo_comms(self.COMM):
rsize, ssize = get_neighbors_count(comm)
for array, typecode in arrayimpl.loop():
for n in range(1, 4):
for v in range(3):
sbuf = array(v, typecode, (ssize, n))
rbuf = array(-1, typecode, (rsize, n))
saddr = MPI.Get_address(sbuf.as_raw())
raddr = MPI.Get_address(rbuf.as_raw())
sdt, rdt = sbuf.mpidtype, rbuf.mpidtype
sdsp = list(
range(0, ssize * n * sdt.extent, n * sdt.extent))
rdsp = list(
range(0, rsize * n * rdt.extent, n * rdt.extent))
sdsp = [saddr + d for d in sdsp]
rdsp = [raddr + d for d in rdsp]
smsg = [MPI.BOTTOM, ([n] * ssize, sdsp), [sdt] * ssize]
rmsg = (MPI.BOTTOM, ([n] * rsize, rdsp), [rdt] * rsize)
try:
comm.Neighbor_alltoallw(smsg, rmsg)
except NotImplementedError:
self.skipTest('mpi-neighbor_alltoallw')
for value in rbuf.flat:
self.assertEqual(value, v)
sbuf.flat[:] = array(v + 1, typecode, (ssize, n)).flat
try:
comm.Ineighbor_alltoallw(smsg, rmsg).Wait()
except NotImplementedError:
self.skipTest('mpi-neighbor_alltoallw')
for value in rbuf.flat:
self.assertEqual(value, v + 1)
comm.Free()
def testNeighborAllgather(self):
for comm in create_topo_comms(self.COMM):
rsize, ssize = get_neighbors_count(comm)
for array, typecode in arrayimpl.loop():
if unittest.is_mpi_gpu('openmpi', array):
# segfault as of OpenMPI 4.1.1; TODO(leofang): why?
if array.backend == 'numba':
continue
for v in range(3):
sbuf = array(v, typecode, 3)
rbuf = array(-1, typecode, (rsize, 3))
comm.Neighbor_allgather(sbuf.as_mpi(), rbuf.as_mpi())
for value in rbuf.flat:
self.assertEqual(value, v)
sbuf = array(v, typecode, 3)
rbuf = array(-1, typecode, (rsize, 3))
comm.Neighbor_allgatherv(sbuf.as_mpi_c(3),
rbuf.as_mpi_c(3))
for value in rbuf.flat:
self.assertEqual(value, v)
sbuf = array(v, typecode, 3)
rbuf = array(-1, typecode, (rsize, 3))
comm.Ineighbor_allgather(sbuf.as_mpi(),
rbuf.as_mpi()).Wait()
for value in rbuf.flat:
self.assertEqual(value, v)
sbuf = array(v, typecode, 3)
rbuf = array(-1, typecode, (rsize, 3))
comm.Ineighbor_allgatherv(sbuf.as_mpi_c(3),
rbuf.as_mpi_c(3)).Wait()
for value in rbuf.flat:
self.assertEqual(value, v)
comm.Free()
def testAlltoallwBottom(self):
size = self.COMM.Get_size()
rank = self.COMM.Get_rank()
for array, typecode in arrayimpl.loop():
for n in range(1, size + 1):
sbuf = array(n, typecode, (size, n))
rbuf = array(-1, typecode, (size, n))
saddr = MPI.Get_address(sbuf.as_raw())
raddr = MPI.Get_address(rbuf.as_raw())
sdt, rdt = sbuf.mpidtype, rbuf.mpidtype
stypes = [
MPI.Datatype.Create_struct([n], [saddr + d],
[sdt]).Commit()
for d in list(
range(0, size * n * sdt.extent, n * sdt.extent))
]
rtypes = [
MPI.Datatype.Create_struct([n], [raddr + d],
[sdt]).Commit()
for d in list(
range(0, size * n * rdt.extent, n * rdt.extent))
]
smsg = (MPI.BOTTOM, ([1] * size, [0] * size), stypes)
rmsg = (MPI.BOTTOM, ([1] * size, [0] * size), rtypes)
try:
self.COMM.Alltoallw(smsg, rmsg)
except NotImplementedError:
self.skipTest('mpi-alltoallw')
finally:
for t in stypes:
t.Free()
for t in rtypes:
t.Free()
for value in rbuf.flat:
self.assertEqual(value, n)
def testScan(self):
size = self.COMM.Get_size()
rank = self.COMM.Get_rank()
# --
for array, typecode in arrayimpl.loop():
# segfault as of OpenMPI 4.1.1
if unittest.is_mpi_gpu('openmpi', array): continue
for op in (MPI.SUM, MPI.PROD, MPI.MAX, MPI.MIN):
if skip_op(typecode, op): continue
buf = array(range(size), typecode)
self.COMM.Iscan(MPI.IN_PLACE,
buf.as_mpi(),
op).Wait()
max_val = maxvalue(buf)
for i, value in enumerate(buf):
if op == MPI.SUM:
if (i * (rank + 1)) < max_val:
self.assertAlmostEqual(value, i * (rank + 1))
elif op == MPI.PROD:
if (i ** (rank + 1)) < max_val:
self.assertAlmostEqual(value, i ** (rank + 1))
elif op == MPI.MAX:
self.assertEqual(value, i)
elif op == MPI.MIN:
self.assertEqual(value, i)
def testExscan(self):
size = self.COMM.Get_size()
rank = self.COMM.Get_rank()
for array, typecode in arrayimpl.loop():
# segfault as of OpenMPI 4.1.1
if unittest.is_mpi_gpu('openmpi', array): continue
for op in (MPI.SUM, MPI.PROD, MPI.MAX, MPI.MIN):
if skip_op(typecode, op): continue
buf = array(range(size), typecode)
try:
self.COMM.Iexscan(MPI.IN_PLACE,
buf.as_mpi(),
op).Wait()
except NotImplementedError:
self.skipTest('mpi-exscan')
if rank == 1:
for i, value in enumerate(buf):
self.assertEqual(value, i)
elif rank > 1:
max_val = maxvalue(buf)
for i, value in enumerate(buf):
if op == MPI.SUM:
if (i * rank) < max_val:
self.assertAlmostEqual(value, i * rank)
elif op == MPI.PROD:
if (i ** rank) < max_val:
self.assertAlmostEqual(value, i ** rank)
elif op == MPI.MAX:
self.assertEqual(value, i)
elif op == MPI.MIN:
self.assertEqual(value, i)
def arrayimpl_loop_io():
openmpi = unittest.mpi_predicate('openmpi(<5.0.0)')
is_i386 = platform.machine() in ('i386', 'i686')
for array, typecode in arrayimpl.loop():
if unittest.is_mpi_gpu('mvapich2', array): continue
if openmpi and is_i386 and typecode in ('g', 'G'): continue
yield array, typecode
def testScan(self):
size = self.COMM.Get_size()
rank = self.COMM.Get_rank()
# --
for array, typecode in arrayimpl.loop():
for op in (MPI.SUM, MPI.PROD, MPI.MAX, MPI.MIN):
if skip_op(typecode, op): continue
sbuf = array(range(size), typecode)
rbuf = array(0, typecode, size)
StartWaitFree(
self.COMM.Scan_init(sbuf.as_mpi(),
rbuf.as_mpi(),
op)
)
max_val = maxvalue(rbuf)
for i, value in enumerate(rbuf):
if op == MPI.SUM:
if (i * (rank + 1)) < max_val:
self.assertAlmostEqual(value, i * (rank + 1))
elif op == MPI.PROD:
if (i ** (rank + 1)) < max_val:
self.assertAlmostEqual(value, i ** (rank + 1))
elif op == MPI.MAX:
self.assertEqual(value, i)
elif op == MPI.MIN:
self.assertEqual(value, i)
def testReduce(self):
size = self.COMM.Get_size()
rank = self.COMM.Get_rank()
for array, typecode in arrayimpl.loop():
for op in (MPI.SUM, MPI.PROD, MPI.MAX, MPI.MIN):
if skip_op(typecode, op): continue
for root in range(size):
sbuf = array(range(size), typecode)
rbuf = array(-1, typecode, size)
self.COMM.Reduce(sbuf.as_mpi(), rbuf.as_mpi(), op, root)
max_val = maxvalue(rbuf)
for i, value in enumerate(rbuf):
if rank != root:
self.assertEqual(value, -1)
continue
if op == MPI.SUM:
if (i * size) < max_val:
self.assertAlmostEqual(value, i * size)
elif op == MPI.PROD:
if (i**size) < max_val:
self.assertAlmostEqual(value, i**size)
elif op == MPI.MAX:
self.assertEqual(value, i)
elif op == MPI.MIN:
self.assertEqual(value, i)
def testReduce(self):
size = self.COMM.Get_size()
rank = self.COMM.Get_rank()
for array, typecode in arrayimpl.loop():
# segfault as of OpenMPI 4.1.1
if unittest.is_mpi_gpu('openmpi', array): continue
for op in (MPI.SUM, MPI.PROD, MPI.MAX, MPI.MIN):
if skip_op(typecode, op): continue
for root in range(size):
count = size
if rank == root:
buf = array(range(size), typecode)
sbuf = MPI.IN_PLACE
rbuf = buf.as_mpi()
else:
buf = array(range(size), typecode)
buf2 = array(range(size), typecode)
sbuf = buf.as_mpi()
rbuf = buf2.as_mpi()
self.COMM.Ireduce(sbuf, rbuf, op, root).Wait()
if rank == root:
max_val = maxvalue(buf)
for i, value in enumerate(buf):
if op == MPI.SUM:
if (i * size) < max_val:
self.assertAlmostEqual(value, i*size)
elif op == MPI.PROD:
if (i ** size) < max_val:
self.assertAlmostEqual(value, i**size)
elif op == MPI.MAX:
self.assertEqual(value, i)
elif op == MPI.MIN:
self.assertEqual(value, i)
def testAllreduce(self):
size = self.COMM.Get_size()
rank = self.COMM.Get_rank()
for array, typecode in arrayimpl.loop():
# segfault as of OpenMPI 4.1.1
if unittest.is_mpi_gpu('openmpi', array): continue
for op in (MPI.SUM, MPI.MAX, MPI.MIN, MPI.PROD):
if skip_op(typecode, op): continue
sbuf = array(range(size), typecode)
rbuf = array(0, typecode, size)
self.COMM.Iallreduce(sbuf.as_mpi(),
rbuf.as_mpi(),
op).Wait()
max_val = maxvalue(rbuf)
for i, value in enumerate(rbuf):
if op == MPI.SUM:
if (i * size) < max_val:
self.assertAlmostEqual(value, i*size)
elif op == MPI.PROD:
if (i ** size) < max_val:
self.assertAlmostEqual(value, i**size)
elif op == MPI.MAX:
self.assertEqual(value, i)
elif op == MPI.MIN:
self.assertEqual(value, i)
def testReduceScatter(self):
size = self.COMM.Get_size()
rank = self.COMM.Get_rank()
for array, typecode in arrayimpl.loop():
for op in (MPI.SUM, MPI.MAX, MPI.MIN, MPI.PROD):
if skip_op(typecode, op): continue
rcnt = list(range(1, size + 1))
if op == MPI.PROD:
rbuf = array([rank + 1] * sum(rcnt), typecode)
else:
rbuf = array([rank] * sum(rcnt), typecode)
self.COMM.Reduce_scatter(MPI.IN_PLACE, rbuf.as_mpi(), rcnt, op)
max_val = maxvalue(rbuf)
for i, value in enumerate(rbuf):
if i >= rcnt[rank]:
if op == MPI.PROD:
self.assertEqual(value, rank + 1)
else:
self.assertEqual(value, rank)
else:
if op == MPI.SUM:
redval = sum(range(size))
if redval < max_val:
self.assertAlmostEqual(value, redval)
elif op == MPI.PROD:
redval = prod(range(1, size + 1))
if redval < max_val:
self.assertAlmostEqual(value, redval)
elif op == MPI.MAX:
self.assertEqual(value, size - 1)
elif op == MPI.MIN:
self.assertEqual(value, 0)
def testReduceScatterBlock(self):
size = self.COMM.Get_size()
rank = self.COMM.Get_rank()
for array, typecode in arrayimpl.loop():
# one of the ranks would fail as of OpenMPI 4.1.1
if unittest.is_mpi_gpu('openmpi', array): continue
for op in (MPI.SUM, MPI.MAX, MPI.MIN, MPI.PROD):
if skip_op(typecode, op): continue
for rcnt in range(size):
if op == MPI.PROD:
rbuf = array([rank + 1] * rcnt * size, typecode)
else:
rbuf = array([rank] * rcnt * size, typecode)
self.COMM.Reduce_scatter_block(MPI.IN_PLACE, rbuf.as_mpi(),
op)
max_val = maxvalue(rbuf)
for i, value in enumerate(rbuf):
if i >= rcnt:
if op == MPI.PROD:
self.assertEqual(value, rank + 1)
else:
self.assertEqual(value, rank)
else:
if op == MPI.SUM:
redval = sum(range(size))
if redval < max_val:
self.assertAlmostEqual(value, redval)
elif op == MPI.PROD:
redval = prod(range(1, size + 1))
if redval < max_val:
self.assertAlmostEqual(value, redval)
elif op == MPI.MAX:
self.assertEqual(value, size - 1)
elif op == MPI.MIN:
self.assertEqual(value, 0)
def testExscan(self):
size = self.COMM.Get_size()
rank = self.COMM.Get_rank()
for array, typecode in arrayimpl.loop():
for op in (MPI.SUM, MPI.PROD, MPI.MAX, MPI.MIN):
if skip_op(typecode, op): continue
sbuf = array(range(size), typecode)
rbuf = array(0, typecode, size)
try:
self.COMM.Exscan(sbuf.as_mpi(), rbuf.as_mpi(), op)
except NotImplementedError:
self.skipTest('mpi-exscan')
if rank == 1:
for i, value in enumerate(rbuf):
self.assertEqual(value, i)
elif rank > 1:
max_val = maxvalue(rbuf)
for i, value in enumerate(rbuf):
if op == MPI.SUM:
if (i * rank) < max_val:
self.assertAlmostEqual(value, i * rank)
elif op == MPI.PROD:
if (i**rank) < max_val:
self.assertAlmostEqual(value, i**rank)
elif op == MPI.MAX:
self.assertEqual(value, i)
elif op == MPI.MIN:
self.assertEqual(value, i)
def testISendrecv(self):
size = self.COMM.Get_size()
rank = self.COMM.Get_rank()
dest = (rank + 1) % size
source = (rank - 1) % size
try:
self.COMM.Isendrecv(
bytearray(1),
dest,
0,
bytearray(1),
source,
0,
).Wait()
except NotImplementedError:
if MPI.Get_version() >= (4, 0): raise
raise unittest.SkipTest("mpi-isendrecv")
for array, typecode in arrayimpl.loop():
with arrayimpl.test(self):
for s in range(0, size + 1):
with self.subTest(s=s):
sbuf = array(s, typecode, s)
rbuf = array(-1, typecode, s + 1)
self.COMM.Isendrecv(
sbuf.as_mpi(),
dest,
0,
rbuf.as_mpi(),
source,
0,
).Wait()
for value in rbuf[:-1]:
self.assertEqual(value, s)
self.assertEqual(rbuf[-1], -1)
def testSendRecv(self):
size = self.COMM.Get_size()
rank = self.COMM.Get_rank()
for array, typecode in arrayimpl.loop():
with arrayimpl.test(self):
if unittest.is_mpi_gpu('openmpi', array): continue
if unittest.is_mpi_gpu('mvapich2', array): continue
for s in range(0, size + 1):
with self.subTest(s=s):
#
sbuf = array(s, typecode, s)
rbuf = array(-1, typecode, s)
mem = array(0, typecode,
2 * (s + MPI.BSEND_OVERHEAD)).as_raw()
if size == 1:
MPI.Attach_buffer(mem)
rbuf = sbuf
MPI.Detach_buffer()
elif rank == 0:
MPI.Attach_buffer(mem)
self.COMM.Ibsend(sbuf.as_mpi(), 1, 0).Wait()
self.COMM.Bsend(sbuf.as_mpi(), 1, 0)
MPI.Detach_buffer()
self.COMM.Send(sbuf.as_mpi(), 1, 0)
self.COMM.Ssend(sbuf.as_mpi(), 1, 0)
self.COMM.Recv(rbuf.as_mpi(), 1, 0)
self.COMM.Recv(rbuf.as_mpi(), 1, 0)
self.COMM.Recv(rbuf.as_mpi(), 1, 0)
self.COMM.Recv(rbuf.as_mpi(), 1, 0)
elif rank == 1:
self.COMM.Recv(rbuf.as_mpi(), 0, 0)
self.COMM.Recv(rbuf.as_mpi(), 0, 0)
self.COMM.Recv(rbuf.as_mpi(), 0, 0)
self.COMM.Recv(rbuf.as_mpi(), 0, 0)
MPI.Attach_buffer(mem)
self.COMM.Ibsend(sbuf.as_mpi(), 0, 0).Wait()
self.COMM.Bsend(sbuf.as_mpi(), 0, 0)
MPI.Detach_buffer()
self.COMM.Send(sbuf.as_mpi(), 0, 0)
self.COMM.Ssend(sbuf.as_mpi(), 0, 0)
else:
rbuf = sbuf
for value in rbuf:
self.assertEqual(value, s)
#
rank = self.COMM.Get_rank()
sbuf = array(s, typecode, s)
rbuf = array(-1, typecode, s)
rreq = self.COMM.Irecv(rbuf.as_mpi(), rank, 0)
self.COMM.Rsend(sbuf.as_mpi(), rank, 0)
rreq.Wait()
for value in rbuf:
self.assertEqual(value, s)
rbuf = array(-1, typecode, s)
rreq = self.COMM.Irecv(rbuf.as_mpi(), rank, 0)
self.COMM.Irsend(sbuf.as_mpi(), rank, 0).Wait()
rreq.Wait()
for value in rbuf:
self.assertEqual(value, s)
def testGetAccumulate(self):
group = self.WIN.Get_group()
size = group.Get_size()
rank = group.Get_rank()
group.Free()
self.WIN.Fence()
obuf = MPI.Alloc_mem(1)
memzero(obuf)
rbuf = MPI.Alloc_mem(1)
memzero(rbuf)
try:
try:
self.WIN.Get_accumulate([obuf, 0, MPI.BYTE],
[rbuf, 0, MPI.BYTE], rank)
self.WIN.Fence()
finally:
MPI.Free_mem(obuf)
MPI.Free_mem(rbuf)
except NotImplementedError:
self.skipTest('mpi-win-get_accumulate')
self.WIN.Fence()
for array, typecode in arrayimpl.loop():
with arrayimpl.test(self):
if unittest.is_mpi_gpu('openmpi', array): continue
if unittest.is_mpi_gpu('mvapich2', array): continue
if typecode in 'FDG': continue
for count in range(10):
for rank in range(size):
with self.subTest(rank=rank, count=count):
ones = array([1] * count, typecode)
sbuf = array(range(count), typecode)
rbuf = array(-1, typecode, count + 1)
gbuf = array(-1, typecode, count + 1)
for op in (
MPI.SUM,
MPI.PROD,
MPI.MAX,
MPI.MIN,
MPI.REPLACE,
MPI.NO_OP,
):
self.WIN.Lock(rank)
self.WIN.Put(ones.as_mpi(), rank)
self.WIN.Flush(rank)
self.WIN.Get_accumulate(sbuf.as_mpi(),
rbuf.as_mpi_c(count),
rank,
op=op)
self.WIN.Flush(rank)
self.WIN.Get(gbuf.as_mpi_c(count), rank)
self.WIN.Flush(rank)
self.WIN.Unlock(rank)
#
for i in range(count):
self.assertEqual(sbuf[i], i)
self.assertEqual(rbuf[i], 1)
self.assertEqual(gbuf[i], op(1, i))
self.assertEqual(rbuf[-1], -1)
self.assertEqual(gbuf[-1], -1)
def testAlltoall(self):
size = self.COMM.Get_size()
rank = self.COMM.Get_rank()
for array, typecode in arrayimpl.loop():
for root in range(size):
sbuf = array(root, typecode, (size, root+1))
rbuf = array( -1, typecode, (size, root+1))
self.COMM.Ialltoall(sbuf.as_mpi(), rbuf.as_mpi_c(root+1)).Wait()
for value in rbuf.flat:
self.assertEqual(value, root)
def testPackSize(self):
for array, typecode in arrayimpl.loop():
with arrayimpl.test(self):
if typecode in self.skipdtype: continue
datatype = array.TypeMap[typecode]
itemsize = datatype.Get_size()
overhead = datatype.Pack_external_size(EXT32, 0)
for count in range(10):
with self.subTest(count=count):
pack_size = datatype.Pack_external_size(EXT32, count)
real_size = pack_size - overhead
def testAllgather(self):
size = self.COMM.Get_size()
rank = self.COMM.Get_rank()
for array, typecode in arrayimpl.loop():
for root in range(size):
sbuf = array(root, typecode, root+1)
rbuf = array( -1, typecode, (size, root+1))
StartWaitFree(
self.COMM.Allgather_init(sbuf.as_mpi(), rbuf.as_mpi())
)
for value in rbuf.flat:
self.assertEqual(value, root)
def testPackSize(self):
for array, typecode in arrayimpl.loop():
with arrayimpl.test(self):
if typecode in self.skipdtype: continue
datatype = array.TypeMap[typecode]
itemsize = datatype.Get_size()
overhead = datatype.Pack_size(0, self.COMM)
for count in range(10):
with self.subTest(count=count):
pack_size = datatype.Pack_size(count, self.COMM)
self.assertEqual(pack_size - overhead,
count * itemsize)
def testBcast(self):
size = self.COMM.Get_size()
rank = self.COMM.Get_rank()
for array, typecode in arrayimpl.loop():
for root in range(size):
if rank == root:
buf = array(root, typecode, root)
else:
buf = array( -1, typecode, root)
self.COMM.Ibcast(buf.as_mpi(), root=root).Wait()
for value in buf:
self.assertEqual(value, root)
def testScatter(self):
size = self.COMM.Get_size()
rank = self.COMM.Get_rank()
for array, typecode in arrayimpl.loop():
for root in range(size):
rbuf = array(-1, typecode, size)
if rank == root:
sbuf = array(root, typecode, (size, size))
else:
sbuf = array([], typecode)
self.COMM.Scatter(sbuf.as_mpi(), rbuf.as_mpi(), root=root)
for value in rbuf:
self.assertEqual(value, root)
