def test_add(self):
"""See if binary ufunc works for a LocalArray."""
d = Distribution.from_shape(comm=self.comm, shape=(16, 16))
a = LocalArray(d, dtype='int32')
b = LocalArray(d, dtype='int32')
a.fill(1)
b.fill(1)
c = localarray.add(a, b)
self.assertTrue(np.all(c.ndarray == 2))
c = localarray.empty_like(a)
c = localarray.add(a, b, c)
self.assertTrue(np.all(c.ndarray == 2))
d0 = Distribution.from_shape(comm=self.comm, shape=(16, 16))
d1 = Distribution.from_shape(comm=self.comm, shape=(20, 20))
a = LocalArray(d0, dtype='int32')
b = LocalArray(d1, dtype='int32')
self.assertRaises(IncompatibleArrayError, localarray.add, a, b)
d0 = Distribution.from_shape(comm=self.comm, shape=(16, 16))
d1 = Distribution.from_shape(comm=self.comm, shape=(20, 20))
a = LocalArray(d0, dtype='int32')
b = LocalArray(d0, dtype='int32')
c = LocalArray(d1, dtype='int32')
self.assertRaises(IncompatibleArrayError, localarray.add, a, b, c)
def test_slicing(self):
distribution = Distribution.from_shape(self.comm, (16, 16),
dist=('b', 'n'))
a = ones(distribution)
if self.comm.Get_rank() == 0:
dd00 = {
"dist_type": 'b',
"size": 5,
"start": 0,
"stop": 3,
"proc_grid_size": 2,
"proc_grid_rank": 0
}
dd01 = {"dist_type": 'n', "size": 16}
new_distribution = Distribution(self.comm, [dd00, dd01])
rvals = a.global_index.get_slice((slice(5, None), slice(None)),
new_distribution=new_distribution)
assert_array_equal(rvals, np.ones((3, 16)))
elif self.comm.Get_rank() == 1:
dd10 = {
"dist_type": 'b',
"size": 5,
"start": 3,
"stop": 5,
"proc_grid_size": 2,
"proc_grid_rank": 1
}
dd11 = {"dist_type": 'n', "size": 16}
new_distribution = Distribution(self.comm, [dd10, dd11])
rvals = a.global_index.get_slice((slice(None, 10), slice(None)),
new_distribution=new_distribution)
assert_array_equal(rvals, np.ones((2, 16)))
def setUp(self):
self.dist_1d = Distribution.from_shape(comm=self.comm,
shape=(7,), grid_shape=(4,))
self.larr_1d = LocalArray(self.dist_1d, buf=None)
self.dist_2d = Distribution.from_shape(comm=self.comm,
shape=(16, 16), grid_shape=(4, 1))
self.larr_2d = LocalArray(self.dist_2d, buf=None)
def setUp(self):
self.dist_1d = Distribution.from_shape(comm=self.comm,
shape=(7, ),
grid_shape=(4, ))
self.larr_1d = LocalArray(self.dist_1d, buf=None)
self.dist_2d = Distribution.from_shape(comm=self.comm,
shape=(16, 16),
grid_shape=(4, 1))
self.larr_2d = LocalArray(self.dist_2d, buf=None)
def test_arecompatible(self):
"""Test if two DistArrays are compatible."""
d0 = Distribution.from_shape(comm=self.comm, shape=(16,16))
a = LocalArray(d0, dtype='int64')
b = LocalArray(d0, dtype='float32')
self.assertTrue(arecompatible(a,b))
da = Distribution.from_shape(comm=self.comm, shape=(16, 16), dist='c')
a = LocalArray(da, dtype='int64')
db = Distribution.from_shape(comm=self.comm, shape=(16, 16), dist='b')
b = LocalArray(db, dtype='float32')
self.assertFalse(arecompatible(a,b))
def test_asdist_like(self):
"""Test asdist_like for success and failure."""
d = Distribution.from_shape(comm=self.comm,
shape=(16, 16), dist=('b', 'n'))
a = LocalArray(d)
b = LocalArray(d)
new_a = a.asdist_like(b)
self.assertEqual(id(a), id(new_a))
d2 = Distribution.from_shape(comm=self.comm,
shape=(16, 16), dist=('n', 'b'))
a = LocalArray(d)
b = LocalArray(d2)
self.assertRaises(IncompatibleArrayError, a.asdist_like, b)
def test_copy_cbc(self):
distribution = Distribution(comm=self.comm,
dim_data=self.ddpr[self.comm.Get_rank()])
a = LocalArray(distribution, dtype=np.int_)
a.fill(12)
b = a.copy()
assert_localarrays_equal(a, b, check_dtype=True)
def test_rank_coords(self):
""" Test that we can go from rank to coords and back. """
d = Distribution.from_shape(comm=self.comm, shape=(4, 4))
la = LocalArray(d)
max_size = self.comm.Get_size()
for rank in range(max_size):
self.round_trip(la, rank=rank)
def test_crazy(self):
"""Can we go from global to local indices and back for a complex case?"""
distribution = Distribution.from_shape(comm=self.comm,
shape=(10, 100, 20),
dist=('b', 'c', 'n'))
la = LocalArray(distribution)
self.round_trip(la)
def test_cyclic(self):
"""Can we go from global to local indices and back for cyclic?"""
distribution = Distribution.from_shape(comm=self.comm,
shape=(8, 8),
dist=('c', 'n'))
la = LocalArray(distribution)
self.round_trip(la)
def test_fromndarray_like(self):
"""Can we build an array using fromndarray_like?"""
d = Distribution.from_shape(comm=self.comm,
shape=(16, 16), dist=('c', 'b'))
la0 = LocalArray(d, dtype='int8')
la1 = localarray.fromndarray_like(la0.ndarray, la0)
assert_localarrays_equal(la0, la1, check_dtype=True)
def test_copy_bn(self):
distribution = Distribution.from_shape(comm=self.comm,
shape=(16, 16), dist=('b', 'n'))
a = LocalArray(distribution, dtype=np.int_)
a.fill(11)
b = a.copy()
assert_localarrays_equal(a, b, check_dtype=True)
def test_ndenumerate(self):
d = Distribution.from_shape(comm=self.comm,
shape=(16, 16, 2),
dist=('c', 'b', 'n'))
a = LocalArray(d)
for global_inds, value in ndenumerate(a):
a.global_index[global_inds] = 0.0
def test_cyclic_simple(self):
"""Can we compute local indices for a cyclic distribution?"""
distribution = Distribution.from_shape(comm=self.comm,
shape=(10,), dist={0: 'c'})
la = LocalArray(distribution)
self.assertEqual(la.global_shape, (10,))
self.assertEqual(la.grid_shape, (4,))
if la.comm_rank == 0:
gis = (0, 4, 8)
self.assertEqual(la.local_shape, (3,))
calc_result = [la.local_from_global((gi,)) for gi in gis]
result = [(0,), (1,), (2,)]
elif la.comm_rank == 1:
gis = (1, 5, 9)
self.assertEqual(la.local_shape, (3,))
calc_result = [la.local_from_global((gi,)) for gi in gis]
result = [(0,), (1,), (2,)]
elif la.comm_rank == 2:
gis = (2, 6)
self.assertEqual(la.local_shape, (2,))
calc_result = [la.local_from_global((gi,)) for gi in gis]
result = [(0,), (1,)]
elif la.comm_rank == 3:
gis = (3, 7)
self.assertEqual(la.local_shape, (2,))
calc_result = [la.local_from_global((gi,)) for gi in gis]
result = [(0,), (1,)]
self.assertEqual(result, calc_result)
def test_ones(self):
size = self.comm_size
nrows = size * 3
d = Distribution.from_shape(comm=self.comm, shape=(nrows, 20))
a = localarray.ones(d)
expected = np.ones((nrows // size, 20))
assert_array_equal(a.ndarray, expected)
def test_rank_coords(self):
""" Test that we can go from rank to coords and back. """
d = Distribution.from_shape(comm=self.comm, shape=(4, 4))
la = LocalArray(d)
max_size = self.comm.Get_size()
for rank in range(max_size):
self.round_trip(la, rank=rank)
def test_pack_unpack_index(self):
distribution = Distribution.from_shape(comm=self.comm,
shape=(16, 16, 2), dist=('c', 'b', 'n'))
a = LocalArray(distribution)
for global_inds, value in ndenumerate(a):
packed_ind = a.pack_index(global_inds)
self.assertEqual(global_inds, a.unpack_index(packed_ind))
def test_crazy(self):
"""Can we go from global to local indices and back for a complex case?"""
distribution = Distribution.from_shape(comm=self.comm,
shape=(10, 100, 20),
dist=('b', 'c', 'n'))
la = LocalArray(distribution)
self.round_trip(la)
def test_asdist_like(self):
"""Test asdist_like for success and failure."""
d = Distribution.from_shape(comm=self.comm,
shape=(16, 16),
dist=('b', 'n'))
a = LocalArray(d)
b = LocalArray(d)
new_a = a.asdist_like(b)
self.assertEqual(id(a), id(new_a))
d2 = Distribution.from_shape(comm=self.comm,
shape=(16, 16),
dist=('n', 'b'))
a = LocalArray(d)
b = LocalArray(d2)
self.assertRaises(IncompatibleArrayError, a.asdist_like, b)
def test_slicing(self):
distribution = Distribution.from_shape(self.comm,
(16, 16),
dist=('b', 'n'))
a = ones(distribution)
if self.comm.Get_rank() == 0:
dd00 = {"dist_type": 'b',
"size": 5,
"start": 0,
"stop": 3,
"proc_grid_size": 2,
"proc_grid_rank": 0}
dd01 = {"dist_type": 'n',
"size": 16}
new_distribution = Distribution(self.comm, [dd00, dd01])
rvals = a.global_index.get_slice((slice(5, None), slice(None)),
new_distribution=new_distribution)
assert_array_equal(rvals, np.ones((3, 16)))
elif self.comm.Get_rank() == 1:
dd10 = {"dist_type": 'b',
"size": 5,
"start": 3,
"stop": 5,
"proc_grid_size": 2,
"proc_grid_rank": 1}
dd11 = {"dist_type": 'n',
"size": 16}
new_distribution = Distribution(self.comm, [dd10, dd11])
rvals = a.global_index.get_slice((slice(None, 10), slice(None)),
new_distribution=new_distribution)
assert_array_equal(rvals, np.ones((2, 16)))
def test_cyclic_simple(self):
"""Can we compute local indices for a cyclic distribution?"""
distribution = Distribution.from_shape(comm=self.comm,
shape=(10, ),
dist={0: 'c'})
la = LocalArray(distribution)
self.assertEqual(la.global_shape, (10, ))
self.assertEqual(la.grid_shape, (4, ))
if la.comm_rank == 0:
gis = (0, 4, 8)
self.assertEqual(la.local_shape, (3, ))
calc_result = [la.local_from_global((gi, )) for gi in gis]
result = [(0, ), (1, ), (2, )]
elif la.comm_rank == 1:
gis = (1, 5, 9)
self.assertEqual(la.local_shape, (3, ))
calc_result = [la.local_from_global((gi, )) for gi in gis]
result = [(0, ), (1, ), (2, )]
elif la.comm_rank == 2:
gis = (2, 6)
self.assertEqual(la.local_shape, (2, ))
calc_result = [la.local_from_global((gi, )) for gi in gis]
result = [(0, ), (1, )]
elif la.comm_rank == 3:
gis = (3, 7)
self.assertEqual(la.local_shape, (2, ))
calc_result = [la.local_from_global((gi, )) for gi in gis]
result = [(0, ), (1, )]
self.assertEqual(result, calc_result)
def test_negative(self):
"""See if unary ufunc works for a LocalArray."""
d = Distribution.from_shape(comm=self.comm, shape=(16, 16))
a = LocalArray(d, dtype='int32')
a.fill(1)
b = localarray.negative(a)
self.assertTrue(np.all(a.ndarray == -b.ndarray))
b = localarray.empty_like(a)
b = localarray.negative(a, b)
self.assertTrue(np.all(a.ndarray == -b.ndarray))
d2 = Distribution.from_shape(comm=self.comm, shape=(20, 20))
a = LocalArray(d, dtype='int32')
b = LocalArray(d2, dtype='int32')
self.assertRaises(IncompatibleArrayError, localarray.negative, b, a)
def test_pack_unpack_index(self):
distribution = Distribution.from_shape(comm=self.comm,
shape=(16, 16, 2),
dist=('c', 'b', 'n'))
a = LocalArray(distribution)
for global_inds, value in ndenumerate(a):
packed_ind = a.pack_index(global_inds)
self.assertEqual(global_inds, a.unpack_index(packed_ind))
def test_abs(self):
"""See if unary ufunc works for a LocalArray."""
d = Distribution.from_shape(comm=self.comm, shape=(16, 16))
a = LocalArray(d, dtype='int32')
a.fill(-5)
a[2, 3] = 11
b = abs(a)
self.assertTrue(np.all(abs(a.ndarray) == b.ndarray))
def test_copy_bn(self):
distribution = Distribution.from_shape(comm=self.comm,
shape=(16, 16),
dist=('b', 'n'))
a = LocalArray(distribution, dtype=np.int_)
a.fill(11)
b = a.copy()
assert_localarrays_equal(a, b, check_dtype=True)
def setUp(self):
# On Python3, an 'int' gets converted to 'np.int64' on copy,
# so we force the numpy type to start with so we get back
# the same thing.
self.int_type = np.int64
self.distribution = Distribution.from_shape(comm=self.comm,
shape=(4,))
self.int_larr = LocalArray(self.distribution, dtype=self.int_type)
self.int_larr.fill(3)
def setUp(self):
d = Distribution.from_shape(comm=self.comm,
shape=(53, 77),
dist={
0: 'c',
1: 'b'
},
grid_shape=(2, 2))
self.larr = LocalArray(d)
def setUp(self):
d = Distribution.from_shape(comm=self.comm,
shape=(30, 60),
dist={
0: 'n',
1: 'b'
},
grid_shape=(1, 4))
self.larr = LocalArray(d)
def setUp(self):
# On Python3, an 'int' gets converted to 'np.int64' on copy,
# so we force the numpy type to start with so we get back
# the same thing.
self.int_type = np.int64
self.distribution = Distribution.from_shape(comm=self.comm,
shape=(4, ))
self.int_larr = LocalArray(self.distribution, dtype=self.int_type)
self.int_larr.fill(3)
def test_save_2d(self):
d = Distribution.from_shape(comm=self.comm, shape=(11, 15))
la = LocalArray(d)
np_arr = numpy.random.random(la.local_shape)
la.ndarray = np_arr
save_hdf5(self.output_path, la, key=self.key, mode='w')
with self.h5py.File(self.output_path, 'r', driver='mpio',
comm=self.comm) as fp:
for i, v in ndenumerate(la):
self.assertEqual(v, fp[self.key][i])
def test_astype_bn(self):
new_dtype = np.float32
d = Distribution.from_shape(comm=self.comm,
shape=(16, 16), dist=('b', 'n'))
a = LocalArray(d, dtype=np.int_)
a.fill(11)
b = a.astype(new_dtype)
assert_localarrays_allclose(a, b, check_dtype=False)
self.assertEqual(b.dtype, new_dtype)
self.assertEqual(b.ndarray.dtype, new_dtype)
def test_astype_cbc(self):
new_dtype = np.int8
d = Distribution(comm=self.comm,
dim_data=self.ddpr[self.comm.Get_rank()])
a = LocalArray(d, dtype=np.int32)
a.fill(12)
b = a.astype(new_dtype)
assert_localarrays_allclose(a, b, check_dtype=False)
self.assertEqual(b.dtype, new_dtype)
self.assertEqual(b.ndarray.dtype, new_dtype)
def _local_call(comm, local_func_name, ddpr, kwargs):
import distarray.localapi.random as local_random
from distarray.localapi.maps import Distribution
local_func = getattr(local_random, local_func_name)
if len(ddpr):
dim_data = ddpr[comm.Get_rank()]
else:
dim_data = ()
dist = Distribution(dim_data=dim_data, comm=comm)
return proxyize(local_func(distribution=dist, **kwargs))
def test_astype_bn(self):
new_dtype = np.float32
d = Distribution.from_shape(comm=self.comm,
shape=(16, 16),
dist=('b', 'n'))
a = LocalArray(d, dtype=np.int_)
a.fill(11)
b = a.astype(new_dtype)
assert_localarrays_allclose(a, b, check_dtype=False)
self.assertEqual(b.dtype, new_dtype)
self.assertEqual(b.ndarray.dtype, new_dtype)
def test_global_limits_cyclic(self):
"""Find the boundaries of a cyclic distribution"""
d = Distribution.from_shape(comm=self.comm,
shape=(16, 16), dist=('c', 'n'))
a = LocalArray(d)
answers = [(0, 12), (1, 13), (2, 14), (3, 15)]
limits = a.global_limits(0)
self.assertEqual(limits, answers[a.comm_rank])
answers = 4 * [(0, 15)]
limits = a.global_limits(1)
self.assertEqual(limits, answers[a.comm_rank])
def test_fromfunction_complicated(self):
"""Can we build an array using fromfunction and a nontrivial function."""
def f(*global_inds):
return sum(global_inds)
d = Distribution.from_shape(comm=self.comm,
shape=(16, 16), dist=('b', 'c'))
a = localarray.fromfunction(f, d, dtype='int64')
self.assertEqual(a.global_shape, (16,16))
self.assertEqual(a.dtype, np.dtype('int64'))
for global_inds, value in localarray.ndenumerate(a):
self.assertEqual(sum(global_inds), value)
def test_indexing_1(self):
"""Can we get and set local elements for a complex dist?"""
distribution = Distribution.from_shape(comm=self.comm,
shape=(16, 16, 2), dist=('c', 'b', 'n'))
a = LocalArray(distribution)
b = LocalArray(distribution)
for i, value in ndenumerate(a):
a.global_index[i] = 0.0
for i, value in ndenumerate(a):
b.global_index[i] = a.global_index[i]
for i, value in ndenumerate(a):
self.assertEqual(b.global_index[i], a.global_index[i])
self.assertEqual(a.global_index[i], 0.0)
def test_global_limits_block(self):
"""Find the boundaries of a block distribution"""
d = Distribution.from_shape(comm=self.comm,
shape=(16, 16), dist=('b', 'n'))
a = LocalArray(d)
answers = [(0, 3), (4, 7), (8, 11), (12, 15)]
limits = a.global_limits(0)
self.assertEqual(limits, answers[a.comm_rank])
answers = 4 * [(0, 15)]
limits = a.global_limits(1)
self.assertEqual(limits, answers[a.comm_rank])
def test_block_simple(self):
"""Can we compute local indices for a block distribution?"""
distribution = Distribution.from_shape(comm=self.comm, shape=(4, 4))
la = LocalArray(distribution)
self.assertEqual(la.global_shape, (4, 4))
self.assertEqual(la.grid_shape, (4, 1))
self.assertEqual(la.local_shape, (1, 4))
row_result = [(0, 0), (0, 1), (0, 2), (0, 3)]
row = la.comm_rank
calc_row_result = [la.local_from_global((row, col)) for col in
range(la.global_shape[1])]
self.assertEqual(row_result, calc_row_result)
def check_op(self, op):
"""Check unary operation for success.
Check the one- and two-arg ufunc versions as well as the method
version attached to a LocalArray.
"""
d = Distribution.from_shape(comm=self.comm,
shape=(16, 16), dist=('b', 'n'))
x = localarray.ones(d)
y = localarray.ones(d)
with warnings.catch_warnings():
warnings.simplefilter("ignore", category=RuntimeWarning)
result0 = op(x, casting='unsafe') # standard form
op(x, y=y, casting='unsafe') # two-arg form
assert_array_equal(result0.ndarray, y.ndarray)
def test_create_comm_with_list(self):
""" Test that create_comm_with_list() works correctly. """
nodes = list(range(self.max_size))
comm = create_comm_with_list(nodes)
if comm == MPI.COMM_NULL:
# Only proceed when not COMM_NULL.
return
# Run a simple test to confirm this comm works.
size = len(nodes)
nrows = size * 3
d = Distribution.from_shape(comm=comm, shape=(nrows, 20))
a = zeros(d)
expected = numpy.zeros((nrows // size, 20))
assert_array_equal(a.ndarray, expected)
# Cleanup.
comm.Free()
def test_block_complex(self):
"""Can we compute local indices for a block distribution?"""
distribution = Distribution.from_shape(comm=self.comm, shape=(8, 2))
la = LocalArray(distribution)
self.assertEqual(la.global_shape, (8, 2))
self.assertEqual(la.grid_shape, (4, 1))
self.assertEqual(la.local_shape, (2, 2))
expected_lis = [(0, 0), (0, 1), (1, 0), (1, 1)]
if la.comm_rank == 0:
gis = [(0, 0), (0, 1), (1, 0), (1, 1)]
elif la.comm_rank == 1:
gis = [(2, 0), (2, 1), (3, 0), (3, 1)]
elif la.comm_rank == 2:
gis = [(4, 0), (4, 1), (5, 0), (5, 1)]
elif la.comm_rank == 3:
gis = [(6, 0), (6, 1), (7, 0), (7, 1)]
result = [la.local_from_global(gi) for gi in gis]
self.assertEqual(result, expected_lis)
def test_empty(self):
size = self.comm_size
nrows = size * 3
d = Distribution.from_shape(comm=self.comm, shape=(nrows, 20))
a = localarray.empty(d)
self.assertEqual(a.global_shape, (nrows, 20))
def test_cyclic(self):
"""Can we go from global to local indices and back for cyclic?"""
distribution = Distribution.from_shape(comm=self.comm,
shape=(8, 8), dist=('c', 'n'))
la = LocalArray(distribution)
self.round_trip(la)
def test_ndenumerate(self):
d = Distribution.from_shape(comm=self.comm,
shape=(16, 16, 2), dist=('c', 'b', 'n'))
a = LocalArray(d)
for global_inds, value in ndenumerate(a):
a.global_index[global_inds] = 0.0
def test_create_localarray(self):
# regression test for issue #144
dist = Distribution.from_shape(comm=self.comm,
shape=(16, 16), dist=('n', 'b'))
la = LocalArray(dist)
def test_bad_global_limits(self):
""" Test that invalid global_limits fails as expected. """
d = Distribution.from_shape(comm=self.comm, shape=(4, 4))
a = LocalArray(d)
with self.assertRaises(InvalidDimensionError):
a.global_limits(-1)