def weighted_map(ipix, npix, weights, localsize, comm):
""" Make a map from particles, for quantities like
W(t) = \int dx delta(t, x) w
Parameters
----------
ipix: array_like
weights : array_like
Returns
-------
Wmap, Nmap; distributed maps
Wmap is the weighted map. Nmap is the number of objects
"""
ipix, labels = numpy.unique(ipix, return_inverse=True)
N = numpy.bincount(labels)
weights = numpy.bincount(labels, weights)
#print("shrink to %d from %d" % (len(ipix), len(labels)))
del labels
pairs = numpy.empty(len(ipix) + 1, dtype=[('ipix', 'i4'), ('N', 'i4'), ('weights', 'f8') ])
pairs['ipix'][:-1] = ipix
pairs['weights'][:-1] = weights
pairs['N'][:-1] = N
pairs['ipix'][-1] = npix - 1 # trick to make sure the final length is correct.
pairs['weights'][-1] = 0
pairs['N'][-1] = 0
disa = DistributedArray(pairs, comm=comm)
disa.sort('ipix')
w = disa['ipix'].bincount(weights=disa['weights'].local, local=False, shared_edges=False)
N = disa['ipix'].bincount(weights=disa['N'].local, local=False, shared_edges=False)
if npix - w.cshape[0] != 0:
if comm.rank == 0:
print('padding -- this shouldnt have occured ', npix, w.cshape)
# pad with zeros, since the last few bins can be empty.
ipadding = DistributedArray.cempty((npix - w.cshape[0],), dtype='i4', comm=comm)
fpadding = DistributedArray.cempty((npix - w.cshape[0],), dtype='f8', comm=comm)
fpadding.local[:] = 0
ipadding.local[:] = 0
w = DistributedArray.concat(w, fpadding)
N = DistributedArray.concat(N, ipadding)
w = DistributedArray.concat(w, localsize=localsize)
N = DistributedArray.concat(N, localsize=localsize)
return w.local, N.local
def test_distributed_array_unique_labels(comm):
from nbodykit.utils import DistributedArray, EmptyRank
data = numpy.array(
comm.scatter([
numpy.array([0, 1, 2, 3], 'i4'),
numpy.array([3, 4, 5, 6], 'i4'),
numpy.array([], 'i4'),
numpy.array([6], 'i4'),
]))
da = DistributedArray(data, comm)
da.sort()
labels = da.unique_labels()
assert_array_equal(numpy.concatenate(comm.allgather(labels.local)),
[0, 1, 2, 3, 3, 4, 5, 6, 6])
def test_distributed_array_unique_labels(comm):
from nbodykit.utils import DistributedArray, EmptyRank
data = numpy.array(comm.scatter(
[numpy.array([0, 1, 2, 3], 'i4'),
numpy.array([3, 4, 5, 6], 'i4'),
numpy.array([], 'i4'),
numpy.array([6], 'i4'),
]))
da = DistributedArray(data, comm)
da.sort()
labels = da.unique_labels()
assert_array_equal(
numpy.concatenate(comm.allgather(labels.local)),
[0, 1, 2, 3, 3, 4, 5, 6, 6]
)
def _assign_labels(minid, comm, thresh):
"""
Convert minid to sequential labels starting from 0.
This routine is used to assign halo label to particles with
the same minid.
Halos with less than thresh particles are reclassified to 0.
Parameters
----------
minid : array_like, ('i8')
The minimum particle id of the halo. All particles of a halo
have the same minid
comm : py:class:`MPI.Comm`
communicator. since this is a collective operation
thresh : int
halo with less than thresh particles are merged into halo 0
Returns
-------
labels : array_like ('i8')
The new labels of particles. Note that this is ordered
by the size of halo, with the exception 0 represents all
particles that are in halos that contain less than thresh particles.
"""
from mpi4py import MPI
dtype = numpy.dtype([
('origind', 'u8'),
('fofid', 'u8'),
])
data = numpy.empty(len(minid), dtype=dtype)
# assign origind for recovery of ordering, since
# we need to work in sorted fofid
data['fofid'] = minid
data['origind'] = numpy.arange(len(data), dtype='u4')
data['origind'] += numpy.sum(comm.allgather(len(data))[:comm.rank], dtype='intp') \
data = DistributedArray(data, comm)
# first attempt is to assign fofid for each group
data.sort('fofid')
label = data['fofid'].unique_labels()
N = label.bincount()
# now eliminate those with less than thresh particles
small = N.local <= thresh
Nlocal = label.bincount(local=True)
# mask == True for particles in small halos
mask = numpy.repeat(small, Nlocal)
# globally shift halo id by one
label.local += 1
label.local[mask] = 0
data['fofid'].local[:] = label.local[:]
del label
data.sort('fofid')
data['fofid'].local[:] = data['fofid'].unique_labels().local[:]
# unique_labels may miss the 0 index representing disconnected
# particles if there are no such particles.
# shift the fofoid by 1 in that case.
anysmall = comm.allreduce(small.sum()) != 0
if not anysmall: data['fofid'].local[:] += 1
data.sort('origind')
label = data['fofid'].local.view('i8').copy()
del data
Nhalo0 = max(comm.allgather(label.max())) + 1
Nlocal = numpy.bincount(label, minlength=Nhalo0)
comm.Allreduce(MPI.IN_PLACE, Nlocal, op=MPI.SUM)
# sort the labels by halo size
arg = Nlocal[1:].argsort()[::-1] + 1
if Nhalo0 > 2**31:
dtype = 'i8'
else:
dtype = 'i4'
P = numpy.arange(Nhalo0, dtype=dtype)
P[arg] = numpy.arange(len(arg), dtype=dtype) + 1
label = P[label]
return label
def _assign_labels(minid, comm, thresh):
"""
Convert minid to sequential labels starting from 0.
This routine is used to assign halo label to particles with
the same minid.
Halos with less than thresh particles are reclassified to 0.
Parameters
----------
minid : array_like, ('i8')
The minimum particle id of the halo. All particles of a halo
have the same minid
comm : py:class:`MPI.Comm`
communicator. since this is a collective operation
thresh : int
halo with less than thresh particles are merged into halo 0
Returns
-------
labels : array_like ('i8')
The new labels of particles. Note that this is ordered
by the size of halo, with the exception 0 represents all
particles that are in halos that contain less than thresh particles.
"""
from mpi4py import MPI
dtype = numpy.dtype([
('origind', 'u8'),
('fofid', 'u8'),
])
data = numpy.empty(len(minid), dtype=dtype)
# assign origind for recovery of ordering, since
# we need to work in sorted fofid
data['fofid'] = minid
data['origind'] = numpy.arange(len(data), dtype='u4')
data['origind'] += numpy.sum(comm.allgather(len(data))[:comm.rank], dtype='intp') \
data = DistributedArray(data, comm)
# first attempt is to assign fofid for each group
data.sort('fofid')
label = data['fofid'].unique_labels()
N = label.bincount()
# now eliminate those with less than thresh particles
small = N.local <= thresh
Nlocal = label.bincount(local=True)
# mask == True for particles in small halos
mask = numpy.repeat(small, Nlocal)
# globally shift halo id by one
label.local += 1
label.local[mask] = 0
data['fofid'].local[:] = label.local[:]
del label
data.sort('fofid')
data['fofid'].local[:] = data['fofid'].unique_labels().local[:]
data.sort('origind')
label = data['fofid'].local.view('i8').copy()
del data
Nhalo0 = max(comm.allgather(label.max())) + 1
Nlocal = numpy.bincount(label, minlength=Nhalo0)
comm.Allreduce(MPI.IN_PLACE, Nlocal, op=MPI.SUM)
# sort the labels by halo size
arg = Nlocal[1:].argsort()[::-1] + 1
if Nhalo0 > 2**31:
dtype = 'i8'
else:
dtype = 'i4'
P = numpy.arange(Nhalo0, dtype=dtype)
P[arg] = numpy.arange(len(arg), dtype=dtype) + 1
label = P[label]
return label