def test_basic():
lmax = 10
nside = 8
rank = MPI.COMM_WORLD.Get_rank()
ms = np.arange(rank, lmax + 1, MPI.COMM_WORLD.Get_size(), dtype=np.int32)
order = libsharp.packed_real_order(lmax, ms=ms)
grid = libsharp.healpix_grid(nside)
alm = np.zeros(order.local_size())
if rank == 0:
alm[0] = 1
elif rank == 1:
alm[0] = 1
map = libsharp.synthesis(grid,
order,
np.repeat(alm[None, None, :], 3, 0),
comm=MPI.COMM_WORLD)
assert np.all(map[2, :] == map[1, :]) and np.all(map[1, :] == map[0, :])
map = map[0, 0, :]
if rank == 0:
healpy.mollzoom(map)
from matplotlib.pyplot import show
show()
def distribute_rings_libsharp(mpi_comm, nside, lmax):
"""Create a libsharp map distribution based on rings
Build a libsharp grid object to distribute a HEALPix map
balancing North and South distribution of rings to achieve
the best performance on Harmonic Transforms
Returns the grid object and the pixel indices array in RING ordering
Parameters
---------
mpi_comm : mpi4py.MPI.Comm
mpi4py communicator
nside : int
nside of the map
Returns
-------
grid : libsharp.healpix_grid
libsharp object that includes metadata about HEALPix distributed rings
local_pix : np.ndarray
integer array of local pixel indices in the current MPI process in RING
ordering
"""
import libsharp
nrings = 4 * nside - 1 # four missing pixels
# ring indices are 1-based
ring_indices_emisphere = np.arange(2 * nside, dtype=np.int32) + 1
local_ring_indices = ring_indices_emisphere[mpi_comm.rank::mpi_comm.size]
# to improve performance, symmetric rings north/south need to be in the same rank
# therefore we use symmetry to create the full ring indexing
if local_ring_indices[-1] == 2 * nside:
# has equator ring
local_ring_indices = np.concatenate(
[local_ring_indices[:-1], nrings - local_ring_indices[::-1] + 1])
else:
# does not have equator ring
local_ring_indices = np.concatenate(
[local_ring_indices, nrings - local_ring_indices[::-1] + 1])
libsharp_grid = libsharp.healpix_grid(nside, rings=local_ring_indices)
# returns start index of the ring and number of pixels
startpix, ringpix, _, _, _ = hp.ringinfo(nside,
local_ring_indices.astype(int))
local_npix = libsharp_grid.local_size()
local_pixels = expand_pix(startpix, ringpix, local_npix).astype(int)
local_m_indices = np.arange(mpi_comm.rank,
lmax + 1,
mpi_comm.size,
dtype=np.int32)
libsharp_order = libsharp.packed_real_order(lmax, ms=local_m_indices)
return local_pixels, libsharp_grid, libsharp_order
def distribute_rings(nside, rank, n_mpi_processes):
"""Create a libsharp map distribution based on rings
Build a libsharp grid object to distribute a HEALPix map
balancing North and South distribution of rings to achieve
the best performance on Harmonic Transforms
Returns the grid object and the pixel indices array in RING ordering
Parameters
---------
nside : int
HEALPix NSIDE parameter of the distributed map
rank, n_mpi_processes, ints
rank of the current MPI process and total number of processes
Returns
-------
grid : libsharp.healpix_grid
libsharp object that includes metadata about HEALPix distributed rings
local_pix : np.ndarray
integer array of local pixel indices in the current MPI process in RING
ordering
"""
if libsharp is None:
raise RuntimeError('libsharp not available')
autotimer = timing.auto_timer()
nrings = 4 * nside - 1 # four missing pixels
# ring indices are 1-based
ring_indices_emisphere = np.arange(2 * nside, dtype=np.int32) + 1
local_ring_indices = ring_indices_emisphere[rank::n_mpi_processes]
# to improve performance, symmetric rings north/south need to be in the same rank
# therefore we use symmetry to create the full ring indexing
if local_ring_indices[-1] == 2 * nside:
# has equator ring
local_ring_indices = np.concatenate(
[local_ring_indices[:-1],
nrings - local_ring_indices[::-1] + 1]
)
else:
# does not have equator ring
local_ring_indices = np.concatenate(
[local_ring_indices,
nrings - local_ring_indices[::-1] + 1]
)
grid = libsharp.healpix_grid(nside, rings=local_ring_indices)
return grid, local_ring_indices
def distribute_rings(nside, rank, n_mpi_processes):
"""Create a libsharp map distribution based on rings
Build a libsharp grid object to distribute a HEALPix map
balancing North and South distribution of rings to achieve
the best performance on Harmonic Transforms
Returns the grid object and the pixel indices array in RING ordering
Parameters
---------
nside : int
HEALPix NSIDE parameter of the distributed map
rank, n_mpi_processes, ints
rank of the current MPI process and total number of processes
Returns
-------
grid : libsharp.healpix_grid
libsharp object that includes metadata about HEALPix distributed rings
local_pix : np.ndarray
integer array of local pixel indices in the current MPI process in RING
ordering
"""
if libsharp is None:
raise RuntimeError('libsharp not available')
autotimer = timing.auto_timer()
nrings = 4 * nside - 1 # four missing pixels
# ring indices are 1-based
ring_indices_emisphere = np.arange(2 * nside, dtype=np.int32) + 1
local_ring_indices = ring_indices_emisphere[rank::n_mpi_processes]
# to improve performance, symmetric rings north/south need to be in the same rank
# therefore we use symmetry to create the full ring indexing
if local_ring_indices[-1] == 2 * nside:
# has equator ring
local_ring_indices = np.concatenate(
[local_ring_indices[:-1], nrings - local_ring_indices[::-1] + 1])
else:
# does not have equator ring
local_ring_indices = np.concatenate(
[local_ring_indices, nrings - local_ring_indices[::-1] + 1])
grid = libsharp.healpix_grid(nside, rings=local_ring_indices)
return grid, local_ring_indices
def test_basic():
lmax = 10
nside = 8
rank = MPI.COMM_WORLD.Get_rank()
ms = np.arange(rank, lmax + 1, MPI.COMM_WORLD.Get_size(), dtype=np.int32)
order = libsharp.packed_real_order(lmax, ms=ms)
grid = libsharp.healpix_grid(nside)
alm = np.zeros(order.local_size())
if rank == 0:
alm[0] = 1
elif rank == 1:
alm[0] = 1
map = libsharp.synthesis(grid, order, np.repeat(alm[None, None, :], 3, 0), comm=MPI.COMM_WORLD)
assert np.all(map[2, :] == map[1, :]) and np.all(map[1, :] == map[0, :])
map = map[0, 0, :]
print(rank, "shape", map.shape)
print(rank, "mean", map.mean())
def test_basic():
lmax = 10
nside = 8
rank = MPI.COMM_WORLD.Get_rank()
ms = np.arange(rank, lmax + 1, MPI.COMM_WORLD.Get_size(), dtype=np.int32)
order = libsharp.packed_real_order(lmax, ms=ms)
grid = libsharp.healpix_grid(nside)
alm = np.zeros(order.local_size())
if rank == 0:
alm[0] = 1
elif rank == 1:
alm[0] = 1
map = libsharp.synthesis(grid,
order,
np.repeat(alm[None, None, :], 3, 0),
comm=MPI.COMM_WORLD)
assert np.all(map[2, :] == map[1, :]) and np.all(map[1, :] == map[0, :])
map = map[0, 0, :]
print(rank, "shape", map.shape)
print(rank, "mean", map.mean())
def test_basic():
lmax = 10
nside = 8
rank = MPI.COMM_WORLD.Get_rank()
ms = np.arange(rank, lmax + 1, MPI.COMM_WORLD.Get_size(), dtype=np.int32)
order = libsharp.packed_real_order(lmax, ms=ms)
grid = libsharp.healpix_grid(nside)
alm = np.zeros(order.local_size())
if rank == 0:
alm[0] = 1
elif rank == 1:
alm[0] = 1
map = libsharp.synthesis(grid, order, np.repeat(alm[None, None, :], 3, 0), comm=MPI.COMM_WORLD)
assert np.all(map[2, :] == map[1, :]) and np.all(map[1, :] == map[0, :])
map = map[0, 0, :]
if rank == 0:
healpy.mollzoom(map)
from matplotlib.pyplot import show
show()
local_ring_indices = np.concatenate(
[local_ring_indices[:-1],
nrings - local_ring_indices[::-1] + 1]
)
else:
# does not have equator ring
local_ring_indices = np.concatenate(
[local_ring_indices,
nrings - local_ring_indices[::-1] + 1]
)
print("rank", rank, "n_rings", len(local_ring_indices))
if not mpi:
local_ring_indices = None
grid = libsharp.healpix_grid(nside, rings=local_ring_indices)
# returns start index of the ring and number of pixels
startpix, ringpix, _, _, _ = hp.ringinfo(nside, local_ring_indices.astype(np.int64))
local_npix = grid.local_size()
def expand_pix(startpix, ringpix, local_npix):
"""Turn first pixel index and number of pixel in full array of pixels
to be optimized with cython or numba
"""
local_pix = np.empty(local_npix, dtype=np.int64)
i = 0
for start, num in zip(startpix, ringpix):
local_pix[i:i+num] = np.arange(start, start+num)
# therefore we use symmetry to create the full ring indexing
if local_ring_indices[-1] == 2 * nside:
# has equator ring
local_ring_indices = np.concatenate(
[local_ring_indices[:-1], nrings - local_ring_indices[::-1] + 1])
else:
# does not have equator ring
local_ring_indices = np.concatenate(
[local_ring_indices, nrings - local_ring_indices[::-1] + 1])
print("rank", rank, "n_rings", len(local_ring_indices))
if not mpi:
local_ring_indices = None
grid = libsharp.healpix_grid(nside, rings=local_ring_indices)
# returns start index of the ring and number of pixels
startpix, ringpix, _, _, _ = hp.ringinfo(nside,
local_ring_indices.astype(np.int64))
local_npix = grid.local_size()
def expand_pix(startpix, ringpix, local_npix):
"""Turn first pixel index and number of pixel in full array of pixels
to be optimized with cython or numba
"""
local_pix = np.empty(local_npix, dtype=np.int64)
i = 0