# Process 0 creates the test matrices for each group which are identical up to
# rescaling. The rescaling is such that the greatest eigenvalue is equal to the
# group index.
if rank == 0:
mat = np.random.standard_normal((n, n))
mat = mat + mat.T
mat = mat / la.eigvalsh(mat).max()
for i in range(world_comm.size):
np.save("testarr_%i.npy" % i, mat * (i + 1))
world_comm.Barrier()
# Create the Process Context
context = pscore.ProcessContext(comm=new_comm, gridsize=[2, 2])
# Create a distributed matrix and find its eigenvalues
dm = pscore.DistributedMatrix.from_npy("testarr_%i.npy" % group_index,
context=context)
evals, evecs = psr.pdsyevd(dm)
# Print out results for testing.
for i in range(world_comm.size):
world_comm.Barrier()
if rank == i:
print "Group %i. Group rank %i. Rank %i. " % (group_index, group_rank,
rank)
print evals[-10:]
print
pscore._blocksize = [b, b]
# Process 0 creates the test matrices for each group which are identical up to
# rescaling. The rescaling is such that the greatest eigenvalue is equal to the
# group index.
if rank == 0:
mat = np.random.standard_normal((n, n))
mat = mat + mat.T
mat = mat / la.eigvalsh(mat).max()
for i in range(world_comm.size):
np.save("testarr_%i.npy" % i, mat * (i + 1))
world_comm.Barrier()
# Create the Process Context
context = pscore.ProcessContext(comm=new_comm, gridsize=[2, 2])
# Create a distributed matrix and find its eigenvalues
dm = pscore.DistributedMatrix.from_npy("testarr_%i.npy" % group_index, context=context)
evals, evecs = psr.pdsyevd(dm)
# Print out results for testing.
for i in range(world_comm.size):
world_comm.Barrier()
if rank == i:
print "Group %i. Group rank %i. Rank %i. " % (group_index, group_rank, rank)
print evals[-10:]
print
def execute(self, nprocesses=1):
# nprocesses doesn't do anything.
params = self.params
# Matrix blocking.
blocksize = (params["blocksize"], params["blocksize"])
# Set up mpi and the process grid.
nproc = comm.Get_size()
rank = comm.Get_rank()
# Number of processors per side on grid.
# usually this should be square.
npx = int(nproc ** 0.5) # Round down intensionally.
npy = npx
pscore.initmpi(gridsize=[npx, npy], blocksize=blocksize)
# Make parent directory and write parameter file.
kiyopy.utils.mkparents(params["output_root"])
parse_ini.write_params(params, params["output_root"] + "params.ini", prefix=prefix)
in_root = params["input_root"]
out_root = params["output_root"]
# Figure out what the band names are.
bands = params["bands"]
for pol_str in params["polarizations"]:
for band in bands:
noise_fname = in_root + "noise_inv_" + pol_str + "_" + repr(band) + ".npy"
if self.feedback > 1 and rank == 0:
print "Using noise inverse: " + noise_fname
# Read the array header as there is some information it it we
# need.
shape, fortran_order, dtype, header_length = npyutils.read_header_data(noise_fname)
if len(shape) != 6 or fortran_order or dtype != "<f8":
msg = "Noise matrix header not as expected."
raise ce.DataError(msg)
if shape[0] != shape[3] or shape[1] != shape[4] or shape[2] != shape[5]:
msg = "Noise matrix not square."
raise ce.DataError(msg)
n = shape[0] * shape[1] * shape[2]
mat_shape = (n, n)
Mat = pscore.DistributedMatrix.from_npy(noise_fname, blocksize=blocksize, shape_override=mat_shape)
# Replace with a Scalapack call.
evals, evects = psroutine.pdsyevd(Mat, destroy=True)
# evals = np.zeros(n)
# evects = Mat
# Now construct meta data for these objects.
evals = algebra.make_mat(evals, axis_names=("mode",), row_axes=(0,), col_axes=(0,))
evects_mat_shape = (shape[0], shape[1], shape[2], n)
evects_mat_axis_names = ("freq", "ra", "dec", "mode")
evects_mat_rows = (0, 1, 2)
evects_mat_cols = (3,)
evects_mat_info = {
"type": "mat",
"axes": evects_mat_axis_names,
"rows": evects_mat_rows,
"cols": evects_mat_cols,
}
# Figure out file names:
evects_fname = out_root + "noise_evects_" + pol_str + "_" + repr(band) + ".npy"
evals_fname = out_root + "noise_evalsinv_" + pol_str + "_" + repr(band) + ".npy"
# Write out.
if rank == 0:
if self.feedback > 1:
print "Writing eigenvectors to " + evects_fname
# Eigenvalues.
algebra.save(evals_fname, evals)
# Eigenvector meta data.
meta_fid = open(evects_fname + ".meta", "w")
try:
meta_fid.write(repr(evects_mat_info))
finally:
meta_fid.close()
# Eigenvectors.
evects.to_npy(evects_fname, fortran_order=False, shape_override=evects_mat_shape)
