def ToHypreParVec(vec):
import mfem.par as mfem
from mpi4py import MPI
if mfem.sizeof_HYPRE_Int() == 4:
dtype = 'int32'
else:
dtype = 'int64'
comm = MPI.COMM_WORLD
num_proc = MPI.COMM_WORLD.size
myid = MPI.COMM_WORLD.rank
vec = vec.flatten()
ml = vec.shape[0]
# collect col array to determin partitioning
m_array = comm.allgather(ml)
cols = [0] + list(np.cumsum(m_array))
glob_size = cols[-1]
col_starts = np.array([cols[myid], cols[myid + 1], glob_size], dtype=dtype)
vec = vec.astype('float', copy=False)
v = mfem.HypreParVector(MPI.COMM_WORLD, glob_size, [vec, col_starts])
return v
def ToHypreParCSR(mat,
check_partitioning=False,
verbose=False,
col_starts=None,
assert_non_square_no_col_starts=True):
'''
convert scipy sparse matrix to hypre
vertically stack csr matrix to generte HYPRE Par CSR
Note:
row partitioning is inferred from distribution of input matrix.
column patitioning needs to be specified col_starts.
If col_starts is not given, column partitioning is chosen
to be the same as row partitioning. This works if matrix is square (M = N).
For an aribtrary rectangular matrix, the column partitioning can be
different from the row partitioning. For example, MFEM mixedbilinearfomr
uses different partitiong rules for row and column.
ToDo: change default assert_non_square_no_col_starts to False
'''
from mpi4py import MPI
import mfem.par as mfem
if mfem.sizeof_HYPRE_Int() == 4:
dtype = 'int32'
else:
dtype = 'int64'
comm = MPI.COMM_WORLD
num_proc = MPI.COMM_WORLD.size
myid = MPI.COMM_WORLD.rank
def verbose_message(m, n, nrows, i, j, data, row_starts, col_starts):
for k in range(num_proc):
MPI.COMM_WORLD.Barrier()
if myid == k:
print 'MyID : ', k
print(m,
n), nrows, len(data), i, j, data, row_starts, col_starts
print 'NNZ', np.sum(data != 0.0)
MPI.COMM_WORLD.Barrier()
from scipy.sparse import csr_matrix
if isinstance(mat, csr_matrix):
mat = mat.astype('float')
ml, nl = mat.shape
n_array = comm.allgather(nl)
else:
raise ValueError("Import Matrix Format should be csr or None")
# collect row array to determin the size of matrix
m_array = comm.allgather(ml)
rows = [0] + list(np.cumsum(m_array))
m = rows[-1]
row_starts = np.array([rows[myid], rows[myid + 1], m], dtype=dtype)
n = nl
nrows = ml
i = mat.indptr.astype(dtype)
j = mat.indices.astype(dtype)
data = mat.data
if col_starts is None and m != nl:
col_starts = get_assumed_patitioning(nl)
if assert_non_square_no_col_starts:
assert False, "col_starts must be specified for non diagonal array"
if col_starts is None:
col_starts = row_starts.copy()
col_starts[-1] = n
if col_starts[0] > n:
col_starts[0] = n
if col_starts[1] > n:
col_starts[1] = n
col_starts[2] = n
else:
# make sure that dtype is right....
col_starts = np.array(col_starts, dtype=dtype)
if check_partitioning:
ch = get_assumed_patitioning(m)
if (row_starts[0] != ch[0] or row_starts[1] != ch[1]
or nrows != ch[2]):
for k in range(num_proc):
MPI.COMM_WORLD.Barrier()
if myid == k:
print 'MyID : ', k
print ch, nrows, row_starts, col_starts
print 'NNZ', np.sum(data != 0.0)
MPI.COMM_WORLD.Barrier()
raise ValueError("partitioning of input matrix is not correct")
if verbose:
verbose_message(m, n, nrows, i, j, data, row_starts, col_starts)
#
# it seems row_starts and col_starts are both to determin
# which part is treated diagnal element.
#
if (m == n and row_starts[0] == col_starts[0]
and row_starts[1] == col_starts[1]):
# this will cause hypre_CSRMatrixReorder call.
M = mfem.HypreParMatrix(MPI.COMM_WORLD, nrows, m, n,
[i, j, data, col_starts])
M.CopyRowStarts()
M.CopyColStarts()
else:
M = mfem.HypreParMatrix(MPI.COMM_WORLD, nrows, m, n,
[i, j, data, row_starts, col_starts])
M.CopyRowStarts()
M.CopyColStarts()
return M