pylops_mpi.DistributedArray#
- class pylops_mpi.DistributedArray(global_shape, base_comm=<mpi4py.MPI.Intracomm object>, partition=Partition.SCATTER, axis=0, local_shapes=None, mask=None, engine='numpy', dtype=<class 'numpy.float64'>)[source]#
Distributed Numpy Arrays
Multidimensional NumPy-like distributed arrays. It brings NumPy arrays to high-performance computing.
Warning
When setting the partition of the DistributedArray to
pylops_mpi.Partition.BROADCAST, it is crucial to be aware that any attempts to make arrays different from rank to rank will be overwritten by the actions of rank 0. This is accomplished internally by broadcasting the content of rank 0 every time a modification of the array is attempted. Such a behaviour does however incur a cost as communication may be not needed if the user ensures not to modify the content of the array in different ranks in a different way. To avoid broadcasting, one can usepylops_mpi.Partition.UNSAFE_BROADCASTinstead.- Parameters:
- global_shape
tupleorint Shape of the global array.
- base_comm
mpi4py.MPI.Comm, optional MPI Communicator over which array is distributed. Defaults to
mpi4py.MPI.COMM_WORLD.- partition
Partition, optional Broadcast, UnsafeBroadcast, or Scatter the array. Defaults to
Partition.SCATTER.- axis
int, optional Axis along which distribution occurs. Defaults to
0.- local_shapes
list, optional List of tuples or integers representing local shapes at each rank.
- mask
list, optional Mask defining subsets of ranks to consider when performing ‘global’ operations on the distributed array such as dot product or norm.
- engine
str, optional Engine used to store array (
numpyorcupy)- dtype
str, optional Type of elements in input array. Defaults to
numpy.float64.
- global_shape
Methods
__init__(global_shape[, base_comm, ...])add(dist_array)Distributed Addition of arrays
add_ghost_cells([cells_front, cells_back])Add ghost cells to the DistributedArray along the axis of partition at each rank.
asarray()Global view of the array
conj()Distributed conj() method
copy()Creates a copy of the DistributedArray
dot(dist_array)Distributed Dot Product
iadd(dist_array)Distributed In-place Addition of arrays
multiply(dist_array)Distributed Element-wise multiplication
norm([ord, axis])Distributed numpy.linalg.norm method
ravel([order])Return a flattened DistributedArray
to_dist(x[, base_comm, partition, axis, ...])Convert A Global Array to a Distributed Array
zeros_like()Creates a copy of the DistributedArray filled with zeros
