Usecase: cluster running with 4 processes. User tries creating a 3D distributed array with empty((10, 20, 30), dist=('c', 'b', 'c')). Currently, this will result in a GridShapeError, since 4 has only one factoring without 1's, namely, 2 * 2. If 1's were allowed in the factoring, then this would not yield a GridShapeError, and the factoring would be (probably) (1, 2, 2). The 0th dimension would be non-distributed "cyclic", but at least it would work in many more cases than it does currently.
This would allow DistArray to always work even if the user fires up a cluster with a prime number of engines, for example. Currently, a prime number of engines will always fail for distributing over more than 1 dimension.
Usecase: cluster running with 4 processes. User tries creating a 3D distributed array with
empty((10, 20, 30), dist=('c', 'b', 'c')). Currently, this will result in aGridShapeError, since 4 has only one factoring without 1's, namely, 2 * 2. If 1's were allowed in the factoring, then this would not yield aGridShapeError, and the factoring would be (probably)(1, 2, 2). The 0th dimension would be non-distributed "cyclic", but at least it would work in many more cases than it does currently.This would allow DistArray to always work even if the user fires up a cluster with a prime number of engines, for example. Currently, a prime number of engines will always fail for distributing over more than 1 dimension.