*sigh*

1.3.0.0

1.2.0.0

[ci skip]

[ci skip]

The permute operation in the PTX backend requires nested calls to `withRemote` in order to copy the defaults array.

Our workload is heavy with deletes, which is not a good workload for the basic hash table.

Use the new `mutate` and `mutateIO` operations, rather than `lookup`/`insert` pairs, which should be more efficient.

In particular, the operation only affects the shape of the innermost dimension. For example, for folds we (now) have the property that:

> shape (fold f z xs) == indexTail (shape xs)

This is required in order to define the instances for AccelerateHS/accelerate#419.