feat[dace][next]: Using the canonicalize_memlet_trees() function
#2351
Conversation
| # into the split transformation. | ||
| # TODO(phimuell): Implement a data cleaner. | ||
| dace_sdutils.canonicalize_memlet_trees(sdfg) | ||
| dace_propagation.propagate_memlets_sdfg(sdfg) |
There was a problem hiding this comment.
Would it be useful to have 2 gt4py wrapper functions?
gt_propagate_memlet_sdfg()that also callscanonicalize_memlet_trees(sdfg)gt_propagate_memlet_map()that also callscanonicalize_memlet_trees_for_scope(map_entry / map_exit)
There was a problem hiding this comment.
I am not fully sure, as we only call canonicalize_memlet_trees() (which seems to have a bug) to fix some other things and in the long run the only reason where it is called is in front of dace simplify or other (known) DaCe transformation.
So I am not sure if it makes sense in the long run.
However I agree that there should be a function that runs the thing for entry and exit but this should life in dace.
I will do a PR tomorrow.
Rather, made some botherline legal behaviour more explicit.
Also removed the reference to `dace_iterator`, which no longer exists from the `pyproject.toml` file.
|
Correctness can be seen here: C2SM/icon4py#928 |
|
I am currently running a performance regression test, but it will take some time until it got through. |
|
Performance is okay, but we have an issue in the ICON4Py CI. |
|
2 participants
Especially older DaCe transformations have trouble handling non standard Memlets. i.e. they expect a particular assignment of
Memlet.dataandMemlet.subset, among them is Memlet propagation.The recently introduced helper function
canonicalize_memlet_trees()allows to transform the SDFG in such a way that it conforms to this format.This PR introduces calls to this function at particular locations, mostly before Memlet propagation happens.
TODO: