dsl: Move SubDimension thickness evaluation to the thicknesses themselves #2470
Conversation
EdCaunt
requested review from
FabioLuporini,
mloubout,
georgebisbas and
ZoeLeibowitz
| @@ -640,6 +642,11 @@ def _prepare_arguments(self, autotune=None, **kwargs): | |||
| for d in reversed(toposort): | |||
| args.update(d._arg_values(self._dspace[d], grid, **kwargs)) | |||
|
|
|||
| # Process SubDimensionThicknesses | |||
| for p in self.parameters: | |||
There was a problem hiding this comment.
I think this should be triggered from within SubDimension._arg_values itself
There was a problem hiding this comment.
No, the issue is that the SubDimension isn't necessarily available to the operator. You can craft legal Operators which only contain the thickness, not the parent subdimension, hence the reworking in this PR
| @@ -536,6 +535,81 @@ def _arg_check(self, *args, **kwargs): | |||
| # The Dimensions below are exposed in the user API. They can only be created by | |||
| # the user | |||
|
|
|||
| class SubDimensionThickness(DataSymbol): | |||
| """A DataSymbol to represent a thickness of a SubDimension""" | |||
There was a problem hiding this comment.
Thought we didn't do those on single line docstrings?
| rtkn = r_rtkn or 0 | ||
|
|
||
| return {i.name: v for i, v in zip(self._thickness_map, (ltkn, rtkn))} | ||
| # SubDimension thicknesses at runtime are calculated by the thicknesses |
There was a problem hiding this comment.
as I wrote in another comment, this SubDimension should trigger the processing of its own arguments, namely its own Thicknesses
There was a problem hiding this comment.
They shouldn't (or at least it would be redundant if it did), see my other comment
Codecov ReportAttention: Patch coverage is
Additional details and impacted files@@ Coverage Diff @@
## master #2470 +/- ##
===========================================
- Coverage 87.17% 57.49% -29.68%
===========================================
Files 239 238 -1
Lines 45160 45155 -5
Branches 4007 4007
===========================================
- Hits 39367 25961 -13406
- Misses 5112 18373 +13261
- Partials 681 821 +140 ☔ View full report in Codecov by Sentry. |
|
|
||
| mapper[d] = d._rebuild(symbolic_min=left, symbolic_max=right, thickness=thickness) | ||
| tkns = tuple(t._rebuild(name=sregistry.make_name(prefix=t.name)) for t in d.tkns) | ||
| mapper.update({t0: t1 for t0, t1 in zip(d.tkns, tkns)}) |
There was a problem hiding this comment.
I would avoid t0 and t1 due to name clash with time buffers.
It's not about correctness, but mostly when grepping
| # Dimension is of type `left`/`right` - compute the offset | ||
| # and then add 1 to get the appropriate thickness | ||
| if self.value is not None: | ||
| tkn = grid.distributor.glb_to_loc(self.root, r_tkn-1, self.side) |
| @property | ||
| def is_abstract(self): | ||
| return all(i is None for i in flatten(self.thickness)) | ||
| tkns = thickness # Shortcut for thickness |
There was a problem hiding this comment.
should this not be a cached propertry?
Fixes issues with standalone use of
SubDimension.symbolic_minorSubDimension.symbolic_maxin an operator. Previously variants onwould suffer from two issues:
ix.symbolic_max) would not be concretised fromx_ltkntox_ltkn0_arg_valuesunless the user explicitly passed it to the operator (and even then, it would not be adjusted to reflect any MPI decomposition)These are rectified by this PR.