You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
logical domain is not the right target for checking validation of reference TensorView.
We care about transformation propagations and coverage, so we should have used initial loop domain instead.
Specifically, this makes it easier to schedule opaque indexing operations, or sparse update operations like scatter. Because the logical domain of the large buffer isn't relevant for scheduling.
The PR replaces getLogicalDomain() with getInitialLoopDomain() in multiple places within areAllInputIdsMappedTo and areAllTargetIdsCoveredBy functions. However, it is unclear if all uses of the logical domain were intended to be replaced, especially in contexts where transformation propagation relies on logical structure. This change may affect correctness in scheduling scenarios involving complex transformations.
for (auto in_id : input_tv->getInitialLoopDomain()) {
if (canIgnoreIndexedInputDomainID(input_tv, in_id, ca_map_)) {
continue;
}
// Permissive map is required for the transpose scheduler to support cases// like T0[I0, b] + T1[b, I1]auto concrete =
ca_map_.getConcreteMappedID(in_id, IdMappingMode::PERMISSIVE);
if (!concrete->isBroadcast() && !in_id->isReduction()) {
in_concrete_ids.insert(concrete);
}
}
// Erase all input concrete IDs mapped to the output domain// Ignore unresolved broadcast dimensionseraseifInputMappedThroughRootDomainAndIndexing(
in_concrete_ids, tv->getInitialLoopDomain());
The comment suggests that broadcast handling should ideally be moved inside a loop using get_source_iter_domains, but this was not implemented due to an open issue. The PR updates the comment to reference getInitialLoopDomain(), but does not address the underlying technical debt or clarify whether this change impacts broadcast handling safety.
// `get_source_iter_domains(target_tv->getInitialLoopDomain())` and skip// broadcast source IDs. currently we have the issue that split/merge does
CI seems to have some issue here. I cannot see the failing tests....
On Naoya's comment here: #5114 (comment)
I'm suspecting TransformPropagator also need to switch from logical domain to initial loop domain, following the same logic here. But that felt like a large hammer... I'll double check after looking at failing log.
jjsjann123
changed the title
Use initial loop domain in domain_map
[DoNotReview] Use initial loop domain in domain_map
Sep 8, 2025
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
logical domain is not the right target for checking validation of reference TensorView.
We care about transformation propagations and coverage, so we should have used initial loop domain instead.
Specifically, this makes it easier to schedule opaque indexing operations, or sparse update operations like scatter. Because the logical domain of the large buffer isn't relevant for scheduling.