commit 406392d08ee8746f474c82cd7d2489da3c91f6c8
author Han-Chung Wang <hanhan0912@gmail.com> Wed Dec 17 00:31:36 2025 +0800
committer GitHub <noreply@github.com> Tue Dec 16 16:31:36 2025 +0000
tree 3792f7178dcb5e8a1273ba60f57f42d828a102a9
parent bb6ec290fe145690db9ec15289dbca6179ef4ae3

[Stream] Selecting unified encodings from encoding resolvers. (#22898)

Add a new `getUnifiedEncoding` method to the `LayoutResolverAttr`
interface that returns a unified encoding given multiple candidate
encodings. This is used by the UnifyEncodingForGlobals pass to select an
appropriate encoding when the same source data has multiple encoded
versions.

Also refactor GlobalEncodingAnalyzer to:
- Set up the layout resolver from dialect interfaces internally
- Compute unified encodings as part of the analysis phase
- Provide a `getUnifiedEncoding(name)` getter for querying results

This simplifies the pass by moving analysis-related logic into the
analyzer class, making the pass focused on applying transformations.

Update lit tests to include device definitions and affinity attributes
on stream.tensor.* ops, which are required for the pass to resolve
layout attributes properly. It also switches identity_resolver to
specialization_resolver, which improves the test quality. Identity
encoding is used in fallback solution.

It is a step towards https://github.com/iree-org/iree/issues/22485

---------

Signed-off-by: hanhanW <hanhan0912@gmail.com>