commit 02d145e5a948283df3cd30289fa68fc9ceb602cb
author Han-Chung Wang <hanhan0912@gmail.com> Wed Jan 08 19:39:49 2025 -0800
committer GitHub <noreply@github.com> Thu Jan 09 03:39:49 2025 +0000
tree 6a37230fa8cb2c4ae1e88451f2c13b50f127c0a6
parent 74f8d3c6b9800b431c42573df523e5d06c9a51d4

[Stream] Implement SpecializeEncodings pass (1/n) (#19502)

There are three major changes in the revision:

- Introduce `AffinityAnalysisDialectInterface` Stream dialect interface.
It is used to fetch attributes that are defined by other dialects. In
the revision, HAL implements the dialect interface, and it can return
whatever attribute attached in HAL::ExecutableTarget attributes. The
main idea of the dialect interface is that Stream **does not** need to
depend on HAL to get the layout information.
- Add `cloneWithLayouts` method to the EncodingAttr. It is used in the
encoding specialization pass where it can resolve the layout
requirements and add it to the `layouts` field. The other optional
parameters are dropped because the layout is already resolved. It can be
a new Encoding dialect attribute because it is just describing the
layout. The stream tensor ops do not need to know the `op_type`,
`element_types` and `operand_index` parameters. It only needs the layout
information, and the attribute should implement the interface method.
- Partially implement the SpecializeEncodings pass. The responsibility
of the pass is large, so I decide to implement it incrementally. This
revision only implements the mechanism of updating stream tensor ops'
encoding, and only stream.tensor.sizeof op is supported. The rest of the
support for other stream tensor op can be added later on. The executable
duplication and the update of dispatch ops will be implemented in
subsequent PRs.

---------

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