Fixing reshape op so it supports reshaping of scalars

[ajakovljevicTT]

As described in #1306, tt-xla produces some reshapes on scalars which are currently not supported by our stableHLO to TTIR dialect lowering (issue #1317). In addition, tt-metal currently throws an error when trying to reshape to a 1-dimensional shape (issue tenstorrent/tt-metal#15201).

To fix both of these issues and fully support scalar testing on tt-xla, I've ingested some of the changes from @mmanzoorTT 's change (#1319) - will merge this PR after it. In addition, I've added that the reshape lowering pass omits the ttir.reshape op completely if the types of the input op and the target shape are the same, which will fix the unnecessary calls to the ttnn reshape op from tt-xla.

Fixes #1306

Edit: In talks with @sdjordjevicTT, he has pointed out the existence of the folder for transform passes for TTIR ops, so I have changed my pass to live there.

[kmitrovicTT]

Run

> pre-commit install
> pre-commit run --all-files

to fix failing check.

[mmanzoorTT]

There is another PR (PR #1252 by @uazizTT) for handling the same issue for broadcast op (remove redundant ops). These two PRs are handling the same problem in two different location. Redundant broadcast op is removed in TTIR->TTIR conversion pass, whereas redundant reshape op is removed in StableHLO->TTIR conversion pass. These two PRs should remove redundant ops in a consistent way.

[mmanzoorTT]

Please add some tests for these changes.

[ajakovljevicTT]

There is another PR (PR #1252 by @uazizTT) for handling the same issue for broadcast op (remove redundant ops). These two PRs are handling the same problem in two different location. Redundant broadcast op is removed in TTIR->TTIR conversion pass, whereas redundant reshape op is removed in StableHLO->TTIR conversion pass. These two PRs should remove redundant ops in a consistent way.

The reshape of here is remove in the TTIR->TTIR decomposition, which is currently a part of the larger TTIR->TTNN conversion, not StableHLO->TTIR. Will look into the referenced PR to get the idea how it is solved there.
I personally prefer the elimination to happen in the TTIR->TTNN part of the pipeline, as it then doesn't disturb the simple StableHLO->TTIR conversion that we agreed on.

[azecevicTT]

Curious why you decided to go with the whole pass, isn't the fold sufficient here?

[ajakovljevicTT]

Curious why you decided to go with the whole pass, isn't the fold sufficient here?

@azecevicTT Just wanted to be in agreement with a PR that does a similar thing for broadcast #1252. Do you think it should be moved to a fold mechanism?

[azecevicTT]

Curious why you decided to go with the whole pass, isn't the fold sufficient here?

@azecevicTT Just wanted to be in agreement with a PR that does a similar thing for broadcast #1252. Do you think it should be moved to a fold mechanism?

I can see, but I think fold is ideal for simple rewritings like this (and broadcast too). In the future we will probably have 10s if not 100s of this kind of rewrting patterns, having a separate pass for all of them can become burdensome from a maintainability standpoint, so it will be best to use an existing infrastructure, or group them under one pass that will operate on all ops, at the moment fold seems like the best option.

[ajakovljevicTT]

Curious why you decided to go with the whole pass, isn't the fold sufficient here?

@azecevicTT Just wanted to be in agreement with a PR that does a similar thing for broadcast #1252. Do you think it should be moved to a fold mechanism?

I can see, but I think fold is ideal for simple rewritings like this (and broadcast too). In the future we will probably have 10s if not 100s of this kind of rewrting patterns, having a separate pass for all of them can become burdensome from a maintainability standpoint, so it will be best to use an existing infrastructure, or group them under one pass that will operate on all ops, at the moment fold seems like the best option.

@azecevicTT Added folding and tested if it works on our current setup with the test in the PR. It seems to be working, so added the ommission as a fold instead of a pass. Thanks for pointing that out!

[ajakovljevicTT]

Hi all, did any of the people who took a look at this PR (or anyone else), have the time to take a second look at this and leave comments and/or approve the PR? Thanks!

[sdjordjevicTT]

Nice and elegant! Thanks.