Non-ideal fix for unreachable scalar variables

[fmalatino]

Description
As authored by @FlorianDeconinck. Current hotfix for scalar variables which are unreachable during DaCe orchestration procedure

(HOT)Fixes # 42
Hot fix until DaCe, GT4Py bridge refactor

Checklist:

• My code follows the style guidelines of this project
• I have performed a self-review of my own code
• I have commented my code, particularly in hard-to-understand areas
• I have made corresponding changes to the documentation
• My changes generate no new warnings
• Any dependent changes have been merged and published in downstream modules
• New check tests, if applicable, are included

[oelbert]

🙃

[twicki]

We take what we need. Can we have reverting this reflected in an issue such that we come back to it once we can?

[FlorianDeconinck]

Logging the full story + examples + patch code fix that partially fixes things here: #70

[bensonr]

Do we need the same construct for y_out/y_in?

[FlorianDeconinck]

Do we need the same construct for y_out/y_in?

No, we are just forcing the stencil to not become empty when all the horizontal region are being culled. Empty stencil is not an issue but we have a signature/argument passing problem that show up when it happens.

I asked Frank to change it to x_out=x_out which is nonsensical but keeps the stencil from being entirely empty.

[bensonr]

The acoef=mysign was put in to fix a similar issue with scalar variable removal. Are these separate, but related issues?

[FlorianDeconinck]

The acoef=mysign was put in to fix a similar issue with scalar variable removal. Are these separate, but related issues?

Exactly... It all boils down to: we have a bridle way to link signature as described by GT4Py, real call signature in code and SDFG expectation for data going in and out. #70 has a series of 3 tests that encompasses all the issues and some code that fix some of it (but not all).

I believe the right way to deal with this is a good top to bottom clean up of the call structure, simplifying some optimization that are micro-opt with little to no performance improvement and complexify the entire system

[xyuan]

this is to fix same issue but with different stencil, this fix doesn't work when scales large scale run, it requires to rewrite the gt4py interface optimization code to allow the void stencil, or void loop nodes for the backend.

…

On Thu, Sep 5, 2024 at 10:49 AM Rusty Benson ***@***.***> wrote: The acoef=mysign was put in to fix a similar if not the same issue with scalar variable removal. Are these separate, but related issues? — Reply to this email directly, view it on GitHub <#69 (comment)>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/AAFCG2BJZKZMRZLH2LQQMM3ZVBVPZAVCNFSM6AAAAABNU4R6BOVHI2DSMVQWIX3LMV43OSLTON2WKQ3PNVWWK3TUHMZDGMZRHEYTAMZQGE> . You are receiving this because you are subscribed to this thread.Message ID: ***@***.***>

[FlorianDeconinck]

Withholding merge on this as I have implemented a proper fix gt4py side.

Working branch: https://github.com/FlorianDeconinck/gt4py/tree/cartesian/fix/missing_parameter

[FlorianDeconinck]

Seems like the proper fix is not quite there. We will be merging this to unblock release.