Cores per column placer 2

[kurtis-b-1]

Building on top of branch: https://github.com/Xilinx/mlir-aie/tree/cores-per-column-placer

The "column-wise" unplaced design now maps the reduction across the columns and at the final column performs reduction across the rows using the column-limited placer. It should be possible to use any number of cores up to 8 for the Phoenix and 16 for the Strix.

Renamed the original designs to "row-wise" designs, which use the mapping from the sequential placer, and kept them the same as before.

[AndraBisca]

The example in this PR looks good. More comments in the Column-Limited placer and just the placers.py in general would be good. How much code duplication is there between this new placer and the Sequential one?

[fifield]

Please add some sort of unit test for the new placer functionality...some small snippets of iron as input, then check the mlir is correct in the outputs. I don't see any such tests added in this or the cores-per-column-placer branch, and I'm not sure if they exist at all for the placer.

Of course what I'd really like is for this functionality to migrate into the core mlir-aie infrastructure, e.g. #2265, but so far nobody is interested in that other than me.

[andrej]

Of course what I'd really like is for this functionality to migrate into the core mlir-aie infrastructure, e.g. #2265, but so far nobody is interested in that other than me.

I am also interested in that and more generally migrating some of the higher-level IRON abstractions into MLIR. I believe @jgmelber expressed to me before that he thinks this would be good too.

[kurtis-b-1]

Please add some sort of unit test for the new placer functionality...some small snippets of iron as input, then check the mlir is correct in the outputs.

I added a test that checks the mapping with Column-Limited placer set to 2 cores per column. Please let me know if I should modify it or if it doesn't make sense.

[AndraBisca]

Thank you for making the changes @kurtis-b-1! I have one last nitpick about the naming, which currently makes me think that this placer limits the number of columns. Maybe something like ComputeLimitedPlacer would be more descriptive?

[kurtis-b-1]

@AndraBisca Thanks for catching those! Updated the documentation as per the comments.