[SplitLogicalObjFifo] Fix split-logicalobjfifo pass to analyse unique producers/consumers ObjFifos

[Abhishek-Varma]

In order to decide the split factor we were basing the inference solely on the number of columns available.
Because of this, for 4x8 array and for the new pipeline, we were splitting L2 buffers :-

1. LHS - 8 times.
2. RHS - 8 times.
3. OUT - 8 times.

As a result, the tiles being assigned were :-

1. LHS : (0,0) -> (7,0)
2. RHS : (0,0) -> (7,0)
3. OUT : (0,0) -> (7,0)

This causes exhaustion of DMA channels. Refer to this thread for the discussion thread.

The L2 buffer split should be :-

1. LHS : (0,0) -> (3,0)
2. RHS : (0,0) -> (7,0)
3. OUT : (0,0) -> (7,0)

So that later on when the tiles are being assigned, the expected no. of tile assignments for LHS/RHS/OUT matches the corresponding L2 buffers.

This PR aims to analyse the number of unique producers/consumers ObjFifos for the ObjFifo being split..

e2e CI test for Matmul both with/without ukernel via pack-peel-4-level-tiling pipeline targeting 4x8 array on Strix have been added.

Signed-off-by: Abhishek Varma abhvarma@amd.com

[yzhang93]

This revision is much more concise and cleaner. I have more comments on the test.

[yzhang93]

I don't understand why this test works as in your description num_cols=2 and num_rows=1, while here it's num_cols=8 and num_rows=4.

[Abhishek-Varma]

That's the intention : how much to split is not only dependent on the number of available columns but also the unique L2<->L1 pairs -> so we use gcd between the two (and in some cases we might need to take gcd between source/target sizes)

"to keep the test case concise it demonstrates a similar splitting strategy for 1 row and 2 columns." means the actual compute is taking place on 1 row and 2 columns. I could change num_cols=2 and num_rows=1 but that'd defeat the purpose.

I've updated the comments.

[jtuyls]

LGTM, just one nit and also update the PR title to producers/consumers instead of L1/L2.