[ObjectFifo] Create a new pass to split L2 buffers

[Abhishek-Varma]

This PR includes 2 commits :-

Commit 1
-- This commit introduces a new pass --iree-amdaie-split-buffers to split L2 buffers for dealing with Matmul+Elementwise.
-- It addresses sub-action 2 as well from #644

Commit 2

• Adds --iree-amdaie-split-buffers pass to the objectFifo pipeline.

Signed-off-by: Abhishek Varma [email protected]

NOTE: Let me know what can we name the pass instead of iree-amdaie-split-buffers - iree-amdaie-split-l2-buffers is a possibility but we clearly aren't splitting ALL L2 buffers.

[Abhishek-Varma]

Previously L2's access patterns were incorrect after being split.

L3 -> L2 requires 4D access pattern. L2 -> L1 requires 6D access pattern. The previous revision was incorrectly forming 4D access pattern for both cases and was doing so by linearizing the L2 buffer.

The current revision takes care of the following :-

1. 4D access pattern for L3 -> L2.
2. 6D access pattern for L2 -> L1.
3. Correct offset (L2 for both L3 -> L2 and L2 -> L1)
4. Correct size (L2 for both L3 -> L2 and L2 -> L1)

I've pushed the current revision as a WIP because now I need to take a look at L3's case as well as strides in the access pattern.

Once verified, only then will I update the lit test case and clean up the code before asking for re-review. But attaching here the IR snippet - now instead of <1024>, I'm forming correct four splits (<1x1x32x32>) of the L2 buffer (<2x2x32x32>).

CC: @jtuyls

[Abhishek-Varma]

Here is the current state of Matmul + Elemwise : e2e IR log with ToM + Splitting LogicalObjectFifos.