commit 3659ccf168725bcc33e2db8c101e222c0a903c35
author MaheshRavishankar <1663364+MaheshRavishankar@users.noreply.github.com> Tue Sep 28 09:28:49 2021 -0700
committer GitHub <noreply@github.com> Tue Sep 28 09:28:49 2021 -0700
tree 48bb593079bc5509d70b043a5a687d224e738687
parent 5c4478401fbed50f1256a59d4c7eb3f71b6230f5

Revisit CPU lowering configuration settings. (#7157)

The current configurations on CPU backends (and all backends) rely on
looking at operations within the dispatch region to pick a
default. This makes it hard to set a good default to ensure good tile

distribute of operations. For every op a configuration method is
needed.
Instead one can just look at the loops and get an idea of the problem
size. These loops are created by tile + distribute at the flow level
using parametric tile sizes. The bounds of the loop are already
determined from the problem sizes. Ideally we would have a loop
representation that does not immediately materialize the distributed
loop, but rather maintains the loop in undistributed form, with
additional operands that specify the id/count/size to use when
materializing the distributed loop (defering the materialization to a
later convenient point).

In absense of this, in this PR, the tiled loop as generated today is
recognized by "lifting" it back to the untiled version based on the
know distribution pattern used at the flow level. From this the
original problem size is known as well and used for default
distribution without having to look at the code within the dispatch
region.