Adding some additional discussion with @csullivan .
We agree that:
- There are different ways to encode layout and padding decisions:
- E0: BufferConstraint(as element in the IR)
- E1: Composing a stage that transforms the layout(a loop that represents
the mapping)
- Non-local rewrites are needed to be able to propagate the layout and padding
decision through out the entire model through constraint solving.
Right now we have some healthy discussions about ways to encode layout and
padding decisions.
Some of my thoughts:
Introducing changes to TIR would needs some additional thoughts that deserves
some extra consideration. Due to the N*M complexity (where N is the TIR
possibilities and M is the number of primitives to be supported).
Right now it is possible to do non-local constraint rewriting flowings as part
of the graph pass. Note that while E1 is indeed less "compact" on one hand, we
can use it to reconstruct THE compact data structure(that represents the layout
decision) that we can use to flow the decisions across the graph node. E1 also
enables some additional capabilities (e.g.) expressing future memory remappings
that do not necessarily fit into padding/packing.
Starting from the graph level allows us to capture learnings, then use some e2e
goals to make an informed decision on TIR level change later if needed.
--
Reply to this email directly or view it on GitHub:
https://github.com/apache/tvm-rfcs/pull/77#issuecomment-1156407362
You are receiving this because you are subscribed to this thread.
Message ID: <apache/tvm-rfcs/pull/77/[email protected]>
