[Mlir] Skip default schedule for linalg.fill - #63
Conversation
54d6356 to
e404b60
Compare
|
rebase done |
|
just wanted to note that this is compatible with fusion since fusion ignores fused producer handles including fills. |
3c39731 to
a63ca4b
Compare
qaco
left a comment
qaco
left a comment
There was a problem hiding this comment.
Ok but at some point in the future I think we should be able to schedule the fill
I think the main reason that we're not scheduling the fill is so that we can fuse them into the main op |
a63ca4b to
5546481
Compare
|
Exactly, I have a pass in SDist that fuse the fill op with the next op. (mandatory to have descent perf in my case).
|
Do not schedule the initialization operation linalg.fill, which is not controllable by the user. The default schedule applied can overconstraint lowering passes trying to fuse/optimize it.
5546481 to
6b4261c
Compare
Codecov Report✅ All modified and coverable lines are covered by tests. 📢 Thoughts on this report? Let us know! |
guillon
left a comment
guillon
left a comment
There was a problem hiding this comment.
Fine for me, but we should change the implementation, the linalg fill should not have an associated node schedule, I will experiment with this and come back to you.
4 participants
Do not schedule the initialization operation linalg.fill, which is not controllable by the user.
Motivation
The default schedule applied can overconstraint lowering passes trying to fuse/optimize it.
Description
Based on the naming convention of Mlir nodes, skip the scheduling of the node corresponding to the linalg.fill.
Commits
Single commit
Discussion
What do you think ?