Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

[WIP] Resize scheduler #3425

Draft
wants to merge 116 commits into
base: main
Choose a base branch
from
Draft

[WIP] Resize scheduler #3425

wants to merge 116 commits into from

Conversation

naoyam
Copy link
Collaborator

@naoyam naoyam commented Nov 16, 2024

naoyam added a commit that referenced this pull request Dec 4, 2024
Currently we check all loop groups, no matter if they correspond to loop
IDs. For loop promotion, we only need to consider loop domains when
checking if the full loop promotion analysis can be skipped. This
refinement was necessary when playing with scheduling approaches for
resize, although it is not currently necessary in #3425. I think this is
generally better than the current method.
Base automatically changed from domain_redundancy to main December 4, 2024 23:20
naoyam added a commit that referenced this pull request Dec 6, 2024
Extracted from #3425 

A simple scheduling utility that replays a single transform expr on the
loop domain of a given tensor. It is conceptually similar to the
existing transform propagator, but this interface can also replay a
transform expr not just as a forward expr but also as a backward expr.

The existing `scheduler_tools::scheduleLoopDomainsLike` can also be used
to propagate loop domains across tensors, but I encountered with a
couple of issues due to the resize mapping [issue](#3455). This
`scheduleLoopDomainsBy` gives more explicit control to set loop domains
with resize ops.
naoyam added a commit that referenced this pull request Dec 10, 2024
Added a scheduler util function schedule a fusion with resize-based ops
such as slice, pad and concat. This propagates resize ops to producers
so that all tensors have the exact-mapped loop domains.

Part of #3425. Extracted so that it can be individually tested.

(There's a follow-up PR: #3555)
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment
Labels
None yet
Projects
None yet
Development

Successfully merging this pull request may close these issues.

1 participant