Skip to content

Conversation

@vtjnash
Copy link
Member

@vtjnash vtjnash commented Sep 14, 2020

Closes #37448, reopens #35566

Uses of switchtupleunion need to be guarded in cost (e.g. with countunionsplit), while there's currently no practical way to limit this. The inference algorithm assumes that the canonical form is that of tuplemerge, and dislikes when we form the switched version implicitly, as it has a higher lattice cost, so it's expected to only appear explicitly.

@JeffBezanson JeffBezanson merged commit b18647e into master Sep 21, 2020
@JeffBezanson JeffBezanson deleted the jn/35600-revert branch September 21, 2020 20:55
vtjnash added a commit that referenced this pull request Sep 23, 2020
@vtjnash
Copy link
Member Author

vtjnash commented Sep 23, 2020

The before (green) / after (blue) times here are pretty cool to see:
image

The #37714 PR will make them a bit worse than the blue again, but not much (not like the green line).

vtjnash added a commit that referenced this pull request Sep 24, 2020
KristofferC pushed a commit that referenced this pull request Oct 21, 2020
This reverts commit cf0c3e0.

(cherry picked from commit b18647e)
@KristofferC KristofferC mentioned this pull request Oct 21, 2020
15 tasks
@vchuravy vchuravy mentioned this pull request Jan 21, 2021
27 tasks
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.

Exploding inference time with map(identity, zip(x...))

4 participants