Skip to content

Conversation

tomwhite
Copy link
Member

Fixes #221

The key idea here is that merge_chunks is implemented using map_direct rather than using rechunk, and is therefore amenable to blockwise fusion.

Here are the before/after plans for lithops-matmul-random.py - note that there are 5 arrays to compute, rather than 7:

cubed-before

cubed-after

There are 20% fewer tasks too, which makes the computation run faster

1689162377_timeline

1689164509_timeline

using `map_direct` rather than using `rechunk`, and is therefore
amenable to blockwise fusion.
@tomwhite tomwhite merged commit 22e57d6 into main Jul 17, 2023
@tomwhite tomwhite deleted the merge-chunks branch July 17, 2023 08:42
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.

Fuse rechunk and blockwise combine operations in reductions
1 participant