Skip to content

Conversation

stkaplan
Copy link
Contributor

This adds another layer of kernel caching, that lives on disk and persists between program runs. This includes the generated source, header, and compiled object files for each kernel.

The file name is a hash of the names, types, and formats of each tensor, plus a serialization of the IndexStmt. This should be sufficient to uniquely identify a kernel.

When the environment variable TACO_CACHE_DIR is set, the value is used as the cache directory. If unset, the disk-level cache is disabled.

This adds another layer of kernel caching, that lives on disk and
persists between program runs. This includes the generated source,
header, and compiled object files for each kernel.

The file name is a hash of the names, types, and formats of each tensor,
plus a serialization of the IndexStmt. This should be sufficient to
uniquely identify a kernel.

When the environment variable TACO_CACHE_DIR is set, the value is used
as the cache directory. If unset, the disk-level cache is disabled.
@stkaplan
Copy link
Contributor Author

I have this caching disabled by default, but we could have it enabled with a default directory like ~/.taco/cache or something. What do you think is best?

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