This repository was archived by the owner on Jul 1, 2025. It is now read-only.
[runtime] Remove all tensor copies and almost all allocs from the executor #2821
Conversation
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Contributor
Author
|
By my measure this cuts a few seconds off the unittest run time as well, hooray |
Contributor
|
Nice! |
nadavrot
reviewed
Apr 30, 2019
nadavrot
reviewed
Apr 30, 2019
Contributor
|
This is great, thanks! |
Contributor
jackm321
approved these changes
Apr 30, 2019
Contributor
There was a problem hiding this comment.
Add a comment about lack of ownership of tensors here
Contributor
Author
There was a problem hiding this comment.
What do you mean here? The Placeholders are owned by module_. (maybe just write the comment you like here and i'll copy it in 😄)
Contributor
Author
|
found another unused propagate function to remove |
Contributor
Author
|
change |
Contributor
|
@nickgg Congrats on landing this PR. This is great! |
Sign up for free
to subscribe to this conversation on GitHub.
Already have an account?
Sign in.
5 participants
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
Description: The ThreadPoolExecutor is doing a lot of copies, which makes it pretty slow - when running realistic size traffic through the runtime we spent more than 3ms in the executor preparing Tensors. Fortunately the executor doesn't need to do any copies at all - since we serialize any nodes that would share an input/output pair, so I've cut them all out. Additionally, we were allocating new Tensor buffers for all sub-graphs which we don't need either.
This has a significant impact on Runtime performance - about a 15x reduction in introduced latency in the test I ran (scales with input size though, obviously):
before:
after:
Testing: all tests under debug, release and ASAN/UBSAN.
Documentation: fixes #2773 and #2776 by removing affected code (I think).