TensorIterator: Avoid nesting two levels of function_ref in for_each#53613
Closed
peterbell10 wants to merge 2 commits intopytorch:masterfrom
Closed
TensorIterator: Avoid nesting two levels of function_ref in for_each#53613peterbell10 wants to merge 2 commits intopytorch:masterfrom
peterbell10 wants to merge 2 commits intopytorch:masterfrom
Conversation
Contributor
💊 CI failures summary and remediationsAs of commit 310fb71 (more details on the Dr. CI page): 💚 💚 Looks good so far! There are no failures yet. 💚 💚 This comment was automatically generated by Dr. CI (expand for details).Follow this link to opt-out of these comments for your Pull Requests.Please report bugs/suggestions to the (internal) Dr. CI Users group. |
swolchok
reviewed
Mar 9, 2021
Contributor
There was a problem hiding this comment.
I especially like this part. IIRC from looking at this previously, SmallVector overhead was a small problem.
Collaborator
Author
There was a problem hiding this comment.
Unfortunately, this doesn't work. data here is potentially shared between multiple threads so it does need to be created inside the lambda after all.
Comment on lines
252
to
260
Contributor
There was a problem hiding this comment.
could we avoid duplicating what used to be LOOP_WRAPPER with a template function of its own that took loop, data, and ntensor (by reference/value as appropriate)?
ezyang
approved these changes
Mar 10, 2021
Contributor
ezyang
left a comment
There was a problem hiding this comment.
This is really nice work, thanks!
Contributor
facebook-github-bot
left a comment
There was a problem hiding this comment.
@ezyang has imported this pull request. If you are a Facebook employee, you can view this diff on Phabricator.
Contributor
xsacha
pushed a commit
to xsacha/pytorch
that referenced
this pull request
Mar 31, 2021
…ytorch#53613) Summary: When calling `TensorIterator::for_each` with a 1d loop, it creates a `function_ref` for the 1D iteration, then wraps it with `LOOP_WRAPPER` to transform it into a 2d loop. That 2d loop then gets wrapped in another `function_ref`. This can result in significant overhead if the 1d inner loop is over a small number of elements. Instead, this wraps the 1d loop before type-erasure so only one level of `function_ref` is introduced. A simple benchmark demonstrates this is a win: ```python import torch a = torch.rand((10000, 2))[::2] %timeit a + a ``` Note the 2D tensor cannot be coalesced into 1D and both `cpu_kernel` and `cpu_kernel_vec` use 1D for_each. On master, this takes 42 us but with this change it's down to 32us. Pull Request resolved: pytorch#53613 Reviewed By: VitalyFedyunin Differential Revision: D26947143 Pulled By: ezyang fbshipit-source-id: 5189ada0d82bbf74170fb446763753f02478abf6
laurentdupin
pushed a commit
to laurentdupin/pytorch
that referenced
this pull request
Apr 24, 2026
…ytorch#53613) Summary: When calling `TensorIterator::for_each` with a 1d loop, it creates a `function_ref` for the 1D iteration, then wraps it with `LOOP_WRAPPER` to transform it into a 2d loop. That 2d loop then gets wrapped in another `function_ref`. This can result in significant overhead if the 1d inner loop is over a small number of elements. Instead, this wraps the 1d loop before type-erasure so only one level of `function_ref` is introduced. A simple benchmark demonstrates this is a win: ```python import torch a = torch.rand((10000, 2))[::2] %timeit a + a ``` Note the 2D tensor cannot be coalesced into 1D and both `cpu_kernel` and `cpu_kernel_vec` use 1D for_each. On master, this takes 42 us but with this change it's down to 32us. Pull Request resolved: pytorch#53613 Reviewed By: VitalyFedyunin Differential Revision: D26947143 Pulled By: ezyang fbshipit-source-id: 5189ada0d82bbf74170fb446763753f02478abf6
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
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
6 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.
When calling
TensorIterator::for_eachwith a 1d loop, it creates afunction_reffor the 1D iteration, then wraps it withLOOP_WRAPPERto transform it into a 2d loop. That 2d loop then gets wrapped in anotherfunction_ref. This can result in significant overhead if the 1d inner loop is over a small number of elements.Instead, this wraps the 1d loop before type-erasure so only one level of
function_refis introduced. A simple benchmark demonstrates this is a win:Note the 2D tensor cannot be coalesced into 1D and both
cpu_kernelandcpu_kernel_vecuse 1D for_each. On master, this takes 42 us but with this change it's down to 32us.