[ET-VK][4/n] Improve codegen for aten.permute#3087
Closed
yipjustin wants to merge 1 commit intogh/yipjustin/6/basefrom
Closed
[ET-VK][4/n] Improve codegen for aten.permute#3087yipjustin wants to merge 1 commit intogh/yipjustin/6/basefrom
yipjustin wants to merge 1 commit intogh/yipjustin/6/basefrom
Conversation
In the generated code, it uses CPU as reference implementation. Tricky part happens when CPU modify the stride for some indexing operations like `permute`, leading the return Tensor with a non-continous stride. When we create a `vk_out` tensor based on this non-continous tensor with `at::empty_like`, the `vk_out` tensor inherits the stride property. Leading to wrong answer when moving data back from staging. As a solution, we add `.continous()` to after `at::empty_like` to revert back to default stride. Differential Revision: [D56095204](https://our.internmc.facebook.com/intern/diff/D56095204/) [ghstack-poisoned]
🔗 Helpful Links🧪 See artifacts and rendered test results at hud.pytorch.org/pr/pytorch/executorch/3087
Note: Links to docs will display an error until the docs builds have been completed. ✅ No FailuresAs of commit 29dbac5 with merge base 4b6d2c3 ( This comment was automatically generated by Dr. CI and updates every 15 minutes. |
facebook-github-bot
added
the
CLA Signed
This was referenced Apr 17, 2024
yipjustin
added a commit
that referenced
this pull request
Apr 17, 2024
In the generated code, it uses CPU as reference implementation. Tricky part happens when CPU modify the stride for some indexing operations like `permute`, leading the return Tensor with a non-continous stride. When we create a `vk_out` tensor based on this non-continous tensor with `at::empty_like`, the `vk_out` tensor inherits the stride property. Leading to wrong answer when moving data back from staging. As a solution, we add `.continous()` to after `at::empty_like` to revert back to default stride. Differential Revision: [D56095204](https://our.internmc.facebook.com/intern/diff/D56095204/) ghstack-source-id: 222417364 Pull Request resolved: #3087
Contributor
|
This pull request was exported from Phabricator. Differential Revision: D56095204 |
This was referenced Apr 17, 2024
SS-JIA
approved these changes
Apr 17, 2024
Contributor
|
This pull request has been merged in 28be9d6. |
Closed
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
3 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.
Stack from ghstack (oldest at bottom):
In the generated code, it uses CPU as reference implementation.
Tricky part happens when CPU modify the stride for some indexing operations like
permute, leading the return Tensor with a non-continous stride.When we create a
vk_outtensor based on this non-continous tensor withat::empty_like, thevk_outtensor inherits the stride property. Leading to wrong answer when moving data back from staging.As a solution, we add
.continous()to afterat::empty_liketo revert back to default stride.Differential Revision: D56095204