[ET-VK][ez] Enable dynamic shape support when using push constants #11253
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
## Changes * Call `encode_execute()` upon resize in `VulkanBackend.cpp` * Minor update to `DispatchNode` to store push constant data array as a persistent member of the class ## Motivation Passing in tensor metadata (i.e. sizes, strides) via push constants is typically more performant than passing them via a UBO (uniform buffer object). However, currently dynamic shapes do not work when push constants are used as I realized that the tensor metadata contained in the push constants do not get updated. It appears that that `vkCmdPushConstants` sets the push constants when encoding the command buffer, however the push constants will not be updated if the command buffer is submitted for execution multiple times. Therefore, to update push constant values **the command buffer needs to be re-encoded**. ## Performance Impact This may add a small performance overhead (i.e. re-encoding the command buffer) when executing models with dynamic shapes. Models that do not trigger tensor resizing will not be impacted. However, I measured the impact on a llama 3.2 1B model and the impact of re-encoding a command buffer appears to be negligible. In any case, re-encoding the command buffer is a "necessary evil" when working with dynamic shapes, otherwise the tensor metadata seen by shaders may never get updated. Furthermore, re-encoding the command buffer can allow an opportunity to adjust global work group sizing to match current tensor sizes, which may have a huge performance impact when maximum tensor sizes far exceeds what tensor sizes will realistically be during inference (one instance of this is for transformer models when the max sequence length is very long). Differential Revision: [D75686051](https://our.internmc.facebook.com/intern/diff/D75686051/) [ghstack-poisoned]
🔗 Helpful Links🧪 See artifacts and rendered test results at hud.pytorch.org/pr/pytorch/executorch/11253
Note: Links to docs will display an error until the docs builds have been completed. ❌ 2 New Failures, 2 PendingAs of commit d966e79 with merge base ce8359d ( NEW FAILURES - The following jobs have failed:
This comment was automatically generated by Dr. CI and updates every 15 minutes. |
This was referenced May 30, 2025
Merged
|
This pull request was exported from Phabricator. Differential Revision: D75686051 |
…onstants" ## Changes * Call `encode_execute()` upon resize in `VulkanBackend.cpp` * Minor update to `DispatchNode` to store push constant data array as a persistent member of the class ## Motivation Passing in tensor metadata (i.e. sizes, strides) via push constants is typically more performant than passing them via a UBO (uniform buffer object). However, currently dynamic shapes do not work when push constants are used as I realized that the tensor metadata contained in the push constants do not get updated. It appears that that `vkCmdPushConstants` sets the push constants when encoding the command buffer, however the push constants will not be updated if the command buffer is submitted for execution multiple times. Therefore, to update push constant values **the command buffer needs to be re-encoded**. ## Performance Impact This may add a small performance overhead (i.e. re-encoding the command buffer) when executing models with dynamic shapes. Models that do not trigger tensor resizing will not be impacted. However, I measured the impact on a llama 3.2 1B model and the impact of re-encoding a command buffer appears to be negligible. In any case, re-encoding the command buffer is a "necessary evil" when working with dynamic shapes, otherwise the tensor metadata seen by shaders may never get updated. Furthermore, re-encoding the command buffer can allow an opportunity to adjust global work group sizing to match current tensor sizes, which may have a huge performance impact when maximum tensor sizes far exceeds what tensor sizes will realistically be during inference (one instance of this is for transformer models when the max sequence length is very long). Differential Revision: [D75686051](https://our.internmc.facebook.com/intern/diff/D75686051/) [ghstack-poisoned]
|
This pull request was exported from Phabricator. Differential Revision: D75686051 |
…onstants" ## Changes * Call `encode_execute()` upon resize in `VulkanBackend.cpp` * Minor update to `DispatchNode` to store push constant data array as a persistent member of the class ## Motivation Passing in tensor metadata (i.e. sizes, strides) via push constants is typically more performant than passing them via a UBO (uniform buffer object). However, currently dynamic shapes do not work when push constants are used as I realized that the tensor metadata contained in the push constants do not get updated. It appears that that `vkCmdPushConstants` sets the push constants when encoding the command buffer, however the push constants will not be updated if the command buffer is submitted for execution multiple times. Therefore, to update push constant values **the command buffer needs to be re-encoded**. ## Performance Impact This may add a small performance overhead (i.e. re-encoding the command buffer) when executing models with dynamic shapes. Models that do not trigger tensor resizing will not be impacted. However, I measured the impact on a llama 3.2 1B model and the impact of re-encoding a command buffer appears to be negligible. In any case, re-encoding the command buffer is a "necessary evil" when working with dynamic shapes, otherwise the tensor metadata seen by shaders may never get updated. Furthermore, re-encoding the command buffer can allow an opportunity to adjust global work group sizing to match current tensor sizes, which may have a huge performance impact when maximum tensor sizes far exceeds what tensor sizes will realistically be during inference (one instance of this is for transformer models when the max sequence length is very long). Differential Revision: [D75686051](https://our.internmc.facebook.com/intern/diff/D75686051/) [ghstack-poisoned]
|
This pull request was exported from Phabricator. Differential Revision: D75686051 |
…onstants" ## Changes * Call `encode_execute()` upon resize in `VulkanBackend.cpp` * Minor update to `DispatchNode` to store push constant data array as a persistent member of the class ## Motivation Passing in tensor metadata (i.e. sizes, strides) via push constants is typically more performant than passing them via a UBO (uniform buffer object). However, currently dynamic shapes do not work when push constants are used as I realized that the tensor metadata contained in the push constants do not get updated. It appears that that `vkCmdPushConstants` sets the push constants when encoding the command buffer, however the push constants will not be updated if the command buffer is submitted for execution multiple times. Therefore, to update push constant values **the command buffer needs to be re-encoded**. ## Performance Impact This may add a small performance overhead (i.e. re-encoding the command buffer) when executing models with dynamic shapes. Models that do not trigger tensor resizing will not be impacted. However, I measured the impact on a llama 3.2 1B model and the impact of re-encoding a command buffer appears to be negligible. In any case, re-encoding the command buffer is a "necessary evil" when working with dynamic shapes, otherwise the tensor metadata seen by shaders may never get updated. Furthermore, re-encoding the command buffer can allow an opportunity to adjust global work group sizing to match current tensor sizes, which may have a huge performance impact when maximum tensor sizes far exceeds what tensor sizes will realistically be during inference (one instance of this is for transformer models when the max sequence length is very long). Differential Revision: [D75686051](https://our.internmc.facebook.com/intern/diff/D75686051/) [ghstack-poisoned]
|
This pull request was exported from Phabricator. Differential Revision: D75686051 |
…onstants" ## Changes * Call `encode_execute()` upon resize in `VulkanBackend.cpp` * Minor update to `DispatchNode` to store push constant data array as a persistent member of the class ## Motivation Passing in tensor metadata (i.e. sizes, strides) via push constants is typically more performant than passing them via a UBO (uniform buffer object). However, currently dynamic shapes do not work when push constants are used as I realized that the tensor metadata contained in the push constants do not get updated. It appears that that `vkCmdPushConstants` sets the push constants when encoding the command buffer, however the push constants will not be updated if the command buffer is submitted for execution multiple times. Therefore, to update push constant values **the command buffer needs to be re-encoded**. ## Performance Impact This may add a small performance overhead (i.e. re-encoding the command buffer) when executing models with dynamic shapes. Models that do not trigger tensor resizing will not be impacted. However, I measured the impact on a llama 3.2 1B model and the impact of re-encoding a command buffer appears to be negligible. In any case, re-encoding the command buffer is a "necessary evil" when working with dynamic shapes, otherwise the tensor metadata seen by shaders may never get updated. Furthermore, re-encoding the command buffer can allow an opportunity to adjust global work group sizing to match current tensor sizes, which may have a huge performance impact when maximum tensor sizes far exceeds what tensor sizes will realistically be during inference (one instance of this is for transformer models when the max sequence length is very long). Differential Revision: [D75686051](https://our.internmc.facebook.com/intern/diff/D75686051/) [ghstack-poisoned]
|
This pull request was exported from Phabricator. Differential Revision: D75686051 |
trivedivivek
approved these changes
May 30, 2025
…onstants" ## Changes * Call `encode_execute()` upon resize in `VulkanBackend.cpp` * Minor update to `DispatchNode` to store push constant data array as a persistent member of the class ## Motivation Passing in tensor metadata (i.e. sizes, strides) via push constants is typically more performant than passing them via a UBO (uniform buffer object). However, currently dynamic shapes do not work when push constants are used as I realized that the tensor metadata contained in the push constants do not get updated. It appears that that `vkCmdPushConstants` sets the push constants when encoding the command buffer, however the push constants will not be updated if the command buffer is submitted for execution multiple times. Therefore, to update push constant values **the command buffer needs to be re-encoded**. ## Performance Impact This may add a small performance overhead (i.e. re-encoding the command buffer) when executing models with dynamic shapes. Models that do not trigger tensor resizing will not be impacted. However, I measured the impact on a llama 3.2 1B model and the impact of re-encoding a command buffer appears to be negligible. In any case, re-encoding the command buffer is a "necessary evil" when working with dynamic shapes, otherwise the tensor metadata seen by shaders may never get updated. Furthermore, re-encoding the command buffer can allow an opportunity to adjust global work group sizing to match current tensor sizes, which may have a huge performance impact when maximum tensor sizes far exceeds what tensor sizes will realistically be during inference (one instance of this is for transformer models when the max sequence length is very long). Differential Revision: [D75686051](https://our.internmc.facebook.com/intern/diff/D75686051/) [ghstack-poisoned]
8000
|
This pull request was exported from Phabricator. Differential Revision: D75686051 |
…onstants" ## Changes * Call `encode_execute()` upon resize in `VulkanBackend.cpp` * Minor update to `DispatchNode` to store push constant data array as a persistent member of the class ## Motivation Passing in tensor metadata (i.e. sizes, strides) via push constants is typically more performant than passing them via a UBO (uniform buffer object). However, currently dynamic shapes do not work when push constants are used as I realized that the tensor metadata contained in the push constants do not get updated. It appears that that `vkCmdPushConstants` sets the push constants when encoding the command buffer, however the push constants will not be updated if the command buffer is submitted for execution multiple times. Therefore, to update push constant values **the command buffer needs to be re-encoded**. ## Performance Impact This may add a small performance overhead (i.e. re-encoding the command buffer) when executing models with dynamic shapes. Models that do not trigger tensor resizing will not be impacted. However, I measured the impact on a llama 3.2 1B model and the impact of re-encoding a command buffer appears to be negligible. In any case, re-encoding the command buffer is a "necessary evil" when working with dynamic shapes, otherwise the tensor metadata seen by shaders may never get updated. Furthermore, re-encoding the command buffer can allow an opportunity to adjust global work group sizing to match current tensor sizes, which may have a huge performance impact when maximum tensor sizes far exceeds what tensor sizes will realistically be during inference (one instance of this is for transformer models when the max sequence length is very long). Differential Revision: [D75686051](https://our.internmc.facebook.com/intern/diff/D75686051/) [ghstack-poisoned]
|
This pull request was exported from Phabricator. Differential Revision: D75686051 |
abdfc57
into
gh/SS-JIA/229/base
98 of 102 checks passed
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
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):
Changes
encode_execute()upon resize inVulkanBackend.cppDispatchNodeto store push constant data array as a persistent member of the classMotivation
Passing in tensor metadata (i.e. sizes, strides) via push constants is typically more performant than passing them via a UBO (uniform buffer object). However, currently dynamic shapes do not work when push constants are used as I realized that the tensor metadata contained in the push constants do not get updated.
It appears that that
vkCmdPushConstantssets the push constants when encoding the command buffer, however the push constants will not be updated if the command buffer is submitted for execution multiple times. Therefore, to update push constant values the command buffer needs to be re-encoded.Performance Impact
This may add a small performance overhead (i.e. re-encoding the command buffer) when executing models with dynamic shapes. Models that do not trigger tensor resizing will not be impacted.
However, I measured the impact on a llama 3.2 1B model and the impact of re-encoding a command buffer appears to be negligible.
In any case, re-encoding the command buffer is a "necessary evil" when working with dynamic shapes, otherwise the tensor metadata seen by shaders may never get updated. Furthermore, re-encoding the command buffer can allow an opportunity to adjust global work group sizing to match current tensor sizes, which may have a huge performance impact when maximum tensor sizes far exceeds what tensor sizes will realistically be during inference (one instance of this is for transformer models when the max sequence length is very long).
Differential Revision: D75686051