[ET-VK] Add mechanism to trigger command buffer re-encode only when necessary #13184
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
…ecessary ## Context Dynamic shape models currently will require the command buffer to be re-encoded every inference. However, this introduces a significant overhead when running models that require dynamic shapes. The reality is that a command buffer re-encode may not be needed every frame. A command buffer re-encode will only be needed when: 1. Shader dispatch parameters change; i.e. new tensor sizes require a completely different compute shader, require new local work group sizing, or require new work group grid size (i.e. global work group size / local work group size) 2. Push constants containing tensor metadata need to be updated This diff aims to reduce the overhead of triggering tensor shape change by detecting when a command buffer re-encode is actually needed. ## Changes `ComputeGraph`: * Introduce `requires_reencode` flag to `ComputeGraph` to indicate when a command buffer re-encode is needed. * Introduce a `std::set<ValueRef>` tracking which values were updated when propagating tensor sizes * "update" can be one of two things: 1) tensor sizes changed 2) symint value changed `DispatchNode`: * When propagating new tensor sizes, only execute the resize function if any of the values participating in the computation have been updated * Mark `requries_reencode` if any push constants associated with tensor metadata need to be udpated `DynamicDispatchNode`: * Only recompute compute shader dispatch params if any of the values participating in the computation have been updated * Mark `requires_reencode` if 1) a new compute shader is required, 2) local work group size changed, 3) work group grid size changed Differential Revision: [D79813237](https://our.internmc.facebook.com/intern/diff/D79813237/) [ghstack-poisoned]
🔗 Helpful Links🧪 See artifacts and rendered test results at hud.pytorch.org/pr/pytorch/executorch/13184
Note: Links to docs will display an error until the docs builds have been completed. ❌ 3 New Failures, 4 Unrelated FailuresAs of commit e90389a with merge base b36d6b6 ( NEW FAILURES - The following jobs have failed:
BROKEN TRUNK - The following jobs failed but were present on the merge base:👉 Rebase onto the `viable/strict` branch to avoid these failures
This comment was automatically generated by Dr. CI and updates every 15 minutes. |
meta-cla
bot
added
the
CLA Signed
|
This pull request was exported from Phabricator. Differential Revision: D79813237 |
This PR needs a
|
msluszniak
reviewed
Aug 7, 2025
…only when necessary" ## Context Dynamic shape models currently will require the command buffer to be re-encoded every inference. However, this introduces a significant overhead when running models that require dynamic shapes. The reality is that a command buffer re-encode may not be needed every frame. A command buffer re-encode will only be needed when: 1. Shader dispatch parameters change; i.e. new tensor sizes require a completely different compute shader, require new local work group sizing, or require new work group grid size (i.e. global work group size / local work group size) 2. Push constants containing tensor metadata need to be updated This diff aims to reduce the overhead of triggering tensor shape change by detecting when a command buffer re-encode is actually needed. ## Changes `ComputeGraph`: * Introduce `requires_reencode` flag to `ComputeGraph` to indicate when a command buffer re-encode is needed. * Introduce a `std::set<ValueRef>` tracking which values were updated when propagating tensor sizes * "update" can be one of two things: 1) tensor sizes changed 2) symint value changed `DispatchNode`: * When propagating new tensor sizes, only execute the resize function if any of the values participating in the computation have been updated * Mark `requries_reencode` if any push constants associated with tensor metadata need to be udpated `DynamicDispatchNode`: * Only recompute compute shader dispatch params if any of the values participating in the computation have been updated * Mark `requires_reencode` if 1) a new compute shader is required, 2) local work group size changed, 3) work group grid size changed Differential Revision: [D79813237](https://our.internmc.facebook.com/intern/diff/D79813237/) [ghstack-poisoned]
SS-JIA
pushed a commit
that referenced
this pull request
Aug 11, 2025
…ecessary Pull Request resolved: #13184 ## Context Dynamic shape models currently will require the command buffer to be re-encoded every inference. However, this introduces a significant overhead when running models that require dynamic shapes. The reality is that a command buffer re-encode may not be needed every frame. A command buffer re-encode will only be needed when: 1. Shader dispatch parameters change; i.e. new tensor sizes require a completely different compute shader, require new local work group sizing, or require new work group grid size (i.e. global work group size / local work group size) 2. Push constants containing tensor metadata need to be updated This diff aims to reduce the overhead of triggering tensor shape change by detecting when a command buffer re-encode is actually needed. ## Changes `ComputeGraph`: * Introduce `requires_reencode` flag to `ComputeGraph` to indicate when a command buffer re-encode is needed. * Introduce a `std::set<ValueRef>` tracking which values were updated when propagating tensor sizes * "update" can be one of two things: 1) tensor sizes changed 2) symint value changed `DispatchNode`: * When propagating new tensor sizes, only execute the resize function if any of the values participating in the computation have been updated * Mark `requries_reencode` if any push constants associated with tensor metadata need to be udpated `DynamicDispatchNode`: * Only recompute compute shader dispatch params if any of the values participating in the computation have been updated * Mark `requires_reencode` if 1) a new compute shader is required, 2) local work group size changed, 3) work group grid size changed ghstack-source-id: 302101273 @exported-using-ghexport Differential Revision: [D79813237](https://our.internmc.facebook.com/intern/diff/D79813237/)
|
This pull request was exported from Phabricator. Differential Revision: D79813237 |
…only when necessary" ## Context Dynamic shape models currently will require the command buffer to be re-encoded every inference. However, this introduces a significant overhead when running models that require dynamic shapes. The reality is that a command buffer re-encode may not be needed every frame. A command buffer re-encode will only be needed when: 1. Shader dispatch parameters change; i.e. new tensor sizes require a completely different compute shader, require new local work group sizing, or require new work group grid size (i.e. global work group size / local work group size) 2. Push constants containing tensor metadata need to be updated This diff aims to reduce the overhead of triggering tensor shape change by detecting when a command buffer re-encode is actually needed. ## Changes `ComputeGraph`: * Introduce `requires_reencode` flag to `ComputeGraph` to indicate when a command buffer re-encode is needed. * Introduce a `std::set<ValueRef>` tracking which values were updated when propagating tensor sizes * "update" can be one of two things: 1) tensor sizes changed 2) symint value changed `DispatchNode`: * When propagating new tensor sizes, only execute the resize function if any of the values participating in the computation have been updated * Mark `requries_reencode` if any push constants associated with tensor metadata need to be udpated `DynamicDispatchNode`: * Only recompute compute shader dispatch params if any of the values participating in the computation have been updated * Mark `requires_reencode` if 1) a new compute shader is required, 2) local work group size changed, 3) work group grid size changed Differential Revision: [D79813237](https://our.internmc.facebook.com/intern/diff/D79813237/) [ghstack-poisoned]
SS-JIA
pushed a commit
that referenced
this pull request
Aug 13, 2025
…ecessary Pull Request resolved: #13184 ## Context Dynamic shape models currently will require the command buffer to be re-encoded every inference. However, this introduces a significant overhead when running models that require dynamic shapes. The reality is that a command buffer re-encode may not be needed every frame. A command buffer re-encode will only be needed when: 1. Shader dispatch parameters change; i.e. new tensor sizes require a completely different compute shader, require new local work group sizing, or require new work group grid size (i.e. global work group size / local work group size) 2. Push constants containing tensor metadata need to be updated This diff aims to reduce the overhead of triggering tensor shape change by detecting when a command buffer re-encode is actually needed. ## Changes `ComputeGraph`: * Introduce `requires_reencode` flag to `ComputeGraph` to indicate when a command buffer re-encode is needed. * Introduce a `std::set<ValueRef>` tracking which values were updated when propagating tensor sizes * "update" can be one of two things: 1) tensor sizes changed 2) symint value changed `DispatchNode`: * When propagating new tensor sizes, only execute the resize function if any of the values participating in the computation have been updated * Mark `requries_reencode` if any push constants associated with tensor metadata need to be udpated `DynamicDispatchNode`: * Only recompute compute shader dispatch params if any of the values participating in the computation have been updated * Mark `requires_reencode` if 1) a new compute shader is required, 2) local work group size changed, 3) work group grid size changed ghstack-source-id: 302596078 @exported-using-ghexport Differential Revision: [D79813237](https://our.internmc.facebook.com/intern/diff/D79813237/)
|
This pull request was exported from Phabricator. Differential Revision: D79813237 |
…only when necessary" ## Context Dynamic shape models currently will require the command buffer to be re-encoded every inference. However, this introduces a significant overhead when running models that require dynamic shapes. The reality is that a command buffer re-encode may not be needed every frame. A command buffer re-encode will only be needed when: 1. Shader dispatch parameters change; i.e. new tensor sizes require a completely different compute shader, require new local work group sizing, or require new work group grid size (i.e. global work group size / local work group size) 2. Push constants containing tensor metadata need to be updated This diff aims to reduce the overhead of triggering tensor shape change by detecting when a command buffer re-encode is actually needed. ## Changes `ComputeGraph`: * Introduce `requires_reencode` flag to `ComputeGraph` to indicate when a command buffer re-encode is needed. * Introduce a `std::set<ValueRef>` tracking which values were updated when propagating tensor sizes * "update" can be one of two things: 1) tensor sizes changed 2) symint value changed `DispatchNode`: * When propagating new tensor sizes, only execute the resize function if any of the values participating in the computation have been updated * Mark `requries_reencode` if any push constants associated with tensor metadata need to be udpated `DynamicDispatchNode`: * Only recompute compute shader dispatch params if any of the values participating in the computation have been updated * Mark `requires_reencode` if 1) a new compute shader is required, 2) local work group size changed, 3) work group grid size changed Differential Revision: [D79813237](https://our.internmc.facebook.com/intern/diff/D79813237/) [ghstack-poisoned]
SS-JIA
pushed a commit
that referenced
this pull request
Aug 13, 2025
…ecessary Pull Request resolved: #13184 ## Context Dynamic shape models currently will require the command buffer to be re-encoded every inference. However, this introduces a significant overhead when running models that require dynamic shapes. The reality is that a command buffer re-encode may not be needed every frame. A command buffer re-encode will only be needed when: 1. Shader dispatch parameters change; i.e. new tensor sizes require a completely different compute shader, require new local work group sizing, or require new work group grid size (i.e. global work group size / local work group size) 2. Push constants containing tensor metadata need to be updated This diff aims to reduce the overhead of triggering tensor shape change by detecting when a command buffer re-encode is actually needed. ## Changes `ComputeGraph`: * Introduce `requires_reencode` flag to `ComputeGraph` to indicate when a command buffer re-encode is needed. * Introduce a `std::set<ValueRef>` tracking which values were updated when propagating tensor sizes * "update" can be one of two things: 1) tensor sizes changed 2) symint value changed `DispatchNode`: * When propagating new tensor sizes, only execute the resize function if any of the values participating in the computation have been updated * Mark `requries_reencode` if any push constants associated with tensor metadata need to be udpated `DynamicDispatchNode`: * Only recompute compute shader dispatch params if any of the values participating in the computation have been updated * Mark `requires_reencode` if 1) a new compute shader is required, 2) local work group size changed, 3) work group grid size changed ghstack-source-id: 302703876 @exported-using-ghexport Differential Revision: [D79813237](https://our.internmc.facebook.com/intern/diff/D79813237/)
andreanicastro
approved these changes
Aug 13, 2025
|
This pull request was exported from Phabricator. Differential Revision: D79813237 |
facebook-github-bot
merged commit a64208e
into
gh/SS-JIA/271/base
98 of 106 checks passed
facebook-github-bot
temporarily deployed
to
cherry-pick-bot
GitHub Actions
Inactive
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):
Context
Dynamic shape models currently will require the command buffer to be re-encoded every inference. However, this introduces a significant overhead when running models that require dynamic shapes.
The reality is that a command buffer re-encode may not be needed every frame. A command buffer re-encode will only be needed when:
This diff aims to reduce the overhead of triggering tensor shape change by detecting when a command buffer re-encode is actually needed.
Changes
ComputeGraph:requires_reencodeflag toComputeGraphto indicate when a command buffer re-encode is needed.std::set<ValueRef>tracking which values were updated when propagating tensor sizesDispatchNode:requries_reencodeif any push constants associated with tensor metadata need to be udpatedDynamicDispatchNode:requires_reencodeif 1) a new compute shader is required, 2) local work group size changed, 3) work group grid size changedDifferential Revision: D79813237