[ET-VK] Refine paritioner to account for storage type and memory layout #6635
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
## Context There are a variety of ways that tensors can be represented in Vulkan. The two main descriptors for how a tensor is laid out in memory is: 1. Storage Type (buffer or texture) 2. Memory Layout (which dim is packed along a texel, which dim has a stride of 1, etc.) Due to the differences between buffers and textures, and the differences between different memory layouts, an implementation for an operator may only support a specific set of (storage type, memory layout) combinations. Furthermore, if an operator implementation supports multiple (storage type, memory layout) combinations, there may be a "preferred" setting which results in optimal performance. These changes lay the foundation for the implementation of a memory metadata tagging graph transform, which will make sure that all tensors participating in an operator call is has a valid/optimal (storage type, memory layout) setting, and insert transition operators to transfer input tensors to the correct memory settings when necessary. An additional change that is required arises from the fact that in Vulkan, there is a limit on texture and buffer sizes. Therefore, the partitioner needs to account for the storage types and memory layouts supported by the operator implementation, and check if all tensors participating in a computation can be represented with some storage type, memory layout combination supported by the implementation. ## Changes Improvements to the operator registry: * Introduce utility functions to check the optimal and enabled storage types and memory layouts for an operator Improvements to the Partitioner: * Account for the storage types and memory layouts supported by an operator when deciding if a node should be partitioned * Improved logic for fusable ops (i.e. the permute/transpose before a mm which can be fused into linear) to check if the final target op is supported in Vulkan, and only partition those nodes if so. Otherwise, don't partition it so that it can be fused by another backend. Differential Revision: [D65428843](https://our.internmc.facebook.com/intern/diff/D65428843/) [ghstack-poisoned]
🔗 Helpful Links🧪 See artifacts and rendered test results at hud.pytorch.org/pr/pytorch/executorch/6635
Note: Links to docs will display an error until the docs builds have been completed. ❌ 2 New FailuresAs of commit a61cab1 with merge base cd565b5 ( NEW FAILURES - The following jobs have failed:
This comment was automatically generated by Dr. CI and updates every 15 minutes. |
facebook-github-bot
added
the
CLA Signed
|
This pull request was exported from Phabricator. Differential Revision: D65428843 |
…memory layout" ## Context There are a variety of ways that tensors can be represented in Vulkan. The two main descriptors for how a tensor is laid out in memory is: 1. Storage Type (buffer or texture) 2. Memory Layout (which dim is packed along a texel, which dim has a stride of 1, etc.) Due to the differences between buffers and textures, and the differences between different memory layouts, an implementation for an operator may only support a specific set of (storage type, memory layout) combinations. Furthermore, if an operator implementation supports multiple (storage type, memory layout) combinations, there may be a "preferred" setting which results in optimal performance. These changes lay the foundation for the implementation of a memory metadata tagging graph transform, which will make sure that all tensors participating in an operator call is has a valid/optimal (storage type, memory layout) setting, and insert transition operators to transfer input tensors to the correct memory settings when necessary. An additional change that is required arises from the fact that in Vulkan, there is a limit on texture and buffer sizes. Therefore, the partitioner needs to account for the storage types and memory layouts supported by the operator implementation, and check if all tensors participating in a computation can be represented with some storage type, memory layout combination supported by the implementation. ## Changes Improvements to the operator registry: * Introduce utility functions to check the optimal and enabled storage types and memory layouts for an operator Improvements to the Partitioner: * Account for the storage types and memory layouts supported by an operator when deciding if a node should be partitioned * Improved logic for fusable ops (i.e. the permute/transpose before a mm which can be fused into linear) to check if the final target op is supported in Vulkan, and only partition those nodes if so. Otherwise, don't partition it so that it can be fused by another backend. Differential Revision: [D65428843](https://our.internmc.facebook.com/intern/diff/D65428843/) [ghstack-poisoned]
|
This pull request was exported from Phabricator. Differential Revision: D65428843 |
…memory layout" ## Context There are a variety of ways that tensors can be represented in Vulkan. The two main descriptors for how a tensor is laid out in memory is: 1. Storage Type (buffer or texture) 2. Memory Layout (which dim is packed along a texel, which dim has a stride of 1, etc.) Due to the differences between buffers and textures, and the differences between different memory layouts, an implementation for an operator may only support a specific set of (storage type, memory layout) combinations. Furthermore, if an operator implementation supports multiple (storage type, memory layout) combinations, there may be a "preferred" setting which results in optimal performance. These changes lay the foundation for the implementation of a memory metadata tagging graph transform, which will make sure that all tensors participating in an operator call is has a valid/optimal (storage type, memory layout) setting, and insert transition operators to transfer input tensors to the correct memory settings when necessary. An additional change that is required arises from the fact that in Vulkan, there is a limit on texture and buffer sizes. Therefore, the partitioner needs to account for the storage types and memory layouts supported by the operator implementation, and check if all tensors participating in a computation can be represented with some storage type, memory layout combination supported by the implementation. ## Changes Improvements to the operator registry: * Introduce utility functions to check the optimal and enabled storage types and memory layouts for an operator Improvements to the Partitioner: * Account for the storage types and memory layouts supported by an operator when deciding if a node should be partitioned * Improved logic for fusable ops (i.e. the permute/transpose before a mm which can be fused into linear) to check if the final target op is supported in Vulkan, and only partition those nodes if so. Otherwise, don't partition it so that it can be fused by another backend. Differential Revision: [D65428843](https://our.internmc.facebook.com/intern/diff/D65428843/) [ghstack-poisoned]
|
This pull request was exported from Phabricator. Differential Revision: D65428843 |
…memory layout" ## Context There are a variety of ways that tensors can be represented in Vulkan. The two main descriptors for how a tensor is laid out in memory is: 1. Storage Type (buffer or texture) 2. Memory Layout (which dim is packed along a texel, which dim has a stride of 1, etc.) Due to the differences between buffers and textures, and the differences between different memory layouts, an implementation for an operator may only support a specific set of (storage type, memory layout) combinations. Furthermore, if an operator implementation supports multiple (storage type, memory layout) combinations, there may be a "preferred" setting which results in optimal performance. These changes lay the foundation for the implementation of a memory metadata tagging graph transform, which will make sure that all tensors participating in an operator call is has a valid/optimal (storage type, memory layout) setting, and insert transition operators to transfer input tensors to the correct memory settings when necessary. An additional change that is required arises from the fact that in Vulkan, there is a limit on texture and buffer sizes. Therefore, the partitioner needs to account for the storage types and memory layouts supported by the operator implementation, and check if all tensors participating in a computation can be represented with some storage type, memory layout combination supported by the implementation. ## Changes Improvements to the operator registry: * Introduce utility functions to check the optimal and enabled storage types and memory layouts for an operator Improvements to the Partitioner: * Account for the storage types and memory layouts supported by an operator when deciding if a node should be partitioned * Improved logic for fusable ops (i.e. the permute/transpose before a mm which can be fused into linear) to check if the final target op is supported in Vulkan, and only partition those nodes if so. Otherwise, don't partition it so that it can be fused by another backend. Differential Revision: [D65428843](https://our.internmc.facebook.com/intern/diff/D65428843/) [ghstack-poisoned]
|
This pull request was exported from Phabricator. Differential Revision: D65428843 |
junpi3
approved these changes
Nov 5, 2024
facebook-github-bot
merged commit af124d1
into
gh/SS-JIA/136/base
37 of 41 checks passed
facebook-github-bot
temporarily deployed
to
cherry-pick-bot
GitHub Actions
Inactive
SS-JIA
added a commit
that referenced
this pull request
Nov 5, 2024
…ut (#6668) Pull Request resolved: #6635 ## Context There are a variety of ways that tensors can be represented in Vulkan. The two main descriptors for how a tensor is laid out in memory is: 1. Storage Type (buffer or texture) 2. Memory Layout (which dim is packed along a texel, which dim has a stride of 1, etc.) Due to the differences between buffers and textures, and the differences between different memory layouts, an implementation for an operator may only support a specific set of (storage type, memory layout) combinations. Furthermore, if an operator implementation supports multiple (storage type, memory layout) combinations, there may be a "preferred" setting which results in optimal performance. These changes lay the foundation for the implementation of a memory metadata tagging graph transform, which will make sure that all tensors participating in an operator call is has a valid/optimal (storage type, memory layout) setting, and insert transition operators to transfer input tensors to the correct memory settings when necessary. An additional change that is required arises from the fact that in Vulkan, there is a limit on texture and buffer sizes. Therefore, the partitioner needs to account for the storage types and memory layouts supported by the operator implementation, and check if all tensors participating in a computation can be represented with some storage type, memory layout combination supported by the implementation. ## Changes Improvements to the operator registry: * Introduce utility functions to check the optimal and enabled storage types and memory layouts for an operator Improvements to the Partitioner: * Account for the storage types and memory layouts supported by an operator when deciding if a node should be partitioned * Improved logic for fusable ops (i.e. the permute/transpose before a mm which can be fused into linear) to check if the final target op is supported in Vulkan, and only partition those nodes if so. Otherwise, don't partition it so that it can be fused by another backend. ghstack-source-id: 251883705 @exported-using-ghexport Differential Revision: [D65428843](https://our.internmc.facebook.com/intern/diff/D65428843/) Co-authored-by: Stephen Jia <ssjia@meta.com>
SS-JIA
added a commit
that referenced
this pull request
Nov 5, 2024
* [ET-VK] Refine paritioner to account for storage type and memory layout Pull Request resolved: #6635 ## Context There are a variety of ways that tensors can be represented in Vulkan. The two main descriptors for how a tensor is laid out in memory is: 1. Storage Type (buffer or texture) 2. Memory Layout (which dim is packed along a texel, which dim has a stride of 1, etc.) Due to the differences between buffers and textures, and the differences between different memory layouts, an implementation for an operator may only support a specific set of (storage type, memory layout) combinations. Furthermore, if an operator implementation supports multiple (storage type, memory layout) combinations, there may be a "preferred" setting which results in optimal performance. These changes lay the foundation for the implementation of a memory metadata tagging graph transform, which will make sure that all tensors participating in an operator call is has a valid/optimal (storage type, memory layout) setting, and insert transition operators to transfer input tensors to the correct memory settings when necessary. An additional change that is required arises from the fact that in Vulkan, there is a limit on texture and buffer sizes. Therefore, the partitioner needs to account for the storage types and memory layouts supported by the operator implementation, and check if all tensors participating in a computation can be represented with some storage type, memory layout combination supported by the implementation. ## Changes Improvements to the operator registry: * Introduce utility functions to check the optimal and enabled storage types and memory layouts for an operator Improvements to the Partitioner: * Account for the storage types and memory layouts supported by an operator when deciding if a node should be partitioned * Improved logic for fusable ops (i.e. the permute/transpose before a mm which can be fused into linear) to check if the final target op is supported in Vulkan, and only partition those nodes if so. Otherwise, don't partition it so that it can be fused by another backend. ghstack-source-id: 251883705 @exported-using-ghexport Differential Revision: [D65428843](https://our.internmc.facebook.com/intern/diff/D65428843/) * [ET-VK] Introduce memory metadata tagging pass Pull Request resolved: #6636 ## Context As title; implements the memory metadata tagging graph transform described in the dependent diff. See the comments for more details. ghstack-source-id: 251884020 @exported-using-ghexport Differential Revision: [D65428842](https://our.internmc.facebook.com/intern/diff/D65428842/) --------- Co-authored-by: Stephen Jia <ssjia@meta.com>
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):
Context
There are a variety of ways that tensors can be represented in Vulkan. The two main descriptors for how a tensor is laid out in memory is:
Due to the differences between buffers and textures, and the differences between different memory layouts, an implementation for an operator may only support a specific set of (storage type, memory layout) combinations.
Furthermore, if an operator implementation supports multiple (storage type, memory layout) combinations, there may be a "preferred" setting which results in optimal performance.
These changes lay the foundation for the implementation of a memory metadata tagging graph transform, which will make sure that all tensors participating in an operator call is has a valid/optimal (storage type, memory layout) setting, and insert transition operators to transfer input tensors to the correct memory settings when necessary.
An additional change that is required arises from the fact that in Vulkan, there is a limit on texture and buffer sizes. Therefore, the partitioner needs to account for the storage types and memory layouts supported by the operator implementation, and check if all tensors participating in a computation can be represented with some storage type, memory layout combination supported by the implementation.
Changes
Improvements to the operator registry:
Improvements to the Partitioner:
Differential Revision: D65428843