[ET-VK] Use dim order as the source of truth for tensor strides #4844
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
… account for small numerical differences ## Context Introduce a `check_close` function in `vulkan_compute_api_test` to account for small numerical differences when executing on the GPU. Otherwise some tests will be reported as failures on some platforms (i.e. Mac). Differential Revision: [D61666459](https://our.internmc.facebook.com/intern/diff/D61666459/) [ghstack-poisoned]
## Context This diff lays the foundation for the implementation of "view operators". These operators will not dispatch any shaders; they will be solely responsible for updating the sizes and strides metadata of the output tensor, which will use the same storage resource as the input tensor. These ops will be implemented by adding a "no-op" `ExecuteNode` instance to the `ComputeGraph`, which does not contain a shader but does contain a resize function to update sizes and strides upon a resize. This diff allows `ShaderInfo` to empty initialize, and add a bool operator to check if the ShaderInfo actually points to valid shader code. This will be used to construct "no-op" `ExecuteNode` instances which can be used to implement view operators. Differential Revision: [D61666460](https://our.internmc.facebook.com/intern/diff/D61666460/) [ghstack-poisoned]
## Context This diff lays the foundation for the implementation of "view operators". These operators will not dispatch any shaders; they will be solely responsible for updating the sizes and strides metadata of the output tensor, which will use the same storage resource as the input tensor. These ops will be implemented by adding a "no-op" ExecuteNode instance to the ComputeGraph, which does not contain a shader but does contain a resize function to update sizes and strides upon a resize. This diff allows `ExecuteNode` to empty initialize. If the `ShaderInfo` of the `ExecuteNode` instance is not valid, then the `execute()` function of the `ExecuteNode` instance will be a no-op. Differential Revision: [D61666465](https://our.internmc.facebook.com/intern/diff/D61666465/) [ghstack-poisoned]
## Context This diff introduces dim order as a core metadata member of `vTensor`. The dim order is a better source of truth for the data layout of a buffer backed tensor because it is less ambiguous than the strides; when multiple dims have a stride of 1, it is ambiguous which dim comes "before" the other. This amibiguity makes it impossible to update the strides correctly when the tensor is resized. ## Changes * Introduce `dim_order_` as a core metadata member of `vTensor` class * Tensor view construction now accepts `new_dim_order` as an argument instead of `new_strides` * Introduce `virtual_reconfigure` member function of `vTensor` which is an extension of `virtual_resize` that allows updating of sizes as well as strides (via updating the dim order) Differential Revision: [D61666464](https://our.internmc.facebook.com/intern/diff/D61666464/) [ghstack-poisoned]
🔗 Helpful Links🧪 See artifacts and rendered test results at hud.pytorch.org/pr/pytorch/executorch/4844
Note: Links to docs will display an error until the docs builds have been completed. ✅ No FailuresAs of commit d34c1a9 with merge base bfc5b17 ( 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: D61666464 |
This was referenced Aug 22, 2024
…th for tensor strides" ## Context This diff introduces dim order as a core metadata member of `vTensor`. The dim order is a better source of truth for the data layout of a buffer backed tensor because it is less ambiguous than the strides; when multiple dims have a stride of 1, it is ambiguous which dim comes "before" the other. This amibiguity makes it impossible to update the strides correctly when the tensor is resized. ## Changes * Introduce `dim_order_` as a core metadata member of `vTensor` class * Tensor view construction now accepts `new_dim_order` as an argument instead of `new_strides` * Introduce `virtual_reconfigure` member function of `vTensor` which is an extension of `virtual_resize` that allows updating of sizes as well as strides (via updating the dim order) Differential Revision: [D61666464](https://our.internmc.facebook.com/intern/diff/D61666464/) [ghstack-poisoned]
…rides" ## Context This diff introduces dim order as a core metadata member of `vTensor`. The dim order is a better source of truth for the data layout of a buffer backed tensor because it is less ambiguous than the strides; when multiple dims have a stride of 1, it is ambiguous which dim comes "before" the other. This amibiguity makes it impossible to update the strides correctly when the tensor is resized. ## Changes * Introduce `dim_order_` as a core metadata member of `vTensor` class * Tensor view construction now accepts `new_dim_order` as an argument instead of `new_strides` * Introduce `virtual_reconfigure` member function of `vTensor` which is an extension of `virtual_resize` that allows updating of sizes as well as strides (via updating the dim order) Differential Revision: [D61666464](https://our.internmc.facebook.com/intern/diff/D61666464/) [ghstack-poisoned]
|
This pull request was exported from Phabricator. Differential Revision: D61666464 |
junpi3
approved these changes
Aug 22, 2024
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
This diff introduces dim order as a core metadata member of
vTensor. The dim order is a better source of truth for the data layout of a buffer backed tensor because it is less ambiguous than the strides; when multiple dims have a stride of 1, it is ambiguous which dim comes "before" the other. This amibiguity makes it impossible to update the strides correctly when the tensor is resized.Changes
dim_order_as a core metadata member ofvTensorclassnew_dim_orderas an argument instead ofnew_stridesvirtual_reconfiguremember function ofvTensorwhich is an extension ofvirtual_resizethat allows updating of sizes as well as strides (via updating the dim order)Differential Revision: D61666464