[ET-VK] Using a single GPU buffer for all tensor uniforms. #7015
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
This diff changes Tensor class to store all uniforms in a single uniform buffer. Entities stored in uniforms ie. size, stride, numel and logical limits are now stored in a single buffer and their offsets are stored as unsigned ints in Tensor class. Other changes includes: Adding a new ctor for ParamsBuffer class to allow allocation with size without data ptr. Adding an offset input to Buffer::data function. Adding an offset parameter to BufferBindInfo ctor, so additional offset can be supplied when binding a buffer. Differential Revision: [D65841750](https://our.internmc.facebook.com/intern/diff/D65841750/) [ghstack-poisoned]
🔗 Helpful Links🧪 See artifacts and rendered test results at hud.pytorch.org/pr/pytorch/executorch/7015
Note: Links to docs will display an error until the docs builds have been completed. ✅ No FailuresAs of commit f3bc1e6 with merge base a04a87f ( 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: D65841750 |
nathanaelsee
approved these changes
Nov 26, 2024
This diff changes Tensor class to store all uniforms in a single uniform buffer. Entities stored in uniforms ie. size, stride, numel and logical limits are now stored in a single buffer and their offsets are stored as unsigned ints in Tensor class. Other changes includes: Adding a new ctor for ParamsBuffer class to allow allocation with size without data ptr. Adding an offset input to Buffer::data function. Adding an offset parameter to BufferBindInfo ctor, so additional offset can be supplied when binding a buffer. Differential Revision: [D65841750](https://our.internmc.facebook.com/intern/diff/D65841750/) [ghstack-poisoned]
trivedivivek
added a commit
that referenced
this pull request
Nov 26, 2024
Pull Request resolved: #7015 This diff changes Tensor class to store all uniforms in a single uniform buffer. Entities stored in uniforms ie. size, stride, numel and logical limits are now stored in a single buffer and their offsets are stored as unsigned ints in Tensor class. Other changes includes: Adding a new ctor for ParamsBuffer class to allow allocation with size without data ptr. Adding an offset input to Buffer::data function. Adding an offset parameter to BufferBindInfo ctor, so additional offset can be supplied when binding a buffer. ghstack-source-id: 255514948 @exported-using-ghexport Differential Revision: [D65841750](https://our.internmc.facebook.com/intern/diff/D65841750/)
|
This pull request was exported from Phabricator. Differential Revision: D65841750 |
This diff changes Tensor class to store all uniforms in a single uniform buffer. Entities stored in uniforms ie. size, stride, numel and logical limits are now stored in a single buffer and their offsets are stored as unsigned ints in Tensor class. Other changes includes: Adding a new ctor for ParamsBuffer class to allow allocation with size without data ptr. Adding an offset input to Buffer::data function. Adding an offset parameter to BufferBindInfo ctor, so additional offset can be supplied when binding a buffer. Differential Revision: [D65841750](https://our.internmc.facebook.com/intern/diff/D65841750/) [ghstack-poisoned]
trivedivivek
added a commit
that referenced
this pull request
Nov 26, 2024
Pull Request resolved: #7015 This diff changes Tensor class to store all uniforms in a single uniform buffer. Entities stored in uniforms ie. size, stride, numel and logical limits are now stored in a single buffer and their offsets are stored as unsigned ints in Tensor class. Other changes includes: Adding a new ctor for ParamsBuffer class to allow allocation with size without data ptr. Adding an offset input to Buffer::data function. Adding an offset parameter to BufferBindInfo ctor, so additional offset can be supplied when binding a buffer. ghstack-source-id: 255516791 @exported-using-ghexport Differential Revision: [D65841750](https://our.internmc.facebook.com/intern/diff/D65841750/)
|
This pull request was exported from Phabricator. Differential Revision: D65841750 |
This diff changes Tensor class to store all uniforms in a single uniform buffer. Entities stored in uniforms ie. size, stride, numel and logical limits are now stored in a single buffer and their offsets are stored as unsigned ints in Tensor class. Other changes includes: Adding a new ctor for ParamsBuffer class to allow allocation with size without data ptr. Adding an offset input to Buffer::data function. Adding an offset parameter to BufferBindInfo ctor, so additional offset can be supplied when binding a buffer. Differential Revision: [D65841750](https://our.internmc.facebook.com/intern/diff/D65841750/) [ghstack-poisoned]
trivedivivek
added a commit
that referenced
this pull request
Nov 27, 2024
Pull Request resolved: #7015 This diff changes Tensor class to store all uniforms in a single uniform buffer. Entities stored in uniforms ie. size, stride, numel and logical limits are now stored in a single buffer and their offsets are stored as unsigned ints in Tensor class. Other changes includes: Adding a new ctor for ParamsBuffer class to allow allocation with size without data ptr. Adding an offset input to Buffer::data function. Adding an offset parameter to BufferBindInfo ctor, so additional offset can be supplied when binding a buffer. ghstack-source-id: 255668611 @exported-using-ghexport Differential Revision: [D65841750](https://our.internmc.facebook.com/intern/diff/D65841750/)
|
This pull request was exported from Phabricator. Differential Revision: D65841750 |
junpi3
approved these changes
Dec 2, 2024
This diff changes Tensor class to store all uniforms in a single uniform buffer. Entities stored in uniforms ie. size, stride, numel and logical limits are now stored in a single buffer and their offsets are stored as unsigned ints in Tensor class. Other changes includes: Adding a new ctor for ParamsBuffer class to allow allocation with size without data ptr. Adding an offset input to Buffer::data function. Adding an offset parameter to BufferBindInfo ctor, so additional offset can be supplied when binding a buffer. Differential Revision: [D65841750](https://our.internmc.facebook.com/intern/diff/D65841750/) [ghstack-poisoned]
|
This pull request was exported from Phabricator. Differential Revision: D65841750 |
This diff changes Tensor class to store all uniforms in a single uniform buffer. Entities stored in uniforms ie. size, stride, numel and logical limits are now stored in a single buffer and their offsets are stored as unsigned ints in Tensor class. Other changes includes: Adding a new ctor for ParamsBuffer class to allow allocation with size without data ptr. Adding an offset input to Buffer::data function. Adding an offset parameter to BufferBindInfo ctor, so additional offset can be supplied when binding a buffer. Differential Revision: [D65841750](https://our.internmc.facebook.com/intern/diff/D65841750/) [ghstack-poisoned]
|
This pull request was exported from Phabricator. Differential Revision: D65841750 |
This diff changes Tensor class to store all uniforms in a single uniform buffer. Entities stored in uniforms ie. size, stride, numel and logical limits are now stored in a single buffer and their offsets are stored as unsigned ints in Tensor class. Other changes includes: Adding a new ctor for ParamsBuffer class to allow allocation with size without data ptr. Adding an offset input to Buffer::data function. Adding an offset parameter to BufferBindInfo ctor, so additional offset can be supplied when binding a buffer. Differential Revision: [D65841750](https://our.internmc.facebook.com/intern/diff/D65841750/) [ghstack-poisoned]
|
This pull request was exported from Phabricator. Differential Revision: D65841750 |
This diff changes Tensor class to store all uniforms in a single uniform buffer. Entities stored in uniforms ie. size, stride, numel and logical limits are now stored in a single buffer and their offsets are stored as unsigned ints in Tensor class. Other changes includes: Adding a new ctor for ParamsBuffer class to allow allocation with size without data ptr. Adding an offset input to Buffer::data function. Adding an offset parameter to BufferBindInfo ctor, so additional offset can be supplied when binding a buffer. Differential Revision: [D65841750](https://our.internmc.facebook.com/intern/diff/D65841750/) [ghstack-poisoned]
|
This pull request was exported from Phabricator. Differential Revision: D65841750 |
This diff changes Tensor class to store all uniforms in a single uniform buffer. Entities stored in uniforms ie. size, stride, numel and logical limits are now stored in a single buffer and their offsets are stored as unsigned ints in Tensor class. Other changes includes: Adding a new ctor for ParamsBuffer class to allow allocation with size without data ptr. Adding an offset input to Buffer::data function. Adding an offset parameter to BufferBindInfo ctor, so additional offset can be supplied when binding a buffer. Differential Revision: [D65841750](https://our.internmc.facebook.com/intern/diff/D65841750/) [ghstack-poisoned]
|
This pull request was exported from Phabricator. Differential Revision: D65841750 |
facebook-github-bot
temporarily deployed
to
cherry-pick-bot
GitHub Actions
Inactive
kirklandsign
pushed a commit
that referenced
this pull request
Dec 5, 2024
Pull Request resolved: #7015 This diff changes Tensor class to store all uniforms in a single uniform buffer. Entities stored in uniforms ie. size, stride, numel and logical limits are now stored in a single buffer and their offsets are stored as unsigned ints in Tensor class. Other changes includes: Adding a new ctor for ParamsBuffer class to allow allocation with size without data ptr. Adding an offset input to Buffer::data function. Adding an offset parameter to BufferBindInfo ctor, so additional offset can be supplied when binding a buffer. ghstack-source-id: 256728325 @exported-using-ghexport Differential Revision: [D65841750](https://our.internmc.facebook.com/intern/diff/D65841750/) Co-authored-by: Vivek Trivedi <[email protected]>
SS-JIA
added a commit
that referenced
this pull request
Jan 2, 2025
…ructor ## Context I discovered this bug when trying to execute the `vulkan_compute_api_test` binary on Windows. Almost all the tests were failing, with compute shaders producing incorrect results. After bisecting the change, it turns out the culprit is #7015. The diff introduced an alternative templated constructor for `ParamsBuffer` which would initialize an empty UBO with a specified size instead of wrapping a pre-existing object. The issue is that these constructors are ambiguous because they both are template constructors and both only accept one argument. Therefore, the original constructor would be called when certain callsites intended to call the new constructor. This results in a UBO being created with an incorrect size, and resulted in the tensor's metadata being passed incorrectly into a compute shader. To fix, I added a dummy argument into the new constructor for disambiguation purposes. I also changed it so that it's not templated, since there's no reason for it to be templated. Differential Revision: [D67770791](https://our.internmc.facebook.com/intern/diff/D67770791/) [ghstack-poisoned]
SS-JIA
added a commit
that referenced
this pull request
Jan 2, 2025
## Context Recently #7015 was implemented so that all tensor metadata (e.g. sizes, strides) would be stored in a single UBO instead of with separate UBO objects. This helps with memory savings presumably due to defragmentation of memory allocations. However, once the change was introduced, I noticed two new warnings produced by the Vulkan Validation Layer. The first complains that the offset of a UBO descriptor is not a multiple of the `minUniformBufferOffsetAlignment` field reported by the physical device properties. The second complains that the range of a UBO descriptor exceeds the offset + range of the underlying UBO object. # Solution To address the first one, instead of using `sizeof(utils::ivec4)` to determine the offset per metadata field, check the `minUniformBufferOffsetAlignment` field of reported by the device and use that instead. The second warning arises because the logic in the constructor of `BufferBindInfo` had a mistake; instead of using the range of the underlying UBO object, it should use the range subtracted by the user specified offset. Differential Revision: [D67770792](https://our.internmc.facebook.com/intern/diff/D67770792/) [ghstack-poisoned]
SS-JIA
added a commit
that referenced
this pull request
Jan 2, 2025
## Context Recently #7015 was implemented so that all tensor metadata (e.g. sizes, strides) would be stored in a single UBO instead of with separate UBO objects. This helps with memory savings presumably due to defragmentation of memory allocations. However, once the change was introduced, I noticed two new warnings produced by the Vulkan Validation Layer. The first complains that the offset of a UBO descriptor is not a multiple of the `minUniformBufferOffsetAlignment` field reported by the physical device properties. The second complains that the range of a UBO descriptor exceeds the offset + range of the underlying UBO object. # Solution To address the first one, instead of using `sizeof(utils::ivec4)` to determine the offset per metadata field, check the `minUniformBufferOffsetAlignment` field of reported by the device and use that instead. The second warning arises because the logic in the constructor of `BufferBindInfo` had a mistake; instead of using the range of the underlying UBO object, it should use the range subtracted by the user specified offset. Differential Revision: [D67770792](https://our.internmc.facebook.com/intern/diff/D67770792/) ghstack-source-id: 260031110 Pull Request resolved: #7480
SS-JIA
added a commit
that referenced
this pull request
Jan 3, 2025
…ructor (#7482) ## Context I discovered this bug when trying to execute the `vulkan_compute_api_test` binary on Windows. Almost all the tests were failing, with compute shaders producing incorrect results. After bisecting the change, it turns out the culprit is #7015. The diff introduced an alternative templated constructor for `ParamsBuffer` which would initialize an empty UBO with a specified size instead of wrapping a pre-existing object. The issue is that these constructors are ambiguous because they both are template constructors and both only accept one argument. Therefore, the original constructor would be called when certain callsites intended to call the new constructor. This results in a UBO being created with an incorrect size, and resulted in the tensor's metadata being passed incorrectly into a compute shader. To fix, I added a dummy argument into the new constructor for disambiguation purposes. I also changed it so that it's not templated, since there's no reason for it to be templated. Differential Revision: [D67770791](https://our.internmc.facebook.com/intern/diff/D67770791/) ghstack-source-id: 260031108 Pull Request resolved: #7478 Co-authored-by: Stephen Jia <[email protected]>
SS-JIA
added a commit
that referenced
this pull request
Jan 3, 2025
…ed (#7483) * [ET-VK][ez] Fix undefined behaviour in ambiguous `ParamsBuffer` constructor ## Context I discovered this bug when trying to execute the `vulkan_compute_api_test` binary on Windows. Almost all the tests were failing, with compute shaders producing incorrect results. After bisecting the change, it turns out the culprit is #7015. The diff introduced an alternative templated constructor for `ParamsBuffer` which would initialize an empty UBO with a specified size instead of wrapping a pre-existing object. The issue is that these constructors are ambiguous because they both are template constructors and both only accept one argument. Therefore, the original constructor would be called when certain callsites intended to call the new constructor. This results in a UBO being created with an incorrect size, and resulted in the tensor's metadata being passed incorrectly into a compute shader. To fix, I added a dummy argument into the new constructor for disambiguation purposes. I also changed it so that it's not templated, since there's no reason for it to be templated. Differential Revision: [D67770791](https://our.internmc.facebook.com/intern/diff/D67770791/) ghstack-source-id: 260031108 Pull Request resolved: #7478 * [ET-VK] Create Pipeline layouts with push constant ranges when required ## Context #7223 added the ability to use push constants in shaders. However, one thing the diff missed was not specifying that the compute pipeline layout needed to include a push constant upon creation. The Vulkan validation layers warns against this, and on certain GPUs such as the integrated Intel GPU on my windows laptop compute shaders will produce incorrect output. This diff makes the change such that the compute pipeline layout will be created with a push constant block if necessary. ## Solution Change the key of the pipeline layout cache to accept an additional push constant size field. The push constant size will be used to create the pipeline layout with a push constant block of the specified size. Differential Revision: [D67770793](https://our.internmc.facebook.com/intern/diff/D67770793/) ghstack-source-id: 260031109 Pull Request resolved: #7479 --------- Co-authored-by: Stephen Jia <[email protected]>
SS-JIA
added a commit
that referenced
this pull request
Jan 3, 2025
* [ET-VK][ez] Fix undefined behaviour in ambiguous `ParamsBuffer` constructor ## Context I discovered this bug when trying to execute the `vulkan_compute_api_test` binary on Windows. Almost all the tests were failing, with compute shaders producing incorrect results. After bisecting the change, it turns out the culprit is #7015. The diff introduced an alternative templated constructor for `ParamsBuffer` which would initialize an empty UBO with a specified size instead of wrapping a pre-existing object. The issue is that these constructors are ambiguous because they both are template constructors and both only accept one argument. Therefore, the original constructor would be called when certain callsites intended to call the new constructor. This results in a UBO being created with an incorrect size, and resulted in the tensor's metadata being passed incorrectly into a compute shader. To fix, I added a dummy argument into the new constructor for disambiguation purposes. I also changed it so that it's not templated, since there's no reason for it to be templated. Differential Revision: [D67770791](https://our.internmc.facebook.com/intern/diff/D67770791/) ghstack-source-id: 260031108 Pull Request resolved: #7478 * [ET-VK] Create Pipeline layouts with push constant ranges when required ## Context #7223 added the ability to use push constants in shaders. However, one thing the diff missed was not specifying that the compute pipeline layout needed to include a push constant upon creation. The Vulkan validation layers warns against this, and on certain GPUs such as the integrated Intel GPU on my windows laptop compute shaders will produce incorrect output. This diff makes the change such that the compute pipeline layout will be created with a push constant block if necessary. ## Solution Change the key of the pipeline layout cache to accept an additional push constant size field. The push constant size will be used to create the pipeline layout with a push constant block of the specified size. Differential Revision: [D67770793](https://our.internmc.facebook.com/intern/diff/D67770793/) ghstack-source-id: 260031109 Pull Request resolved: #7479 * [ET-VK] Fix metadata UBO VVL warnings ## Context Recently #7015 was implemented so that all tensor metadata (e.g. sizes, strides) would be stored in a single UBO instead of with separate UBO objects. This helps with memory savings presumably due to defragmentation of memory allocations. However, once the change was introduced, I noticed two new warnings produced by the Vulkan Validation Layer. The first complains that the offset of a UBO descriptor is not a multiple of the `minUniformBufferOffsetAlignment` field reported by the physical device properties. The second complains that the range of a UBO descriptor exceeds the offset + range of the underlying UBO object. # Solution To address the first one, instead of using `sizeof(utils::ivec4)` to determine the offset per metadata field, check the `minUniformBufferOffsetAlignment` field of reported by the device and use that instead. The second warning arises because the logic in the constructor of `BufferBindInfo` had a mistake; instead of using the range of the underlying UBO object, it should use the range subtracted by the user specified offset. Differential Revision: [D67770792](https://our.internmc.facebook.com/intern/diff/D67770792/) ghstack-source-id: 260031110 Pull Request resolved: #7480 --------- Co-authored-by: Stephen Jia <[email protected]>
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):
This diff changes Tensor class to store all uniforms in a single uniform buffer.
Entities stored in uniforms ie. size, stride, numel and logical limits are now stored in a single buffer and their offsets are stored as unsigned ints in Tensor class.
Other changes includes:
Adding a new ctor for ParamsBuffer class to allow allocation with size without data ptr.
Adding an offset input to Buffer::data function.
Adding an offset parameter to BufferBindInfo ctor, so additional offset can be supplied when binding a buffer.
Differential Revision: D65841750