pytorch
Support int8 texture tensors without requiring int8 buffers
Summary:
Context
By default, storage buffers in Vulkan must contain 32 bit data types; using 8 bit and 16 bit data types in buffers can be enabled optionally by supporting the VK_KHR_8bit_storage extension or the VK_KHR_16bit_storage extension respectively.
Previously, 8-bit and 16-bit tensors were enabled by using those extensions; however, this meant that 8-bit and 16-bit tensors could not be used if the Vulkan driver does not support the corresponding extension.
This diff adds support for 8-bit texture-backed tensors without the need for the VK_KHR_8bit_storage extension. This is done by introducing shaders that manually pack and repack 4 8-bit integers into a single int32 value. Once the tensor data has been transferred to an image texture (which will use the VK_FORMAT_R8G8B8A8_SINT image format) the extension will no longer be required.
Differential Revision: D60536832
:link: Helpful Links
:test_tube: See artifacts and rendered test results at hud.pytorch.org/pr/pytorch/executorch/4485
- :page_facing_up: Preview Python docs built from this PR
Note: Links to docs will display an error until the docs builds have been completed.
:white_check_mark: No Failures
As of commit c269cf972a097429325515c962a15638749d5897 with merge base 1090bcd26e683188ebc1ebb056f174b35930392c (
This comment was automatically generated by Dr. CI and updates every 15 minutes.
![pytorch-bot[bot] avatar](https://avatars.githubusercontent.com/in/40112?v=4 "Published by a pub.dev verified publisher"){kind=small}
This pull request has been merged in pytorch/executorch@448c7d3c07125c1f13e24ff81de7473a6a446916.