[ET-VK][INT4-MM] Move QMat2 to buffer storage and scales_and_zeros to Channels Packed #5515
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
… Channels Packed
Storing QMat2 in a texture gives way to two main problems:
- Indexing is a mess and additional computation is required to take into account the fact that we are reading ivec4's and only using half of the values
- There is no texel fetching in int8. The texel is read in int32 and needs to be casted
Keeping QMat2 in a buffer performs better because, although reading from buffers is slower, removing the extra computation compensates for this.
{F1863459327}
This diff also moves the scales_and_zeros tensor to Channels Packed in texture implementations because it just makes more sense, I had done some terrible indexing shennanigans before.
Differential Revision: [D62504978](https://our.internmc.facebook.com/intern/diff/D62504978/)
[ghstack-poisoned]
🔗 Helpful Links🧪 See artifacts and rendered test results at hud.pytorch.org/pr/pytorch/executorch/5515
Note: Links to docs will display an error until the docs builds have been completed. ✅ No FailuresAs of commit 2144171 with merge base 0ec003b ( This comment was automatically generated by Dr. CI and updates every 15 minutes. |
facebook-github-bot
added
the
CLA Signed
SS-JIA
added a commit
that referenced
this pull request
Sep 20, 2024
… Channels Packed
Storing QMat2 in a texture gives way to two main problems:
- Indexing is a mess and additional computation is required to take into account the fact that we are reading ivec4's and only using half of the values
- There is no texel fetching in int8. The texel is read in int32 and needs to be casted
Keeping QMat2 in a buffer performs better because, although reading from buffers is slower, removing the extra computation compensates for this.
{F1863459327}
This diff also moves the scales_and_zeros tensor to Channels Packed in texture implementations because it just makes more sense, I had done some terrible indexing shennanigans before.
Differential Revision: [D62504978](https://our.internmc.facebook.com/intern/diff/D62504978/)
ghstack-source-id: 243833256
Pull Request resolved: #5515
|
This pull request was exported from Phabricator. Differential Revision: D62504978 |
…nd_zeros to Channels Packed"
Storing QMat2 in a texture gives way to two main problems:
- Indexing is a mess and additional computation is required to take into account the fact that we are reading ivec4's and only using half of the values
- There is no texel fetching in int8. The texel is read in int32 and needs to be casted
Keeping QMat2 in a buffer performs better because, although reading from buffers is slower, removing the extra computation compensates for this.
{F1863459327}
This diff also moves the scales_and_zeros tensor to Channels Packed in texture implementations because it just makes more sense, I had done some terrible indexing shennanigans before.
Differential Revision: [D62504978](https://our.internmc.facebook.com/intern/diff/D62504978/)
[ghstack-poisoned]
|
This pull request was exported from Phabricator. Differential Revision: D62504978 |
SS-JIA
added a commit
that referenced
this pull request
Sep 23, 2024
… Channels Packed Pull Request resolved: #5515 Storing QMat2 in a texture gives way to two main problems: - Indexing is a mess and additional computation is required to take into account the fact that we are reading ivec4's and only using half of the values - There is no texel fetching in int8. The texel is read in int32 and needs to be casted Keeping QMat2 in a buffer performs better because, although reading from buffers is slower, removing the extra computation compensates for this. {F1863459327} This diff also moves the scales_and_zeros tensor to Channels Packed in texture implementations because it just makes more sense, I had done some terrible indexing shennanigans before. ghstack-source-id: 244258611 @exported-using-ghexport Differential Revision: [D62504978](https://our.internmc.facebook.com/intern/diff/D62504978/)
yipjustin
approved these changes
Sep 23, 2024
|
This pull request has been merged in 2eae7a9. |
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):
Storing QMat2 in a texture gives way to two main problems:
Keeping QMat2 in a buffer performs better because, although reading from buffers is slower, removing the extra computation compensates for this.
{F1863459327}
This diff also moves the scales_and_zeros tensor to Channels Packed in texture implementations because it just makes more sense, I had done some terrible indexing shennanigans before.
Differential Revision: D62504978