[ET-VK] Efficient tiled int8 matmul #9766
Merged
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
Introduce a optimized tiled implementation for computing the weight int8-quantized linear operation.
This implementation takes advantage of the following principles to squeeze out performance:
* Compute an output tile with each thread, rather than a single output element. This allows for better memory re-use of loaded input tensor data.
* Compute the output tile by iteratively loading tiles of the input matrices, caching them in registers, and then performing the `fma` accumulations to obtain a partial output. By splitting the data loading and computation into distinct steps, the GPU is able to perform latency hiding more effectively, i.e. switching to a warp that needs to perform compute when the current warp is waiting on data load
* Use a work group size of `{N, 1, 1}`. This makes it so that all the threads in a work group load the same row of the input matrx, and consecutive columns of the weight matrix. This way, the row of the input is kept hot in the cache, and accesses to the weight matrix can be coalesced due to the previous diff un-transposing the weight matrix.
Differential Revision: [D72066587](https://our.internmc.facebook.com/intern/diff/D72066587/)
[ghstack-poisoned]
🔗 Helpful Links🧪 See artifacts and rendered test results at hud.pytorch.org/pr/pytorch/executorch/9766
Note: Links to docs will display an error until the docs builds have been completed. ❌ 2 New Failures, 1 Cancelled JobAs of commit 7a20e01 with merge base 2aa7748 ( NEW FAILURES - The following jobs have failed:
CANCELLED JOB - The following job was cancelled. Please retry:
This comment was automatically generated by Dr. CI and updates every 15 minutes. |
SS-JIA
added a commit
that referenced
this pull request
Mar 31, 2025
## Context
Introduce a optimized tiled implementation for computing the weight int8-quantized linear operation.
This implementation takes advantage of the following principles to squeeze out performance:
* Compute an output tile with each thread, rather than a single output element. This allows for better memory re-use of loaded input tensor data.
* Compute the output tile by iteratively loading tiles of the input matrices, caching them in registers, and then performing the `fma` accumulations to obtain a partial output. By splitting the data loading and computation into distinct steps, the GPU is able to perform latency hiding more effectively, i.e. switching to a warp that needs to perform compute when the current warp is waiting on data load
* Use a work group size of `{N, 1, 1}`. This makes it so that all the threads in a work group load the same row of the input matrx, and consecutive columns of the weight matrix. This way, the row of the input is kept hot in the cache, and accesses to the weight matrix can be coalesced due to the previous diff un-transposing the weight matrix.
Differential Revision: [D72066587](https://our.internmc.facebook.com/intern/diff/D72066587/)
ghstack-source-id: 275129678
Pull Request resolved: #9766
facebook-github-bot
added
the
CLA Signed
|
This pull request was exported from Phabricator. Differential Revision: D72066587 |
This PR needs a
|
## Context
Introduce a optimized tiled implementation for computing the weight int8-quantized linear operation.
This implementation takes advantage of the following principles to squeeze out performance:
* Compute an output tile with each thread, rather than a single output element. This allows for better memory re-use of loaded input tensor data.
* Compute the output tile by iteratively loading tiles of the input matrices, caching them in registers, and then performing the `fma` accumulations to obtain a partial output. By splitting the data loading and computation into distinct steps, the GPU is able to perform latency hiding more effectively, i.e. switching to a warp that needs to perform compute when the current warp is waiting on data load
* Use a work group size of `{N, 1, 1}`. This makes it so that all the threads in a work group load the same row of the input matrx, and consecutive columns of the weight matrix. This way, the row of the input is kept hot in the cache, and accesses to the weight matrix can be coalesced due to the previous diff un-transposing the weight matrix.
Differential Revision: [D72066587](https://our.internmc.facebook.com/intern/diff/D72066587/)
[ghstack-poisoned]
SS-JIA
added a commit
that referenced
this pull request
Mar 31, 2025
Pull Request resolved: #9766 ## Context Introduce a optimized tiled implementation for computing the weight int8-quantized linear operation. This implementation takes advantage of the following principles to squeeze out performance: * Compute an output tile with each thread, rather than a single output element. This allows for better memory re-use of loaded input tensor data. * Compute the output tile by iteratively loading tiles of the input matrices, caching them in registers, and then performing the `fma` accumulations to obtain a partial output. By splitting the data loading and computation into distinct steps, the GPU is able to perform latency hiding more effectively, i.e. switching to a warp that needs to perform compute when the current warp is waiting on data load * Use a work group size of `{N, 1, 1}`. This makes it so that all the threads in a work group load the same row of the input matrx, and consecutive columns of the weight matrix. This way, the row of the input is kept hot in the cache, and accesses to the weight matrix can be coalesced due to the previous diff un-transposing the weight matrix. Differential Revision: [D72066587](https://our.internmc.facebook.com/intern/diff/D72066587/) ghstack-source-id: 275180032
|
This pull request was exported from Phabricator. Differential Revision: D72066587 |
trivedivivek
approved these changes
Apr 1, 2025
facebook-github-bot
merged commit 219e746
into
gh/SS-JIA/205/base
80 of 84 checks passed
facebook-github-bot
temporarily deployed
to
cherry-pick-bot
GitHub Actions
Inactive
SS-JIA
pushed a commit
that referenced
this pull request
Apr 1, 2025
Pull Request resolved: #9766 ## Context Introduce a optimized tiled implementation for computing the weight int8-quantized linear operation. This implementation takes advantage of the following principles to squeeze out performance: * Compute an output tile with each thread, rather than a single output element. This allows for better memory re-use of loaded input tensor data. * Compute the output tile by iteratively loading tiles of the input matrices, caching them in registers, and then performing the `fma` accumulations to obtain a partial output. By splitting the data loading and computation into distinct steps, the GPU is able to perform latency hiding more effectively, i.e. switching to a warp that needs to perform compute when the current warp is waiting on data load * Use a work group size of `{N, 1, 1}`. This makes it so that all the threads in a work group load the same row of the input matrx, and consecutive columns of the weight matrix. This way, the row of the input is kept hot in the cache, and accesses to the weight matrix can be coalesced due to the previous diff un-transposing the weight matrix. Differential Revision: [D72066587](https://our.internmc.facebook.com/intern/diff/D72066587/)
kirklandsign
pushed a commit
that referenced
this pull request
Apr 11, 2025
Pull Request resolved: #9766 ## Context Introduce a optimized tiled implementation for computing the weight int8-quantized linear operation. This implementation takes advantage of the following principles to squeeze out performance: * Compute an output tile with each thread, rather than a single output element. This allows for better memory re-use of loaded input tensor data. * Compute the output tile by iteratively loading tiles of the input matrices, caching them in registers, and then performing the `fma` accumulations to obtain a partial output. By splitting the data loading and computation into distinct steps, the GPU is able to perform latency hiding more effectively, i.e. switching to a warp that needs to perform compute when the current warp is waiting on data load * Use a work group size of `{N, 1, 1}`. This makes it so that all the threads in a work group load the same row of the input matrx, and consecutive columns of the weight matrix. This way, the row of the input is kept hot in the cache, and accesses to the weight matrix can be coalesced due to the previous diff un-transposing the weight matrix. Differential Revision: [D72066587](https://our.internmc.facebook.com/intern/diff/D72066587/)
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
Introduce a optimized tiled implementation for computing the weight int8-quantized linear operation.
This implementation takes advantage of the following principles to squeeze out performance:
fmaaccumulations to obtain a partial output. By splitting the data loading and computation into distinct steps, the GPU is able to perform latency hiding more effectively, i.e. switching to a warp that needs to perform compute when the current warp is waiting on data load{N, 1, 1}. This makes it so that all the threads in a work group load the same row of the input matrx, and consecutive columns of the weight matrix. This way, the row of the input is kept hot in the cache, and accesses to the weight matrix can be coalesced due to the previous diff un-transposing the weight matrix.Differential Revision: D72066587