[ET-VK][qlinear] Faster weight only quantized linear gemv kernel #12444
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
## Changes * Introduce a new compute shader for int4 linear's gemv cases that performs much better than the existing shader. This shader is inspired from MNN's gemv_1x1_conv_buf.cl shader. With this compute kernel, transformer models' text generation can execute much faster than before. On Samsung Galaxy S24 for Llama 3.2 1B, generating 128 tokens: Before: ~25 tok/s After: ~49 tok/s ## Why this new shader is faster The biggest reason is due to vectorized loading of the uint4 weight buffer. This new shader loads the weight buffer as a buffer/image of `uvec4`, whereas the old shader loads the weight buffer as a buffer/image of `u8vec4`. Using the Adreno Offline Compiler, I found that in the former, only one load instruction was used to load from the weight tensor, whereas in the latter 16 load instructions were used to load from the weight tensor. It appears that the data loading was not being vectorized at the assembly level. This is potentially behaviour that can be approved in the SPIR-V shader compiler. An additional factor is better weight packing layout. The new prepacking routine results in better memory coalescing between threads in a work group. The final major factor is the use of tree based reduction to co-operatively reduce partial results into the final output. Previously, a single thread was responsible for the final reduction. ## Future Work * Introduce faster shader for int4 linear gemm cases * Update QCSNW to also use these updated shaders Differential Revision: [D78275584](https://our.internmc.facebook.com/intern/diff/D78275584/) [ghstack-poisoned]
🔗 Helpful Links🧪 See artifacts and rendered test results at hud.pytorch.org/pr/pytorch/executorch/12444
Note: Links to docs will display an error until the docs builds have been completed. ❌ 2 New FailuresAs of commit 7f88379 with merge base dd4488d ( NEW FAILURES - The following jobs have failed:
This comment was automatically generated by Dr. CI and updates every 15 minutes. |
SS-JIA
added a commit
that referenced
this pull request
Jul 14, 2025
## Changes * Introduce a new compute shader for int4 linear's gemv cases that performs much better than the existing shader. This shader is inspired from MNN's gemv_1x1_conv_buf.cl shader. With this compute kernel, transformer models' text generation can execute much faster than before. On Samsung Galaxy S24 for Llama 3.2 1B, generating 128 tokens: Before: ~25 tok/s After: ~49 tok/s ## Why this new shader is faster The biggest reason is due to vectorized loading of the uint4 weight buffer. This new shader loads the weight buffer as a buffer/image of `uvec4`, whereas the old shader loads the weight buffer as a buffer/image of `u8vec4`. Using the Adreno Offline Compiler, I found that in the former, only one load instruction was used to load from the weight tensor, whereas in the latter 16 load instructions were used to load from the weight tensor. It appears that the data loading was not being vectorized at the assembly level. This is potentially behaviour that can be approved in the SPIR-V shader compiler. An additional factor is better weight packing layout. The new prepacking routine results in better memory coalescing between threads in a work group. The final major factor is the use of tree based reduction to co-operatively reduce partial results into the final output. Previously, a single thread was responsible for the final reduction. ## Future Work * Introduce faster shader for int4 linear gemm cases * Update QCSNW to also use these updated shaders Differential Revision: [D78275584](https://our.internmc.facebook.com/intern/diff/D78275584/) ghstack-source-id: 296055475 Pull Request resolved: #12444
facebook-github-bot
added
the
CLA Signed
|
This pull request was exported from Phabricator. Differential Revision: D78275584 |
This PR needs a
|
…kernel" ## Changes * Introduce a new compute shader for int4 linear's gemv cases that performs much better than the existing shader. This shader is inspired from MNN's gemv_1x1_conv_buf.cl shader. With this compute kernel, transformer models' text generation can execute much faster than before. On Samsung Galaxy S24 for Llama 3.2 1B, generating 128 tokens: Before: ~25 tok/s After: ~49 tok/s ## Why this new shader is faster The biggest reason is due to vectorized loading of the uint4 weight buffer. This new shader loads the weight buffer as a buffer/image of `uvec4`, whereas the old shader loads the weight buffer as a buffer/image of `u8vec4`. Using the Adreno Offline Compiler, I found that in the former, only one load instruction was used to load from the weight tensor, whereas in the latter 16 load instructions were used to load from the weight tensor. It appears that the data loading was not being vectorized at the assembly level. This is potentially behaviour that can be approved in the SPIR-V shader compiler. An additional factor is better weight packing layout. The new prepacking routine results in better memory coalescing between threads in a work group. The final major factor is the use of tree based reduction to co-operatively reduce partial results into the final output. Previously, a single thread was responsible for the final reduction. ## Future Work * Introduce faster shader for int4 linear gemm cases * Update QCSNW to also use these updated shaders Differential Revision: [D78275584](https://our.internmc.facebook.com/intern/diff/D78275584/) [ghstack-poisoned]
SS-JIA
added a commit
that referenced
this pull request
Jul 14, 2025
Pull Request resolved: #12444 ## Changes * Introduce a new compute shader for int4 linear's gemv cases that performs much better than the existing shader. This shader is inspired from MNN's gemv_1x1_conv_buf.cl shader. With this compute kernel, transformer models' text generation can execute much faster than before. On Samsung Galaxy S24 for Llama 3.2 1B, generating 128 tokens: Before: ~25 tok/s After: ~49 tok/s ## Why this new shader is faster The biggest reason is due to vectorized loading of the uint4 weight buffer. This new shader loads the weight buffer as a buffer/image of `uvec4`, whereas the old shader loads the weight buffer as a buffer/image of `u8vec4`. Using the Adreno Offline Compiler, I found that in the former, only one load instruction was used to load from the weight tensor, whereas in the latter 16 load instructions were used to load from the weight tensor. It appears that the data loading was not being vectorized at the assembly level. This is potentially behaviour that can be approved in the SPIR-V shader compiler. An additional factor is better weight packing layout. The new prepacking routine results in better memory coalescing between threads in a work group. The final major factor is the use of tree based reduction to co-operatively reduce partial results into the final output. Previously, a single thread was responsible for the final reduction. ## Future Work * Introduce faster shader for int4 linear gemm cases * Update QCSNW to also use these updated shaders Differential Revision: [D78275584](https://our.internmc.facebook.com/intern/diff/D78275584/) ghstack-source-id: 296122483
|
This pull request was exported from Phabricator. Differential Revision: D78275584 |
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):
run_prepack()toprepack()and replaceencode_prepack()+prepack()with justprepack()#12443Changes
With this compute kernel, transformer models' text generation can execute much faster than before.
On Samsung Galaxy S24 for Llama 3.2 1B, generating 128 tokens:
Before: ~25 tok/s
After: ~49 tok/s
Why this new shader is faster
The biggest reason is due to vectorized loading of the uint4 weight buffer. This new shader loads the weight buffer as a buffer/image of
uvec4, whereas the old shader loads the weight buffer as a buffer/image ofu8vec4. Using the Adreno Offline Compiler, I found that in the former, only one load instruction was used to load from the weight tensor, whereas in the latter 16 load instructions were used to load from the weight tensor. It appears that the data loading was not being vectorized at the assembly level. This is potentially behaviour that can be approved in the SPIR-V shader compiler.An additional factor is better weight packing layout. The new prepacking routine results in better memory coalescing between threads in a work group.
The final major factor is the use of tree based reduction to co-operatively reduce partial results into the final output. Previously, a single thread was responsible for the final reduction.
Future Work
Differential Revision: D78275584