Fix quantized_conv1d_ncl padding bug in HiFi kernel (#18783)#18783
Merged
meta-codesync[bot] merged 1 commit intopytorch:mainfrom Apr 9, 2026
Merged
Fix quantized_conv1d_ncl padding bug in HiFi kernel (#18783)#18783meta-codesync[bot] merged 1 commit intopytorch:mainfrom
meta-codesync[bot] merged 1 commit intopytorch:mainfrom
Conversation
🔗 Helpful Links🧪 See artifacts and rendered test results at hud.pytorch.org/pr/pytorch/executorch/18783
Note: Links to docs will display an error until the docs builds have been completed. ⏳ 1 Pending, 2 Unrelated FailuresAs of commit 2c626b7 with merge base 21d9c64 ( BROKEN TRUNK - The following jobs failed but were present on the merge base:👉 Rebase onto the `viable/strict` branch to avoid these failures
This comment was automatically generated by Dr. CI and updates every 15 minutes. |
meta-cla
Bot
added
the
CLA Signed
Contributor
|
@ethansfng has exported this pull request. If you are a Meta employee, you can view the originating Diff in D100069524. |
This PR needs a
|
meta-codesync
Bot
changed the title
ethansfng
added a commit
to ethansfng/executorch
that referenced
this pull request
Apr 8, 2026
Summary: The HiFi optimized int8 path in `op_quantized_conv1d_ncl.cpp` simulates 1D convolution as 2D (height=1) using `xa_nn_conv2d_per_chan_sym8sxasym8s`. The NNLib convention is x=width, y=height, but the code had them swapped: x_stride/x_padding were set to 1/0 (height values) while y_stride/y_padding held the actual 1D stride/padding (width values). This caused any conv1d with padding > 0, dilation == 1, stride == 1 to compute without padding, producing incorrect results. This was introduced by D95279330 which added the quantized_conv1d_ncl kernel. Models with conv1d at dilation=1 and padding>0 (e.g., first TemporalBlock in TCN-based models like microgestures) hit this bug. Fix: swap the axes so x (width) gets the actual 1D stride/padding and y (height) gets the trivial 1/0 values. Differential Revision: D100069524
ethansfng
force-pushed
the
export-D100069524
branch
from
fd409fd to
74d3d64
Compare
Summary: Pull Request resolved: pytorch#18783 The HiFi optimized int8 path in `op_quantized_conv1d_ncl.cpp` simulates 1D convolution as 2D (height=1) using `xa_nn_conv2d_per_chan_sym8sxasym8s`. The NNLib convention is x=width, y=height, but the code had them swapped: x_stride/x_padding were set to 1/0 (height values) while y_stride/y_padding held the actual 1D stride/padding (width values). This caused any conv1d with padding > 0, dilation == 1, stride == 1 to compute without padding, producing incorrect results. This was introduced by D95279330 which added the quantized_conv1d_ncl kernel. Models with conv1d at dilation=1 and padding>0 (e.g., first TemporalBlock in TCN-based models like microgestures) hit this bug. Fix: swap the axes so x (width) gets the actual 1D stride/padding and y (height) gets the trivial 1/0 values. Differential Revision: D100069524
ethansfng
force-pushed
the
export-D100069524
branch
from
74d3d64 to
2c626b7
Compare
aliafzal
approved these changes
Apr 8, 2026
jpiat
pushed a commit
to jpiat/executorch
that referenced
this pull request
Apr 14, 2026
Differential Revision: D100069524 Pull Request resolved: pytorch#18783
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
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
2 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.
Summary:
The HiFi optimized int8 path in
op_quantized_conv1d_ncl.cppsimulates 1Dconvolution as 2D (height=1) using
xa_nn_conv2d_per_chan_sym8sxasym8s.The NNLib convention is x=width, y=height, but the code had them swapped:
x_stride/x_padding were set to 1/0 (height values) while y_stride/y_padding
held the actual 1D stride/padding (width values). This caused any conv1d with
padding > 0, dilation == 1, stride == 1 to compute without padding, producing
incorrect results.
This was introduced by D95279330 which added the quantized_conv1d_ncl kernel.
Models with conv1d at dilation=1 and padding>0 (e.g., first TemporalBlock in
TCN-based models like microgestures) hit this bug.
Fix: swap the axes so x (width) gets the actual 1D stride/padding and y
(height) gets the trivial 1/0 values.
Differential Revision: D100069524