Arm backend: Fix ceil_mode pooling when the last window is dropped by EylonKrause · Pull Request #24893 · pytorch/executorch

EylonKrause · 2026-11-11T13:28:31Z

Summary

PyTorch drops a last ceil_mode pooling window that would start in the right padding, so e.g. max_pool2d(kernel=2, stride=4, padding=0, ceil_mode=True) on a width of 7 gives 2 output columns, the same as floor mode. The input then has a remainder that no window uses.

SizeAdjustInputPass skipped every ceil_mode pool, so that remainder was never sliced off and adjust_pooling_pad_if_needed returned a negative post padding (-1 above). MAX_POOL2D and AVG_POOL2D were then emitted with a negative pad, which the TOSA reference model rejects ("At least one pad is smaller than zero"). For avg_pool2d with count_include_pad=True, DecomposeAvgPool2dPass fell back to padding with the full symmetric pad instead, which kept the dropped window and produced one output element too many.

This change:

slices the unused input for ceil_mode pools with static spatial shapes in every dimension where ceil_mode does not add an output element (when the last window is dropped, the ceil and floor output sizes are equal, so the floor-mode slice is exact);
does not add post padding in DecomposeAvgPool2dPass when the needed post padding is negative.

Dynamic (SymInt) spatial shapes keep the previous behaviour.

Release notes: arm

Test plan

New cases ceil_mode_last_window_dropped / ceil_mode_last_window_dropped_padding in test_max_pool.py and ceil_mode_last_window_dropped / ceil_mode_last_window_dropped_count_include_pad in test_avg_pool2d.py. The TOSA FP/INT/INT a16w8 variants fail on main and pass with this change.
Random sweep of ceil_mode max_pool2d/avg_pool2d configurations through TosaPipelineFP and the TOSA reference model: 146/300 fail on main; the remaining 17 failures with this change are all divisor_override + ceil_mode value mismatches, a separate issue (separate PR). Floor mode: 200/200 pass before and after.
pytest -c backends/arm/test/pytest.ini backends/arm/test/ops/test_max_pool.py backends/arm/test/ops/test_avg_pool2d.py backends/arm/test/passes (excluding u55/u85/vgf): no new failures.
Not run locally: the u55/u85/vgf variants (no Corstone FVP / Vela installed here), so those rely on CI.
ufmt check and flake8 clean on the changed files.

cc @digantdesai @freddan80 @per @zingo @oscarandersson8218 @mansnils @Sebastian-Larsson @robell @rascani @nil-is-all

PyTorch drops a last ceil_mode pooling window that would start in the right padding, so e.g. max_pool2d(kernel=2, stride=4, padding=0, ceil_mode=True) on a width of 7 gives 2 output columns, the same as floor mode. The input then has a remainder that no window uses. SizeAdjustInputPass skipped every ceil_mode pool, so that remainder was never sliced off and adjust_pooling_pad_if_needed returned a negative post padding (-1 above). MAX_POOL2D and AVG_POOL2D were emitted with a negative pad, which the TOSA reference model rejects ("At least one pad is smaller than zero"). For avg_pool2d with count_include_pad=True, DecomposeAvgPool2dPass fell back to padding with the full symmetric pad instead, which made ceil_mode keep the dropped window and produced one output element too many. Slice the unused input for ceil_mode pools with static spatial shapes in every dimension where ceil_mode does not add an output element, and do not add post padding in DecomposeAvgPool2dPass when the needed post padding is negative. Test plan: Lowered 400 ceil_mode max_pool2d/avg_pool2d configurations (kernel 1-3, stride 1-4, padding 0-1, count_include_pad, divisor_override, input height 4-7) through TosaPipelineFP: 169 failed before, the remaining 40 failures are all padding=0 with divisor_override, which is a separate issue. Added test cases to test_max_pool.py and test_avg_pool2d.py. Signed-off-by: Eylon Eliyahu Krause <64223519+EylonKrause@users.noreply.github.com>

pytorch-bot · 2026-11-11T13:28:35Z

🔗 Helpful Links

🧪 See artifacts and rendered test results at hud.pytorch.org/pr/pytorch/executorch/24893

📄 Preview Python docs built from this PR

Note: Links to docs will display an error until the docs builds have been completed.

✅ No Failures

As of commit 28af32a with merge base b3b565b ():
💚 Looks good so far! There are no failures yet. 💚

This comment was automatically generated by Dr. CI and updates every 15 minutes.

linux-foundation-easycla · 2026-11-11T13:28:38Z

The committers listed above are authorized under a signed CLA.

✅ login: EylonKrause / name: Eylon Eliyahu Krause (b9e74bd)

pytorch-bot · 2026-11-11T13:28:48Z

~~Workflows were awaiting approval.~~ CI has now been triggered for the ciflow labels on this PR.

zingo · 2026-11-11T14:48:16Z

Thanks a lot for this PRs :) I kicked of some extended testing. As this is the first contribution to the project you will need to sign a EasyCLA, I hope that is OK, instructions are in a comment in this issue see #24893 (comment) :)

Return a plain bool from _ceil_mode_adds_output (mypy: the comparison of two adjust_pooling_pad_if_needed results is typed bool | SymBool), and rewrap its docstring as docformatter expects. Signed-off-by: Eylon Eliyahu Krause <64223519+EylonKrause@users.noreply.github.com>

EylonKrause · 2026-11-13T13:18:39Z

@pytorchbot label "release notes: arm"

EylonKrause · 2026-11-13T13:18:40Z

Thanks @zingo! I pushed a follow-up commit for the lintrunner findings: _ceil_mode_adds_output now returns a plain bool (mypy saw bool | SymBool), and its docstring is rewrapped for docformatter. The WebGPU jobs failed with curl: (22) download errors, and the unittest failures are also on the merge base, so those look unrelated.

zingo

Thanks a lot for the fixes :) OK To merge! I'm not sure if you have access, if not let me know and I press the button :)

EylonKrause · 2026-11-15T12:10:40Z

Don't think I have the authority yet, if you can do it yourself it would very much help, thanks :D

zingo · 2026-11-15T13:18:26Z

MacOS fails are unrelated and probably fixed with #24903 Ill rebase and check

…dding (#24894) ### Summary `DecomposeAvgPool2dPass` lowers `avg_pool2d` with a `divisor_override` to `avg_pool2d(count_include_pad=False)` on a zero padded input, followed by a multiplication by `kernel_size / divisor_override`. That is only exact when every window is full. `_compute_post_pad` returned early for `padding=0`, so with `ceil_mode=True` the last, partial window was averaged over its valid elements only. Example: `avg_pool2d(kernel=(1, 2), stride=(1, 2), padding=0, ceil_mode=True, divisor_override=2)` on `[1, 2, 3, 4, 5]`: PyTorch gives `[1.5, 3.5, 2.5]` (5 / 2), the decomposition gave `[1.5, 3.5, 5.0]`. This is a silent numerical error in the lowered model. This change computes the post padding for `ceil_mode` with a `divisor_override` also when `padding=0`, and inserts the pad whenever any pad value is non-zero rather than only when the symmetric pre-padding is. Behaviour only changes when `padding=0`, `ceil_mode=True` and `divisor_override` are all set. Independent of #24893 (it touches a different function in the same file; the two branches merge cleanly). Release notes: arm ### Test plan - New cases `divisor_override_ceil_mode_no_padding` and `divisor_override_ceil_mode_no_padding_count_include_pad` in `test_avg_pool2d.py`; they fail on main in the TOSA FP and INT variants and pass with this change. - Sweep over kernel 2-3, stride 1-3, padding 0-1, ceil_mode, count_include_pad, divisor_override in {None, 1, 2, 4, 7}, input height 4-15: 176 configurations with padding=0 fail on main (all with ceil_mode and a divisor_override), 0 with this change. - `test_avg_pool2d.py`, `test_decompose_avg_pool2d_pass.py`, `test_rewrite_avg_pool2d_pass.py`, `test_adaptive_avg_pool2d.py`, `test_mean_dim.py` (excluding u55/u85/vgf): no new failures. - Not run locally: the u55/u85/vgf variants (no Corstone FVP / Vela installed here), so those rely on CI. - `ufmt check` and flake8 clean on the changed files. Signed-off-by: Eylon Eliyahu Krause <64223519+EylonKrause@users.noreply.github.com>

zingo · 2026-11-15T15:23:12Z

Both merge! Super thanks and congratulation for you first contribution to the project!

EylonKrause requested review from digantdesai and rascani as code owners November 11, 2026 13:28

meta-cla Bot added the CLA Signed This label is managed by the Facebook bot. Authors need to sign the CLA before a PR can be reviewed. label Nov 11, 2026

EylonKrause mentioned this pull request Nov 11, 2026

Arm backend: Fix avg_pool2d divisor_override with ceil_mode and no padding #24894

Merged

github-actions Bot added ciflow/trunk module: arm Issues related to arm backend labels Nov 11, 2026

executorch-triage Bot added the community: contribution PRs coming from community (excluding hardware partners) label Nov 11, 2026

zingo added the partner: arm For backend delegation, kernels, demo, etc. from the 3rd-party partner, Arm label Nov 11, 2026

EylonKrause force-pushed the fix/pool-ceil-mode-dropped-window branch from 91dc013 to 4f22175 Compare November 13, 2026 13:18

pytorch-bot Bot added the release notes: arm Changes to the ARM backend delegate label Nov 13, 2026

zingo approved these changes Nov 15, 2026

View reviewed changes

Merge branch 'main' into fix/pool-ceil-mode-dropped-window

28af32a

zingo merged commit 42f42f3 into pytorch:main Nov 15, 2026
434 checks passed

Navigation Menu

Repository navigation

Arm backend: Fix ceil_mode pooling when the last window is dropped - #24893

Arm backend: Fix ceil_mode pooling when the last window is dropped#24893
zingo merged 3 commits into
pytorch:mainfrom
EylonKrause:fix/pool-ceil-mode-dropped-window

EylonKrause commented Nov 11, 2026 •

edited by pytorch-bot Bot

Loading

Uh oh!

pytorch-bot Bot commented Nov 11, 2026 •

edited

Loading

Uh oh!

linux-foundation-easycla Bot commented Nov 11, 2026 •

edited

Loading

Uh oh!

pytorch-bot Bot commented Nov 11, 2026 •

edited

Loading

Uh oh!

zingo commented Nov 11, 2026

Uh oh!

EylonKrause commented Nov 13, 2026

Uh oh!

EylonKrause commented Nov 13, 2026

Uh oh!

zingo left a comment

Uh oh!

EylonKrause commented Nov 15, 2026 •

edited

Loading

Uh oh!

zingo commented Nov 15, 2026 •

edited

Loading

Uh oh!

Uh oh!

zingo commented Nov 15, 2026

Uh oh!

Reviewers

Uh oh!

Assignees

Labels

Projects

Milestone

Development

Uh oh!

2 participants

Conversation

EylonKrause commented Nov 11, 2026 • edited by pytorch-bot Bot Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Summary

Test plan

Uh oh!

pytorch-bot Bot commented Nov 11, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

🔗 Helpful Links

🧪 See artifacts and rendered test results at hud.pytorch.org/pr/pytorch/executorch/24893

✅ No Failures

Uh oh!

linux-foundation-easycla Bot commented Nov 11, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

pytorch-bot Bot commented Nov 11, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

zingo commented Nov 11, 2026

Uh oh!

EylonKrause commented Nov 13, 2026

Uh oh!

EylonKrause commented Nov 13, 2026

Uh oh!

zingo left a comment

Choose a reason for hiding this comment

Uh oh!

EylonKrause commented Nov 15, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

zingo commented Nov 15, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

Uh oh!

zingo commented Nov 15, 2026

Uh oh!

Reviewers

Uh oh!

Assignees

Labels

Projects

Milestone

Development

Uh oh!

2 participants

EylonKrause commented Nov 11, 2026 •

edited by pytorch-bot Bot

Loading

pytorch-bot Bot commented Nov 11, 2026 •

edited

Loading

linux-foundation-easycla Bot commented Nov 11, 2026 •

edited

Loading

pytorch-bot Bot commented Nov 11, 2026 •

edited

Loading

EylonKrause commented Nov 15, 2026 •

edited

Loading

zingo commented Nov 15, 2026 •

edited

Loading