Vulkan: hoist the pointwise conv spatial divides out of the K loop

[msluszniak]

Summary

conv2d_pw_tiled recovers the (x, y) texel coordinate from the flattened spatial index with a divide and a modulo by W_out. W_out is a runtime uniform, so neither strength reduces to a shift, and both sat inside the loop over K: once per (m, k4) on the input loads, and again per output texel on the store. The coordinates depend only on m, so a conv with K4 = 4 issued four times as many integer divides as the arithmetic they fed.

This computes them once per invocation and passes them down. No change to the tiling, the dispatch, or the math.

Measurements

Three models, both GPU vendors. Arms interleaved within one session, per-dispatch GPU timestamps from the query pool, median of 3 rounds. Mali-G76 is a Galaxy S10+, Adreno 840 a Galaxy S26 Ultra.

model	device	conv2d_pw share	conv2d_pw	whole graph GPU
lraspp-mobilenet-v3-large fp16 @520	Mali-G76	71%	256.3 -> 72.0 ms (3.56x)	284.1 -> 100.2 ms (2.83x)
selfie-segmentation fp16 @256	Mali-G76	66%	8.02 -> 4.18 ms (1.92x)	12.2 -> 8.3 ms (1.47x)
selfie-segmentation fp16 @256	Adreno 840	66%	0.657 -> 0.519 ms (1.27x)	1.179 -> 1.041 ms (1.13x)
PP-OCRv6 fp16	Adreno 840	41%	32.08 -> 23.64 ms (1.36x)	65.76 -> 57.34 ms (1.15x)

The gain tracks how much of the graph is pointwise convolution and how deep the K loop is, which is why lraspp gains most: more channels means more K iterations, so more redundant divides per invocation. A transformer with no pointwise convs (CLIP ViT-B/32 vision, 0 such dispatches) is unaffected at 1.000x, as expected.

End to end on lraspp this is 310.3 -> 114.1 ms on Mali. For reference the XNNPACK int8 build of the same model, which is what that backend ships as its Android default, measures 93.0 ms there.

Interleaving matters on the Mali part: the same binary reads 25.2 ms hot and 8.0 ms settled, so arms measured in separate sessions are not comparable. Within a session, round to round spread is under 1%.

Correctness

Output is bit identical before and after on every model measured, max |diff| = 0.0, including on inputs chosen to produce non-degenerate outputs.

[pytorch-bot]

🔗 Helpful Links

🧪 See artifacts and rendered test results at hud.pytorch.org/pr/pytorch/executorch/22780

• 📄 Preview Python docs built from this PR

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

⚠️ 15 Awaiting Approval

As of commit 2e6e14e with merge base 14dab10 ():

AWAITING APPROVAL - The following workflows need approval before CI can run:

• Build documentation (gh)
• Build Presets (gh)
• Lint (gh)
• periodic (gh)
• pull (gh)
• Test ARM Backend (gh)
• Test CoreML Backend (gh)
• Test Cortex-M Backend (gh)
• Test NXP Backend (gh)
• Test QNN Backend (gh)
• Test QNN Windows MSVC build (gh)
• Test Vulkan Backend (gh)
• Test Vulkan Backend (specialized runners) (gh)
• Test WebGPU Backend (gh)
• Test XNNPACK Backend (gh)

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

[github-actions]

This PR needs a release notes: label

If your change should be included in the release notes (i.e. would users of this library care about this change?), please use a label starting with release notes:. This helps us keep track and include your important work in the next release notes.

To add a label, you can comment to pytorchbot, for example
@pytorchbot label "release notes: none"

For more information, see
https://github.com/pytorch/pytorch/wiki/PyTorch-AutoLabel-Bot#why-categorize-for-release-notes-and-how-does-it-work.