[rascani]

Channel-broadcast add, mul, and sub on contiguous NCHW inputs currently
fall back to float in explicit layout. Recognize these patterns and use
the existing layout pass to place their channels in the innermost
dimension. Preserve that layout during copy removal and let the existing
add and mul kernels consume channel vectors in either supported layout.

Cover both operand orders, singleton dimensions, chained broadcasts,
convolution producers, and unsupported spatial broadcasts. MobileNetV3
now retains all 28 quantized multiplies with one input/output
quantization boundary. All 165 host tests and 25 M55 simulator cases
passed; eight existing cases remain expected failures. Targeted lint
also passed.

Authored with AI using Codex.

[rascani]

Stack from ghstack (oldest at bottom):

• -> Support channel broadcasting in Cortex-M explicit layout #24761
• Preserve portable fallback in Cortex-M explicit layout #24760

[pytorch-bot]

🔗 Helpful Links

🧪 See artifacts and rendered test results at hud.pytorch.org/pr/pytorch/executorch/24761

• 📄 Preview Python docs built from this PR

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

❌ 1 New Failure

As of commit 7977a82 with merge base 472953c ():

NEW FAILURE - The following job has failed:

• Lint / lintrunner-mypy (gh)
  >>> Lint for backends/cortex_m/passes/explicit_layout_pass.py:

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