Arm backend: Move rescale ops out of node visitors

[martinlsm]

Some TOSA ops do not support INT8 as inputs and outputs. Instead, only INT32 is supported as a whole number type. Prior to this patch, affected node visitors inserted rescale ops between the data types INT8 and INT32 before and after the operator such that it will accept its input and output.

Change this by moving the insertion of the rescale ops to a new pass called InsertRescaleInt32Pass. This will further enable optimizations to the graph by fusing the rescale nodes.

Only comparison, ABS, MAXIMUM and MINIMUM operators are handled in this patch; the remaining ones are left out to be done in another patch.

Test plan

This is refactoring which means that external behavior is not altered. A new pass InsertRescaleInt32Pass has been added and it comes with a new unit test in backends/arm/test/passes/test_insert_rescale_i32_pass.py.

cc @digantdesai @freddan80 @per @zingo @oscarandersson8218

[pytorch-bot]

🔗 Helpful Links

🧪 See artifacts and rendered test results at hud.pytorch.org/pr/pytorch/executorch/14584

• 📄 Preview Python docs built from this PR

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

❌ 2 New Failures, 1 Unrelated Failure

As of commit 0b65f38 with merge base f81e834 ():

NEW FAILURES - The following jobs have failed:

• pull / test-samsung-models-linux / linux-job (gh)
  RuntimeError: Command docker exec -t 0ade18a3a200c52c49c6a37691f483bc54d919ffcea530744094cfe51e9f1d9a /exec failed with exit code 1
• trunk / test-llama-runner-mac (fp32, coreml) / macos-job (gh)
  RuntimeError: Command bash /Users/ec2-user/runner/_work/_temp/exec_script failed with exit code 1

BROKEN TRUNK - The following job failed but were present on the merge base:

👉 Rebase onto the `viable/strict` branch to avoid these failures

• Apple / build-demo-ios / macos-job (gh) (trunk failure)
  RuntimeError: Command bash /Users/runner/work/_temp/exec_script failed with exit code 74

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

[martinlsm]

@pytorchbot label ciflow/trunk

[martinlsm]

@pytorchbot label "partner: arm"

[martinlsm]

failures look unrelated

[digantdesai]

may be this boat has sailed already but for logical grouping perspective, I would rather keep the util call in respective ops instead of a pass.. but that's just me. Not a strong objection.

[martinlsm]

@digantdesai Not sure what you mean fully. Can you elaborate a bit?

[digantdesai]

Instead of inserting rescale ops in the tosa graph through this pass, I would prefer (mild preference) of keeping the way things were, i.e. an op lowering (op_.py) would inject an op when needed.

This is mainly to keep things together logically and keep the tosa graph 'sound' after each op lowering.

That said I understand the code duplication aspect, but that could be tackled with keeping these insert rescale as util functions.

Again, not a strong objection. Stamping to unblock you, and leaving it up to you to decide.

Thanks!