ARM Ethos-U55: Enable support for `aten.upsample_nearest2d.vec`

[etrommer]

🚀 The feature, motivation and pitch

Several Ultralytics Yolo variants use 2x2 nearest neighbor upsampling operators, which are explicitly blacklisted for lowering to U55 (for U85, they can be lowered without issues):

executorch/backends/arm/operator_support/ethos_u55_support.py

Line 212 in 73b46d3

exir_ops.edge.aten.upsample_nearest2d.vec, # RESIZE

For TFlite exports, this operator is correctly ran on the NPU, while for ExecuTorch, this is currently the only remaining operator in the model running on CPU, leading to a much higher runtime and insertion of Q/DQ nodes in the graph.

According to the documentation, it should be possible to run this operator on U55: https://gitlab.arm.com/artificial-intelligence/ethos-u/ethos-u-vela/-/blob/main/SUPPORTED_OPS.md?ref_type=heads#ethos-u55-and-ethos-u65-tflite-resize_nearest_neighbor-constraints

Since Yolo models are quite a common use case for platforms with U55, I think it would make sense to add the necessary plumbing for this operator in the ExecuTorch -> U55 path.

Alternatives

No response

Additional context

Comparison of ExecuTorch graph of Yolo v26 on Ethos-U55 vs. TFlite graph:

RFC (Optional)

No response

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

[etrommer]

I simply commented out the line with the operator in the unsupported_ops list to see what would break and it seems that this operator is lowered just fine with the existing flow. Model runs on U55 and the accuracy of the output model also seems unaffected. At least for Yolo v26, it appears as if this is already supported. (Disclaimer: Haven't tested if there any adverse side effects for other models)

[perheld]

I simply commented out the line with the operator in the unsupported_ops list to see what would break and it seems that this operator is lowered just fine with the existing flow. Model runs on U55 and the accuracy of the output model also seems unaffected. At least for Yolo v26, it appears as if this is already supported. (Disclaimer: Haven't tested if there any adverse side effects for other models)

Nice that it wasnt a blocker! I still I think we should support it properly with checking of batchsizes and scaling so there is some code and tests on the way specific for U55.

[etrommer]

Yes, wasn't intending for this to be a productive fix, just a useful observation 😄