Support for older ONNX operators (opset11)

I've reviewed the onnx-opset14/15/16/17-support.md documentation, but I couldn't find clear information regarding the support for mathematical functions (specifically sin, cos, pow, div, and mod).

Hi ​@Hiroyuki Kojima! I’ve been asking the guys internally, and Opset 11 is supported, which is good news! The recommendation is to use Opset 17 as much as possible, but previous Opsets are supported 😃

However, I can’t see any mention of support for the mathematical functions, unfortunately. Let me know if this helps, or I can help to dig deeper if needed!

HI, Spanner-san

I'm working on deploying RTMO-Pose’s ONNX on Metis-AIPU.

I'd like to use the NPU to perform as much processing as possible, but I'm struggling because the model head uses mathematical instructions to calculate the final keypoint and box coordinates.

Attached is a list of layers and ONNX instructions. I'd appreciate any advice or suggestions on the best way to go about it.

Best, regards
Hiroyuki Kojima

Hi ​@Hiroyuki Kojima ! I found some more details on Opset 11, and it doesn’t look as though those instructions are supported on the AIPU. I’ve attacted it so you can take a look, and see if there’s anything else that might help you out with this.

Possibly the model could be split, so the majority of the compute is on the NPU but the mathematical post-processing is moved to the CPU?

HI, Spanner-san

Thank you for the reply.

The supported operators list has been reviewed.

In our model, LayerNorm is applied in the neck (encoder); however, as it is not supported, this imposes constraints.

Similar requests have been raised in other threads. Kindly treat this message as an additional request for LayerNorm support.

Best, regards
Hiroyuki Kojima