QNN Error with GroupNorm

[escorciav]

Hi,
I was wondering how are you guys dealing with GroupNorm. I discovered the following weird situation in S8G2 DSP (HTP, chipset: SM8550).

GroupNorm(groups=32)	1, 64, 512, 512	1, 64, 256, 256	1, 64, 128, 128	1, 64, 64, 64
msec	1528.3	371.1	error	0.794

The test used QNN 2.16.

Do you know how can I report that bug to the QNN team?

Thanks!

[quic-hitameht]

Tagging @quic-akinlawo @quic-mangal here.

[escorciav]

It may sound odd, but what can I do it's what it's, i.e., hardware & on-device stuff aka too many moving wheels

I can't replicate the error for input 1,64,128,128. The model using such op compile & latency performance is shown in the screenshot below. First column is msec & second the version of QNN

Details (for record)

• I cannot rplicate the QNN compilation error anymore
• I tried different versions of QNN & onnx opset version.
• It's possible that the version of onnx & PyTorch itself plays a role. I wasted some time testing things out. But, decided to move on & enjoy the good news.