According to the documentation, the mobile benchmarking binary only runs on the mobile cpu and doesn’t using hardware accelerators like DSPs and GPUs. Is there a way of benchmarking pytorch models on a mobile phone which is able to leverage these hardware accelerators?
Also, another question, I’ve been successfully able to run the RepNet model on my phone using this binary but as I understand qnnpack and xnnpack do not support Conv3D or attention layers, is there a fallback in the pytorch mobile runtime that is used when the model contains unsupported layers because the model runs without any issues?