Hi,
We implemented an unsupported operation using NNAPI extensions to run on ARM Neon. since the rest of the graph is executed on the DSP, there are DSP->CPU->DSP copies.
We are considering implementing this using SNPE UDO on DSP if it avoid the DSP->CPU->DSP copies.
Is our assumption correct? if the previous operation executed on the DSP, the UDO will be executed on the DSP as well without any copies?
Regards