Zephyr-embARC_MLI internal review

[SiyuanCheng-CN]

No description provided.

[ruuddw]

Thanks, I'll have a closer look later. Quick feedback: for the earlier PR, Zephyr TSC feedback was that they did not like pulling in another MPI copy/module as there already is one. Could you do a comparison between these? Would it e.g. be feasible to merge the ARC MLI changes into that one, or are they too different?

[SiyuanCheng-CN]

In my understanding, what the TSC said for another ML module is Tensorflow. In Tensorflow, other than generic ML code, it also has some platform-specific library, including embARC_MLI to accelerate some part of kernel functions.
From what I can see now:

1. The version of embARC_MLI in Tensorflow is v1.1, now the latest version is v2.0. Difference between them is big.
2. embARC_MLI now support plenty of networks, example in Tensorflow only accelerate one network.
3. Example in Tensorflow only targets emsdp.
4. Seems like example can only run in Tensorflow, not in Zephyr.

For context switch, it only support non-preemptive switching, without adding DSP registers, it also work. If you want me to add DSP registers in callee-saved register, I can help.