Hi guys,
I’ve deployed and run the model successfully in both X86 and ARM target machine. When I checked the CPU utilization, they were both higher than 100% (approx 200% in X86 and 780% in ARM, depending on the performance of CPU, not the key point). I’d like to try run it with only one core for comparing time cost. Except the cgroup and the function for setting CPU affinity in Linux, I wonder know if there is any switch to control it in tvm code, or the only way to achieve my goal is the method I mentioned above?
I checked the code and thought the switch should be in threading_backend.cc. I set both of the num_workers_ and exclude_worker0 to 1. The CPU utilization was 100% but the model didn’t work, no inference result displayed.
Thanks.