multi-thread problem. about qnnpack HOT 7 CLOSED

i found qnnp_run_operator run faster when set the threadpool var to nullptr than set the threadpool to 1~4 thread num.

from qnnpack.

so how many thread is opened by default within the qnnpack_run_operator func.

from qnnpack.

in my experiments, the default performance（set the threadpool to nullptr） is similar to set the threadpool to 4 threads.

from qnnpack.

which thread pool implementation do you use? pthreadpool library is not very efficient, but Caffe2 reimplements pthreadpool interface on top of its own (a lot more efficient) thread pool.

from qnnpack.

yes, i used pthreadpool. i‘ll look into caffe2.

from qnnpack.

In the past few days, I have rewrite the threadpool like caffe2, and the computation speed has been greatly improved, so I would like to know if you have any plans to add a custom threadpool library.

from qnnpack.

I have plans to upstream Caffe2 threadpool modifications to pthreadpool, but this is not the first priority right now.

from qnnpack.