[AutoTVM] Use popenpool in local_executor

[shingjan]

This PR intends to replace multiprocessing.pool with popenpool in autotvm/local_executor.
@junrushao1994 @vinx13 @tqchen

[shingjan]

Author's note:

1. There is a weird ModuleNotFoundError: No module named 'test_autotvm_executor' during the execution of test_autotvm_executor.py. Still working on a fix.

2. The reason why PopenPoolExecutor is used here instead of PopenWorker is bcoz using PopenWorker requires the re-implementaion of LocalFuture as well as a usable replacement of multiprocessing.Queue, which are all taken care of in PopenPoolExecutor.

[vinx13]

Similar to the fix you did for auto scheduler, test_autotvm_executor is not really a module and it shouldn't be imported. Anything common for tests should probably be moved to tvm.testing.autotvm
Also, we may think again whether LocalExecutor and LocalFuture is really needed, if possible we prefer using PopenWorker or PopenPoolExecutor directly

[shingjan]

@vinx13 just moved all testing utility functions of autotvm to tvm.testing. About LocalFuture and LocalExecutor, they do seem very much like PopenPoolExecutor now. Maybe we can sync offline about this.

[shingjan]

@vinx13 The refactor is finished. I moved the two test cases for LocalExecutor to test_popen_pool.py as I do believe they can be of use, e.g. make sure the PopenPoolExecutor is functioning correctly. I also move the LocalFutureNoFork to autotvm.executor as one test case depends on it and it is for convenience's sake that we keep it.

[vinx13]

Can you confirm autotvm work both in the terminal and jupyter? You can run this tutorial

[vinx13]

The test case can be updated if the only usage of LocalFutureNoFork is the test case

[shingjan]

On 64-core linux machine, each did 5 runs:

tune_conv2d_cuda.py, n_trial = 192	This PR		Main
Local builder + Local runner	mean	std	mean	std
Build time	10.438	1.935	8.21	3.165
Run time	27.452	3.349	26.248	6.114
Local builder + RPC runner
Build time	10.304	0.925	7.844	1.838
Run time	25.098	3.229	27.644	3.836

[vinx13]

@shingjan can you also check if custom build func still works for autotvm?

[vinx13]

found some changes needed

[shingjan]

With a custom build function like the one below, this PR runs fine for tune_conv2d_cuda:

from tvm.contrib import tar


def custom_build(*args):
    return tar.tar(*args)


custom_build.output_format = "tar"

measure_option = autotvm.measure_option(
    builder=autotvm.LocalBuilder(build_func=custom_build),
    runner=autotvm.LocalRunner(repeat=3, min_repeat_ms=100, timeout=4),
)