Add autoscheduler support to tvmc

[giuseros]

• Add an autoschedule module to tvmc
• Extract common tuning option between autotuner and autoscheduler
• Add testing

[giuseros]

cc @leandron @merrymercy

[giuseros]

Hi @comaniac , @leandron ,
Thanks for your comments, now the code is way more sound. I thought of summing up here some issues left:

• It's now hard to make autotuner vs autoscheduler options mutually exclusive. So if I specify --tuner with --enable-autoscheduling, the tuner I selected won't be considered. This is because we are using default values, so a tuner will be always defined.
• Extracting tasks. @comaniac , you flagged the function, but this is actually how we do for the autotuner. I guess this is because we want to unit-test the function
• Testing: I removed the test_autoscheduler.py test. When I wrote it, I was inspired by the test_autotuner.py. Can we agree on a set of tests we want to execute? @comaniac , I agree we should not test the autotuner, but only the user interface. But I fear to test the user interface, we will need to test a bit the functionality as well. What do you guys think?

[comaniac]

It's much clean now and overall LGTM. Thanks!
The last batch of comments are mostly about refactoring.

[comaniac]

Hi @comaniac , @leandron ,
Thanks for your comments, now the code is way more sound. I thought of summing up here some issues left:

• It's now hard to make autotuner vs autoscheduler options mutually exclusive. So if I specify --tuner with --enable-autoscheduling, the tuner I selected won't be considered. This is because we are using default values, so a tuner will be always defined.

I'm ok with this for now.

• Extracting tasks. @comaniac , you flagged the function, but this is actually how we do for the autotuner. I guess this is because we want to unit-test the function

No worries. It is no longer an issue based on the latest implementation. It is reasonable to have separate functions for two frameworks launched by the same CLI command.

• Testing: I removed the test_autoscheduler.py test. When I wrote it, I was inspired by the test_autotuner.py. Can we agree on a set of tests we want to execute? @comaniac , I agree we should not test the autotuner, but only the user interface. But I fear to test the user interface, we will need to test a bit the functionality as well. What do you guys think?

You're right. It's always hard to keep the proper unit test scope of TVMC. We should have a thread to discuss all of them and make a agreement. cc @leandron

[giuseros]

Hi @comaniac , @leandron ,
I updated the PR as follows:

• I added two more options (--log-estimated-latency and --include-simple-tasks)
• I added the tests back (as they were before).
  Please, let me know if those tests are ok (or if I need to add more or less), and if there is any other autoscheduler relevant option I need to consider.

Thanks,

[comaniac]

Overall LGTM. Thanks. Just a naming nit.

[comaniac]

Thanks @giuseros @leandron @insop.

[merrymercy]

For x86 cpu, it is recommended to use enable_cpu_cache_flush=True in LocalRunner, as shown in the tutorial.

tvm/tutorials/auto_scheduler/tune_network_x86.py

Line 187 in 54c995d

runner=auto_scheduler.LocalRunner(repeat=10, enable_cpu_cache_flush=True),

This can make the measurement faster and more accurate. We should add this to tvmc.

[giuseros]

Hi @merrymercy ,
Thanks for the comment! Sure, I will open a new PR to add this in