[DO NOT MERGE] Revert "[AUTOTVM] Use opt level 3 when extracting tasks"

[anwang2009]

Reverts #10065.

After doing a sweep on nvidia t4 with 2000 autotvm trials, I'm seeing regressions across the board as compared to before this commit. We should revert and investigate why this is happening before we enable this change broadly

[anwang2009]

cc @tkonolige @masahi

[masahi]

I dont like reverting this. This will put us back to the wrong state and there is no guarentee that it will be restored later unless someone takes an action.

So instead of reverting this to workaround the immediate perf problem, can we pursue a more long term solution. Have you tried enabling alter layout? If this is cuda, the issue is most likely related to winograd.

[anwang2009]

Yeah, we've tried enabling alter layout. It has similar results as shown. I'm seeing regressions on broadwell as well as nvidia tesla t4, but the regressions are more pronounced on the t4. Currently running a sweep to confirm if the problem was actually this PR or another one on the stack.

[AndrewZhaoLuo]

@anwang2009 can you open an issue for tracking and then maybe we will be amenable to merging?

[anwang2009]

Awaiting results of this sweep before further action

[anwang2009]

Closing this. Sweeps show that the relay opt level 3 is the right approach