[Strategy][ARM CPU] Remove contrib spatial pack schedule of depthwise convolution

[FrozenGene]

The spatial pack schedule of depthwise convolution is under the tag of contrib before introducing op strategy. However, as this discuss: https://discuss.tvm.ai/t/autotuner-incorrect-result-after-tuning-mobilenetv2-on-arm-cpu/6088, we maybe will meet incorrect result use this schedule. I ever met this problem on some hardware platform too. However, it is not 100% could produce and happened rarely. Correct result should be firstly, so let us make our previous depthwise schedule be the default, not the contrib spatial pack schedule.

@icemelon9 Could you help to review the code?

[FrozenGene]

After thinking it a bit more, lower the plevel could only change the default schedule. However, when the users tune it by themselves, they will still introduce this contrib schedule into their tuning task. Before we could solve this problem, I would like to comment it out (but keep the code there). What about your advices? @comaniac @icemelon9

[comaniac]

After thinking it a bit more, lower the plevel could only change the default schedule. However, when the users tune it by themselves, they will still introduce this contrib schedule into their tuning task. Before we could solve this problem, I would like to comment it out (but keep the code there). What about your advices? @comaniac @icemelon9

You are right. That's the intention of the plevel in strategy. If this schedule has a potential correctness issue, we could have the following 3 solutions:

1. Suggest users enabling correctness checking when tuning (i.e., let AutoTVM check correctness for each schedule config), but this will also check the other task that does not have this issue, making the tuning time even longer.

2. Add an optional when registering an AutoTVM task to enforce correctness checking, but this requires a change to not only the registration but AutoTVM mechanism.

3. Comment out the schedule and let it back after the issue has been resolved. The problem is that we may never let it back if no one is dedicated to fix it...

@icemelon9 do you have any suggestions?

[icemelon]

Yes, let's comment out this implementation temporarily. Could you add a Todo with pointer to the issue in the discussion to both arm cpu strategy and the topi file?

[FrozenGene]

Yes, let's comment out this implementation temporarily. Could you add a Todo with pointer to the issue in the discussion to both arm cpu strategy and the topi file?

Sure. I have added comments as your advice. @comaniac @icemelon9 please have a look again

[FrozenGene]

Thanks @icemelon9 @comaniac for reviewing and giving advices.