[Relay][AutoTVM] Bug Fix for ARM CPUs. Lower strict assumption.

[bernhardklein]

Hello together,
since last changes I had a issue compiling stuff for my ARM CPUs( Rasberry Pi 3b+). I got the error message that stride_h is not known. Seems understandable in the code, so I guess that the required assumption here is to strict. It seems to work well with removed stride assumption. However, the original authors of this code can you take a look and check if I am right.
Thanks, Bernhard

[bernhardklein]

cc @icemelon9 @kevinthesun

[icemelon]

Thanks @bernhardklein. This is now merged.

[FrozenGene]

I think the fix is not correct. Winograd doesn't support stride > 1. The issue here is we don't have code stride_h, stride_w = strides and just have strides = attrs.get_int_tuple("strides"). Please fix it.

[bernhardklein]

@FrozenGene Thank you for your response. I think the assumption a view lines obove assert strides == (1, 1), "Do not support strides now" already ensures that stride is 1. A second check should be unnecessary. Correct me if I'm wrong.

[FrozenGene]

@FrozenGene Thank you for your response. I think the assumption a view lines obove assert strides == (1, 1), "Do not support strides now" already ensures that stride is 1. A second check should be unnecessary. Correct me if I'm wrong.

Yeah, you are correct! I didn't realize we assert here. Thanks for pointing it out.