Vectorization cleanup

[zasdfgbnm]

Currently, vectorization analysis is done by first calling getInnerDimVectorizableWidth to compute teh vectorization factor of the innermost dimension, then expand that to contiguous merge of getMaxVectorizableWidth. However, these two functions are very similar, almost a copy-paste of each other. Therefore, it makes no sense to use getInnerDimVectorizableWidth then getMaxVectorizableWidth, we should remove getInnerDimVectorizableWidth and go directly to getMaxVectorizableWidth.

This PR removes getInnerDimVectorizableWidth, anded a boolean option contig_merge to getMaxVectorizableWidth to reproduce the behavior of getInnerDimVectorizableWidth. The contig_merge=false is only used in the transpose scheduler, where the which dimensions are contiguously merge to which ID is more complicated and is not considered in heuristics.

Unfortunately, because the variable name change, I can not diff the kernel to check if they match exactly or not, but I do briefly skimmed through the diff by eye, and I don't find any change in the vectorization factor.

The V100 failure is OOM.

[zasdfgbnm]

A big portion of getVectorizationFactor has been removed, so I just cut-paste the function body of this function into getVectorizationFactor and remove this function.

[zasdfgbnm]

!build

[zasdfgbnm]

!build

[zasdfgbnm]

!build

[naoyam]

The changes make sense to me. Have you found any change in actual vectorization?

[zasdfgbnm]

The changes make sense to me. Have you found any change in actual vectorization?

No, I didn't find any. But I didn't check programmatically. I just used my eye to inspect the output of

difft --language cpp ../Fuser3/old-kernels/ new-kernels/ | grep load

and didn't find any.