You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
The fusion segmenter only considers fusion of a producer expr group with a consumer expr group. Here, there's no such relationship between the sin and cos exprs. They are both consumers of tv0, and that relationship is not currently considered for fusion.
This isn't an issue if the whole fusion can be scheduled without segmentation. The repro above has segment_set, which is there just to enforce the segmentation step.
I encountered this case when testing #3659, which further promotes the input forwarding. That helps generate better segmentation results in some cases when segmentation is unavoidable, but it also exposed the above issue, resulting in segmenting previously fused expr groups.
The text was updated successfully, but these errors were encountered:
This fusion is segmented to 3 segments with each expr as a distinctive segment. We should fuse
tv1
andtv2
, but that's not what happens.The fusion segmenter only considers fusion of a producer expr group with a consumer expr group. Here, there's no such relationship between the sin and cos exprs. They are both consumers of
tv0
, and that relationship is not currently considered for fusion.This isn't an issue if the whole fusion can be scheduled without segmentation. The repro above has
segment_set
, which is there just to enforce the segmentation step.I encountered this case when testing #3659, which further promotes the input forwarding. That helps generate better segmentation results in some cases when segmentation is unavoidable, but it also exposed the above issue, resulting in segmenting previously fused expr groups.
The text was updated successfully, but these errors were encountered: