If I’m not mistaken under relay, we can have multiple strategies for each operator. This allows usage of multiple different implementations of the same operator where each can be specialized to handle certain cases of input with overlaps.
In addition to that, autotvm under relay picks up the best strategy after tuning.
Which raises my question about Relax, does Relax support some version of strategy?
My impression is that metaschedule works only with legalized ops and can provide multiple schedule templates using schedule rules. This doesn’t account for different implementations say specific computes made for convolution such as winograd or one that’s optimized for depthwise and such thereby tuning multiple strategies for the same operation and selecting the best out of them.