Grainsize, reduce_sum, reduce_sum_static

[sonicking]

Hi, I understand that when using reduce_sum, setting grainsize = 1 uses an internal scheduler to decide on the slice size.

But what exactly happens when you use a different grainsize with reduce_sum?

Also, what exactly happens when you use reduce_sum_static?

I never find using grainsize = 1 gives the quickest run. I also fail to get a good pattern as I follow the suggestion in the manual. Could someone give a better explanation? Much appreciated! Thanks.

[wds15]

Understanding reduce_sum efficiency

Thanks, this is helpful. So the block sorting is “random” in the sense that there is no particular reason for the way the blocks end up being sequenced, but they nonetheless end up in some overall sequence. For dataset construction purposes then, you can just pick an order that is a mixture of group sizes and probably end up with something reasonably optimal, if I am reading you correctly. In terms of exactly which group order ends up being truly optimal within that general approach, that’s so…