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
It would be really nice to have a scatter_mul op, which is identical to scatter except using multiplication reduction. Similarly, scatter_max/min would be nice. There are other functions (select_assign) that could probably benefit from this as well. I suspect it might be better to have an enum for reduction strategy, rather than many functions, and just pass it as an arg to scatter?
Feature motivation
Pytorch has this functionality, so it would be nice for parity reasons.
Personally, I've wanted this a few times while working on #1998 but it isn't a pressing issue. From my looking around it seems like it would be quite a bit of work, so for now I'm not doing it myself (just making a feature request). Once I'm done with all the other sparse stuff I hope I will be able to get around to this, if it hasn't been done by then.
The text was updated successfully, but these errors were encountered:
Hi @McArthur-Alford
I'll be working on porting both reduction kernels and scatter/select operations using CubeCL in the coming days. I'll try to think of an elegant way to generalize this.
Feature description
It would be really nice to have a
scatter_mul
op, which is identical to scatter except using multiplication reduction. Similarly,scatter_max/min
would be nice. There are other functions (select_assign
) that could probably benefit from this as well. I suspect it might be better to have an enum for reduction strategy, rather than many functions, and just pass it as an arg to scatter?Feature motivation
Pytorch has this functionality, so it would be nice for parity reasons.
Personally, I've wanted this a few times while working on #1998 but it isn't a pressing issue. From my looking around it seems like it would be quite a bit of work, so for now I'm not doing it myself (just making a feature request). Once I'm done with all the other sparse stuff I hope I will be able to get around to this, if it hasn't been done by then.
The text was updated successfully, but these errors were encountered: