ENH: new function replace #992

Activity

1. lucascolley

   added

   enhancementNew feature or requestNew feature or request

   new function

   on Sep 17, 2026
2. lucascolley

   added the

   Feature

   issue type on Sep 17, 2026
3. kgryte commented on Sep 21, 2026

   

   kgryte

   on Sep 21, 2026

   More actions

   Bikeshed: might be worthwhile considering the current constellation of names among various array libraries and whether we want to introduce new vocabulary or build on existing vocabulary. E.g., NumPy already has fill, so fill_equal could be an option.

   Additionally, the proposed operation is a functional equivalent to

   x[x==old] = new

   which is already supported by the standard vis-a-vis boolean indexing. In principle, one could have a dedicated kernel for this operation, thus avoiding the intermediate boolean array, but there is a question whether this API is actually needed given support for boolean indexing.
4. rgommers commented on Oct 29, 2026

   

   rgommers

   on Oct 29, 2026

   Member

   More actions

   Additionally, the proposed operation is a functional equivalent to x[x==old] = new

   Kinda but not really, since that doesn't work with JAX or other immutable implementations. The more generic equivalent is where(x == old, new, x), as in the issue description.

   I think it makes sense to add replace to array-api-extra, probably backed by an xp.where-based generic implementation. It provides:

   1. A cleaner API than the where equivalent,
   2. If implemented directly in array libraries with an optimized implementation that avoids constructing the boolean intermediate array, better eager performance

   Once (1) is in, we can use it immediately in scipy, since we vendor array_api_extra. There are probably a couple of dozen places where it can be used. Doing that will give a better feel for whether (2) is worth doing.