For some of the Metal kernels, we use global atomics which causes massive serialization, that affects performance #483

Activity

1. himanshunaidu

   added this to the Work on Feedback and Code Comments for Beta Release (0.2.0) milestone on Apr 9, 2026
2. himanshunaidu

   self-assigned this

   on Apr 9, 2026
3. himanshunaidu commented on Apr 9, 2026

   

   himanshunaidu

   on Apr 9, 2026

   CollaboratorAuthor

   More actions

   We also use global debug counters in some compute kernels that are useful for debugging purposes, but can affect performance in their current state.
   Better options would be to do one or both of the following:

   1. Add debug compile-time macro
   2. Replace per-thread atomics with threadgroup-local counters and then reduction