Skip to content

Latest commit

 

History

History
16 lines (13 loc) · 617 Bytes

note.md

File metadata and controls

16 lines (13 loc) · 617 Bytes

CUDA

CUDA threaded blocks assignment

  • map thread blocks to CUDA cores using dynamic scheduling policy

  • A warp is that different contexts that happens to run the same instruction.

Persistent thread

code in data parallel way, instead of the sheduling strategy You can fill all the thread with __shared__, while(1) {if (id == 0) idx += block}...(See in ppt)

There are smater assignment strategy

  • CUDA is dynamic assignment
    • there is a queue between CPU and GPU, kernel says 'give me more'****
    • assign more once a time -> save locker
    • the bigger stask first
  • SIPC is static assignment