hipBLASLt icon indicating copy to clipboard operation
hipBLASLt copied to clipboard

Published 20 hours ago • verified publisher icon ROCm

[tensilelite] Update NumWarmups and MaxEnqueuesPerSync

Open AndySu12 opened this issue 10 months ago • 2 comments

  • NumWarmups: This factor will be restricted to 30ms if set, and also affect MaxEnqueuesPerSync.
  • MaxEnqueuesPerSync: max solution enqueues to perform per synchronization. This factor will be restricted if NumWarmups touches 30ms.

AndySu12 avatar Apr 22 '24 07:04 AndySu12

Question: should warmup introduce GPU time like enqueue?

AndySu12 avatar Apr 22 '24 07:04 AndySu12

FYI, there is an experimental implementation for GPU time-based warmup mechanism for tensilelite and hipblaslt-bench.

Serge45 avatar Apr 22 '24 08:04 Serge45