Why don't we just set all the fences to dirty?

[Dustray]

I noticed a lock before resetFenceDirty() in iHipWaitActiveStreams(3c6505c). The question I have is why don't we just set all the fences to dirty after dispatchBarrierPacket(

clr/rocclr/device/rocm/rocvirtual.cpp

Line 1053 in b825dec

if (cache_state == amd::Device::kCacheStateSystem) {

).

Would adding another lock and loop for this hurt performance?

Why can't other CacheState just set dirty like kCacheStateSystem?

Thanks and hope you answer!

[ppanchad-amd]

Hi @Dustray. Apologies for the lack of response. Internal ticket has been created to assist with your question. Thanks!

[schung-amd]

Hi @Dustray, is this question still relevant? I'll reach out to our internal teams if it is, but our codebase has changed in the meanwhile and I don't see iHipWaitActiveStreams anymore. Let me know if this has just been refactored and your question is still relevant, or if you have any questions about the current code.

[Dustray]

Ok I found that the locking judgment against isFenceDirty() has been removed in iHipWaitActiveStreams（now is Device::WaitActiveStreams()）.

[schung-amd]

Thanks for following up!