08:14 <icecream95> alyssa: To get most of the benefit, CRC should be enabled for exported resources as well

08:15 <icecream95> Without a side-channel to communicate CRC state, GPU blits will always have to be used for writes so the checksum is always correct

08:19 <icecream95> One problem is what happens when two exported resources are written to at once--the hardware can only write CRC for one render target

08:25 <icecream95> A solution could be zeroing the checksum for the other RTs in a compute shader and hoping that nothing rendered later has a checksum of zero...

08:27 <icecream95> An assert might be fine too, util we find an application that actually needs to render to two window surfaces at once

09:41 <daniels> well, you have two options there

09:41 <daniels> one is to declare an extra modifier which has the CRC state in an auxiliary plane, to guarantee that both sides understand and update the CRC

09:42 <daniels> the other is to not do that, and assume that anything external is unaware of the CRC and that it might be trashed when it comes back to you

09:43 <daniels> this is basically what happens with Intel/AMD external 'compression' (i.e. memory bandwidth optimisation, not AFBC-style inline compression), where you do a full resolve when it goes out, then assume when it comes back that the external state is invalid

09:46 <macc24> how's AFBC support on midgard?

09:48 <icecream95> daniels: Would that be of much help for a tiled renderer, where you (hopefully) write to each resource once before sending to the compositor?

09:54 <icecream95> macc24: It works and is enabled by default, but currently forces much more GPU synchronisation than needed, so a couple of games (such as DarkPlaces) are unplayable

09:56 <daniels> icecream95: in Wayland you can assume that your resource is not going to be modified, yeah

09:56 <daniels> that's not necessarily true of X11, and it's very not true of GBM, where people do use it for alloc + export and then import elsewhere to mangle it

11:38 <alyssa> icecream95: I don't know winsys stuff enough to comment here, deferring to daniels on this one :)

11:40 <alyssa> But I'm skeptical of solutions requiring a compute shader to manually hack on the CRC

12:42 <alyssa> I don't believe the blob uses CRCs for winsys btw

14:38 <alyssa> On gnu/linux,anyway

14:38 <alyssa> maybe on android