08:26 <kusma> zmike: Just gave it a whack, and just immediately returning from copy_scanout raises my glxgears fps from 4k to 6.5k, so it seems winsys stuff might indeed get us much closer to match iris for this very important benchmark use-case ;)

08:27 <kusma> The interesting bit IMO is that copy_scanout doesn't come that high up in profiling, so I actually wonder if we're GPU bound in this case...

08:28 <kusma> (Yes, we are, scaling down the window from the default 300x300 makes the numbers go up)

08:28 <kusma> That's actually kinda impressive, given how simple the rendering is...

08:36 <kusma> kopper sadly segfaults in util_fast_urem32 called from zink_kopper_displaytarget_create()...

08:38 <kusma> That looks like a regression in the hash-tables to me... Seems hash_table_search doesn't work if the hash-table is empty...

08:39 <kusma> But if I fix that, I get one frame of glxgears, and then "X connection to :0 broken (explicit kill or server shutdown)."

08:42 <kusma> Hmm, looking more at it, that's not a regression, it seems hash_table_search with an empty table was just never really legal...

08:43 <kusma> (for some definition of "legal"; we've been doing this with this assert for about 3 years, and before that we did "% 0" instead, which isn't exactly well-defined either IIRC)

13:01 <zmike> hm I probably never saw this since I have hash table patches in my branch