21:20Manoa: hi, sayin if I use glamor in DDX, will it give more performance in GL applications ?
21:22imirkin_: probably less.
21:24Manoa: thank, is thare anything else I can do to give more GL performance beside boost and pstates ?
21:25imirkin_: buy an AMD gpu
21:25joepublic: knew that was coming :)
21:25imirkin_: or figure out how to better tune nouveau somehow
21:26imirkin_: or do the necessary RE on zcull
21:26imirkin_: so that we can make use of it
21:28Manoa: I guess thare is littel point to ask if I add another card of same type that it will give any performance ? the matrix show no way to use additional cards :x I guess nothing new in mesa 20 in this department ?
21:29Manoa: ok, thank :)
21:29imirkin_: esp for modern renderers, it's unclear how to make use of multiple GPUs
21:33Manoa: you know I do wanne buy AMD card :) but I don't like 5700XT I wanne waith for the next version from AMD nevi 2 or whatever so it will give significant performance to run 4K
21:36imirkin_: yes, the next hardware will always be better than the current hardware
21:37karolherbst: I've heard that we still lack support for memory compression in a lot of cases
21:38Manoa: I don't understand this 2 things suposed to be in hardware no ? 780Ti have hardware culling and memory compression no ?
21:38imirkin_: controlled by software.
21:38imirkin_: ZCULL is an optimization for depth buffers
21:38imirkin_: probably akin to "compression" of some sort
21:39imirkin_: and there's various "compression" on color too
21:39Manoa: so it's thare but is not automatic :x
21:40imirkin_: yeah, i mean obviously culling is there
21:43karolherbst: well, we have to configure certain parts from userspace as well
21:44karolherbst: sadly, most of the feature aren't enabled automatically :p
21:44karolherbst: imirkin_: we donly do depth buffer compression or so, right?
21:44karolherbst: if at all
21:45imirkin_: and msaa
21:47karolherbst: zcull will be fun to implement :/ I think it also needs firmware support, because context switching
21:48imirkin_: on nv50, yes
21:48imirkin_: but i think nvc0, it's all outside of the ctx
21:56karolherbst: imirkin_: okay.. I think, what happens is, that we have to allocate a bo and bind it to the graph context and this is essentially the buffer for the zcull context, and what actually context switches that doesn't matter from userspace perspective anyway and I assume nvidias firmware will already do whatever it has to do
21:58karolherbst: ehh, volta has 4 perf counters related to zcull as it seems
22:01karolherbst: I think.. might be something else though
22:02karolherbst: heh NV_PGRAPH_CLASS_ERROR_CODE_ERROR_ZCULL_SUBREGION_LIMBO 0x0000000e
22:16karolherbst: imirkin_: mhhh, if I would have to make an assumption, I think we need to manage the zcull buffer per rendertarget and probably bind a new zcull buffer each time the rendertarget changes, but this also only seems to really make sense for fb and we could probably ignore it for textures or something
22:17imirkin_: well, per zeta buffer
22:17imirkin_: coz, you know, zcull
22:17imirkin_: not colorcull
22:20karolherbst: mhh, right
22:20karolherbst: so, more like if the depth buffer changes, we probably also want to switch the zcull buffer as well
22:21imirkin_: yeah, i suspect they shoudl be linked
22:21imirkin_: again, for nvc0
22:21imirkin_: on nv50, it's different
22:21imirkin_: it's more global
22:21imirkin_: and less clear.
22:22Manoa: thank for look into this problem, guys :) if you wanne test thing and you don't have 780Ti, I can compile and testing, both kerenal thing and mesa (meson and ninja are realy fast :)
22:26imirkin_: Manoa: mostly what's needed is development power
22:26imirkin_: not testing power