00:18mhenning: karolherbst: is the way the call/return stack stuff works on V100+ written down anywhere, btw? I think I have a reasonable understanding of the pre-V100 stuff but all I know about the later stuff is *handwaves* uniform registers
00:20karolherbst: volta doesn't have that anymore
00:21karolherbst: the stack is completely mananged in the shader now
00:21karolherbst: we just don't do much of it atm
00:21karolherbst: and besides bad perf it doens't matter all that much
00:21karolherbst: normally you push/pop the masks when diverging/converging using the barrier registers, but I didn't get to it yet
00:22karolherbst: mhenning: check how I do quadon/quadpop for volta+
00:22karolherbst: that's the stuff we have to use
00:22mhenning: alright, I'll take a look
00:22karolherbst: for quadon we have to force quad operation mode, so that's what I do there :)
00:22karolherbst: that's importnat for tld lowering
00:23karolherbst: ehh txd
00:23karolherbst: GM107LoweringPass::handleManualTXD
00:37mhenning: ah, it looks like we save masks to registers in FILE_BARRIER and then FILE_BARRIER participates in RA?
00:37mhenning: makes sense to me
00:47karolherbst: yep
00:47karolherbst: I wired up all the core stuff already, because I needed it for TXD
00:47karolherbst: but we also have to use it for joinat/join which I have a branch for
01:23karolherbst: mhenning: needs rebasing, but: https://gitlab.freedesktop.org/karolherbst/mesa/-/commits/nv_volta_barriers/
01:23karolherbst: part of that landed because of TXD
01:24karolherbst: I am actually surprised how that missing joins don't even matter all that much
03:07xerpi: Hi! I already asked a few days ago, but does anybody know if it's possible for the Maxwell B (0xB197) to render to a pitch linear color target, and block tiled depth/stencil target? (0xB197 only supports block tiled depth/stencil)
22:49karolherbst: in case anybody wants to see support for Vulkan on Fermi GPUs, I guess this just become super low priority, because it's quite some work compared to Kepler. So if anybody wants to do that, feel free to raise your hand and I'll tell what needs to be ported over :)