01:10DemiMarie: How bad is it to copy all rendered buffers from the GPU back to RGBA buffers on the CPU and then back to the GPU? I know it isn’t great, but I want to know just how bad it is.
01:16HdkR: Sucking framebuffers through a straw because PCIe is so bandwidth constrained :)
02:47DemiMarie: HdkR: how much of a performance hit are we talking about, quantitatively speaking?
02:49DemiMarie: I ask because SR-IOV and other GPU virtualization solutions may not provide support for GPU-side memory sharing, whereas CPU-side shared memory is much easier.
02:49DemiMarie: This is especially true for solutions that were originally intended for compute workloads.
02:51HdkR: Looking Glass's KVM solution does framebuffer readbacks for that same reason
02:52DemiMarie: How well does it perform?
02:52HdkR: It performs pretty well in testing. A bit worse on consumer platforms because of the limited PCIe lanes and DRAM bandwidth
02:53HdkR: Probably single digit percentage hit as long as you're not saturating your DRAM on the CPU side?
02:54HdkR: Ideally they could utilize P2P DMA at some point
02:59DemiMarie: In Qubes OS we have our own GUI protocol but it requires CPU-side shared memory. I’m wondering how important it is to move to a protocol that supports GPU-side shared memory, and if so, how one could be written that does not require writing a whole bunch of code for each different GPU.
03:06HdkR: You'll find that lower end hardware will become very sad doing copies to system ram constantly :P
03:23DemiMarie: HdkR: what do you mean by “lower end”?
03:24DemiMarie: In the PC world most low-end systems have integrated graphics so system ram and VRAM are one and the same.
03:27HdkR: Some of those weirdo low end laptops with like single-channel memory plus the lowest end dGPU that can be stuffed in to the chassis
03:28DemiMarie: What kind of dGPU is that? Are these x86 or Arm (Qubes only supports x86)? How much RAM do these systems have?
03:29DemiMarie: In other words, are these systems one could realisticly run Qubes OS on, or would they be very sad anyway for other reasons, assuming that they even booted in the first place?
03:31HdkR: These aren't old systems. Companies pushing out like $500 x86 machines with 8GB of RAM and a GTX 1650
03:32HdkR: with iGPU on the CPU of course
03:33HdkR: Or I guess GTX 1630 class is more likely :D
03:36DemiMarie: Ouch, nvidia :D
03:37DemiMarie: any chance that nouveau will get decent performance in the near future?
03:37HdkR: Theoretically with the new GSP stuff
03:54DemiMarie: Practically?
14:35tertl8: hi
14:36tertl8: if I want to install nvk with the open source modules, i install the modules first right?
14:36tertl8: or is it included?
14:40mareko: do you also remember the days when glxgears was the only test people ran before git push
14:43penguin42: what else would you need?
14:44zmike: now it's specviewperf
16:16DodoGTA: tertl8: NVK only works with nouveau kernel module
16:16DodoGTA: There's a MR to add some OGK support but it's very out-of-date and needs quite a lot of rebasing
16:21mareko: zmike: you wish