01:12 <daniels> my officemate does design work for Lego and absolutely loves it, says they're great to work with and is looking forward to going to Billund

01:12 <daniels> not sure how NIR ties in with their ethos, but stranger things have happened

02:21 <mareko> graphitemaster: radeonsi is able to run as a compute-only driver not stalling graphics, but currently only as clover, but an env-var-based option can be added, the problem is that some basic things like pipe->blit are unavailable on compute-only contexts

02:21 <mareko> gawin: I don't remember anything like that

02:24 <mareko> graphitemaster: also you can boot the kernel with amdgpu.mcbp=1, which is preemption at the command buffer level, i.e. it can't stop shaders; AMD_DEBUG=info glxgears should then print: mid_command_buffer_preemption_enabled = 1

02:36 <gawin> mareko: example: https://gitlab.freedesktop.org/-/snippets/4195 (extra writing to output[2])

02:37 <mareko> gawin: it's for gl_FragCoord

02:38 <mareko> also if gl_FragCoord isn't used, it wastes performance I think

08:12 <mslusarz> hakzsam: WRT MR 11779, Eric and Dylan posted alternative MRs for the release calendar, but they can't decide which one to merge, and I'm not sure whether Joshua comment is a nack or not...

10:58 <graphitemaster> There's a lot of weird pseudo suggestions for how to get around preemption issues which all boil down to "if you have this extremely specific non-standard setup" as if to suggest that I'm the user trying to solve a problem - when really I'm a developer trying to ensure our software actually works on any reasonable Linux distribution. The focus is not "how do I work around this as a user", nor is it "what info do I have to provide

10:58 <graphitemaster> to users to make things work". I'm more interested in "how do we make this work for everyone running Linux without them ever being aware it was a problem in the first place".

10:59 <graphitemaster> Which I know is asking a lot, but that should be the goal here, at least I think it should be.

11:36 <ishitatsuyuki> I missed a large part of the discussion, but in my experience things will work pretty nicely as long as you do long-running work on the async compute queue

11:37 <ishitatsuyuki> The idea of preemption is significantly harder and more complex due to limited scheduling ability on GPUs

13:37 <graphitemaster> The goal is pretty simple, keep the desktop user-experience responsive. The solutions presented this far all seem to be what developers can do to make that possible (e.g compute-only queues, which is not available in OpenGL, only Vulkan and OpenCL), but furthermore - if you do specifically wait for the result because you need it, you're back to square one, the whole desktop goes poof. I just don't see how any of this is a

13:37 <graphitemaster> solution. It doesn't work to begin with (only applies for a subset of all the possible APIs, still relies on devs to do the right thing (i.e do not wait, use a compute-only queue, etc), a correct solution here would apply generally and never allow any graphics or compute application, no matter how poorly written from making user-space unresponsive.

13:38 <graphitemaster> But also the notion of user-space responsiveness does not mean losing the context when things time out, it means slicing and dicing the workload to keep things moving forward.

13:39 <graphitemaster> Anyways this is all very frustrating.

13:43 <Venemo> I think everyone here agrees that it's very frustrating

14:16 <daniels> and everyone here wishes they had around 24 more hours in their days, or we had 5x the number of developers working on low-level infrastructure

14:19 <ccr> only 24 hours more?

16:46 <pinchartl> daniels: I wish everybody else had 16h less in their days, that would make the e-mail influx more manageable

16:49 <daniels> pinchartl: a friend of mine was complaining last night that he gets sometimes upwards of ten emails per day

16:49 <daniels> …

19:58 <mareko> the most useful feature of mailboxes is "Mark as Read"

19:59 <pinchartl> mareko: I thought it was delete ? -):

22:02 <jekstrand> graphitemaster: Yes, it's a serious problem that needs to be solved. No one's arguing it isn't.

22:06 <jekstrand> daniels: You just need to learn to control time distortions.

22:09 <jekstrand> daniels: Just slow time down by about 50% and work a 12-hour day (to others) and that gets you an extra 16 hours.

22:41 <graphitemaster> jekstrand, So what is being done about it and how can I help

23:01 <FLHerne> jekstrand: Is NIR "New Intermediate Representation"? I have a vague memory of it becoming a recursive acronym instead

23:01 <FLHerne> [re. https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/14560 ]

23:41 <daniels> FLHerne: it got de-acronymed

23:49 <idr> I think the world learned not to name things "New" clear back with New Coke.

23:50 <imirkin> idr: like the New iPad?

23:54 <jekstrand> FLHerne: Yes, that was the original acronym. These days, it's just NIR.

23:58 <daniels> Windows New Technology