01:35 olivial: how do people usually go about evaluating the performance tradeoffs of a shader variant?
01:36 olivial: typical benchmarking where you look at mean frame times aren't going to really catch the cost of extra shader compilation, which I would expect to show up as increased startup time or stutter
02:12 chirs241097: cheako: will do. Thanks.
06:31 llyyr: is gitlab down? I get 504
07:11 Sachiel: yes, been down for several hours now
08:09 karolherbst: olivial: the thing is, that shader variants are often cheap. You don't have to wait until the compilation finished if you are smart about it, you just replace the shader on the fly once it's ready
08:12 olivial: oh, hmm, do you have any examples of that kind of thing in mesa I could look at?
08:13 olivial: context for this question was https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/34666, where I don't think there's a straightforward way to compile the variant asynchronously
08:13 olivial: you need it as soon as the blend state changes
08:14 olivial: oh, I guess I could precompile the variant for all shaders ahead of time, and then it would be ready if needed...
11:33 randrik: hi all again
11:34 randrik: I used to be AndrewR but apparently I forgot my password or it expired, so i re-registered as randrik
13:18 randrik: interesting. I am testing my "new" AMD GPU, RX550, it connexts 8x @ pcie 2.0 9old mobo). ffmpeg 7.1.1 can pull frames out of hw decoder via vaapi at 10 fps. cinelerra-gg (NLE I help maintain) pulls 13-14 fps! with display (x11). probably our agressive caching helps here. File itself is 10bit HEVC sample, 4k/60 fps :} mpv now plays it ok, but any vram->sys->vram roundtrip kill perf. Not really surprizing, but interesting to observe
14:18 Lynne: randrik: you can avoid the roundtrip by mapping hardware frames as software frames, and if you have to "upload" them, you can just detect that the frame is backed by hardware
14:19 Lynne: vaapi is entirely synchronous too, with vulkan you'd get free asynchronous uploads if you still do a roundtrip
18:44 randrik: Lynne, I think vulkan on gfx8/Polaris still not enabled due to hangs? (after looking into mes MRs)
18:45 randrik: https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/28988
20:42 randrik: Ah, test run worked for decode via vulkan and tonemap/scale via libplacebo, then encode on cpu with x264. 10 fps, so faster than opencl tonemapping with vaapi!
20:43 randrik: Lynne, thanks, I was sceptical about this vulkan h264/hevc video decode, but it working! :)