15:50pac85: Hi, I'm the author of this MR https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/21238 and I'm getting a failure for the test KHR-GL46.shader_image_load_store.basic-allTargets-loadStoreTES on the zink-radv-vangogh-valve 2/3 runner. I was not able to reproduce the issue (I tried running it on a steamdeck, on rdna3 and vega) Is there something special about that envirnment?
16:46agd5f: gekret005, there wouldn't be any pre-OS environment in a VM AFAIK. I guess resizable BARs would probably be a hypervisor issue? Not sure off hand
16:48JoshuaAshton: Seems like there may be some regression in amd-staging-drm-next wrt Doom Eternal perf. I have filed a bug here: https://gitlab.freedesktop.org/drm/amd/-/issues/2467
17:26Venemo: agd5f: what is the current status of the user-mode queue submission patch series? I haven't heard anything about it since last year
17:29agd5f: Venemo, shashank is still working on it. made a lot of progress over the last few weeks. I think he's planning to send out a new revision next week
17:29Venemo: I'd be interested in implementing a user space side for that in RADV once you guys believe the kernel side is ready for use
17:31agd5f: great
17:48Venemo: agd5f: speaking of queue submission, we've been wondering, is it still worth it for userspace to chain the submitted IBs? I think bnieuwenhuizen said that at least this was a significant perf improvement in the past, but we're not sure if still needed or not?
18:07agd5f: Venemo, you mean as opposed to just bigger IBs?
18:08Venemo: agd5f: what we do currently in RADV is that instead of submitting several IBs to the kernel, we "chain" them and submit fewer IBs
18:10agd5f: I'm not sure. It's probably faster in that you get more work into a single command submission to the kernel so you don't have the extra overhead of scheduling multiple jobs
18:48Venemo: is each IB considered a different job?
19:01gekret005: https://bpa.st/QUE7Q sadly I spoke too soon looks like the 7900xtx does exhibit reset issues
19:17Venemo: gekret005: that log doesn't even mention amdgpu
19:17gekret005: well those are the other devices of the gpu
19:20cheako: I'm trying to look into presentation, but I couldn't locate `radv_CreateSwapchain` or `radv_GetSwapchainImages`. or rather what's radv_entrypoints?
19:48bnieuwenhuizen: agd5f: I think the Q is chaining IBs vs. submitting multiple IBs in a single IOCTL
19:49bnieuwenhuizen: but to answer Venemo I think it should be pretty close these days
19:49Venemo: yes that is exactly the question
19:49bnieuwenhuizen: though I think the number of IBs in a single submission has a limit due to the ringbuffer size
19:49bnieuwenhuizen: it is just so big I don't think anybody hit it yet