06:30 phasta: Can someone give me an R-b for this trivial patch pls?
06:30 phasta: https://lore.kernel.org/dri-devel/20260625104903.1713213-2-phasta@kernel.org/
07:45 mripard: pq, emersion, daniels: any chance you could review https://lore.kernel.org/r/20260918-drm-reset-state-flag-v4-0-5ad106370f05@kernel.org ?
13:12 thellstrom: mripard, mlankhorst: Ack to merge this through drm-xe: It's an urgent fix for a drm_pagemap patch that also got merged through drm-xe-next. A bit unclear how: https://patchwork.freedesktop.org/series/174556/
13:34 mripard: thellstrom: yep, works for me
13:35 thellstrom: Thanks.
15:35 Lynne: yo, why are compute shader objects not cached to the drive like compute pipelines?
15:35 Lynne: RADV doesn't
15:35 Lynne: actually none of the mesa drivers cache shader objects
16:39 alyssa: Lynne: because unlike 1.0 pipelines, shader objects are designed to allow the app to do its own caching properly
16:40 glehmann: I would argue nothing about ESO was designed properly
16:40 zmike: hater detected
16:43 glehmann: proudly so
16:46 cwabbott: glehmann: you are now cursed with a thousand GPL state merging bugs
16:46 cwabbott: you must ponder the spec as penance until you repent the sinfulness of your ways
16:47 glehmann: at least it has state
16:47 kisak: death by a thousand pointers instead of needles?
16:48 mareko: which driver has perf results for GPO vs GPL vs ESO/zink other than RADV?
16:49 glehmann: ESO is just big green propaganda targeted at d3d11/gl application devs that want to eat their cake and have it too
16:50 mareko: the RADV perf is/was GPO > GPL > ESO; if you want it slow, use ESO
16:50 alyssa: mareko: hater detected
16:50 mareko: it's a conclusion based on data
16:51 alyssa: OOI, why is GPL perf different than ESO on radv?
16:51 mareko: RADV has optimizations for GPO and static state
16:52 mareko: also: static state > dynamic state
16:52 alyssa: GPL fast link vs ESO i mean
16:52 mareko: not sure what fast link is
16:54 glehmann: fast link is using the three parts (VS prolog + pre-raster stages + fs) without the link time optimization thing
16:55 mareko: the GPL perf tests that I have use only separate libraries
16:57 mareko: the perf issues discovered were in pipeline throughput: fixed-func pixel and primitive throughput of the GPU decreases the more "dynamic" and "separate" things are used
16:59 mareko: a lot of it is fixable (and some of it has been fixed), but it shows the importance of perf testing and validation
17:10 mareko: the final observed perf is actually the result of a min function of different pipeline pieces, e.g. for the pixel pipeline, it's min(fs_threshold/fs_execution_time, pixel_pipeline_throughput), the goal is to measure all inputs of the min function separately; compiler folks care about shader execution time, state management folks care about pipeline throughput, and rarely anybody needs FPS results other
17:10 mareko: than as a "canary in the mine" type of validation
18:55 orowith2os: https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/44265 is good to review and/or merge whenever someone wants to get around to it
19:08 cwabbott: dj-death: ok, I think for my input attachment MR I've mostly got everybody except intel passing CI
19:08 cwabbott: I'll try to figure out the intel failures but if I don't then it would help if you looked into it
19:09 dj-death: cwabbott: sure
19:09 dj-death: a bit sad
19:10 dj-death: I don't think we can backport your MR
19:10 dj-death: so I'll have disable EXT_descriptor_heap until we can it
19:12 cwabbott: my laptop is TGL, so there's a chance I can figure it out myself
19:12 dj-death: I'll only look at it in about ~10h
20:11 cwabbott: dj-death: nvm I figured it out, I think
20:11 cwabbott: hoping this next CI run will be entirely green
20:12 cwabbott: https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/44601/diffs?diff_id=5900644&start_sha=8d54b226f73f4e4f5a793b703a78ce2784e0fbc9
21:06 cwabbott: dj-death: ok, after that fix there are still some failures, in dEQP-VK.rasterization.rasterization_order_attachment_access.* on RPL... I don't have that so I might not be able to figure it out
21:07 cwabbott: e.g. https://gitlab.freedesktop.org/mesa/mesa/-/jobs/111005587