06:06 nmschulte: Looking to use OpenCL under Debian (Bookworm/Sid/Experimental ...) w/ gfx909, gfx1102, or gfx1103_r1 -- but using mesa-opencl-icd (v22, or v23) in Debian seems to produce OpenCL program build failures ...
06:06 nmschulte: I've symlinked gfx909/... .bc files in /usr/lib/clc/ to tahiti... but still no luck.
06:15 HdkR: Have you tried rusticl?
06:16 nmschulte: unsure ... clinfo reports two platforms: Clover / Mesa / OpenCL 1.1 Mesa 23.1.0 / FULL_PROFILE and rusticl / Mesa/X.org / OpenCL 3.0 / FULL_PROFILE
06:16 nmschulte: is there an env var to choose the other platform?
06:17 airlied: rusticl won't work on that gpu that old a mesa
06:18 HdkR: 23.1.0 is too old?
06:18 nmschulte: I admit I am not well versed w/ all this; just trying to do some photogrammetry -- and concur; 23.1 is the latest I can find in Debian, but open to it being "old."
06:18 nmschulte: should I look to AMD GPU PRO stuff?
06:19 HdkR: Or just Tahiti being too old?
06:20 nmschulte: I think it's that Tahiti is too old.
06:21 HdkR: 23.1.0 came out like three weeks ago
06:21 nmschulte: https://github.com/llvm/llvm-project/issues/52665 -- is where I got the suggestion, but it seems the "error: -11" is the same problem I am receiving
06:21 nmschulte: and gfx909 is about the oldest GPU I have available right now.
06:22 airlied: ah maybe rusticl could work on 23.1.0 , RUSTICL_ENABLE=radeonsi is the thing to try
06:23 HdkR: clinfo should also say if it works.
06:23 HdkR: Platform Name rusticl
06:23 HdkR: Number of devices 0
06:23 HdkR: If it fails example
06:24 nmschulte: should I keep the tahiti symlinks in the rusticl case too?
06:25 airlied: shouldn't matter for rusticl, but I think radeonsi support is only in master
06:26 airlied: HdkR: 23.1 branched about 2 months ago though
06:26 HdkR: ah
06:26 HdkR: That would make sense then
06:27 nmschulte: heh, damn. I give up. Seems it just wasn't mean to be, photogrammetry and me (in GNU/Linux w/ AMD).
06:28 nmschulte: clinfo seems happy w/ mesa + tahiti symlinks; RUSTICLE_ENABLE=radeonsi clinfo seems happy as well, and Meshroom shows the device, but says it fails to find the platform when I select just the rusticl device.
06:29 nmschulte: is it possible to disable the Mesa platform w/ rusticle enabled?
06:36 airlied: I doubt rusticl will do what you want in that package, so I'd avoid the enable there
06:37 airlied: clover might be possible to drag over the line, what errors do you get on build?
06:37 airlied: but really opencl isn't really fit for consumption on any Linux driver, amdgpu pro might get you some of the way but might also fail in strange ways
06:37 nmschulte: all I get is: error: -11 -- nothing more.
06:38 nmschulte: I found this, am trying ROCm: https://math.dartmouth.edu/~sarunas/amdgpu-opencl.html
06:49 nmschulte: have to sign off now; thanks for the assistance.
15:36 hakzsam: mareko: pepp: fyi, arb_pipeline_statistics_query-frag is failing on rdna2+ because of VRS flat shading
15:57 mareko: hakzsam: that's expected and correct
15:58 mareko: tests can't make predictions about how many PS invocactions execute
18:58 mareko: RADV is likely CP-bound: https://www.phoronix.com/benchmark/result/amd-radeon-rx-7600-series-linux-benchmarks/hitman-3-1920-x-1080-ultra.svgz
18:58 mareko: for 7900 XTX