00:01karolherbst[d]: the `off.is_zero()` check is failing, but..: `%7 = @ldc_nv (%6 (0x0), %6 (0x0)) (base=0, access=none, align_mul=4, align_offset=0)`
00:08karolherbst[d]: uhhh.. I see what's happening
00:40karolherbst[d]: `CodeSize: 8929970880 -> 9122223216 (+2.15%); split: -0.00%, +2.15%` that's certainly better..
11:10marysaka[d]: mhenning[d]: let's disable compression for now and backport it on 26.0, I am not confident we will fix it before 6.19 release tbh
12:05karolherbst[d]: so...:
12:05karolherbst[d]: CodeSize: 9508362112 -> 9545524464 (+0.39%); split: -0.47%, +0.86%
12:05karolherbst[d]: Static cycle count: 6187073609 -> 6163350156 (-0.38%); split: -0.63%, +0.24%
12:05karolherbst[d]: Max warps/SM: 50607760 -> 50620712 (+0.03%); split: +0.03%, -0.00%
12:05karolherbst[d]: it's getting somewhere
12:15karolherbst[d]: I forgot `nir_intrinsic_load_global_constant_nv` :blobcatnotlikethis:
12:53karolherbst[d]: Totals from 119433 (10.27% of 1162627) affected shaders:
12:53karolherbst[d]: CodeSize: 2680608336 -> 2633982640 (-1.74%); split: -1.77%, +0.03%
12:53karolherbst[d]: Static cycle count: 3785670319 -> 3745346323 (-1.07%); split: -1.12%, +0.05%
12:53karolherbst[d]: Max warps/SM: 4235768 -> 4245328 (+0.23%); split: +0.23%, -0.01%
13:23karolherbst[d]: okay.. one of the "regressions" is interesting. NAK is able to identify more uniform sources, but has to revert it later, because it's in a simple if/else clause...
13:24karolherbst[d]: mhenning[d]: in case you are interested. I think this is a case we should be completely safe to keep the uniform register: https://gist.github.com/karolherbst/bebcce7ce641d95da07db59562e945f5
13:25karolherbst[d]: or uhm... can tld even take uniform registers?
13:26karolherbst[d]: aahh yeah, for the handle
13:26karolherbst[d]: though here it's a coordinate mhhh
13:27karolherbst[d]: but it's kinda funnky how the reversal of the uniform reg gives me worse stats
13:27karolherbst[d]: 70 -> 81 instructions and 314 -> 332 cycles
13:28karolherbst[d]: the original version just sets the zero once
13:28karolherbst[d]: or well.. less often
13:29karolherbst[d]: but also all those movs...
14:05phomes_[d]: karolherbst[d]: assuming this is !39525. Some intersting test results so far: Atomic heart drops 60->45 fps. Deep rock galactic improves 94->101 fps. Lego builders journey crashes
14:07karolherbst[d]: phomes_[d]: have you pulled my latest updates of that MR>
14:07karolherbst[d]: ?
14:08karolherbst[d]: phomes_[d]: do you have any details on the crash?
14:09phomes_[d]: it is the latest yes. Let me just reboot to not have preempt=full and then I will do all the usual tests in the sheet
14:10phomes_[d]: `MESA: warning: ../src/nouveau/vulkan/nvkmd/nouveau/nvkmd_nouveau_ctx.c:149: DRM_NOUVEAU_EXEC failed: No such device (VK_ERROR_DEVICE_LOST)
14:10phomes_[d]: MESA: warning: ../src/nouveau/vulkan/nvk_queue.c:306: Submit failed (VK_ERROR_DEVICE_LOST)
14:10phomes_[d]: MESA: debug: ../src/vulkan/runtime/vk_device.c:380: Timeline mode is ASSISTED.
14:10phomes_[d]: nouveau 0000:01:00.0: gsp: rc engn:00000001 chid:24 gfid:0 level:2 type:109 scope:1 part:233 fault_addr:0000000000000000 fault_type:00000000
14:10phomes_[d]: nouveau 0000:01:00.0: fifo:c00000:0018:0018:[Builder's Journ[98893]] errored - disabling channel
14:10phomes_[d]: nouveau 0000:01:00.0: Builder's Journ[98893]: channel 24 killed!`
14:12karolherbst[d]: ahh.. a NULL pointer
14:12karolherbst[d]: I should run the CTS and see if there are any trivial regressions before I dig into why games are crashing
14:30karolherbst[d]: phomes_[d]: has atomic heart any visual glitches? kinda annoying it's slower 😄
14:31karolherbst[d]: I don't think I have atomic heart fossils...
14:38phomes_[d]: yes there is actually some small visual glitches
14:40phomes_[d]: https://cdn.discordapp.com/attachments/1034184951790305330/1465355587997929668/Screenshot_From_2026-01-26_15-33-30.png?ex=6978ce3d&is=69777cbd&hm=ca5bcd9b42931dd7fb712736c2fb6d1da1b9d101da1d2365bf065a8dd6d42b15&
14:40phomes_[d]: the vertical stripes on the ground
14:40karolherbst[d]: okay..
14:40karolherbst[d]: so let me run the CTS and fix a bunch of bugs then 😄
14:41phomes_[d]: okay. I will wait with further testing then
14:43karolherbst[d]: I'm actually surprised it isn't blowing up completely
15:50karolherbst[d]: why is it always a nightmare to build the VK CTS :blobcatnotlikethis:
16:00karolherbst[d]: seems like the fix for the coop matrix tests has been merged into the CTS: https://github.com/KhronosGroup/VK-GL-CTS/commit/926657ed4bc9b04320a2cb3d7928861690a74183
17:32mhenning[d]: karolherbst[d]: Yeah, I;m still thinking about general ways to fix that
17:32karolherbst[d]: yeah... though I think my phi vec RA stuff might also help with some of the movs there, but I couldn't come up with a proper way of doing that without regressing even more annoying to fix RA problems
17:32karolherbst[d]: at least it would mitigate the movs I see there
17:33karolherbst[d]: also.. `140505, Fail: 12, Crash: 29, Warn: 1, Skip: 147026, Timeout: 427, Duration: 52:49, Remaining: 8:00:19` since when did we arrive at 8 hours CTS run time 🙃
17:34karolherbst[d]: mhhh
17:34karolherbst[d]: https://cdn.discordapp.com/attachments/1034184951790305330/1465399575820108110/image.png?ex=6978f734&is=6977a5b4&hm=5decdcbdeb66a2240412a07f0efa5caf5deba05312d007dc002a7add462e02f1&
17:34karolherbst[d]: is that the GSP RPC starvation issue?
17:35mhenning[d]: I can still CTS in two hours
17:37mhenning[d]: I tried setting DEQP_LOG_NODE_SOURCE=1 recently but it turns out it adds several hours to a cts run so I turned it off
17:37karolherbst[d]: mhhh
17:37karolherbst[d]: it's OFF here
17:38karolherbst[d]: ohh wait... I know this issue...
17:39karolherbst[d]: mhh no
17:44karolherbst[d]: it's faster with a user gnome session 🙃
17:44karolherbst[d]: or something
20:59karolherbst[d]: okay.. I found a bug...
21:08karolherbst[d]: ahh it's `nak_nir_lower_non_uniform_ldcx` that needs fixing
21:15karolherbst[d]: phomes_[d]: I've pushed a fix to the MR. Mind checking if the games with the issues are fixed now?
23:34phomes_[d]: all issues are gone. Performance in atomic heart is also back to normal
23:42karolherbst[d]: is Deep rock galactic still faster?
23:43karolherbst[d]: but yeah.. CTS looks clean now, so if it there are still issues I'll have to debug games 🥲
23:47phomes_[d]: it is still faster but not as much. 94 -> 98 fps
23:53karolherbst[d]: sounds more reasonable tbh.. the stats aren't that much better so any significant gain is already unexpected.. especially as it's like part 1 out of a part 5 series of optimizations 🙃
23:53karolherbst[d]: and this one is the smallest one out of all of them
23:59phomes_[d]: I got a proton update in the middle of testing. That could also explain it