#
…

2022-12-03

00:01chengsun_ has joined #asahi-gpu
00:01capta1nt0ad has joined #asahi-gpu
00:02chengsun has quit [Ping timeout: 480 seconds]
00:48chengsun_ has quit [Ping timeout: 480 seconds]
00:49capta1nt0ad has quit [Remote host closed the connection]
01:44chadmedalyssa: works on kernel 332ca551eb93
02:20Emantor_ has quit []
02:20Emantor has joined #asahi-gpu
02:35zalyx has quit [Quit: later alligator]
02:37zalyx has joined #asahi-gpu
02:38user982492 has joined #asahi-gpu
02:57ma has joined #asahi-gpu
03:03ma4 has quit [Ping timeout: 480 seconds]
03:37user982492 has quit [Quit: My MacBook has gone to sleep. ZZZzzz…]
03:41mini0n has joined #asahi-gpu
03:44alyssa has joined #asahi-gpu
03:45alyssaoh lol, pool allocations in magic.c mean the BO handles are wrong, I assume that's why macOS+mesa was unstable
03:45alyssa(and why Linux was fine)
04:31mini0n has quit []
06:09SSJ_GZ has joined #asahi-gpu
06:13corion has joined #asahi-gpu
06:30corion has quit [Quit: Page closed]
07:30mattgirv has quit [Server closed connection]
07:30mattgirv_ has joined #asahi-gpu
07:30cr1901 has quit [Read error: No route to host]
07:31cr1901 has joined #asahi-gpu
07:40linajannau: Is this with yesterday's clustering fixes, or only the incomplete ones from wednesday?
07:40linasteven: I replied on #asahi-stream but we should probably talk about texture issues here since alyssa is here ^^
08:05jannaulina: current gpu/rust-wip is broken. not sure that clustering itself is the problem as setting WaitForPowerOff avoids the issue as well
08:10corion has joined #asahi-gpu
08:37n1c has quit [Server closed connection]
08:37n1c has joined #asahi-gpu
08:41robinp has joined #asahi-gpu
09:13jannaulina: bisected to bedf1a6cc218669 "drm/asahi: render,buffer: Implement TVB auto-expansion"
09:16corion has quit [Quit: Page closed]
09:29nuup has quit [Server closed connection]
09:29nuup has joined #asahi-gpu
09:37corion has joined #asahi-gpu
09:50corion has quit [Remote host closed the connection]
09:52corion has joined #asahi-gpu
09:53corion_ has joined #asahi-gpu
09:54corion has quit [Remote host closed the connection]
09:56corion_ has quit []
10:36linaInteresting!
10:36linaI did that late last night so it didn't get much testing...
10:36linaI think I'll have to go back to tracing macOS VM accesses around that part of the code, there's probably something I'm doing wrong...
10:37linajannau: setting initial_tvb_size large enough should also avoid the problem, right?
10:37linaYou can also add debug_flags=0x2000 to get messages when TVB growth happens, which might be useful
10:38linaThis might just be that the TVB is too small to work properly and the logic for the initial size is wrong, or missing some factors
10:39linaI know it's missing tile size but I think that mesa branch doesn't have varying tile size yet I think?
10:45chengsun has joined #asahi-gpu
10:47anuejn has joined #asahi-gpu
10:48jannaudoesn't look like mesa:asahi/main has varying tile size
10:49jannauif the initial tvb size was shouldn't I see already issues with "drm/asahi: buffer,render: Add TVB minimum size formula and enforce it"?
10:52corion has joined #asahi-gpu
10:58vup has joined #asahi-gpu
11:00jannaulina: https://www.jannau.net/asahi/gpu/2022-12-03_gpu_tvb_buffer_grow_supertuxkart_start_race_bedf1a6cc21.log, supertuxkart start at "Dez 03 11:50:30", start race at "Dez 03 11:50:57" I think
11:14linajannau: If the growth is what breaks then setting the initial TVB size higher would avoid that
11:16linajannau: Okay, that's weird. I don't think supertuxkart is more than one process? So that should be one TVB, so it not growing linearly is wrong...
11:16linaUnless it creates multiple GL contexts by opening the device more than once somehow?
11:18Dcow has quit [Ping timeout: 480 seconds]
11:20linaOh wait, those are other apps aren't they
11:21linajannau: The grep cut off the fault info, can you get me that? ^^
11:22Dcow has joined #asahi-gpu
11:25LinuxM1 has joined #asahi-gpu
11:51jannaulina: still broken with an insane initial_tvb_size=0x200, no tvb resizes with that, fault is the same with a different vm_slot
11:52jannauis 'address: 0xc' expected? I guess not
11:57LinuxM1 has quit [Quit: Leaving]
12:27corion has quit [Quit: Page closed]
12:34c10l has quit [Quit: Bye o/]
12:52goldsoultheory has quit [Quit: My MacBook has gone to sleep. ZZZzzz…]
13:04c10l has joined #asahi-gpu
13:04xiaojun has joined #asahi-gpu
13:04xiaojun has quit []
13:20chadmed_ has joined #asahi-gpu
13:25yamii has joined #asahi-gpu
13:27linajannau: That sounds like a null pointer, but then it can't bisect to that commit because that commit's code is never running...
13:28goldsoultheory has joined #asahi-gpu
13:31jannaulina: bisect runtested 918e75228b2fe6c
13:31jannaulina: bisect tested 918e75228b2fe6c explictly without reproducing the issue
13:32linaThis stuff sometimes isn't reproducible... yesterday's clustering stuff broke glmark2-refract at 1080p initially but not later
13:32linaBut larger sizes always broke
13:33linaThe only part of bedf1a6cc218669 that changed that doesn't only run with TVB resizes is the 4 lines of code around @@ -360,10 +386,6 @@ impl Buffer::ver {
13:34linaCan you try reverting (re-adding) those and seeing if it does anything?
13:34linaIf that's not it then I don't understand
13:35linaYou can also try removing the matching store in auto_grow() which also runs unconditionally
13:35lina(I basically just moved that store)
14:34goldsoultheory has quit [Quit: My MacBook has gone to sleep. ZZZzzz…]
14:47jannausigh, bisect seems be bad, error reproducible with bedf1a6cc2186 reverted
14:54chadmed_ has quit [Remote host closed the connection]
15:03jannautesting supertuxkart now with kms instead of wayland and now everything is bad, landing now on "drm/asahi: Add overflow detection/guard areas to allocator" but probably only since I marke the previous commit as good without testing
15:07jannaudisabling Clustering still fixes it
15:10linaI don't think the previous commit could've caused it?
15:10linaIt sounds like this is probably broken ever since I enabled clustering (so 44113d0e62)
15:15jannauI'd guess so too but unfortunately not consistently broken
15:33goldsoultheory has joined #asahi-gpu
15:41goldsoultheory has quit [Quit: My MacBook has gone to sleep. ZZZzzz…]
16:00linaalyssa: If you get a chance, try echo 0x01000000000000 > /sys/module/asahi/parameters/debug_flags
16:01linaIt seems to make things 5-10% faster (at least on M1 Ultra), but I have no idea what it does or whether always enabling that has a downside
16:14linajannau: I can reproduce it, by the way ^^
16:14linaBut I still have no idea what it is...
16:14SSJ_GZ has quit [Ping timeout: 480 seconds]
16:14linaOkay, now I know what it is
16:15linaasahi.debug_flags=0xf00002000 at boot time (important) makes it work, and also spam dmesg with buffer overflow errors
16:15linaYay, my allocator debug stuff is useful! ^^
16:16linaIt's the cluster tilemaps!
16:16linaSo why are those undersized...
16:17linaOr maybe I still got the stride somewhere wrong...
16:18SSJ_GZ has joined #asahi-gpu
16:19linaActually I think in this particular case you don't even need to set it at boot time, that only matters for the kernel allocators anyway
16:32rappet has quit [Server closed connection]
16:32rappet has joined #asahi-gpu
16:46Rayyan has quit [Server closed connection]
16:46Rayyan has joined #asahi-gpu
16:49corion has joined #asahi-gpu
16:50linaOK, looks like we have double the tilemaps as usual?
16:52user982492 has joined #asahi-gpu
16:55ayke has quit [Server closed connection]
16:56ayke has joined #asahi-gpu
16:56corion has quit [Quit: Page closed]
17:01Dcow has quit [Remote host closed the connection]
17:02Dcow has joined #asahi-gpu
17:02deflax has quit [Server closed connection]
17:02deflax has joined #asahi-gpu
17:07jn has quit [Server closed connection]
17:07jn has joined #asahi-gpu
17:22os has quit [Server closed connection]
17:22os has joined #asahi-gpu
17:30paddatrapper_ has quit [Server closed connection]
17:30paddatrapper_ has joined #asahi-gpu
17:41linajannau: Should be fixed, but I'm really not happy with the magic *2 factor...
17:53linaIf you ever notice weird breakage again (especially clustering-related), try that flag. It should quickly find if a buffer is undersized somewhere.
17:59corion has joined #asahi-gpu
19:07c10l has quit [Read error: Connection reset by peer]
19:08c10l has joined #asahi-gpu
20:57goldsoultheory has joined #asahi-gpu
21:30ChaosPrincess has quit [Quit: WeeChat 3.7.1]
21:32ChaosPrincess has joined #asahi-gpu
23:21SSJ_GZ has quit [Ping timeout: 480 seconds]
23:37Dcow has quit [Remote host closed the connection]
…