Repository navigation
Luna Abyss (PPSA36257) #18334
Replies: 4 comments
|
Longer run with the workaround, same
AI-assisted: yes (Claude Code) |
|
The "AUDIO ONLY" transmission after the opening cut-scene is not a rendering problem: the title #18342 lets a bare module name fall back to So with #18342 and AI-assisted: yes (Claude Code) |
|
Longer run with #18342 (
Not yet seen: free movement or combat, since the scripted input only advances dialogue here. AI-assisted: yes (Claude Code) |
|
Where the interior's ~12 FPS goes (same setup, 480 s run, trace over 30 s from 400 s after the first present plus The largest GPU wait is the CPU-access path: These runs set AI-assisted: yes (Claude Code) |
Uh oh!
There was an error while loading. Please reload this page.
Existing thread
Title
Luna Abyss (PPSA36257)
Content version tested
01.000.005
AnyPS5 commit
main774111474c.System tested
Linux 7.2.14 (Arch), Ryzen 9 7950X, Radeon RX 6650 XT, RADV Mesa 26.2.6, GCC 16.2.1, Release.
Furthest point
Front end: the photosensitivity warning (40 s) and, after New Game, the Difficulty Select menu (62 s) render correctly. With
APS5_PLAYGO_UNLISTED_CHUNKS=installedand Cross every 15 s from 30 s, the title then stops about 66 s in on most runs (see below). Runs that avoid the crash reach the scheduled close at 122 s; what they show after 62 s was not captured.What is shown
First frame about 10 s in. Over 60–120 s with Cross pressed every 15 s from 30 s, the title runs at 51.6 to 60 FPS (59.95 on the
mainabove, 51.6 under gdb).Intermittent crash about 66 s in. On
774111474c, 2 of 3 runs under gdb and 1 of 1 plain run stopped at 66–68 s; on earliermainrevisions it stopped in 1 of 3. The two captured crash sites differ:libc.prxabort()(NIDL1SBTkC+Cvw), called from the title at0x1421cafafter a failed check (cmp byte [rax],1, then a call that does not return). Thatabort()raisesint $0x45, which Linux reports as SIGSEGV, exit 139.0x1785398(mov (%rdi),%rax) withrdi = 0x4ff3fb172181002, which is not a canonical address.Two different sites at the same point, one of them a non-canonical pointer, look like corrupted guest data rather than one missing function.
The crash is fix(agc): land GPU writes over guest memory before a host write into it #6926's.
mainfb26c5e4e0: 3 of 4 runs crashed, all at 66 s; the same commit with only fix(agc): land GPU writes over guest memory before a host write into it #6926 reverted: 0 of 4 (51.7-54.0 FPS). On774111474c,APS5_NO_UNIT_SHADOW=1also stops it: 0 of 4 (51.8-53.0 FPS), against 3 of 4 without it. The same race is described in fix(agc): publish a host write's whole units first so a unit shadow cannot land over it #17787, which may cover this title too; that PR is untested here.No AnyPS5 driver failure or
NotImplementedthrow is logged before either crash.Current blocker
Intermittent crash about 66 s in, from #6926's unit shadow race (see above). Workaround on current
main:APS5_NO_UNIT_SHADOW=1.How it was run
APS5_PLAYGO_UNLISTED_CHUNKS=installed. Modules insce_moduleare converted with the title.AI-assisted testing and report: yes, Claude Code.
All reactions