Repository navigation
Silksong (PPSA12544): "Act 1: Pharloom" title card never shows; no frame is presented for 68 s while the first area loads #3695
Description
Investigated and written by Claude Code (AI), and posted by me after review.
Title: Hollow Knight: Silksong (PPSA12544), Linux
What happens on the console: after the intro cinematic, the game shows the "Act 1: Pharloom" title card before gameplay starts. Reference (PS5 Pro): https://youtu.be/FZ41MMcxj50 at 1:23.
What happens here: the cinematic ends, the screen stays black, and gameplay appears directly. The card is never seen. The text is otherwise clean (menu and profile text are fixed since #1599), so this is not a font or atlas problem.
What the log shows: the game presents no frame for about 68 s right after the cinematic, while the first area loads:
- The frame dump (
APS5_DUMP_FRAMES_EVERY=30) has a frame at 223.3 s, then none until 291.2 s (67.9 s), then one more at 330.6 s. Normal pacing only returns around 354 s. - During that gap, Unity logs
Forcing call to sce::Agc::suspendPoint to avoid TRC R5089 breachevery 3 s (30 times, from t+255 to t+343 s). That message means the game's render loop has not submitted a frame for a long time.
So the card most likely plays while nothing is presented, or the game skips it because its timer ran out during the load. It is a symptom of the long load, not a missing feature.
Build: a local build from 2026-10-10: main cba8d86 plus what Silksong needs to reach gameplay on Linux. That is #846 (plugin conversion with --module-dir Media/Plugins), #1599, and the sceKernelGetModuleInfoFromAddr st_size change since merged as #3297. The run was user-driven (New Game, cinematic, then gameplay), with the window visible, on an i7-8750H with a GTX 1050 Ti and NVIDIA 580. I have not rerun it on current main yet.
Next step (not done yet): sample the main thread's stack during the black gap to see what the load waits on. The first suspect is synchronous file reads (APR/AIO run synchronously, one stream per read fragment).
I can attach the full log or the window recording if useful.
Activity
oneandonlydean commentedon Oct 13, 2026
Does the 68 s gap without a presented frame still happen on current main? Please post the frame-dump times from one run there.
adhimiw commentedon Oct 14, 2026
Regarding the frame presentation gap during the initial scene loading:
On current
main, during the asset streaming / initial scene transition, Unity executes intensive synchronous asset reads and recompiler tasks, causing render submissions to pause while background worker threads load assets. In composited environments (e.g. Wayland),vkAcquireNextImageKHRcan returnVK_TIMEOUTorVK_NOT_READYduring these prolonged intervals.Previously, this triggered an unhandled exception in
VulkanDevice::AcquireImageviacheck(acquired, "vkAcquireNextImageKHR"). We have submitted PR #5863 to cleanly treatVK_TIMEOUTandVK_NOT_READYas frame-drop conditions so the presentation loop smoothly resumes once scene loading completes.awake-devel commentedon Oct 14, 2026
I have tested on main yes
oneandonlydean commentedon Oct 14, 2026
Thanks. Could you post the frame-dump times from that run and the main commit it was on? That shows whether the gap is still 68 s or shorter now.
mrfelek commentedon Oct 15, 2026
Silksong on current
main: no frame-dump times yet, the run crashes at startupSetup
origin/main44df21632(2026-10-15), clean worktree, Release, Linuxorigin/mainrelinkerResults
origin/mainlibraries,origin/mainrelink (APS5_DUMP_FRAMES_EVERY=30, New Game pressed by script)globalgamemanagers.assetsand the Unity resources load (jump to address 0)origin/mainlibraries, relink from the Oct 10 runtimeorigin/mainrelink, older runtime librariesmain.origin/mainrelink neededMedia/Modules/Il2cppUserAssemblies.prxandPS5Util.prxconverted by hand. The plugin conversion from feat(relinker): add --module-dir for modules loaded at run time #846 (--module-dir Media/Plugins) produced no plugin entries in our run.cba8d86dwith feat(relinker): add --module-dir for modules loaded at run time #846, fix(agc): store GPU results over CPU-stamped bytes the CPU left unchanged #1599, fix(libkernel): ignore the input st_size of sceKernelGetModuleInfoFromAddr #3297). The three PRs don't apply cleanly to that base. Also, fix(agc): store GPU results over CPU-stamped bytes the CPU left unchanged #1599 is titled "store GPU results over CPU-stamped bytes", not a text fix, so I may have the wrong number for the menu-text fix.Do you still need help with this? If yes, the next step is finding the null call in the new relink. I can post the crash log and the three run logs.
AI-assisted; the builds and runs are from my machine.
mrfelek commentedon Oct 16, 2026
Update on current
mainmain(af7db81e6) still crashes at startup, right after loadingglobal-metadata.dat. A bisect points to feat(libc): provide the runtime scripting allocator hooks #5092, and fix(libc): give the scripting heap hooks blocks from libc's own heap #6842 fixes it (results on #6842). With fix(libc): give the scripting heap hooks blocks from libc's own heap #6842 the title runs at about 1,080–1,200 presents per 10 s.Media/ModulesandMedia/Pluginsconverted with--module-dir. IfMediais only a link to the dump, the title fails withunable to load /app0/Media/Modules/PS5Util.prx, error:0x80020002.AI-assisted: this comment was drafted with the help of Claude Code and checked against the logs by us.
MarcZX10 commentedon Oct 16, 2026
Thanks for testing the #5092 startup fix on Silksong. The #6842 result is useful: the build without the fix crashes before presenting frames, while the build with #6842 runs at about 1,080–1,200 presents per 10 seconds. That separates the startup crash from the original loading-gap report, but the steady rate does not show whether presentation stalled during the cinematic-to-first-area transition or whether the title card appeared. When convenient, could you add the frame-dump timestamps around that transition from the #6842 build and note its exact tested commit? That would close the remaining question without inferring the gap from the later presentation rate. AI-assisted: yes.