Repository navigation
Hogwarts Legacy (PPSA01593) #23362
birdmarty
started this conversation in
Game bring-up
Replies: 1 comment 1 reply
|
A data point for the Hogwarts Legacy (PPSA01603, 01.000.013), Ubuntu 26.04.1 (kernel 7.0), RTX 4070 Ti on NVIDIA 615.71, i7-13700K, GCC 15.2 Release. 600 s runs with scripted presses through the Accessibility screen, the EULA and the reward screens.
No commit between the two AI-assisted: yes |
1 reply
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Uh oh!
There was an error while loading. Please reload this page.
Uh oh!
There was an error while loading. Please reload this page.
Existing thread
Title
Hogwarts Legacy (PPSA01593)
Content version tested
01.000.013
AnyPS5 commit
main 6efc0f7 + #20827 + #9915 (head 1749190) + two local fixes (see Current blocker); libraries built from 66d27207a5
System tested
Windows 11 Pro (build 26300), AMD Radeon RX 6800 (AMD proprietary driver 26.10.1 / 32.0.21045.16010, Vulkan 1.4.315), Intel Core i9-9900K, 32 GB, MinGW-w64 GCC 15.3.0, Release
Furthest point
In game
What is shown
It boots through the logos, the menus and character creation. It plays the intro cutscenes and reaches gameplay: the cave, then the cliff path with Professor Fig. The last run went 518 s and presented 3,118 frames. I closed it and it exited cleanly. An earlier run with the same fixes, except the DCC one below, went 1,274 s without a crash.
Current blocker
1. Soft lock: Professor Fig stops following his path (not diagnosed)
About 20 s along the cliff path after the cave, the "Follow Professor Fig" objective stalls. Fig runs in a loop a few metres wide at one spot by the cliff edge and never moves on, so the quest cannot continue. The video is from the 518 s run.
5-follow-fig-soft-lock.mp4
I can't tell yet whether this is AnyPS5 or the title. One likely reason is the frame rate. The video has 62 distinct game frames in 15.2 s: a median of 0.2 s apart, and up to 1.0 s. So Fig covers a lot of ground between two frames. If his path following checks once per frame whether he has reached the next point, he can step past it every frame and turn back, which would look like this loop.
2. Stops already fixed, and what they need
AGC graphics: missing register in context bank at DWORD 0x2e2PA_SU_POLY_OFFSET_FRONT/BACK_SCALE/OFFSET(0x2e0-0x2e3) at 0, their clear-state value, as #1086 did for 0x2dfAGC graphics: vertex fetch exceeds descriptor record count: OOB_SELECT 2, stride 12, NUM_RECORDS 83, vertex index 114#9915 is someone else's PR.
0x2e2. A diagnostic build counted every write to 0x2de-0x2e3 over a whole run:
FRONT_SCALEandFRONT_OFFSET: 274,062 writes each, all throughSET_CONTEXT_REG_INDIRECT.BACK_SCALEandBACK_OFFSET: 0 writes, and the queue was never reset.PA_SU_SC_MODE_CNTL=0x1a40(bias on for both faces, no culling) and a D16 depth target, so most likely shadow passes.The title never writes the back registers, so the console uses their clear-state value, 0. Linux's GFX10 clear-state table has 0 for all four.
OOB_SELECT 2 fetch. A diagnostic build logged every direct draw that fetches past
NUM_RECORDS, over 393 s from a save at the cliff.OOB_SELECT2/3 and 584 withOOB_SELECT0 read past their records.GE_INDX_OFFSET32 or 33. In each, the highest index in the index buffer is exactlyNUM_RECORDS - 1, so only the index offset takes the fetch past the end. The stop above fits this: 82 + 32 = 114.Stretched triangles at the cliff.
CB_COLOR_VIEW. AnyPS5 moved the color address to slice N but left the DCC key address at slice 0. So only one slice per frame got its clear, and the other five kept adding to last frame's values until they overflowed.Suspects ruled out on the way:
NUM_RECORDS: skipping the direct and the indirect draws that do it, or widening shader buffer bindings past it, changed nothing;How it was run
relinker.exe --windows --to-intel eboot.bin app.exe
All reactions