Ghostrunner 2 (PPSA09911)

[saarthakpant]

• I searched this category for the title ID and no thread exists for this title on any platform.

Title

Ghostrunner 2 (PPSA09911)

Content version tested

01.000.000

AnyPS5 commit

main 50efb5c (2026-10-21) with #8336 merged on top, no other local changes.

System tested

Windows 11 (26200), GeForce RTX 2060 SUPER (NVIDIA 617.42, subgroup size 32), Intel Core i5-9600K, 32 GB RAM, MinGW-w64 GCC 15.2.0 r7 (UCRT), Release libs.

Furthest point

In game

What is shown

Logos, title screen, menus and gameplay at 3840x2160: the first level plays, the character moves, dies and respawns on keyboard input (Space through the loading screen, then a fixed key cycle with the respawn key; controller not tested). Gameplay runs at 43 presents per 10 s (median of 18 steady windows, about 4.3 FPS), GPU busy 173 ms per present. Picture and input work; audio was not checked. No stop in a 5-minute run.

Current blocker

Performance, not a stop. Two things bound the frame on this main:

• CPU copies of buffers outside the host imports: 1.55 GiB per 10 s window (25.9 GB over the run, 22.9 GB of it from the region at 0x2050000000), about 2.5 s of every 10 s on queue worker 0. This is the case perf(agc-driver): join every range a bound buffer crosses #5372 removes (open, conflicting with main): with it merged locally on main 4b462ef the copies were zero, GPU busy per present 75 to 83 ms and the frame 171 ms at the median.
• The draw committer thread: about 40,000 draw packets per 10 s at 77 to 86 µs each (capture 18.6, recompile lookup 19.2, lock wait 20.1 µs per draw), the problem perf(agc): the queue-0 draw committer bounds draw-heavy titles on the CPU — is parallel draw recording planned? #7895 tracks. The draw cache hits 0.0 % of draws in gameplay: 97 % of the misses are draws whose only difference is per-draw pointers in the user data (Draw cache misses on almost every draw when a guest driver puts per-draw pointers in the user data (RADV titles) #4054), and relocation (perf(agc): find cached draws across moved user-data pointers #4943) does not reach them because every vertex stage reads its fetch tables through such pointers and the vertex shaders build their buffer descriptors from the pointer words; on local prototypes stacked on perf(agc): reuse cached draw stages when only buffer base addresses moved #5609 and perf(agc): relocate cached draws across user words that only reach the shader as values #6539 the hit rate stays at 0.5 to 0.7 % because the per-stage alias is keyed on SRT contents that differ per draw. The counters will follow as a comment for the owners of perf(agc): reuse cached draw stages when only buffer base addresses moved #5609 and perf(agc): relocate cached draws across user words that only reach the shader as values #6539.

GPU time per present on this run (APS5_PROFILE_GPU=1, program ranges): 0x200fae0000 20.1 ms over 8 dispatches, 0x2013ba0000 12.6 ms, 0x206fd50000 9.8 ms, 0x200e480000 6.2 ms, the skin-cache kernel 0x200fa10000 4.8 ms over 15 dispatches with #8336 (55 ms without it: its 2.5 and 3.3 MiB outputs were written in place into the host import at 4.2 ms per dispatch).

Open pull requests and local changes the title needs or needed:

• fix(agc): keep resident DCC targets watched under aligned block imports #8172 (merged 2026-10-20): before it the title stopped before gameplay at the write-watch refusal of Palworld (PPSA23126): abort at compute shader 0x15c1486e00, "storing storage image results pending over memory outside the write watch without the guest bytes saved before them is not implemented" #7017 ("storing storage image results pending over memory outside the write watch without the guest bytes saved before them is not implemented"); the closed fix(agc): save pending storage images before a partly watched import unwatches them #7687 was the local workaround. On this main the stop is gone.
• perf(agc-driver): join every range a bound buffer crosses #5372 (open, conflicting): removes the CPU copies above.
• perf(agc): stage written buffer elements up to 8 MiB on discrete GPUs #8336 (open): stages the skin-cache outputs; frame p50 308 → 171 ms and presents 196 → 261 in a 60 s window on the earlier stack.
• perf(agc): replace the indexed 16-byte copy kernel with a buffer copy #7968 (open): the indexed 16-byte copy kernel as a buffer copy; GPU time per frame roughly halved when measured, frame time unchanged.

How it was run

eboot.bin, the seven sce_module modules (libc, libSceFace, libSceFaceTracker, libSceFontGsm, libSceJobManager, libSceNpCppWebApi, libScePfs) and the title's two prx libraries (libfmod, libfmodstudio) unwrapped with tools/unwrap_self.py and relinked with relinker --windows --registry. Import audit (tools/import_audit.py --libs <built libs> --modules _relink_input/prx --modules _relink_input/sce_module over the ten registries, against the 357 libraries of this main): 3,099 unique imports, 0 absent, 49 stubs (libSceAgc, libSceAgcDriver, libSceFont-module, libSceGameLiveStreaming, libSceWebBrowserDialog, libkernel), 1,254 resolved by the title's own modules; no stub is hit in gameplay. Launched from the desktop session with APS5_PLAYGO_UNLISTED_CHUNKS=installed (chunks the title queries that its PlayGo data does not list are reported installed); input scripted (Space every 1.5 s through the loading screen, then I, Q, E, T, G, H held 300 ms with 1.2 s gaps, so every run reaches the same first level and respawns in it). Measurements with APS5_PROFILE_GPU=1 APS5_PROFILE_DRAW=1 (10 s windows) and APS5_TRACE_SECONDS=120:60 for the performance-report captures; 7 shaders are deferred at registration (program address not in the shader registers) and prepared on use.

AI-assisted: yes (Claude Code).