Killing Floor 3 (PPSA08427) on Linux

[hypermoist]

Title: Killing Floor 3, PPSA08427, 02.011.000 (Unreal Engine 5)
Host: CachyOS (kernel 7.2.9), Ryzen 5 5600G, RTX 3060 12 GB (NVIDIA 615.78.08, open kernel module), GCC 16.2.1, 31 GiB RAM

Updated 2026-10-12.

On unmodified main

• Stops at start-up: pros.sdk.Prospero.prx.guest.prx: undefined symbol: y7JD1gfwyfA (sceNpEntitlementAccessGetGameTrialsFlag). feat(libSceNpEntitlementAccess): declare sceNpEntitlementAccessGetGameTrialsFlag #3449 declares it.
• With feat(libSceNpEntitlementAccess): declare sceNpEntitlementAccessGetGameTrialsFlag #3449 every module loads, and the title stops before its first frame at a draw that exports clip distances (clip distances, layer, viewport or auxiliary vertex exports is unsupported (register 0x207 = 0x0040000f)); feat(agc): vertex clip and cull distances #423 covers clip and cull distances.

On our fork branch wip/ppsa08427-sync-fix (main of 2026-10-11 plus local changes; the install is relinked against it)

• Boots, plays the Tripwire logo movie (about 12 fps), reaches the key-art title screen and the main menu.
• Title screen and menus run at about 2-2.6 fps (20-27 presents per 10 s).
• Play Solo loads into a match: the HUD draws, the world stays black (tested on the branch before the merge with main).
• The later intro movies appear to skip: they play while the 3D menu loads in the background, the GPU queue manages 3-7 presents per 10 s for about 20 s, and the movie clock keeps running.

Host-import spans and the NVIDIA driver (found while merging main)
With the span imports from 70d7aa5 the title crashed within seconds: segfault at 160 ... in libnvidia-glcore.so, MMU faults (Xid 31) and, in the kernel log, Cannot map memory with base addr 0x3014900000 and size of 0x330 pages. The title registers a heap 1 MiB at a time; each new piece extends the run, so the whole run is retired and pinned again (0x3014900000+0x120000, +0x220000, +0x320000), and pinning the next one fails inside the driver, which then crashes instead of returning an error. Our branch imports only the ranges a request needs (from the range containing it to its end, growing an existing import from its base), and hands a retired import to the recorder even when it is idle, since a draw being built can still hold its buffer from an earlier lookup. With both the title runs as before.

Play Solo crash and its workaround
Without a config change the game crashes with SIGSEGV about 80 s into the level load (eboot vaddr 0x3b6227, a null read inside UE garbage collection). UE's shipping build gives up after 10 s without progress on a pending FinishDestroy (a primitive component's render-thread fence), and its error path crashes. Our render thread is that slow during the load. With this in download0/nightfall/saved/ps5/engine.ini (and the same pair as -ini:Engine:[ConsoleVariables]:... on app0/uecommandline.txt) the match loads:

[ConsoleVariables]
gc.AdditionalFinishDestroyTimeGC=600
gc.EnableTimeoutOnPendingDestroyedObjectInShipping=0

What limits it (measured with APS5_PROFILE_DRAW, APS5_PROFILE_GPU and perf)

• One driver thread processes graphics queue 0 and is busy the whole time; the async compute queue spends about 60% of its time waiting on labels from it, and the game's RHI thread waits on end-of-pipe events from both. Each compute dispatch costs about 0.4 ms of driver time and each indirect one about 1.4 ms; about 90% of that is resource setup, not capture or recompilation.
• GPU time is about 110-130 ms per frame on the title screen. The largest part is storage-image traffic. The title aliases transient render targets: at one address a 2560x1440 RGBA16F target, a 2560x1440 8-bit target and an 80x67 target take turns, and small buffers (4-byte counters, 4 KiB constant blocks) are placed in the same memory. Each refresh of one image stores the others' pending results to guest memory and uploads the surface again, about 400 MiB per frame. With APS5_UNIT_SHADOW_MIB=2048 (and a 3 GiB storage cache to fit VRAM) every write-back gets a unit shadow and about 5 GiB per 10 s of those uploads come from the device shadow instead of the host import, but the frame rate does not change, so at the menu the CPU side is the limit for now.
• Recompiled SPIR-V is large: 40 to 930 words per guest word (a 68-word copy kernel becomes 63k words, mostly unreachable blocks and duplicate constants), which mostly costs pipeline creation time.
• In the match, about 8,000 draws and dispatches per run fail vkAllocateMemory (out of device memory) on 12 GB.

Environment used
APS5_SKIP_FAILED_WORK=1 (local debug aid on our branch; without it an unsupported mesh draw, "mesh draw contains no complete primitive", aborts the game), APS5_STORAGE_CACHE_MIB=4096 (2048 thrashes: 26 GiB of re-uploads per 10 s), APS5_HOST_IMPORT_MIB=8192, APS5_STAGING_POOL_MIB=512, APS5_BATCH_CAP=4, APS5_RELAXED_FLOAT=1, APS5_SAMPLER_FALLBACK=1.

Sound not checked. The investigation was done with an AI assistant (Claude); the measurements are from runs on the machine above.