Ghostrunner (PPSA03685)

[frankea]

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

Title

Ghostrunner (PPSA03685)

Content version tested

01.000.005

AnyPS5 commit

main 922dc89 (2026-10-17), no local changes.

System tested

Windows 11 (26200), GeForce RTX 4090 (NVIDIA 616.92, Vulkan 1.4.341, subgroup size 32), AMD Ryzen 7 5800X, 32 GB RAM, MinGW-w64 GCC 15.2.0 r7 (UCRT), Release libs.

Furthest point

Logos or loading screen

What is shown

The 3D Realms logo and the opening cinematic, 3840x2160 presented into a 2446x1376 window at 57.9 to 59.9 FPS with no input, for about 100 s; picture only, audio and input not checked.

Current blocker

vkWaitForFences present: Vulkan result -4 (VK_ERROR_DEVICE_LOST) about 100 s into the cinematic, after a burst of new compute dispatches; checkpoint program 0x146f38800 at stages 0x1 and 0x2000. Tracked in #5951. No open pull request known to address it; nothing local.

How it was run

eboot.bin and the six sce_module modules unwrapped with tools/unwrap_self.py, relinked with relinker --windows --registry; app.exe from the desktop session with ANYPS5_REQUIRE_VULKAN=1 and the PlayGo answer, no controller, no input, 300 s cap; run 2 and 3 also with APS5_GPU_CHECKPOINTS=1 APS5_PROFILE_DRAW=1 APS5_DUMP_COMPUTE_PIPELINES=1.

---

(Original report of 2026-10-17 follows unchanged.)

Ghostrunner (PPSA03685, 01.000.005, retail dump, owned, Unreal Engine 4) on Windows / NVIDIA, stock main of today. Short version: the title boots with no input at all, plays its intro cinematic at a locked 58 to 60 FPS at 3840x2160, and then loses the Vulkan device in the same place and with the same checkpoint program as the RTX 5070 and RTX 4080 SUPER runs in #5951. The 127,329-word compute pipeline from that issue is created at start-up in tens of milliseconds here, and a compile-only replay of its capture compiles it in 2.1 s with a fresh application cache and in 1.5 ms with the exported cache imported, exit 0 with SPIRV-Tools validation on, so on this card the pipeline's creation is not what loses the device.

Setup. Windows 11 (26200), GeForce RTX 4090 (NVIDIA 616.92, Vulkan 1.4.341, subgroup size 32), AMD Ryzen 7 5800X, 32 GB RAM, MinGW-w64 GCC 15.2.0 r7 (UCRT), Release libs. Tree: main 922dc89 (2026-10-17, includes #6236), no local changes. eboot.bin and the six sce_module modules (libc, libSceFace, libSceFaceTracker, libSceJobManager, libSceNpCppWebApi, libScePfs) unwrapped with tools/unwrap_self.py, relinked with relinker --windows --registry (no --to-intel, the host is AMD; no --module-dir, the modules are in sce_module). Import audit (tools/import_audit.py --modules sce_module over the seven registries): 2,502 unique imports, 0 absent, 30 stubs (libSceAjm, libSceAudioIn, libSceGameLiveStreaming, libSceVoiceQoS, libSceWebBrowserDialog and a few libSceAgc/libkernel entries), 1,149 resolved by the title's own modules. The dump has no sce_sys/playgo-chunk.dat and no playgo-scenario.json, so with #6236 on main the title needs APS5_PLAYGO_UNLISTED_CHUNKS=absent (without it scePlayGo stops at the first chunk query, as @xXPBSXx reported on #5951); all runs below set it. Launch: app.exe from the desktop session with ANYPS5_REQUIRE_VULKAN=1, no controller, nobody at the keyboard, 900 s cap. Run 1 cold shader cache and no other variables; runs 2 and 3 warm, with APS5_GPU_CHECKPOINTS=1 APS5_PROFILE_DRAW=1 APS5_DUMP_COMPUTE_PIPELINES=<empty dir> APS5_DUMP_COMPUTE_MIN_WORDS=120000, i.e. what @MarcZX100 asked for on #5951.

Result. 3/3 the same device loss. The title needs no input to get there: it goes from the 3D Realms logo into the intro cinematic by itself (the opening cutscene with the protagonist), renders it at 3840x2160 presented into a 2446x1376 window at 57.9 to 59.9 FPS (596 presents per 10 s, 0 dropped), and after roughly 100 s of cinematic a burst of new compute dispatches starts (230 in 10 s, 0 cache hits, 35 compute pipelines created) and the device is lost:

[device-fault] checkpoint: program 0x146f38800 at stage 0x1
[device-fault] checkpoint: program 0x146f38800 at stage 0x2000
[device-fault] result 5:
[device-fault] instruction pointer unknown at 0x300112580 (precision 0x10)
... (12 addresses in 0x300112580..0x30012bdb0; the cold run printed 57 in 0x300124180..0x3001344c0)
[gpu] driver failure: vkWaitForFences present: Vulkan result -4
[videoout] presentation failed: vkWaitForFences present: Vulkan result -4
terminate called after throwing an instance of 'std::runtime_error'
  what():  vkWaitForFences present: Vulkan result -4
FATAL: abort() (SIGABRT)

run	shader cache	variables	launch to exit	frames at the last capture	stop
1	cold (2,759 compiled in ~60 s)	PlayGo answer only	3 min 12 s (fault at t=144 s)	n/a (console captured)	57 fault addresses, first failure vkQueueSubmit recorder: Vulkan result -4, then the present wait
2	warm (2,757 hits)	+ checkpoints, profile, pipeline dump	1 min 57 s	5,892 (57.9 FPS)	checkpoint 0x146f38800 at 0x1 and 0x2000, then vkWaitForFences present: Vulkan result -4
3	warm (2,757 hits)	same as 2	1 min 58 s	5,880 (57.9 FPS)	same as 2 (149 fault addresses in 0x300112540..0x300123570)

The 127,329-word pipeline. With the warm shader cache the driver creates it at start-up, before the window exists, when it replays its pipeline journal: [dispatch] created a pipeline for 127329 SPIR-V words in 31 ms (program 0x0) in run 2 and in 21 ms in run 3, with the 19 MB vk-pipelines driver cache loaded. The title then runs for almost two minutes before the loss, so pipeline creation and the device loss are separated by the whole cinematic here. The capture (compute-f3439387e2a8dbd5/capture.aps5cp, 509,563 bytes, written by APS5_DUMP_COMPUTE_MIN_WORDS=120000, the only module above the threshold) replayed with agc_compute_pipeline_replay built from the same tree with -DANYPS5_ENABLE_SPIRV_TOOLS=ON, compile only, no dispatch: pass a, fresh application cache: pipeline_create 2,077 ms (replay_total 2,151 ms), exported cache 314,091 bytes, exit 0; pass b, that cache imported: pipeline_create 1.5 ms, exit 0; a second fresh-cache pass in a new process right after (the run 3 copy of the same capture, same identity f3439387e2a8dbd5) took 12.0 ms, so the 2.1 s of the first pass was the one real compile and the NVIDIA driver cache served the rest. required_subgroup_size 32, pipeline_flags 16, 1 descriptor set, 1 push-constant range, 0 specialization constants, capture and replay device NVIDIA GeForce RTX 4090, driver 2585198592, pipeline cache UUID 234d00caf7fd8b6cadcf90d3a502d334. So the pipeline compiles and validates on this card; what loses the device is in the dispatch or synchronisation path of the burst that follows, and the checkpoint reaching both stage markers for 0x146f38800 says it is not necessarily that dispatch (as @MarcZX100 noted). No [dispatch] <phase> took line over 1 s appears in either profiled run.

Other lines seen before the stop, none of them a stop: 304 shaders deferred at registration (s_swappc_b64 calls a computed target), 69 unsupported registered shader type 8 (logged and skipped since #1892, listed in #6675), sceSysmoduleLoadModule unknown id: 295 once per run, Project file not found: ../../../Ghostrunner/Ghostrunner.uproject from the title's own boot log (it reads ue4commandline.txt and carries on). No stub is hit. Device-local memory at the stop: about 300 MiB of the 23,370 MiB budget.

Next from here: the dispatch burst is where to look. APS5_PROFILE_DRAW prints the dispatch phases but not which programs the 230 dispatches were; if a switch exists (or a branch appears) that logs the compute programs of the last submission before VK_ERROR_DEVICE_LOST, or that skips/serialises the dispatch of 0x146f38800, I can run it on this card; the layout and warm caches are in place.

Trimmed stderr of run 2 (profiled) and the fault block of the cold run: #7320 (comment)

AI-assisted: yes (Claude Code).

[frankea]

Run 2 (warm shader cache, APS5_GPU_CHECKPOINTS=1 APS5_PROFILE_DRAW=1 APS5_DUMP_COMPUTE_PIPELINES APS5_DUMP_COMPUTE_MIN_WORDS=120000 APS5_PLAYGO_UNLISTED_CHUNKS=absent), main 922dc89, RTX 4090 / 616.92. Launch 19:01:19, window 19:01:24, abort 19:03:16 (box time). The 304 could not be prepared at registration and 69 unsupported registered shader type 8 lines, the per-10 s profile blocks and the pm4 ring dump are cut; the blocks right before the fault are kept.

Run 2 stderr, start-up to the window (trimmed)

[time] scheduler tick raised to 0.5 ms (was 0.50 ms, now 0.50 ms, timed waits fine)
[memory] configured flexible memory: 0x1c000000 bytes (448 MiB)
[guestmem] read 0 calls 0 MiB 0.0 s compare 0 calls 0 MiB 0.0 s write 0 calls 0 MiB 0.0 s write-changed 0 calls 0 MiB 0.0 s collect 0 calls 0 MiB 0.0 s verify 0 calls 0 MiB 0.0 s query 1 calls 0 MiB 0.0 s collect-memo hits 0, collect epochs 0 collect-dirty 0 tracker waits 0 / 1 | arena 0x200000000+0xfa00000000 image 0x140000000+0x8bde000 forgets 0 (0 MiB) | read callers: | write callers: | read si
[write-watch] host imports of write-watched memory keep the write watch: yes (4096 of 4096 scratch pages reported written at the import, 0 by its 4-byte settle read, 0 after it, 0 after a full GPU read, 0 after a second, 0 after a GPU write, 1 after a CPU store); direct host imports of private memory are settled and stay watched, separate shared aliases retain guest write tracking
[shader-disk-cache] C:\src\ghostrunner\run\shader_cache\730242ea3a26f142 (source version 730242ea3a26f142, format 36)
...
[write-watch] host imports of write-watched memory keep the write watch: yes (4096 of 4096 scratch pages reported written at the import, 0 by its 4-byte settle read, 0 after it, 0 after a full GPU read, 0 after a second, 0 after a GPU write, 1 after a CPU store); direct host imports of private memory are settled and stay watched, separate shared aliases retain guest write tracking
[pipeline-cache] loaded 19250.0 KiB from C:\src\ghostrunner\run\shader_cache\vk-pipelines-10de-2684-9a170000-234d00caf7fd8b6cadcf90d3a502d334.bin
[pipeline-journal] loaded 39 pipelines and 42 modules (2.7 MiB) from C:\src\ghostrunner\run\shader_cache\730242ea3a26f142\pipelines.journal in 2 ms
[compute-capture] C:\src\ghostrunner\pipelines-main-2\compute-f3439387e2a8dbd5\capture.aps5cp
[dispatch] creating a pipeline for 127329 SPIR-V words (program 0x0)
[dispatch] created a pipeline for 127329 SPIR-V words in 31 ms (program 0x0)
[pipeline-journal] loaded 39 pipelines, 42 modules (2.7 MiB, 2 ms); replay on 8 threads: 4 graphics and 35 compute built, 0 present, 0 skipped (store full), 0 rejected, 0 failed in 57 ms; recorded this run: 0 graphics, 0 compute, 0 modules (0.0 MiB), 0 write failures
[capture] 1 captures: 0 word reads, 0 pages fetched (0 read word by word over pending GPU writes: 0 words waited 0.00 s, 0 read raw on evidence, 0 verified with 0 mismatches, 0 served from known values, 0 hook reads not observed: no GPU wait), capture (plan lookup + materialize) 0.0 s, 0 KiB registered code this capture; sub-phases (s, us per capture): source resolve 0.00/0.0, walk 0.00/38.9, spec
[presync] 1 dispatches checked, 0 pre-syncs waited 0 ms (cumulative)
[buffers] descriptor elements bound: 1 total, 0 written, 1 read-only; 1 pending-write notes skipped
[mouse] video driver windows: guest 3840x2160 presented at 2446x1376, delta scale 1.5699 x 1.5698
[textures] sampled created: 0 from storage images, 0 snapshots (0 snapshot reads over recorded writes inside cachedTexture, 0 storage-path fallbacks); stage-A records 0: 0 fast hits, 0 full lookups; storage images 0 hits, 13 created, 0 own-object refreshes; image tables 0 reused, 0 resolved; untracked snapshots 0 kept by the write-watch hint, 0 compared
[gpu] sampled texture cache budget 14152 MiB: device-local heap 0 has a VK_EXT_memory_budget budget of 23370 MiB and uses 294 MiB, 0 MiB of it sampled textures (0 MiB in the cache) and 32 MiB cached storage images; storage image cache budget 5842 MiB
[resources] images took 69 ms (1 textures, 0 storage images, 3 buffers, bda 0)
[rescache] revalidate 1 calls 0.6 ms: fast 0, full 1, failed 0; fast-fail by reason: collect 1; full walks: T1 refreshed 0 (0 storage images, 0 view sources, 0 views served by their own pending source, 0 overlaps left by the refresh accepted), full by reason: collect 1, T1 fallback by reason:, proofs verified 0; fast-fail pending image: own 0 (the query excepts the own object), foreign 0, snapshot
[texture] 50 storage images (102 MiB): create 50 ms, write-back 1 ms (alloc 0, host copy 0, gpu 4, store 0), 186 reused, 76 direct uploads, 14 direct write-backs, clears 4 recorded / 0 waited

Run 2 stderr, the last 10 s block and the device fault

[texture] 5400 textures (66 copied from storage images, 5334 uploads recorded): allocate+image 1408 ms, guest read 1 ms, detile+copy 128 ms, view+buffer release 29 ms
[textures] sampled created: 66 from storage images, 5344 snapshots (0 snapshot reads over recorded writes inside cachedTexture, 0 storage-path fallbacks); stage-A records 5797: 408 fast hits, 5389 full lookups; storage images 20904 hits, 95 created, 0 own-object refreshes; image tables 0 reused, 0 resolved; untracked snapshots 2849 kept by the write-watch hint, 5501 compared
[dcc] 11997 scans, 174.5 MiB scanned, 0 memo hits, 0 flush syncs, 13.3 ms; uncompressed keys stored: 29 on the GPU (0 with an unaligned head or tail), 0 on the CPU; key proofs: 0 proved, 8892 scanned, 8892 unstable
[rescache] revalidate 14648 calls 2448.6 ms: fast 5894, full 5882, failed 2872; fast-fail by reason: collect 8754; full walks: T1 refreshed 0 (0 storage images, 0 view sources, 0 views served by their own pending source, 0 overlaps left by the refresh accepted), full by reason: collect 8754, T1 fallback by reason:, proofs verified 0; fast-fail pending image: own 0 (the query excepts the own object), foreign 0, snapshot 0, T1 eligible 0; keys proven: 5857 sampled, 0 storage; epoch gate: 5852 registry scans skipped, 11513 import loops skipped
[texture-cache] (10 s): 494 textures cached, 0 used uncached (first use this frame, the least recently used entry used this frame or the last), 0 evicted, 0 held ones passed over, 494 found changed; 30 MiB in 117 entries, budget 13913 MiB, frame 5867
[resources] 3000 builds, phase totals: bindings=35ms precollect=24ms guest memory upload=80ms descriptors=34ms stage A=101ms images=1771ms bda=1ms stage B=1862ms
[gpu] worker at 99 s: dispatch 0.0 s, draw 0.0 s, wait 0.0 s
[gpu] worker at 105 s: dispatch 0.0 s, draw 0.0 s, wait 0.0 s
[rescache] miss churn (dispatch), first differing word by role: GuestBuffers 155 (V# word 155/0/0/0); layout 0, draw target/index 0, same key 5 (not inserted or evicted), first seen 33
[rescache] miss churn (draws), first differing word by role:; layout 60, draw target/index 4, same key 0 (not inserted or evicted), first seen 3
[presync] 230 dispatches checked, 0 pre-syncs waited 0 ms (cumulative)
[rescache] 37 hits, 193 misses, 0 invalidated, 160 entries (dispatch + draw); template hits: 0 refreshed the data buffers (revalidate 0.0 ms), 37 had the same words; Find calls 15542, Touch calls 0
[vk] deviceProc lookups inside the device call: 8403 (36.53 per dispatch); pre-dispatch barriers recorded 37, skipped 192; compute pipelines created 35 (0 under the GPU mutex)
[recipe] dispatch (10 s): hits 0; pre-check misses: no recipe 0, device 0, template gone 0, objects gone 0, not recordable 0; proofs: fast 0, T1 refreshed 0, full 0; rebuilds: imports 0, evicted 0, pending 0, changed 0, keys 0, other 0, device 0; presyncs from the recipe 0 (recomputed 0, waited 0); data refreshes recorded 0 / skipped by hash 0 (0 decided by words: data hits); restarts 0, attaches 116; us per hit: pre-check 0.0, proof 0.0, record 0.0; verified 0 (object replaced 0)
[recipe] indirect (10 s): hits 0; pre-check misses: no recipe 0, device 0, template gone 0, objects gone 0, not recordable 0; proofs: fast 0, T1 refreshed 0, full 0; rebuilds: imports 0, evicted 0, pending 0, changed 0, keys 0, other 0, device 0; presyncs from the recipe 0 (recomputed 0, waited 0); data refreshes recorded 0 / skipped by hash 0 (0 decided by words: data hits); restarts 0, attaches 0; us per hit: pre-check 0.0, proof 0.0, record 0.0; verified 0 (object replaced 0)
[recipe] draw (10 s): hits 0; pre-check misses: no recipe 2466, device 0, template gone 0, targets gone 0, not recordable 0; proofs: fast 0, T1 refreshed 0, full 0; rebuilds: imports 0, evicted 0, pending 0, changed 0, keys 0, other 0, device 0; presyncs from the recipe 0 (recomputed 0, waited 0); data refreshes recorded 0 / skipped by hash 0 (0 decided by words: data hits); restarts 0, attaches 94; us per hit: pre-check 0.0, proof 0.0, record 0.0; verified 0 (object replaced 0)
[gpu] worker at 105 s: dispatch 0.7 s, draw 0.0 s, wait 0.0 s
[dispatch] driver phases queue 0 direct (10 s, 230 dispatches, 0 cache hits), us per dispatch: prologue 102.5 key 0.2 lookup 0.3 validate 0.0 validate GPU wait 0.0 relock 0.0 capture 206.0 recompile 678.4 insert 1.2 queued-label check 0.2 snapshots 0.2 prepare: key 1.2 prepare: find 1.1 prepare: precollect 0.0 prepare: presync 1.1 prepare: stage A (miss) 109.1 prepare: other 3.6 recipe pre-check 0.0 lock wait 0.4 labels 0.1 note writers 0.3 device call 1890.9 tail 0.1 epilogue 1.4, total 2998.4 (689.6 ms); validate 0.0 us per validation (0 validations; GPU waits inside it 0.0 us per validation
[buffers] staging: 219 regions staged device-local (49 with atomics, 27 of reused builds), 73189 KiB copied in, 72927 KiB copied back, 0 without copy-back, 0 shadows refused; last 10 s: 217 regions, 72927 KiB in, 72927 KiB back
[buffers] image mirrors: 1 ranges (0.2 MiB, 0 writable, 0 heap 0.0 MiB), 6 address-based builds served, 0 descriptor sub-ranges bound, 0 sparse heaps bound, 1 rebuilds, 0 refreshes: 0 blocks compared, 0 copied, 0 refresh syncs; heap refills 0, unwatched heaps 0, 0 heap mirrors released for imports (0.0 MiB)
[address] 6 address-based builds (10 s), us per build: lease 134, imports pass 30646, mirror prepare 483, compare 1 (0 blocks, 0.0 copied), snapshots 3; space hits 1 / rebuilds: generation 2, epoch 2, device 0, waiter drop 0, first 1; unpublished 0, dissolved: overlap 2, imports 1; [bda-table] hits 0 / misses 5 (space tables 0), first entry: expired 3, hash differs low 0 / heap 0, same hash 0, none held 2
[address-sync] leases: 6 released at completion, 0 synced at once; 0 pin-contention waits by guest threads (0 finished the lease batch, 0 drained the recorder, 0 cache-only drops) 0.0 s; BDA table cache 0 hits / 6 misses (1 tables held); snapshot compares: 0 skipped (live-backed region), 48 made
[device-fault] checkpoint: program 0x146f38800 at stage 0x1
[device-fault] checkpoint: program 0x146f38800 at stage 0x2000
[device-fault] result 5: 
[device-fault] instruction pointer unknown at 0x300122680 (precision 0x10)
[device-fault] instruction pointer unknown at 0x30012bdb0 (precision 0x10)
[device-fault] instruction pointer unknown at 0x300112580 (precision 0x10)
[device-fault] instruction pointer unknown at 0x300113b20 (precision 0x10)
[device-fault] instruction pointer unknown at 0x3001234d0 (precision 0x10)
[device-fault] instruction pointer unknown at 0x300112ea0 (precision 0x10)
[device-fault] instruction pointer unknown at 0x300113d50 (precision 0x10)
[device-fault] instruction pointer unknown at 0x3001147a0 (precision 0x10)
[device-fault] instruction pointer unknown at 0x3001127b0 (precision 0x10)
[device-fault] instruction pointer unknown at 0x300120470 (precision 0x10)
[device-fault] instruction pointer unknown at 0x300113ad0 (precision 0x10)
[device-fault] instruction pointer unknown at 0x300113a70 (precision 0x10)
[gpu] driver failure: vkWaitForFences present: Vulkan result -4
[videoout] presentation failed: vkWaitForFences present: Vulkan result -4
[resources] 3207 builds, phase totals: bindings=38ms precollect=28ms guest memory upload=224ms descriptors=35ms stage A=127ms images=1872ms bda=2ms stage B=2097ms
terminate called after throwing an instance of 'std::runtime_error'
  what():  vkWaitForFences present: Vulkan result -4
FATAL: abort() (SIGABRT) on thread 6648
    #0 0x00007fff9aad2bd8 libc.prx+0x12bd8
    #1 0x00007fff9aba5af1 libc.prx+0xe5af1
... (SIGABRT thread dump cut)

Run 1 (cold shader cache, no profile variables): the fault block

[log] 10.0 s: 700 lines (700 stdout, 0 stderr), 81.4 KiB written, 0 repeats coalesced; callers spent 2.28 ms queueing and formatting (0.023% of one thread), logger thread 106.90 ms, queue peak 0.3 KiB
[log] 10.0 s: 702 lines (702 stdout, 0 stderr), 81.6 KiB written, 0 repeats coalesced; callers spent 2.24 ms queueing and formatting (0.022% of one thread), logger thread 109.67 ms, queue peak 0.3 KiB
[pipeline-cache] keeping C:\src\ghostrunner\run\shader_cache\vk-pipelines-10de-2684-9a170000-234d00caf7fd8b6cadcf90d3a502d334.bin: the merged cache has 17132.3 KiB, less than the 19250.0 KiB on disk or loaded at start
[log] 10.2 s: 332 lines (332 stdout, 0 stderr), 38.7 KiB written, 0 repeats coalesced; callers spent 0.97 ms queueing and formatting (0.009% of one thread), logger thread 49.00 ms, queue peak 0.3 KiB
[device-fault] result 5: 
[device-fault] instruction pointer unknown at 0x300133a70 (precision 0x10)
[device-fault] instruction pointer unknown at 0x3001263e0 (precision 0x10)
[device-fault] instruction pointer unknown at 0x300128840 (precision 0x10)
[device-fault] instruction pointer unknown at 0x300125f40 (precision 0x10)
[device-fault] instruction pointer unknown at 0x300132940 (precision 0x10)
[device-fault] instruction pointer unknown at 0x3001275b0 (precision 0x10)
... (57 addresses, 0x300124180 to 0x3001344c0)
[gpu] driver failure: AGC graphics: vkQueueSubmit recorder: Vulkan result -4
[shader-disk-cache] (10 s): 0 hits, 2 misses, 2 writes (0.0 MiB), 0 load failures, 0 write failures; resource plans: 0 hits, 1 misses, 1 writes, 0 load failures; 0.0 MiB read in 0.3 ms (totals: 0 hits, 2759 misses, 2759 writes; plans 0 hits, 2485 misses)
[videoout] presentation failed: vkWaitForFences present: Vulkan result -4
terminate called after throwing an instance of 'std::runtime_error'
  what():  vkWaitForFences present: Vulkan result -4
FATAL: abort() (SIGABRT) on thread 23528
    #0 0x00007fff9aad2bd8 libc.prx+0x12bd8
    #1 0x00007fff9aba5af1 libc.prx+0xe5af1
    #2 0x00000001ab17f910

Run 3: same as run 2

[dispatch] created a pipeline for 127329 SPIR-V words in 21 ms (program 0x0)
[gpu] worker at 105 s: dispatch 0.7 s, draw 0.0 s, wait 0.0 s
[dispatch] driver phases queue 0 direct (10 s, 230 dispatches, 0 cache hits), us per dispatch: prologue 99.9 key 0.2 lookup 0.3 validate 0.0 validate GPU wait 0.0 relock 0.0 capture 192.9 recompile 687.4 insert 1.2 queued-label check 0.2 snapshots 0.3 prepare: key 1.3 prepare: find 1.0 prepare: prec
[buffers] staging: 217 regions staged device-local (49 with atomics, 27 of reused builds), 72832 KiB copied in, 72571 KiB copied back, 0 without copy-back, 0 shadows refused; last 10 s: 215 regions, 72571 KiB in, 72571 KiB back
[buffers] image mirrors: 1 ranges (0.2 MiB, 0 writable, 0 heap 0.0 MiB), 6 address-based builds served, 0 descriptor sub-ranges bound, 0 sparse heaps bound, 1 rebuilds, 0 refreshes: 0 blocks compared, 0 copied, 0 refresh syncs; heap refills 0, unwatched heaps 0, 0 heap mirrors released for imports (
[address] 6 address-based builds (10 s), us per build: lease 157, imports pass 31206, mirror prepare 705, compare 1 (0 blocks, 0.0 copied), snapshots 3; space hits 1 / rebuilds: generation 2, epoch 2, device 0, waiter drop 0, first 1; unpublished 0, dissolved: overlap 3, imports 1; [bda-table] hits 
[address-sync] leases: 6 released at completion, 0 synced at once; 0 pin-contention waits by guest threads (0 finished the lease batch, 0 drained the recorder, 0 cache-only drops) 0.0 s; BDA table cache 0 hits / 6 misses (1 tables held); snapshot compares: 0 skipped (live-backed region), 48 made
[device-fault] checkpoint: program 0x146f38800 at stage 0x1
[device-fault] checkpoint: program 0x146f38800 at stage 0x2000
[device-fault] result 5: 
... (149 addresses, 0x300112540 to 0x300123570)
[gpu] driver failure: vkWaitForFences present: Vulkan result -4
[videoout] presentation failed: vkWaitForFences present: Vulkan result -4
[resources] 2207 builds, phase totals: bindings=26ms precollect=20ms guest memory upload=200ms descriptors=23ms stage A=96ms images=1163ms bda=1ms stage B=1354ms
terminate called after throwing an instance of 'std::runtime_error'
  what():  vkWaitForFences present: Vulkan result -4
FATAL: abort() (SIGABRT) on thread 3448

Compile-only replay of compute-f3439387e2a8dbd5/capture.aps5cp (509,563 bytes, identical in runs 2 and 3)

START Sat 10/17/2026 19:11:27.46 dir=C:\src\ghostrunner\pipelines-main-2 
C:\src\ghostrunner\pipelines-main-2\compute-f3439387e2a8dbd5
C:\src\ghostrunner\pipelines-main-2\compute-f3439387e2a8dbd5\capture.aps5cp
=== capture 1: C:\src\ghostrunner\pipelines-main-2\compute-f3439387e2a8dbd5\capture.aps5cp (509563 bytes) 
--- pass a (fresh application cache, driver cache uncontrolled) 19:11:27.48 
{
  "schema_version":1,
  "source":"C:\\src\\ghostrunner\\pipelines-main-2\\compute-f3439387e2a8dbd5\\capture.aps5cp",
  "capture_identity":"f3439387e2a8dbd5",
  "capture_identity_algorithm":"fnv1a64",
  "spirv_words":127329,
  "descriptor_sets":1,
  "push_constant_ranges":1,
  "specialization_constants":0,
  "required_subgroup_size":32,
  "pipeline_flags":16,
  "stage_flags":0,
  "captured_device":{"name":"NVIDIA GeForce RTX 4090","vendor_id":4318,"device_id":9860,"driver_version_raw":2585198592,"api_version_raw":4211039,"pipeline_cache_uuid":"234d00caf7fd8b6cadcf90d3a502d334"},
  "replay_device":{"name":"NVIDIA GeForce RTX 4090","vendor_id":4318,"device_id":9860,"driver_version_raw":2585198592,"api_version_raw":4211039,"pipeline_cache_uuid":"234d00caf7fd8b6cadcf90d3a502d334"},
  "application_cache_policy":"fresh",
  "cache_import_path":"",
  "cache_import_bytes":0,
  "cache_export_path":"C:\\src\\ghostrunner\\runs\\main-2-replay-1.cache",
  "cache_export_bytes":314091,
  "driver_cache_state":"uncontrolled",
  "dispatch_count":0,
  "pipeline_statistics":false,
  "milliseconds":{"shader_module":0.2182,"layout":9.8893,"pipeline_create":2077.1659,"cache":0.0764,"replay_total":2151.1613}
}
exit=0 19:11:29.92 
--- pass b (imported application cache, warm driver cache) 19:11:29.93 
{
  "schema_version":1,
  "source":"C:\\src\\ghostrunner\\pipelines-main-2\\compute-f3439387e2a8dbd5\\capture.aps5cp",
  "capture_identity":"f3439387e2a8dbd5",
  "capture_identity_algorithm":"fnv1a64",
  "spirv_words":127329,
  "descriptor_sets":1,
  "push_constant_ranges":1,
  "specialization_constants":0,
  "required_subgroup_size":32,
  "pipeline_flags":16,
  "stage_flags":0,
  "captured_device":{"name":"NVIDIA GeForce RTX 4090","vendor_id":4318,"device_id":9860,"driver_version_raw":2585198592,"api_version_raw":4211039,"pipeline_cache_uuid":"234d00caf7fd8b6cadcf90d3a502d334"},
  "replay_device":{"name":"NVIDIA GeForce RTX 4090","vendor_id":4318,"device_id":9860,"driver_version_raw":2585198592,"api_version_raw":4211039,"pipeline_cache_uuid":"234d00caf7fd8b6cadcf90d3a502d334"},
  "application_cache_policy":"imported",
  "cache_import_path":"C:\\src\\ghostrunner\\runs\\main-2-replay-1.cache",
  "cache_import_bytes":314091,
  "cache_export_path":"",
  "cache_export_bytes":0,
  "driver_cache_state":"uncontrolled",
  "dispatch_count":0,
  "pipeline_statistics":false,
  "milliseconds":{"shader_module":0.221,"layout":0.9313,"pipeline_create":1.4538,"cache":0.1199,"replay_total":64.9506}
}
exit=0 19:11:30.23 

run 3 capture (same bytes, run right after):
  "milliseconds":{"shader_module":0.2203,"layout":0.7291,"pipeline_create":12.0115,"cache":0.1294,"replay_total":78.9641}
exit=0 19:11:30.61 
  "milliseconds":{"shader_module":0.2203,"layout":0.7291,"pipeline_create":12.0115,"cache":0.1294,"replay_total":78.9641}
  "milliseconds":{"shader_module":0.2182,"layout":0.7185,"pipeline_create":1.486,"cache":0.1345,"replay_total":66.9779}
exit=0 19:11:30.92 
  "milliseconds":{"shader_module":0.2182,"layout":0.7185,"pipeline_create":1.486,"cache":0.1345,"replay_total":66.9779}

Title's own stdout (both runs): Final commandline = ../../../Ghostrunner/Ghostrunner.uproject, Project file not found: ../../../Ghostrunner/Ghostrunner.uproject, sceSysmoduleLoadModule unknown id: 295, then nothing until the abort.

[bipinkrish]

First bring-up record for Ghostrunner (PPSA03685) from an owner's dump: UE4 boot starts, the engine's memory manager exhausts the direct-memory pool and stops with a fatal. No frame is rendered yet.

Reported in the "Reporting a result" order from GAME_TESTING.md.

1. Build

• main at e7e3e9c198, targets libs relinker (Linux), no local changes.
• Two --trace-imports conversions (with APS5_TRACE_CALLS=all / =libkernel and APS5_TRACE_MEMORY=1 runs) were used to localise the failure; the boot sequence and the fatal are identical with and without tracing.

2. Title

Ghostrunner, PPSA03685. The dump's param.json carries no content-version field; contentids.json lists UP4040-PPSA03682_00-GHOSTRUNNER00000 / EP4040-PPSA03685_00-GHOSTRUNNER00000 / JP1561-PPSA04462_00-GHOSTRUNNER00000.

3. Host

AMD Ryzen 7 7735HS, Ubuntu 24.04.4 LTS, kernel 7.0.0-34-generic, AC power. GPU not selected — the crash happens before AnyPS5's device selection (no Selected GPU name= line); host GPU AMD Radeon 680M (RADV REMBRANDT, mesa 25.2.8).

4. Command and environment

Conversion: eboot + the six sce_module modules (libc, libSceFace, libSceFaceTracker, libSceJobManager, libSceNpCppWebApi, libScePfs) all relink cleanly. No application-bundled modules outside sce_module/ (no other .prx in the dump, no %s/Modules/-style dlopen paths in the eboot).

relinker --registry --sce-module-path <parent-of-sce_module> <unwrapped-eboot> app.elf
cd <folder> && APS5_PROFILE_DRAW=1 APS5_PLAYGO_UNLISTED_CHUNKS=installed timeout 60 ./app.elf > run.log 2>&1

• APS5_PLAYGO_UNLISTED_CHUNKS=installed is required: this dump has neither sce_sys/playgo-chunk.dat nor playgo-scenario.json, so the first run without it stops with "chunk ID 1 cannot be checked". With it the title proceeds.
• No other APS5_* variables set. ASLR not disabled. /dev/udmabuf open Permission denied → copies fallback (as in the Silt report; write-watch probe unknown, imported ranges compared).

5. How far it got

1. UE4 boots: reads /app0/ue4commandline.txt, final command line ../../../Ghostrunner/Ghostrunner.uproject exactly as the dump ships it. The .uproject does not exist in the packaged dump; the engine logs Project file not found and continues (opening /app0/.. once).
2. The engine's memory manager then allocates the entire 13.5 GiB direct-memory pool in ~20 libkernel calls (halving pattern 6.75 GiB → 3.38 → 1.69 → 0.84 → 0.42 → …, then map direct/remap fixed clearing, APS5_TRACE_MEMORY=1 traced all of it).
3. It then aborts before GPU selection, asking for 5,565,054,976 bytes with 64 KiB alignment that nothing can answer:

PS5 Out of Memory:
    AvailablePhysical: 4902879232
    UsedPhysical:      9277997056

LowLevelFatalError [File:Unknown] [Line: 186]
Ran out of memory allocating 5565054976 bytes with alignment 65536

Full final run log (43 lines)

[src/HostCpu.cpp:193 operator()] host CPU AMD Ryzen 7 7735HS with Radeon Graphics: has SSE, SSE2, SSE3, SSSE3, SSE4.1, SSE4.2, POPCNT, LZCNT, BMI1, BMI2, MOVBE, CMPXCHG16B, LAHF/SAHF, AVX, AVX2, FMA, F16C, AES, PCLMULQDQ, RDRAND, RDSEED, ADX, SHA, CLFLUSHOPT, CLWB, RDPID, XSAVEOPT, XSAVEC, SSE4A, PREFETCHW, CLZERO, OS AVX state (XCR0); lacks none
Memory va range 1000000000 - 8fc0000000
Frame Buffer va range 8fc0000000 - 9000000000
FBootAllocator used 2192936 out of 2195456 bytes of memory, wasting 2520 bytes.
Used memory before allocating anything was 22.19MB
[src/Open.cpp:266 sceKernelOpen] path=/app0/ue4commandline.txt flags=0x0 nativeFlags=0x0 mode=00
File root is /app0/
Final commandline = ../../../Ghostrunner/Ghostrunner.uproject 
[src/Open.cpp:266 sceKernelOpen] path=/app0/deepfiles flags=0x20000 nativeFlags=0x10000 mode=0555
[src/Open.cpp:284 sceKernelOpen] open failed path=/app0/deepfiles resolvedPath=/tmp/opencode/gr/out3/app0/deepfiles result=0x80020002
Project file not found: ../../../Ghostrunner/Ghostrunner.uproject
[src/Open.cpp:266 sceKernelOpen] path=/app0/.. flags=0x20000 nativeFlags=0x10000 mode=0555

PS5 Out of Memory:
    AvailablePhysical: 4902879232
    UsedPhysical:      9277997056

LowLevelFatalError [File:Unknown] [Line: 186] 
Ran out of memory allocating 5565054976 bytes with alignment 65536
LowLevelFatalError [File:Unknown] [Line: 186] 
Ran out of memory allocating 5565054976 bytes with alignment 65536
0x0000000001409128 UnknownFunction []
0x0000000001409dda UnknownFunction []
0x000000000041dfae UnknownFunction []
0x0000000001399e18 UnknownFunction []
0x0000000001370bff UnknownFunction []
0x000000000136e7d5 UnknownFunction []
0x000000000136eded UnknownFunction []
0x0000000002f6f43a UnknownFunction []
0x0000000002f6f71e UnknownFunction []
0x0000000002f6f7e8 UnknownFunction []
0x0000000000400f5f UnknownFunction []
0x0000000000401c5e UnknownFunction []
0x000000000040156f UnknownFunction []
0x0000000000413949 UnknownFunction []
0x0000000000414078 UnknownFunction []
0x0000000000414052 UnknownFunction []
0x00000000004000af UnknownFunction []
0x0000000009c859ae UnknownFunction []



timeout: the monitored command dumped core

4. Guest stack is all UnknownFunction frames (the guest binary ships no symbols) → SIGSEGV, exit 139.

What is not yet attributable: the engine's budget arithmetic — AvailablePhysical + UsedPhysical = 14,180,876,288 = our 13.5 GiB direct-memory size (DIRECT_MEMORY_SIZE = 13,824 MiB) minus 318,682,112. Neither the sysctl answers on this tree (hw.physmem/hw.realmem report host physical, 32 GiB here; hw.usermem answers the 16 KiB page size) nor the image-segment flexible charge (172.9 MiB for the eboot + six modules) accounts for that gap. Pinning the query that produced AvailablePhysical needs the import trace to survive the segfault (its abort report only fires on SIGABRT, this path dies by SIGSEGV), or a console fixture. No implementation claim is made beyond what the traces show.

6. Frames per 10 s and memory budget

None — no [present] line ever printed. The engine-side [memory] cascade reported above is the only memory data: 13.50 GiB allocated from a 13.5 GiB pool, then the 5.5 GiB single request fails.

7. Runs

Six runs today, same fatal in every one that reached the engine: PlayGo-check stop (45 s, no APS5_PLAYGO_UNLISTED_CHUNKS, kept as data), 60 s normal, 40 s traced (TRACE_CALLS=all, survived the timeout with no further output, inconclusive), 90 s traced (=libkernel, fatal), 60 s APS5_TRACE_MEMORY=1 (fatal + allocator trace), 60 s final main build (fatal). All ended by timeout kill or segfault, none by user action.

8. Date

2026-10-25.