Dreaming Sarah (PPSA02929) on Windows: boots to first render, crashes in amdvlk64 on RX 6700 XT

[poonasor]

Environment

• OS: Windows 11 Pro 23H2 (build 22631)
• AnyPS5: built from main 57f96fd, WinLibs GCC 15.2.0 r7 (15.2.0posix-14.0.0-ucrt-r7), Ninja, Release, BUILD_TESTING=ON
• Game/Executable: Dreaming Sarah (PPSA02929-app0/eboot.bin)
• GPU: AMD Radeon RX 6700 XT (Navi 22, device 0x73df, Adrenalin 25.10.45.05, driver 32.0.21045.5002 dated 2026-08-16, Vulkan 1.4.315), the only Vulkan device, auto-selected

Conversion

relinker --windows eboot.bin game.exe succeeds. 815 NID input references. The bundled sce_module/libc.prx converts as a guest module. One dumping-tool note: the dumper left decrypted ELF copies as eboot.bin.esbak/libc.prx.esbak beside the encrypted files, and the .esbak inside sce_module/ is scanned as a second guest module, failing with Ambiguous guest dependency name: libc.prx (both files carry SONAME libc.prx). Moving the .esbak files out of the module directory is enough.

Import audit (exit 0): 553 unique imports, 918 references - 367 implemented, 16 stub, 0 absent, 170 module (the title's own libc.prx). The 16 stubs are capture/diagnostics functions (sceAgcDcb*/sceAgcAcb* capture, sceAgcDriverRequestCapture*, backtrace/heap-report postfixes, __tls_get_addr postfix).

Runtime

The title boots through the whole stack: ELF entry, guest threads, save data mount (mount3 -> /savedata0), full Vulkan device init on the RX 6700 XT, shader disk cache creation, resource loading (language files, media/error.flac background load), render target allocation up to 14 MiB and texture creation (1280x720 storage images, 320x512/980x347/250x250 textures). It then crashes at the first real graphics submission:

FATAL: unhandled exception 0xc0000005 on thread 15024 ''
  rip 0x00007ffbeacdef34 amdvlk64.dll+0x21eef34
  read of 0xffffffffffffffff

The faulting offset moves between runs inside amdvlk64.dll: +0x21eef34, +0x1bbfa4, +0xab8dc1, +0x21ee870, +0x1ba6e5 (the last two recurring). Every run aborts with 0xc0000005 at the same progression point.

APS5_NO_TEXTURE_CACHE=1, APS5_NO_WRITE_WATCH=1 and APS5_STORE_AT_FLIP=1 all still crash in amdvlk64.dll at shifted offsets, so the texture cache, write tracking and store-at-flip paths are each not the single cause.

Relation to #2944

amdvlk64.dll+0x21eef34 is the exact rip of the agc_driver_vertex_fetch SEGFAULT on this machine and driver (#2944); nine agc_driver_* execution tests segfault there. The game crash and the test crashes look like the same driver failure: on a GTX 1050 Ti the title is verified playable at 60 FPS, so this is specific to the AMD Windows driver (current Adrenalin 25.10.45.05).

I can rerun anything on this machine, including other Adrenalin versions.

AI-assisted (ZCode).

[poonasor]

Update: playable on the RX 6700 XT with one change.

The crash at the first graphics submission was the background link-time-optimized pipeline link (#2713) on AMD's Windows driver, tracked in #2944 (also #2755, #2777). The fix is #3149, which was closed automatically by the PR intake pause, not on review; the commit is 8acd0113 on poonasor:fix/agc-library-module-lifetime.

Relinked and run against main 379d8b4e + 8acd0113, same machine and driver as above (RX 6700 XT, Adrenalin 25.10.45.05, Windows 11 23H2):

• Boots past the point where it faulted in amdvlk64.dll+0x21eef34 and reaches gameplay.
• Renders correctly at about 60 FPS (window title FPS: 59.67), first forest level: character, platforms, background and fog layers all drawn.
• Audio works, controls work.
• No fault in the log.

Without the fix, APS5_NO_PIPELINE_LTO=1 avoids the faulting link in the agc_ tests; I have not run the title that way.

Only the opening area was played so far.

AI-assisted (Claude Code).