Assassin's Creed Shadows (PPSA20397): boot cinematics render, stops at the stencil replay-safety gate

[akkadbakkad89]

Existing thread

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

Title

Assassin's Creed Shadows (PPSA20397)

Content version tested

01.019.076 (masterVersion 01.00), full image — all forge packs mounted.

AnyPS5 commit

main 2185dbf92 plus the feat/psml-mfsr-queries branch commits (the psml MFSR exports of #8290) and the libSceNpGameIntent app-launcher lifecycle of #8355. Windows build of that tree, all runtime libraries one vintage.

System tested

Windows 11 Pro 10.0.26300, Intel Core i7-10700, NVIDIA GeForce RTX 3070 (driver 617.42, native VK_KHR_fragment_shader_barycentric), WinLibs GCC 15.2.0 UCRT POSIX SEH. Earlier bring-up runs were on Linux/WSL2 + llvmpipe on the same host.

Furthest point

Boot cinematics — an in-engine rendered intro sequence, the Assassin's Creed logo splash and the auto-save notice, presented at ~25 FPS. The run stops before the title menu.

What is shown

The window presents real rendered frames: an in-engine cinematic (character tumble through the bright chamber), the AC logo over the sky backdrop, and the "This game saves data automatically" notice. The title bar frame counter climbs past 1,300 at ~25 FPS. About 9,000-9,300 shaders compile on a cold run; texture streaming/detiling and save-journal writes continue until the abort.

Current blocker

A presentation draw aborts the run on this GPU:

shader 0x407325e400 stage 5: ShaderRecompiler::Recompile:
stencil reference export emulation requires a replay-safe pixel shader with late tests
[videoout] presentation failed

The draw writes stencil unconditionally (front+back compare ALWAYS, all ops REPLACE, masks 0xff) at 4xMSAA with a reference exported per pixel. That reference is a texture-derived value (ImageRead), so it is unbounded: on a device without VK_EXT_shader_stencil_export the recompiler takes the all-256-value replay path, but the shader then fails StencilReplaySafe — maskStable() cannot prove the Ballot/WqmU64 helper-lane mask stable across replays — and the draw also trips the samples == 1 replay gate. Full shader/draw analysis is in #8514.

Resolved along the way: the early-boot dead-handle SIGSEGV of #8514 was a launch-subset content artifact — a forge pack absent from the partial app0/ populated the handle-table slot. With the complete image it never reproduces.

How it was run

relinker --windows --to-intel --module-dir . eboot.elf app.exe

over the complete app0/; libaegir_f/libmemorywrapper_f stay title-deferred at the app0/ root, libs/ holds the same-vintage Windows build of every prx plus the three MinGW runtime DLLs. --to-intel is required on non-SSE4a hosts (the guest uses insertq/extrq). Two environment variables only for logging (APS5_LOG_TIMESTAMPS, APS5_CRASH_STACK_BYTES).

Prior progress notes live in #8514 — launch-subset forensics, the SSE4a fix, the deferred-module layout, and the llvmpipe mesh-output budget wall (36,864 B vs the 32,768 B device limit) that the 3070's native barycentrics bypassed.

AI-assisted: yes (Devin).

[akkadbakkad89]

Status update from local in-title runs (RTX 3070, driver 617.42, Windows 11, Khronos validation layer 1.4.363 armed, full consistent .prx set).

The stencil replay-safety rejection reported here is cleared: on a union of the open recompiler chain (#9317 + #9393/#9904/#9905/#9906), shader 0x407325e400 (stage 5) recompiles and enters stencil replay — a 256-reference unbounded export. The title then progressed through four further gates, mapped in #4276:

1. MSAA stencil draw — a local relaxation of the single-sample requirement passed it (each replay pass is an ordinary draw; hardware has no per-sample stencil export, so MSAA is safe by construction — argument posted on AGC: pixel shaders that export a stencil reference value abort on devices without VK_EXT_shader_stencil_export (NVIDIA; Demon's Souls, 1080p renderconfig) #4276).
2. Guest-buffer offsets outside push constants — covered by open fix(agc): put a stencil replay shader's guest buffer offsets where its shader data is #8615.
3. sampled depth-surface feedback — the shader samples previous-pass depth surfaces (0x41b3290000 / 0x41b3b00000, 1920x1080) while the bound target is a different surface (960x540); the check is over-broad and a bound-target-scoped narrowing passes it locally.
4. CPU read over a drawn 4x MSAA color target — current frontier. Validated open fix(agc): retire an undrawn filled multisampled color target that a guest access overlaps #9186 against it today: the target is host-only-drawn (not an undrawn fill), so that PR correctly still refuses. What remains is real guest access to host-resident MSAA samples — the feat(agc): draw PS4 multisampled color targets as host-resident targets #9892 direction or a CPU-read resolve-back path.

So the blocker reported in this discussion is resolved in-tree; what stands between here and gameplay is MSAA color-target guest access, which already has upstream work in flight.