Big Helmet Heroes (PPSA19943) — tutorial boots on Linux; black artifacts and low FPS

[aryan0078]

Big Helmet Heroes — PPSA19943, v01.000.400 (Linux)

Status: reaches the tutorial; severe black rendering artifacts and low performance. Not playable in this test.

Environment: Ubuntu 26.04.1 LTS, kernel 7.0.0-38-generic, GCC 15.2 Release, Core i7-12700, 30 GiB RAM, Radeon RX 6700 (owner-confirmed), PCI 1002:73df / RADV NAVI22, 10 GiB VRAM, Mesa 26.0.8. The GPU's shared PCI family name includes other models; it is not an RX 6700 XT.

The executable and 10 bundled modules were converted with --to-intel --registry. Controller A/Cross selects New Game, difficulty and one player. Main-menu UI, tutorial HUD, character and some vegetation render; much of the scene remains black. The owner confirmed entry into the tutorial, and a subsequent screenshot shows the tutorial HUD and character. Window FPS is approximately 4–5 in the menu and 0.7–1.2 in the tutorial; these are observations, not a controlled benchmark. Loading screens can report approximately 60 FPS and should not be mistaken for gameplay performance.

Startup fix

#3854 resolves EOSSDK-PS5-Shipping.debug_prx through the exact SCE module identity declared by the bundled eossdk-ps5-shipping.prx. The manual filename alias was removed for testing. Standalone on main a11c8f7: 114/114 relinker tests pass, with a failing-before/passing-after dependency regression. This fix alone does not supply the rendering features below.

Local renderer integration

Tested integration head ed2c56d, based on main a11c8f7, published for reproducibility at https://github.com/aryan0078/AnyPS5/tree/fix/agc-big-helmet-bringup. It combines the existing work in #1077/#3763, #425, #2904, #454, #2434 and #2053, retaining their authors. This is not a standalone compatibility claim for any one PR.

• Main initially refuses multi-slice 3D color targets. feat(agc): render layered draws into 3D color targets #1077/fix(agc): render all attachment layers with dynamic rendering #3763 allow this pass.
• feat(agc): depth planes sampled as 2D arrays, cubes and integer views #425 supplies depth-surface storage reads.
• Loading then fails on an incompatible reused depth allocation: an 8192x2048 D16 surface viewed as a 1024x1024 R8 2D array. fix(graphics): allow storage textures in reused depth memory #2904/fix(libSceAgcDriver): retire a depth surface once another writer covers its memory #454, plus the integration correction below, pass that point and enter the tutorial.
• Import probes on RADV show host-pointer imports resolve write protection while dma-buf imports preserve it; fix(agc): import guest memory through dma-buf before host pointer when host imports lose write tracking #2434 selects dma-buf. A 497 MiB import also exceeded this host's 64 MiB udmabuf size limit. Testing with size_limit_mb=1024 removes that allocation refusal. Remaining non-shared allocations still use copies.

Two small integration corrections are published separately as commits:

1. e61e955: descriptor-based depth lookup must skip retired surfaces, matching fix(libSceAgcDriver): retire a depth surface once another writer covers its memory #454's other lookups. A test with two different extents at one address fails before the change and passes afterward.
2. ed2c56d: fix(agc): take depth fast clears from HTILE written by any kernel #2053's metadata scan and fixture must use main's padded HTILE allocation. The original fixture was undersized and the partial-clear regression failed; using HtileMetadataBytes and a correctly sized fixture passes it.

The patched runtime libraries were regenerated with the explicit libs target and tested, rather than running the unpatched test library as the guest driver.

Validation and remaining blockers

Focused tests pass: patched_libraries, agc_driver_depth_htile_write, agc_driver_depth_storage_reuse, agc_driver_depth_surface_last_writer and agc_driver_host_import_dma_buf_watch (5/5, 0.28 s). Source conventions and diff whitespace checks pass. The full renderer suite and Windows title execution were not validated; no full CI-green claim.

The game remains mostly black and too slow. The log also reports ignored layer/viewport exports (PA_CL_VS_OUT_CNTL=0x01240000), but that warning has not been established as the cause. Additional graphics correctness and performance work is needed. No game binaries, shader captures or private desktop screenshots are included in this report.

AI-assisted investigation and integration: OpenAI Codex. Further graphics corrections are on the integration branch, not folded into the independent relinker PR. CONTRIBUTING currently limits authors with no merged PRs to one open PR; #3854 is that PR.