The Smurfs – Dreams (PPSA21607)

[JJenkx]

Existing thread

• Searched this category for PPSA21607 and the game name; no existing title thread found.

Title

The Smurfs – Dreams (PPSA21607)

Content version tested

01.002.000

AnyPS5 commit

Main 3aa35dc014a84b783c2fbe642f8e08341feb1a30, compared with #6959 at 7f39ff42fbe39b386a7c99633757dca2f1099be6 (that main base plus the PR). This report is pinned to these revisions, not a claim about newer main.

System tested

Arch Linux, GCC 16.2.1, AMD RX 6650 XT, RADV Mesa 26.2.4. Native Linux, virtual controller input file. Window on the secondary monitor; keyboard focus was not held.

Furthest point

Interactive village gameplay: analog input moved the character and camera from spawn to the right fence on main3aa35 + #6959. The opening sequence progresses and held Cross skips it. Adding #6948 removes the severe white village washout on RX 6650 XT; see the pinned follow-up below. Audio and full-title progression remain unverified.

Initial menu comparison (superseded progression status)

Main displays readable menus but the Play/Options/Credits background is nearly black. #6959 renders the sky, clouds and surrounding characters in that menu. Both main and #6959 reached a colorful opening cinematic image. The candidate's captures at 225 and 285 seconds are identical, so reaching that image must not be read as successful video playback. Audio has not been checked.

Both completed by ordinary window close with native launcher/supervisor exit 0, no remaining game PID, and the temporary udmabuf limit restored to its previous value. Menu frame rates reported by the window were low, roughly 1–3 FPS. These are sparse observations, not a measured performance comparison or a speed-up claim.

Main menu	#6959 menu

Initial investigation leads (superseded by follow-up below)

• Menu rendering: fix(agc): render layered mesh targets within device limits #6959 supplies the visible improvement above on this GPU.
• Gameplay and cinematic progression remain to be tested; no fatal/NotImplemented line identified in this run. The opening-image observation alone does not diagnose its cause.
• Related work: fix(agc): store linear compute writes behind their dispatch #6948 reports a linear/small-texture readback correction for washed-out Smurfs village lighting on another setup. It has not been applied or validated in these local runs, and is listed as a lead rather than a demonstrated dependency here.

How it was run

The supplied title has an AMPRPAK4 index with one packed file. A first run stopped because smurf_dreams/content/paks/pakchunk0-ps5.ucas was absent from the loose-file view. Unpacking with ampr_emu's AMPR tool, python3 ampr_pack.py unpack --index /path/to/ampr_assets.index --output /path/to/unpacked, recovered the packed file; the unmodified original dump is kept separately. That preparation alone allowed main to reach menus.

Native library paths point to the respective main/PR build-full/core/libs/libs. Common settings: VK_DRIVER_FILES=/usr/share/vulkan/icd.d/radeon_icd.json, APS5_HOST_IMPORT_MIB=16384, APS5_DRAW_CACHE_ENTRIES=4096, APS5_PIPELINED_DRAWS=1, APS5_PAD_INPUT_FILE=<runtime>/native-pad.txt. Each run used isolated empty shader/Mesa cache directories and the same saved-game checkpoint for the displayed menu comparison. Capture timing differs, and no performance comparison is claimed.

Evidence packet: exact command arrays, native reports, terminal results, selected screenshots and hashes. Main menu is sync203 at actual 115 seconds; PR menu is sync201 at actual 86 seconds. An additional sync202 run had an iconic window during its menu interval; delayed screenshots marked 115/155 were actually captured at 212 seconds and are not menu-time evidence. No game assets, binaries or saves are included.

Follow-up: village reached and PR6948 lighting validation

Independent PPSA21607 (The Smurfs – Dreams 01.002.000) rendering check on RX 6650 XT / RADV Mesa 26.2.4, Arch Linux / GCC 16.2.1: applying this PR removes the severe white village washout in my local comparison.

Baseline 7f39ff42fbe39b386a7c99633757dca2f1099be6 is main 3aa35dc014a84b783c2fbe642f8e08341feb1a30 plus #6959. Candidate 97a9225d722842fc1e893ceabca3969617230e47 adds the exact #6948 diff at 2c4d1388a17bb84784bb877514b6dd1e19478526. This isolates this PR on that integration base; it is not a result against newest main.

Both started from the same checkpoint with separate initially empty shader/Mesa caches, identical input through the compared 345-second village capture, and no keyboard-focus holding. Both reached the village and closed normally (native/supervisor exit 0, PID gone, udmabuf restored to 64 MiB). Baseline subsequently accepted analog movement, moving the character and camera from spawn to the right fence. Movement on the candidate was not tested. The opening sequence also progresses beyond the previously reported still image. Audio and full-title progression remain unverified.

Baseline village	With #6948

Focused tests on the same GPU: baseline 6/6; candidate 8/8 including agc_driver_linear_readback_label and its expected-failure disabled control. Mandatory Vulkan, core validation and shader cache disabled: no skips, missing-layer messages or VUIDs. The candidate follows this PR's test configuration, including APS5_NO_LABEL_READBACK_STORE=1 for clamped_global_fill.

Performance is still very slow: the candidate window reported about 0.9–1 FPS in the village. These are observations, not a controlled performance comparison or a speed-up claim. Removing the visible washout does not establish pixel-perfect reference correctness.

Evidence packet: exact commands, pins, native reports, focused-test summaries, selected captures and hashes. No game assets, binaries or saves included. Game discussion: #8158.