Super Meat Boy 3D

Observed result

Twelve 60-second level 1-1 runs split randomly between fast and slow runs on both main revisions and PR #8756; no single build determined the outcome. The 10-second present count was 30–42 in faster runs versus 5–9 in slower runs.

The GPU depth plane of 31.9 MiB is flushed as a whole when a read-only shader buffer's range intersects about 3 KiB of its stale data. Slow runs logged 1,547–1,837 depth stores (48–57 GB) per 10 seconds, versus 306–455 (2.2–6.7 GB) in fast runs. The reporter had not yet determined whether the shader needs to read the overlapping bytes.

Blockers & limits

• Repeated full-plane store-back on a small descriptor overlap greatly reduces rendering throughput for some allocations.
• Optimizing this cannot simply skip required depth synchronization because another title can read depth through a raw buffer; the correct narrow-range solution is under investigation.

Observed result

Four alternating base/patch pairs of 60-second level 1-1 runs: patch 33–46 presents per ten-second window across all four paired runs; base 2–27 presents per window. In the pair showing the original slow mode, base wrote back 47–60 GB of depth data per ten seconds, while the patched run wrote 2.5–2.8 GB.

The fix tracks 64 KiB depth-plane units, copies only those intersecting the read, and retains guest overrides. Screenshots in pairs 2 and 4 matched. Core depth and graphics tests also passed; no full-game run or stable-release binary retest was reported.

Blockers & limits

• PR #9026 is merged to upstream main but not included in the existing tagged v0.1.3 release; a new built main must be distinguished from release downloads.
• The matched comparison covers one level and one NVIDIA/Linux host; the patch author did not test the title on Windows, AMD GPUs or other levels.

Observed result

After entering gameplay, player/terrain/foliage are uniformly cyan-white [224,248,255] with black shadows; 2D UI and FMV are full color. Texture dumps showed intact diffuse textures; an MRT[3] red-output witness did not alter the final frame. DCC-key zeros and deferred-attachment routing are investigated but not a proven root cause.

Blockers & limits

• Graphics are visually incorrect despite reaching gameplay; a Windows RX 6700 XT retest on the same current main commit as the Linux comparison has not been reported.
• The tests do not separate a Windows/AMD-specific fault from source-commit differences; no verified correction or official Playable result.

Observed result

About three minutes into level 1, the tester saw green foliage, red character, purple crystals and brown ground with shadows, at about 5 FPS. This comparison does not reproduce the cyan-white visuals, but uses a different OS, GPU and commit than the Windows RX 6700 XT observation.

Blockers & limits

• Throughput remains slow at approximately 5 FPS in that first-level scene.
• Cross-platform observations are not a matched same-device A/B, so the cause of the Windows rendering difference remains unresolved.

Observed result

The earlier Windows reporter observed green foliage, red character, brown terrain and correctly rendered lighting and shadows. Main changed and the capture layer was inactive, so the individual cause was not isolated.

Blockers & limits

• The exact cause of the earlier cyan-white rendering was not isolated; this follow-up changed both the build and capture-layer conditions.