Repository navigation
Pokopia: shader compilation produces temporary freezes and FPS is typically ~30 #15
Description
Affected build
Frameropia v0.0.1 (Windows), tester-reported.
Observed
- Gameplay usually sits near 30 FPS.
- An occasional maximum around 42 FPS was reported; there is no evidence of a hard 42 FPS cap.
- Compilation of new shaders causes major stutter and sometimes a temporary complete freeze until it finishes.
Expected
Shader compilation should minimize pauses without dropping required draws or causing graphics regressions.
Suggested validation
- Record exact GPU/driver and Vulkan settings, especially asynchronous shaders.
- Compare cold-cache and warm-cache traversal of the same area.
- Record shader compilation logs, frametimes and FPS in identical scenes.
- Test async compilation separately for graphics accuracy; skipped pipelines can cause temporary missing objects.
No performance or shader fix has been confirmed. See test plan.
Activity
Framerr commentedon Nov 7, 2026
Performance research: notsapinho build 24 (2026-11-08)
Reporter tested notsapinho/Eglantine build 24 and observed steady 60 FPS in Pokopia, compared with Frameropia's performance issues. The actual cause(s) are not isolated: 17 commits between builds 21 and 24.
Concrete candidates (author measurements, not our runtime validation):
e3aed3effc: skip buffer cache rescans for clean ranges, update dirty vertex streams only; reported per-draw buffer update cost ~440 → 398 ns. ToggleEDEN_NX2_SYNC_SKIP=0.a95161229e: avoid redundant graphics pipeline state lookup; author reports skipping 88.6% of 17.5M lookups with no mismatches in Pokopia; per-draw lookup ~132 → 82 ns.EDEN_NX2_PIPELINE_SKIP=0disables.6e73445ccb: linear image uploads directly from guest memory, ~194 → 120 ns per draw.a2042312ee: defer large linear image uploads to Vulkan worker thread; author reports ~25.6 → 2.8 ms/s back-thread upload cost, additional ~7ms/s worker waiting. ToggleEDEN_NX2_DEFER_UPLOADS=0.84e061ce05: far shadow cascade restored by default; optionalEDEN_NX2_SHADOW_FAR=0reduces GPU thread work ~17% but visually removes far shadows, so do not silently enable it as a speed improvement.Do not cherry-pick everything without examining dependencies, graphics regressions, compiler tests and same-scene FPS. First isolate photo PR #17; evaluate performance commits separately. No builds or releases requested yet.
Framerr commentedon Nov 8, 2026
Performance experiments B/C rejected after reporter testing (2026-11-08)
A successful build is not evidence of runtime correctness. Do not reuse #19's deferred texture upload change without isolating/fixing pink-sky issue. Investigate remaining independent performance candidates (pipeline lookups and clean-range cache scans) carefully; no user-verified 60 FPS improvement in Frameropia from B/C. The approved photo DMA fix #17 was merged separately into nx2.
Framerr commentedon Nov 8, 2026
2026-11-08 fork scan: further performance leads
The last notsapinho release remains build-24 (2026-11-07). Rejected Frameropia direct linear texture uploads (#18) and deferred upload (#19) must not be reintroduced unmodified; #19 showed intensely pink sky. Remaining isolated candidates are notsapinho pipeline-state cache a951612 (upstream author measured 88.6% skipped lookups, 132→82 ns/draw) and clean-region synchronization e3aed3e (440→398 ns per draw in author's sample). These are microbenchmarks, not observed FPS gains, and not demonstrated in Frameropia. Prefer independent A/B opt-out branches, first maintain Pokémon Pokopia image correctness.
Framerr commentedon Nov 8, 2026
Two independent v0.0.3 performance builds for reporter A/B (2026-11-08)
HintColdmaintenance to exclude guest ranges reuploaded repeatedly. No texture layout changes, no experiment: test faster linear texture uploads for Pokopia #18/experiment: defer large linear Pokopia texture uploads to Vulkan worker #19 regression.GPU::TickWorkatomic pending-request fast path from notsapinho a1008ebd, omitting unrelated macro JIT change.Both branches start from immutable release v0.0.3, are independent, draft, unmerged, artifact-only; test FPS, visible stutters, first-capture photo correctness, sky/geometry and stability before approving. GitHub Windows builds still in progress at issue update.