Diagnostic cinematic · rendering defects
October 16: shared resource preparation uses fewer copies, faster content checks and shorter scalar walks. PPSA21564 v1.018 has a new title measurement; cinematic rendering defects remain.
Animated title: 20.07 FPS over 2,640 frames / 131.56 s; 60-frame blocks 18.39–21.68 FPS. Ryzen 7 7700 / RTX 3070 Ti, 1080p output, game-controlled 4K source, retained caches. This is a title-screen average, not a minimum or gameplay result. The earlier crowd sample was 4.60 FPS over 10 s.
NVIDIA GeForce RTX 3070 Ti
Result
Each entry names the release that includes the reported changes. Work not yet included in a published release is marked as a development build. Dates and FPS remain those of the original runs.
The renderer avoids 26.125 MiB of obsolete host copies per frame; queue discovery avoids about 117 repeated pointer checks. The uninterrupted title measurement is 20.07 FPS over 2,640 frames, with 60-frame blocks at 18.39–21.68. 680 HLE tests, 274 Vulkan tests and 64 native overwrite configurations pass. Twelve unresolved storage bindings and cinematic defects remain; character control is unverified.
Read the report →
Borrowed shader data, compact lookup keys, content fingerprints and cached loop writes reduce CPU work. Omitting inert CPU scalar steps measures 17.85 → 17.88 FPS, an inconclusive difference. The final 2,640-frame run averages 17.83 FPS; 582 GPU/Vulkan tests and five native probes pass. Twelve unresolved storage bindings and cinematic defects remain; 20 FPS and character control are unverified.
Read the report →
Packed linear texture reads, smaller scalar scratch and static-query caches reduce CPU work. Counter-assisted GPU waiting measures 17.78 → 18.02 FPS in a balanced comparison; the uninterrupted 2,640-frame confirmation averages 18.05 FPS. 573 GPU/Vulkan tests and native validation pass. Twelve unresolved storage bindings and cinematic defects remain; character control is unverified.
Read the report →
AnyPS5/KytyPS5 review: omit 49 unused CPU proof reads per frame and reuse compute texture instruction indices. Later title comparison: 14.95 versus 14.91 FPS, 660 frames per mode; no established gain, 25 FPS not reached. 560 GPU/Vulkan tests and native validation pass; 12 unresolved storage bindings remain. Graphics uniform specialization was also tested and remains disabled because it reduced FPS.
Read the report →
Reviewed AnyPS5 and KytyPS5; removed redundant preparation for proven entry scalar reads. Title: 14.92 versus 15.02 FPS, 660 frames per mode; no measured gain, 25 FPS not reached. About 411 buffer preparations avoided per frame. 558 GPU/Vulkan tests and native graphics validation pass.
Read the report →
AnyPS5 review led to fixes for unused scalar-buffer fields and bounded texture descriptors assembled from multiple loads. Title: 14.97 FPS over 660 frames. 556 GPU/Vulkan tests and native probes pass; ten broader RDNA2 failures also occur on the baseline. New capture; 25 FPS and character control remain unverified.
Read the report →
Title: 15.33 versus 15.22 FPS, 660 frames per setting. Unusable replay data is no longer copied or hashed. A larger cache did not improve FPS; the default stays unchanged. All 553 tests and native Vulkan checks pass. New title capture; 25 FPS and character control remain unverified.
Read the report →
Title comparison: 14.48 versus 14.26 FPS, 660 frames per setting. GPU layout conversion preserves padding; rectangular overwrite proofs avoid redundant uploads. Shared scalar walks use exact instruction budgets. All 551 tests and native validation probes pass. New title capture; 25 FPS and character control remain unverified.
Read the report →
Title comparison: 14.25 versus 14.19 FPS, 660 frames per setting. Reused immutable call/NGG analysis, cheaper buffer queries and bounded Vulkan waits. Added verified 8/16-bit GPU layout transforms; storage uploads stay disabled after a slower comparison. 548 tests pass. New title capture; 25 FPS and character control are not confirmed.
Read the report →
Balanced title comparison: 13.79 versus 13.43 FPS, 660 frames per setting. Less repeated shader analysis, indexed dirty-buffer completion, queued satisfied waits and faster large copies. 542 GPU/Vulkan tests pass. New title capture; 25 FPS is not reached and character control remains unverified.
Read the report →
A 250 μs command-recording budget improved the balanced title comparison from 11.69 to 12.43 FPS. Later storage changes measured 12.55 versus 12.49 FPS: no clear extra gain, but 26.125 MiB of redundant hashing is avoided per frame. 537 GPU/Vulkan tests and 64 overwrite configurations pass. New title capture; 25 FPS and character control are not verified.
Read the report →Register snapshots are half the previous size while retaining dependency tracking. Scalar uploads write only the words used by the shader. GPU/Vulkan tests pass 535/535; native checks use nonzero poisoned buffers and cover changing inputs, loops, pixels and worker/shared preparation. A 1,320-frame title run measured 11.62 FPS. The balanced upload comparison measured 11.79 versus 11.69 FPS, 660 frames per setting: only a 0.9% observed difference. No fresh screenshot or character-control check; 25 FPS remains unmet.
Read the report →Exact texel digests replace retained image copies, and checked guest fingerprints avoid the temporary snapshot. Balanced title comparisons: 11.19 versus 10.83 FPS for digest baselines; 11.51 versus 11.36 FPS for direct fingerprints, 660 frames per setting. Each step removes a 26.125 MiB CPU copy per frame. Native checks pass 32 configurations covering padding, newer CPU writes, unavailable fingerprints, aborted preparation and fallback; mip publication also passes. No new capture or visual/control verification. The 25 FPS target remains unmet.
Read the report →M0-indexed register moves now preserve indirect reads/writes, masked lanes and loops. Native checks pass 128 GPU dispatches; GPU/Vulkan tests pass 534/534. The warmed animated title measured 10.85 FPS across 1,320 frames, comparable to the previous build. No new draw/dispatch failures appeared in the log. No fresh screenshot or visual/control verification; the 25 FPS target remains unmet.
Read the report →A conservative shader proof avoids uploading old output pixels while preserving padding, newer CPU writes and recovery after aborted preparation. Native GPU checks pass with immediate and batched commands; 532 GPU/Vulkan unit tests pass. Balanced title samples: 9.45 FPS enabled versus 9.36 disabled, 660 frames per setting, with 25.38 MiB of uploads avoided per frame. The first old-raster readback comparison measured 9.64 versus 9.67 FPS but skipped no readbacks. The 25 FPS target remains unmet; no new screenshot or character-control result. With DCC-enabled raster sources included under the same coverage checks, the final comparison reaches 10.95 FPS versus 9.68 FPS (660 frames per setting, about 13.1% faster), avoiding one raster readback per frame. Native tests also preserve DCC metadata.
Read the report →Native image transfers preserve texels, padding and newer CPU writes in GPU tests. Exact buffer reads preserve published holes; completion batches reuse valid prefetched copies. Vulkan tests pass 259/259. Balanced title comparisons measured 9.42 FPS with native storage transfers versus 9.55 with the CPU path, and 9.50 versus 9.55 after the prefetch completion fix. Those experimental transfers remain disabled by default. Exact interior buffer reads recorded no uses in their measured interval. The 25 FPS target is not reached; no new screenshot or character-control result is claimed.
Read the report →Deferred GDS snapshots preserve counter values across later resets; transfer batching and full-image overwrite handling pass native checks. The follow-up warmed title baseline is 9.20–9.65 FPS over 20 seconds. Enabled/disabled comparisons remain around 9–10 FPS and do not establish a speedup. Vulkan tests pass 256/256. Incompatible image views still cause about 20 MiB of render-target readback per frame. No fresh screenshot or character-control check: window automation was unavailable. The 25–30 FPS target remains unmet.
Read the report →October 14 development build: warmed title observations are about 9–10 FPS; the crowd scene presents 46 frames in 10 seconds (4.60 FPS). Shared command banks, bounded resource analysis, cache recovery and native GPU transfers reduce overhead. GPU tests pass 272/272 and Vulkan tests 256/256. The final 2D storage-to-color transfer passes native probes but has no live benchmark yet. Rendering defects remain, character control is unverified and the 25–30 FPS target is not reached. No new capture; the displayed image is from October 13.
Read the report →Whole-wave scalar branches and converged LDS barriers fix the repeating bands. BVH bounds and opaque child pointers are corrected; opt-in PSR v9 snapshots complete the captured query with a 16,384-block budget. Later sampled frames report no draw/dispatch failures, but four warm-up programs were rejected before the switch was enabled. Dark/patchy surfaces remain; gameplay and a sustained FPS gain are unverified.
Read the report →
JSON, APR, font and shader fixes reach the diagnostic hub cinematic using local font aliases and disabled acoustic propagation. The SDWA correction removes severe green corruption; this first checkpoint still rejects four BVH compute programs. Character control is not established.
Read the report →Captures
Frames produced by the emulator itself during the runs above.
Storage copies reduced; title averages 20.07 FPS
October 16, 2026 · Development build
Resource preparation updated; title averages 17.83 FPS
October 16, 2026 · Development build
Resource costs reduced; 18.05 FPS title average
October 15, 2026 · Development build
Reduce uniform-analysis reads and texture scans
October 15, 2026 · Development build
Reuse proven scalar reads
October 15, 2026 · Development build
Scalar reads and split texture descriptors
October 14, 2026 · Development build
Draw capture and cache comparison
October 14, 2026 · Development build
Storage ownership and scalar loop budgets
October 14, 2026 · Development build
Static shader analysis and narrow texture layouts
October 14, 2026 · Development build
Scalar analysis and queued completion
October 14, 2026 · Development build
Bounded recording and exact storage proofs
October 14, 2026 · Development build
Wave synchronization removes horizontal lighting bands
October 13, 2026 · Development build
Startup fixes and removal of green corruption
October 13, 2026 · Development build
20
Recorded runs
NVIDIA GeForce RTX 3070 Ti
0.3.5
Current build
October 16, 2026
Intro / menus
Result
Intro and menus render
Playthrough reports carry the date the maintainer confirmed them. Every timing refers to the NVIDIA GeForce RTX 3070 Ti development machine and will differ on other hardware. Game content is supplied locally and is never distributed here.
