Demon’s Souls

Last updated: 2026-10-12. For setup, see installation and boot; for current limitations and availability, see Testing and Known Issues.

KytyPS5-v20261013 test — 2026-10-13

Release KytyPS5-v20261013, source 446b9f60.

Correction (2026-10-13): an earlier version of this section showed six area results. Those runs never reached the areas: the scripted route pressed its buttons during the logos, and every run measured the game's opening cinematic. The results are withdrawn. A corrected area test, with cutscenes skipped and each area verified from screenshots, will replace this section.

Verified on this build (Ryzen 7 7800X3D, RTX 5080, NVIDIA 591.86, Windows 11):

• 15-minute precompile: finished in about 17 minutes including the first-start preparation and stored 1,869 MB.
• Game start: the game started from the release ZIP and played its intro without a crash.
• Background precompile: it continued during a session (58 batches in 4½ minutes) and stopped when the game closed. A script reading the launch script's output returned when the game exited. Its effect on gameplay frame times has not been measured.

Preview 2 test — 2026-10-11

Item	Verified value
Source	b3266317 (release)
Game	Demon’s Souls PS5, PPSA01341 version 01.007.000
Platform/hardware	Windows 11, AMD Ryzen 7 7800X3D, RTX 5080, NVIDIA 591.86
Start	Release ZIP extracted fresh; launcher.cmd double-clicked; package defaults (1920×1080 fullscreen, on-demand recording)
Data	Copied Boletarian Palace save; a compatible cache from the previous build (888 inputs / 803 recipes, 87.6 MB of driver caches), reused without regeneration
Progress	Title 60 FPS → Continue → fog load (about 5½ minutes) → play in the save’s area
Play	Mostly 56–60 FPS standing; dips to 30–48 FPS and severe hitches in new areas and when fights start
End	640 seconds; normal close, exit 0; cache grew to 2,300 inputs / 2,142 pipelines

The tester reported loading and play felt the same as earlier builds. The game thread waited on 738 pipeline creations of 100 ms or more, about 350 seconds in total; the longest stalled frame was 76.6 seconds. A later start with the grown cache was reported to load the fog scene much faster; it was not timed. See Testing and Known Issues for the cause of the pauses and the limits.

Preview 1 test — 2026-10-11

Item	Verified value
Source	635289e8
Game	Demon’s Souls PS5, PPSA01341 version 01.007.000
Platform/hardware	Windows 11, AMD Ryzen 7 7800X3D, RTX 5080, NVIDIA591.86
Display	Requested and measured client 1280×720, fixed bordered; maximize did not enlarge it; internal raster unmeasured
Profile	On-demand record, prefetch 0, intensity 100, fog_enable=false, doIndexBufferCulling=false
Data	Copied save and fresh matching cache; originals preserved
Progress	Title, main menu, Continue/Continue Offline, Nexus character/HUD/world
Input	Screenshot comparisons after forward movement and camera rotation verified visible response
End	604.662 seconds; normal close at planned limit, exit 0/no device loss/no forced cleanup

Cold startup and Continue required minutes while pipelines compiled. Shader-table binding and image-set refusal messages remain unresolved. Title readings from 1–60 FPS and a cinematic 60 FPS sample do not establish a gameplay rate; low readings occurred after a new camera view. Combat, campaign coverage, complete rendering correctness, warm-restart improvement and fresh physical vibration were not tested. Complete first-use script preparation was not part of the direct game run. Earlier cutscene/flicker failures have not been demonstrated fixed by this run.

Windows/Linux build-and-test, five GPU regression groups and isolated Windows CLI/Qt startup checks passed for this exact source. The hidden Qt check required a controlled stop, separately from the game’s normal exit. See the compatibility table for the scope of each result.

Unreleased 1080p fullscreen checkpoint — 2026-10-11

A separate local candidate reached Nexus character/world/HUD in a captured 1920×1080 borderless fullscreen client after reusing application caches, without a full precompile. The initial fresh-cache run remained at Bluepoint; both tests closed normally without reported device loss. New-content creation hitches remained. No stable FPS or complete controlled-route comparison was measured. Boletarian Palace 1-1 was prepared but not launched in that checkpoint. See the scoped test and GPU-translation findings. These historical observations do not certify the published previews’ fullscreen behavior.

Parallel loading development checkpoint — 2026-10-12

A separate unreleased candidate reached Gates of Boletaria and Nexus at 1280×720 windowed on Windows 11, Ryzen 7 7800X3D and RTX 5080/NVIDIA 591.86, using PPSA01341 version 01.007.000. Boletaria testing confirmed player/world/HUD, movement, camera, combat, healing and death/reload. The fresh-cache Nexus replay confirmed movement between floor, steps and NPC positions and camera response during about 29 minutes of mixed play and observation. It included about 78 seconds of requested input holds, not continuous traversal throughout the session. A normal close during loading also succeeded. No device-loss report was detected in these bounded runs; originals were preserved.

Parallel creation reduced the observed Continue Offline-to-world interval in a controlled 720p comparison, while first-use pauses remain. Complete rendering, stable FPS, campaign progress and results at other resolutions or on other hardware remain unverified. This candidate is not in the published previews or currently published source. See the comparison and its cache limitations.

Tested configuration

The following configuration belongs to earlier local experiments; each recorded run retains its actual settings.

Item	Recorded value
Game	Demon’s Souls, PS5
Serial / version	PPSA01341 / 01.007.000
Test platform	Windows 11
CPU	AMD Ryzen 7 7800X3D
GPU / driver	NVIDIA RTX 5080 / 591.86 (32.0.15.9186)
Compiler / configuration	clang-cl 19.1.5, Release with interprocedural optimization
Recorded output	Earlier opening/play and controller-feedback sessions: 1280×720; Nexus-fog and Outpost replays: 1920×1080. Latest Nexus replay requested 1280×720, then used a maximized bordered window; actual client/render dimensions unmeasured. Mailbox presentation
Other recorded settings	English US, AMD CPU compatibility patches, Windows red-zone protection, shader validation enabled
Data isolation	Separate test save, configuration and cache directories
Experimental source basis	Chen snapshot 75aa4b9e; see source provenance

Earlier test requests used 1280×720. The run history identifies each session’s actual output dimensions. Window dimensions do not establish the game’s internal raster resolution or guarantee a reduction in GPU workload. This hardware identifies the tested configuration; it does not establish minimum system requirements.

The experimental integration uses Chen’s coherent runtime, shader tooling and dependency payload. Its results do not establish identical behavior in an unmodified Chen checkout or in another build. The regression failure and pending verification described here belong to that historical source, before the preview’s corrections and complete tests.

Recorded runs

The following dated records describe earlier local experiments. Their executable identities, settings, incomplete preparation and verification limits are historical; the preview result above is separate.

Earlier local baseline — 2026-10-10

A separate local baseline stopped during startup before its first rendered frame, after approximately 4.56 seconds, with process exit code 321. No device loss was recorded. A failed 2 MiB direct allocation and unresolved NpCppWebApi imports appeared before the exception; neither has been established as the cause.

Executable SHA-256:

F6FABC7374FEA6778B3AD2855F6E950081130C01A12C5633283B81B77B948745

This earlier local baseline is separate from both the original public parent snapshot and the tested preview.

Experimental opening run — 2026-10-10

The locally compiled experimental executable rendered the opening narration. The run lasted approximately 307 seconds and ended at a configured test deadline, rather than a recorded crash. No device loss was recorded. Introductory-screen telemetry showed 55 FPS at one capture and approximately 60 FPS earlier; this does not measure sustained gameplay.

Experimental executable SHA-256, also used for the play session below:

2DB7DEB499D32B0F4F21EFFBE399B9D896D1A37CEDD9DAD4DC1C982E9C563041

Experimental play session — 2026-10-10

The same executable was restarted without an automated cutoff, reusing the caches from the opening run. The tester reported completing character creation, with severe and inconsistent frame drops. After cutscenes, the display stayed black while background audio continued; the displayed FPS reached zero. The session lasted approximately 331 seconds and exited with code 0 when closed. No device loss was recorded.

This is an unresolved progression failure. A normal process exit does not demonstrate successful gameplay, and the available observations do not establish whether shader compilation, guest logic, rendering or another subsystem caused the black screen. The later copied-save replay below follows a different entry path and does not prove that this original transition was corrected.

Experimental Nexus-fog replay — 2026-10-10

A copied-save replay at 1920×1080 displayed a fog-loading screen, without verified controllable area access. The session lasted approximately 207 seconds and was closed manually, with exit code 0 and no recorded device loss. Long frames were heavily attributed to compilation, including synchronous pipeline creation lasting approximately 30 seconds. Its displayed FPS could remain stale during long frames.

The replay used the same experimental executable digest above, warm caches and an isolated save copy. Cache coverage grew from 893 inputs/796 recipes to 960 inputs/859 recipes. This is additional encounter coverage, rather than a complete preload or a solution to the prior progression failure.

Controllable Outpost Passage replay — 2026-10-10

The same locally compiled experimental executable was replayed at 1920×1080, reusing the 960-input/859-recipe encounter set in a fresh save/configuration/cache directory. Outpost Passage rendered with the character and HUD. The tester confirmed movement and camera control, reporting below 10 FPS while moving, up to 40 while stationary, and black flickering. The approximately 269-second session ended when closed manually: exit code 0, no automated cutoff and no recorded device loss.

The intended Nexus save fixture led to Outpost Passage. A valid save checksum therefore did not prove the selected destination or story flags. Nexus traversal and the other world-entry copies remain unverified. The moving and stationary intervals were not marked for a controlled comparison; the reported rates should not be read as a repeatable performance benchmark.

Pipeline compilation continued in the world scene. One approximately 9.397-second logged frame attributed 9.364 seconds to compilation; a later 6.362-second frame attributed 6.156 seconds to compilation. The cache grew to 1,338 inputs and 1,182 recipes. These observations show compilation contributes to some pauses; they do not establish the cause of every frame-rate drop or flicker.

Indoor hub and confirmed controller feedback — 2026-10-10

A subsequent locally compiled experimental candidate was tested at 1280×720, with controller intensity set to 100. The test supplied fog_enable=false; the setting’s actual visual effect was not independently established. A copied-save replay rendered the character, HUD and NPCs in an indoor hub. The scene appeared to be Nexus, but the exact area name and traversal coverage remain unconfirmed.

The tester explicitly confirmed vibration working in game. This is physical feedback for this configuration, rather than only a software-test result. The observation does not isolate ordinary motor rumble from DualSense waveform haptics, and does not verify adaptive triggers, speaker output or feedback latency. Haptic-channel attribution and complete candidate acceptance remain pending.

The tester reported severe fog/continue-loading lag, movement still below 10 FPS, subjective improvement over the earlier 1080p replay, improving FPS as the scene warmed, and random indoor drops. The source, window dimensions, fog setting and cache state changed together; this is not a matched comparison and cannot attribute improvement to resolution, fog or cache reuse. The previous black-flickering report remains unresolved.

The session lasted 521.56 seconds and exited with code 0 when closed manually. There was no automated cutoff or recorded device loss. One logged frame lasted approximately 17.008 seconds, with 16.940 seconds attributed to compilation. Encounter caches grew from 1,338 inputs/1,182 recipes to 1,729 inputs/1,540 recipes. These observations establish additional coverage and some compilation stalls, rather than stable gameplay or complete preloading.

Experimental executable SHA-256:

34F51D20271E6C6D17384CB7D424950BF339DF9035B3CAFC06B6FC992D5490F7

This executable and its experimental integration have not been uploaded as public source or a custom release. The result does not establish identical behavior in the public fork’s current revision.

Shader preparation and cache observations

Stage	Observation	What it establishes
Seed discovery	20,141 programs and 53,522 pipeline recipes discovered in 79 seconds	Discovery completed; full static-input validation was interrupted
Restart after the first run	160 shader inputs and 156 recipes loaded	A previously encountered subset persisted across restart
End of the play session	893 inputs and 796 pipelines saved	More encountered inputs were recorded
Driver-cache output	54,854,232 cache bytes and 28,010,599 table-cache bytes saved	Cache output was written; successful reuse for every pipeline is unverified
Later Outpost replay	Started with 960 inputs/859 recipes; ended with 1,338 inputs/1,182 recipes	New world-scene encounters still required compilation; full preload remains unverified
Latest controller-feedback replay	Started with 1,338 inputs/1,182 recipes; ended with 1,729 inputs/1,540 recipes	Additional indoor-scene encounters recorded; reported warming improvement has no controlled measurement

One cold compute-pipeline compilation took approximately 2,361 ms. A later play-session compute compilation still took approximately 1,044 ms. Long shader and driver-pipeline compilations therefore remain an observed source of stalls. Persistence alone does not prove a measured cold-versus-warm improvement or explain the progression failure.

The discovery counts do not imply full shader coverage across every map, effect or game state. Cache behavior depends on GPU, driver, build and shader inputs. Full offline driver-pipeline preloading has not been verified. No extracted shaders or cache files are attached to this wiki.

Newer experimental preparation checkpoint — 2026-10-10

A newer locally compiled, unreleased candidate passed focused checks and began known-input preparation. Its discovery pass found 20,141 programs and 53,522 graphics recipes in approximately 50 seconds. This is a separate attempt from the earlier 79-second discovery record above; the counts do not establish complete game coverage.

The first preparation attempt failed while saving cache files under a deeply nested Windows folder. A focused file-creation check reproduced the long-path failure. With the same executable and a fresh short isolated cache location, the previously failing batch completed and saved its output successfully. Original data and the failed result were preserved. This verifies a configuration workaround for that observed failure; permanent native long-path handling has not been corrected or verified.

At this checkpoint, the full preparation batch was incomplete. No fullscreen game run, rendering result, preparation-wide success, compatible-build migration benefit or measured warm-restart improvement is recorded for this candidate. No fixed completion time is established.

Newer experimental executable SHA-256:

07168585A4C95888F01C57120270CD795F77227131F1B96AC6A18338B1B271DB

This exact experimental executable was not published. The preview’s preparation tools are documented in the installation guide; these older sample results do not verify another executable.

Later experimental preparation stopped — 2026-10-10

A later unreleased preparation attempt was intentionally stopped after approximately 49 minutes and 6 seconds. It did not complete. Saved partial outputs and the original data were preserved, the preparation workers ended, and no device loss was reported in the preserved result. Its automatic game continuation did not run.

This is elapsed time for an interrupted attempt, not the duration of a completed preparation run or a measured speedup. Partial batch progress does not establish unique successful pipeline coverage or a fully prepared cache.

A subsequent local evaluation has now started in fullscreen at 1280×720, using preserved partial cache data and on-demand compilation instead of waiting for full offline preparation. The runtime reported loading stored pipeline data and adopting recorded inputs without bulk startup preparation. Early frame progression was recorded, but no visible menu/area, stable gameplay frame rate, smoothness, newly verified controller feedback or warm-restart benefit is established at this launch checkpoint. First-use compilation and stalls can remain.

Preparation tool SHA-256 for this stopped attempt:

507DD761777871B958FB7CE21EC6A3E9FA62E341A2B0ECD76BC1036D07758FD2

Experimental emulator SHA-256 for the subsequent on-demand launch:

36E7F575C70A101A2E5AE87BEE6FB2468445B284D288F2F27EB83BCDEA32E450

That exact experimental executable was not shipped as the preview. The earlier documented physical-vibration result belongs to the earlier controller-feedback candidate. See understanding shader preparation for the difference between shader inputs, complete pipelines and deferred first-use work.

On-demand exit and replay display — 2026-10-10

The first on-demand attempt ended normally after approximately 144 seconds, without an automatic deadline or reported device loss. It saved cache data and input records. Normal exit does not establish a visually verified scene or stable gameplay performance.

The logs still contain long first-use events: one frame took about 13.24 seconds, with about 13.23 seconds attributed to compilation, and a compute-pipeline creation took about 9.76 seconds. These are selected logged events, not a full frame-time distribution, a matched speed comparison or proof of which worker pool caused the delay.

A replay has started with the same experimental executable and copied saved cache data. Its display was changed from fullscreen to a maximized window with a title bar and borders, without restarting the game. The requested test dimensions remain 1280×720; the actual maximized client size and internal render resolution have not been measured. Visible area, controls, physical feedback, performance and warm-restart improvement remain unverified at this replay checkpoint. The stopped full preparation remains incomplete.

Nexus replay and unresolved stutters — 2026-10-10

The replay described above reached Nexus, as confirmed by the tester. It ran for 400.631 seconds and exited normally with code 0, without an automated cutoff or reported device loss. It used the same experimental executable with SHA-256 36E7F575C70A101A2E5AE87BEE6FB2468445B284D288F2F27EB83BCDEA32E450, on the configuration listed above. The tester reported that it remained laggy after warming, similar to the previous attempt. This is an area-identification result, not verified traversal, stable gameplay or a performance improvement. No new physical-feedback observation was recorded for this replay.

The display changed from requested 1280×720 fullscreen to a maximized window with borders during this run. Actual client and internal render dimensions remain unmeasured. The session finished with 1,737 shader inputs / 1,548 recipes recorded; these encounter counts do not establish complete preparation or coverage.

Logs recorded only frames exceeding 100 milliseconds: 76 slow frames across the run, including 17 lasting at least one second. Compilation scopes accounted for at least 80% of frame wall time in 61 of those 76 frames. The longest logged frame lasted 45.489 seconds, with 45.460 seconds attributed to compilation scopes. Those scopes can include waits and other work; they are not pure shader-translation or driver-call timings and do not identify a proven contention mechanism.

After 300 seconds from launch, four thresholded slow frames remained, ranging from 138.6 to 347.2 milliseconds. The largest late frame included 323.7 milliseconds in compilation scopes. Faster frames were omitted, and exact loading, movement and scene timestamps were not recorded. These observations are not average FPS or a full frame-time distribution. Late displayed FPS does not override the tester’s report of movement lag.

The complete preparation attempt remains interrupted. At that checkpoint, further live testing awaited implemented and verified improvements. The later preview verification is recorded above; these diagnostic results alone did not establish a fix. This candidate has not been uploaded as public source or a custom release, and no warm-restart benefit has been verified.

Pipeline-reuse research remains experimental. See understanding cache reuse reports for the distinction between loaded entries, saved data and successful in-game reuse. No new real-driver reuse, gameplay improvement or warm-restart benefit has been verified since this completed replay.

Remaining coverage

Stable sustained gameplay, confirmed Nexus traversal, other areas, campaign progression and completion remain unverified. The earlier controller-feedback candidate has both focused software evidence and tester-confirmed physical vibration in game; that observation does not verify physical feedback in the preview. Standard rumble, DualSense waveform haptics, adaptive triggers and speaker output require separate attribution and verification. Save-copy tools can prepare a map position, but a valid checksum does not establish story flags, unlocks or actual playability.

DLSS reconstruction and frame generation are being investigated separately. The earlier Outpost executable had no enabled Streamline backend, and no DLSS performance benefit has been verified. No reconstruction or frame-generation result is claimed for the latest controller-feedback candidate. Generated display frames must not be presented as guest gameplay/simulation FPS. The completed test with fog_enable=false does not establish that the setting took effect or improved performance.

See testing and known issues for current limitations and sources and credits for the code and tooling used. The dated experimental results retain their historical status; the preview record above identifies the tested release separately. Consult GitHub Releases for versioned packages.

Additional loading experiment — 2026-10-12

An unreleased development experiment examined whether physical-device format capability queries could return directly while earlier CPU recording continued. Four Nexus runs used the same executable, copied starting save, settings and empty emulator caches on Windows 11, Ryzen 7 7800X3D, RTX 5080 and NVIDIA 591.86 at 1280×720 windowed/Mailbox. The fully drawn world, player and HUD appeared in overlapping intervals: about 74–87 seconds after confirming Offline with the original scheduling, and about 74–92 seconds with the experimental scheduling. Ten-second captures limit those intervals; global driver-cache isolation was not established.

The named format-query waits disappeared, but waits at queue capacity and pipeline creation increased. Thresholded diagnostic totals overlap and omit faster frames; they are not elapsed savings or average FPS. No clear loading or hitch improvement was demonstrated, so the experimental bypass was not retained.

All four runs reached the Nexus, responded to bounded movement and closed normally. One enabled run additionally showed camera response and movement down the steps. A healing input was delivered, but consumption was not confirmed before shutdown. The first control later spent hours idle while observation was inactive; that idle duration is excluded from performance and gameplay claims. The final enabled repeat briefly lacked foreground before the title prompt, which limits its startup comparison. No second-area, combat, full-campaign or steady-FPS result is implied by this experiment.

Settings and source saves were preserved; no existing cache was cleared or retagged. Normal global driver-cache writes were not excluded. This result does not change the available public source, previews or releases. See loading comparison limits.