Skip to content

Releases

GSF Kedavra 1.7.0

Choose a tag to compare

@github-actions github-actions released this 18 Nov 15:32

What's new in GSF Kedavra 1.7.0

  • Auto targets 30 FPS ×2 at 60 or 59.94 Hz, including when a lower floor was saved for another display.
  • At fixed 30→60, Silk can show a ready real frame up to one refresh earlier when the same pair would otherwise generate a second intermediate frame.

GSF Kedavra 1.6.2

Choose a tag to compare

@github-actions github-actions released this 11 Nov 23:35

GSF Kedavra 1.6.2

Restores centred HUD panels used by the existing GFG Benchmark interface. Touch coordinates follow the displayed panel. Includes the 1.6.1 shared-controller lease and native frame telemetry. Focus, shaders, interpolation and the Silk scheduler are unchanged.

Install this ZIP in Decky, reinstall Gamescope from Setup and reboot. If GFG is also installed, update both plugins and install the matching GFG runtime last. Standalone Kedavra yields engine control while GFG is active without losing preferences.

Rollback: remove the custom Gamescope runtime in Setup, reboot, then reinstall the prior plugin and its matching runtime. Do not pair an old standalone controller with a new GFG controller.

Automated checks do not establish real Steam Deck touch, power or frame-pacing performance. No physical Deck was available for this build.

GSF Kedavra 1.6.1

Choose a tag to compare

@github-actions github-actions released this 11 Nov 09:21

GSF Kedavra 1.6.1

  • Adds compositor HUD touch input and native frame interval telemetry for the GFG Extreme 3.4.1 in-game Benchmark.
  • Adds controller ownership handoff: standalone Kedavra pauses its HUD and controller commands while a compatible GFG Extreme session owns Gamescope.
  • Preserves the 1.6.0 Focus shader, optical flow, frame warping and Silk scheduler.

Install: Decky Loader -> Install Plugin from ZIP. If also using GFG Extreme 3.4.1, update both plugins, then install/reinstall the Gamescope runtime from GFG Setup and reboot once. Do not overwrite GFG's runtime afterward using the standalone installer.

Software CI and virtual runtime checks do not constitute Steam Deck hardware validation.

GSF Kedavra 1.6.0

Choose a tag to compare

@github-actions github-actions released this 10 Nov 10:03

What's New in 1.6.0

  • Kedavra Focus: Adaptive edge sharpening with halo protection.
  • Simplified Controls: One ON/OFF switch, fixed 100% strength.
  • Frame Generation: Focus works on real and generated frames.
  • Upscaler Compatibility: Independent of FSR, XeSS, DLSS and OptiScaler.
  • GPU Efficiency: Single-pass processing without extra full-resolution buffers.

Focus is SDR-only. HDR is bypassed.

GSF 1.5.0 Kedavra

Choose a tag to compare

@github-actions github-actions released this 08 Nov 21:29

GSF Kedavra 1.5.0

GSF Kedavra 1.5.0 develops the Silk 1.4.0 engine used by GFG Extreme 3.2.2. This release focuses on keeping fine details sharper during fast motion while retaining the existing reconstruction for uncertain regions.

What improved

  • Sharper fast motion. The engine reduces its existing motion blur when the two reconstructed source samples agree, the motion estimate has a low matching error, and neighbouring motion vectors are consistent.
  • Conservative handling of difficult regions. Areas with ambiguous motion, conflicting samples or inconsistent neighbouring vectors continue to use the previous reconstruction.
  • A small change to the GPU pipeline. The improvement uses samples already available to the reconstruction shader. It adds no GPU passes, intermediate buffers or texture reads.
  • Existing engine behaviour is preserved. Motion estimation, frame scheduling, generation modes and SDR/HDR support remain based on Silk.

The earlier experimental local colour-range clamp is not included in this release.

Validation and measured results

In a reproducible 1280×800 synthetic fast-pan test running through the Vulkan engine, generated-frame PSNR improved by 3.68–4.20 dB. Mean absolute RGB error was approximately 80% lower; at the first ×3 intermediate phase, SSIM increased from 0.965 to 0.987.

The normal and low-source-FPS software test suites passed, including checks for both ×3 intermediate frames, ×2, static content, HUD preservation, scene cuts and HDR. CI also passed scheduler tests, Decky tests, the Gamescope build and a runtime compatibility check against an older glibc.

These results describe the tested synthetic scenes. They are not measurements of performance or image quality in every game.

Known limitations

Block-shaped artifacts can still appear during complex camera turns below 30 source FPS. Their reduction in the tested difficult scenes was small; this release does not claim to solve that problem.

Steam Deck GPU cost, power consumption, physical display pacing and input-to-photon latency have not been validated on hardware for this release. No additional texture reads does not mean zero additional shader cost.

Installation and rollback

  1. Download GSF-Kedavra-1.5.0.zip and optionally verify it against the supplied SHA256 file.
  2. In Decky Loader, open Settings → Developer → Install Plugin from ZIP.
  3. Install the ZIP, open Kedavra Setup and update/reinstall its Gamescope runtime.
  4. Reboot when prompted and confirm that Setup reports the installed runtime matching the plugin.

Keep your previous working installer and settings backup. To roll back, use Setup to restore the stock Gamescope, reboot, then install the previous plugin and its runtime.

GSF 1.4.0 Kedavra

Choose a tag to compare

@github-actions github-actions released this 22 Oct 17:08
6176caf

GSF Kedavra

GSF 1.4.0 Kedavra "Silk": smoother frame pacing with the same latency and the same GPU work.

Install: Decky → Settings → Developer → Install plugin from zip → GSF-Kedavra-1.4.0.zip, then Setup → Reinstall Gamescope KEDAVRA and reboot (the engine changed). The zip installs over 1.3.x in place. Details: README.

New in 1.4.0: silk pacing

A Deck report showed where the remaining stutter comes from: the game misses its cap now and then (1.7 % of flips were repeats), and about every 14 seconds the scheduler gave latency back by skipping a whole frame. Both show as a hitch: the picture stops for one refresh or jumps by two.

  • Catch-up instead of a skip. Content positions now have eight steps per refresh. When the display is behind (after a late game frame, or when it gives a refresh of latency back), it runs 1.25× fast for a few refreshes (1.5× when far behind) instead of jumping a whole frame.
  • Slow-down instead of a freeze. The scheduler looks a few refreshes ahead at what will be ready: the pairs whose motion analysis will be done by then, and the next game frame where the game's own rhythm puts it. When the picture would run out of frames on the way, it slows down evenly (half speed at most) and catches up once the frame is there.
  • No extra cost. Every refresh is still one warp or one copy, as in 1.3.x: no extra GPU work and no extra power. The steady latency is the same or lower.
  • Pacing analysis counts hitches. The trace analysis now reports motion (even steps, freezes, jumps, speed-ups, slow-downs) and hitches_per_min. The engine status has a new catchup counter.

Scheduler simulation (3 seeds × 5 minutes per case; hitch = a freeze or a jump of the picture):

Case Hitches per minute, 1.3.1 → 1.4.0 Jumps (5 min × 3) Median latency
×2 at 60 Hz, base 30, game near its cap (like the Deck report) 12.0 → 5.3 58 → 0 57.1 → 55.5 ms
×3 at 90 Hz, base 30, game near its cap 22.2 → 12.8 92 → 8 50.7 → 49.2 ms
×2 at 90 Hz, base 45, game near its cap 19.8 → 11.1 78 → 6 37.8 → 37.7 ms
×3 at 90 Hz, game spikes 21.1 → 11.0 79 → 0 46.7 → 46.7 ms
×3 at 90 Hz, stable game 0.8 → 0.4 6 → 0 46.7 → 46.7 ms

Low base FPS (economy modes) gains the most:

Case Hitches per minute, 1.3.1 → 1.4.0 Median latency
×3.75 at 90 Hz, base 24, game near its cap 97.7 → 32.6 61.6 → 53.2 ms
×3.75 at 90 Hz, base 24, game with room 18.7 → 10.7 58.1 → 58.2 ms
×3 at 90 Hz, base 30, game with room 65.9 → 15.0 42.6 → 42.6 ms
×2.5 at 60 Hz, base 24, game near its cap 32.9 → 8.3 61.8 → 61.8 ms
×3 at 60 Hz, base 20, game near its cap 17.3 → 16.0 74.4 → 74.0 ms

At base 20 most of what is left are long game stalls (over 100 ms), which no pacing can hide.

The freezes that are left are game frames that arrive far too late to cover: no frame, nothing to move to. All 16 pacing scenarios (3 seeds) and the unit tests pass.

GSF 1.3.1 Kedavra

Choose a tag to compare

@github-actions github-actions released this 22 Oct 09:40
c87e3ae

GSF Kedavra

GSF 1.3.1 Kedavra: frame generation in Gaming Mode, inside a patched Gamescope, controlled from Decky. Now for the Steam Machine on a TV as well as the Steam Deck. No Lossless Scaling. No launch command.

Install: Decky → Settings → Developer → Install plugin from zip → GSF-Kedavra-1.3.1.zip, then Setup → Reinstall Gamescope KEDAVRA and reboot (the engine changed). The zip installs over 1.3.0 in place. Details: README.

New in 1.3.1: Steam Machine

A Steam Machine usually drives a TV at 60 Hz (some at 120 Hz), and its games run at 1080p by default, at 4K in lighter games. 1.3.1 brings the Deck engine there without spending the bigger GPU on it.

  • Motion analysis capped by size. Up to 1536×960 pixels (1.44× the Deck's screen) the analysis runs at the game's own size, as before. Above that, it runs on the frame scaled down to that size. Only the final warp runs at the game's own resolution. On the Deck nothing changes: at 1280×800 the output is identical to 1.3.0, bit for bit.

    Cost per pair of game frames at ×3, lavapipe (a CPU Vulkan driver, objects scene), relative to the Deck:

    Game resolution Analysis capped (1.3.1) Full-size analysis
    1280×800 (Deck) 1.0× 1.0×
    1920×1080 1.37× 1.8×
    3840×2160 1.74× (most of it is the 4K warp) 6.9×

    Image at 1080p against a full-size analysis (same scenes, same motion, lavapipe):

    • at ×3, the camera's worst frame is 24.0 dB against 23.9, and occlusion 34.9 against 34.2 dB;
    • slow pans are 32.6 against 34.2 dB;
    • at base 24, most scenes are better (near object 29.7 against 27.4 dB, dolly 29.2 against 28.4 dB) and slow pans are equal.

    At the Deck's own size (1280×800) slow pans on a 1080p TV lost 4.5 dB. That is why the cap is 1.44× the Deck and not 1×.

  • Auto on a TV. Auto never goes above ×3:

    • 60 Hz: the game is held at 30 and doubled (×2). It no longer drops to base 20 at ×3.
    • 120 Hz: the cap goes from 40 to 60 (×3 to ×2).

    Allow ×2.5 at 60 Hz lets Auto hold a game that cannot keep 30 at 24. ×4 at 120 Hz (base 30) is not recommended. It is off and lives in Setup → Allow ×4 in Auto. On the Deck, Allow ×3.75 stays where it was.

  • VRR is paused while generation is on. Generated frames need a fixed refresh. With Auto or Custom selected, Gamescope keeps adaptive sync off, and the panel says so. Off gives VRR back. It switches with the mode, not per game, because each switch is a modeset (a black screen on many TVs).

  • Panel and HUD follow the screen.

    • The Custom slider goes up to the display's refresh rate.
    • The low-base toggle names the real multiplier for this refresh rate.
    • The in-game HUD keeps its share of the screen on a 1080p or 4K TV.
  • Status: analysis_width, analysis_height, vrr_held and output_height are new fields. kfg_analysis_pixels (0 = full size) and kfg_hold_vrr are new engine options for testing.

Status

Experimental. Nothing in 1.3.1 has run on a Steam Machine: there is no hardware here, and every number above is software (lavapipe). The Steam Machine ships SteamOS with Gamescope and Decky works there, so the plugin installs the same way, but that is unconfirmed too. On the Deck the engine output is unchanged from 1.3.0. To check on a Steam Machine:

  1. Play a game at 60 Hz in Auto, with HDR on and with HDR off.
  2. Record a trace in Pace and press Save report to Desktop.
  3. Send gpu_us_avg, gpu_analysis_us_avg and analysis_width from the report.

GSF 1.3.1 Kedavra: генерация кадров в игровом режиме внутри пропатченного Gamescope, управление из Decky. Теперь не только для Steam Deck, но и для Steam Machine с телевизором. Без Lossless Scaling. Без команд запуска.

Установка: Decky → Настройки → Разработчик → Установить плагин из zip → GSF-Kedavra-1.3.1.zip, затем Setup → Reinstall Gamescope KEDAVRA и перезагрузка (движок изменился). Zip ставится поверх 1.3.0. Подробно: README.

Что нового в 1.3.1: Steam Machine

Steam Machine обычно подключена к телевизору на 60 Гц (иногда 120 Гц). Игры по умолчанию идут в 1080p, лёгкие в 4K. 1.3.1 переносит туда движок Deck и не тратит на него большой GPU.

  • Анализ движения ограничен по размеру. До 1536×960 пикселей (1,44 экрана Deck) анализ, как и раньше, идёт в собственном разрешении игры. Выше этого он идёт по кадру, уменьшенному до этого размера. В полном разрешении игры работает только финальный warp. На Deck ничего не меняется: в 1280×800 результат совпадает с 1.3.0 бит в бит.

    Стоимость пары кадров игры на ×3, lavapipe (Vulkan на процессоре, сцена objects), относительно Deck:

    Разрешение игры Анализ ограничен (1.3.1) Полный анализ
    1280×800 (Deck) 1,0× 1,0×
    1920×1080 1,37× 1,8×
    3840×2160 1,74× (в основном это warp в 4K) 6,9×

    Картинка в 1080p по сравнению с полным анализом (те же сцены и то же движение, lavapipe):

    • на ×3 худший кадр камеры 24,0 дБ против 23,9 дБ, перекрытие 34,9 против 34,2 дБ;
    • медленные панорамы 32,6 против 34,2 дБ;
    • на базе 24 большинство сцен лучше (near object 29,7 против 27,4 дБ, dolly 29,2 против 28,4 дБ), медленные панорамы одинаковые.

    При размере Deck (1280×800) медленные панорамы на телевизоре 1080p теряли 4,5 дБ. Поэтому предел взят 1,44 экрана Deck, а не 1.

  • Auto на телевизоре. Auto никогда не поднимается выше ×3:

    • 60 Гц: игра держится на 30 и удваивается (×2). Она больше не падает до базы 20 на ×3.
    • 120 Гц: лимит от 40 до 60 (от ×3 до ×2).

    Allow ×2.5 на 60 Гц позволяет Auto держать на 24 игру, которая не выдерживает 30. ×4 на 120 Гц (база 30) не рекомендуется. Он выключен и находится в Setup → Allow ×4 in Auto. На Deck Allow ×3.75 остался на прежнем месте.

  • VRR на паузе, пока работает генерация. Сгенерированным кадрам нужна постоянная частота. Когда выбран Auto или Custom, Gamescope держит адаптивную синхронизацию выключенной, и панель об этом пишет. Off возвращает VRR. Переключение идёт вместе с режимом, а не для каждой игры, потому что каждое переключение это modeset (на многих телевизорах чёрный экран).

  • Панель и HUD подстраиваются под экран.

    • Ползунок Custom доходит до частоты экрана.
    • Переключатель низкой базы показывает настоящий множитель для этой частоты.
    • HUD в игре занимает ту же долю экрана на телевизоре 1080p или 4K.
  • Статус: новые поля analysis_width, analysis_height, vrr_held и output_height. Новые настройки движка для проверки: kfg_analysis_pixels (0 = полный размер) и kfg_hold_vrr.

Статус

Экспериментально. Ничего из 1.3.1 не запускалось на Steam Machine: железа здесь нет, все цифры выше программные (lavapipe). Steam Machine поставляется со SteamOS и Gamescope, Decky там работает, поэтому плагин ставится так же, но это тоже не подтверждено. На Deck результат движка не изменился по сравнению с 1.3.0. Как проверить на Steam Machine:

  1. Сыграйте в игру на 60 Гц в Auto, с HDR и без него.
  2. Запишите трассу в Pace и нажмите Save report to Desktop.
  3. Пришлите из отчёта gpu_us_avg, gpu_analysis_us_avg и analysis_width.

GSF 1.3.0 Kedavra

Choose a tag to compare

@github-actions github-actions released this 22 Oct 03:58
3a6b114

GSF Kedavra

GSF 1.3.0 Kedavra: frame generation for Steam Deck in Gaming Mode, inside a patched Gamescope, controlled from Decky. No Lossless Scaling. No launch command.

Install: Decky → Settings → Developer → Install plugin from zip → GSF-Kedavra-1.3.0.zip, then Setup → Reinstall Gamescope KEDAVRA and reboot (the engine changed). The zip installs over 1.2.0 in place. Details: README.

New in 1.3.0

This release is about GPU cost. On a Deck, Witcher 3 at ×3.75 lost 10 to 16 game FPS to generation and filters. 1.3.0 does the expensive work once per pair of game frames instead of once per shown frame.

  • Motion choice once per pair. Choosing the vector for each block and repairing the field (select and two repair passes) used to run for every generated frame. It now runs once per pair, at the middle of the pair, at the end of the analysis. Each generated frame only projects that field to its own phase and warps. On lavapipe (objects scene, median of two runs, software timings):
    • per generated frame: 56 → 5.4 ms (about 10× less)
    • analysis per pair: 90 → 144 ms (it now holds the choice)
    • whole pair at ×3: about 202 → 155 ms (−23%); at ×3.75: about 244 → 159 ms (−35%)
  • Filters once per game frame. Colour looks, sharpening and SMAA are now applied while the game frame is copied in, so a shown real or generated frame does not filter again. At ×3 that is three times fewer filter passes; at ×3.75, 3.75 times fewer. This matters to GFG Extreme, which uses the filters; they stay off in the standalone plugin.
  • Shorter motion search. The search stops early on a near-perfect candidate (a predictor, zero motion or the previous pair's vector), and the warp skips a frame whose weight is zero.
  • Image. The quality is about the same as 1.2.0, not better overall. At ×3 every test scene stays within ±0.6 dB (camera sequence 27.8 dB on average, worst frame 23.3 → 23.2 dB). At ×3.75 (base 24), the two hardest scenes for moving objects improve: dolly 28.9 → 30.2 dB, near object 29.1 → 30.1 dB, occlusion 33.5 → 34.3 dB; dolly over stairs drops 19.9 → 19.4 dB, and the camera's worst frame 23.1 → 22.9 dB.
  • Analysis time in the status. gpu_analysis_us_avg is the GPU time of one pair's analysis. gpu_us never counted it, so the real cost on a Deck could not be read before.

Status

Experimental. All numbers above are software results (lavapipe, a CPU Vulkan driver); the Deck's GPU will give other absolute times. Nothing in 1.3.0 has been measured on a Steam Deck yet, including whether the game gets back its FPS. To check: the same game and mode as before, base FPS with and without generation, gpu_us_avg and gpu_analysis_us_avg from Save report to Desktop, and a Pace trace.


GSF 1.3.0 Kedavra: генерация кадров для Steam Deck в игровом режиме внутри пропатченного Gamescope, управление из Decky. Без Lossless Scaling. Без команд запуска.

Установка: Decky → Настройки → Разработчик → Установить плагин из zip → GSF-Kedavra-1.3.0.zip, затем Setup → Reinstall Gamescope KEDAVRA и перезагрузка (движок изменился). Zip ставится поверх 1.2.0. Подробно: README.

Что нового в 1.3.0

Этот выпуск про нагрузку на GPU. На Deck Witcher 3 на ×3,75 терял от 10 до 16 FPS игры из-за генерации и фильтров. В 1.3.0 дорогая работа делается один раз на пару кадров игры, а не на каждый показанный кадр.

  • Выбор движения один раз на пару. Выбор вектора для каждого блока и починка поля (select и два прохода repair) раньше шли на каждый сгенерированный кадр. Теперь они идут один раз на пару, для середины пары, в конце анализа. Каждый сгенерированный кадр только переносит это поле на свою фазу и делает warp. На lavapipe (сцена objects, медиана двух прогонов, программные замеры):
    • на сгенерированный кадр: 56 → 5,4 мс (примерно в 10 раз меньше)
    • анализ на пару: 90 → 144 мс (теперь в нём выбор)
    • вся пара на ×3: примерно 202 → 155 мс (−23%), на ×3,75: примерно 244 → 159 мс (−35%)
  • Фильтры один раз на кадр игры. Цветовые профили, резкость и SMAA теперь накладываются при копировании кадра игры, и показанный настоящий или сгенерированный кадр второй раз не фильтруется. На ×3 проходов фильтров в три раза меньше, на ×3,75 в 3,75 раза. Это важно для GFG Extreme, который использует фильтры; в отдельном плагине они выключены.
  • Короче поиск движения. Поиск останавливается раньше на почти идеальном кандидате (предсказание, нулевое движение или вектор прошлой пары), а warp пропускает кадр с нулевым весом.
  • Картинка. Качество примерно как в 1.2.0, в целом не лучше. На ×3 все тестовые сцены в пределах ±0,6 дБ (последовательность камеры в среднем 27,8 дБ, худший кадр 23,3 → 23,2 дБ). На ×3,75 (база 24) две самые тяжёлые сцены с движущимися объектами стали лучше: dolly 28,9 → 30,2 дБ, near object 29,1 → 30,1 дБ, occlusion 33,5 → 34,3 дБ; dolly над лестницей 19,9 → 19,4 дБ, худший кадр камеры 23,1 → 22,9 дБ.
  • Время анализа в статусе. gpu_analysis_us_avg показывает время анализа одной пары на GPU. gpu_us его никогда не учитывал, поэтому реальную нагрузку на Deck раньше нельзя было увидеть.

Статус

Экспериментально. Все цифры выше получены программно (lavapipe, Vulkan на процессоре); на GPU Deck абсолютные времена будут другими. Ничего из 1.3.0 на Steam Deck ещё не измерено, в том числе возвращает ли игра свои FPS. Как проверить: та же игра и тот же режим, что раньше, базовый FPS с генерацией и без, gpu_us_avg и gpu_analysis_us_avg из Save report to Desktop и трасса в Pace.

GSF 1.2.0 Kedavra

Choose a tag to compare

@github-actions github-actions released this 22 Oct 02:15
89cacf2

GSF Kedavra

GSF 1.2.0 Kedavra: frame generation for Steam Deck in Gaming Mode, inside a patched Gamescope, controlled from Decky. No Lossless Scaling. No launch command.

Install: Decky → Settings → Developer → Install plugin from zip → GSF-Kedavra-1.2.0.zip, then Setup → Reinstall Gamescope KEDAVRA and reboot (the engine changed). The zip installs over 1.1.x in place. Details: README.

New in 1.2.0

This release brings the engine that GFG Extreme 3.0 runs to the standalone plugin, plus a fix for fast turns at a low base frame rate.

  • HDR games. HDR10 (PQ) and scRGB games get generation in float frames. Before, they passed through without it. HDR games in the panel turns this off. Tested in software only, not on a Deck.

  • ×3.75 in Auto. With Allow ×3.75 on, Auto may hold a game that cannot keep 30 at 24 FPS (90 Hz: ×3.75, each real frame shown 3 or 4 refreshes) instead of repeating frames. The panel and the HUD show ×3.75 exactly. It is off by default: at 24 the image is stretched further, so artifacts are easier to see.

  • Fast turns at a low base frame rate. At base 24 a fast camera turn moves 150 px and more per game frame, past the reach of the motion search. The engine then lost the camera and fell back to repeated frames, or produced broken frames. The coarse search now follows the camera motion of the previous frame pair. In a test sequence turning at up to 175 px per game frame (base 24, phases of ×3.75, lavapipe):

    • worst frame 14.1 → 23.1 dB
    • frame-to-frame change error 14.5 → 5.9
    • false scene cuts 3 → 0

    At ×3 the camera sequence is practically unchanged (27.8 dB on average; worst frame 23.5 → 23.3 dB). The other test scenes give the same numbers as 1.1.2.

  • Engine from GFG Extreme 3.0: filters after generation (kfg_post, used by GFG; off in the standalone plugin), a policy revision in the status, and Gamescope built with the /usr prefix so SteamOS's display scripts load. That last one is what makes HDR available on the Deck OLED.

  • The plugin restores its engine options after Gamescope restarts.

  • Repair fix (found in an external review): the block repair pass read one neighbour before loading it, so whether a block got repaired depended on an undefined value. All nine neighbours are now loaded first. Threshold checks in warp and repair compare squared lengths instead of taking square roots. lavapipe: within ±0.02 dB in every scene.

  • Stricter Pace analysis. Refresh slots with no flip row at all now count as missed. Before, only rows marked "repeat" did. Long unexplained gaps are no longer dropped from the statistics. As a result, older traces score worse than before: the best Auto trace goes from 0.27% to an estimated 0.78% missed.

Status

Experimental. The image numbers are software results (lavapipe). Nothing in 1.2.0 has been measured on a Steam Deck yet, HDR included. Record a trace in Pace with ×3 and with ×3.75, then send them with Save report to Desktop.


GSF 1.2.0 Kedavra: генерация кадров для Steam Deck в игровом режиме внутри пропатченного Gamescope, управление из Decky. Без Lossless Scaling. Без команд запуска.

Установка: Decky → Настройки → Разработчик → Установить плагин из zip → GSF-Kedavra-1.2.0.zip, затем Setup → Reinstall Gamescope KEDAVRA и перезагрузка (движок изменился). Zip ставится поверх 1.1.x. Подробно: README.

Что нового в 1.2.0

В этом выпуске отдельный плагин получает движок, на котором работает GFG Extreme 3.0, и исправление быстрых поворотов при низком базовом FPS.

  • HDR-игры. Игры в HDR10 (PQ) и scRGB теперь получают генерацию в кадрах с плавающей точкой. Раньше они проходили без неё. Переключатель HDR games на панели её выключает. Проверено только программно, на Deck не проверялось.

  • ×3,75 в Auto. Если включить Allow ×3.75, Auto может держать игру, которая не выдерживает 30, на 24 FPS (90 Гц: ×3,75, каждый настоящий кадр показывается 3 или 4 обновления), вместо повторов кадров. Панель и HUD показывают ровно ×3,75. По умолчанию выключено: на 24 картинка растягивается сильнее, и артефакты заметнее.

  • Быстрые повороты при низком базовом FPS. На базе 24 быстрый поворот камеры сдвигает картинку на 150 px и больше за кадр игры, дальше, чем доставал поиск движения. Движок терял камеру и откатывался к повторам кадров или выдавал сломанные кадры. Теперь грубый поиск следует за движением камеры из прошлой пары кадров. Тестовая последовательность с поворотом до 175 px за кадр игры (база 24, фазы ×3,75, lavapipe):

    • худший кадр 14,1 → 23,1 дБ
    • ошибка изменения между кадрами 14,5 → 5,9
    • ложных смен сцены 3 → 0

    На ×3 последовательность камеры практически не изменилась (в среднем 27,8 дБ, худший кадр 23,5 → 23,3 дБ). Остальные тестовые сцены дают те же цифры, что и 1.1.2.

  • Движок из GFG Extreme 3.0:

    • фильтры после генерации (kfg_post: их использует GFG, в отдельном плагине они выключены);
    • номер политики в статусе;
    • сборка Gamescope с префиксом /usr, чтобы загружались скрипты дисплея SteamOS. Именно это открывает HDR на Deck OLED.
  • После перезапуска Gamescope плагин восстанавливает свои настройки движка.

  • Исправление repair (найдено при внешнем ревью): проход починки блоков читал одного соседа до того, как загрузил его, поэтому решение чинить блок или нет зависело от неопределённого значения. Теперь сначала загружаются все девять соседей. Пороговые проверки в warp и repair сравнивают квадраты длин вместо корней. На lavapipe отличие в пределах ±0,02 дБ во всех сценах.

  • Строже анализ Pace. Слоты обновления, для которых нет строки flip, теперь считаются пропущенными. Раньше считались только строки с пометкой «repeat». Длинные необъяснённые паузы больше не выбрасываются из статистики. Поэтому старые трассы оцениваются хуже, чем раньше: лучшая трасса Auto вместо 0,27% получает оценку 0,78% пропусков.

Статус

Экспериментально. Цифры качества получены программно на lavapipe. Ничего из 1.2.0 на Steam Deck ещё не измерено, включая HDR. Запишите трассу в Pace с ×3 и с ×3,75 и пришлите их через Save report to Desktop.

GSF 1.1.2 Kedavra

Choose a tag to compare

@wavessevaw wavessevaw released this 21 Oct 12:57
42924cc

GSF Kedavra

GSF 1.1.2 Kedavra: x2 / x3 frame generation for Steam Deck in Gaming Mode, inside a patched Gamescope, controlled from Decky. No Lossless Scaling. No launch command.

Install: Decky → Settings → Developer → Install plugin from zip → GSF-Kedavra-1.1.2.zip, then Setup → Reinstall Gamescope KEDAVRA and reboot. The zip installs over 1.1.1 in place. Details: README.

New in 1.1.2

  • Cheaper motion analysis. Each block's own window is read once per search instead of once per candidate, which halves the texture reads of the search. On lavapipe the analysis of a frame pair goes from about 154 to 85–91 ms (old and new run alternately). The generated images are unchanged in every test scene. Less GPU time per pair should mean fewer repeats from late analysis (me_late), but that is not measured on the Deck yet.
  • Cheaper vector selection. Identical candidates are matched once, and the sub-pixel fit reuses the best candidate's error: 10–15% less on lavapipe, same images (within 0.01 dB).
  • Stopping a Pace recording. It has restarted Gaming Mode on the Deck. The trace file is now opened, written and closed on its own thread, so stopping a trace does no file work in the compositor, and the plugin analyses the trace only after Gamescope has closed it. The cause of the restart is not confirmed, so this may not be the fix.
  • Crash records in the report. Save report to Desktop now includes coredumpctl (the stack of a crashed Gamescope or Steam) and the kernel GPU log, so a crash can be told apart from a GPU reset. If Gaming Mode restarts, save a report right after and send it.

Status

Experimental. The speed-ups are measured on lavapipe (software Vulkan). They have not been measured on a Steam Deck. Run Setup → GPU benchmark and record a trace in Pace, then send them with Save report to Desktop.


GSF 1.1.2 Kedavra: генерация кадров x2 / x3 для Steam Deck в игровом режиме внутри пропатченного Gamescope, управление из Decky. Без Lossless Scaling. Без команд запуска.

Установка: Decky → Настройки → Разработчик → Установить плагин из zip → GSF-Kedavra-1.1.2.zip, затем Setup → Reinstall Gamescope KEDAVRA и перезагрузка. Zip ставится поверх 1.1.1. Подробно: README.

Что нового в 1.1.2

  • Анализ движения дешевле. Окно самого блока читается один раз на поиск, а не для каждого кандидата, так что выборок из текстур при поиске вдвое меньше. На lavapipe анализ пары кадров занимает 85–91 мс вместо примерно 154 (старая и новая версии запускались попеременно). Сгенерированные кадры во всех тестовых сценах не изменились. Меньше времени GPU на пару должно означать меньше повторов из-за опоздавшего анализа (me_late), но на Deck это ещё не измерено.
  • Выбор вектора дешевле. Одинаковые кандидаты сравниваются один раз, а субпиксельное уточнение берёт уже посчитанную ошибку лучшего кандидата: на lavapipe это на 10–15% меньше работы, картинка та же (в пределах 0,01 дБ).
  • Остановка записи в Pace. Раньше она перезапускала игровой режим на Deck. Теперь файл трассы открывается, пишется и закрывается в отдельном потоке, и композитор при остановке с файлом не работает. Плагин разбирает трассу только после того, как Gamescope её закрыл. Причина перезапуска не подтверждена, так что это может быть ещё не исправление.
  • Записи о падениях в отчёте. Save report to Desktop теперь добавляет coredumpctl (стек упавшего Gamescope или Steam) и журнал GPU из ядра, чтобы падение можно было отличить от сброса GPU. Если игровой режим перезапустится, сразу сохраните отчёт и пришлите его.

Статус

Экспериментально. Ускорение измерено на lavapipe (программный Vulkan). На Steam Deck оно не измерено. Запустите Setup → GPU benchmark и запишите трассу в Pace, затем пришлите всё через Save report to Desktop.