Repository navigation
GSF 1.3.0 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_avgis the GPU time of one pair's analysis.gpu_usnever 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_avgпоказывает время анализа одной пары на GPU.gpu_usего 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.
