vulkan: initialize presentation upscaling independently of frame generation #417

Merged

monsterlag merged 3 commits from fix/upscale-without-framegen into main

2026-11-08 05:20:57 +00:00

monsterlag commented

2026-11-08 00:24:31 +00:00

(Migrated from github.com)

The scaler has independent initialization/failure state and rejects invalid crop/destination sizes. Partial allocation failure clears stale size metadata and falls back to the existing blit. No renderer API, aspect policy, decoder or game-generation change is included.

The branch normally merges upstream e40306e. It preserves all three upstream queue flushes before interpolation/scaler resize waits, including locked flushes before both scaler idle waits.

Validation:

Actual production initializer/scaler functions compile with authored API spies. Catmull, Lanczos, sharpness-only, both frame-generation environment states, all-off, cached initialization failure/cleanup, allocation failure/recovery and crop clamping/recovery pass. The spies check locked flush ordering before resize waits. These checks also passed with the new official bundled GNU Clang23.1.5 toolchain.
The original production scaler fails the first FG-off dispatch assertion at runtime. This negative was local; Linux CI runs both environment configurations against the fix.
Existing frame-generation-off gate checks pass. Direct-view forwarding/age bookkeeping has an explicitly source-based wiring check; the spy fixture does not establish GPU lifetime safety.
All four CI jobs pass at ad725a9: Windows/Linux source checks, SPIR-V fixture and Linux AppImage renderer build.
A fresh normal lift and complete matching Windows host/API4 renderer build using the no-visual-studio release toolkit passed in 992 seconds with zero failures. This live candidate combines this branch and #205 on e40306e; API4 belongs to #205.
RTX5090/Windows, isolated checkpoint, 100% game resolution: actual 3840x1080 and 5120x1440 windows each display the Kokiri scene/HUD with Catmull and Lanczos, FG off. Lanczos cases explicitly set RX_VK_FRAMEGEN=0. Successful-dispatch diagnostics confirm requested modes and centered 1920x1080/2560x1440 destinations; interpolation does not initialize.
Sharpness-only (mode0/sharpness30/FG off) and Catmull/FG2 also dispatch and display correctly. The 3840x1080 Catmull/FG-off case preserves the scene through actual 5120x2160 fullscreen and back. All six sessions close normally.

Opt into RX_VK_UPSCALE_LOG=1 to log the first successful scale dispatch's mode, sharpness, source, requested crop and destination once.

These are compatibility checks, not native60+FPS evidence: warm gameplay was roughly 48-52FPS at 100% resolution. The first cold run sustained 11.6-11.8FPS after shader jobs completed, recovering to 46-50FPS after F11; that separate readback-wait anomaly remains unexplained and is not retroactively passed.

Physical32:9 fullscreen, the reporter's RTX4070, AMD/Intel/Linux runtime, wider progression and physical audio/controller behavior remain unverified. This fixes the established FG-off bypass; no reported black-screen fix is claimed. The sampled view retains the cached image's existing retirement policy; a later presentation-fence ownership model and aggressive eviction overrides remain outside this change.

Upscaling and Sharpness silently stop working when Frame generation is Off. Forward the direct-present sampled view under the replay lock and initialize the scaler independently, so Catmull-Rom, Lanczos and sharpening work without interpolation's copy/luma work, frame sets or display thread. The scaler has independent initialization/failure state and rejects invalid crop/destination sizes. Partial allocation failure clears stale size metadata and falls back to the existing blit. No renderer API, aspect policy, decoder or game-generation change is included. The branch normally merges upstream e40306e. It preserves all three upstream queue flushes before interpolation/scaler resize waits, including locked flushes before both scaler idle waits. Validation: - Actual production initializer/scaler functions compile with authored API spies. Catmull, Lanczos, sharpness-only, both frame-generation environment states, all-off, cached initialization failure/cleanup, allocation failure/recovery and crop clamping/recovery pass. The spies check locked flush ordering before resize waits. These checks also passed with the new official bundled GNU Clang23.1.5 toolchain. - The original production scaler fails the first FG-off dispatch assertion at runtime. This negative was local; Linux CI runs both environment configurations against the fix. - Existing frame-generation-off gate checks pass. Direct-view forwarding/age bookkeeping has an explicitly source-based wiring check; the spy fixture does not establish GPU lifetime safety. - All four CI jobs pass at ad725a9: Windows/Linux source checks, SPIR-V fixture and Linux AppImage renderer build. - A fresh normal lift and complete matching Windows host/API4 renderer build using the no-visual-studio release toolkit passed in 992 seconds with zero failures. This live candidate combines this branch and #205 on e40306e; API4 belongs to #205. - RTX5090/Windows, isolated checkpoint, 100% game resolution: actual 3840x1080 and 5120x1440 windows each display the Kokiri scene/HUD with Catmull and Lanczos, FG off. Lanczos cases explicitly set RX_VK_FRAMEGEN=0. Successful-dispatch diagnostics confirm requested modes and centered 1920x1080/2560x1440 destinations; interpolation does not initialize. - Sharpness-only (mode0/sharpness30/FG off) and Catmull/FG2 also dispatch and display correctly. The 3840x1080 Catmull/FG-off case preserves the scene through actual 5120x2160 fullscreen and back. All six sessions close normally. Opt into RX_VK_UPSCALE_LOG=1 to log the first successful scale dispatch's mode, sharpness, source, requested crop and destination once. These are compatibility checks, not native60+FPS evidence: warm gameplay was roughly 48-52FPS at 100% resolution. The first cold run sustained 11.6-11.8FPS after shader jobs completed, recovering to 46-50FPS after F11; that separate readback-wait anomaly remains unexplained and is not retroactively passed. Physical32:9 fullscreen, the reporter's RTX4070, AMD/Intel/Linux runtime, wider progression and physical audio/controller behavior remain unverified. This fixes the established FG-off bypass; no reported black-screen fix is claimed. The sampled view retains the cached image's existing retirement policy; a later presentation-fence ownership model and aggressive eviction overrides remain outside this change.

Commenting is not possible because the repository is archived.

No reviewers