Add an opt-in native scene base scale #578

Closed

oscarrc91 wants to merge 5 commits from fix/native-scene-base-scale into main

oscarrc91 commented

2026-11-08 12:43:25 +00:00

(Migrated from github.com)

Fixed Resolution 100% can still produce a 2048x1152 scene from a 2560x1440 base because graphics mode 1 has a separate 0.8 base multiplier on each axis. Add the default-off HOS_SCENE_NATIVE=1 experiment to replace that one entry with 1.0 and request the engine's original resource-refresh path.

This is an experimental, opt-in quality option, separate from FSR in #577. It relates to #68 and addresses a different layer from the stale fixed-scale cache in #329.

Dependency: deterministic fixed 1440p currently also needs #329 (HOS_FIX_RES=1), or an equivalent upstream correction. This branch changes the base table; it does not fix the separately cached scale. A comparison run without #329 confirmed the table edits but still rendered 1920x1080 into 2560x1440. Its implementation is not duplicated here. Please coordinate these changes before merging; FSR itself is not required for the base-scale option.

Scope

Patch the configuration-update function in the user's locally lifted code after verifying the complete stock (1.0, 0.8, 2/3) table in each of the two profiles.
Change only mode 1's scale. Preserve the selected mode and run the original assignment, dirty flags and callbacks once when the table changes. Unknown/already-patched tables stay untouched.
Fail before any write on unexpected/duplicate translated-code layouts. Add a game-free compiled fixture, optional local generated-code verification, CI wiring and documentation.
Do not force a display size, antialiasing, texture filters or frame-generation settings. No lifted game source is committed.

Validation

Gameplay verified on 2026-11-08: isolated integration with #577 and the existing #329 commit on upstream 3f5b0aa (source fingerprint 9882ef6b). Windows, RX 9070 XT, driver 32.0.32015.7029, Ryzen 7 7800X3D; Kokiri Forest near Saria, fixed Resolution 100%, 16:9, handheld/ultrawide and frame generation off.
Both base-table edits executed. The native network report shows 2560x1440 input/output, and FSR 4.1.2 completed at least 2,520 dispatches guarded by matching color/depth/output dimensions. No FSR SDK error, dispatch failure or device-lost entry appeared in the captured log. The player approved the image and stability.
Compiled checks pass for the disabled path, both profiles, complete-table guards, all three modes, one-time refresh, normal later transitions, idempotence and atomic rejection of duplicate/unexpected layouts. The same checks also pass against mode-selection/notification blocks from the local owned v131072 lift. The text-speed regression passes.
Updated to upstream ddba87c: resolved the host-global declaration conflict while retaining the new upstream occlusion option. The merged Windows host passes a syntax compilation, and native-scale/text-speed regressions pass. These later upstream changes have not had another gameplay run.

The #329 author/commit is preserved in private integration history and is not included in this PR's diff. GitHub Actions for the fork contribution awaits maintainer approval; local results do not imply upstream CI has passed.

Rendering 2560x1440 instead of 2048x1152 adds 56.25% more pixels and increases GPU/memory work. The verified combination's 19 gameplay presentation samples had a median of 52.2 FPS (47.4-54.8), with frame generation off. This is not a controlled before/after benchmark of this hook alone.

Remaining checks

Exercise menu/area/mode transitions, resolution percentages, handheld/ultrawide and other hardware. A table-patched log alone does not establish actual scene dimensions. The #329 proposal documents unresolved lower-percentage issues; the verified setup targets fixed 100% only. Keep the feature off by default and coordinate with #68/#329 before extending the scope.

AI assistance: OpenAI Codex; recorded in commit trailers. No game data, lifted sources, saves, caches or binaries are included.

Fixed Resolution 100% can still produce a 2048x1152 scene from a 2560x1440 base because graphics mode 1 has a separate **0.8 base multiplier on each axis**. Add the default-off `HOS_SCENE_NATIVE=1` experiment to replace that one entry with 1.0 and request the engine's original resource-refresh path. This is an **experimental, opt-in quality option**, separate from FSR in #577. It relates to #68 and addresses a different layer from the stale fixed-scale cache in #329. **Dependency:** deterministic fixed 1440p currently also needs #329 (`HOS_FIX_RES=1`), or an equivalent upstream correction. This branch changes the base table; it does not fix the separately cached scale. A comparison run without #329 confirmed the table edits but still rendered 1920x1080 into 2560x1440. Its implementation is not duplicated here. Please coordinate these changes before merging; FSR itself is not required for the base-scale option. ### Scope - Patch the configuration-update function in the user's locally lifted code after verifying the complete stock (1.0, 0.8, 2/3) table in each of the two profiles. - Change only mode 1's scale. Preserve the selected mode and run the original assignment, dirty flags and callbacks once when the table changes. Unknown/already-patched tables stay untouched. - Fail before any write on unexpected/duplicate translated-code layouts. Add a game-free compiled fixture, optional local generated-code verification, CI wiring and documentation. - Do not force a display size, antialiasing, texture filters or frame-generation settings. No lifted game source is committed. ### Validation - **Gameplay verified on 2026-11-08:** isolated integration with #577 and the existing #329 commit on upstream `3f5b0aa` (source fingerprint `9882ef6b`). Windows, RX 9070 XT, driver 32.0.32015.7029, Ryzen 7 7800X3D; Kokiri Forest near Saria, fixed Resolution 100%, 16:9, handheld/ultrawide and frame generation off. - Both base-table edits executed. The native network report shows **2560x1440 input/output**, and FSR 4.1.2 completed at least **2,520 dispatches** guarded by matching color/depth/output dimensions. No FSR SDK error, dispatch failure or device-lost entry appeared in the captured log. The player approved the image and stability. - Compiled checks pass for the disabled path, both profiles, complete-table guards, all three modes, one-time refresh, normal later transitions, idempotence and atomic rejection of duplicate/unexpected layouts. The same checks also pass against mode-selection/notification blocks from the local owned v131072 lift. The text-speed regression passes. - **Updated to upstream `ddba87c`:** resolved the host-global declaration conflict while retaining the new upstream occlusion option. The merged Windows host passes a syntax compilation, and native-scale/text-speed regressions pass. These later upstream changes have not had another gameplay run. The #329 author/commit is preserved in private integration history and is not included in this PR's diff. GitHub Actions for the fork contribution awaits maintainer approval; local results do not imply upstream CI has passed. Rendering 2560x1440 instead of 2048x1152 adds 56.25% more pixels and increases GPU/memory work. The verified combination's 19 gameplay presentation samples had a median of 52.2 FPS (47.4-54.8), with frame generation off. This is not a controlled before/after benchmark of this hook alone. ### Remaining checks Exercise menu/area/mode transitions, resolution percentages, handheld/ultrawide and other hardware. A table-patched log alone does not establish actual scene dimensions. The #329 proposal documents unresolved lower-percentage issues; the verified setup targets fixed 100% only. Keep the feature off by default and coordinate with #68/#329 before extending the scope. AI assistance: OpenAI Codex; recorded in commit trailers. No game data, lifted sources, saves, caches or binaries are included.

NotRexed commented

2026-11-09 05:29:48 +00:00

(Migrated from github.com)

Thank you! #849 (merged) now sets the graphics scale to x1 for a chosen output, which covers the 0.8 base multiplier this changed, so I'm closing it as superseded. If current main still renders below the output at 100%, please open a new PR on top of it.

Commenting is not possible because the repository is archived.

No reviewers