WIP: feat(vk): add support for FSR 4 "Lite" model #1989

Draft
prfd wants to merge 1 commit from prfd/Longshot:feat/fsr4-lite into main
First-time contributor

Summary

AMD's FSR 4 library ships four networks (FP8, full INT8, the Navi 2 INT8 model and Lite). FSR 4's SDK doesn't expose a
model selector: its provider walk calls each network's IsSupported, which answers yes only for the GPU's native model.

This scans AMD's loaded Vulkan DLL for the four provider classes, checks their 4.1.2 vtable shape, and patches only
each provider's IsSupported entry (vtable[2]) in this process's copy, under VirtualProtect: Auto still calls
AMD's original, Lite answers yes only for the Lite provider.

A Graphics-tab FSR 4 model row (Auto or Lite) and RX_VK_FSR4_MODEL choose it for one run. The model joins the
upscaler context's key, so a change builds one new context at the next frame, and a refused Lite context falls back to
Auto once. Auto stays the default and AMD's own selection is unchanged.

Read first: gpu/vk/fsr4_models.cpp (scan_image, install), gpu/vk/vk_fsr4.cpp (fsr4_model,
fsr4_lite_available, the context key) and gpu/vk/menu/upscaler.cpp (the row).

What changed

  • Scanner and hook — gpu/vk/fsr4_models.h/.cpp (new).
    • scan_image() reads a PE image as bytes (no Vulkan, no GPU) and finds the four . ?AVffxProvider_FSR4* classes,
      their Complete Object Locators, vtables and the static provider registry. It checks the 4.1.2 shape: entries 0, 1
      and 3..7 shared, entry 2 distinct; fewer than two providers or a shifted entry fails the check.
    • On Windows, scan_loaded_modules() scans every loaded module at the first FSR 4 use. The probe logs each module's
      providers, shape, registry and the native model, asked of each provider's own IsSupported.
    • install() patches only vtable[2] under VirtualProtect, once per module, and only when the shape is understood,
      a Lite provider is present and at least two providers are found. If entry 2 is not AMD's original it fails closed
      and rolls back. Auto delegates to AMD's original (recording the first yes as the native model); Lite answers true
      only for the Lite provider. Nothing is written to AMD's file.
  • Engine — gpu/vk/vk_fsr4.cpp/.h.
    • At load (Windows) the probe runs and the hook is installed. fsr4_lite_available() is true only for an AMD GPU
      with a hooked Lite provider.
    • Fsr4Model Auto/Lite, RX_VK_FSR4_MODEL=auto|lite (also 0|1) and the menu's live value; the switch wins for the
      run.
    • The context key includes the model, so a change makes one new context at the next frame and logs whether it is
      native or forced, and tuned or AMD's shaders.
    • A forced Lite context on a card whose native model is not Lite suppresses the tuned buffer/shader/pipeline hooks
      (AMD's own shaders); an older Auto context keeps its tuned pipelines, so cmd_dispatch is not suppressed.
    • A refused Lite context or a failed dispatch falls back to Auto once; the saved row stays Lite and the log says why.
  • Menu — the row is a module (gpu/vk/menu/upscaler.cpp, #1633), after DLSS model: Auto (AMD's) and Lite
    (INT8), saved as fsr4_model (gpu/vk/vk_menu.h), with its on_change updating the engine. vk_menu.cpp only reads
    RX_VK_FSR4_MODEL at start and sets the live value. The row is shown only where FSR 4 runs and a hooked Lite provider
    is present, and is locked when the Upscaler is not FSR 4. The help and all 13 language tables carry it.
  • Switches and docs — RX_VK_FSR4_MODEL in gpu/vk/switches.def; the user guide's Graphics row and FSR 4 section
    (docs/user/graphics-and-controls.md, docs/user/advanced.md) and the developer reference
    (docs/dev/reference/switches.md, docs/dev/reference/upscalers.md, "FSR 4 models").
  • Tests — tests/test_fsr4_models.py with tests/fsr4_models_test.cpp (new) covers the scanner on synthetic PE
    images laid out like 4.1.2: the four providers and their order, their Complete Object Locators/vtables, the registry,
    the shape check, a missing Lite, one provider, a shifted shared entry, an unknown provider, a non-PE image and the
    native-model helper; on Windows it also exercises install, Auto/Lite/Auto answers, install twice and a second scan
    (its image's pages are made executable for that call, which Windows does not do for heap pages). tests/test_fsr4_tune.py
    checks the suppression pass-through, that cmd_dispatch is not suppressed, and that prepare rebuilds on a model
    change. tests/test_upscaler_rows.cpp covers the row's values and shown/fixed rules; tests/test_menu_ini_keys.py
    freezes the new key, and tests/test_menu_env_overrides.py and tests/test_hdr_encode.py cover the menu_init and
    make_context(..., want) changes.

Testing

Tested:

  • In the game on Windows, AMD Radeon RX 6600 (RDNA 2), amd_fidelityfx_upscaler_vk.dll 4.1.2.4555 (FSR SDK 2.4), with
    FSR 4 on: the FSR 4 model row selected Auto and Lite in Kokiri Forest, 1440x810 -> 1920x1080, and the FSR 4
    upscaling GPU time was read from longshot_log.txt for both.

  • The game-free suite on WSL2 Ubuntu 24.04.4 LTS (the suite reports linux) at 7dffea2b (on 6e1521a7), on the same
    PC and checkout as the Windows runs (/mnt/d/Longshot is D:\Longshot), python3 -I -B tests/run_all.py: 720
    passed, 70 failed, 115 skipped
    (767 s, 12 at once). The 69 failures below are the same checks with the same error on
    a worktree of clean 6e1521a7, so none comes from this branch; the 70th is byte_compile_tracked, which fails only
    because this working tree holds tests/test_vbytes.py out locally (as tests/test_vbytes.py.skip; the test itself
    runs longer than the runner's 600 s check timeout on this WSL/drvfs checkout) and passes with the file in place.

  • tests/test_fsr4_models.py and the branch's menu checks that run here (test_menu_ini_keys.py,
    test_menu_env_overrides.py, test_hdr_encode.py, test_menu_i18n.py) pass. tests/test_fsr4_tune.py is skipped
    without spirv-val; test_upscaler_rows.py, test_menu_modules.py, test_menu_sections.py,
    test_settings_table.py, test_menu_module_slots.py and test_pads_after_menu.py fail on clean main too.

    The 69 checks that also fail on clean `6e1521a7` (this PC)

    tests/test_appimage_player_build.py, tests/test_bug_report_defaults.py, tests/test_capture_read_once.py,
    tests/test_check_third_party.py, tests/test_crash_dialog.py, tests/test_crash_record.py,
    tests/test_crt_options.py, tests/test_crt_shaders.py, tests/test_damage_debug_menu.py,
    tests/test_dev_ab.py, tests/test_dev_packs.py, tests/test_device_lost_exit.py, tests/test_discover.py,
    tests/test_display_startup_failure.py, tests/test_dlss_feature_size.py,
    tests/test_failed_compile_keeps_objects.py, tests/test_flush_nothing_held.py, tests/test_frame_overlap.py,
    tests/test_gpu_wait_stats.py, tests/test_guest_imported.py, tests/test_high_fps_fixes.py,
    tests/test_host_guest_writes.py, tests/test_image_reaper_threads.py, tests/test_libc_hooks.py,
    tests/test_lift_cache.py, tests/test_lift_lookups.py, tests/test_linux_preflight.py,
    tests/test_linux_pressure.py, tests/test_linux_tools.py, tests/test_linux_update.py,
    tests/test_log_open_failure.py, tests/test_longshot_saves.py, tests/test_make_game_cache.py,
    tests/test_make_game_ngx.py, tests/test_menu_module_slots.py, tests/test_menu_modules.py,
    tests/test_menu_sections.py, tests/test_mesh_conv.py, tests/test_mise_build.py, tests/test_missing_nup.py,
    tests/test_mods_list.py, tests/test_mods_size_table.py, tests/test_ngx_sdk.py,
    tests/test_oot_shadow_fit.py, tests/test_pads_after_menu.py, tests/test_pending_guest_import.py,
    tests/test_pipeline_play_cache.py, tests/test_pipelines_outlive_cache_format.py,
    tests/test_prepare_cache_save.py, tests/test_profiling_mode.py, tests/test_program_cache_cleanup.py,
    tests/test_program_cache_one_variant.py, tests/test_psychov35_fsr3_fg.py,
    tests/test_readback_timeout_source.py, tests/test_reflex_swapchain_lifetime.py,
    tests/test_renderer_new_handler.py, tests/test_renhdrum.py, tests/test_settings_table.py,
    tests/test_slow_frames.py, tests/test_song_note_cue.py, tests/test_sparse_mirror.py,
    tests/test_spirv_pack.py, tests/test_split_giants.py, tests/test_sprint.py, tests/test_thinlto_tls.py,
    tests/test_upscaler_rows.py, tests/test_vcore_escape.py, tests/test_vcore_mask0.py, tests/test_walk.py

Still to test:

  • Other AMD cards (RDNA 3/4): with FSR 4 on, expect the vk: AMD FSR 4: models:.

Measurements

Verdict: Lite is about twice as fast as the default Navi 2 INT8 model on a RX 6600, and it provides superior image quality compared to both FSR 3.1.6 and XeSS-DP4a, specially with hair and moving objects.

Setup: Windows 10, AMD RX 6600, 26.10.1; Kokiri Forest, (75%) 1440x810 -> 1920x1080

Measurement Auto (INT8 Navi 2) Lite (INT8) Δ
upscaling GPU time 5.608 ms 2.899 ms -2.709 ms (-48.3%)

Flags and settings

One new menu row and one new switch; both default to Auto.

Flag What it is Before this PR After this PR Effect Helps / doesn't help / untested Recommendation
fsr4_model (FSR 4 model row) The Graphics tab's network choice, saved in longshot_options.ini - Auto (0) or Lite (1); Auto Auto lets AMD pick for the GPU; Lite forces the Lite provider Helps cards AMD does not natively run Lite on; other cards untested Auto
RX_VK_FSR4_MODEL The switch over the row, for one run - unset; auto or lite (0/1) auto keeps AMD's pick; lite forces the Lite provider and locks the row For A/B runs and testing; Linux has no Lite, so it does nothing there unset (Auto)

Risks

  • The hook is in AMD's process copy. A DLL with the same class names but another vtable layout would be misread; the
    RTTI names and 4.1.2 shape check are the gate, and anything not understood is left alone (no hook, row hidden, Auto).
  • Only IsSupported (vtable entry 2) is replaced, and AMD's original is still called for Auto. A library that gives
    that slot another meaning is not supported.
  • A refused Lite context or failed dispatch falls back to Auto once per run; the row still says Lite and the log says
    why.
  • The row is absent where there is no Lite provider, on Linux, or on a DLL whose shape is not understood;
    RX_VK_FSR4_MODEL=lite there logs and stays Auto.
  • Turning it off: Auto on the FSR 4 model row (fsr4_model=0), or leave RX_VK_FSR4_MODEL unset. Reverting the
    commit removes the row and the hook.
  • Compatibility: no AMD file is changed on disk and nothing new is shipped; no Vulkan interface or ABI version
    changes. longshot_options.ini gains fsr4_model (0 Auto, 1 Lite); an options file from before gets Auto. Saves,
    caches and the other upscalers are unchanged.

Checklist

Ready for review

  • One topic, on current main; the title is a Conventional Commit.
  • Tested on my own PC, in the game for a change the game runs, as Testing says.
  • mise run test passes (any failure that also fails on main is named).
  • Measured against main, or Measurements says why there is nothing to measure.
  • No game files, build output or hand-edited generated files; third_party/ untouched; nothing GPL-incompatible.

When it applies

  • A switch added or changed: it ships at the default the measurements chose, with its row in gpu/vk/switches.def.
  • A check under tests/ that reads changed source is updated, not weakened.
  • Docs updated.
  • AI assistance recorded with Assisted-by: / Co-Authored-By: trailers.
  • Screenshots of the same spot on main and here, for a change that shows on screen: the row is new, and the
    in-game screenshot is still to take.

Ready to merge

  • Everything in Still to test has run, its results in Tested.
  • Tier 1 hardware it can affect is tested (docs/dev/standards.md, "Hardware tiers").
  • Approved in review, every comment answered.
  • CI passes on the head; on current main, measured again or a comment says why the numbers still hold.
## Summary AMD's FSR 4 library ships four networks (FP8, full INT8, the Navi 2 INT8 model and Lite). FSR 4's SDK doesn't expose a model selector: its provider walk calls each network's `IsSupported`, which answers yes only for the GPU's native model. This scans AMD's loaded Vulkan DLL for the four provider classes, checks their 4.1.2 vtable shape, and patches only each provider's `IsSupported` entry (`vtable[2]`) in this process's copy, under `VirtualProtect`: Auto still calls AMD's original, Lite answers yes only for the Lite provider. A Graphics-tab **FSR 4 model** row (Auto or Lite) and `RX_VK_FSR4_MODEL` choose it for one run. The model joins the upscaler context's key, so a change builds one new context at the next frame, and a refused Lite context falls back to Auto once. Auto stays the default and AMD's own selection is unchanged. Read first: `gpu/vk/fsr4_models.cpp` (`scan_image`, `install`), `gpu/vk/vk_fsr4.cpp` (`fsr4_model`, `fsr4_lite_available`, the context key) and `gpu/vk/menu/upscaler.cpp` (the row). ### What changed - **Scanner and hook** — `gpu/vk/fsr4_models.h/.cpp` (new). - `scan_image()` reads a PE image as bytes (no Vulkan, no GPU) and finds the four `. ?AVffxProvider_FSR4*` classes, their Complete Object Locators, vtables and the static provider registry. It checks the 4.1.2 shape: entries 0, 1 and 3..7 shared, entry 2 distinct; fewer than two providers or a shifted entry fails the check. - On Windows, `scan_loaded_modules()` scans every loaded module at the first FSR 4 use. The probe logs each module's providers, shape, registry and the native model, asked of each provider's own `IsSupported`. - `install()` patches only `vtable[2]` under `VirtualProtect`, once per module, and only when the shape is understood, a Lite provider is present and at least two providers are found. If entry 2 is not AMD's original it fails closed and rolls back. Auto delegates to AMD's original (recording the first yes as the native model); Lite answers true only for the Lite provider. Nothing is written to AMD's file. - **Engine** — `gpu/vk/vk_fsr4.cpp/.h`. - At load (Windows) the probe runs and the hook is installed. `fsr4_lite_available()` is true only for an AMD GPU with a hooked Lite provider. - `Fsr4Model` Auto/Lite, `RX_VK_FSR4_MODEL=auto|lite` (also 0|1) and the menu's live value; the switch wins for the run. - The context key includes the model, so a change makes one new context at the next frame and logs whether it is native or forced, and tuned or AMD's shaders. - A forced Lite context on a card whose native model is not Lite suppresses the tuned buffer/shader/pipeline hooks (AMD's own shaders); an older Auto context keeps its tuned pipelines, so `cmd_dispatch` is not suppressed. - A refused Lite context or a failed dispatch falls back to Auto once; the saved row stays Lite and the log says why. - **Menu** — the row is a module (`gpu/vk/menu/upscaler.cpp`, #1633), after **DLSS model**: Auto (AMD's) and Lite (INT8), saved as `fsr4_model` (`gpu/vk/vk_menu.h`), with its `on_change` updating the engine. `vk_menu.cpp` only reads `RX_VK_FSR4_MODEL` at start and sets the live value. The row is shown only where FSR 4 runs and a hooked Lite provider is present, and is locked when the Upscaler is not FSR 4. The help and all 13 language tables carry it. - **Switches and docs** — `RX_VK_FSR4_MODEL` in `gpu/vk/switches.def`; the user guide's Graphics row and FSR 4 section (`docs/user/graphics-and-controls.md`, `docs/user/advanced.md`) and the developer reference (`docs/dev/reference/switches.md`, `docs/dev/reference/upscalers.md`, "FSR 4 models"). - **Tests** — `tests/test_fsr4_models.py` with `tests/fsr4_models_test.cpp` (new) covers the scanner on synthetic PE images laid out like 4.1.2: the four providers and their order, their Complete Object Locators/vtables, the registry, the shape check, a missing Lite, one provider, a shifted shared entry, an unknown provider, a non-PE image and the native-model helper; on Windows it also exercises install, Auto/Lite/Auto answers, install twice and a second scan (its image's pages are made executable for that call, which Windows does not do for heap pages). `tests/test_fsr4_tune.py` checks the suppression pass-through, that `cmd_dispatch` is not suppressed, and that `prepare` rebuilds on a model change. `tests/test_upscaler_rows.cpp` covers the row's values and shown/fixed rules; `tests/test_menu_ini_keys.py` freezes the new key, and `tests/test_menu_env_overrides.py` and `tests/test_hdr_encode.py` cover the `menu_init` and `make_context(..., want)` changes. ## Testing **Tested:** - In the game on Windows, AMD Radeon RX 6600 (RDNA 2), `amd_fidelityfx_upscaler_vk.dll` 4.1.2.4555 (FSR SDK 2.4), with FSR 4 on: the FSR 4 model row selected Auto and Lite in Kokiri Forest, 1440x810 -> 1920x1080, and the FSR 4 upscaling GPU time was read from `longshot_log.txt` for both. - The game-free suite on WSL2 Ubuntu 24.04.4 LTS (the suite reports `linux`) at `7dffea2b` (on `6e1521a7`), on the same PC and checkout as the Windows runs (`/mnt/d/Longshot` is `D:\Longshot`), `python3 -I -B tests/run_all.py`: **720 passed, 70 failed, 115 skipped** (767 s, 12 at once). The 69 failures below are the same checks with the same error on a worktree of clean `6e1521a7`, so none comes from this branch; the 70th is `byte_compile_tracked`, which fails only because this working tree holds `tests/test_vbytes.py` out locally (as `tests/test_vbytes.py.skip`; the test itself runs longer than the runner's 600 s check timeout on this WSL/drvfs checkout) and passes with the file in place. - `tests/test_fsr4_models.py` and the branch's menu checks that run here (`test_menu_ini_keys.py`, `test_menu_env_overrides.py`, `test_hdr_encode.py`, `test_menu_i18n.py`) pass. `tests/test_fsr4_tune.py` is skipped without `spirv-val`; `test_upscaler_rows.py`, `test_menu_modules.py`, `test_menu_sections.py`, `test_settings_table.py`, `test_menu_module_slots.py` and `test_pads_after_menu.py` fail on clean `main` too. <details> <summary>The 69 checks that also fail on clean `6e1521a7` (this PC)</summary> `tests/test_appimage_player_build.py`, `tests/test_bug_report_defaults.py`, `tests/test_capture_read_once.py`, `tests/test_check_third_party.py`, `tests/test_crash_dialog.py`, `tests/test_crash_record.py`, `tests/test_crt_options.py`, `tests/test_crt_shaders.py`, `tests/test_damage_debug_menu.py`, `tests/test_dev_ab.py`, `tests/test_dev_packs.py`, `tests/test_device_lost_exit.py`, `tests/test_discover.py`, `tests/test_display_startup_failure.py`, `tests/test_dlss_feature_size.py`, `tests/test_failed_compile_keeps_objects.py`, `tests/test_flush_nothing_held.py`, `tests/test_frame_overlap.py`, `tests/test_gpu_wait_stats.py`, `tests/test_guest_imported.py`, `tests/test_high_fps_fixes.py`, `tests/test_host_guest_writes.py`, `tests/test_image_reaper_threads.py`, `tests/test_libc_hooks.py`, `tests/test_lift_cache.py`, `tests/test_lift_lookups.py`, `tests/test_linux_preflight.py`, `tests/test_linux_pressure.py`, `tests/test_linux_tools.py`, `tests/test_linux_update.py`, `tests/test_log_open_failure.py`, `tests/test_longshot_saves.py`, `tests/test_make_game_cache.py`, `tests/test_make_game_ngx.py`, `tests/test_menu_module_slots.py`, `tests/test_menu_modules.py`, `tests/test_menu_sections.py`, `tests/test_mesh_conv.py`, `tests/test_mise_build.py`, `tests/test_missing_nup.py`, `tests/test_mods_list.py`, `tests/test_mods_size_table.py`, `tests/test_ngx_sdk.py`, `tests/test_oot_shadow_fit.py`, `tests/test_pads_after_menu.py`, `tests/test_pending_guest_import.py`, `tests/test_pipeline_play_cache.py`, `tests/test_pipelines_outlive_cache_format.py`, `tests/test_prepare_cache_save.py`, `tests/test_profiling_mode.py`, `tests/test_program_cache_cleanup.py`, `tests/test_program_cache_one_variant.py`, `tests/test_psychov35_fsr3_fg.py`, `tests/test_readback_timeout_source.py`, `tests/test_reflex_swapchain_lifetime.py`, `tests/test_renderer_new_handler.py`, `tests/test_renhdrum.py`, `tests/test_settings_table.py`, `tests/test_slow_frames.py`, `tests/test_song_note_cue.py`, `tests/test_sparse_mirror.py`, `tests/test_spirv_pack.py`, `tests/test_split_giants.py`, `tests/test_sprint.py`, `tests/test_thinlto_tls.py`, `tests/test_upscaler_rows.py`, `tests/test_vcore_escape.py`, `tests/test_vcore_mask0.py`, `tests/test_walk.py` </details> **Still to test:** - Other AMD cards (RDNA 3/4): with FSR 4 on, expect the `vk: AMD FSR 4: models:`. ## Measurements **Verdict:** Lite is about twice as fast as the default Navi 2 INT8 model on a RX 6600, and it provides superior image quality compared to both FSR 3.1.6 and XeSS-DP4a, specially with hair and moving objects. Setup: Windows 10, AMD RX 6600, 26.10.1; Kokiri Forest, (75%) 1440x810 -> 1920x1080 | Measurement|Auto (INT8 Navi 2) | Lite (INT8) | Δ | |---|---:|---:|---: | upscaling GPU time | 5.608 ms | 2.899 ms | -2.709 ms (-48.3%) | ### Flags and settings One new menu row and one new switch; both default to Auto. | Flag | What it is | Before this PR | After this PR | Effect | Helps / doesn't help / untested | Recommendation | |---|---|---|---|---|---|---| | `fsr4_model` (FSR 4 model row) | The Graphics tab's network choice, saved in `longshot_options.ini` | - | Auto (0) or Lite (1); Auto | Auto lets AMD pick for the GPU; Lite forces the Lite provider | Helps cards AMD does not natively run Lite on; other cards untested | Auto | | `RX_VK_FSR4_MODEL` | The switch over the row, for one run | - | unset; `auto` or `lite` (`0`/`1`) | `auto` keeps AMD's pick; `lite` forces the Lite provider and locks the row | For A/B runs and testing; Linux has no Lite, so it does nothing there | unset (Auto) | ## Risks - The hook is in AMD's process copy. A DLL with the same class names but another vtable layout would be misread; the RTTI names and 4.1.2 shape check are the gate, and anything not understood is left alone (no hook, row hidden, Auto). - Only `IsSupported` (vtable entry 2) is replaced, and AMD's original is still called for Auto. A library that gives that slot another meaning is not supported. - A refused Lite context or failed dispatch falls back to Auto once per run; the row still says Lite and the log says why. - The row is absent where there is no Lite provider, on Linux, or on a DLL whose shape is not understood; `RX_VK_FSR4_MODEL=lite` there logs and stays Auto. - **Turning it off:** Auto on the FSR 4 model row (`fsr4_model=0`), or leave `RX_VK_FSR4_MODEL` unset. Reverting the commit removes the row and the hook. - **Compatibility:** no AMD file is changed on disk and nothing new is shipped; no Vulkan interface or ABI version changes. `longshot_options.ini` gains `fsr4_model` (0 Auto, 1 Lite); an options file from before gets Auto. Saves, caches and the other upscalers are unchanged. ## Checklist ### Ready for review - [x] One topic, on current `main`; the title is a Conventional Commit. - [x] Tested on my own PC, in the game for a change the game runs, as Testing says. - [x] `mise run test` passes (any failure that also fails on `main` is named). - [x] Measured against `main`, or Measurements says why there is nothing to measure. - [x] No game files, build output or hand-edited generated files; `third_party/` untouched; nothing GPL-incompatible. ### When it applies - [x] A switch added or changed: it ships at the default the measurements chose, with its row in `gpu/vk/switches.def`. - [x] A check under `tests/` that reads changed source is updated, not weakened. - [x] Docs updated. - [x] AI assistance recorded with `Assisted-by:` / `Co-Authored-By:` trailers. - [ ] Screenshots of the same spot on `main` and here, for a change that shows on screen: the row is new, and the in-game screenshot is still to take. ### Ready to merge - [ ] Everything in Still to test has run, its results in Tested. - [ ] Tier 1 hardware it can affect is tested (docs/dev/standards.md, "Hardware tiers"). - [ ] Approved in review, every comment answered. - [ ] CI passes on the head; on current `main`, measured again or a comment says why the numbers still hold.
feat(vk): add support for FSR 4 "Lite" model
Some checks are pending
CI / Conventional Commit title (pull_request) Blocked by required conditions
CI / Python source checks (ubuntu-24.04, part 1/2) (pull_request) Blocked by required conditions
CI / Python source checks (ubuntu-24.04, part 2/2) (pull_request) Blocked by required conditions
CI / SPIR-V fixtures (global narrow load/store, direct surface atomics, F2I immediate) (pull_request) Blocked by required conditions
CI / Workflow lint (actionlint) (pull_request) Blocked by required conditions
CI / Website build and checks (pull_request) Blocked by required conditions
Guide screenshots / Guide images and local UI source review (pull_request) Blocked by required conditions
Third-party parts / As their notes say (pull_request) Blocked by required conditions
Third-party parts / The same as upstream (pull_request) Blocked by required conditions
Website / Website on Cloudflare (pull_request) Blocked by required conditions
PR description / Description follows the template (pull_request_target) Successful in 6s
Website / Website on Cloudflare (pull_request_target) Has been skipped
3cbaf907fb
## Summary

AMD's FSR 4 library ships four networks (FP8, full INT8, the Navi 2 INT8 model and Lite). FSR 4's SDK doesn't expose a
model selector: its provider walk calls each network's `IsSupported`, which answers yes only for the GPU's native model.

This scans AMD's loaded Vulkan DLL for the four provider classes, checks their 4.1.2 vtable shape, and patches only
each provider's `IsSupported` entry (`vtable[2]`) in this process's copy, under `VirtualProtect`: Auto still calls
AMD's original, Lite answers yes only for the Lite provider.

A Graphics-tab **FSR 4 model** row (Auto or Lite) and `RX_VK_FSR4_MODEL` choose it for one run. The model joins the
upscaler context's key, so a change builds one new context at the next frame, and a refused Lite context falls back to
Auto once. Auto stays the default and AMD's own selection is unchanged.

Read first: `gpu/vk/fsr4_models.cpp` (`scan_image`, `install`), `gpu/vk/vk_fsr4.cpp` (`fsr4_model`,
`fsr4_lite_available`, the context key) and `gpu/vk/menu/upscaler.cpp` (the row).

### What changed

- **Scanner and hook** — `gpu/vk/fsr4_models.h/.cpp` (new).
  - `scan_image()` reads a PE image as bytes (no Vulkan, no GPU) and finds the four `. ?AVffxProvider_FSR4*` classes,
    their Complete Object Locators, vtables and the static provider registry. It checks the 4.1.2 shape: entries 0, 1
    and 3..7 shared, entry 2 distinct; fewer than two providers or a shifted entry fails the check.
  - On Windows, `scan_loaded_modules()` scans every loaded module at the first FSR 4 use. The probe logs each module's
    providers, shape, registry and the native model, asked of each provider's own `IsSupported`.
  - `install()` patches only `vtable[2]` under `VirtualProtect`, once per module, and only when the shape is understood,
    a Lite provider is present and at least two providers are found. If entry 2 is not AMD's original it fails closed
    and rolls back. Auto delegates to AMD's original (recording the first yes as the native model); Lite answers true
    only for the Lite provider. Nothing is written to AMD's file.
- **Engine** — `gpu/vk/vk_fsr4.cpp/.h`.
  - At load (Windows) the probe runs and the hook is installed. `fsr4_lite_available()` is true only for an AMD GPU
    with a hooked Lite provider.
  - `Fsr4Model` Auto/Lite, `RX_VK_FSR4_MODEL=auto|lite` (also 0|1) and the menu's live value; the switch wins for the
    run.
  - The context key includes the model, so a change makes one new context at the next frame and logs whether it is
    native or forced, and tuned or AMD's shaders.
  - A forced Lite context on a card whose native model is not Lite suppresses the tuned buffer/shader/pipeline hooks
    (AMD's own shaders); an older Auto context keeps its tuned pipelines, so `cmd_dispatch` is not suppressed.
  - A refused Lite context or a failed dispatch falls back to Auto once; the saved row stays Lite and the log says why.
- **Menu** — the row is a module (`gpu/vk/menu/upscaler.cpp`, #1633), after **DLSS model**: Auto (AMD's) and Lite
  (INT8), saved as `fsr4_model` (`gpu/vk/vk_menu.h`), with its `on_change` updating the engine. `vk_menu.cpp` only reads
  `RX_VK_FSR4_MODEL` at start and sets the live value. The row is shown only where FSR 4 runs and a hooked Lite provider
  is present, and is locked when the Upscaler is not FSR 4. The help and all 13 language tables carry it.
- **Switches and docs** — `RX_VK_FSR4_MODEL` in `gpu/vk/switches.def`; the user guide's Graphics row and FSR 4 section
  (`docs/user/graphics-and-controls.md`, `docs/user/advanced.md`) and the developer reference
  (`docs/dev/reference/switches.md`, `docs/dev/reference/upscalers.md`, "FSR 4 models").
- **Tests** — `tests/test_fsr4_models.py` with `tests/fsr4_models_test.cpp` (new) covers the scanner on synthetic PE
  images laid out like 4.1.2: the four providers and their order, their Complete Object Locators/vtables, the registry,
  the shape check, a missing Lite, one provider, a shifted shared entry, an unknown provider, a non-PE image and the
  native-model helper; on Windows it also exercises install, Auto/Lite/Auto answers, install twice and a second scan
  (its image's pages are made executable for that call, which Windows does not do for heap pages). `tests/test_fsr4_tune.py`
  checks the suppression pass-through, that `cmd_dispatch` is not suppressed, and that `prepare` rebuilds on a model
  change. `tests/test_upscaler_rows.cpp` covers the row's values and shown/fixed rules; `tests/test_menu_ini_keys.py`
  freezes the new key, and `tests/test_menu_env_overrides.py` and `tests/test_hdr_encode.py` cover the `menu_init` and
  `make_context(..., want)` changes.

## Testing

**Tested:**
- In the game on Windows, AMD Radeon RX 6600 (RDNA 2), `amd_fidelityfx_upscaler_vk.dll` 4.1.2.4555 (FSR SDK 2.4), with
  FSR 4 on: the FSR 4 model row selected Auto and Lite in Kokiri Forest, 1440x810 -> 1920x1080, and the FSR 4
  upscaling GPU time was read from `longshot_log.txt` for both.
- The game-free suite on WSL2 Ubuntu 24.04.4 LTS (the suite reports `linux`) at `7dffea2b` (on `6e1521a7`), on the same
  PC and checkout as the Windows runs (`/mnt/d/Longshot` is `D:\Longshot`), `python3 -I -B tests/run_all.py`: **720
  passed, 70 failed, 115 skipped** (767 s, 12 at once). The 69 failures below are the same checks with the same error on
  a worktree of clean `6e1521a7`, so none comes from this branch; the 70th is `byte_compile_tracked`, which fails only
  because this working tree holds `tests/test_vbytes.py` out locally (as `tests/test_vbytes.py.skip`; the test itself
  runs longer than the runner's 600 s check timeout on this WSL/drvfs checkout) and passes with the file in place.
- `tests/test_fsr4_models.py` and the branch's menu checks that run here (`test_menu_ini_keys.py`,
  `test_menu_env_overrides.py`, `test_hdr_encode.py`, `test_menu_i18n.py`) pass. `tests/test_fsr4_tune.py` is skipped
  without `spirv-val`; `test_upscaler_rows.py`, `test_menu_modules.py`, `test_menu_sections.py`,
  `test_settings_table.py`, `test_menu_module_slots.py` and `test_pads_after_menu.py` fail on clean `main` too.

  <details>
  <summary>The 69 checks that also fail on clean `6e1521a7` (this PC)</summary>

  `tests/test_appimage_player_build.py`, `tests/test_bug_report_defaults.py`, `tests/test_capture_read_once.py`,
  `tests/test_check_third_party.py`, `tests/test_crash_dialog.py`, `tests/test_crash_record.py`,
  `tests/test_crt_options.py`, `tests/test_crt_shaders.py`, `tests/test_damage_debug_menu.py`,
  `tests/test_dev_ab.py`, `tests/test_dev_packs.py`, `tests/test_device_lost_exit.py`, `tests/test_discover.py`,
  `tests/test_display_startup_failure.py`, `tests/test_dlss_feature_size.py`,
  `tests/test_failed_compile_keeps_objects.py`, `tests/test_flush_nothing_held.py`, `tests/test_frame_overlap.py`,
  `tests/test_gpu_wait_stats.py`, `tests/test_guest_imported.py`, `tests/test_high_fps_fixes.py`,
  `tests/test_host_guest_writes.py`, `tests/test_image_reaper_threads.py`, `tests/test_libc_hooks.py`,
  `tests/test_lift_cache.py`, `tests/test_lift_lookups.py`, `tests/test_linux_preflight.py`,
  `tests/test_linux_pressure.py`, `tests/test_linux_tools.py`, `tests/test_linux_update.py`,
  `tests/test_log_open_failure.py`, `tests/test_longshot_saves.py`, `tests/test_make_game_cache.py`,
  `tests/test_make_game_ngx.py`, `tests/test_menu_module_slots.py`, `tests/test_menu_modules.py`,
  `tests/test_menu_sections.py`, `tests/test_mesh_conv.py`, `tests/test_mise_build.py`, `tests/test_missing_nup.py`,
  `tests/test_mods_list.py`, `tests/test_mods_size_table.py`, `tests/test_ngx_sdk.py`,
  `tests/test_oot_shadow_fit.py`, `tests/test_pads_after_menu.py`, `tests/test_pending_guest_import.py`,
  `tests/test_pipeline_play_cache.py`, `tests/test_pipelines_outlive_cache_format.py`,
  `tests/test_prepare_cache_save.py`, `tests/test_profiling_mode.py`, `tests/test_program_cache_cleanup.py`,
  `tests/test_program_cache_one_variant.py`, `tests/test_psychov35_fsr3_fg.py`,
  `tests/test_readback_timeout_source.py`, `tests/test_reflex_swapchain_lifetime.py`,
  `tests/test_renderer_new_handler.py`, `tests/test_renhdrum.py`, `tests/test_settings_table.py`,
  `tests/test_slow_frames.py`, `tests/test_song_note_cue.py`, `tests/test_sparse_mirror.py`,
  `tests/test_spirv_pack.py`, `tests/test_split_giants.py`, `tests/test_sprint.py`, `tests/test_thinlto_tls.py`,
  `tests/test_upscaler_rows.py`, `tests/test_vcore_escape.py`, `tests/test_vcore_mask0.py`, `tests/test_walk.py`

  </details>

**Still to test:**
- Other AMD cards (RDNA 3/4): with FSR 4 on, expect the `vk: AMD FSR 4: models:`.

## Measurements

**Verdict:** Lite is about twice as fast as the default Navi 2 INT8 model on a RX 6600, and it provides superior image quality compared to both FSR 3.1.6 and XeSS-DP4a, specially with hair and moving objects.

Setup: Windows 10, AMD RX 6600, 26.10.1; Kokiri Forest, (75%) 1440x810 -> 1920x1080

| Measurement|Auto (INT8 Navi 2) | Lite (INT8) | Δ |
|---|---:|---:|---:
| upscaling GPU time | 5.608 ms | 2.899 ms | -2.709 ms (-48.3%) |

### Flags and settings

One new menu row and one new switch; both default to Auto.

| Flag | What it is | Before this PR | After this PR | Effect | Helps / doesn't help / untested | Recommendation |
|---|---|---|---|---|---|---|
| `fsr4_model` (FSR 4 model row) | The Graphics tab's network choice, saved in `longshot_options.ini` | - | Auto (0) or Lite (1); Auto | Auto lets AMD pick for the GPU; Lite forces the Lite provider | Helps cards AMD does not natively run Lite on; other cards untested | Auto |
| `RX_VK_FSR4_MODEL` | The switch over the row, for one run | - | unset; `auto` or `lite` (`0`/`1`) | `auto` keeps AMD's pick; `lite` forces the Lite provider and locks the row | For A/B runs and testing; Linux has no Lite, so it does nothing there | unset (Auto) |

## Risks

- The hook is in AMD's process copy. A DLL with the same class names but another vtable layout would be misread; the
  RTTI names and 4.1.2 shape check are the gate, and anything not understood is left alone (no hook, row hidden, Auto).
- Only `IsSupported` (vtable entry 2) is replaced, and AMD's original is still called for Auto. A library that gives
  that slot another meaning is not supported.
- A refused Lite context or failed dispatch falls back to Auto once per run; the row still says Lite and the log says
  why.
- The row is absent where there is no Lite provider, on Linux, or on a DLL whose shape is not understood;
  `RX_VK_FSR4_MODEL=lite` there logs and stays Auto.
- **Turning it off:** Auto on the FSR 4 model row (`fsr4_model=0`), or leave `RX_VK_FSR4_MODEL` unset. Reverting the
  commit removes the row and the hook.
- **Compatibility:** no AMD file is changed on disk and nothing new is shipped; no Vulkan interface or ABI version
  changes. `longshot_options.ini` gains `fsr4_model` (0 Auto, 1 Lite); an options file from before gets Auto. Saves,
  caches and the other upscalers are unchanged.

## Checklist

### Ready for review

- [x] One topic, on current `main`; the title is a Conventional Commit.
- [x] Tested on my own PC, in the game for a change the game runs, as Testing says.
- [x] `mise run test` passes (any failure that also fails on `main` is named).
- [x] Measured against `main`, or Measurements says why there is nothing to measure.
- [x] No game files, build output or hand-edited generated files; `third_party/` untouched; nothing GPL-incompatible.

### When it applies

- [x] A switch added or changed: it ships at the default the measurements chose, with its row in `gpu/vk/switches.def`.
- [x] A check under `tests/` that reads changed source is updated, not weakened.
- [x] Docs updated.
- [x] AI assistance recorded with `Assisted-by:` / `Co-Authored-By:` trailers.
- [ ] Screenshots of the same spot on `main` and here, for a change that shows on screen: the row is new, and the
  in-game screenshot is still to take.

### Ready to merge

- [ ] Everything in Still to test has run, its results in Tested.
- [ ] Tier 1 hardware it can affect is tested (docs/dev/standards.md, "Hardware tiers").
- [ ] Approved in review, every comment answered.
- [ ] CI passes on the head; on current `main`, measured again or a comment says why the numbers still hold.

Assisted-by: DeepSeek V4.1 Flash (OpenCode)
Author
First-time contributor

Tested in game: RX 6600, Windows 10, AMD 26.10.1

Tested in game: RX 6600, Windows 10, AMD 26.10.1
Member

Possibly related, for context:

  • PR #1886 perf(vk): warm FSR 4 pipelines during shader preparation — Open FSR 4 work also changes vk_fsr4 context preparation; possible integration overlap, but distinct goal.
  • issue #1319 Game: FSR 4 runs slower than native 4K on RX 7900 XTX — Open AMD FSR 4 performance report; Lite selection may be relevant to performance testing.
  • PR #1752 perf(vk): AMD FSR 4's unrolled layers in a pipeline cache of their own, kept across updates, and taken from it with no memory asked — Merged FSR 4 pipeline-cache change underlying the tuning/context path; useful implementation context.
Possibly related, for context: - PR #1886 perf(vk): warm FSR 4 pipelines during shader preparation — Open FSR 4 work also changes vk_fsr4 context preparation; possible integration overlap, but distinct goal. - issue #1319 Game: FSR 4 runs slower than native 4K on RX 7900 XTX — Open AMD FSR 4 performance report; Lite selection may be relevant to performance testing. - PR #1752 perf(vk): AMD FSR 4's unrolled layers in a pipeline cache of their own, kept across updates, and taken from it with no memory asked — Merged FSR 4 pipeline-cache change underlying the tuning/context path; useful implementation context.
BeanBot left a comment

Recommendation: Request changes. A failed Lite dispatch disables FSR 4 before the promised Auto retry can run. The Windows probe also reads and patches loaded DLLs without establishing their lifetime or verifying the private ABI, so it can break the default Auto path.

Adds a saved FSR 4 model choice and a run-only override. The RX 6600 report shows a promising reduction in upscaling GPU time. The feature adds a private DLL hook at first use and can rebuild the upscaler during play, with costs and compatibility still to establish.

@fix will do (21)

  1. L1 gpu/vk/vk_fsr4.cpp:1374: The next-frame Auto retry is unreachable in gameplay. (more on the line)
  2. L2 gpu/vk/fsr4_models.cpp:460: A Toolhelp snapshot does not retain a module. (more on the line)
  3. L3 gpu/vk/fsr4_models.cpp:211: SizeOfImage bounds the reserved image range. (more on the line)
  4. L4 gpu/vk/fsr4_models.cpp:237: The shared-entry pattern does not identify the tested ABI. (more on the line)
  5. L5 gpu/vk/fsr4_models.cpp:276: Any writable word holding a vtable address becomes a candidate object. (more on the line)
  6. L6 gpu/vk/fsr4_models.cpp:538: A caller can reach the patched slot before g_hooked_count publishes its record at line 544. (more on the line)
  7. L7 gpu/vk/fsr4_models.cpp:513: Rollback writes to a slot even when making its page writable fails. (more on the line)
  8. L8 gpu/vk/vk_fsr4.cpp:1130: Availability means that any scanned DLL installed a Lite hook. (more on the line)
  9. L9 gpu/vk/vk_fsr4.cpp:1162: Suppression is a process-wide setting that remains in place after context creation. (more on the line)
  10. L10 gpu/vk/vk_fsr4.cpp:1303: A live model change can call tune::device_proc again while detached Auto unroll jobs are active. (more on the line)
  11. L11 gpu/vk/vk_fsr4.cpp:1199: The code treats an error with a non-null output handle as a failure, then clears the handle before retrying. (more on the line)
  12. L12 gpu/vk/vk_fsr4.cpp:1302: A model-only change retires a working Auto context before the replacement is known to work. (more on the line)
  13. L13 (minor) gpu/vk/fsr4_models.cpp:456: GetModuleFileNameA logs a lossy path on Windows and can silently truncate it at MAX_PATH. (more on the line)
  14. L14 (minor) tests/fsr4_models_test.cpp:38: The Windows fixture makes pages of a std::vector heap allocation executable and writable, potentially including unrelated heap allocations. (more on the line)
  15. L15 (minor) tests/test_fsr4_models.py:32: The harness compiles with the selected C++ compiler, but it can fail to start with llvm-mingw when the compiler's runtime DLL directory is off PATH. (more on the line)
  16. L16 tests/test_fsr4_tune.py:204: The new source checks can pass if the tuning guard is inverted, and there is no integrated test that can catch the broken Auto retry. (more on the line)
  17. L17 (minor) tests/test_menu_env_overrides.py:67: The fixture stubs the model override as always absent and discards the live update. (more on the line)
  18. L18 (its part an edit can do) gpu/vk/vk_fsr4.cpp:1108: Limit discovery to the verified AMD upscaler, defer mutation until Lite is requested, and log the switch-frame create time and both contexts' reported memory during overlap.
  19. L19 (minor): The description checks off measurement against main and passing checks, but shows Auto versus Lite without build revisions and cites a suite run from 7dffea2b with a locally omitted test and a skipped tuning check. Update the checklist and testing claims to match the evidence, name the revisions for the game runs, and give each outstanding test an owner and expected result. Correct the Auto off instructions to reflect whether the hook still installs.
  20. L20 (minor) gpu/vk/menu/upscaler.cpp:172: The help promises lower memory use and better quality than FSR 3 for every player. (more on the line)
  21. L21 (minor) docs/user/images/guide/manifest.json:845: The Graphics capture tracks menu files this PR changes, but the PR includes no source review of the capture. Check the guide on a configuration where the new row is visible and on the existing UI fixture, then record which capture still matches and replace any affected capture that no longer does. Include the same-spot UI comparison required by the PR template.

For the author (1) (a run, a measurement or a decision)

  • L18 (minor) gpu/vk/vk_fsr4.cpp:1108: Measure the default path and the live switch. (more on the line)

Checks: Still running: CI / Conventional Commit title (pull_request), CI / Python source checks (ubuntu-24.04, part 1/2) (pull_request), CI / Python source checks (ubuntu-24.04, part 2/2) (pull_request), CI / SPIR-V fixtures (global narrow load/store, direct surface atomics, F2I immediate) (pull_request), CI / Workflow lint (actionlint) (pull_request), CI / Website build and checks (pull_request), Guide screenshots / Guide images and local UI source review (pull_request), Third-party parts / As their notes say (pull_request), Third-party parts / The same as upstream (pull_request), Website / Website on Cloudflare (pull_request).

More: does it help, other systems, switches, what was read

Does it help: Unproven. The PR reports 5.608 ms for Auto and 2.899 ms for Lite on Windows 10, RX 6600, driver 26.10.1, in Kokiri Forest at 1440×810 to 1920×1080. That is 2.709 ms less FSR 4 GPU time, or 48.3%. Neither build revision nor interleaved run results are given for that comparison. Base versus head Auto, frame pacing, VRAM and the claimed image-quality gain are unmeasured in the supplied evidence.

Other systems: The Windows probe runs during startup or first menu use even with Auto selected. It can touch driver and overlay DLLs and runs before the AMD GPU gate. FSR 4 and its Native AA mode share the context path, including native HDR. Context switches can overlap deferred GPU retirement and shader compilation. Linux and Steam Deck skip the hook and use Auto. Other upscaler engines retain their existing selection logic. The diff leaves save formats, the Vulkan API and shader cache formats unchanged.

Switches: fsr4_model defaults to Auto (0) and saves Lite as 1. RX_VK_FSR4_MODEL defaults to unset, overrides the row for one run and accepts auto|lite or 0|1. Invalid nonempty values lock the row to Auto. Auto is the conservative model default. It still installs the Windows hook, so the default path needs its own safety and performance evidence. The switch table has the new entry.

Reviewed: 3cbaf907fb against 6e1521a70a: Read AGENTS.md, CONTRIBUTING.md, the Forgejo PR template and developer standards. No docs/review-checklist.md was present. Read the full diff and commit, both new production files and both new test files in full, the changed menu module and FSR 4 engine paths, and the relevant temporal caller, tuning worker, menu persistence, build discovery, vendor headers and test harness code. Skimmed the 13 translation tables and guide capture records. The rest of the large menu and temporal files, vendor binaries and remote PRs were outside the read. No raw run logs were supplied.; nothing run (a code review).

Checked and fine:

  • The single commit has the required Assisted-by trailer.
  • No game-derived data, vendor binaries, third-party edits, build artifacts or generated shader edits appear in the diff. The public Vulkan interface is unchanged.
  • The row lives in the upscaler module, stays out of presets, has a frozen INI key and appears in the guide's menu order. All 13 translation tables add the row strings.
  • The normal replacement path uses the existing deferred retirement queue, and a newly made context resets temporal history through the existing prepare result.

Testing: Read-only review. The author reports Windows RX 6600 gameplay and earlier WSL2 suite results of 720 passed, 70 failed and 115 skipped. No native Windows harness result or exact-head baseline comparison was supplied. The listed CI checks are still running.

Overlaps and merge order: None found in the local history. The PR builds on the existing #1633 menu modules. Remote overlap and the cause of the reported non-mergeable status could not be verified from the checkout.


Automated code review by BeanBot (gpt-6.2-sol), on the pull request's opening. Nothing was run; a maintainer decides.

**Recommendation: Request changes.** A failed Lite dispatch disables FSR 4 before the promised Auto retry can run. The Windows probe also reads and patches loaded DLLs without establishing their lifetime or verifying the private ABI, so it can break the default Auto path. Adds a saved FSR 4 model choice and a run-only override. The RX 6600 report shows a promising reduction in upscaling GPU time. The feature adds a private DLL hook at first use and can rebuild the upscaler during play, with costs and compatibility still to establish. **`@fix` will do (21)** 1. **L1** `gpu/vk/vk_fsr4.cpp:1374`: The next-frame Auto retry is unreachable in gameplay. (more on the line) 2. **L2** `gpu/vk/fsr4_models.cpp:460`: A Toolhelp snapshot does not retain a module. (more on the line) 3. **L3** `gpu/vk/fsr4_models.cpp:211`: `SizeOfImage` bounds the reserved image range. (more on the line) 4. **L4** `gpu/vk/fsr4_models.cpp:237`: The shared-entry pattern does not identify the tested ABI. (more on the line) 5. **L5** `gpu/vk/fsr4_models.cpp:276`: Any writable word holding a vtable address becomes a candidate object. (more on the line) 6. **L6** `gpu/vk/fsr4_models.cpp:538`: A caller can reach the patched slot before `g_hooked_count` publishes its record at line 544. (more on the line) 7. **L7** `gpu/vk/fsr4_models.cpp:513`: Rollback writes to a slot even when making its page writable fails. (more on the line) 8. **L8** `gpu/vk/vk_fsr4.cpp:1130`: Availability means that any scanned DLL installed a Lite hook. (more on the line) 9. **L9** `gpu/vk/vk_fsr4.cpp:1162`: Suppression is a process-wide setting that remains in place after context creation. (more on the line) 10. **L10** `gpu/vk/vk_fsr4.cpp:1303`: A live model change can call `tune::device_proc` again while detached Auto unroll jobs are active. (more on the line) 11. **L11** `gpu/vk/vk_fsr4.cpp:1199`: The code treats an error with a non-null output handle as a failure, then clears the handle before retrying. (more on the line) 12. **L12** `gpu/vk/vk_fsr4.cpp:1302`: A model-only change retires a working Auto context before the replacement is known to work. (more on the line) 13. **L13** (minor) `gpu/vk/fsr4_models.cpp:456`: `GetModuleFileNameA` logs a lossy path on Windows and can silently truncate it at MAX_PATH. (more on the line) 14. **L14** (minor) `tests/fsr4_models_test.cpp:38`: The Windows fixture makes pages of a `std::vector` heap allocation executable and writable, potentially including unrelated heap allocations. (more on the line) 15. **L15** (minor) `tests/test_fsr4_models.py:32`: The harness compiles with the selected C++ compiler, but it can fail to start with llvm-mingw when the compiler's runtime DLL directory is off PATH. (more on the line) 16. **L16** `tests/test_fsr4_tune.py:204`: The new source checks can pass if the tuning guard is inverted, and there is no integrated test that can catch the broken Auto retry. (more on the line) 17. **L17** (minor) `tests/test_menu_env_overrides.py:67`: The fixture stubs the model override as always absent and discards the live update. (more on the line) 18. **L18** (its part an edit can do) `gpu/vk/vk_fsr4.cpp:1108`: Limit discovery to the verified AMD upscaler, defer mutation until Lite is requested, and log the switch-frame create time and both contexts' reported memory during overlap. 19. **L19** (minor): The description checks off measurement against main and passing checks, but shows Auto versus Lite without build revisions and cites a suite run from 7dffea2b with a locally omitted test and a skipped tuning check. Update the checklist and testing claims to match the evidence, name the revisions for the game runs, and give each outstanding test an owner and expected result. Correct the Auto off instructions to reflect whether the hook still installs. 20. **L20** (minor) `gpu/vk/menu/upscaler.cpp:172`: The help promises lower memory use and better quality than FSR 3 for every player. (more on the line) 21. **L21** (minor) `docs/user/images/guide/manifest.json:845`: The Graphics capture tracks menu files this PR changes, but the PR includes no source review of the capture. Check the guide on a configuration where the new row is visible and on the existing UI fixture, then record which capture still matches and replace any affected capture that no longer does. Include the same-spot UI comparison required by the PR template. **For the author (1)** (a run, a measurement or a decision) - **L18** (minor) `gpu/vk/vk_fsr4.cpp:1108`: Measure the default path and the live switch. (more on the line) **Checks:** Still running: `CI / Conventional Commit title (pull_request)`, `CI / Python source checks (ubuntu-24.04, part 1/2) (pull_request)`, `CI / Python source checks (ubuntu-24.04, part 2/2) (pull_request)`, `CI / SPIR-V fixtures (global narrow load/store, direct surface atomics, F2I immediate) (pull_request)`, `CI / Workflow lint (actionlint) (pull_request)`, `CI / Website build and checks (pull_request)`, `Guide screenshots / Guide images and local UI source review (pull_request)`, `Third-party parts / As their notes say (pull_request)`, `Third-party parts / The same as upstream (pull_request)`, `Website / Website on Cloudflare (pull_request)`. <details> <summary>More: does it help, other systems, switches, what was read</summary> **Does it help:** Unproven. The PR reports 5.608 ms for Auto and 2.899 ms for Lite on Windows 10, RX 6600, driver 26.10.1, in Kokiri Forest at 1440×810 to 1920×1080. That is 2.709 ms less FSR 4 GPU time, or 48.3%. Neither build revision nor interleaved run results are given for that comparison. Base versus head Auto, frame pacing, VRAM and the claimed image-quality gain are unmeasured in the supplied evidence. **Other systems:** The Windows probe runs during startup or first menu use even with Auto selected. It can touch driver and overlay DLLs and runs before the AMD GPU gate. FSR 4 and its Native AA mode share the context path, including native HDR. Context switches can overlap deferred GPU retirement and shader compilation. Linux and Steam Deck skip the hook and use Auto. Other upscaler engines retain their existing selection logic. The diff leaves save formats, the Vulkan API and shader cache formats unchanged. **Switches:** `fsr4_model` defaults to Auto (0) and saves Lite as 1. `RX_VK_FSR4_MODEL` defaults to unset, overrides the row for one run and accepts `auto|lite` or `0|1`. Invalid nonempty values lock the row to Auto. Auto is the conservative model default. It still installs the Windows hook, so the default path needs its own safety and performance evidence. The switch table has the new entry. **Reviewed:** 3cbaf907fb5c against 6e1521a70a32: Read AGENTS.md, CONTRIBUTING.md, the Forgejo PR template and developer standards. No docs/review-checklist.md was present. Read the full diff and commit, both new production files and both new test files in full, the changed menu module and FSR 4 engine paths, and the relevant temporal caller, tuning worker, menu persistence, build discovery, vendor headers and test harness code. Skimmed the 13 translation tables and guide capture records. The rest of the large menu and temporal files, vendor binaries and remote PRs were outside the read. No raw run logs were supplied.; nothing run (a code review). **Checked and fine:** - The single commit has the required Assisted-by trailer. - No game-derived data, vendor binaries, third-party edits, build artifacts or generated shader edits appear in the diff. The public Vulkan interface is unchanged. - The row lives in the upscaler module, stays out of presets, has a frozen INI key and appears in the guide's menu order. All 13 translation tables add the row strings. - The normal replacement path uses the existing deferred retirement queue, and a newly made context resets temporal history through the existing prepare result. **Testing:** Read-only review. The author reports Windows RX 6600 gameplay and earlier WSL2 suite results of 720 passed, 70 failed and 115 skipped. No native Windows harness result or exact-head baseline comparison was supplied. The listed CI checks are still running. **Overlaps and merge order:** None found in the local history. The PR builds on the existing #1633 menu modules. Remote overlap and the cause of the reported non-mergeable status could not be verified from the checkout. </details> --- <sub>Automated code review by BeanBot (gpt-6.2-sol), on the pull request's opening. Nothing was run; a maintainer decides.</sub>
@ -0,0 +208,4 @@
if (read_u64(s.begin + i) != col_va) continue;
const uintptr_t vtable = s.begin + i + 8;
const Section* vt_section = im.section_at(vtable);
if (!vt_section || vt_section->code || !im.in_image(vtable, kVtableEntries * 8)) continue;
Member

L3: SizeOfImage bounds the reserved image range. A discarded, guarded or unreadable page can sit inside it, and the candidate vtable check only checks the section containing its first byte. A table at the end of a readable section can make the eight-entry read cross into the next mapping. Validate committed readable spans and executable targets against the live image mapping before parsing, and require the whole table to fit within its validated section.

**L3**: `SizeOfImage` bounds the reserved image range. A discarded, guarded or unreadable page can sit inside it, and the candidate vtable check only checks the section containing its first byte. A table at the end of a readable section can make the eight-entry read cross into the next mapping. Validate committed readable spans and executable targets against the live image mapping before parsing, and require the whole table to fit within its validated section.
@ -0,0 +234,4 @@
}
/// The 4.1.2 shape: entries 0, 1 and 3..7 are the same across every provider, entry 2 is distinct.
bool check_shape(const std::vector<Provider>& providers, const char** reason) {
Member

L4: The shared-entry pattern does not identify the tested ABI. An SDK update can keep those RTTI names and put a different per-provider virtual method in slot 2 while retaining seven shared entries. The patch would then invoke and replace that method as bool(self, device). Require a verified fingerprint for the tested AMD binary and its known provider layout before either calling a private entry or patching it. Add a rejection test for a layout that still passes the shared-entry comparison.

**L4**: The shared-entry pattern does not identify the tested ABI. An SDK update can keep those RTTI names and put a different per-provider virtual method in slot 2 while retaining seven shared entries. The patch would then invoke and replace that method as `bool(self, device)`. Require a verified fingerprint for the tested AMD binary and its known provider layout before either calling a private entry or patching it. Add a rejection test for a layout that still passes the shared-entry comparison.
@ -0,0 +273,4 @@
for (const Section& s : im.sections) {
if (s.code || !s.writable) continue;
for (size_t i = 0; i + 8 <= s.size; i += 8)
if (read_u64(s.begin + i) == vtable) candidates.emplace(s.begin + i, pi);
Member

L5: Any writable word holding a vtable address becomes a candidate object. An unrelated pointer table can contain an equally long ordered run, and native_model() then calls the private method with those inferred addresses as this. Remove the synthetic object calls and observe the real provider walk, or use an independently verified registry and instance layout for the exact supported DLL.

**L5**: Any writable word holding a vtable address becomes a candidate object. An unrelated pointer table can contain an equally long ordered run, and `native_model()` then calls the private method with those inferred addresses as `this`. Remove the synthetic object calls and observe the real provider walk, or use an independently verified registry and instance layout for the exact supported DLL.
@ -0,0 +453,4 @@
do {
ModuleScan m;
char path[MAX_PATH] = {};
if (GetModuleFileNameA(me.hModule, path, MAX_PATH)) m.name = path;
Member

L13 (minor): GetModuleFileNameA logs a lossy path on Windows and can silently truncate it at MAX_PATH. For example, a runtime under a long Unicode game folder may be misidentified in the log. Use the wide API with a growable buffer and convert to UTF-8, or log the verified module's basename.

**L13** (minor): `GetModuleFileNameA` logs a lossy path on Windows and can silently truncate it at MAX_PATH. For example, a runtime under a long Unicode game folder may be misidentified in the log. Use the wide API with a growable buffer and convert to UTF-8, or log the verified module's basename.
@ -0,0 +457,4 @@
else m.name = "?";
m.base = reinterpret_cast<uintptr_t>(me.modBaseAddr);
m.size = me.modBaseSize;
m.result = scan_image(me.modBaseAddr, me.modBaseSize);
Member

L2: A Toolhelp snapshot does not retain a module. An overlay DLL can unload between enumeration and the scan, or between the scan and the later calls and writes through the returned pointers. Resolve the AMD upscaler used by this loader, acquire a live reference before reading it and keep it for as long as any hook or original pointer can be called. Tie installation state to that lifetime.

**L2**: A Toolhelp snapshot does not retain a module. An overlay DLL can unload between enumeration and the scan, or between the scan and the later calls and writes through the returned pointers. Resolve the AMD upscaler used by this loader, acquire a live reference before reading it and keep it for as long as any hook or original pointer can be called. Tie installation state to that lifetime.
@ -0,0 +510,4 @@
const auto rollback = [&] {
for (auto it = patched.rbegin(); it != patched.rend(); ++it) {
DWORD back = 0;
VirtualProtect(it->slot, sizeof(void*), PAGE_READWRITE, &back);
Member

L7: Rollback writes to a slot even when making its page writable fails. The forward path also ignores failure to restore page protection, and grows two vectors after mutation has begun. A later slot failure can fault during rollback, or a protection failure can still report a successful install. Reserve bookkeeping before patching, check every protection change and restore, and maintain valid Auto delegation if a rollback cannot restore a slot.

**L7**: Rollback writes to a slot even when making its page writable fails. The forward path also ignores failure to restore page protection, and grows two vectors after mutation has begun. A later slot failure can fault during rollback, or a protection failure can still report a successful install. Reserve bookkeeping before patching, check every protection change and restore, and maintain valid Auto delegation if a rollback cannot restore a slot.
@ -0,0 +535,4 @@
h.vtable = p.vtable;
h.original = p.is_supported;
h.model = p.model;
p.vtable[2] = (void*)&is_supported_hook;
Member

L6: A caller can reach the patched slot before g_hooked_count publishes its record at line 544. In that window the hook returns false even for AMD's supported provider. The install mutex serializes installers only. Require a quiescent install, or publish immutable delegation records before redirecting any slot and keep them valid for callers already in flight during rollback. Enable Lite only after the module is fully patched.

**L6**: A caller can reach the patched slot before `g_hooked_count` publishes its record at line 544. In that window the hook returns false even for AMD's supported provider. The install mutex serializes installers only. Require a quiescent install, or publish immutable delegation records before redirecting any slot and keep them valid for callers already in flight during rollback. Enable Lite only after the module is fully patched.
@ -161,0 +169,4 @@
static const char* const t[] = {
"AMD's own choice for this GPU: FP8 on RDNA 4, full INT8 on RDNA 3 and Strix Halo, the Navi 2 INT8 model on RX 6000, "
"Lite on Strix Point and Van Gogh.",
"AMD's smaller FSR 4 network: less GPU time and memory with better quality than FSR 3. On a GPU AMD already runs Lite "
Member

L20 (minor): The help promises lower memory use and better quality than FSR 3 for every player. The supplied measurement covers only GPU time on an RX 6600. Qualify the help and translations to match the demonstrated trade-off, and make the guide explicit that forcing Lite requires a supported Windows runtime. Linux can keep the saved value but cannot force Lite.

**L20** (minor): The help promises lower memory use and better quality than FSR 3 for every player. The supplied measurement covers only GPU time on an RX 6600. Qualify the help and translations to match the demonstrated trade-off, and make the guide explicit that forcing Lite requires a supported Windows runtime. Linux can keep the saved value but cannot force Lite.
@ -1168,2 +1299,4 @@
if (!s.context || s.context->width != ow || s.context->height != oh || s.context->hdr != hdr ||
s.context->model != want) {
if (s.context && s.context->hdr != hdr) logf("vk: AMD FSR 4: made again: the game's native HDR turned %s", hdr ? "on" : "off");
release();
Member

L12: A model-only change retires a working Auto context before the replacement is known to work. Near the VRAM limit, Lite and its Auto retry can both fail while the previous Auto allocation is still in flight, and the caller then disables FSR 4. When the old context still matches the size and HDR requirements, retain and reuse it if the model change fails.

**L12**: A model-only change retires a working Auto context before the replacement is known to work. Near the VRAM limit, Lite and its Auto retry can both fail while the previous Auto allocation is still in flight, and the caller then disables FSR 4. When the old context still matches the size and HDR requirements, retain and reuse it if the model change fails.
@ -1018,2 +1105,4 @@
return no(FSR_NO_MODEL);
}
#ifdef _WIN32
const auto modules = probe_models(l.device_context);
Member

L18 (minor): Measure the default path and the live switch. The startup probe walks every loaded image and can hold load() while the main menu waits. A switch can also pay for an SDK create on a gameplay frame while the previous context is still resident. Compare base Auto, head Auto and head Lite on the same build settings, with at least two interleaved runs each. Include first probe time, the switch frame, peak RAM and VRAM, average FPS, 1% lows, median and p99 frame times. Include AMD RDNA 3 and 4 and the affected Tier 1 Windows and Linux paths.

**L18** (minor): Measure the default path and the live switch. The startup probe walks every loaded image and can hold `load()` while the main menu waits. A switch can also pay for an SDK create on a gameplay frame while the previous context is still resident. Compare base Auto, head Auto and head Lite on the same build settings, with at least two interleaved runs each. Include first probe time, the switch frame, peak RAM and VRAM, average FPS, 1% lows, median and p99 frame times. Include AMD RDNA 3 and 4 and the affected Tier 1 Windows and Linux paths.
@ -1035,6 +1127,12 @@ bool load() {
return no(FSR_NO_MODEL);
}
logf("vk: AMD FSR 4: upscalers %s; using %u.%u.%u", listed.c_str(), best[0], best[1], best[2]);
g_lite_available = dev.props.vendorID == 0x1002 && fsr4_models::hook_installed() &&
Member

L8: Availability means that any scanned DLL installed a Lite hook. The override ID still comes from the pre-hook Auto enumeration, and a successful ffxCreateContext is logged as Lite without checking which provider created it. With two runtime copies, or a loader that resolves the override to a specific provider, the requested Lite context can succeed using Auto and lose tuning. Bind availability and native-model state to the runtime servicing this loader, confirm the actual selected provider before reporting Lite, and leave the row unavailable when that cannot be established.

**L8**: Availability means that any scanned DLL installed a Lite hook. The override ID still comes from the pre-hook Auto enumeration, and a successful `ffxCreateContext` is logged as Lite without checking which provider created it. With two runtime copies, or a loader that resolves the override to a specific provider, the requested Lite context can succeed using Auto and lose tuning. Bind availability and native-model state to the runtime servicing this loader, confirm the actual selected provider before reporting Lite, and leave the row unavailable when that cannot be established.
@ -1048,0 +1159,4 @@
}
c->effective = effective;
const bool suppress = effective == FSR4_MODEL_LITE && fsr4_models::native_model() != fsr4_models::MODEL_LITE;
tune::suppress(suppress);
Member

L9: Suppression is a process-wide setting that remains in place after context creation. An older Auto context doing deferred resource creation while Lite is active would see Lite's setting, as would work in a Lite context after switching back. Scope the provider override to the intended creation and give each context its own Vulkan dispatch policy so its callbacks keep the correct behavior throughout its lifetime.

**L9**: Suppression is a process-wide setting that remains in place after context creation. An older Auto context doing deferred resource creation while Lite is active would see Lite's setting, as would work in a Lite context after switching back. Scope the provider override to the intended creation and give each context its own Vulkan dispatch policy so its callbacks keep the correct behavior throughout its lifetime.
@ -1075,0 +1196,4 @@
tune::suppress(false);
fsr4_models::set_model(fsr4_models::SELECT_AUTO);
c->tuned = tune::on();
c->context = nullptr;
Member

L11: The code treats an error with a non-null output handle as a failure, then clears the handle before retrying. If AMD has returned a partially created context, the retry can leak its resources and fail under the same memory pressure. Keep each attempt's handle separate and clean up any returned context according to the SDK ownership contract before attempting Auto. Cover that case with the fake runtime.

**L11**: The code treats an error with a non-null output handle as a failure, then clears the handle before retrying. If AMD has returned a partially created context, the retry can leak its resources and fail under the same memory pressure. Keep each attempt's handle separate and clean up any returned context according to the SDK ownership contract before attempting Auto. Cover that case with the fake runtime.
@ -1168,3 +1301,3 @@
if (s.context && s.context->hdr != hdr) logf("vk: AMD FSR 4: made again: the game's native HDR turned %s", hdr ? "on" : "off");
release();
s.context = make_context(ow, oh, hdr);
s.context = make_context(ow, oh, hdr, want);
Member

L10: A live model change can call tune::device_proc again while detached Auto unroll jobs are active. Its plain writes to function pointers such as real_csm at line 1002 race with reads in build_unrolled() at line 760, even if the address being written is the same. Initialize the tuning dispatch table once before workers start and make subsequent lookups read-only.

**L10**: A live model change can call `tune::device_proc` again while detached Auto unroll jobs are active. Its plain writes to function pointers such as `real_csm` at line 1002 race with reads in `build_unrolled()` at line 760, even if the address being written is the same. Initialize the tuning dispatch table once before workers start and make subsequent lookups read-only.
@ -1238,1 +1371,4 @@
logf("vk: AMD FSR 4: its dispatch failed (%d)", (int)r);
if (s.context->effective == FSR4_MODEL_LITE && !g_lite_failed.exchange(true)) {
logf("vk: AMD FSR 4: the Lite model's dispatch failed; using Auto");
release(); // (the next prepare makes an Auto context; the saved value stays Lite)
Member

L1: The next-frame Auto retry is unreachable in gameplay. After Lite dispatch fails, evaluate() releases the context and returns false. draw_engine() at vk_temporal.cpp:1679 then marks FSR 4 failed because its engine has no retry policy. Return a recoverable result for this transition so the game's scaler can draw the failed frame and the next frame builds Auto with fresh history. Preserve the existing handling of a subsequent Auto failure.

**L1**: The next-frame Auto retry is unreachable in gameplay. After Lite dispatch fails, `evaluate()` releases the context and returns false. `draw_engine()` at vk_temporal.cpp:1679 then marks FSR 4 failed because its engine has no retry policy. Return a recoverable result for this transition so the game's scaler can draw the failed frame and the next frame builds Auto with fresh history. Preserve the existing handling of a subsequent Auto failure.
@ -0,0 +35,4 @@
// default, so the image's pages are made executable here)
Image() {
DWORD old = 0;
if (!VirtualProtect(bytes.data(), bytes.size(), PAGE_EXECUTE_READWRITE, &old)) {
Member

L14 (minor): The Windows fixture makes pages of a std::vector heap allocation executable and writable, potentially including unrelated heap allocations. It never restores their protection and executes generated instructions without flushing the instruction cache. Use a dedicated virtual-memory allocation with separate code and data page protections, flush the instruction cache before calling the stubs, and give the fixture a lifetime that covers the installed hooks.

**L14** (minor): The Windows fixture makes pages of a `std::vector` heap allocation executable and writable, potentially including unrelated heap allocations. It never restores their protection and executes generated instructions without flushing the instruction cache. Use a dedicated virtual-memory allocation with separate code and data page protections, flush the instruction cache before calling the stubs, and give the fixture a lifetime that covers the installed hooks.
@ -0,0 +29,4 @@
with tempfile.TemporaryDirectory() as td:
exe = Path(td) / ("fsr4_models_test.exe" if os.name == "nt" else "fsr4_models_test")
cmd = [args.cxx, "-std=c++20", "-O2", "-I", str(VK), str(VK / "fsr4_models.cpp"), str(HARNESS), "-o", str(exe)]
Member

L15 (minor): The harness compiles with the selected C++ compiler, but it can fail to start with llvm-mingw when the compiler's runtime DLL directory is off PATH. The adjacent FSR 4 tuning test links statically on Windows for this case. Apply the same Windows toolchain handling and keep the existing --cxx override.

**L15** (minor): The harness compiles with the selected C++ compiler, but it can fail to start with llvm-mingw when the compiler's runtime DLL directory is off PATH. The adjacent FSR 4 tuning test links statically on Windows for this case. Apply the same Windows toolchain handling and keep the existing `--cxx` override.
@ -201,6 +201,19 @@ def source_problems():
problems.append("tune::close does not stop the jobs and wait for the one building")
if "void fsr4_close() { tune::close(); }" not in src or src.count("void fsr4_close()") != 2:
problems.append("fsr4_close is not defined in both builds (with and without FSR)")
for fn, real in [("create_buffer", "real_cb"), ("create_shader_module", "real_csm"),
Member

L16: The new source checks can pass if the tuning guard is inverted, and there is no integrated test that can catch the broken Auto retry. Add a game-free test with fake SDK and Vulkan callbacks that drives Auto to Lite and back, observes context creation and tuning calls, then injects Lite creation and dispatch failures through the temporal caller. Assert that fallback occurs once and that the next successful context starts with fresh history.

**L16**: The new source checks can pass if the tuning guard is inverted, and there is no integrated test that can catch the broken Auto retry. Add a game-free test with fake SDK and Vulkan callbacks that drives Auto to Lite and back, observes context creation and tuning calls, then injects Lite creation and dispatch failures through the temporal caller. Assert that fallback occurs once and that the next successful context starts with fresh history.
@ -64,2 +64,4 @@
#include "display_modes.h" // (gpu/vk/: the sizes the Output resolution offers)
#include "vk_fsr4.h"
using namespace mkwvk;
namespace mkwvk { int fsr4_model_env() { return -1; } void fsr4_set_model(Fsr4Model) {} }
Member

L17 (minor): The fixture stubs the model override as always absent and discards the live update. Both new menu_init calls can be removed and this test still passes. Exercise a saved Auto value with a Lite override and assert that the live value is Lite while an unrelated save keeps Auto. Also assert that a saved Lite value reaches the engine with the environment cleared.

**L17** (minor): The fixture stubs the model override as always absent and discards the live update. Both new `menu_init` calls can be removed and this test still passes. Exercise a saved Auto value with a Lite override and assert that the live value is Lite while an unrelated save keeps Auto. Also assert that a saved Lite value reaches the engine with the environment cleared.
Some checks are pending
CI / Conventional Commit title (pull_request) Blocked by required conditions
CI / Python source checks (ubuntu-24.04, part 1/2) (pull_request) Blocked by required conditions
CI / Python source checks (ubuntu-24.04, part 2/2) (pull_request) Blocked by required conditions
CI / SPIR-V fixtures (global narrow load/store, direct surface atomics, F2I immediate) (pull_request) Blocked by required conditions
CI / Workflow lint (actionlint) (pull_request) Blocked by required conditions
CI / Website build and checks (pull_request) Blocked by required conditions
Guide screenshots / Guide images and local UI source review (pull_request) Blocked by required conditions
Third-party parts / As their notes say (pull_request) Blocked by required conditions
Third-party parts / The same as upstream (pull_request) Blocked by required conditions
Website / Website on Cloudflare (pull_request) Blocked by required conditions
PR description / Description follows the template (pull_request_target) Successful in 6s
Website / Website on Cloudflare (pull_request_target) Has been skipped
This pull request is marked as a work in progress.
Some workflows are waiting to be reviewed.
This branch is out-of-date with the base branch
View command line instructions

Checkout

From your project repository, check out a new branch and test the changes.
git fetch -u 'feat/fsr4-lite:prfd-feat/fsr4-lite'
git switch 'prfd-feat/fsr4-lite'
Sign in to join this conversation.
No reviewers
No milestone
No project
No assignees
2 participants
Notifications
Due date
The due date is invalid or out of range. Please use the format "yyyy-mm-dd".

No due date set.

Dependencies

No dependencies set

Reference
MagicBeanDev/Longshot!1989
No description provided.