[FIocker]

The upscaling (#1831) and neural rendering (#1833) work needs menu rows, renderer hooks and build changes; this PR is those. It also fixes two upload hitches and adds the lock diagnostics docs/performance.md is built on.

#1831, #1833 and this PR build only together. This PR includes vk_dlssg.h, vk_ffx.h, framegen_methods.h, vk_dlssnr*.h and vk_opennr.h, and the other two use this one's rows and hooks. Please review and merge the three as one unit; FIocker:super-build/all is the three together, and it's what was built and tested. #1830 is separate and builds alone.

Part	Branch	Builds
Shader fixes	super-build/shaders (#1830)	on its own
Upscaling and frame generation	super-build/upscaling-framegen (#1831)	with #1833 and this PR
DLSS 5 neural rendering	super-build/neural (#1833)	with #1831 and this PR
Menu, renderer and build plumbing (this PR)	super-build/core (#1834)	with #1831 and #1833

What it does

Insert menu (vk_menu.*):

• The Neural tab is TAB_NEURAL = 5, after main's Audio tab (TAB_COUNT = 6), with its rows, including Stage (after or before the upscaler).
• Saved keys: the tab's rows are saved as nr_<name> (nr_model, nr_style, nr_preset, …) by one rule in ini_key. The switch itself keeps neural_rendering.
• Frame generation method: Auto, NVIDIA DLSS or Built-in, of what runs here (framegen_methods.h).
• Upscaler row: AMD's FidelityFX FSR 4 counts as FSR 4 where main's FSR 4 engine isn't built (ffx::upscale_supported).

Renderer:

• vk_main.cpp: device features and extensions for NGX and OpenDLSS-NR.
• vk_replay.cpp: the present's NR pass is skipped when the chain already ran it. The headless present now uses the same call and guard (nr::apply behind !dlss_nr_before_hud()), so a headless run with Stage = Before upscaler no longer applies NR twice.
• vk_draw.cpp: the scene's view-projection capture, and skinned-draw detection for the new actor formats.
• The scene depth handed to NR is now the largest depth target of this frame, not the largest ever (scene_depth.frame). scene_frame_inputs answers only for this frame's or the last one's. scene_depth and scene_vp are written under a lock, since RX_VK_PARALLEL's workers resolve targets too.

Upload hitches (vk_texture.cpp, vk_memory.cpp):

• No import for one upload: upload_unswizzle_gpu uses the GPU path only from a pool that is already imported (memory_guest_contained). An import made for one upload took 22-25 ms on the replay thread, and three of them, when the camera turns to the Deku Tree, made a 60 ms hitch. RX_VK_UNSWIZZLE_IMPORT=1 brings back the import.
• Threaded CPU unswizzle from 4 MB instead of 16 MB. The scene-size targets uploaded at that moment (2779 × 1318 RGBA8, 15 MB) took about 20 ms each on one thread.

Lock diagnostics (vk_internal.h, vk_memory.cpp, vk_texture.*): ParLock counts by file and line, and memory_pool_lock_reasons and texture_touch_misses count why the two busiest sites take the lock. These cost nothing while parallel replay is off. docs/performance.md has the numbers; it no longer contradicts itself and names functions instead of line numbers.

NVIDIA's shader disk cache is opt-in: nvidia_shader_cache_gb=<GB> in hookshot_options.ini, with no menu row. It sets __GL_SHADER_DISK_CACHE_SIZE and turns trimming off, only when asked and only where the environment doesn't set them. It's documented in docs/performance.md.

Host: nn_vi.c's HOS_DISPLAY_VRR=1 (opt-in) tells the game the display does variable refresh, which lifts its 60 fps limiter, and SetDisplay120HzRefreshRateCapability is answered.

Build (nx2/build.py): NGX flags for every file that sees NVIDIA's headers (vk_dlss, vk_ngx, vk_dlssg, vk_dlssnr_ngx), OpenDLSS-NR's PTX kernels, and Streamline's SDK path for the NR backend.

Tests: main's fixtures that cut our changed functions out get the stubs they now need (direct_buffer_writes, dlss_sdk_cache, framegen_preset, present_geometry_trace, present_source_crop, stick_dead_zone, temporal_jitter, upscale_independent, audio_output_menu, rumble_menu, shader_stamp_cache). None of their checks are weakened.

Taken out since the first review

Everything below is gone from the series:

• Fork release and update plumbing: semantic-release and --repository-url, package.json, repository.txt / HOOKSHOT_REPOSITORY, the prebuilt-renderer path, and the Linux workflow's NGX SDK fetch. Main has moved to release-please, and the rest was inactive on Windows anyway.
• The old Upscaling-row migration (old_upscaling 3/4) and the extra tx_upscale labels. tx_upscale is main's.
• hos_sdl_refresh_hz() and the HOS_VSYNC_HZ mention: neither was used.
• The duplicate includes in vk_main.cpp and vk_menu.cpp.
• Streamline frame generation: its row and vk_slfg.cpp, nx2_sl.h and its hooks in Reflex, the device and the present (see feat(vk): DLSS frame generation (NGX), more actor motion for DLSS, FidelityFX FSR 4 fallback (builds with #1833, #1834) #1831).

Testing

• Tests: tests/run_all.py on the three together (super-build/all), locally (clang targeting MSVC, the CPM Vulkan headers, glslang 11.12): 328 passed. The only failures are the same 18 that main at d7a7306 has on this machine, all down to this PC's toolchain.
• Build and game: Windows 11, NVIDIA RTX, OoT v131072, MSVC build with the NGX SDK; results in feat(vk): DLSS frame generation (NGX), more actor motion for DLSS, FidelityFX FSR 4 fallback (builds with #1833, #1834) #1831 and feat(vk): DLSS 5 neural rendering through NGX, in the temporal chain before the HUD #1833.
• Screenshots of the Neural tab and the Frame generation method row will follow in a comment.

🤖 Generated with Claude Code

[FIocker]

Thanks for going through the whole series; it changed a lot. I rebuilt it on main at d7a7306. The descriptions are rewritten.

How the series fits together: #1830 now builds on its own. Its vk_texture.cpp changes, memory_guest_contained and texture_touch_misses moved here, and I checked it with the full test suite locally. #1831, #1833 and this one still need each other, and all three descriptions now say so plainly, at the top, with a table. Splitting them along file lines would mean shipping half-wired rows, so I'd rather they're reviewed as one unit; super-build/all is the three together. The vk_texture.h line in the old description was wrong; it's fixed.

What the rebase hit:

• Tabs: TAB_NEURAL = 5 after main's Audio tab, TAB_COUNT = 6, and Audio kept in tab_titles.
• Release plumbing gone: semantic-release, --repository-url, the npm dependency, package.json and .releaserc.json are all gone (main uses release-please). So are linux-appimage.yml's changes, including the NGX SDK fetch.
• run_all.py: main's per-test RUN_ALL_ARGS / RUN_ALL_SKIP. The opennr tests and test_framegen_methods.py declare theirs; nothing goes in the retired tables, and the duplicate test_fsr3_shaders.py entry went with them.

Bugs:

• test_rumble_menu.py: the leftover wheel line is gone.
• scene_depth: now the largest depth of this frame (scene_depth.frame), and scene_frame_inputs answers only for this frame's or the last one's. Both scene_depth and scene_vp are written under a lock.
• Headless present: nr::apply behind !dlss_nr_before_hud(), the same call and guard as present_image.
• HOS_VSYNC_HZ and hos_sdl_refresh_hz(): both removed, since they were unused.

Your list for NotRexed:

• FSR frame generation: main itself made FSR 3.1's interpolation opt-in on 2026-11-10 (RX_VK_FSR_FG=1, FrameGen and FSR causes blur and ghosting while moving #1635), so the series follows main there. feat(vk): DLSS frame generation (NGX), more actor motion for DLSS, FidelityFX FSR 4 fallback (builds with #1833, #1834) #1831 says why FidelityFX FSR 4's isn't offered either, and the method enum keeps both values so a saved choice reads as Auto.
• Streamline frame generation: removed completely (vk_slfg.cpp, nx2_sl.h, its row, and its hooks in Reflex, the device and the present).
• Linux NGX SDK in CI: removed.
• The 10 GB NVIDIA cache: now opt-in, nvidia_shader_cache_gb=<GB> in the options file, documented in docs/performance.md.
• Fork-aware updates and the prebuilt renderer: removed from the series.

Nits:

• Upscaling migration: old_upscaling and the extra tx_upscale labels are gone.
• ini keys: the Neural rows are saved as nr_<name> by one rule in ini_key.
• Comments: the nr_backend comment lists 0-3.
• Docs: docs/performance.md no longer contradicts itself and names functions, not line numbers.
• Includes: the duplicates in vk_main.cpp and vk_menu.cpp are gone.
• Screenshots of the Neural tab and the Frame generation method row will follow from an in-game run.

Tests: tests/run_all.py on the three together, locally (clang targeting MSVC, the CPM Vulkan headers, glslang 11.12), gives 328 passed. The failures are only the same 18 main has on this machine. Main's fixtures that cut our functions out now get stubs for them, with no checks weakened.

[Joey-Magnus]

With #1831 closed in favour of #2118, this plumbing no longer has the series it was built for, so I'd close it. It includes vk_dlssg.h, vk_ffx.h and framegen_methods.h from #1831 and vk_dlssnr*.h / vk_opennr.h from #1833. It also conflicts with main in seven files: vk_menu.cpp, vk_menu.h, vk_replay.cpp, nx2/build.py, two present tests and test_shader_stamp_cache.py. Two pieces stand on their own and would be welcome as small PRs on main. One is the upload hitch fix: no pool import for a single upload_unswizzle_gpu, and the threaded CPU unswizzle from 4 MB. The other is the lock diagnostics. If the vk_main.cpp default for __GL_SHADER_DISK_CACHE_SIZE / SKIP_CLEANUP comes back, it needs its own PR and a maintainer decision, as hckrman101 noted, since it changes disk use for every Windows NVIDIA player.

[FIocker]

Closing, as suggested. With #1831 gone this plumbing has no series left, so its pieces went where they belong:

• The upload hitch fix is perf(vk): no pool import for one GPU unswizzle, and the CPU unswizzle threaded from 4 MB #2785: no pool import for one GPU unswizzle (now read with env_on), and the CPU unswizzle threaded from 4 MB.
• The lock diagnostics and docs/performance.md are feat(vk): ParLock counted by site, and why its two busiest sites take it #2786.
• The menu rows, nr_settings() and the chain and present hooks that neural rendering needs are being folded into feat(vk): DLSS 5 neural rendering through NGX, in the temporal chain before the HUD #1833, rebuilt on main on top of feat(dlss): DLSS Frame Generation through NGX, evenly paced at 3x and up #2118's NGX session.
• The NVIDIA shader-cache default isn't coming back in any of these. If it ever does, it'll be its own PR for a maintainer decision.

Thanks for the reviews.