Fork touchpoints

July 30, 2026 · View on GitHub

This index lists every upstream file the fork touches. It is the authoritative inventory of fork-vs-upstream surface area, maintained as fork edits are added or removed.

See docs/CONTRIBUTING.md for the fork-touchpoint discipline this index supports.

Conventions

Fork-owned file naming

Fork-owned files use the rtx_fork_* prefix with a subsystem suffix (e.g. rtx_fork_api_entry.cpp, rtx_fork_atmosphere.cpp, rtx_fork_overlay.cpp, rtx_fork_light.cpp). Single prefix keeps the convention simple and grep-friendly.
All fork-owned files live under src/dxvk/rtx_render/ (or the subsystem-appropriate equivalent directory).
Hook functions are declared in the fork_hooks:: namespace (src/dxvk/rtx_render/rtx_fork_hooks.h) and implemented in their respective fork-owned .cpp files.

Fridge-list invariant

Every edit to an upstream file must have a fridge-list entry in the same commit. The PR-template bullet reminds contributors; a future CI check will enforce it if discipline slips.

Entry types

Hook — upstream file contains a one-line call into fork-owned code. The fork logic itself lives in the fork-owned file referenced by the entry.
Inline tweak — upstream file contains a small fork-introduced change (typically <= 20 LOC) that was not worth lifting into its own fork file.

Upstream files touched by the fork

public/include/remix/remix.h

Pre-refactor fork footprint: +101 / -9 LOC (audit 2026-04-18)

Category: migrate

Block at Interface class (method declarations) — ~13 LOC, planned target N/A (public header) in N/A (public header). Declares fork-added C++ wrapper methods: CreateMeshBatched, GetUIState, SetUIState, AddTextureHash, RemoveTextureHash, dxvk_GetTextureHash, CreateLightBatched, UpdateLightDefinition, SetGameValue.
Block at Interface::CreateMeshBatched (inline definition) — ~9 LOC, planned target N/A (public header) in N/A (public header). Inline C++ wrapper that calls m_CInterface.CreateMeshBatched for the batched mesh submission API slot.
Block at Interface::GetUIState / Interface::SetUIState (inline definitions) — ~16 LOC, planned target N/A (public header) in N/A (public header). Inline C++ wrappers for UI state query/set API, guarding on nullptr slot before dispatching.
Block at Interface::AddTextureHash / Interface::RemoveTextureHash (inline definitions) — ~16 LOC, planned target N/A (public header) in N/A (public header). Inline C++ wrappers for texture-hash category mutation API.
Block at Interface::dxvk_GetTextureHash (inline definition) — ~13 LOC, planned target N/A (public header) in N/A (public header). Inline C++ wrapper that retrieves the DXVK image hash for a D3D9 texture via the fork's dxvk-specific extension slot.
Block at Interface::CreateLightBatched / Interface::UpdateLightDefinition (inline definitions) — ~22 LOC, planned target N/A (public header) in N/A (public header). Inline C++ wrappers for batched light creation and deferred light-definition update.
Block at UIState enum (file scope) — ~5 LOC, planned target N/A (public header) in N/A (public header). Adds UIState C++ enum mirroring remixapi_UIState for the C++ API surface.
Block at Interface::SetGameValue (declaration + inline definition) — ~8 LOC, planned target N/A (public header) in N/A (public header). C++ wrapper for the remixapi_SetGameValue C API slot introduced in workstream 10 (plugin-injected game-state write). Wrapper guards on nullptr vtable slot before dispatching, matching the SetConfigVariable shape. Companion readers are graph components GameValueReadBool / GameValueReadNumber; backing store lives in rtx_fork_game_state.h.
Block at remixapi_Interface static_assert updates (file scope) — ~3 LOC (three separate assert sizes), planned target N/A (public header) in N/A (public header). Updates sizeof(remixapi_Interface) static_asserts in the C++ header to match each successive vtable extension (208 → 240 → 272 → 280 → 288).

public/include/remix/remix_c.h

Pre-refactor fork footprint: +154 / -29 LOC (audit 2026-04-18)

Category: migrate

Block at REMIXAPI_VERSION_MAJOR/MINOR/PATCH (file scope) — ~3 LOC, planned target N/A (public header) in N/A (public header). Sets the Remix Plus ABI version to 0.1000.0 (reserved MINOR 1000), distinct from stock NVIDIA 0.6.x. Because isVersionCompatible treats each minor as breaking while MAJOR==0, this makes the runtime reject binaries built against stock Remix 0.6.x or older Remix Plus 0.6.x — whose remixapi_Interface layout and category-bit ABI differ. Bump MINOR on any further breaking ABI change.
Block at remixapi_StructType enum (file scope) — ~3 LOC, planned target N/A (public header) in N/A (public header). Adds REMIXAPI_STRUCT_TYPE_TEXTURE_INFO, INSTANCE_INFO_PARTICLE_SYSTEM_EXT, and INSTANCE_INFO_GPU_INSTANCING_EXT enumerators.
Block at remixapi_TextureHandle typedef (file scope) — ~1 LOC, planned target N/A (public header) in N/A (public header). Declares the opaque remixapi_TextureHandle_T* handle type for the texture upload API.
Block at REMIXAPI_INSTANCE_CATEGORY_BIT_* enum (file scope) — ~16 LOC, planned target N/A (public header) in N/A (public header). Bit values match upstream NVIDIA exactly (reverted 2026-06-27 from an earlier fork build that shifted IGNORE_ALPHA_CHANNEL to bit 8 to mirror the internal InstanceCategories order). The C↔internal mapping in toRtCategories() is by-name, so the public bit values are free to match upstream and now do. No remaining fork delta — the enum now matches upstream exactly. (The misleadingly-named LEGACY_EMISSIVE alias of bit 24 / SMOOTH_NORMALS was removed 2026-06-28: its name implied emissive behavior but it routed to SmoothNormals, so callers got a silent wrong-category result; removing it converts that into a compile error.)
Block at IDirect3DTexture9 forward declaration (file scope) — ~1 LOC, planned target N/A (public header) in N/A (public header). Forward-declares IDirect3DTexture9 so the dxvk-extension function signatures compile without pulling in d3d9 headers.
Block at remixapi_InstanceInfo.isDynamic field (struct) — ~1 LOC, planned target N/A (public header) in N/A (public header). Adds isDynamic bool field to remixapi_InstanceInfo to control temporal accumulation behavior.
Block at remixapi_InstanceInfo.ignoreViewModel field (struct) — ~1 LOC, planned target N/A (public header) in N/A (public header). Adds ignoreViewModel bool field to remixapi_LightInfo so API-submitted lights can opt out of view-model geometry lighting.
Block at PFN_remixapi_AddTextureHash / PFN_remixapi_RemoveTextureHash typedefs (file scope) — ~8 LOC, planned target N/A (public header) in N/A (public header). Declares function-pointer types for texture-hash category mutation (add/remove a texture hash from a named option set).
Block at remixapi_Format enum + remixapi_TextureInfo struct (file scope) — ~28 LOC, planned target N/A (public header) in N/A (public header). Declares the texture upload type system: format enum mapping to VkFormat values, and the remixapi_TextureInfo struct carrying pixel data for CreateTexture.
Block at PFN_remixapi_CreateTexture / PFN_remixapi_DestroyTexture typedefs (file scope) — ~6 LOC, planned target N/A (public header) in N/A (public header). Declares function-pointer types for the texture upload/destroy lifecycle.
Block at PFN_remixapi_dxvk_GetTextureHash typedef (file scope) — ~4 LOC, planned target N/A (public header) in N/A (public header). Declares the dxvk-specific extension function to retrieve the GPU image hash from a D3D9 texture object.
Block at PFN_remixapi_CreateMeshBatched typedef (file scope) — ~7 LOC, planned target N/A (public header) in N/A (public header). Declares the batched mesh-creation function-pointer type and stub comment noting slot is currently nullptr.
Block at remixapi_UIState enum + remixapi_GetUIState / remixapi_SetUIState declarations (file scope) — ~16 LOC, planned target N/A (public header) in N/A (public header). Declares the UI state enum and the two API entry points for reading/setting the ImGui overlay visibility level.
Block at PFN_remixapi_DrawScreenOverlay typedef (file scope) — ~9 LOC, planned target N/A (public header) in N/A (public header). Declares the function-pointer type for compositing a plugin-supplied pixel buffer over the final frame.
Block at PFN_remixapi_BridgeCallback + PFN_remixapi_RegisterCallbacks + PFN_remixapi_AutoInstancePersistentLights + PFN_remixapi_UpdateLightDefinition (file scope) — ~22 LOC, planned target N/A (public header) in N/A (public header). Declares frame-boundary callback registration, the persistent-light auto-instance helper, and the deferred light-definition update function.
Block at PFN_remixapi_CreateLightBatched typedef (file scope) — ~4 LOC, planned target N/A (public header) in N/A (public header). Declares the batched light-creation function-pointer type.
Block at PFN_remixapi_dxvk_GetSharedD3D11TextureHandle typedef (file scope) — ~5 LOC, planned target N/A (public header) in N/A (public header). Declares the dxvk-specific extension function for the DX11 shared-texture export path (stub in this fork; slot populated for ABI layout compatibility).
Block at PFN_remixapi_SetGameValue typedef (file scope) — ~14 LOC (including the contract doc block), planned target N/A (public header) in N/A (public header). Declares the function-pointer type for the plugin-injected game-state write API introduced in workstream 10. The entrypoint stores a single string/string pair under a caller-chosen key in a fork-owned thread-safe map; graph components GameValueReadBool / GameValueReadNumber read those values by name. The contract doc block above the typedef describes key/value semantics, validation, and lifetime (store survives Shutdown / re-init).
Block at remixapi_Interface vtable additions (struct fields) — ~15 LOC spread across the vtable struct, planned target N/A (public header) in N/A (public header). Appends new function-pointer slots to remixapi_Interface: AddTextureHash, RemoveTextureHash, CreateTexture, DestroyTexture, dxvk_GetTextureHash, CreateMeshBatched, GetUIState/SetUIState, DrawScreenOverlay, RegisterCallbacks, AutoInstancePersistentLights, UpdateLightDefinition, CreateLightBatched, dxvk_GetSharedD3D11TextureHandle, SetGameValue. 2026-06-27: the upstream SetCameraMediumMaterial slot was moved out of the middle of the struct (it had been inherited between SetupCamera and DrawInstance) to immediately after Present, mirroring upstream's canonical tail layout (upstream 2bac8874); fork slots remain appended after it. Size-neutral move — the sizeof sentinel is unchanged. An append-at-end warning comment was restored above the struct.

src/d3d9/d3d9_device.cpp

Pre-refactor fork footprint: +5 / -5 LOC (audit 2026-04-18)

Category: index-only

Inline tweak at D3D9DeviceEx::PresentEx (~line 3748) — 4-line addition for [RTX-Diag] entry log on PresentEx. Logs hwnd override and swapchain pointer at the top of PresentEx to correlate with the Remix API present chain during diagnostics.

src/d3d9/d3d9_rtx.cpp

Pre-refactor fork footprint: +11 / -11 LOC (audit 2026-04-18)

Category: index-only

Inline tweak at D3D9Rtx::EndFrame (~line 1216) — 5-line addition for [RTX-Diag] entry log on EndFrame. Logs targetImage pointer and callInjectRtx flag at the top of EndFrame, plus a second log after the CS lambda is emitted, to trace the frame-end dispatch chain.

src/d3d9/d3d9_swapchain.cpp

Pre-refactor fork footprint: +14 / -4 LOC (audit 2026-04-18)

Category: index-only

Inline tweak at D3D9SwapChainEx::Present (~line 441) — 4-line addition for [RTX-Diag] entry log. Logs hDestWindowOverride pointer at the top of D3D9SwapChainEx::Present to trace the native present path during diagnostics.
Inline tweak at D3D9SwapChainEx::Present (after remix API call site) — 6-line addition for remixapi_AutoInstancePersistentLights flush. Calls remixapi_AutoInstancePersistentLights() once per frame on the native D3D9 present path so persistent lights submitted via the Remix C API are auto-instanced even when the caller bypasses remixapi_Present. The callee early-outs cheaply for native-only consumers (see its body entry below), so this call is effectively free unless the C-API light path is in use.

src/d3d9/d3d9_swapchain_external.cpp

Pre-refactor fork footprint: +5 / -5 LOC (audit 2026-04-18)

Category: index-only

Inline tweak at D3D9SwapchainExternal::Present (~line 44) — 4-line addition for [RTX-Diag] entry log. Logs VkImage pointer and framebuffer extent at the top of D3D9SwapchainExternal::Present to trace the external swapchain present path during diagnostics.

src/dxvk/dxvk_limits.h

Category: index-only

Inline tweak at the MaxPushConstantSize enum value (~line 21) — 1-LOC value change + multi-line rationale comment. Bumps MaxPushConstantSize from 128 to 256. The fork-side tonemap apply args struct (ToneMappingApplyToneMappingArgs in src/dxvk/shaders/rtx/pass/tonemap/tonemapping.h) grew to 144 bytes across Workstream 2 commits 3–4 (Hable + AgX operator params). With the old cap, DxvkContext::pushConstants() silently overflowed the per-bank storage by 16 bytes on every Global / Direct tonemap dispatch, corrupting adjacent bank state and crashing the NVIDIA driver later. Every RTX-class GPU reports maxPushConstantsSize >= 256; pipeline layouts derive per-shader push-constant ranges from shader reflection, so the larger cap is harmless to existing shaders.

src/dxvk/imgui/dxvk_imgui.cpp

Pre-refactor fork footprint: +236 / -71 LOC (audit 2026-04-18) Post-refactor footprint: 3 hook call sites + #include "rtx_render/rtx_fork_hooks.h" (migrated 2026-04-18)

Category: migrate

Block at ImGUI::wndProcHandler (top of function + dispatch points) — ~30 LOC, planned target fork_hooks::wndProcHandlerDiag in rtx_fork_overlay.cpp. REVERTED before migration (commit 664a9ba4). Not present in current file. No hook created.
Block at ImGUI::processHotkeys (top) — ~10 LOC, planned target fork_hooks::processHotkeysDiag in rtx_fork_overlay.cpp. REVERTED before migration (commit 664a9ba4). Not present in current file. No hook created.
Block at ImGUI::checkHotkeyState (alt-chord logging branch) — ~22 LOC, planned target fork_hooks::checkHotkeyStateDiag in rtx_fork_overlay.cpp. REVERTED before migration (commit 664a9ba4). Not present in current file. No hook created.
Block at ImGUI::wndProcHandler (context pin at entry) — ~2 LOC. Migrated to fork_hooks::imguiContextPin in rtx_fork_overlay.cpp. Pins ImGui and ImPlot contexts at the top of wndProcHandler to prevent context corruption when plugin activity drifts GImGui off the dev menu's context between frames. Call site passes m_context and m_plotContext directly — no friend declaration needed.
Block at ImGUI::showRenderingSettings (sky mode UI section) — ~154 LOC. Migrated to fork_hooks::showAtmosphereUI in rtx_fork_atmosphere.cpp. Adds the Sky Mode combo (Skybox Rasterization / Numos), atmosphere preset buttons (Earth, Mars, Clear Sky, Polluted/Hazy, Alien World, Desert Planet), and the full atmosphere parameter tree (sun, density sliders, advanced coefficients) under the Sky Tuning collapsing header. The skyModeCombo static was moved from dxvk_imgui.cpp into the fork-owned atmosphere file. No friend declaration needed.
Block at ImGUI::showMainMenu (wrapper tab handling) — ~6 LOC. Partially migrated to fork_hooks::wrapperTabDraw in rtx_fork_overlay.cpp. The kTab_Wrapper guard (remixapi_imgui_HasDrawCallback() check + continue) remains as an inline tweak in the tab loop (structural control flow, not extractable). The case body (remixapi_imgui_InvokeDrawCallback()) is wrapped as fork_hooks::wrapperTabDraw(). No friend declaration needed.
Hook at tonemapper ImGui settings → fork_hooks::showTonemapOperatorUI / fork_hooks::showLocalTonemapOperatorUI in rtx_fork_tonemap.cpp. Also remove the standalone RemixGui::Checkbox("Use Legacy ACES", ...) at ~line 3888 — its RtxOption rtx.useLegacyACES is being deleted. Operator combo + per-operator sliders replace the old ACES checkbox. "Use Legacy ACES" reachable via TonemapOperator::ACESLegacy enum value.
Inline tweak at ImGUI::showRenderingSettings "Tonemapping" header — removed the Tonemapping Mode combo (Global / Local / Direct) and the standalone "User Brightness" / "User Brightness EV Range" sliders. The header body is now a single always-visible metaToneMapping().showImguiSettings() call between two separators. Tuning Mode (tone curve sliders) is also removed from the panel. 2026-05-13 tonemap refactor: mode selector removed; operator dropdown is now the primary control. 2026-05-15: local tonemap path removed entirely, so no per-path UI gate remains.

src/dxvk/imgui/dxvk_imgui_about.cpp

Category: index-only

Rationale: Fork additions are string-literal entries in the in-game About panel's "GitHub Contributors" list — they live inside a curly-braced initializer list and can't be lifted into a separate TU.

Inline tweak at ImGuiAbout::Credits::Credits constructor (GitHub Contributors string list, ~lines 91-107) — one string literal per fork contributor, sorted alphabetically per the inline comment. Per CONTRIBUTING.md, contributors add their own entry when their PR adds something visible. Tracked here per the fridge-list invariant. Each entry is in the format "FirstName 'Handle' LastName" or handle-only ("BrunchyChineapple", "Dayton 'watbulb'"). The list is the canonical record of community contributors visible in the About panel.

src/dxvk/imgui/dxvk_imgui.h

Pre-refactor fork footprint: +2 / -1 LOC (audit 2026-04-18)

Category: index-only

Inline tweak at tabNames array / kTab_Wrapper enum (in ImGUI class) (~line 112) — 2-line addition for wrapper tab constant and name. Adds kTab_Wrapper enumerator and "Plugin" entry to tabNames[] to expose the plugin-drawn ImGui tab in the Remix dev menu.

src/dxvk/imgui/imgui_impl_win32.cpp

Pre-refactor fork footprint: +1 / -1 LOC (audit 2026-04-18)

Category: index-only

Inline tweak at ImGui_ImplWin32_UpdateMouseCursor (~line 236) — 1-line modification to GetAsyncKeyState call. Changes key-state sampling to use GetAsyncKeyState (async hardware state) instead of the previous synchronous variant to fix key-state detection on the overlay's WndProc path.

src/dxvk/dxvk_limits.h

Category: inline-tweak

Inline tweak at MaxPushConstantSize constant — increased from 128 to 256. ToneMappingApplyToneMappingArgs grew past the original 128-byte limit once per-operator parameter blocks were added (current size 176 B — see tonemapping.h static_assert). Changing to 256 keeps the constant larger than any current push-constant struct and stays within the 256-byte push-constant size supported by all target GPUs.

src/dxvk/meson.build

Pre-refactor fork footprint: +4 / -0 LOC (audit 2026-04-18)

Category: index-only

Inline tweak at dxvk_src files list (~line 187) — 2-line addition registering atmosphere sources. Registers rtx_render/rtx_atmosphere.cpp and rtx_render/rtx_atmosphere.h in the DXVK build.
Inline tweak at dxvk_src files list (~line 400) — 2-line addition registering ImGui export sources. Registers imgui/imgui_remix_exports.cpp and imgui/imgui_remix_exports.h in the DXVK build.
Inline tweak — register src/dxvk/rtx_render/rtx_fork_tonemap.cpp in the rtx_render source list. The fork-owned tonemap operator headers (aces.slangh, adaptation_v1.slangh, agx.slangh, fork_tonemap_operators.slangh, gt7.slangh, hable.slangh, lottes.slangh, neutwo.slangh, psycho17.slangh) live under src/dxvk/shaders/rtx/pass/tonemap/ and are picked up via the shader-include glob; no explicit meson.build entry is required for those. Fork-owned tonemap module.
Inline tweak at dxvk_src files list (rtx_render block) — 2-line addition registering weather sources. Registers 'rtx_render/rtx_fork_weather.cpp' and 'rtx_render/rtx_fork_weather.h' in the DXVK build source list.

src/dxvk/rtx_render/graph/rtx_component_list.h

Category: index-only

Inline tweak at components/ include list (~line 56) — 2-line addition. Not migrated: the include manifest is the intended extension point for new components, and adding two alphabetically-placed #include lines is the canonical way to register fork-owned graph components. Registers components/game_value_read_bool.h and components/game_value_read_number.h in the component manifest. Both are fork-owned Sense components introduced in workstream 10 (plugin-injected game-state readers); their backing store is the fork-owned rtx_fork_game_state.h.

src/dxvk/rtx_render/rtx_camera_manager.cpp

Pre-refactor fork footprint: +10 / -10 LOC (audit 2026-04-18)

Category: index-only

Inline tweak at CameraManager::processExternalCamera (~line 179) — 6-line addition for [RTX-Diag] entry log. Logs the old and new frameLastTouched values when an external camera is processed, to diagnose camera-validity skip conditions.

src/dxvk/rtx_render/rtx_context.cpp

Pre-refactor fork footprint: +211 / -26 LOC (audit 2026-04-18) Post-refactor footprint: 5 hook call sites + #include "rtx_fork_hooks.h" (migrated 2026-04-18)

Note on Block 4: The audit listed this as "physical atmosphere sky skip + diag logs" in injectRTX. The [RTX-Diag] portion was reverted before migration (commits ff61f77d and b0d2da33); what remains is only the 5-line early-return guard in rasterizeSky. Migrated as fork_hooks::injectRtxAtmosphereSkySkip. The audit's method name (injectRTX) was incorrect — the actual call site is rasterizeSky.

Note on Block 6: The endFrame [RTX-Diag] log block was reverted (commit b0d2da33) before this migration ran. endFrameDiag was not implemented; no hook call site exists in endFrame.

Hook at RtxContext::RtxContext constructor (atmosphere init) → fork_hooks::initAtmosphere in rtx_fork_atmosphere.cpp Constructs the RtxAtmosphere instance during RtxContext initialization.
Hook at RtxContext::updateRaytraceArgsConstantBuffer (sky mode + atmosphere section) → fork_hooks::updateAtmosphereConstants in rtx_fork_atmosphere.cpp Sets constants.skyMode, detects sky mode transitions (clearing skybox buffers on switch to Numos), and calls m_atmosphere->initialize / computeLuts / getAtmosphereArgs to populate the atmosphere constant block.
Hook at RtxContext::bindCommonRayTracingResources (atmosphere LUT bindings) → fork_hooks::bindAtmosphereLuts in rtx_fork_atmosphere.cpp Ensures the atmosphere object is initialized and binds the three atmosphere LUT textures (BINDING_ATMOSPHERE_TRANSMITTANCE_LUT, BINDING_ATMOSPHERE_MULTISCATTERING_LUT, BINDING_ATMOSPHERE_SKY_VIEW_LUT) for all passes that declare them in common_bindings.
Hook at RtxContext::rasterizeSky (physical atmosphere sky skip) → fork_hooks::injectRtxAtmosphereSkySkip in rtx_fork_atmosphere.cpp Returns early from rasterized sky rendering when Numos mode is active. No private-member access; no friend declaration required.
Hook at RtxContext::dispatchScreenOverlay (method body + ScreenOverlayShader class) → fork_hooks::dispatchScreenOverlay in rtx_fork_overlay.cpp ScreenOverlayShader lifted to rtx_fork_overlay.cpp; dispatchScreenOverlay is now a one-line delegate. The hook alpha-composites a plugin-uploaded RGBA buffer over the final tone-mapped image using the compute shader.
Inline tweak at RtxContext::dispatchTonemapping — removed the TonemappingMode::Global || TonemappingMode::Direct dispatch gate; the tonemapper now always runs (always operator-only). The DxvkLocalToneMapping dispatch block was removed entirely (2026-05-15). 2026-05-13 tonemap refactor: global tone curve removed. 2026-05-15: local tonemapper removed.
Inline tweak at (file scope) (weather header include) — 1-line addition near the existing rtx_fork_*.h includes. Adds #include "rtx_fork_weather.h" so WeatherBlender and the fork_weather namespace are available in this translation unit.
Inline tweak at RtxContext::RtxContext constructor (weather blender init) — ~1 LOC. Adds m_weatherBlender = std::make_unique<fork_weather::WeatherBlender>(); so the blender is constructed alongside the atmosphere object.
Hook at RtxContext per-frame entry (weather blender update) — fork_hooks::updateWeatherBlender in rtx_fork_weather.cpp. Calls fork_hooks::updateWeatherBlender(*this, GlobalTime::get().deltaTime()) once per frame so the blender can read trigger keys, advance the lerp timeline, and write blended values to the Derived RTX_OPTION layer.

src/dxvk/rtx_render/rtx_context.h

Pre-refactor fork footprint: +26 / -0 LOC (audit 2026-04-18) Post-refactor fork footprint: +26 / -0 LOC inline tweaks + 4 friend declarations added (migrated 2026-04-18)

Inline tweak at RtxContext class (member declarations — atmosphere) — ~4 LOC. Adds m_lastSkyMode and m_atmosphere member fields to RtxContext. These remain as class members; the fork_hooks functions access them via friend declarations.
Inline tweak at RtxContext::setScreenOverlayData and dispatchScreenOverlay declarations — ~5 LOC. Declares the two overlay-path methods. setScreenOverlayData remains a standalone public method; dispatchScreenOverlay is now a one-line delegate to fork_hooks::dispatchScreenOverlay.
Inline tweak at RtxContext class (member declarations — screen overlay state) — ~11 LOC. Adds ScreenOverlayFrame struct, m_pendingScreenOverlay, m_screenOverlayImage, m_screenOverlayView, m_screenOverlayWidth, m_screenOverlayHeight, and m_screenOverlayFormat. These remain as class members accessed via friend declarations.
Inline tweak at RtxContext class body (just before closing };) — 4-line block of friend declarations plus a forward-declaration block above the class. Grants fork_hooks::initAtmosphere, fork_hooks::updateAtmosphereConstants, fork_hooks::bindAtmosphereLuts, and fork_hooks::dispatchScreenOverlay access to private members.
Inline tweak at (file scope) (weather forward declarations) — ~2 LOC above the class definition. Adds namespace fork_weather { class WeatherBlender; } forward declaration and a void updateWeatherBlender(class RtxContext& ctx, float deltaTimeSeconds) forward declaration inside the fork_hooks namespace block, so the private member and friend declaration below can reference the type.
Inline tweak at RtxContext class body (private member declarations — weather) — ~1 LOC. Adds std::unique_ptr<fork_weather::WeatherBlender> m_weatherBlender; as a private member of RtxContext.
Inline tweak at RtxContext class body (friend declarations block) — ~1 LOC addition to the existing friend block. Adds friend void fork_hooks::updateWeatherBlender(RtxContext&, float); so the hook can access the private m_weatherBlender member.

src/dxvk/rtx_render/rtx_game_capturer.cpp

Pre-refactor fork footprint: +94 / -28 LOC (audit 2026-04-18) Post-refactor footprint: 2 hook call sites + inline tweaks + #include "rtx_fork_hooks.h" (migrated 2026-04-18)

Note on Block 1 (materialLookupHash selection): The material-lookup-hash block is pervasive inline tweaks throughout GameCapturer::newInstance (computing materialLookupHash, keying bIsNewMat, the captureMaterial call, meshes[meshHash]->matHash, and instance.matHash). Lifting this into a hook would require threading too many in/out parameters. Kept as inline tweaks; tracked below.

Inline tweak at GameCapturer::newInstance (materialLookupHash computation and usage) — ~7 LOC distributed through the function. Computes materialLookupHash = material.getHash() and substitutes it for the raw BLAS matHash in the bIsNewMat guard, the captureMaterial call, meshes[meshHash]->matHash, and instance.matHash, so USD capture references align with runtime replacement lookup for API-submitted materials.
Hook at GameCapturer::captureMaterial (method body) → fork_hooks::captureMaterialApiPath in rtx_fork_capture.cpp GameCapturer::captureMaterial is now a one-line delegate; all logic lives in the hook. Exports the albedo texture for both D3D9 materials (color texture valid — direct export) and API-submitted materials (fallback: resolves texture hash via the texture-manager table and exports by hash). Access to private m_exporter and m_pCap is granted via a friend declaration — see the rtx_game_capturer.h entry below.
Hook at GameCapturer::prepExport (coord-system transform block) → fork_hooks::captureCoordSystemSkip in rtx_fork_capture.cpp Skips the view/proj handedness inversion for the global USD export transform when the game is configured as a left-handed coordinate system, since API-submitted geometry is already in consistent Y-up space.

src/dxvk/rtx_render/rtx_game_capturer.h

Post-refactor fork footprint: forward decl + friend declaration (added 2026-04-18)

Category: index-only

Inline tweak at file scope (just before class GameCapturer) — 13-line forward declaration of fork_hooks::captureMaterialApiPath so the friend declaration inside GameCapturer can name the fork-owned hook. Companion to the rtx_fork_capture.cpp hook that needs private-member access to m_exporter and m_pCap.
Inline tweak at GameCapturer class body (top of class, before public:) — 4-line friend declaration granting fork_hooks::captureMaterialApiPath access to private members. Canonical pattern for hooks that must read/write private upstream state — one inline tweak per such hook, tracked here.

src/dxvk/rtx_render/rtx_light_manager.cpp

Pre-refactor fork footprint: +126 / -12 LOC (audit 2026-04-18) Post-refactor footprint: 7 hook call sites + #include "rtx_fork_hooks.h" (migrated 2026-04-18)

Category: migrate

Hook at LightManager::prepareSceneData (pending-mutation flush block) → fork_hooks::flushPendingLightMutations in rtx_fork_light.cpp At frame start, applies queued external-light erases (clearing replacement links), applies queued updates (erase-then-emplace to handle union-type changes), registers pending active-light activations, and auto-instances all persistent lights. Access to private members granted via friend declaration — see rtx_light_manager.h.
Hook at LightManager::updateExternallyTrackedLight (indexed static-sleep path) → fork_hooks::updateLightStaticSleep in rtx_fork_light.cpp Shared static-sleep logic: tracks isStaticCount, skips copy when motionless for N frames, always updates dynamic lights. Restores externallyTrackedLightId when externalId is valid.
Hook at LightManager::addExternalLight (hash-map static-sleep path) → fork_hooks::updateLightStaticSleep in rtx_fork_light.cpp Second call site for the same hook (Block 2's two-copies reduction). Passes kInvalidExternallyTrackedLightId so the id-restore branch is skipped.
Hook at LightManager::addExternalLight (new-light emplace branch) → fork_hooks::setExternalLightEmplace in rtx_fork_light.cpp Emplaces the new external light and stamps frameLastTouched. Access to m_externalLights via friend declaration.
Hook at LightManager::removeExternalLight (queue erase) → fork_hooks::disableExternalLightQueue in rtx_fork_light.cpp Queues the handle for deferred erase instead of immediate removal. Access to m_pendingExternalLightErases via friend declaration.
Hook at LightManager::registerPersistentExternalLight → fork_hooks::registerPersistentLight in rtx_fork_light.cpp Inserts the handle into m_persistentExternalLights. Access via friend declaration.
Hook at LightManager::unregisterPersistentExternalLight → fork_hooks::unregisterPersistentLight in rtx_fork_light.cpp Removes the handle from m_persistentExternalLights. Access via friend declaration.
Hook at LightManager::queueAutoInstancePersistent → fork_hooks::queueAutoInstancePersistent in rtx_fork_light.cpp Copies all persistent-light handles into m_pendingExternalActiveLights. Access via friend declaration.

src/dxvk/rtx_render/rtx_light_manager.h

Pre-refactor fork footprint: +10 / -0 LOC (audit 2026-04-18) Post-refactor footprint: forward decls + friend declarations (updated 2026-04-18)

Category: index-only

Inline tweak at LightManager class (public method declarations) (~line 101) — 4-line addition. Declares registerPersistentExternalLight, unregisterPersistentExternalLight, and queueAutoInstancePersistent in the public API of LightManager.
Inline tweak at LightManager class (private member declarations) (~line 129) — 6-line addition. Adds four deferred-mutation member containers: m_pendingExternalLightErases, m_pendingExternalLightUpdates, m_pendingExternalActiveLights, and m_persistentExternalLights.
Inline tweak at file scope (just before struct LightManager) — forward declarations of all seven fork_hooks:: functions that need LightManager or RtLight access, plus struct RtLight forward decl. Required so the friend declarations inside the class can name the fork-owned hooks.
Inline tweak at LightManager class body (top of class, before public:) — 7-line block of friend declarations granting the light hooks access to private members. Canonical pattern for hooks that must read/write private upstream state — one friend line per hook, tracked here.

src/dxvk/rtx_render/rtx_lights.cpp

Pre-refactor fork footprint: +25 / -17 LOC (audit 2026-04-18) Post-refactor fork footprint: +25 / -17 LOC inline tweaks (reclassified 2026-04-18)

Category: index-only

Rationale: All fork changes are signature modifications and single-line flag-packing additions woven into the middle of existing writeGPUData function bodies, plus a one-line copy in the copy constructor. There is no standalone block to lift; the ignoreViewModel parameter is intrinsic to each function's signature and the flag-set line (if (ignoreViewModel) flags |= 1 << 1;) is inseparably interleaved with the upstream flags-assembly code. A hook would require passing the entire flags word in and out, making it structurally equivalent to rewriting each function — not a meaningful extraction.

Inline tweak at RtSphereLight::writeGPUData (ignoreViewModel parameter + bit 1 flag) — ~3 LOC. Adds bool ignoreViewModel = false parameter; sets bit 1 of the flags word when set. Companion signature change tracked in rtx_lights.h.
Inline tweak at RtRectLight::writeGPUData (ignoreViewModel parameter + bit 1 flag) — ~3 LOC. Same pattern for rect lights.
Inline tweak at RtDiskLight::writeGPUData (ignoreViewModel parameter + bit 1 flag) — ~3 LOC. Same pattern for disk lights.
Inline tweak at RtCylinderLight::writeGPUData (ignoreViewModel parameter + bit 1 flag) — ~3 LOC. Same pattern for cylinder lights; refactors the previously direct writeGPUHelper call to use a local flags variable so the bit can be conditionally set.
Inline tweak at RtDistantLight::writeGPUData (ignoreViewModel parameter + bit 1 flag) — ~4 LOC. Same pattern for distant lights; same refactor of the direct helper call to a local flags variable.
Inline tweak at RtLight::writeGPUData (dispatch passes this->ignoreViewModel) — ~5 LOC (5 call-site updates). Each per-type writeGPUData call now forwards this->ignoreViewModel as the third argument.
Inline tweak at RtLight::copyFrom (ignoreViewModel copy) — ~1 LOC. Copies ignoreViewModel in copyFrom, called by the copy constructor.

src/dxvk/rtx_render/rtx_lights.h

Pre-refactor fork footprint: +6 / -5 LOC (audit 2026-04-18)

Category: index-only

Inline tweak at RtSphereLight::writeGPUData declaration (~line 133) — 1-line modification to add ignoreViewModel default parameter. Adds bool ignoreViewModel = false parameter to the writeGPUData declaration.
Inline tweak at RtRectLight::writeGPUData declaration (~line 197) — 1-line modification (same pattern). Same default parameter addition for RtRectLight.
Inline tweak at RtDiskLight::writeGPUData declaration (~line 275) — 1-line modification (same pattern). Same default parameter addition for RtDiskLight.
Inline tweak at RtCylinderLight::writeGPUData declaration (~line 350) — 1-line modification (same pattern). Same default parameter addition for RtCylinderLight.
Inline tweak at RtDistantLight::writeGPUData declaration (~line 412) — 1-line modification (same pattern). Same default parameter addition for RtDistantLight.
Inline tweak at RtLight struct (~line 645) — 1-line addition of ignoreViewModel field. Adds bool ignoreViewModel = false member to RtLight to carry the per-light view-model exclusion flag across the GPU data write.

src/dxvk/rtx_render/rtx_instance_manager.cpp

Category: index-only

Inline tweak at InstanceManager::onInstanceUpdated surface-meta block (~line 988, beside isAnimatedWater) — REVERTED 2026-06-19, comment-only. Previously set currentInstance.surface.isDecalCategory for the cloud-shadow zenith gate. The gate and flag were deleted when the cloud shadow moved onto the sun term in the NEE (no geometry test needed); only a "removed" comment remains.

src/dxvk/rtx_render/rtx_materials.h

Category: index-only

Inline tweak at RtSurface flags block (~line 336) + RtSurface::writeGPUData flags0 packing (~line 121) — REVERTED 2026-06-19, comment-only. Previously added bool isDecalCategory packed into flags0 bit 2 for the cloud-shadow zenith gate. Removed with the gate; flags0 bit 2 is free again. Only "removed" comments remain.

src/dxvk/rtx_render/rtx_options.h

Pre-refactor fork footprint: +32 / -0 LOC (audit 2026-04-18) Post-refactor fork footprint: +32 / -0 LOC inline tweaks (reclassified 2026-04-18)

Category: index-only

Rationale: All fork additions are an enum definition and RTX_OPTION(...) macro declarations inside the RtxOptions class body. RTX_OPTION expands to an inline static member declaration — it is structurally part of the class definition and cannot be lifted into a separate TU or wrapped in a hook. There is no function body to extract.

Inline tweak at (file scope namespace dxvk) (SkyMode enum) — ~5 LOC. Declares the SkyMode enum class (SkyboxRasterization = 0, Numos = 1). Required by RtxOptions::skyMode below and by atmosphere hook code in rtx_fork_atmosphere.cpp (via the rtx_options.h include chain).
Inline tweak at RtxOptions class body (skyMode RTX_OPTION) — ~2 LOC. Declares RTX_OPTION("rtx", SkyMode, skyMode, SkyMode::SkyboxRasterization, ...) immediately after the existing sky-related options block. Consumed by fork_hooks::updateAtmosphereConstants in rtx_fork_atmosphere.cpp.
Inline tweak at RtxOptions class body (atmosphere RTX_OPTIONs block) — ~25 LOC for the original 17 options + ~5 LOC for night-sky + ~52 LOC for the DECLARE_MOON_OPTIONS(N) macro and 4 invocations + ~11 LOC for the cloud block + ~13 LOC for the cloud-enhancement block (including cloudVerticalProfile, cloudCurvature). Sun/star position fields (sunElevation, sunRotation, starBrightness, starRotation) and per-moon pose fields (elevationN, rotationN, phaseN) are game-drivable per-frame but persist when saved (runtime push is the last writer; cold start uses the saved value until any push lands). Cloud defaults tuned from artist iteration. Declares the original 17 atmosphere tuning options under the rtx.atmosphere prefix (sunDisc, sunSize, sunIntensity, sunElevation, sunRotation, altitude, airDensity, aerosolDensity, ozoneDensity, planetRadius, atmosphereThickness, mieAnisotropy, rayleighScattering, mieScattering, ozoneAbsorption, ozoneLayerAltitude, ozoneLayerWidth, sunIlluminance), plus the night-sky block (starBrightness, starDensity, starTwinkleSpeed, nightSkyBrightness, nightSkyColor), plus a per-moon block declared via the DECLARE_MOON_OPTIONS(N) macro for N in 0..MAX_MOONS-1 (each block: enabledN, angularRadiusN, brightnessN, colorN, surfaceStyleN, craterDensityN, surfaceContrastN, surfaceNoiseScaleN, darkSideBrightnessN, roughnessAmountN, plus pose fields elevationN/rotationN/phaseN), plus a cloud block (cloudEnabled, cloudDensity, cloudAltitude, cloudScale, cloudColor, cloudWindSpeed, cloudWindDirection, cloudShadowStrength, cloudAnisotropy, plus NoSave-flagged cloudCoverage for game-driven weather), plus a cloud-enhancement block (cloudViewSamples, cloudThickness, cloudDetailWeight, cloudShadowTint, cloudShadowTintStrength, cloudSunsetWarmth, cloudVariance, cloudVarianceScale, cloudVerticalProfile, cloudCurvature) for volumetric ray-march tuning, color polish, vertical-shape character, and sky-dome curvature. All consumed by RtxAtmosphere::getAtmosphereArgs() and the atmosphere UI hook in rtx_fork_atmosphere.cpp.
Inline tweak at RtxOptions class body (cloud spatial-variation block) — +21 / -11 net LOC. Adds cloudTypeMean, cloudTypeSpread, cloudTypeNoiseScale, cloudCoverageMean, cloudCoverageSpread, cloudCoverageNoiseScale, cloudAnvilBias, cloudWindShearStrength RTX_OPTIONs (Nubis-style spatial variation, spec 2026-05-06; cloudWindShearStrength added as a tunable knob on the existing wind-shear UV perturbation in sampleCloudDensity). Replaces retired cloudCoverage, cloudVariance, cloudVarianceScale, cloudVerticalProfile.
Inline tweak at RtxOptions class body (cloud aerial-perspective extinction) — +13 LOC. Adds cloudAerialExtinctionPerKm RTX_OPTION (default 0.2) under the rtx.atmosphere prefix (2026-05-16). Distance-based atmospheric extinction applied to cloud samples: aerialT = exp(-cloudAerialExtinctionPerKm * t) multiplies both per-step radiance contribution AND per-step extinction so horizon-grazing rays don't accumulate through ~100 km of cloud volume into a solid white wall. Applied identically in marchCloudSlab (cloud_render.comp.slang) and the legacy analytical evalClouds path (atmosphere_sky.slangh) so the two render modes stay visually consistent. Setting the option to 0 reverts to legacy (no aerial perspective) behavior. Consumed by RtxAtmosphere::getAtmosphereArgs() — the new field repurposes the existing pad_cloudWorley_0 slot in AtmosphereArgs, preserving 16-byte struct alignment — and the atmosphere UI hook in rtx_fork_atmosphere.cpp.
Inline tweak at RtxOptions class body (moon-lighting strength sliders + cloud-look shape) — +10 LOC (Phase 1) + +18 LOC (Phase 3 Task 1) + +28 LOC (Phase 3 Task 2). Phase 1 (2026-05-07) added moonNeeStrength (world-side master, default 1.0) and moonAtmosphericCouplingStrength (sky-side, default 1.0) RTX_OPTIONs. Phase 3 Task 1 (2026-05-08) added per-path stylistic multipliers: surfaceMoonBrightness (default 50.0), cloudMoonBrightness (default 2.0), haloMoonBrightness (default 15.0) — empirically tuned by in-game testing on 2026-05-08 against the Fallout: New Vegas test scene at m.brightness=1.0. Setting all three to 1.0 reverts to architecturally-pure physical-baseline output. Phase 3 Task 2 (2026-05-08) exposed five cloud-look + halo shape constants previously hardcoded in atmosphere_sky.slangh: moonCloudDiffuseGain (0.10), moonCloudPhaseGain (0.30), moonCloudAnisotropy (0.85), moonHaloMagnitude (0.0015), moonAmbientAirglow (0.0015). Defaults preserved; exposure is for in-game tuning without shader rebuilds. All consumed across evalAtmosphereRadiance, evalClouds, evalMoonDisk's halo, and sampleAtmosphereMoonLight.
Inline tweak at RtxOptions class body (Phase 2 default migration) — net 0 LOC, value/text changes only. Phase 2 (2026-05-08) shifts the per-moon brightness##N default from 4.0 → 1.0 (physical neutral; was magic-number magnitude-cheat) and the per-moon color##N default from (0.85, 0.87, 0.92) → (0.12, 0.12, 0.12) (neutral lunar Bond albedo; the prior cool-blue tint was magnitude-cheating). Retires the cloudMoonBrightness RTX_OPTION (its job -- scaling the cloud path's magic-number magnitude -- was eliminated by the Phase 2 unified physical irradiance scaffold). See 2026-05-08-moon-physical-irradiance-design.md.
Inline tweak at RtxOptions class body (cloud-look master multipliers) — +8 LOC. Sky/moon ImGui simplification pass (2026-05-21) added moonSilverLiningIntensity and moonHaloGlowStrength RTX_OPTIONs (both default 1.0). Applied C++-side in RtxAtmosphere::getAtmosphereArgs() as multipliers on the existing cloud-look + halo + airglow knob values, so the five Phase-3-Task-2 fine knobs collapse to three ImGui sliders (Silver Lining Intensity / Sharpness / Halo Glow) while the underlying ratio constants remain .conf-tunable. Shaders unchanged — masters apply at args population. Default 1.0 = byte-identical to pre-2026-05-21 behavior. See 2026-05-20-sky-moon-imgui-persistence-and-reorg-design.md.
Inline tweak at (file scope) (weather header include) -- 1-line addition near the existing rtx_fork_*.h includes. Adds #include "rtx_fork_weather.h" so the DECLARE_ALL_WEATHER_PRESETS() macro is in scope before it is used inside the RtxOptions class body.
Inline tweak at RtxOptions class body (weather preset RTX_OPTION block) -- 1-line macro invocation + 14-line undef block. Invokes DECLARE_ALL_WEATHER_PRESETS() inside the RtxOptions struct body to expand all 504 RTX_OPTION declarations (12 presets x 42 fields). The 14 #undef lines immediately following clean up the binder macros so they do not leak into downstream includes.
Inline tweak -- cloud RTX_OPTION audit cleanup (2026-05-19). Removes three dead-knob RTX_OPTIONs whose values were written into AtmosphereArgs but never read by any shader: cloudScale (pre-3D-texture era; replaced by cloudNoiseTileKm), cloudDetailWeight (legacy FBM detail-fade; Nubis Cubed sampler has no detail-vs-base split), and cloudWindShearStrength (legacy analytical-only wind shear; textured sampler intentionally drops it). Also flips cloudShadowStrength default 0.0 → 1.0 so the voxel-grid cloud-on-terrain shadow system (cloudVoxelShadowsEnable, cloudShadowMarchStrength, cloudShadowFactorStrength) is not silently muted at boot. Args struct slots preserved as padDead* placeholders to maintain 16-byte alignment until a repack pass; the 24 weather-preset entries (12 presets x 2 fields) and the matching WeatherSnapshot plumbing in rtx_fork_weather.{h,cpp} are dropped along with the RTX_OPTION declarations and the ImGui sliders.
Inline tweak -- cloud system iteration #2 (2026-05-19, second pass). Coverage gate floor lowered from 0.15 to 0.0 in sampleCloudDensityTextured so max-Coverage truly fills the sky (was clipping the bottom 15% of the noise range). Aerial perspective split into two independent RTX_OPTIONs: cloudAerialHazePerKm (renamed from cloudAerialExtinctionPerKm; dims cloud radiance with distance — visual softness) and new cloudAerialFadePerKm (softens cloud extinction with distance — prevents horizon white wall). Two new layer-2 spread RTX_OPTIONs (cloudLayer2CoverageSpread, cloudLayer2TypeSpread) plus a shader-side seed-offset (cloudLayer2NoiseSeed, default 1000) added to smoothNoise2D calls for layer 2 so the two layers generate fully decorrelated horizontal weather patterns instead of stacking. Position-shift approach considered first but reverted because computeCloudHeightFraction uses length() against planet center, so any XZ shift > ~10 km pushes samples out of the slab via spherical-distance inflation. Four previously hidden knobs re-added to the menu (cloudCoverageNoiseScale, cloudTypeNoiseScale, cloudCurvature, cloudPhaseG2) since they were silently driving the look at their defaults. 16 RTX_OPTION defaults retuned from in-game FNV tuning. Args slot recycling: pad6/pad7/three padDead* slots are now all live fields (cloudLayer2NoiseSeed, cloudAerialFadePerKm, cloudLayer2CoverageSpread, cloudLayer2TypeSpread, cloudMsScale). Struct size unchanged.
Inline tweak — remove rtx.useLegacyACES + rtx.showLegacyACESOption RtxOptions (superseded by TonemapOperator::ACESNarkowicz enum value). Both options live at the rtx namespace (not rtx.tonemap); removed in the enum refactor.
Inline tweak — remove TonemappingMode enum (Global / Local / Direct) and tonemappingMode RTX_OPTION. The dynamic tone curve (histogram + curve passes) is removed; the apply pass dispatches the operator directly. Local tonemapping (DxvkLocalToneMapping, useLocalToneMapping RTX_OPTION, rtx.localtonemap.*) removed entirely on 2026-05-15. The vestigial directOperatorMode CB field was removed in the 2026-05-XX cleanup along with the dead histogram / tone-curve dispatch passes and the ACES enum rename (ACES/ACESLegacy → ACESHill/ACESNarkowicz). 2026-05-13 tonemap refactor: simplified from three-mode selector to global operator dropdown. 2026-05-15: local tonemap path removed entirely. 2026-05-XX: dead-code cleanup + snake_case shader rename.
Inline tweak — weather-preset cold-default alignment (2026-05-26). Ten cold defaults aligned to WEATHER_PRESET_VALUES_overcast in rtx_fork_weather.h (the macro the codebase comments call "current default look"): cloudShadowStrength 1.0 → 0.10, cloudCoverageMean 0.85 → 0.64, cloudCoverageSpread 1.0 → 0.16, cloudTypeMean 0.75 → 0.5, cloudTypeSpread 0.5 → 0.2, cloudTypeNoiseScale 0.001 → 0.0034, cloudDensity 1.65 → 1.8, cloudThickness 2.75 → 3.05, aerosolDensity 1.0 → 1.1, sunIlluminance (20,20,20) → (15,15,15). Fixes the regression introduced by the 2026-05-19 cloudShadowStrength 0→1 flip: users sitting on the dormant "(none / dormant)" weather preset saw ground geometry crushed dark by full-strength cloud-voxel shadows over 85% default coverage. The dormant blender path leaves cold RTX_OPTIONs untouched, so the cold values themselves had to move.

src/dxvk/rtx_render/rtx_global_volumetrics.cpp

Category: index-only

Inline tweak at RtxGlobalVolumetrics::dispatch (volumeArgs population block) — 2 LOC (2026-05-26; Numos-gated 2026-07-04). Populates volumeArgs.fogSunVisibilityGain and volumeArgs.volumetricConsumerGain from the matching RTX_OPTIONs, forced to 1.0 unless RtxOptions::skyMode() == SkyMode::Numos — both gains are Numos-atmosphere compensations and must not perturb stock volumetrics in rasterized-sky mode. Adjacent to the existing volumetricFogAnisotropy populate. Companion to the fields in volume_args.h and the options in rtx_global_volumetrics.h.
Inline tweak at RtxGlobalVolumetrics::showImguiSettings (advanced material section) — 3 widgets (2026-05-26/2026-06-28/2026-07-05). Fork sliders "Fog Sun Visibility Gain", "Atmosphere Sun Fog Scale", and "Volumetric Consumer Gain" alongside the stock advanced material widgets, exposing the fork gain options for tuning.

src/dxvk/rtx_render/rtx_global_volumetrics.h

Pre-refactor fork footprint: N/A — value-only cold-default tweaks

Category: index-only

Inline tweak — weather-preset cold-default alignment (2026-05-26). Three RTX_OPTION cold defaults aligned to the WEATHER_PRESET_VALUES_overcast block in rtx_fork_weather.h: transmittanceColor (0.999, 0.999, 0.999) → (0.995, 0.995, 0.995), transmittanceMeasurementDistanceMeters 200.0 → 500.0, anisotropy 0.0 → 0.05 (mapped from volumetricAnisotropy in the preset). Companion to the matching rtx_options.h block — same rationale: the dormant "(none / dormant)" weather preset path leaves cold RTX_OPTIONs untouched, so the cold defaults themselves had to move to match overcast.
Inline tweak — new fogSunVisibilityGain RTX_OPTION (2026-05-26; default lowered 5.0→1.0 on 2026-06-15). rtx.volumetrics.fogSunVisibilityGain (default 1.0, range 0.0–50.0) replaces the historical hardcoded artistic gain (x5 with a misleading "10x" comment in the gmod-rtx port) that was previously baked into the per-cache-write expression in fork-owned atmosphere_common.slangh. Default is now physical (1.0 = no boost) rather than the gmod-era ~5x. Read by volume_composite_helpers.slangh::integrateVolumetricNEE (consumer-side fog application only — surface consumers still read the cache straight). Companions: rtx_global_volumetrics.cpp (CB populate), volume_args.h (CB field), submodule fork edit at rtxdi-sdk/include/volumetrics/rtx/algorithm/volume_composite_helpers.slangh.

src/dxvk/rtx_render/rtx_overlay_window.cpp

Pre-refactor fork footprint: +57 / -35 LOC (audit 2026-04-18) Post-refactor footprint: 1 hook call site + 1 #include "rtx_fork_hooks.h" (migrated 2026-04-18)

Note on diag blocks: The fridge list originally listed 3 [RTX-Diag] blocks (blocks 2-4). These were introduced and then immediately reverted (commit 664a9ba4 reverted 0d590fb4) before this migration ran. They are not present in the current file; no action was taken. Only block 1 (keyboard-forward) was active and required migration.

Hook at GameOverlay::gameWndProcHandler (after hwnd guard) → fork_hooks::overlayInputForward in rtx_fork_overlay.cpp Forwards keyboard (WM_KEY*, WM_CHAR, WM_SYSCHAR) AND mouse (WM_MOUSEMOVE, WM_{L,R,M,X}BUTTON*, WM_MOUSE{,H}WHEEL) messages to ImGui_ImplWin32_WndProcHandler on the legacy WndProc path so ImGui keyboard + mouse state stays in sync when a game menu captures raw input or when the plugin HUD pulls focus via the Remix API. Mouse coords in lParam are translated from gameHwnd to overlayHwnd client-space when the two differ; wheel lParam is screen-space and forwards without translation. Access to private m_hwnd is granted via a friend declaration — see the rtx_overlay_window.h entry below. Previously named overlayKeyboardForward (keyboard-only); renamed + expanded 2026-04-19 when the plugin-API mouse-input bug was diagnosed.

src/dxvk/rtx_render/rtx_overlay_window.h

Post-refactor fork footprint: forward decl + friend declaration (added 2026-04-18)

Category: index-only

Inline tweak at file scope (just before class GameOverlay) — 6-line forward declaration of fork_hooks::overlayInputForward so the friend declaration inside GameOverlay can name the fork-owned hook. Companion to the rtx_fork_overlay.cpp hook that needs private-member access to m_hwnd. Renamed from overlayKeyboardForward on 2026-04-19 when the hook's scope expanded to cover mouse messages.
Inline tweak at GameOverlay class body (top of class, before public:) — 3-line friend declaration granting fork_hooks::overlayInputForward access to m_hwnd. Canonical pattern for hooks that must read/write private upstream state — one inline tweak per such hook, tracked here.

src/dxvk/rtx_render/rtx_remix_api.cpp

Pre-refactor fork footprint: +1277 / -118 LOC (audit 2026-04-18) Post-refactor footprint (fully migrated — migrations #7a, #7b, #7c, #7d done): 23 hook call sites + 1 #include "rtx_fork_hooks.h" + inline tweaks listed below. All extractable fork blocks have been migrated to rtx_fork_api_entry.cpp.

Inline tweak at (file scope) (includes block) — ~8 LOC added. Not migrated: include lines don't get hooks — they either stay inline or the fork-owned file pulls them for its own code. Tracked here per the fridge-list invariant. Adds includes for dxvk_objects.h, dxvk_imgui.h, rtx_context.h, rtx_option_layer.h, util_hash_set_layer.h, xxhash.h, algorithm, and d3d9_texture.h to support fork-added API functions, plus rtx_fork_hooks.h added in migration #7a and rtx_fork_game_state.h added in workstream 10 for the remixapi_SetGameValue entry point.
Inline tweak at (file scope) (PendingScreenOverlay struct + s_pendingScreenOverlay) — Removed in migration #7b. Both the struct and the optional are now defined exclusively in rtx_fork_api_entry.cpp (anonymous namespace). A comment marking the removal remains in the upstream file for auditability. The struct held staging buffer, dimensions, format, and opacity; the optional was the hand-off point between the API thread (writer: drawScreenOverlay) and the render thread (reader: presentScreenOverlayFlush). Both now live in the fork-owned TU.
Inline tweak at (anonymous namespace) — s_inFrame, s_beginCallback, s_endCallback, s_presentCallback — Removed in migration #7c. All four vars now live in rtx_fork_api_entry.cpp (anonymous namespace). A comment block marking the removal remains in the upstream file for auditability. Previously used inline at 6 call sites in rtx_remix_api.cpp (DrawInstance, DrawLightInstance, Shutdown, Present×3). All call sites are now one-liner hook delegates.
Hook at convert::toRtMaterialFinalized::preloadTexture lambda (inside the MaterialDataType::Opaque / Translucent / Portal texture preload path) → fork_hooks::textureHashPathLookup in rtx_fork_api_entry.cpp (migrated 2026-04-18, migration #7a). Adds a "0x..." hex-path shortcut inside the upstream texture-path resolver so API-uploaded textures can be referenced by hash string in material JSON without creating a real file path. Hook returns true and writes the resolved TextureRef when the path parses as hex and matches a registered texture; caller returns immediately. Falls through to the normal AssetDataManager lookup otherwise.
Hook at (anonymous namespace) remixapi_AddTextureHash / remixapi_RemoveTextureHash → fork_hooks::mutateTextureHashOption in rtx_fork_api_entry.cpp (migrated 2026-04-18, migration #7a). Looks up an RtxOption<fast_unordered_set> by full option name and adds or removes a hash via the user config layer. Call sites acquire s_mutex then delegate to the hook (which internally takes the RtxOption update mutex — lock order documented alongside s_mutex). The call-site signature replaced the local TextureHashMutation enum with a plain bool add parameter.
Inline tweak at convert::toRtDrawState (skinning hash computation) — 1 LOC, not worth a hook. Not migrated. Calls skinningData.computeHash() on the prototype after building skinning data so the skinning hash participates in geometry deduplication.
Inline tweak at convert::toRtDrawState (blend-weight/index buffer stride fix) — 2 LOC across two call sites, not worth a hook. Not migrated. Fixes blendWeightBuffer and blendIndicesBuffer strides to use bonesPerVertex-based byte widths rather than fixed-width placeholders.
Inline tweak at remixapi_SetupCamera (devLock RAII guard) — 1 LOC. Not extracted to a hook. The LockDevice() guard is scope-tied to the function body (its destructor must run at end-of-function), so a hook cannot own it without rewriting the entire function. Tracked here per the fridge-list invariant. Adds auto devLock = remixDevice->LockDevice() so the EmitCs call that submits external camera data is race-safe.
RETIRED 2026-07-29 (upstream sync) — the remixapi_DrawInstance devLock RAII guard is gone. Upstream restructured the function to serialize on the api-level s_mutex instead of the device lock, hoisting the toRtDrawState conversion outside the lock and emitting with EmitCs<false> (AllowFlush=false). The fork's LockDevice() guard was dropped in favour of upstream's scheme; remixapi_SetupCamera above still carries its own guard because upstream did not restructure that one. remixapi_DrawInstance is now upstream's function verbatim plus the two fork hook calls below.
Hook at remixapi_DrawInstance (beginScene dispatch) → fork_hooks::notifyBeginScene in rtx_fork_api_entry.cpp (migrated 2026-04-18, migration #7c). One-liner call. Atomically exchanges s_inFrame to true and fires s_beginCallback on the first frame submission.
Hook at remixapi_DrawLightInstance (beginScene dispatch) → fork_hooks::notifyBeginScene in rtx_fork_api_entry.cpp (migrated 2026-04-18, migration #7c). Same hook as DrawInstance; lights-only frames also trigger the beginScene callback.
Hook at remixapi_Shutdown (callback + frame-state clear) → fork_hooks::shutdownCallbacks in rtx_fork_api_entry.cpp (migrated 2026-04-18, migration #7c). One-liner call replacing the 4-line null/false reset. Clears s_beginCallback, s_endCallback, s_presentCallback, and s_inFrame.
Hook at remixapi_Present (screen overlay flush — inner namespace path) → fork_hooks::presentScreenOverlayFlush in rtx_fork_api_entry.cpp (migrated 2026-04-18, migration #7b). One-liner call to the fork-owned flush hook. State (PendingScreenOverlay + s_pendingScreenOverlay) was unified in the same migration.
Hook at remixapi_Present (endScene callback, before native Present) → fork_hooks::presentEndSceneDispatch in rtx_fork_api_entry.cpp (migrated 2026-04-18, migration #7c). Fires s_endCallback if s_inFrame is set, immediately before the native remixDevice->Present() call.
Hook at remixapi_Present (present callback + s_inFrame reset, after native Present) → fork_hooks::presentCallbackDispatch in rtx_fork_api_entry.cpp (migrated 2026-04-18, migration #7c). Fires s_presentCallback and resets s_inFrame to false after a successful native Present.
Hook at extern "C" remixapi_AutoInstancePersistentLights (screen overlay flush path) → fork_hooks::presentScreenOverlayFlush in rtx_fork_api_entry.cpp (migrated 2026-04-18, migration #7b). Same flush hook called on the C-export AutoInstancePersistentLights path, which also drains the pending overlay once per frame.
Hook at remixapi_DrawScreenOverlay (function body) → fork_hooks::drawScreenOverlay in rtx_fork_api_entry.cpp (migrated 2026-04-18, migration #7b). Upstream wrapper acquires device + mutex, then delegates to the hook. The reverted [RTX-Diag] FIRST-CALL log block is absent (never present in current HEAD).
Hook at remixapi_GetUIState (function body) → fork_hooks::getUiState in rtx_fork_api_entry.cpp (migrated 2026-04-18, migration #7b). One-liner delegate passing tryAsDxvk() to the hook.
Hook at remixapi_SetUIState (function body) → fork_hooks::setUiState in rtx_fork_api_entry.cpp (migrated 2026-04-18, migration #7b). One-liner delegate.
Hook at remixapi_dxvk_GetSharedD3D11TextureHandle (function body) → fork_hooks::getSharedD3D11TextureHandle in rtx_fork_api_entry.cpp (migrated 2026-04-18, migration #7b). Stub returning GENERAL_FAILURE. One-liner delegate.
Hook at remixapi_dxvk_GetTextureHash (function body) → fork_hooks::dxvkGetTextureHash in rtx_fork_api_entry.cpp (migrated 2026-04-18, migration #7b). One-liner delegate.
Hook at remixapi_CreateTexture (function body) → fork_hooks::createTexture in rtx_fork_api_entry.cpp (migrated 2026-04-18, migration #7b). Upstream wrapper acquires s_mutex, then delegates. Full GPU resource creation lives in the hook.
Hook at remixapi_DestroyTexture (function body) → fork_hooks::destroyTexture in rtx_fork_api_entry.cpp (migrated 2026-04-18, migration #7b). Upstream wrapper acquires s_mutex, then delegates.
Hook at remixapi_RegisterCallbacks (function body) → fork_hooks::registerCallbacks in rtx_fork_api_entry.cpp (migrated 2026-04-18, migration #7c). One-liner delegate. Body now lives in the fork-owned TU where the callback state vars live.
Hook at remixapi_RequestVramCompaction (function body) → fork_hooks::requestVramCompaction in rtx_fork_api_entry.cpp (migrated 2026-04-20, migration #7d). One-liner delegate passing tryAsDxvk() to the hook. Hook does its own null check and sets SceneManager's atomic VRAM-compaction flag; render thread consumes it in manageTextureVram. Lock-free — s_mutex not taken.
Hook at remixapi_RequestTextureVramFree (function body) → fork_hooks::requestTextureVramFree in rtx_fork_api_entry.cpp (migrated 2026-04-20, migration #7d). One-liner delegate. Hook sets SceneManager's atomic texture-VRAM-free flag; the render-thread tick calls textureManager.clear(), matching the DX9 scene-transition behavior exposed to plugins. Lock-free — s_mutex not taken.
Hook at remixapi_GetVramStats (function body) → fork_hooks::getVramStats in rtx_fork_api_entry.cpp (migrated 2026-04-20, migration #7d). One-liner delegate. Hook fills remixapi_VramStats with per-category DXVK totals plus driver-view heap info (driverAllocatedBytes / driverBudgetBytes) and the fork-side RtxTextureManager::getTextureTable().size() (forkTextureCacheCount). Driver-view numbers match Task Manager / nvidia-smi; the gap vs totalAllocatedBytes exposes non-DXVK allocations (NGX, RT pipeline state, descriptor pools, NRC).
Inline tweak at (anonymous namespace) frame-boundary callback infrastructure — s_pendingLightCreates, s_pendingLightUpdates, s_pendingDomeUpdates, s_pendingLightDestroys, s_pendingMeshCreates, s_handlesDeletedThisFrame. Not migrated. The pending-queue state stays in upstream because it is accessed by too many anonymous-namespace functions (flushPendingMeshes, remixapi_CreateMeshBatched, remixapi_CreateLight, remixapi_DestroyLight, remixapi_Present, remixapi_UpdateLightDefinition) — moving it would require either lifting all those callers or exposing a wide accessor surface. Tracked here per the fridge-list invariant.
Inline tweak at remixapi_AutoInstancePersistentLights / remixapi_UpdateLightDefinition bodies (extern-C fork-owned functions) — not extracted to hooks. These are REMIXAPI-exported entry points; their bodies are the fork's implementation of those API calls. The pending-queue state they access is documented as staying inline above. Tracked here per the fridge-list invariant. 2026-06-27: remixapi_AutoInstancePersistentLights now early-outs (skips the per-frame LockDevice+EmitCs) when no C-API scene work is queued and no external light has ever been registered, gated on the file-scope sticky s_externalLightApiUsed (set in remixapi_CreateLight, remixapi_CreateLightBatched, remixapi_UpdateLightDefinition). Fixes native-only consumers seeing all lights flicker from the empty per-frame dispatch on the native present path. The self-gating overlay flush still runs.
Inline tweak at bit-24 category routing (in toRtCategories) — ~1 LOC. Not migrated. Routes REMIXAPI_INSTANCE_CATEGORY_BIT_SMOOTH_NORMALS (bit 24, upstream name) to InstanceCategories::SmoothNormals. (The misleadingly-named LEGACY_EMISSIVE alias of this bit was removed 2026-06-28 — it routed to SmoothNormals despite its name, a silent footgun.)
Inline tweak at (anonymous namespace) remixapi_SetGameValue — ~15 LOC. Not migrated (fork-owned store + direct inline body fits the surrounding remixapi_SetConfigVariable pattern; no anonymous-namespace state to share with other TUs). Implements the plugin-injected game-state write API introduced in workstream 10. Validates args, constructs std::string copies of the incoming C strings, and forwards to dxvk::fork_game_state::GameStateStore::get().set(key, value). Does not take s_mutex — the store owns its own lock, and funnelling high-frequency plugin writes through the API-wide mutex has no benefit.
Block at extern "C" vtable init block (fork-added anonymous-namespace slots) — ~11 LOC inline assignment block in remixapi_InitializeLibrary. Not fully hookable: the anonymous-namespace function pointers have internal linkage and cannot be named from another TU. Tracked here per the fridge-list invariant. The three extern-C-linked fork slots (RegisterCallbacks, AutoInstancePersistentLights, UpdateLightDefinition) are assigned via fork_hooks::remixApiVtableInit (migrated 2026-04-18, migration #7c). Registers all fork-added API functions into the remixapi_Interface vtable. The inline block assigns the anonymous-namespace slots (including SetGameValue added in workstream 10); the hook fills the three externally-linked ones.
Inline tweak at extern "C" vtable size static_assert — 1 LOC. Not migrated (fridge-listed). The static_assert(sizeof(interf) == 288, ...) sentinel is the final value in the chain (208 → 240 → 272 → 280 → 288 across five workstreams). Retained inline in remixapi_InitializeLibrary as a size sentinel.

src/dxvk/rtx_render/rtx_remix_specialization.inl

Pre-refactor fork footprint: +3 / -1 LOC (audit 2026-04-18)

Category: index-only

Inline tweak at pnext::detail specialization list (~line 95) — 2-line addition. Adds remixapi_CameraInfoParameterizedEXT and remixapi_TextureInfo to the pnext type-list so pnext::chain can traverse these new struct types.
Inline tweak at pnext::detail::ToEnum specialization (~line 123) — 1-line addition. Maps remixapi_TextureInfo to REMIXAPI_STRUCT_TYPE_TEXTURE_INFO in the sType enum specialization table.

src/dxvk/rtx_render/rtx_resources.cpp

Pre-refactor fork footprint: +18 / -0 LOC (audit 2026-04-18)

Category: index-only

Inline tweak at Resources::getAtmosphereTransmittanceLut (new method body) (~line 808) — 5-line addition. Stub accessor returning m_atmosphereTransmittanceLut; LUT is populated by RtxAtmosphere.
Inline tweak at Resources::getAtmosphereMultiscatteringLut (new method body) (~line 813) — 5-line addition. Stub accessor returning m_atmosphereMultiscatteringLut.
Inline tweak at Resources::getAtmosphereSkyViewLut (new method body) (~line 818) — 5-line addition. Stub accessor returning m_atmosphereSkyViewLut.

src/dxvk/rtx_render/rtx_resources.h

Pre-refactor fork footprint: +7 / -0 LOC (audit 2026-04-18)

Category: index-only

Inline tweak at Resources class (public method declarations) (~line 393) — 3-line addition. Declares getAtmosphereTransmittanceLut, getAtmosphereMultiscatteringLut, and getAtmosphereSkyViewLut on the Resources class.
Inline tweak at Resources class (private member fields) (~line 469) — 4-line addition. Adds m_atmosphereTransmittanceLut, m_atmosphereMultiscatteringLut, and m_atmosphereSkyViewLut storage fields to Resources.

src/dxvk/rtx_render/rtx_scene_manager.cpp

Pre-refactor footprint: +73 / -2 LOC (migrated 2026-04-18) Post-refactor footprint: 4 hook call sites + 1 #include "rtx_fork_hooks.h"

Hook at SceneManager::submitExternalDraw (before submesh loop) → fork_hooks::externalDrawMeshReplacement in rtx_fork_submit.cpp Checks for USD mesh/light replacements keyed on the API mesh handle hash; call site handles the early-exit + drawReplacements dispatch since those are private SceneManager methods. 2026-07-29 (upstream sync): the call site's replacement-template setup now goes through DrawCallState::modifyGeometryData() instead of writing replacementDrawCall.geometryData directly — upstream moved geometryData into the private section behind that accessor, and SceneManager is not a friend of DrawCallState (the fork hooks and D3D9Rtx are). Same for modifyMaterialData() at the setHashOverride site below.
Hook at SceneManager::submitExternalDraw (inside if (material != nullptr), before setHashOverride) → fork_hooks::externalDrawMaterialReplacement in rtx_fork_submit.cpp Checks for USD material replacements via getReplacementMaterial() and updates the material pointer in-place if one is found.
Hook at SceneManager::submitExternalDraw (inside if (material != nullptr), after setHashOverride) → fork_hooks::externalDrawTextureCategories in rtx_fork_submit.cpp Resolves albedo texture hash from the API material's opaque data and auto-applies all texture-based instance categories (Sky, Ignore, WorldUI, WorldMatte, Particle, Beam, DecalStatic, Terrain, AnimatedWater, IgnoreLights, IgnoreAntiCulling, IgnoreMotionBlur, Hidden).
Hook at SceneManager::submitExternalDraw (after particle setup, before processDrawCallState) → fork_hooks::externalDrawObjectPicking in rtx_fork_submit.cpp Stores per-draw texture hash metadata in m_drawCallMeta when object picking is active. Access to the private m_drawCallMeta member is granted via a friend declaration — see the rtx_scene_manager.h entry below.

src/dxvk/rtx_render/rtx_scene_manager.h

Post-refactor fork footprint: forward decl + friend declaration (added 2026-04-18)

Category: index-only

Inline tweak at file scope (just before class SceneManager) — 9-line forward declaration of fork_hooks::externalDrawObjectPicking so the friend declaration inside SceneManager can name the fork-owned hook. Companion to the rtx_fork_submit.cpp hook that needs private-member access to m_drawCallMeta.
Inline tweak at SceneManager class body (top of class, before public:) — 5-line friend declaration granting fork_hooks::externalDrawObjectPicking access to m_drawCallMeta. Canonical pattern for hooks that must read/write private upstream state — one inline tweak per such hook, tracked here.

src/dxvk/rtx_render/rtx_sky.h

Pre-refactor fork footprint: +6 / -0 LOC (audit 2026-04-18)

Category: index-only

Inline tweak at tryHandleSky (~line 145) — 6-line addition for physical atmosphere sky skip. Returns TryHandleSkyResult::SkipSubmit early for any draw with cameraType == CameraType::Sky when Numos mode is active, preventing rasterized skybox geometry from being submitted.

src/dxvk/rtx_render/rtx_tone_mapping.cpp

Hook calls at DxvkToneMapping::dispatchApplyToneMapping (args-population) and DxvkToneMapping::showImguiSettings (ImGui panel) → fork_hooks::populateTonemapOperatorArgs + fork_hooks::showTonemapOperatorUI in rtx_fork_tonemap.cpp. Routes global tonemap through the fork operator dispatcher.

src/dxvk/rtx_render/rtx_tone_mapping.h

Inline tweak — remove rtx.tonemap.finalizeWithACES RtxOption (superseded by rtx.tonemap.tonemapOperator in rtx_fork_tonemap.cpp); add #include "rtx_fork_tonemap.h". Adopts the fork operator enum.

src/dxvk/shaders/rtx/algorithm/geometry_resolver.slangh

Pre-refactor fork footprint: +171 / -1 LOC (audit 2026-04-18)

Category: migrate

Block at geometryResolverVertex (miss handler — sky radiance branch) — ~30 LOC (active) + ~60 LOC (commented-out deprecated decals-on-sky block), planned target fork_hooks::geoResolverAtmosphereMiss in rtx_fork_atmosphere.slangh. Adds a conditional atmosphere sky-radiance evaluation (evalSkyRadiance) in the geometry-resolver miss path when cb.skyMode == 1, selecting between dome light, physical atmosphere, and skybox rasterization. The commented-out block documents the deprecated enableDecalsOnSky feature. Cloud temporal smoothing (2026-05-09): the primary view ray's evalSkyRadiance call now passes enableCloudTemporalSmoothing=true plus the motion-vector + screen-extent args needed to reproject and EMA-blend the cloud layer against the previous frame's history at slots 206/207. PSR and indirect callers continue to pass false (their pixelCoord refers to a non-primary direction; reusing primary screen-space cloud history would smear). Sky-matte primary sampling (2026-07-05): in skybox-rasterization mode (skyMode==0, no dome light) camera rays now sample the screen-space SkyMatte at cameraPixelCoordinateToScreenUV instead of the SkyProbe cubemap — since the composite-side SkyMatte add was removed (double-sky fix), the resolver write is the visible sky, and the cube parameterization's gradient discontinuity showed as seams along face edges; the matte is the game's own raster for that exact pixel. Rays with directionAltered set (crossed a ray portal, so their direction no longer matches the rasterized view) keep the direction-based probe, as do PSR and indirect misses.
Block at geometryResolverVertex (hit path — occluder comment block) — ~42 LOC (fully commented out), planned target fork_hooks::geoResolverOccluder in rtx_fork_atmosphere.slangh. Preserves the design for the deprecated isOccluder surface property that would have shown sky behind occluder surfaces; kept commented for future reference.
Block at geometryPSRResolverVertex (PSR hit — atmosphere sky radiance) — ~9 LOC, planned target fork_hooks::geoResolverPsrAtmosphere in rtx_fork_atmosphere.slangh. Adds atmosphere sky-radiance evaluation in the PSR resolver's emissive radiance accumulation path when physical atmosphere mode is active. Volumetric attenuation (2026-07-04): the three sky branches (dome light / Numos atmosphere / SkyProbe) were unified into a common skyRadiance that is multiplied by calcVolumetricAttenuation before accumulation, matching the primary-miss path — fixes reflected sky on water being brighter than the directly-visible sky under fog.
Block at geometryPSRResolverVertex (PSR hit — occluder comment block) — ~45 LOC (fully commented out), planned target fork_hooks::geoResolverPsrOccluder in rtx_fork_atmosphere.slangh. Same occluder design-preservation comment block for the PSR path.
Block at (file scope) (atmosphere include) — ~4 LOC, planned target fork_hooks::atmosphereInclude in rtx_fork_atmosphere.slangh. Adds #include "rtx/pass/atmosphere/atmosphere_common.slangh" at the top of the file, plus #include "rtx/pass/atmosphere/atmosphere_sky.slangh" gated by #ifdef ATMOSPHERE_AVAILABLE for the sky-radiance evaluation paths.

src/dxvk/shaders/rtx/algorithm/integrator_direct.slangh

Pre-refactor fork footprint: +120 / -2 LOC (audit 2026-04-18)

Category: migrate

Block at (file scope) (atmosphere include) — ~1 LOC, planned target fork_hooks::atmosphereInclude in rtx_fork_atmosphere.slangh. Adds #include "rtx/pass/atmosphere/atmosphere_common.slangh". 2026-06-27: moved this include ahead of rtxcr_material.slangh so the SSS NEE code can call sampleCloudGroundShadow_OptionB for the SSS cloud-shadow fold (see the rtxcr_material.slangh touchpoint).
Block at evalAtmosphereSunNEE (full function) — ~40 LOC, planned target fork_hooks::evalAtmosphereSunNEEDirect in rtx_fork_atmosphere.slangh. Implements primary-bounce sun NEE for physical atmosphere: samples sun direction + cone angle, traces multiple jittered shadow rays for soft shadows, averages visibility, evaluates BRDF split-weight, and accumulates diffuse/specular sun radiance. As of the 2026-06-19 sun-only cloud-shadow re-architecture this function no longer touches clouds at all — the cloud-on-terrain shadow folds onto the sun's radiance inside sampleAtmosphereSunLight (atmosphere_common.slangh), so the per-pixel PrimaryCloudShadowFactor write, the sealed-interior zenith up-ray gate, and the viewmodel/decal/normal-flip origin corrections were all deleted here. 2026-06-20: taught the shadow ray about thin-opaque subsurface — it now traces with visibilityModeEnableSubsurfaceMaterials and skips the NdotL <= 0 early-out (both the up-front check and the per-sample jitter skip) for thin-opaque materials, so backlit translucency works under the physical-atmosphere sun (it is the only sun NEE in skyMode==1; the SSS-capable standard path only runs for RTXDI/RIS lights).
Block at evalAtmosphereMoonNEE (full function) — ~100 LOC, planned target fork_hooks::evalAtmosphereMoonNEEDirect in rtx_fork_atmosphere.slangh. Primary-bounce moon NEE -- mirror of evalAtmosphereSunNEE for the moon. Calls sampleAtmosphereMoonLight with a u_pick blue-noise sample so one of the enabled, above-horizon moons is importance-picked per ray (weight = brightness × phaseGlow × elevation). Soft-shadow cone jitter via getJitteredSunDirection (direction-agnostic). Accumulated contribution divided by moonSample.solidAnglePdf (discrete pick PDF) so multi-moon importance sampling stays unbiased over many frames. Added by 2026-05-07 moon sun-parity workstream. 2026-06-20: same thin-opaque subsurface fix as the sun path (visibilityModeEnableSubsurfaceMaterials + thin-opaque-gated NdotL <= 0 early-outs) so moonlit translucency matches.
Block at integrateDirectPath (atmosphere sun NEE call site) — ~14 LOC, planned target fork_hooks::directPathAtmosphereSunCall in rtx_fork_atmosphere.slangh. Calls evalAtmosphereSunNEE in the direct-path integrator when cb.skyMode == 1. 2026-06-21 (experiment branch): the gate also requires debugSkyBisectFlags bit 2 clear — when the sun/moon are injected as real distant lights (rtx.atmosphere.useDirectionalLights), bit 2 is set and this bespoke sun+moon NEE block is skipped to avoid double-counting.
Block at integrateDirectPath (atmosphere moon NEE call site) — ~12 LOC, planned target fork_hooks::directPathAtmosphereMoonCall in rtx_fork_atmosphere.slangh. Calls evalAtmosphereMoonNEE immediately after evalAtmosphereSunNEE in the direct-path integrator when cb.skyMode == 1. Sun and moon NEE are independent samples -- both can be valid at twilight, both invalid during pure daytime / pure-night-with-no-moons; each early-outs cheaply when invalid. Added by 2026-05-07 moon sun-parity workstream.
Block at integrateDirectPath (sky radiance miss branch) — ~8 LOC, planned target fork_hooks::directPathAtmosphereMiss in rtx_fork_atmosphere.slangh. Adds #ifdef ATMOSPHERE_AVAILABLE branch in the miss sky-radiance evaluation to call evalSkyRadiance in physical atmosphere mode.
Block at integrateDirectPath / sampleLightRIS call sites (customIndex for view-model) — ~3 LOC, planned target fork_hooks::directPathViewModelCustomIndex in rtx_fork_light.slangh. Synthesizes a customIndex carrying CUSTOM_INDEX_IS_VIEW_MODEL from geometryFlags.isViewModel and threads it through to evalDirectLighting call sites so view-model geometry skips ignoreViewModel lights.

src/dxvk/shaders/rtx/algorithm/integrator_indirect.slangh

Pre-refactor fork footprint: +138 / -4 LOC (audit 2026-04-18)

Category: migrate

Block at (file scope) (atmosphere include) — ~4 LOC, planned target fork_hooks::atmosphereInclude in rtx_fork_atmosphere.slangh. Adds #include "rtx/pass/atmosphere/atmosphere_common.slangh", plus #include "rtx/pass/atmosphere/atmosphere_sky.slangh" gated by #ifdef ATMOSPHERE_AVAILABLE for the sky-radiance evaluation in the indirect-path miss handler.
Block at evalAtmosphereSunNEESecondary (full function) — ~100 LOC, planned target fork_hooks::evalAtmosphereSunNEEIndirect in rtx_fork_atmosphere.slangh. Secondary-bounce variant of the atmosphere sun NEE function: uses half the sample count for performance, otherwise identical structure to the direct-path version.
Block at evalAtmosphereMoonNEESecondary (full function) — ~100 LOC, planned target fork_hooks::evalAtmosphereMoonNEEIndirect in rtx_fork_atmosphere.slangh. Secondary-bounce variant of the moon NEE function -- mirror of evalAtmosphereSunNEESecondary. Same structure as the direct-path moon NEE but with the indirect shadow-mask flags, half the sample count, and additive diffuseLight / specularLight accumulation (no throughput multiplier; caller folds throughput in). Added by 2026-05-07 moon sun-parity workstream.
Block at integratePathVertex (atmosphere moon NEE call site) — ~16 LOC, planned target fork_hooks::indirectPathAtmosphereMoonCall in rtx_fork_atmosphere.slangh. Calls evalAtmosphereMoonNEESecondary after the existing sun call when cb.skyMode == 1 && isNeeEnabledOnBounce. Accumulates the returned diffuseLight + specularLight via accumulateRadiance. Added by 2026-05-07 moon sun-parity workstream.
Block at integratePathVertex (atmosphere sky radiance in miss) — ~8 LOC, planned target fork_hooks::indirectPathAtmosphereMiss in rtx_fork_atmosphere.slangh. Adds the #ifdef ATMOSPHERE_AVAILABLE sky-radiance branch in the indirect path miss handler.
Block at integratePathVertex (secondary bounce atmosphere sun NEE call) — ~18 LOC, planned target fork_hooks::indirectPathAtmosphereSunCall in rtx_fork_atmosphere.slangh. Calls evalAtmosphereSunNEESecondary for secondary bounces when physical atmosphere mode is active and NEE is enabled on the bounce.
Block at integratePathVertex (customIndex for view-model lights) — ~4 LOC, planned target fork_hooks::indirectPathViewModelCustomIndex in rtx_fork_light.slangh. Synthesizes customIndex from rayInteraction.isViewModel at both RTXDI and advanced-RIS call sites in the indirect path.

src/dxvk/shaders/rtx/algorithm/lighting.slangh

Pre-refactor fork footprint: +27 / -5 LOC (audit 2026-04-18)

Category: migrate

Block at sampleLightRTXDI signature + ignoreViewModel filter — ~12 LOC, planned target fork_hooks::sampleLightViewModelFilter in rtx_fork_light.slangh. Adds a customIndex parameter (default 0) to sampleLightRTXDI and inserts a guard that returns false when the sampled reservoir light has ignoreViewModel set and the caller's customIndex has CUSTOM_INDEX_IS_VIEW_MODEL.
Block at sampleLightAdvancedRIS signature + ignoreViewModel filter — ~10 LOC, planned target fork_hooks::sampleLightViewModelFilter in rtx_fork_light.slangh. Same customIndex parameter and ignoreViewModel skip guard in the advanced-RIS sampling loop.
Block at sampleLightRIS dispatch (propagate customIndex) — ~3 LOC, planned target fork_hooks::sampleLightViewModelFilter in rtx_fork_light.slangh. Updates the sampleLightRIS dispatcher to pass customIndex through to sampleLightAdvancedRIS.

src/dxvk/shaders/rtx/algorithm/rtxcr/rtxcr_material.slangh

Pre-refactor fork footprint: +11 / -4 LOC (audit 2026-04-18)

Category: index-only

Inline tweak at evalSssDiffusionProfile (~line 167) — 3-line addition for view-model customIndex. Synthesizes customIndex from geometryFlags.isViewModel and threads it to the SSS diffusion-profile light-sampling calls.
Inline tweak at evalSingleScatteringTransmission (first call site, ~line 323) — 3-line addition for view-model customIndex. Same customIndex pattern for the first single-scattering transmission light sample.
Inline tweak at evalSingleScatteringTransmission (second call site, ~line 423) — 3-line addition for view-model customIndex. Same customIndex pattern for the second single-scattering transmission light sample.
Block at file scope (sssApplyAtmosphereCloudShadow helper) + 3 call sites — fork — 2026-06-27. Folds the cloud-on-terrain shadow onto SSS NEE. The opaque direct/indirect integrators fold the per-pixel cloud transmittance onto the flagged atmosphere sun's lightSample.radiance (the atmosphereCloudShadowed real-light fold), but evalSssDiffusionProfile / evalSingleScatteringTransmission sample their own lights internally, so the diffusion-profile, transmission, and single-scattering terms stayed fully sunlit under heavy cloud while surrounding opaque diffuse darkened. Adds a file-local sssApplyAtmosphereCloudShadow(inout LightSample, lightIdx, surfacePos) mirroring the integrator fold (same skyMode==1 / cloudVoxelShadowsEnable / atmosphereCloudShadowed / cloudShadowFactorStrength gating) and calls it at all three SSS light-sample sites. Depends on sampleCloudGroundShadow_OptionB from atmosphere_common.slangh, which is why the integrator_direct.slangh include of atmosphere_common was reordered ahead of this header.

src/dxvk/shaders/rtx/concept/light/light.h

Pre-refactor fork footprint: +1 / -0 LOC (audit 2026-04-18)

Category: index-only

Inline tweak at DecodedPolymorphicLight struct (~line 56) — 1-line addition of ignoreViewModel field. Adds bool ignoreViewModel to the decoded-light struct so the GPU-side light filter can read the flag after decode.

src/dxvk/shaders/rtx/concept/light/polymorphic_light.slangh

Pre-refactor fork footprint: +2 / -0 LOC (audit 2026-04-18)

Category: index-only

Inline tweak at decodePolymorphicLight (~line 60) — 1-line addition for ignoreViewModel decode. Extracts bit 1 of the flags word into decodedPolymorphicLight.ignoreViewModel during polymorphic-light decode.

src/dxvk/shaders/rtx/concept/surface/surface.h

Category: index-only

Inline tweak at Surface flags properties (~line 306, after isVertexColorBakedLighting) — REVERTED 2026-06-19, comment-only. Previously added the Surface::isDecalCategory property (data0b.z bit 2) read by the cloud-shadow zenith gate. Removed with the gate; bit 2 is free again. Only a "freed" comment remains.

src/dxvk/shaders/rtx/pass/common_binding_indices.h

Pre-refactor fork footprint: +9 / -1 LOC (audit 2026-04-18)

Category: index-only

Inline tweak at (file scope) (atmosphere binding index defines) (~line 49) — 3-line addition. Defines BINDING_ATMOSPHERE_TRANSMITTANCE_LUT (200), BINDING_ATMOSPHERE_MULTISCATTERING_LUT (201), and BINDING_ATMOSPHERE_SKY_VIEW_LUT (202) at high slot numbers to avoid conflicts with pass-specific bindings.
Inline tweak at COMMON_BINDING_DEFINITION_LIST macro (~line 96) — 3-line addition for common-binding list. Adds TEXTURE2D entries for the three atmosphere LUT bindings to the common-binding definition macro so they appear in all passes that include common_bindings.
Inline tweak at (file scope) (atmosphere binding index defines) (~line 56) and COMMON_RAYTRACING_BINDINGS macro (~line 103) — Stage C addition. Adds BINDING_ATMOSPHERE_CLOUD_NOISE_3D = 203 and a corresponding TEXTURE3D entry in the macro list for the prebaked 3D cloud noise volume (256³ R8 Perlin). No consumer yet; resource and bake pass land in subsequent Stage C tasks.
Inline tweak at (file scope) (atmosphere binding index defines, ~line 57) and COMMON_RAYTRACING_BINDINGS macro (~line 104) — Stage C Task 8a addition. Adds BINDING_ATMOSPHERE_CLOUD_NOISE_SAMPLER = 204 and a corresponding SAMPLER entry in the macro list. The linear/REPEAT sampler is bound alongside the cloud noise SRV in bindAtmosphereLuts and consumed by sampleCloudDensityTextured at call sites (Task 8b).
Inline tweak at (file scope) (atmosphere binding index defines, ~line 58) and COMMON_RAYTRACING_BINDINGS macro (~line 108) — FAST-noise jitter (2026-05-09). Adds BINDING_ATMOSPHERE_FAST_NOISE = 205 and a corresponding TEXTURE2DARRAY entry in the macro list. Resource is the EA Importance-Sampled FAST noise (128×128×32 RG8) uploaded once by RtxFastNoise and bound in bindAtmosphereLuts. Consumed by the fastJitter() helper in atmosphere_common.slangh for cloud view-march jitter (channel x) and sun-shadow tap jitter (channel y).
Inline tweak at (file scope) (atmosphere binding index defines) and COMMON_RAYTRACING_BINDINGS macro — cloud history temporal smoothing (2026-05-09, age channel added 2026-05-13). Adds BINDING_ATMOSPHERE_CLOUD_HISTORY_PREV = 206 (TEXTURE2D) and BINDING_ATMOSPHERE_CLOUD_HISTORY_CURR = 207 (RW_TEXTURE2D). RGBA16F screen-space ping-pong owned by RtxAtmosphere and bound in bindAtmosphereLuts. Consumed by evalSkyRadiance in atmosphere_sky.slangh when called with enableCloudTemporalSmoothing=true (currently only the primary view ray in geometry_resolver.slangh miss path). Smooths per-frame FAST-noise jitter variance to give DLSS a stable signal. The age-channel companion BINDING_ATMOSPHERE_CLOUD_HISTORY_FRAME_ID_PREV = 212 (TEXTURE2D) and BINDING_ATMOSPHERE_CLOUD_HISTORY_FRAME_ID_CURR = 213 (RW_TEXTURE2D) carries the frame index at which each pixel was last refreshed by the sky-miss path; R16_UINT, cleared to 0xFFFF "never written" sentinel at allocation. The shader rejects history whose stored frame-id != (frameIdx - 1) & 0xFFFFu, which fixes the multi-frame bright-trail ghosting that the alpha-only disocclusion guard previously left exposed once the 2026-05-13 Nubis Cubed rewrite drove cloud radiance higher.
Inline tweak at (file scope) (atmosphere binding index defines) and COMMON_RAYTRACING_BINDINGS macro — cloud voxel grids (Nubis Cubed 2023, 2026-05-12). Adds BINDING_ATMOSPHERE_CLOUD_D_SUN = 210 and BINDING_ATMOSPHERE_CLOUD_D_AMBIENT = 211, both TEXTURE3D. 256x256x32 R16F camera-centered tile-wrapped voxel grids storing summed optical depth along the sun direction (D_sun) and zenith (D_ambient). Round-robin baked every 8 frames by cloud_sun_density_grid.comp.slang / cloud_ambient_density_grid.comp.slang dispatched from RtxAtmosphere::computeLuts; bound via fork_hooks::bindAtmosphereLuts. Sampled at shade time via sampleDSun / sampleDAmbient helpers in atmosphere_common.slangh. No consumer in this commit — the Nubis Cubed cloud-lighting rewrite (C4-C6 of the 2026-05-12 workstream) reads them.
Inline tweak at COMMON_BINDING_DEFINITION_LIST macro (~line 91) — 1-line addition for sampler readback buffer. Adds RW_STRUCTURED_BUFFER(BINDING_SAMPLER_READBACK_BUFFER) to the common binding list (upstream omission fixed).

src/dxvk/shaders/rtx/pass/common_bindings.slangh

Pre-refactor fork footprint: +10 / -0 LOC (audit 2026-04-18)

Category: index-only

Inline tweak at (file scope) (atmosphere LUT texture declarations) (~line 118) — 7-line addition. Declares AtmosphereTransmittanceLut, AtmosphereMultiscatteringLut, and AtmosphereSkyViewLut as Texture2D resources bound at the three atmosphere binding slots.
Inline tweak at (file scope) (atmosphere FAST-noise texture declaration) (~line 138) — 2-line addition (2026-05-09). Declares AtmosphereFastNoise as a Texture2DArray<float2> resource bound at BINDING_ATMOSPHERE_FAST_NOISE (slot 205). Used by the fastJitter() helper in atmosphere_common.slangh for cloud ray-march sample-distribution jitter.

src/dxvk/shaders/rtx/pass/composite/composite.comp.slang

Category: migrate

Inline tweak at applySkyContribution (sky-miss composition) — 2026-07-05. Removes the SkyMatte SkyLight.SampleLevel else-branch on primary miss. Since sync 1448e4e77, geometry_resolver.slangh already writes rasterized sky (skyMode==0; SkyMatte for camera rays / SkyProbe for portal-crossed rays since 2026-07-05) or Numos evalSkyRadiance (skyMode==1) into SharedRadiance on primary miss; composite was double-counting the same sky via SkyMatte. Commit 49229eda4 fixed the double sky by removing the geometry_resolver branch, which broke DLSS-RR particle ordering — this composite-side skip restores official 83150626ef particle behavior while keeping single sky. Dome-light path unchanged.
Block at compositePixel (cloud-shadow application) — REMOVED 2026-06-19, comment-only. The entire screen-space cloud-shadow application in composite is gone. The indirect multiply was removed 2026-06-18 (issue #37, double-count + geometry-blind). The direct multiply (pow(PrimaryCloudShadowFactor, cloudShadowFactorStrength) onto post-denoise primary direct radiance) was removed 2026-06-19 when the cloud shadow was re-architected onto the SUN's radiance pre-denoise inside sampleAtmosphereSunLight — it now darkens only the sun (correct indoors for all surface types, no per-pixel gate). The PrimaryCloudShadowFactor texture binding is also removed; only a removal comment remains in the shader.

src/dxvk/shaders/rtx/pass/composite/composite_args.h

Category: index-only

Inline tweak — cloud-shadow CB fields, now both reserved pads. Two former cloud-shadow CB slots are now reserved float pad1 / float pad2 (CB layout ABI-unchanged). pad1 held cloudShadowFactorStrength until 2026-06-19, when the cloud shadow moved onto the sun term in the NEE and the composite application was deleted (the knob now lives in atmosphere_args.h::cloudShadowFactorStrength, reusing the former pad_artistic0 slot there). pad2 held cloudShadowIndirectStrength until 2026-06-18 (issue #37 indirect multiply removal). Rename-to-pad rather than delete keeps the slots ABI-stable.

src/dxvk/shaders/rtx/pass/gbuffer/gbuffer.slang

Pre-refactor fork footprint: +16 / -0 LOC (audit 2026-04-18)

Category: index-only

Inline tweak at (file scope module-level pragmas) (ATMOSPHERE_AVAILABLE defines) (~lines 35–249) — 16-line addition spread across many pragmas. Adds //!> ATMOSPHERE_AVAILABLE Slang module dependency annotation lines so the gbuffer module can access atmosphere functionality when the define is active.

src/dxvk/shaders/rtx/pass/gbuffer/gbuffer_miss.rmiss.slang

Pre-refactor fork footprint: +1 / -0 LOC (audit 2026-04-18)

Category: index-only

Inline tweak at (file scope) (~line 44) — 1-line addition. Adds #define ATMOSPHERE_AVAILABLE so the gbuffer miss shader can reference atmosphere evaluation functions.

src/dxvk/shaders/rtx/pass/gbuffer/gbuffer_psr_miss.rmiss.slang

Pre-refactor fork footprint: +1 / -0 LOC (audit 2026-04-18)

Category: index-only

Inline tweak at (file scope) (~line 44) — 1-line addition. Adds #define ATMOSPHERE_AVAILABLE so the gbuffer PSR miss shader can reference atmosphere evaluation functions.

src/dxvk/shaders/rtx/pass/integrate/integrate_direct.slang

Category: index-only

Inline tweak at (file scope) (~line 36) — 1-line addition. Adds #define ATMOSPHERE_AVAILABLE so the direct-integration pass compiles against the real evalCloudGroundShadow body. The macro gates a binding-free fallback intended for atmosphere LUT compute shaders that lack the cloud-noise SRV; this pass already includes common_bindings.slangh (which declares AtmosphereCloudNoise3D + sampler), so the fallback over-suppresses cloud shadow on terrain surface NEE. Without this define, evalAtmosphereSunNEE → sampleAtmosphereSunLight → getTransmittanceToSun → evalCloudGroundShadow short-circuits to 1.0 and terrain never darkens under clouds regardless of cloudShadowStrength.

src/dxvk/shaders/rtx/pass/integrate/integrate_indirect.slang

Category: index-only

Inline tweak at (file scope) (~line 233) — 1-line addition. Adds #define ATMOSPHERE_AVAILABLE so the indirect-integration pass evaluates the real evalCloudGroundShadow for secondary-bounce surface NEE (evalAtmosphereSunNEESecondary). Same rationale as integrate_direct.slang: the cloud-noise SRV is bound via common_bindings.slangh here, so the binding-free fallback is unnecessary.

src/dxvk/shaders/rtx/pass/integrate/integrate_indirect_closesthit.rchit.slang

Category: index-only

Inline tweak at (file scope) (~line 241) — 1-line addition. Adds #define ATMOSPHERE_AVAILABLE so the closest-hit variant of indirect integration evaluates the real evalCloudGroundShadow. Same rationale as the sibling integrate_indirect.slang entry — common_bindings.slangh provides the cloud-noise SRV.

src/dxvk/shaders/rtx/pass/integrate/integrate_indirect_miss.rmiss.slang

Pre-refactor fork footprint: +1 / -0 LOC (audit 2026-04-18)

Category: index-only

Inline tweak at (file scope) (~line 66) — 1-line addition. Adds #define ATMOSPHERE_AVAILABLE so the indirect miss shader can evaluate atmosphere sky radiance on rays that miss all geometry.

src/dxvk/shaders/rtx/pass/raytrace_args.h

Pre-refactor fork footprint: +3 / -0 LOC (audit 2026-04-18)

Category: index-only

Inline tweak at RaytraceArgs struct (atmosphereArgs + skyMode) (~lines 153, 363) — 2-line addition. Adds AtmosphereArgs atmosphereArgs and uint skyMode fields to RaytraceArgs so the atmosphere parameters and sky mode flag are available in all ray-tracing passes via the constant buffer.
Inline tweak at (file scope) (atmosphere args include) (~line 35) — 1-line addition. Adds #include "rtx/pass/atmosphere/atmosphere_args.h" so AtmosphereArgs is defined.

src/dxvk/shaders/rtx/pass/rtxdi/restir_gi_reuse_binding_indices.h

Pre-refactor fork footprint: +20 / -20 LOC (audit 2026-04-18)

Category: migrate

Block at (file scope) (binding index renumbering) — ~20 LOC replacing 20 LOC, planned target fork_hooks::restirGiBindingIndices in rtx_fork_atmosphere.slangh. Renumbers the ReSTIR GI reuse pass binding indices (WORLD_SHADING_NORMAL_INPUT through RESERVOIR_INPUT_OUTPUT) to make room for the three atmosphere LUT bindings at slots 200-202, avoiding conflicts introduced by the common-bindings expansion.

src/dxvk/shaders/rtx/pass/tonemap/aces.slangh

Fork-owned — new file. ACES operator implementations: acesHill (Stephen Hill ACES fit) and acesNarkowicz (Krzysztof Narkowicz ACES approximation). Included by fork_tonemap_operators.slangh; dispatched as tonemapOperatorACESHill / tonemapOperatorACESNarkowicz. Fork-owned ACES operator implementations.

src/dxvk/shaders/rtx/pass/tonemap/adaptation_v1.slangh

Fork-owned — new file. Tiny helper namespace adaptation::v1 exposing ExponentialBlend and AdaptAsymmetric for asymmetric exponential eye adaptation (light-tau / dark-tau). Consumed by auto_exposure.comp.slang. Renodx-attributed (Carlos Lopez Jr., MIT 2025). Fork-owned eye-adaptation primitive.

src/dxvk/shaders/rtx/pass/tonemap/agx.slangh

Fork-owned — new file (renamed from AgX.hlsl). AgX Minimal display rendering transform by Benjamin Wrensch (MIT 2024) — agxMinimalToneMapping(color, saturation, look). Depends on neutwo.slangh for the max-channel pre-scale that normalizes HDR input into the curve's [0, 1] domain. Included by fork_tonemap_operators.slangh. Fork-owned AgX operator implementation (Minimal variant).

src/dxvk/shaders/rtx/pass/tonemap/auto_exposure.comp.slang

Block at (file scope) — full rewrite of the resolve stage to a perceptual observer model. Reads the log-Yf histogram emitted by auto_exposure_histogram.comp.slang, computes a count-weighted log mean of Yf as the adapted scene level, and derives the target exposure scale from a first-site cone-contrast law exposure = targetAdaptedYf / (Y_adapt + Y_noise) with Y_noise = 0.0032 (Stockman & Brainard 2010 cone-system noise floor — caps the dark-scene boost without an arbitrary clamp). targetAdaptedYf (default 0.18, i.e. mid-gray reflectance) is a push-constant sourced from cb.targetAdaptedYf (RtxOption rtx.autoExposure.targetAdaptedYf). The raw target is then clamped to cb.maxExposure (RtxOption rtx.autoExposure.maxExposure, default 8.0) before the temporal blend so the smoother converges to the cap directly. Asymmetric exponential blending then runs in log-exposure space so the time-constants are invariant to absolute scene level. Spatial center-weighting (the previously-per-bin log-Yf Gaussian) was lifted into auto_exposure_histogram.comp.slang; bin counts arrive pre-weighted so this pass is a plain count-weighted log mean. Replaces the prior Naka-Rushton-in-resolve form (per-bin Gaussian + 1 / (L + sigma)). Auto-exposure resolve now shares an achromatic basis (Yf) with the psycho17 tonemap operator and exposes a tunable mid-gray target plus a hard exposure cap.

src/dxvk/shaders/rtx/pass/tonemap/auto_exposure_histogram.comp.slang

Inline tweak at inputToHistogramBucket — bin Stockman-Sharpe CIE 170-2 luminosity Yf (via renodx::tonemap::psycho::yf::from_BT709) instead of BT.709 photometric luminance, so the resolve pass and the psycho17 observer share a single physiological achromatic measure. Also collapses NaN / negative inputs to bin 0 via !(yf >= eps). Adds #include "rtx/pass/tonemap/psycho17.slangh" for the Yf helper. Yf is rescaled by kYfRefWhite (Yf-of-BT.709-white, ~1.0504) so 18% linear reflectance lands at Yf == 0.18 exactly — without it rtx.autoExposure.targetAdaptedYf = 0.18 would silently target ~0.171 reflectance.
Inline tweak at main — replaces the unweighted InterlockedAdd(g_localData[bin], 1) with spatial Gaussian center-weighted metering. Distance is normalised by the shorter screen axis so the falloff is circular in pixel space and the horizontal periphery on 16:9 / 21:9 doesn't pin adaptation on the sky; sigma = 0.25 of the short axis. Per-pixel weight is fixed-point (gauss × 256, floored to 1u) so any pixel still nudges the histogram and the resolve's sum(weight*logYf) / sum(weight) cancels the scale. Histogram now lives in observer-model space (normalised so 18% reflectance ⇔ Yf 0.18) and meters center-weighted instead of uniformly — adaptation tracks the subject the gaze fixates on rather than the unweighted scene mean.

src/dxvk/shaders/rtx/pass/tonemap/fork_tonemap_operators.slangh

Fork-owned — new file. Hosts the applyTonemapOperator(uint op, float3 color, bool suppressBlackLevelClamp, ..., float3 adaptiveStateBT709) dispatcher. Branches on the operator enum and forwards into the operator-specific headers (aces.slangh, hable.slangh, agx.slangh, lottes.slangh, psycho17.slangh, gt7.slangh). The trailing adaptiveStateBT709 parameter carries the perceptual-AE observer adaptive state into psycho17's current_adaptive_state_bt709 / current_background_state_bt709 slots; the other operators ignore it. Fork-owned shader header: operator dispatch lives here so upstream passes shrink to one-line calls.

src/dxvk/shaders/rtx/pass/tonemap/gt7.slangh

Fork-owned — new file (2026-05-XX). Slang port of the Polyphony Digital "GT7 Tone Mapping" reference (MIT 2025, SIGGRAPH 2025 supplemental). SDR mode, peak hardcoded to 1.0, ICtCp UCS. Wired into the dispatcher via tonemapOperatorGT7 (= 7). Fork-owned GT7 operator implementation.

src/dxvk/shaders/rtx/pass/tonemap/hable.slangh

Fork-owned — new file. Hable Filmic (Uncharted 2) tonemap operator — hableFilmicToneMapping(color, exposureBias, A, B, C, D, E, F, W). Split out of fork_tonemap_operators.slangh so each operator has its own header. Included by fork_tonemap_operators.slangh. Fork-owned Hable Filmic operator implementation.

src/dxvk/shaders/rtx/pass/tonemap/lottes.slangh

Fork-owned — new file (renamed from Lottes.hlsl). Lottes 2016 tonemap operator — lottesToneMapping(color, hdrMax, contrast, shoulder, midIn, midOut). Lottes shares Hable Filmic's 8 param slots in the shader args struct (the two operators are mutually exclusive); slot mapping is documented at the struct definition in tonemapping.h. Included by fork_tonemap_operators.slangh. Fork-owned Lottes operator implementation.

src/dxvk/shaders/rtx/pass/tonemap/neutwo.slangh

Fork-owned — new file. Slang port of the renodx "Neutwo" max-channel pre-scale helper — neutwo_ComputeMaxChannelScale(color) returns a channel-coherent scale that brings HDR-range input into the [0, 1] curve domain; neutwo_Neutwo(x) = x * rsqrt(x*x + 1) is the underlying saturation kernel. Currently consumed by agx.slangh. Renodx-attributed (Carlos Lopez Jr., MIT 2025). Fork-owned curve-normalization helper.

src/dxvk/shaders/rtx/pass/tonemap/psycho17.slangh

Fork-owned — new file. Self-contained Slang port of the renodx "Psycho Test 17" operator and its required color-pipeline dependencies (Stockman-Sharpe LMS, CIE 170-2 MacLeod-Boynton + gamut, Naka-Rushton, color grading). Dispatched as tonemapOperatorPsycho17 (UI label: PsychoV17_Beta). Renodx-attributed (Carlos Lopez Jr., MIT 2025). Fork-owned Psycho Test 17 operator implementation.

src/dxvk/shaders/rtx/utility/pq.slangh

Fork-owned — new file (2026-05-XX). Shared SMPTE ST.2084 (PQ) constants + PQDecode / PQEncode (vec3, donut-attributed) + scalar pq_eotfSt2084 / pq_inverseEotfSt2084 (GT7-style, frame-buffer units). Extracted from temporal_aa.comp.slang so both the TAA pass and the GT7 tonemap operator can share the same math. Fork-owned shared PQ math.

src/dxvk/shaders/rtx/pass/temporal_aa/temporal_aa.comp.slang

Inline tweak at the include block + PQ constants/helpers (~lines 22-83) — replaced the inlined PQ constants and PQDecode / PQEncode definitions with #include "rtx/utility/pq.slangh" so the GT7 tonemap operator can share them. The function signatures and constant values are unchanged. PQ helpers extracted to a shared header; donut attribution preserved in utility/pq.slangh.

src/dxvk/shaders/rtx/pass/tonemap/tonemapping.h

Inline tweak at (file scope) (operator constants) — add tonemapOperatorNone / tonemapOperatorACESHill / tonemapOperatorACESNarkowicz / tonemapOperatorHableFilmic / tonemapOperatorAgX / tonemapOperatorLottes / tonemapOperatorPsycho17 / tonemapOperatorGT7 (renamed from the original tonemapOperatorACES / tonemapOperatorACESLegacy in the 2026-05-XX cleanup; Psycho11 was renamed to Psycho17 when the operator was replaced with a port of renodx Psycho Test 17 — see src/dxvk/shaders/rtx/pass/tonemap/psycho17.slangh for the MIT attribution to Carlos Lopez Jr. The UI dropdown label is PsychoV17_Beta. tonemapOperatorGT7 was added 2026-05-XX for the Polyphony Digital GT7 SDR port — see src/dxvk/shaders/rtx/pass/tonemap/gt7.slangh).
Inline tweak at ToneMappingAutoExposureArgs — doc comment updated to describe the new perceptual pipeline (log2-Yf histogram + geometric mean + first-site cone-contrast law + log-space asymmetric blend) instead of the previous BT.709-luminance + Gaussian + Naka-Rushton-in-resolve description. Fields: 2 of the 3 trailing uint pad slots were replaced with float targetAdaptedYf (mid-gray adaptation target, sourced from rtx.autoExposure.targetAdaptedYf) and float maxExposure (hard ceiling on the auto-exposure multiplier, sourced from rtx.autoExposure.maxExposure); 1 uint pad0 remains to preserve the 20-byte struct size.
Inline tweak at ToneMappingApplyToneMappingArgs struct — swap finalizeWithACES/useLegacyACES uints for tonemapOperator + per-operator param blocks (Hable, AgX, Psycho17). The vestigial directOperatorMode field and the legacy histogram / tone-curve bindings (TONEMAPPING_HISTOGRAM_*, TONEMAPPING_TONE_CURVE_*, TONEMAPPING_APPLY_TONEMAPPING_TONE_CURVE_INPUT) and the structs ToneMappingHistogramArgs / ToneMappingCurveArgs were removed in the 2026-05-XX cleanup. static_assert(sizeof(...) == 176) pins the current struct size: AgX block is 16 B (saturation + look + 2 pad floats) since the AgX Minimal operator only consumes those two fields; Psycho17 block is 64 B (14 floats + 2 trailing pad floats for 16-byte alignment). Global tonemap shader-shared header adopts the operator enum.

src/dxvk/shaders/rtx/pass/tonemap/tonemapping_apply_tonemapping.comp.slang

Inline tweak at applyToneMapping — replace if (cb.finalizeWithACES) { color = ACESFilm(color, cb.useLegacyACES); } with color = applyTonemapOperator(cb.tonemapOperator, color, false, ..., adaptiveStateBT709);. Add #include "rtx/pass/tonemap/fork_tonemap_operators.slangh". The 2026-05-XX cleanup also stripped the dead helpers (reinhardToneMapper, filmicToneMapper, dynamicToneMapper, lumaAverage, setSaturationAverage) and the InToneCurve binding. adaptiveStateBT709 is the observer adaptive state in post-AE-exposure BT.709 space — currently (0.18, 0.18, 0.18) because the perceptual auto-exposure brings the geometric-mean scene Yf to mid-gray; consumed by psycho17, ignored by other operators. Global apply pass routes through the fork dispatcher for operator selection.
Inline tweak at applyToneMapping (dither gate) — fork — 2026-06-27 (experimental). Adds cb.performSRGBConversion > 0 && to the enableDithering argument of the inline ditherTo8Bit call. The fork's tonemap refactor defers sRGB conversion + dithering to the dedicated srgb_dither pass, calling the tonemapper with performSRGBConversion=false. The inline dither was left active, so it added its ±0.5/255 perturbation in LINEAR space; the later linearToGamma in srgb_dither amplifies that ~13x in the near-black sRGB toe, producing a heavy boiling grain visible only in dark regions (plus a redundant double-dither with the dedicated pass). Gating on performSRGBConversion disables the inline dither in the current pipeline (the gamma-space srgb_dither pass owns dithering) and keeps it correct should the tonemapper ever do its own sRGB encode. VALIDATED in-game 2026-06-27 — dark-region grain gone.

src/dxvk/shaders/rtx/pass/volume_args.h

Category: index-only

Inline tweak at the VolumeArgs struct tail (replaces the vec2 pad0 slot) — 1 LOC delta (2026-05-26). Adds float fogSunVisibilityGain; consuming 4 bytes of the existing 8-byte pad0 slot; pad0 shrinks to float pad0; to preserve 16-byte struct alignment. Consumed by the submodule fork edit at rtxdi-sdk/include/volumetrics/rtx/algorithm/volume_composite_helpers.slangh (named knob replaces the previous hardcoded * 10.0f artistic gain at the fog-render consumer site). Companions: rtx_global_volumetrics.{h,cpp}.

src/dxvk/rtx_render/rtx_fork_atmosphere.cpp

Category: fork-owned (modifications by weather preset workstream)

Note: This is a fork-owned file. It is listed here because the weather preset workstream added a call site inside showAtmosphereUI(), extending the fork-owned ImGui surface.

Inline tweak at fork_hooks::showAtmosphereUI (weather UI call site) — ~1 LOC. Calls fork_hooks::showWeatherUI() between the Moons and Clouds collapsing-header tree blocks so the Weather Presets panel appears in the correct visual position in the Atmosphere dev menu.
Block (anonymous namespace) + call from fork_hooks::updateAtmosphereConstants (directional sun/moon lights — experiment, branch experiment/atmosphere-directional-sun, 2026-06-21) — ~180 LOC. When rtx.atmosphere.useDirectionalLights is on and skyMode==Numos, injects the sun + each enabled moon as externally-tracked RtDistantLights (via LightManager::createExternallyTrackedLight / updateExternallyTrackedLight) so they go through the standard NEE/RTXDI path and SSS/decals/viewmodels are handled by the unified pipeline. Radiance is the CPU port of sampleAtmosphereSunLight/sampleAtmosphereMoonLight ÷ π (distant-light irradiance convention); transmittance is the CPU port of the closed-form getAtmosphericTransmittanceForDir so sunset reddening is preserved. The bespoke evalAtmosphereSunNEE/MoonNEE are gated off via debugSkyBisectFlags bit 2. KNOWN v1 limitation: no cloud-on-terrain shadow.

src/dxvk/rtx_render/rtx_fork_hooks.h

Category: fork-owned (forward declaration additions)

Note: This is a fork-owned file. It is listed here because the weather preset workstream added two new forward declarations to the fork_hooks namespace block.

Inline tweak at fork_hooks namespace block (forward declarations) — ~2 LOC. Adds void updateWeatherBlender(class RtxContext& ctx, float deltaTimeSeconds) and void showWeatherUI() forward declarations. These allow rtx_context.cpp and rtx_fork_atmosphere.cpp to call the weather hook without including the full rtx_fork_weather.h header at those call sites.

submodules/rtxdi/rtxdi-sdk/include/volumetrics/rtx/algorithm/volume_integrator.slangh

Category: submodule (fork-controlled — RemixProjGroup/RTXDI branch remix)

Note: This file lives in the fork-controlled RTXDI submodule. Edits land via PR/commit to RemixProjGroup/RTXDI branch remix and then a sibling dxvk-remix commit bumps the submodule pointer (mirror of 96c56d5). The audit script scripts/audit-fork-touchpoints.sh does NOT inspect submodule files; this entry exists for rebase-safety / human discoverability.

Inline tweak at the top of the file (atmosphere helper include) — 1-line addition. Adds #include "rtx/pass/atmosphere/atmosphere_common.slangh" so the per-froxel atmosphere-sun NEE block and the sky-ambient hemisphere integration block can call atmosphere helpers (sampleAtmosphereSunLightVolume, sampleSkyAmbientForVolume, hgPhase).
REMOVED 2026-06-28 — the bespoke atmosphere sun NEE block (cb.skyMode == 1, ~35 LOC, originally CattaRappa RTXDI commit 2ff8c57) that built an AtmosphereVolumeSunSample via sampleAtmosphereSunLightVolume, traced its own visibility ray, and added the sun to radianceSH. Why removed: since the directional-light graduation the atmosphere sun/moons are real Remix RtDistantLights sampled by the volume NEE/ReSTIR loop like any light (no distant-light exclusion — rtx_light_manager.cpp gives every non-empty type volumeRISSampleCount ≥ 1), so this block double-counted the sun into the froxel cache (once via NEE, once here) with mismatched phase/visibility math — making volumetric fog impossible to balance (the user had to crank fogSunVisibilityGain to 0.01 and it still looked wrong; issue #35). This mirrors the earlier removal of the bespoke surface sun NEE. atmosphereSunVolumetricRadianceScale (which scaled only this block) was left behind with no consumer and was removed 2026-07-26: the option, its AtmosphereArgs slot (now padRetired11), the "Atmosphere Sun Fog Scale" ImGui slider, and the atmosphereSunFogScale weather-preset field across all 12 presets. Use rtx.volumetrics.fogSunVisibilityGain to scale fog in-scattering.
Inline addition 2026-06-28 — applyAtmosphereCloudShadowVolume(inout RAB_LightSample, uint lightIdx, vec3 froxelWorldPos) helper + one call in each NEE branch (ReSTIR + non-ReSTIR fallback) of integrateVolume, before evalNEE. Preserves cloud-on-terrain shadowing on volumetric god-rays now that the bespoke block (which applied it) is gone. Mirrors the surface fold in integrator_direct.slangh: when skyMode == 1 && cloudVoxelShadowsEnable and the sampled light's atmosphereCloudShadowed flag (distant-light GPU flag bit 2) is set, multiplies the light sample's radiance by pow(sampleCloudGroundShadow_OptionB(froxelPos, dirToLight, …), cloudShadowFactorStrength). Guards against RTXDI_INVALID_LIGHT_INDEX. Reuses only bindings the removed block already required (lights[], atmosphere helpers via the file's existing ATMOSPHERE_AVAILABLE define).
Inline tweak at the end of integrateVolume (sky-ambient hemisphere integration block, cb.skyMode == 1 && cloudSkyAmbientStrength > 0) — ~50 LOC. (Workstream 2026-05-12.) Fixed 6-direction upper-hemisphere integration (zenith + 5 mid-elevation at 30° elevation, 72° azimuth spacing) of sampleSkyAmbientForVolume(dir, args, AtmosphereSkyViewLut, AtmosphereCloudSkyTransmittanceLut, sampler) weighted by HG phase against the volumetric anisotropy (0.3). Results scaled by cloudSkyAmbientStrength, firefly-filtered, and stored as a single SH entry with zenith as the dominant direction. Gated on cb.skyMode == 1 and on the strength knob being > 0 so the baseline ships with zero behavior change (cloudSkyAmbientStrength default = 0). Consumes the sky-view LUT (slot 202), cloud-sky-transmittance LUT (slot 208), and the cloud-noise sampler (slot 204 — REPEAT, correct on azimuth, never sampled below-horizon). See docs/superpowers/specs/2026-05-12-volumetric-sky-ambient-design.md.

Commit C4 — Cloud render compute pass + Nubis Cubed equations (fork — 2026-05-12)

The C4 commit of the 2026-05-12 cloud-lighting workstream lands the per- pixel screen-space cloud raymarch with the Nubis Cubed 2023 lighting equations (paper pp. 137, 142) and a debug view (enum 876) for standalone A/B against the existing analytical evalClouds rendering. No production consumer yet — the sky-miss composite still calls analytical clouds; the composite gate lands in C5.

src/dxvk/shaders/rtx/pass/atmosphere/cloud_render.comp.slang — new file (fork-owned). Per-pixel view-direction raymarch through the cloud slab using the Nubis Cubed lighting equations. Reconstructs viewDir from CPU-pushed Y-up basis vectors (cloudRenderForwardYUp / RightYUp / UpYUp in AtmosphereArgs, pre-scaled by tan(halfFovX/Y) and aspect). Intersects the curvature-adjusted base/top cloud shells (intersectSphere) to get [tEntry, tExit]; marches with per-pixel FAST-noise jitter. At each density-passing sample calls evalNubisCubedSample for the page-137 two-HG-lobe direct term + page-142 ambient exp(-D_ambient). Writes premultiplied rgb + view-ray transmittance alpha to AtmosphereCloudRender at slot 209.
src/dxvk/shaders/rtx/pass/atmosphere/atmosphere_args.h — fork-owned additions. Adds 8 floats of Nubis Cubed lighting params (cloudPhaseG1/G2, cloudMsSunDotMax, cloudMsSigmaShallow/Deep, cloudMsSdfDepth) + cloudRenderFrameIdx + pad; plus 3 × (vec3 + pad) for the cloud-render camera basis (cloudRenderForwardYUp, cloudRenderRightYUp, cloudRenderUpYUp). All consumed exclusively by cloud_render.comp.slang; the basis is pushed CPU-side from updateAtmosphereConstants before computeLuts runs so the values land in m_constantsBuffer in time.
src/dxvk/shaders/rtx/pass/atmosphere/atmosphere_common.slangh — fork-owned additions. Adds five Nubis Cubed lighting helpers (~120 LOC) right after sampleDAmbient: sampleDimProfile (proxy on cloudTypeProfile), sampleCloudSdf (slab-distance + density-weighted-depth proxy, returns negative-inside meters clamped to [-cloudMsSdfDepth4, 0]), hgPhaseNubis (paper-flavored HG with denom guard — distinct from the existing hgPhase to avoid perturbing non-Nubis callers), NubisCubedLighting struct, and evalNubisCubedSample (the paper-page-137 two-HG-lobe direct term + page-142 ambient exp(-D_ambient)). Calls sampleDSun / sampleDAmbient from C1.*
src/dxvk/shaders/rtx/pass/common_binding_indices.h — index-only, fork. Adds BINDING_ATMOSPHERE_CLOUD_RENDER_RT = 209 and a TEXTURE2D entry in COMMON_RAYTRACING_BINDINGS. Slot 209 was reserved between 208 (cloud-sky-transmittance LUT) and 210 (cloud D_sun); fills the gap.
src/dxvk/shaders/rtx/pass/common_bindings.slangh — index-only, fork. Declares Texture2D<float4> AtmosphereCloudRender at slot 209 with a 6-line comment block. Consumed by the cloud render RT debug view (enum 876) and — in C5 — by the sky-miss composite path.
src/dxvk/rtx_render/rtx_atmosphere.h — fork-owned additions. Adds m_cloudRenderRT (Resources::Resource), m_cloudRenderExtent (VkExtent2D), m_cloudRenderForwardYUp / RightYUp / UpYUp (Vector3), m_cloudRenderFrameIdx (uint32_t); plus public methods getCloudRenderRT(), ensureCloudRenderRT(ctx, downscaleExtent), setCloudRenderCameraBasis(forward, right, up, frameIdx); plus private dispatchCloudRender(ctx).
src/dxvk/rtx_render/rtx_atmosphere.cpp — fork-owned additions. Adds #include <rtx_shaders/cloud_render.h>, the CloudRenderShader ManagedShader class (7-slot binding parameter list), ensureCloudRenderRT (resize-aware alloc of RGBA16F at downscale extent), setCloudRenderCameraBasis (member-state setter), dispatchCloudRender (rebuilds the args buffer, binds the 7 slots, dispatches 8×8 thread groups), populates the 6 Nubis Cubed lighting fields + 3-basis-vector camera fields + frameIdx into AtmosphereArgs in getAtmosphereArgs(), calls dispatchCloudRender(ctx) from computeLuts after the voxel grid bakes, and binds slot 209 (BINDING_ATMOSPHERE_CLOUD_RENDER_RT) in bindResources (which mirrors the active bindAtmosphereLuts site).
src/dxvk/rtx_render/rtx_options.h — fork-owned additions. Adds 6 RTX_OPTION declarations in the rtx.atmosphere cluster: cloudPhaseG1 (default 0.8), cloudPhaseG2 (0.3), cloudMsSunDotMax (0.9), cloudMsSigmaShallow (0.25), cloudMsSigmaDeep (0.05), cloudMsSdfDepth (128.0 meters). All surface as ImGui sliders in the Nubis Cubed Lighting collapsing block.
src/dxvk/rtx_render/rtx_fork_atmosphere.cpp — fork-owned additions. In updateAtmosphereConstants: reads RtCamera basis (forward/right/up + position) + fov + aspect, applies isZUp swap, pre-scales right/up by tan(halfFovX/Y) and aspect ratio, and pushes via setCloudRenderCameraBasis before computeLuts. Also calls ensureCloudRenderRT with the current downscale extent. In bindAtmosphereLuts: adds the cloud render RT bind at slot 209. Adds new getCloudRenderRT(ctx) accessor for the debug view. Adds a "Nubis Cubed Lighting (fork — 2026-05-12)" ImGui collapsing header inside the Clouds tree with 6 sliders mapping to the 6 new RTX_OPTIONs.
src/dxvk/rtx_render/rtx_context.h — fork-touchpoint inline tweak. Adds forward declaration Resources::Resource fork_hooks::getCloudRenderRT(RtxContext&) and a matching friend line inside class RtxContext. Mirrors the existing getCloudDSun / getCloudDAmbient pattern.
src/dxvk/shaders/rtx/utility/debug_view_indices.h — index-only, fork. Adds DEBUG_VIEW_CLOUD_RENDER_RT = 876 with a 4-line comment block.
src/dxvk/shaders/rtx/pass/debug_view/debug_view.comp.slang — fork-owned addition. Adds a [[vk::binding]]-decorated Texture2D<float4> DebugViewCloudRenderRT declaration and a case DEBUG_VIEW_CLOUD_RENDER_RT arm in the main switch that samples the RT via Load and returns its rgb (alpha is the view-ray transmittance, not relevant to the standalone debug view).
src/dxvk/shaders/rtx/pass/debug_view/debug_view_binding_indices.h — index-only, fork. (Inventory substitution: not listed in the Task 4 spec, but structurally required for the debug view case to access the cloud render RT — mirrors the D_sun/D_ambient pattern at slots 35/36.) Adds DEBUG_VIEW_BINDING_CLOUD_RENDER_RT_INPUT = 37.
src/dxvk/rtx_render/rtx_debug_view.cpp — fork-owned addition. Adds a TEXTURE2D(DEBUG_VIEW_BINDING_CLOUD_RENDER_RT_INPUT) line in the debug-view shader's BEGIN_PARAMETER block, binds the cloud render RT each dispatch via fork_hooks::getCloudRenderRT, and adds a label + multi-line description block to the debug-view selector list ("Atmosphere: Cloud Render RT (Nubis Cubed)").

Commit C5 — Sky-miss composite of cloud RT (gated, default-off) (fork — 2026-05-12)

The C5 commit wires the Nubis Cubed cloud render RT (from C4) as the primary-ray sky-miss cloud source, gated by a default-off RTX_OPTION (cloudRenderRTEnable). With the gate off, rendering is bit-identical to pre-C5 — analytical evalClouds continues to run at every site. With the gate on, primary-ray sky-miss reads from the prerendered RT while indirect, PSR, and reflection rays continue to use analytical clouds (the RT is at primary-ray pixel coordinates, sampling it for a non-primary ray direction would return the wrong cloud).

src/dxvk/rtx_render/rtx_options.h — fork-owned addition. Adds RTX_OPTION("rtx.atmosphere", bool, cloudRenderRTEnable, false, …) in the rtx.atmosphere cluster directly after the C4 Nubis Cubed lighting options. Default false; flipped on in C7 after visual gate.
src/dxvk/shaders/rtx/pass/atmosphere/atmosphere_args.h — fork-owned addition. Adds uint cloudRenderRTEnable plus three uint pads at the end of AtmosphereArgs for 16-byte alignment. Sits after the C4 cloud-render camera basis block; no existing field offsets change.
src/dxvk/shaders/rtx/pass/atmosphere/atmosphere_sky.slangh — fork-owned addition. In evalSkyRadiance: adds a trailing default-false bool isPrimaryRay parameter, and a primary-ray-only branch that reads AtmosphereCloudRender.Load(int3(pixelCoord, 0)) and inverts its transmittance alpha into opacity (vec4(rgb, 1 - cloudRT.a)) so the downstream temporal-smoothing / mix composite operates uniformly on either source. Gate is args.cloudRenderRTEnable != 0u && isPrimaryRay. Falls through to analytical evalClouds when the gate is off OR the caller is non-primary.
src/dxvk/shaders/rtx/algorithm/geometry_resolver.slangh — fork-touchpoint inline tweak. Primary sky-miss call site (cb.skyMode == 1 block, formerly ending at the historyResolution argument) now passes /*isPrimaryRay=*/ true as the trailing argument to evalSkyRadiance. The PSR call site (line ~2553) keeps its 5-argument call shape and gets isPrimaryRay=false via default. ~1 LOC.
src/dxvk/rtx_render/rtx_atmosphere.cpp — fork-owned addition. In getAtmosphereArgs() (right after the C4 camera-basis populate block): sets args.cloudRenderRTEnable from RtxOptions::cloudRenderRTEnable() and zeros the three pad slots. ~5 LOC.
src/dxvk/rtx_render/rtx_fork_atmosphere.cpp — fork-owned addition. Adds a "Master gate (C5)" separator + RemixGui::Checkbox("Composite cloud RT at sky-miss", …) widget at the end of the "Nubis Cubed Lighting" ImGui collapsing header (just below the MS SDF Depth slider). Wired to RtxOptions::cloudRenderRTEnableObject(); tooltip explains the primary-ray-only behavior. ~8 LOC.

Commit C6 — Voxel-grid cloud-on-terrain shadows at NEE (gated) (fork — 2026-05-12)

The C6 commit wires the C3 helper sampleCloudGroundShadow_OptionB into the production surface and volumetric NEE entry points via a multiplicative ratio correction that replaces the legacy evalCloudGroundShadow uniform-dimmer with the rich 3D D_sun voxel-grid lookup. Terrain gains cumulus-shaped drifting shadow patches that match the cloud positions overhead. Gated on a default-off RTX_OPTION (cloudVoxelShadowsEnable).

This commit also fixes two pre-existing concerns in the C3 helper that the diagnostic surfaced: (1) units mismatch — the helper assumed worldPos was in km, but the G-buffer feeds it in engine game units; the helper now converts via the new worldUnitsPerKm field. (2) camera-relative-vs-world-absolute frame — the voxel grid is camera-centered, so the helper now subtracts the camera position (pushed CPU-side via the new setCloudShadowCameraPosition setter) before the cloudVoxelWorldToUVW call. The wire-in is intentionally at the NEE entry points (NOT at getTransmittanceToSun) so the sentinel-position getTransmittanceToSun call from computeGroundReflectionAnalytical continues to consume the legacy uniform-dimmer shadow — preserving the cloud-shadow-map post-mortem's hard-won correctness invariant.

src/dxvk/rtx_render/rtx_options.h — fork-owned addition. Adds two RTX_OPTION declarations in the rtx.atmosphere cluster directly after the C5 cloudRenderRTEnable: cloudVoxelShadowsEnable (default false) and cloudShadowMarchStrength (default 1.0). The strength knob is the Beer-Lambert exponent multiplier inside sampleCloudGroundShadow_OptionB; the C3 commit had to substitute a literal because the field didn't exist on main.
src/dxvk/shaders/rtx/pass/atmosphere/atmosphere_args.h — fork-owned addition. Adds two 16-byte rows at the end of AtmosphereArgs after the C5 block: (cloudVoxelShadowsEnable, cloudShadowMarchStrength, worldUnitsPerKm, pad_c6_0) and (cameraWorldPosYUpKm.xyz, pad_c6_1). All consumed exclusively by sampleCloudGroundShadow_OptionB; the camera world position is pushed CPU-side from updateAtmosphereConstants mirroring the existing setCloudRenderCameraBasis pattern.
src/dxvk/shaders/rtx/pass/atmosphere/atmosphere_common.slangh — fork-owned additions. Two changes. (1) Inside sampleCloudGroundShadow_OptionB_impl: replaces the C3 unit-naive math with a worldPos * (1 / worldUnitsPerKm) game-units → km conversion, a cloudEntryPosKm - args.cameraWorldPosYUpKm camera-relative reframe before cloudVoxelWorldToUVW, and folds args.cloudShadowMarchStrength into the Beer-Lambert exponent (replacing the C3 literal). The legacy cloudShadowStrength mix at the end is preserved. (2) Adds an #ifdef ATMOSPHERE_AVAILABLE-gated ratio correction block inside both sampleAtmosphereSunLight (surface NEE, after result.radiance is set) and sampleAtmosphereSunLightVolume (volumetric NEE, after result.radiance is set) that — when args.cloudVoxelShadowsEnable != 0u — divides out the evalCloudGroundShadow contribution baked into the analytical path and multiplies in the sampleCloudGroundShadow_OptionB result. Skips the correction when the old shadow is below 0.001 to guard against divide-by-zero.
src/dxvk/shaders/rtx/pass/volumetrics/volume_integrate.comp.slang — fork-owned addition. Adds #define ATMOSPHERE_AVAILABLE before the common_bindings.slangh include so the atmosphere helpers consumed by the volumetric pass (sampleAtmosphereSunLightVolume → sampleCloudGroundShadow_OptionB, plus the existing sampleSkyAmbientForVolume, sampleDSun, fastJitter) resolve to bound globals. Matches the #define already present in integrate_direct.slang and integrate_indirect.slang.
src/dxvk/shaders/rtx/pass/volumetrics/volume_restir.comp.slang — fork-owned addition. Same #define ATMOSPHERE_AVAILABLE addition as volume_integrate.comp.slang; the four ReSTIR-stage variants (INITIAL / VISIBILITY / TEMPORAL / SPATIAL_REUSE) all need the cloud voxel-grid bindings available for the per-froxel atmosphere-sun NEE path.
src/dxvk/rtx_render/rtx_atmosphere.h — fork-owned additions. Adds Vector3 m_cameraWorldPosYUpKm (default zero) member and public setCloudShadowCameraPosition(Vector3) setter to support the per-frame push of the camera world position from fork_hooks::updateAtmosphereConstants ahead of computeLuts. Mirrors the existing setCloudRenderCameraBasis plumbing.
src/dxvk/rtx_render/rtx_atmosphere.cpp — fork-owned additions. Implements setCloudShadowCameraPosition (member-state setter). In getAtmosphereArgs() (right after the C5 sky-miss-composite block): populates the four new C6 fields — gate + strength from RTX_OPTIONs, worldUnitsPerKm = 100000 * sceneScale (canonical conversion), and cameraWorldPosYUpKm from the cached member.
src/dxvk/rtx_render/rtx_fork_atmosphere.cpp — fork-owned additions. In updateAtmosphereConstants (immediately after setCloudRenderCameraBasis): reads the camera world position in game units, applies the same toYUp swap used for the basis vectors, converts to km via kmPerWorldUnit = 1 / (100000 * sceneScale), and pushes via setCloudShadowCameraPosition. In the "Nubis Cubed Lighting" ImGui collapsing block (after the Master gate (C5) section): adds a "Cloud-on-terrain shadows (C6)" separator + checkbox bound to cloudVoxelShadowsEnableObject() + a DragFloat slider bound to cloudShadowMarchStrengthObject() with tooltips for both. ~25 LOC total.

Workstream — Cloud system slides 1+3 lift + samplePos parallax fix (fork — 2026-05-15)

Three lifts from "Real-Time Rendering of Volumetric Clouds in Red Dead Redemption 2" (Bauer et al., SIGGRAPH 2019) sit on top of the shipped Nubis Cubed pipeline, plus a one-line bug fix that unblocks all three visually:

Slide 3 — Cloud height LUT. 64×128 RG8 baked once at startup by cloud_height_lut_baker.comp.slang. R = per-altitude density envelope multiplier. G = per-altitude coverage threshold scale (the lever with visible silhouette teeth — lowers the coverage gate near cumulus tops to widen the mushroom-cap horizontally). Sampled by cloud_render.comp.slang via the new cloudHeightProfileFull helper in atmosphere_common.slangh; the procedural cloudTypeProfile is the fallback for the voxel grid bakers and the analytical evalClouds path.
Slide 1 — Two-layer cloud map. The per-pixel march body is extracted into a marchCloudSlab helper with slab altitude / thickness / type / coverage / density-scale parameters. Layer 1 (primary cumulus) marches first; layer 2 (cirrus deck by default) marches after with residual transmittance. Voxel-grid shadows, ground-shadow NEE, and moon shadows remain layer-1-only — cirrus is optically thin enough that the precompute cost isn't justified.
Schneider15 Worley carve. Three new periodic-3D-Worley helpers in atmosphere_common.slangh feed the noise bake. Worley FBM is subtracted from the Perlin base to carve cell silhouettes (chunky cauliflower cumulus instead of smooth Perlin pancakes).
samplePos parallax fix. The per-step world position was being computed as viewDirYUp * t (camera implicitly at origin), gluing the noise field to the camera and letting only wind translate it. Anchoring samplePos to args.cameraWorldPosYUpKm gives real parallax between near and far parts of the volume as the player moves through the world. The three lifts above were inert without this; together they restore the depth cues that make clouds read as 3D volumes instead of cardboard cutouts translating across the sky.

src/dxvk/shaders/rtx/pass/atmosphere/cloud_height_lut_baker.comp.slang — fork-only addition. New one-shot bake compute pass. 8×8 thread groups over a 64×128 R8G8 RWTexture2D. R channel emits the per-type density envelope (trapezoid + Gaussian anvil bump for type > 0.6); G channel emits the per-type coverage threshold scale (1.0 = no effect, drops to ~0.30 at hf ≈ 0.80 for cumulus). Curves are tuned so type values 0 / 0.5 / 1 land close to the procedural cloudTypeProfile shape — keeps default-on visual parity with pre-LUT scenes.
src/dxvk/shaders/rtx/pass/atmosphere/cloud_render.comp.slang — fork-owned additions. Adds two binding slots (10 = Texture2D<float2> height LUT, 11 = SamplerState linear/CLAMP); defines ATMOSPHERE_CLOUD_HEIGHT_LUT_AVAILABLE before the common header so cloudHeightProfileFull resolves to LUT samples. Extracts the per-pixel raymarch into a marchCloudSlab(slabAltKm, slabThickKm, slabTypeMean, slabCoverageMean, slabDensityScale, ctx, args, inout accumColor, inout viewTransmittance) helper plus a CloudShadeContext struct bundling the sun/moon/sky precomputes. main() calls the helper once for layer 1 (always) and once for layer 2 (when args.cloudLayer2Enable != 0u). Anchors samplePos = args.cameraWorldPosYUpKm + viewDirYUp * t (parallax fix). The "trivially clear sky" early-out widens to max(layer1, layer2) coverage so a cirrus-only preset doesn't get short-circuited.
src/dxvk/shaders/rtx/pass/atmosphere/rtx_cloud_noise_baker.comp.slang — fork-owned additions. Adds a worleyFbm3DPeriodic(worldPosKm * worleyFreq, worleyOctaves, 5.0f, basePeriodWorley) tap alongside the existing Perlin base + detail FBM, with worleyFreq / worleyOctaves from args.cloudWorleyFrequency / cloudWorleyOctaves. Output is saturate(baseDensity - worley * args.cloudWorleyCarveStrength) instead of the old smooth Perlin sum.
src/dxvk/shaders/rtx/pass/atmosphere/atmosphere_common.slangh — fork-owned additions. Adds three periodic-3D-Worley helpers (worleyFeaturePoint3D, worleyNoise3DPeriodic, worleyFbm3DPeriodic) used only by the bake. Adds cloudHeightProfileFull (vec2) + sampleCloudHeightLUT gated by ATMOSPHERE_CLOUD_HEIGHT_LUT_AVAILABLE, with a procedural fallback via the existing cloudTypeProfile. Extends sampleCloudDensityTextured and sampleCloudDensityForShadow with slab-parametric overloads (slab altitude / thickness / density-scale) plus thin args-default wrappers preserving the original signature; the slab-parametric versions apply the LUT G channel to the coverage-threshold step and the LUT R channel to the density envelope. Also switches the Y texcoord in both density samplers from slab-relative to isotropic / args.cloudNoiseTileKm (correctness alignment with the isotropic bake).
src/dxvk/shaders/rtx/pass/atmosphere/atmosphere_args.h — fork-owned additions. Adds two 16-byte rows after the C6 camera block: (cloudHeightLutEnable, cloudLayer2Enable, cloudLayer2Altitude, cloudLayer2Thickness) and (cloudLayer2TypeMean, cloudLayer2CoverageMean, cloudLayer2DensityScale, pad_cloudLayer2_0), plus one row for Worley (cloudWorleyCarveStrength, cloudWorleyFrequency, cloudWorleyOctaves, pad_cloudWorley_0).
src/dxvk/rtx_render/rtx_atmosphere.cpp — fork-owned additions. Adds #include <rtx_shaders/cloud_height_lut_baker.h>, the CloudHeightLutBakerShader ManagedShader class (single RW_TEXTURE2D(0)), and dispatchCloudHeightLutBake (one-shot, called from initialize() right after dispatchCloudNoise3DBake). Allocates m_cloudHeightLut (64×128 R8G8_UNORM) in createLutResources. In dispatchCloudRender: creates a linear/CLAMP heightLutSampler, binds slot 10 (LUT view) + slot 11 (sampler), and adds the resource-tracking line. In getAtmosphereArgs: populates the height LUT toggle + six layer-2 fields + three Worley fields from RTX_OPTIONs.
src/dxvk/rtx_render/rtx_atmosphere.h — fork-owned additions. Adds Resources::Resource m_cloudHeightLut member, getCloudHeightLut() accessor, dispatchCloudHeightLutBake declaration, and kCloudHeightLutWidth = 64 / kCloudHeightLutHeight = 128 constants.
src/dxvk/rtx_render/rtx_options.h — fork-owned additions. Adds 10 new RTX_OPTIONs in the rtx.atmosphere cluster: cloudHeightLutEnable (default on), cloudLayer2Enable (default off) + 5 layer-2 tuning floats, and cloudWorleyCarveStrength (0.6) / cloudWorleyFrequency (1.0) / cloudWorleyOctaves (3). The Worley trio is tagged "CHANGE APPLIES ON GAME RELAUNCH" since the noise bake is one-shot at init.
src/dxvk/rtx_render/rtx_fork_atmosphere.cpp — fork-owned additions. Adds Height LUT (1 checkbox), Layer 2 (1 checkbox + 5 sliders), and Worley carve (2 DragFloat + 1 DragInt) ImGui subsections inside the existing Clouds tree, each with tooltips describing the slide source and the relaunch requirement where applicable.

Workstream — Multiscattering: preset-faithful default + physical-blend knob (fork — 2026-05-26)

Owen's 3e37062b (fix(atmosphere): fix multiscattering to match reference two-term model) on canonical replaced the numerical hemisphere integration in computeMultiscattering with an analytical-only fit using heavily blue-biased coefficients (vec3(0.217, 0.347, 0.594) * 0.02), and switched evalAtmosphereRadiance to sample the multiscattering LUT instead of calling the inline analytical helpers. Two consequences:

Cloud-vs-sky color mismatch at sunset. cloud_render.comp.slang reads the sky-view LUT at the sun direction as the warm ambient source for cumulus. The new bake pumped extra blue energy into the LUT at every elevation, so the cumulus ambient term lost its warm tint and read white against an orange sky.
Preset color washed across all defaults. With the LUT consumed by evalAtmosphereRadiance, the hemisphere integration's wavelength bias amplified each preset's Rayleigh into the sky (Earth too blue, Desert blue-ish, Mars desaturated). The previously-inline getAnalyticalMultiscattering was a tame curve fit that the presets were calibrated against.

Fork resolution: restore the numerical hemisphere integration in the LUT bake AND keep the inline analytical multiscattering as the default in evalAtmosphereRadiance, with a per-preset knob to blend in the LUT-based physical version when realism is wanted.

src/dxvk/shaders/rtx/pass/atmosphere/multiscattering_lut.comp.slang — restoration to pre-3e37062b state. Restores computeMultiscattering to the 64-direction × 20-march-sample hemisphere integration (Hillaire 2nd-order scattering primary term) plus ground reflection plus analytical fit. Restores computeAnalyticalMultiscattering coefficients to the toned-down vec3(0.35, 0.38, 0.45) × 0.01 ("more neutral, less blue-heavy") tuning that prevents purple cast when combined with sunset orange.
src/dxvk/shaders/rtx/pass/atmosphere/atmosphere_common.slangh — fork-owned inline tweak. Inside evalAtmosphereRadiance's per-sample loop: replaces the post-3e37062b LUT-only multiscattering with a lerp(contribAnalytical, contribLut, args.multiScatterPhysicalStrength) blend. contribAnalytical calls the existing inline getAnalyticalMultiscattering (preset-faithful, the pre-3e37062b shape); contribLut samples the multiscattering LUT (the now-restored hemisphere integration). Default 0.0 = byte-identical to pre-3e37062b; 1.0 = full physical.
src/dxvk/shaders/rtx/pass/atmosphere/atmosphere_args.h — fork-owned addition. Replaces the existing uint pad2 slot at the end of the LUT-dims 16-byte row with float multiScatterPhysicalStrength. No struct layout change; existing field offsets unchanged.
src/dxvk/rtx_render/rtx_options.h — fork-owned addition. Adds RTX_OPTION("rtx.atmosphere", float, multiScatterPhysicalStrength, 0.0f, …) immediately after sunIlluminance in the rtx.atmosphere cluster.
src/dxvk/rtx_render/rtx_atmosphere.cpp — fork-owned addition. Inside getAtmosphereArgs(): sets args.multiScatterPhysicalStrength = RtxOptions::multiScatterPhysicalStrength(). Removes the now-stale args.pad2 = 0 write (slot is the new typed field). ~2 LOC net.
src/dxvk/rtx_render/rtx_fork_atmosphere.cpp — fork-owned addition. Adds a RemixGui::DragFloat("Multiscatter Physical Strength", …, 0.0f, 1.0f, "%.2f", sliderFlags) widget at the end of the Atmosphere → Advanced ImGui tree (right after Ozone Layer Width), with a tooltip explaining the artistic-vs-physical tradeoff. ~6 LOC.

Workstream — Cloud realism: edge detail erosion, bottom darkening, vertical stretch (fork — 2026-06-10)

Three perceptual cloud upgrades, all live-tunable, all reverting to the prior look at their zero/identity values. (1) Edge detail erosion — a second tap of the prebaked Worley-carved noise volume at cloudDetailScale× the base frequency perturbs the density field before the coverage gate, displacing the silhouette iso-surface with 60–500 m cauliflower teeth while saturated cores stay solid. (2) Bottom darkening — a vertical light gradient on the Nubis Cubed multi-scatter and ambient terms; the paper's M term barely attenuates (sigma_ms ≈ 0.05–0.25) and pow(1 - dim_profile, 0.5) is top/bottom symmetric, so undersides previously read uniformly lit. The direct beam is exempt so silver linings survive. (3) Vertical stretch — lowers the Y noise-sample frequency relative to horizontal so cloud bodies become convective columns rather than round blobs (towering cumulus). New args ride in repurposed pad slots; constant-buffer layout unchanged.

src/dxvk/shaders/rtx/pass/atmosphere/atmosphere_common.slangh — fork-owned additions. In sampleCloudDensityTextured: step 4b applies the detail-erosion tap pre-threshold (density += strength * 0.6 * (detailNoise - 0.6), erosion-biased), and the texcoord mapping divides Y frequency by cloudVerticalStretch. sampleCloudDensityForShadow applies the identical Y-stretch (the baked D_sun / D_ambient grids must describe the same shapes the view-march renders) but skips the detail tap (cheap path; un-eroded grids over-shadow edges conservatively). In evalNubisCubedSample: bottomGradient = mix(1 - cloudBottomDarkening, 1, smoothstep(0, cloudBottomDarkeningHeight, heightFraction)) multiplies the multi-scatter M and L_ambient terms only.
src/dxvk/shaders/rtx/pass/atmosphere/atmosphere_args.h — fork-owned additions. Repurposes pad_cloudVoxel0..2 as cloudBottomDarkening / cloudBottomDarkeningHeight / cloudDetailStrength, pad_nubisCubed0 as cloudDetailScale, and pad_cloudLayer2_0 as cloudVerticalStretch. No layout change (same pattern as cloudNoiseWarpStrength).
src/dxvk/rtx_render/rtx_options.h — fork-owned additions. Adds 5 RTX_OPTIONs to the rtx.atmosphere cluster: cloudDetailStrength (0.6), cloudDetailScale (4.3), cloudBottomDarkening (0.55), cloudBottomDarkeningHeight (0.65), cloudVerticalStretch (1.6).
src/dxvk/rtx_render/rtx_atmosphere.cpp — fork-owned additions. Inside getAtmosphereArgs(): populates the five new args from RtxOptions, replacing the former pad zero-writes. ~5 LOC net.
src/dxvk/rtx_render/rtx_fork_atmosphere.cpp — fork-owned additions. Adds "Edge Detail" + "Vertical Stretch" DragFloats to the Clouds → Coverage & Shape ImGui section and "Bottom Darkening" to Clouds → Lighting. The fine-tune knobs (cloudDetailScale, cloudBottomDarkeningHeight) stay user.conf-only per the 2026-05-19 menu-simplification discipline.

Workstream — Cloud march/bake perf pass + vertical-coherence rework (fork — 2026-06-10)

Performance pass on the per-frame cloud noise evaluation, each piece validated in-game via a step-by-step bisect ladder after an earlier all-at-once landing broke the cloudscape (root causes: an uncompensated warp-amplitude loss and config drift contaminating the first bisect — see the cloud-realism memory notes). Also replaces the rev-1 cloudVerticalStretch Y-domain stretch (stacked-puffs artifact) with a fixed-slice coherence blend, shipped EXPERIMENTAL/default-inert because its look (vertical smearing at high values) isn't approved either.

src/dxvk/shaders/rtx/pass/atmosphere/atmosphere_common.slangh — fork-owned additions. (1) Anti-tile warp refactored into offset form (computeCloudNoiseWarpOffsetKm) and HOISTED to once per ray in the two voxel-grid bake integrals (was 4 perlin evals per tap × 8 taps × 2.1M voxels × 2 grids per frame; the warp's ~25 km shortest wavelength makes per-tap recompute pure waste). sampleCloudDensityForShadow gains a warpOffsetKm caller-provided parameter. (2) Warp octaves switched from constant-z perlinNoise3D planes to new perlinNoise2D (4 corners, branch-free gradient pick, ~2.5× cheaper) with kWarp2DGain = 1.4 compensating the unit-vs-sqrt(2) gradient amplitude — the uncompensated version visibly weakened the warp's load-bearing field shred. (3) The per-sample 720 cycles/km analytic perlin detail is removed (1.4 m wavelength vs ≥100 m steps = unresolvable grain), preserving its hidden −0.05 mean bias as a constant. (4) cloudVerticalStretch rev 2: fixed-Y-slice coherence blend with variance re-expansion anchored on the bake's ~0.4 mean, in BOTH density samplers so the self-shadow grids match.
src/dxvk/shaders/rtx/pass/atmosphere/cloud_render.comp.slang — fork-owned additions. Coverage/type Worley control fields hoisted out of the march loop: evaluated at slab entry + exit (4 evals/ray), lerped per step (was 2 × 9-hash evals per step × 32 steps).
src/dxvk/rtx_render/rtx_options.h — fork-owned addition. cloudVerticalStretch default 1.6 → 1.0 (bit-exact identity) and description rewritten to mark the feature EXPERIMENTAL pending a sky-system-level solution to towering cumulus.

Workstream — Secondary-ray cloud LUT (fork — 2026-06-10, perf)

Every indirect / PSR / reflection ray reaching sky-miss previously ran the full analytical evalClouds march in evalSkyRadiance — a hidden per-ray cost estimated to rival the visible cloud pass. Those rays now sample a 256×128 RGBA16F dome LUT (azimuth × elevation, horizon-concentrated elevation = (π/2)·v²) baked once per frame with the same Nubis Cubed march the visible cloud RT uses. Deliberate look change: secondary rays now see the same clouds the primaries see instead of the legacy Wrenninge analytical approximation; rtx.atmosphere.cloudSecondaryLutEnable = False restores the legacy per-ray march. Cloud parallax across scene-scale ray-origin offsets is negligible versus km-scale cloud distances — the same camera-anchored approximation the per-frame voxel grids already make.

src/dxvk/shaders/rtx/pass/atmosphere/cloud_march_common.slangh — NEW fork-owned file. Shared cloud-march library extracted from cloud_render.comp.slang: CloudShadeContext + buildCloudShadeContext (the pixel-independent sun/sky/moon precompute that previously lived in main()), sampleCloudSunOpticalDepth_local, marchCloudSlab (with generic tMinClamp / tMaxClamp segment params; 0/0 reproduces the pre-extraction march bit-for-bit), and the marchCloudLayers layer-1+2 wrapper. Included after atmosphere_common.slangh under the established binding-then-include layout.
src/dxvk/shaders/rtx/pass/atmosphere/cloud_secondary_lut.comp.slang — NEW fork-owned file. Per-frame dome LUT bake. Bindings 0–11 in lockstep with cloud_render.comp.slang (slot 6 is its own RW output). Marches the full slab per dome texel at cloudViewSamples steps via the shared marchCloudLayers; clear-sky early-out mirrors cloud_render; frame-constant per-texel jitter keeps the LUT temporally stable.
src/dxvk/shaders/rtx/pass/atmosphere/cloud_render.comp.slang — fork-owned changes. March implementation moved to cloud_march_common.slangh; main() now builds the shade context via buildCloudShadeContext and calls marchCloudLayers with 0/0 clamps — behavior-neutral extraction, bit-identical output.
src/dxvk/shaders/rtx/pass/atmosphere/atmosphere_common.slangh — fork-owned addition. cloudDomeUvToDir / cloudDomeDirToUv dome-direction mapping shared by the LUT writer and the evalSkyRadiance sampler (lives here because the sample side does not include the march library).
src/dxvk/shaders/rtx/pass/atmosphere/atmosphere_sky.slangh — fork-owned addition. New middle branch in evalSkyRadiance's cloud-layer selection: non-primary rays with cloudSecondaryLutEnable sample AtmosphereCloudSecondaryLut via cloudDomeDirToUv + the sky-view sampler (REPEAT-U wraps the azimuth seam) and reconcile alpha exactly like the existing cloud-RT branch (opacity = 1 − transmittance). Below-horizon directions clamp to the horizon row.
src/dxvk/shaders/rtx/pass/atmosphere/atmosphere_args.h — fork-owned addition. Repurposes pad_c5_0 as uint cloudSecondaryLutEnable. No layout change (same pattern as cloudNoiseWarpStrength).
src/dxvk/shaders/rtx/pass/common_binding_indices.h — upstream-touched, inline tweak (~12 LOC). Adds BINDING_ATMOSPHERE_CLOUD_SECONDARY_LUT 215 next to the other fork atmosphere bindings and appends TEXTURE2D(BINDING_ATMOSPHERE_CLOUD_SECONDARY_LUT) to COMMON_RAYTRACING_BINDINGS.
src/dxvk/shaders/rtx/pass/common_bindings.slangh — upstream-touched, inline tweak (~10 LOC). Declares Texture2D<float4> AtmosphereCloudSecondaryLut at the new binding, alongside the other fork atmosphere declarations.
src/dxvk/rtx_render/rtx_options.h — fork-owned addition. Adds RTX_OPTION("rtx.atmosphere", bool, cloudSecondaryLutEnable, true, …) directly before the C5 cloudRenderRTEnable.
src/dxvk/rtx_render/rtx_atmosphere.h / rtx_atmosphere.cpp — fork-owned additions. CloudSecondaryLutShader class, m_cloudSecondaryLut resource (256×128 RGBA16F) + getCloudSecondaryLut() accessor + kCloudSecondaryLut* constants, dispatchCloudSecondaryLut (per-frame, after the voxel-grid bakes behind the existing write→read barrier, gated on the option), getAtmosphereArgs populates the gate from RtxOptions, and bindResources binds the new slot.
src/dxvk/rtx_render/rtx_fork_atmosphere.cpp — fork-owned additions. fork_hooks::bindAtmosphereLuts binds the LUT at BINDING_ATMOSPHERE_CLOUD_SECONDARY_LUT; adds the "Fast Cloud Reflections" checkbox to the Clouds ImGui tree.

Workstream — Half-res cloud render RT (fork — 2026-06-11, perf)

The visible cloud march ran once per DLSS-input pixel at up to 32 steps. Clouds are soft, low-frequency content, so the cloud RT is now allocated at cloudRenderResolutionScale (default 0.5) of the downscale extent and bilinearly upsampled at the sky-miss composite — ~4× fewer marched pixels at the default. The temporal-smoothing path runs after the upsample at full downscale resolution, so its stabilization is unchanged. Scale 1.0 lands the sample uv on texel centers of a same-size RT and matches the legacy Load to float precision (live A/B via the "Cloud Render Scale" slider).

src/dxvk/shaders/rtx/pass/atmosphere/atmosphere_sky.slangh — fork-owned change. The primary-ray cloud-RT branch in evalSkyRadiance normalizes pixelCoord by args.cloudRenderFullDimX/Y and bilinearly samples the RT via the sky-view sampler (uv clamped a half-texel inside so screen content never wraps through REPEAT-U); falls back to the legacy Load while the published extent is still zero (first frames).
src/dxvk/shaders/rtx/pass/atmosphere/atmosphere_args.h — fork-owned addition. Repurposes pad_c5_1/2 as cloudRenderFullDimX/Y (the downscale extent the RT is composited into). No layout change.
src/dxvk/rtx_render/rtx_options.h — fork-owned addition. Adds RTX_OPTION("rtx.atmosphere", float, cloudRenderResolutionScale, 0.5f, …).
src/dxvk/rtx_render/rtx_atmosphere.h / rtx_atmosphere.cpp — fork-owned additions. ensureCloudRenderRT records the full downscale extent in new m_cloudRenderFullExtent and allocates the RT at the clamped scale (a live scale change reallocates via the existing extent-mismatch path); getAtmosphereArgs publishes the full extent. Dispatch group math already keys off the (now scaled) m_cloudRenderExtent.
src/dxvk/rtx_render/rtx_fork_atmosphere.cpp — fork-owned addition. Adds the "Cloud Render Scale" DragFloat to the Clouds ImGui tree.

Workstream — Cloud noise hex de-tiling, stage A (fork — 2026-06-11)

Root-cause fix for the prebaked noise volume's periodic repeat (the artifact the anti-tile warp was added to hide). The horizontal plane is partitioned into an equilateral-triangle lattice (golden-ratio multiple of the tile period, so lattice and texture can't resonate); each vertex carries a hash-derived random texture-space XZ offset, and samples blend their triangle's three offset taps with Heitz & Neyret 2018's variance-preserving operator (mean/variance of the field preserved → look unchanged; periodicity destroyed). Weights are pow-4 sharpened so triangle interiors (~70% of area) take a 1-tap fast path; only edge bands pay 3 trilinear taps. Ships alongside the UNCHANGED anti-tile warp — stage B walks the warp down with bake-frequency compensation after in-game validation (the warp's frequency-multiplying Jacobian is load-bearing for the look; see the cloud-realism memory notes).

src/dxvk/shaders/rtx/pass/atmosphere/atmosphere_common.slangh — fork-owned additions. New CloudHexTile / computeCloudHexTile (skewed triangle grid + per-vertex hash offsets + pow-4 weight sharpening + dominant-vertex sort), kCloudNoiseFieldMean (0.4 — shared anchor with the vertical-coherence re-expansion), and sampleCloudNoiseHexBlend (3-tap variance-preserving blend). sampleCloudDensityTextured's base tap branches: hex fast path = tricubic at the dominant offset, blend zones = 3-tap trilinear blend, off = legacy tricubic. sampleCloudDensityForShadow mirrors the same lattice/offsets/blend trilinearly so the baked D_sun / D_ambient grids shadow the same de-tiled field. The vertical-coherence footprint slice follows the dominant vertex's offset in both samplers. The detail-erosion tap stays single-tap (its sub-3 km repeat is not the visible artifact; revisit at stage B).
src/dxvk/shaders/rtx/pass/atmosphere/atmosphere_args.h — fork-owned addition. Repurposes padCloudLook0 as float cloudHexTilingEnable. No layout change.
src/dxvk/rtx_render/rtx_options.h — fork-owned addition. Adds RTX_OPTION("rtx.atmosphere", bool, cloudHexTilingEnable, true, …) before cloudNoiseWarpStrength.
src/dxvk/rtx_render/rtx_atmosphere.cpp — fork-owned addition. getAtmosphereArgs populates the gate from RtxOptions (replacing the former pad zero-write).
src/dxvk/rtx_render/rtx_fork_atmosphere.cpp — fork-owned addition. Adds the "Seamless Cloud Field" checkbox to the Clouds ImGui tree, next to the Anti-Tile Warp slider it will eventually retire.

Workstream — Cloud noise hex de-tiling, stage B enabler (fork — 2026-06-11)

Exposes the bake's base/detail FBM frequency as a live-rebake multiplier so the anti-tile-warp walk-down can be tuned interactively: the warp's Jacobian multiplies the field's horizontal frequency 2-3x and the tuned look depends on that shred, so as cloudNoiseWarpStrength comes down, cloudNoiseBaseFreqScale goes up to fold equivalent frequency content into the bake itself. Default 1.0 = bit-identical bake; the actual walk-down is an in-game tuning session, not a code change.

src/dxvk/shaders/rtx/pass/atmosphere/rtx_cloud_noise_baker.comp.slang — fork-owned change. baseFreq / detailFreq become 0.5 / 2.0 × clamp(args.cloudNoiseBaseFreqScale, 0.25, 4) (4x ratio preserved); the periodic-lattice snapping already handles non-default products.
src/dxvk/shaders/rtx/pass/atmosphere/atmosphere_args.h — fork-owned addition. Repurposes padCloudLook1 as float cloudNoiseBaseFreqScale. No layout change.
src/dxvk/rtx_render/rtx_options.h — fork-owned addition. Adds RTX_OPTION("rtx.atmosphere", float, cloudNoiseBaseFreqScale, 1.0f, …).
src/dxvk/rtx_render/rtx_atmosphere.h / rtx_atmosphere.cpp — fork-owned additions. getAtmosphereArgs populates the new field; needsCloudNoiseRebake / cacheCloudNoiseBakeInputs gain the new input (m_cachedBaseFreqScale) so dragging the slider re-bakes the volume live.
src/dxvk/rtx_render/rtx_fork_atmosphere.cpp — fork-owned addition. Adds the "Noise Frequency" DragFloat next to the Anti-Tile Warp slider with the walk-down pairing explained in its tooltip.

Workstream — Cloud noise hex de-tiling, stage B landing (fork — 2026-06-11)

The in-game warp walk-down validated cleanly at warp 0 with the bake frequency left at its default 1.0 — with hex-tiling killing the repeat at the source, no frequency compensation turned out to be needed. The warp default is retired accordingly; the knob survives as a pure organic-distortion effect. Warp 0 also takes the early-out in applyCloudNoiseAntiTileWarp, refunding 4 perlin2D evaluations per march sample.

src/dxvk/rtx_render/rtx_options.h — fork-owned change. cloudNoiseWarpStrength default 0.75 → 0.0; doc strings for it and cloudNoiseBaseFreqScale updated to describe the post-retirement roles (warp = optional look knob, hex-tiling owns de-tiling).

Workstream — Anti-tile warp removal (fork — 2026-06-11)

Follow-up to the stage B landing: with the warp retired at 0 and hex de-tiling user-validated as the root-cause fix, the entire warp implementation is dead code and is removed. All deleted code was fork-owned; no upstream lines return.

src/dxvk/shaders/rtx/pass/atmosphere/atmosphere_common.slangh — fork-owned removal. Deletes perlinGradient2D / perlinNoise2D (warp-only helpers), computeCloudNoiseWarpOffsetKm / applyCloudNoiseAntiTileWarp, the warp call in sampleCloudDensityTextured, the warpOffsetKm parameter from both sampleCloudDensityForShadow overloads, and the per-ray warp hoists in the D_sun / D_ambient bake integrals.
src/dxvk/shaders/rtx/pass/atmosphere/atmosphere_args.h — fork-owned change. cloudNoiseWarpStrength reverts to padding (padCloudC2, its original slot) — CB layout unchanged.
src/dxvk/rtx_render/rtx_options.h — fork-owned removal. Drops the cloudNoiseWarpStrength RTX_OPTION; cloudHexTilingEnable / cloudNoiseBaseFreqScale doc strings scrubbed of warp pairing references.
src/dxvk/rtx_render/rtx_atmosphere.cpp — fork-owned change. getAtmosphereArgs zeroes padCloudC2 instead of populating the warp strength.
src/dxvk/rtx_render/rtx_fork_atmosphere.cpp — fork-owned removal. Drops the "Anti-Tile Warp" slider from the Clouds ImGui tree; "Seamless Cloud Field" / "Noise Frequency" tooltips scrubbed of warp references.
src/dxvk/shaders/rtx/pass/atmosphere/rtx_cloud_noise_baker.comp.slang — comment-only cleanup (walk-down rationale removed).

Workstream — Sky perf: split sky-LUT cache keys (fork — 2026-06-11)

First change of the non-cloud sky optimization workstream. The three sky LUT bakes (transmittance / multiscatter / sky-view) shared one memcmp gate over the whole normalized AtmosphereArgs, with two per-frame failure modes: the game-driven sidereal starRotation (pushed every frame at night, feeds no LUT bake) re-baked the full cascade every frame, and a moving time-of-day sun re-baked the heavy transmittance + multiscatter pair (multiscatter alone: 32×32 texels × 64 dirs × 20 steps of transmittance taps) even though neither reads sun direction. Each bake now compares a key normalized down to the fields it actually reads. Gated by skyLutCacheKeySplitEnable (default on; legacy single-gate path preserved verbatim for A/B).

src/dxvk/rtx_render/rtx_options.h — fork-owned addition. Adds RTX_OPTION("rtx.atmosphere", bool, skyLutCacheKeySplitEnable, true, …) next to the other atmosphere perf gates.
src/dxvk/rtx_render/rtx_atmosphere.h / rtx_atmosphere.cpp — fork-owned additions. New normalizeForSkyViewLutKey (base normalize + star / Milky Way fields zeroed) and normalizeForTransmittanceMsKey (additionally zeroes sun direction / illuminance / disk size, Mie g, MS blend weight, and all moon fields), cached as m_cachedSkyViewKey / m_cachedTransmittanceMsKey. computeLuts gates the transmittance+MS pair and the sky-view bake independently (tms-dirty implies sky-view-dirty; barrier ordering unchanged); both branches keep the other path's caches coherent so the option can be toggled live without a spurious re-bake.
src/dxvk/rtx_render/rtx_fork_atmosphere.cpp — fork-owned addition. Adds the "Minimal Sky LUT Re-bakes" checkbox to the Atmosphere → Advanced ImGui tree.

Note: rtx.atmosphere.cloudNoiseWarpStrength lines in existing user confs become unknown-option no-ops after this change.

Workstream — Sky perf: per-dispatch bisect toggles (fork — 2026-06-11, diagnostic)

Measurement aid for the sky perf workstream. The atmosphere pass runs several per-frame dispatches that no production option can skip — dispatchCloudRender in particular runs whenever its RT is valid, independent of cloudRenderRTEnable — so frame-time A/B via the production toggles mis-attributes their cost. Three default-ON debugDispatch* toggles let a live ImGui session skip each dispatch individually and read the frame-time delta. Skipping leaves consumers reading stale data (frozen clouds / shadows): diagnostic only, defaults unchanged rendering.

src/dxvk/rtx_render/rtx_options.h — fork-owned additions. debugDispatchCloudVoxelGrids / debugDispatchCloudRender / debugDispatchCloudSkyTransmittance, all default true.
src/dxvk/rtx_render/rtx_atmosphere.cpp — fork-owned change. computeLuts wraps the D_sun + D_ambient bakes, the cloud render dispatch, and the cloud-sky-transmittance bake in their respective toggles (barriers move inside the gates; ordering unchanged when enabled).
src/dxvk/rtx_render/rtx_fork_atmosphere.cpp — fork-owned addition. New "Perf Bisect (Diagnostic)" ImGui tree with the three checkboxes + pointers to the existing production toggles useful in the same session.

Workstream — Sky perf: NEE shadow-ray budget clamps (fork — 2026-06-11)

Third change of the sky perf workstream — the first to target where the milliseconds actually are. Bisect results showed all atmosphere compute dispatches together ≈ the known 2 ms cloud budget; the remaining ~2.7 ms (day) / ~3.7 ms (night) of the skyMode A/B delta lives in the integrators: with physical atmosphere on, sun NEE traces an anisotropy-driven 1–12 soft-shadow visibility rays per primary pixel (getSunSoftShadowParams; ~4 at typical Mie g) plus half that per indirect bounce vertex, and moon NEE adds a constant 4 (+2 indirect) at night. The denoised pipeline temporally converges one blue-noise-jittered ray per frame, so the loops are oversampled. Two clamp knobs cap the counts at the single source (sampleAtmosphereSunLight / sampleAtmosphereMoonLight), so primary and the half-rate secondary paths both inherit. Defaults 0 = legacy uncapped (bit-identical); set 1 for the perf win after visual validation.

src/dxvk/shaders/rtx/pass/atmosphere/atmosphere_args.h — fork-owned change. Repurposes padMoonNee0 / padMoonNee1 as uint sunShadowMaxSamples / uint moonShadowMaxSamples. No layout change.
src/dxvk/shaders/rtx/pass/atmosphere/atmosphere_common.slangh — fork-owned additions. Clamp applied after getSunSoftShadowParams in sampleAtmosphereSunLight and after the constant-4 assignment in sampleAtmosphereMoonLight; 0 = uncapped.
src/dxvk/rtx_render/rtx_options.h — fork-owned additions. RTX_OPTION("rtx.atmosphere", int, sunShadowMaxSamples, 0, …) / moonShadowMaxSamples.
src/dxvk/rtx_render/rtx_atmosphere.cpp — fork-owned change. getAtmosphereArgs populates the two fields (replacing the pad zero-writes).
src/dxvk/rtx_render/rtx_fork_atmosphere.cpp — fork-owned addition. "Sun Shadow Ray Cap" / "Moon Shadow Ray Cap" DragInts in the Atmosphere → Advanced ImGui tree.

Workstream — Sky perf: shader-path bisect toggles (fork — 2026-06-11, diagnostic)

Extends the per-dispatch bisect kit into the integrators: with the dispatch costs known (~1.8 ms ≈ cloud budget) and the NEE rays capped, the residual skyMode delta splits across two shader paths that host-side toggles can't reach. Two default-true diagnostic options pack into one CB gate word (debugSkyBisectFlags, former padMoonNee2 slot — layout unchanged): bit 0 skips atmosphere sun+moon NEE at both integrator call sites (sun/moon lighting goes black), bit 1 makes evalSkyRadiance return flat grey immediately (isolates the full per-ray miss-path cost; the skipped temporal-history write self-heals on re-enable via the frame-id age check). Defaults = bits clear = production paths, bit-identical.

src/dxvk/shaders/rtx/pass/atmosphere/atmosphere_args.h — fork-owned change. Repurposes padMoonNee2 as uint debugSkyBisectFlags. No layout change.
src/dxvk/shaders/rtx/pass/atmosphere/atmosphere_sky.slangh — fork-owned addition. Flat-grey early return at the top of evalSkyRadiance behind bit 1.
src/dxvk/shaders/rtx/algorithm/integrator_direct.slangh / integrator_indirect.slangh — fork-owned change (upstream files, one-line condition edits). The cb.skyMode == 1 atmosphere-NEE gates additionally require bit 0 clear.
src/dxvk/rtx_render/rtx_options.h — fork-owned additions. debugEnableAtmosphereNee / debugEnableSkyMissShading, both default true.
src/dxvk/rtx_render/rtx_atmosphere.cpp — fork-owned change. getAtmosphereArgs packs the two options into the flag word.
src/dxvk/rtx_render/rtx_fork_atmosphere.cpp — fork-owned addition. "Atmosphere Sun/Moon NEE" / "Full Sky Miss Shading" checkboxes in the Perf Bisect (Diagnostic) ImGui tree.

Follow-up (same day): debugDispatchSkyLuts (rtx_options.h, default true) freezes the whole LUT bake cascade — the last untoggleable GPU piece of the atmosphere; with a moving time-of-day sun the sky-view LUT re-bakes every frame by design, and this measures that. computeLuts short-circuits the gate block when off (rtx_atmosphere.cpp); "Sky LUT Bakes" checkbox added to the Perf Bisect tree (rtx_fork_atmosphere.cpp).

Workstream — Sky perf: sky-view re-bake granularity (fork — 2026-06-11)

Production fix for the per-frame sky-view re-bake the frozen-cascade bisect identified (objective frame-time improvement, no visual hit). With a continuously-animating time-of-day sun, the sky-view cache key sees a new sunDirection every frame. skyViewRebakeGranularityDeg (default 0 = legacy every-frame behavior; recommended 0.1) quantizes the sun and moon directions INSIDE the sky-view cache key, so continuous motion flips the memcmp only when a direction crosses a granularity step — the bake itself always uses exact current values, so each re-bake is exact and staleness between re-bakes is bounded by the step angle. All non-direction fields stay exact (sliders / presets re-bake immediately). Applies to the split-key path (skyLutCacheKeySplitEnable, on by default).

src/dxvk/rtx_render/rtx_options.h — fork-owned addition. RTX_OPTION("rtx.atmosphere", float, skyViewRebakeGranularityDeg, 0.0f, …).
src/dxvk/rtx_render/rtx_atmosphere.cpp — fork-owned addition. quantizeDirComponent helper; normalizeForSkyViewLutKey quantizes sunDirection + all moon directions by the option's step (no-op at 0). The transmittance/MS key is unaffected (it zeroes those fields outright).
src/dxvk/rtx_render/rtx_fork_atmosphere.cpp — fork-owned addition. "Sky Re-bake Granularity" DragFloat in the Atmosphere → Advanced ImGui tree.

Workstream — Sky perf: pin validated defaults + retire workstream UI (fork — 2026-06-11)

Closing commit of the sky perf workstream. In-game validation confirmed the two production levers (sun/moon NEE shadow-ray caps at 1, sky-view re-bake granularity at 0.1°) deliver the measured wins with no visual hit, so their defaults are pinned and the workstream's ImGui surface is retired per the user's call ("this is in a good enough spot now"). All options remain conf-tunable; only the UI is removed.

src/dxvk/rtx_render/rtx_options.h — fork-owned change. Defaults: sunShadowMaxSamples 0 → 1, moonShadowMaxSamples 0 → 1, skyViewRebakeGranularityDeg 0.0 → 0.1. Doc strings note the validation; 0 still selects legacy behavior via conf.
src/dxvk/rtx_render/rtx_fork_atmosphere.cpp — fork-owned removal. Drops from the UI: the "Minimal Sky LUT Re-bakes" checkbox, the "Sun/Moon Shadow Ray Cap" DragInts, the "Sky Re-bake Granularity" DragFloat (Atmosphere → Advanced), and the entire "Perf Bisect (Diagnostic)" tree. The underlying options (skyLutCacheKeySplitEnable, the caps, the granularity, and the six debug* bisect toggles) stay declared and conf-tunable for future regression hunting.

Workstream — Cloud perf: voxel-grid re-bake granularity (fork — 2026-06-11)

Post-workstream follow-up at the user's request: the perf-bisect freeze of the per-frame D_sun / D_ambient bakes showed a large win with no short-term visual change — the staleness only accumulates as wind / camera / sun move. Same pattern as the sky-view fix: a cache key (normalizeForVoxelGridKey) quantizes the grid bake's motion inputs — wind scroll + camera position by cloudVoxelGridRebakeGranularityKm, sun/moon directions by the existing skyViewRebakeGranularityDeg — so the grids re-bake once per step of actual motion instead of every frame, staleness bounded by the step. Cloud parameter changes stay exact in the key (same-frame re-bake) and a cloud-noise re-bake force-clears the cached key. In-game validated at 0.1 km (default): ~0.7 ms saved, no visible stepping in cloud lighting or terrain shadows; 0 = legacy every-frame via conf. Note this lever lives on the cloud side — full-rate bakes were the deliberate 2026-05-19 decision against the stuttery fixed-cadence 8-frame round-robin; motion-quantized bounded-error steps passed where fixed cadence failed. Conf-only like the other workstream knobs (no UI).

src/dxvk/rtx_render/rtx_options.h — fork-owned addition. RTX_OPTION("rtx.atmosphere", float, cloudVoxelGridRebakeGranularityKm, 0.1f, …).
src/dxvk/rtx_render/rtx_atmosphere.h / rtx_atmosphere.cpp — fork-owned additions. normalizeForVoxelGridKey (sky-view key base + wind/camera re-injected quantized), m_cachedVoxelGridKey, the dirty gate around the two grid dispatches in computeLuts, and the force-clear on cloud-noise re-bake.

Workstream — Per-column cloud model (column-shaping rework) (fork — 2026-06-12)

Root-cause rework of the "stacked separated layers" read that made the volumetric cloud system look like flat 2D planes. Previously EVERY vertical shaping signal — the height-LUT density envelope (which at cumulus types even baked TWO density peaks separated by a thinner neck), the coverage-threshold scale band, the anvil pow gate, the Nubis dim profile and the bottom-darkening gradient — keyed on the GLOBAL slab height fraction, so every cloud in the sky shared one vertical recipe pinned to absolute altitude (flat global base/top planes, horizontal lighting bands), while the thresholded isotropic 3D noise placed mass independently per altitude (vertically disconnected stacked puffs within a column). The rework introduces a baked 512×512 RGBA8 cloud placement map (R = Worley-cell cluster field, G = per-cloud top-height jitter, B = base lift; tiled at cloudNoiseTileKm, de-tiled by the SAME hex lattice as the 3D noise so cluster and texture travel together) and derives per-column cloud presence + local base/top from it; the sample height is re-normalized to each column's own [base, top] and ALL vertical shaping plus the Nubis lighting proxies key on that per-cloud height. The 3D noise keeps its texture/erosion role but no longer decides placement. Gated by cloudColumnShapingEnable (default on; legacy global-slab path preserved bit-exact for A/B).

src/dxvk/shaders/rtx/pass/atmosphere/cloud_placement_map_baker.comp.slang — NEW fork-owned file. Periodic 2D Worley + value-noise FBM construction of the cluster / top-jitter / base-lift fields; integer cells-per-tile snapping from cloudCellSizeKm so the map tiles seamlessly.
src/dxvk/shaders/rtx/pass/atmosphere/atmosphere_common.slangh — fork-owned additions + changes. New kCloudPlacementFieldMean, sampleCloudPlacement (hex-de-tiled tap; R variance-preserving blend, G/B weighted average), CloudColumn / computeCloudColumn (coverage remap of the cluster field over a cloudColumnFeather band; base lift; presence^cloudColumnTopShape × jitter top). Both density samplers hoist the hex tile, sample placement, early-out outside the column (before any 3D taps — a perf win at partial coverage), re-normalize the height fraction to the column span, and gate on per-column presence instead of weather coverage; the primary overload returns the per-cloud shapeHeightFraction via a new out param. Both samplers + both voxel-bake optical-depth helpers gain a Texture2D<float4> placementMap parameter.
src/dxvk/shaders/rtx/pass/atmosphere/cloud_march_common.slangh — fork-owned change. marchCloudSlab consumes the out-param overload and passes the per-cloud shapeHf to evalNubisCubedSample, so the dim profile / SDF proxy / bottom darkening track each cloud's own base/top instead of painting global altitude bands. The moon-shadow tap plumbs the placement map.
src/dxvk/shaders/rtx/pass/atmosphere/cloud_height_lut_baker.comp.slang — fork-owned change. Gains the args CB (binding 0; output moves to 1) and a second curve family for column mode: single-lobe envelope (smoothstep rise, fall reaching exactly 0 at hf=1 for cumulus) replacing the legacy trapezoid + secondary anvil bump — the two-peak profile was itself a baked-in stacked-strata structure. Legacy family preserved verbatim for the gate-off path; G channel (threshold widening near each cloud's own top) unchanged and now carries the anvil look alone in column mode.
src/dxvk/shaders/rtx/pass/atmosphere/atmosphere_sky.slangh — fork-owned change. evalClouds + sampleCloudSunOpticalDepth pass AtmosphereCloudPlacementMap (from common bindings) into the density sampler — the analytical fallback gets the column model automatically.
src/dxvk/shaders/rtx/pass/atmosphere/cloud_render.comp.slang / cloud_secondary_lut.comp.slang — fork-owned additions. Declare AtmosphereCloudPlacementMap at local binding 12 (sampled with the slot-2 cloud-noise sampler).
src/dxvk/shaders/rtx/pass/atmosphere/cloud_sun_density_grid.comp.slang / cloud_ambient_density_grid.comp.slang — fork-owned additions. Declare the placement map at local binding 4 and pass it through the optical-depth helpers so the baked D_sun / D_ambient grids shadow the same column shapes the view march renders.
src/dxvk/shaders/rtx/pass/atmosphere/atmosphere_args.h — fork-owned change. Repurposes padCloudLook2 → cloudColumnShapingEnable, padCloudC2 → cloudCellSizeKm, pad_cr0..2 → cloudColumnTopVariation / cloudColumnTopShape / cloudColumnBaseVariation, pad_c6_0 → cloudColumnFeather. No CB layout change.
src/dxvk/shaders/rtx/pass/common_binding_indices.h — fork-owned addition (index-only). Adds BINDING_ATMOSPHERE_CLOUD_PLACEMENT_MAP = 216 + a TEXTURE2D entry in COMMON_RAYTRACING_BINDINGS.
src/dxvk/shaders/rtx/pass/common_bindings.slangh — fork-owned addition (index-only). Declares AtmosphereCloudPlacementMap at the new slot for the raytrace TUs (analytical evalClouds fallback).
src/dxvk/rtx_render/rtx_options.h — fork-owned additions. cloudColumnShapingEnable (true), cloudCellSizeKm (2.0), cloudColumnTopVariation (0.45), cloudColumnTopShape (0.6), cloudColumnBaseVariation (0.12), cloudColumnFeather (0.35).
src/dxvk/rtx_render/rtx_atmosphere.h / rtx_atmosphere.cpp — fork-owned additions + changes. m_cloudPlacementMap (512² RGBA8) + kCloudPlacementMapSize, dispatchCloudPlacementMapBake (init + live re-bake via needsCloudPlacementRebake on cloudCellSizeKm / cloudNoiseTileKm), height-LUT re-bake on cloudColumnShapingEnable flip (m_cachedHeightLutColumnMode; the baker now binds the args CB at 0 / output at 1), voxel-grid key force-clear when either shape input re-bakes, placement binds: slot 12 in dispatchCloudRender / dispatchCloudSecondaryLut, slot 4 in both voxel-grid bakes, BINDING_ATMOSPHERE_CLOUD_PLACEMENT_MAP in bindResources, and getAtmosphereArgs fills for the six repurposed fields. Shader-class parameter lists updated accordingly.
src/dxvk/rtx_render/rtx_fork_atmosphere.cpp — fork-owned additions. bindAtmosphereLuts binds the placement map at the common slot; new "Cloud Columns" ImGui group (Volumetric Cloud Columns checkbox + Cloud Cell Size / Top Variation / Top Shape / Base Undulation / Edge Feather sliders); Vertical Stretch tooltip notes columns supersede it.

Workstream — Per-column underside contrast (column-shaping rev 2) (fork — 2026-06-12)

First in-game pass on the column model reported a residual "layered" read on the overcast underside. Diagnosis: the deck base's dominant light is the Nubis multi-scatter M term, which is horizontally near-uniform at production settings (sigma_ms is tiny so exp(-sigma_ms·D_sun) ≈ 1; the bottom-darkening gradient is a pure function of height) — so the underside renders as one flat-lit sheet with only fine erosion lace on top: a dark lace plane over a bright plane reads as two stacked 2D layers. Real deck undersides carry km-scale cellular relief because base brightness follows per-column water depth (thick columns occlude downwelling light, thin columns transmit it). The fix scales the bottom-darkening amount by each sample's column span (already derived by the column model), so the underside brightness is driven by the same placement field that shapes the clouds — moving with wind, matching the baked shadows.

src/dxvk/shaders/rtx/pass/atmosphere/atmosphere_common.slangh — fork-owned changes. sampleCloudDensityTextured's primary overload gains an out float columnSpan (topFrac − baseFrac; 1.0 legacy); evalNubisCubedSample gains a columnSpan parameter and scales cloudBottomDarkening by mix(1, span, cloudColumnUndersideContrast) before building the bottom gradient. Contrast 0 or span 1 reproduce the uniform legacy darkening exactly.
src/dxvk/shaders/rtx/pass/atmosphere/cloud_march_common.slangh — fork-owned change. marchCloudSlab threads the span from the density sampler into the lighting call.
src/dxvk/shaders/rtx/pass/atmosphere/atmosphere_args.h — fork-owned change. Repurposes pad_c6_1 → cloudColumnUndersideContrast. No CB layout change.
src/dxvk/rtx_render/rtx_options.h — fork-owned addition. RTX_OPTION("rtx.atmosphere", float, cloudColumnUndersideContrast, 0.65f, …).
src/dxvk/rtx_render/rtx_atmosphere.cpp — fork-owned change. getAtmosphereArgs fills the new field (replacing the pad zero-write).
src/dxvk/rtx_render/rtx_fork_atmosphere.cpp — fork-owned addition. "Underside Contrast" DragFloat in the Cloud Columns ImGui group.

(Superseded the same day by rev 3 below — the flat per-column scaling produced flat dark smudges in-game; the option and its mechanism were replaced before any release.)

Workstream — Analytic downwelling light field (column-shaping rev 3) (fork — 2026-06-12)

Second in-game pass: rev 2's span-scaled darkening produced big FLAT dark smudges — user verdict "worse… changed the point at which this becomes visible, never fixing the root cause." Root cause, finally named precisely: the underside LIGHT FIELD has no 3D structure. Under a deck, T_primary ≈ 0; the dominant M term varies ≤~20-40% (sigma_ms 0.05–0.25 over D_sun ≈ 2) and that residual comes from the 8-tap voxel-grid D_sun (km-scale mush); ambient's exp(-D_ambient) saturates to ~0. With per-point illumination near-constant, the only visible pattern is opacity silhouetted against a flat backdrop — every density/shaping change just swaps the wallpaper. Rev 3 computes the downwelling light analytically from the column model: the height LUT gains a B channel holding the cumulative envelope integral from each per-cloud height to the cloud top, so downTau = ∫envelope × columnSpan × cloudThickness × cloudDensity is the exact macroscopic water above any sample at full resolution (no voxel grid, no taps); exp(-cloudUndersideLightSigma × downTau) (a diffusion-flavored sigma, far below beam extinction) lights the multi-scatter and ambient terms. Underside brightness now varies continuously with the actual water distribution — dark cores, bright thin spots, in-cell gradients. Legacy mode (column shaping off, or sigma 0) keeps the constant bottom-darkening gradient bit-for-bit.

src/dxvk/shaders/rtx/pass/atmosphere/cloud_height_lut_baker.comp.slang — fork-owned change. Output RG8 → RGBA8; B = midpoint-rule ∫_hf^1 envelope(u) du of the active curve family (re-baked on mode flip as before).
src/dxvk/shaders/rtx/pass/atmosphere/atmosphere_common.slangh — fork-owned changes. New cloudTypeProfileIntegralFromTop (closed-form trapezoid integral, the no-LUT fallback) + cloudHeightProfileDownIntegral (LUT B channel / fallback). evalNubisCubedSample: rev 2's spanFactor block replaced by the verticalLight selector — column mode = analytic Beer-Lambert on downTau, legacy = the constant gradient; M multiplies verticalLight; ambient = ambient_shape × verticalLight in column mode (replacing the saturating exp(-D_ambient) tap — same quantity, better estimate; multiplying both would double-count).
src/dxvk/shaders/rtx/pass/atmosphere/cloud_render.comp.slang / cloud_secondary_lut.comp.slang — fork-owned change. AtmosphereCloudHeightLut declarations Texture2D<float2> → Texture2D<float4>.
src/dxvk/shaders/rtx/pass/atmosphere/atmosphere_args.h — fork-owned change. cloudColumnUndersideContrast (rev 2, pad_c6_1) → cloudUndersideLightSigma. No CB layout change.
src/dxvk/rtx_render/rtx_options.h — fork-owned change. Drops cloudColumnUndersideContrast; adds RTX_OPTION("rtx.atmosphere", float, cloudUndersideLightSigma, 0.12f, …).
src/dxvk/rtx_render/rtx_atmosphere.cpp — fork-owned changes. Height LUT resource format R8G8 → R8G8B8A8; args fill swaps to the new option.
src/dxvk/rtx_render/rtx_fork_atmosphere.cpp — fork-owned change. "Underside Contrast" slider replaced by "Underside Shading" (sigma). Tooltip notes it supersedes Bottom Darkening while Cloud Columns are on.

Workstream — Adaptive cloud-march sampling (fork — 2026-06-12)

Companion fix shipped with column-shaping rev 3, prompted by a community diagram suggesting the march be confined to the shell volume — it already is (marchCloudSlab intersects base + top shells), but the adjacent real defect is the FIXED step count across that volume: 32 uniform steps over a span that varies from ~4 km (zenith) to 50+ km (horizon-grazing through the curved shell) puts ~1.6 km steps against ~2 km cloud features at low elevations. The aliasing, averaged by jitter + the temporal EMA, renders as soft horizontal BANDS concentrated toward the horizon — a direct contributor to the "stacked layers" read. The march now holds a target step LENGTH: count = span / cloudViewStepKm, floored at cloudViewSamples (zenith cost unchanged) and capped at cloudViewSamplesMax; the column model's pre-tap early-outs keep added steps cheap where the slab is empty. cloudViewStepKm = 0 restores the legacy fixed count exactly.

src/dxvk/shaders/rtx/pass/atmosphere/cloud_march_common.slangh — fork-owned change. marchCloudSlab computes effSampleCount from the clamped span and the step target; loop + rayFrac use it. Applies to both consumers (cloud render RT + secondary dome LUT).
src/dxvk/shaders/rtx/pass/atmosphere/atmosphere_args.h — fork-owned change. Repurposes pad_cloudSunsetAmbient0 → cloudViewStepKm and padStarCloud0 → cloudViewSamplesMax. No CB layout change.
src/dxvk/rtx_render/rtx_options.h — fork-owned additions. cloudViewStepKm (0.3) + cloudViewSamplesMax (128).
src/dxvk/rtx_render/rtx_atmosphere.cpp — fork-owned change. getAtmosphereArgs fills both (replacing the pad zero-writes).
src/dxvk/rtx_render/rtx_fork_atmosphere.cpp — fork-owned addition. "March Step Size" DragFloat next to Cloud Render Scale; the cap stays conf-only.
Clarity pass (same day, post-validation): the in-game fix was user-validated with an accepted perf cost, so the controls now state it plainly: slider renamed "Cloud Sample Spacing" with a PERFORMANCE paragraph in the tooltip (cost scales up to cap/32 ≈ 4x on horizon-heavy views at defaults; how to trade it back), and the cap is exposed as a "Max Cloud Samples" DragInt (the perf governor). Option docstrings updated to match.

Workstream — Night cloud-lighting knob rebase (fork — 2026-06-14)

Night clouds read too bright. The earlier suspicion (the moon-cloud directional term) was wrong: the dominant contributor is the STAR-coupling term, added uniformly and uncolored to every cloud sample (nightLight += nightSkyColor × starBrightness × starAmbientCouplingStrength × nightFactor), so it lifts the whole deck — and tints it blue via nightSkyColor — across the entire sky, not just near the moon. At the old default (0.03 × starBrightness 0.5 × nightSkyColor) it produced ~0.006 blue, larger and more pervasive than the moon's ~0.003 silver-lining peak. The knobs were also mis-scaled: the sub-0.01 night-radiance smallness lived inside the user-facing gains, so the only sane values were ~0.001 ("having 0.001 be the only reasonable one is painful"). Fix = rebase to O(1): the smallness moves into internal shader constants so the knobs read as a "multiple of the calibrated night level," and the star default drops to a user-tested 0.25. moonAmbientAirglow is unchanged in effect (it was never the problem). Both the RT march and the legacy analytical evalClouds get the identical edit so the day→night crossfade stays bit-matched.

src/dxvk/shaders/rtx/pass/atmosphere/cloud_march_common.slangh — fork-owned change. buildCloudShadeContext nightLight: internal kStarCloudCoupling = 0.008f and kMoonAirglowScale = 0.0015f factored out of the star-coupling and moon-airglow terms so the user-facing gains become O(1).
src/dxvk/shaders/rtx/pass/atmosphere/atmosphere_sky.slangh — fork-owned change. Identical kStarCloudCoupling / kMoonAirglowScale factoring in the legacy evalClouds nightLight block (crossfade parity with the RT path).
src/dxvk/rtx_render/rtx_options.h — fork-owned change. starAmbientCouplingStrength default 0.03 → 0.25; moonAmbientAirglow default 0.0015 → 1.0; both docstrings rewritten as "multiple of the calibrated night level."
src/dxvk/rtx_render/rtx_fork_atmosphere.cpp — fork-owned change. "Star Ambient Coupling" DragFloat range 0–0.1 / %.4f → 0–3 / %.2f so the O(1) default isn't clamped; tooltip updated.

Workstream — Cloud edge detail via threshold modulation (rev 4) (fork — 2026-06-14)

The additive edge-detail (rev 3) was trapped in a thin blobby rind hugging the silhouette: its containment window (smoothstep(coverageThreshold - 0.15, coverageThreshold, density)) keyed on the pre-detail base density and was one-sided, so detail could only displace the iso-surface from INSIDE — it could never grow billows outward. Widening the window traded the rind for a half-res smear halo (the additive positive lobe lifts a broad sub-threshold clear-sky margin into faint density on soft edges, which the half-res render + bilinear upsample smear). Rev 4 changes the mechanism: the detail tap now wobbles the COVERAGE THRESHOLD (the gate position) instead of the density field, zero-mean about kCloudNoiseFieldMean (0.4). This beats both modes at once — (a) the threshold can drop below the base value so billows grow OUTWARD past the silhouette (no rind), and (b) the gate self-localizes the effect so far-outside samples stay hard-0 (no faint margin to smear) and far-inside stay 1 (no interior holes); the modulation only bites in the gate transition band, so no explicit edge window is needed. Unlike the rejected rev-2 erosion remap it never touches the noise field — it only moves the gate, the same slide-3 anvil lever step 4a uses at low frequency. User-validated in-game ("this is perfect"). Supersedes the prior rind-vs-smear tradeoff; rev 3's window and the rejected tunable-reach knob are both gone.

src/dxvk/shaders/rtx/pass/atmosphere/atmosphere_common.slangh — fork-owned change. sampleCloudDensityTextured step 4b: additive density += + edgeWindow replaced by coverageThreshold -= cloudDetailStrength * 0.5 * (detailNoise - kCloudNoiseFieldMean), clamped ≥ 0. Internal kEdgeDetailThreshold = 0.5 (wobble→threshold scale). No new options; cloudDetailStrength = 0 is bit-identical. Shadow sampler still skips detail.

Workstream — Artistic sunset color controls (fork — 2026-06-14)

Two artistic knobs to recover sunset warmth/saturation lost when commit 3e37062b moved sunset reddening onto the physical Hillaire two-term LUT model: the broadband multiscatter "fill" reads pale-blue and desaturates the warm single-scatter, so the physically-correct sunset renders undersaturated. Both apply inside evalAtmosphereRadiance (the sky-view LUT bake integral), so the baked LUT carries them and clouds inherit the warmer ambient for free (cloud warm ambient samples the sky-view LUT). Both default to 1.0 = physical (no change). sunsetSaturation ramps in only near the horizon (off above ~24°, sin 0.4) so midday sky is untouched; multiScatterStrength is a global scale on the fill term. This is artistic control on top of the physical model — NOT a revert to the pre-3e37062b analytical air-mass reddening.

src/dxvk/shaders/rtx/pass/atmosphere/atmosphere_common.slangh — fork-owned additions. In evalAtmosphereRadiance's per-sample loop: multiScatterContrib *= args.multiScatterStrength after the analytical/LUT blend. Before return L: a luma-preserving saturation boost lerp(vec3(luma), L, satGain) where satGain = lerp(1, args.sunsetSaturation, 1 - smoothstep(0, 0.4, sunDir.y)). Both no-ops at the 1.0 defaults.
src/dxvk/shaders/rtx/pass/atmosphere/atmosphere_args.h — fork-owned additions. Repurposes the trailing pad_cloudEdge1 slot as multiScatterStrength (completes the cloud-edge 16-byte row, layout unchanged), then appends a new 16-byte row sunsetSaturation + pad_artistic0..2. CB grows by one row; all prior field offsets unchanged.
src/dxvk/rtx_render/rtx_options.h — fork-owned additions. Adds multiScatterStrength (1.0) and sunsetSaturation (1.0) RTX_OPTIONs to the rtx.atmosphere cluster, immediately after multiScatterPhysicalStrength.
src/dxvk/rtx_render/rtx_atmosphere.cpp — fork-owned additions. getAtmosphereArgs(): sets both args from RtxOptions unconditionally (after multiScatterPhysicalStrength), so the sky reddens with clouds disabled. normalizeForTransmittanceMsKey(): zeroes both (like multiScatterPhysicalStrength) — they feed only the sky-view bake, not the transmittance/MS LUTs, so changing them must not re-bake the heavy pair.
src/dxvk/rtx_render/rtx_fork_atmosphere.cpp — fork-owned additions. Adds "Multiscatter Strength" (0–2) and "Sunset Saturation" (0–3) DragFloats to the Atmosphere → Advanced ImGui tree, immediately after "Multiscatter Physical Strength", each with an explanatory tooltip. ~10 LOC.

Workstream — Sun-only cloud-on-terrain shadow re-architecture (fork — 2026-06-19)

Moves the cloud-on-terrain shadow from a post-denoise screen-space texture (multiplied onto the combined direct buffer in composite) to a pre-denoise fold-in on the SUN's radiance inside the sun NEE. This deletes the whole geometry-blindness compensation stack (sealed-interior zenith up-ray gate, viewmodel/decal origin correction, camera-side normal flip, dusk/dawn horizon gate) and the screen-space plumbing it required. Several of the removed screen-space pieces (the PrimaryCloudShadowFactor resource + its binding indices/tables, debug view 878) were never previously inventoried — a prior fridge-list gap, now closed by removal.

src/dxvk/shaders/rtx/pass/atmosphere/atmosphere_common.slangh — fork-owned change. AtmosphereSunSample: drops the cloudShadowFactor member. sampleAtmosphereSunLight: folds the cloud shadow onto the sun by result.radiance *= pow(sampleCloudGroundShadow_OptionB(...), args.cloudShadowFactorStrength) (gated on cloudVoxelShadowsEnable); the early issue-#37 factor computation and the long screen-space wire-in comment are gone. sampleAtmosphereSunLightVolume: same pow(..., cloudShadowFactorStrength) curve added for parity (was a plain multiply). sampleCloudGroundShadow_OptionB_impl: horizon gate removed (the grazing-sun OD blowup now only darkens the already-attenuated sun, never lamps).
src/dxvk/shaders/rtx/algorithm/integrator_direct.slangh — fork-owned change. evalAtmosphereSunNEE: deletes the sealed-interior zenith up-ray gate, the viewmodel/decal camera-origin branches, the triangle-normal flip, and the PrimaryCloudShadowFactor[pixelCoordinate] write. The cloud shadow now arrives folded into sunSample.radiance.
src/dxvk/shaders/rtx/pass/atmosphere/atmosphere_args.h — fork-owned change. Repurposes the trailing pad_artistic0 slot as cloudShadowFactorStrength (CB layout ABI-unchanged); the knob moved here from composite_args.h.
src/dxvk/rtx_render/rtx_atmosphere.cpp — fork-owned change. getAtmosphereArgs(): populates args.cloudShadowFactorStrength = max(RtxOptions::cloudShadowFactorStrength(), 0) in the cloud-shadow block.
src/dxvk/shaders/rtx/pass/composite/composite.comp.slang — fork-owned change. Removes the post-denoise pow(PrimaryCloudShadowFactor, cloudShadowFactorStrength) multiply on primary direct radiance (and the texelFetch). Only an architecture comment remains.
src/dxvk/shaders/rtx/pass/composite/composite_args.h + src/dxvk/rtx_render/rtx_composite.cpp — fork-owned change. cloudShadowFactorStrength CB slot renamed to reserved pad1 (ABI-stable); the composite populate of it is removed.
src/dxvk/shaders/rtx/pass/composite/composite_bindings.slangh + composite_binding_indices.h + rtx_composite.cpp — index-only, fork. Removes the PrimaryCloudShadowFactor Texture2D declaration, the two TEXTURE2D(COMPOSITE_PRIMARY_CLOUD_SHADOW_FACTOR_INPUT) parameter-list entries, and the bindResourceView. Binding slot 19 left reserved (number not reused).
src/dxvk/shaders/rtx/pass/integrate/integrate_direct.slangh + integrate_direct_bindings.slangh + integrate_direct_binding_indices.h + rtx_pathtracer_integrate_direct.cpp — index-only, fork. Removes the per-frame PrimaryCloudShadowFactor clear, the RWTexture2D<float> PrimaryCloudShadowFactor declaration, the RW_TEXTURE2D parameter-list entry, and the bindResourceView. Binding slot 73 left reserved.
src/dxvk/rtx_render/rtx_resources.cpp + rtx_resources.h — index-only, fork. Removes the m_primaryCloudShadowFactor Resource member and its createImageResource alloc.
src/dxvk/shaders/rtx/utility/debug_view_indices.h + debug_view.comp.slang + debug_view_binding_indices.h + rtx_debug_view.cpp — index-only, fork. Removes debug view 878 (DEBUG_VIEW_CLOUD_SHADOW_FACTOR_RAW): the enum define (number burned, not reused), the case arm, the DebugViewPrimaryCloudShadowFactor declaration, the binding-index define (slot 38 reserved), the TEXTURE2D parameter-list entry, the bindResourceView, and the selector-list label. Views 875/877 (D_sun grid reads) stay.
src/dxvk/rtx_render/rtx_instance_manager.cpp + rtx_materials.h + src/dxvk/shaders/rtx/concept/surface/surface.h — index-only, fork (REVERTED). Reverts the Surface::isDecalCategory flag (CPU set, flags0 bit 2 pack, shader property) added 2026-06-18; bit 2 is free again. Only "removed/freed" comments remain.

Workstream — Cloud ImGui simplification (fork — 2026-06-15)

Restructured the flat ~45-control Clouds menu (7 non-collapsible TextDisabled dividers, everything always expanded) into a workflow-grouped tree of collapsible sub-nodes. No options, CB fields, or shader behavior change — pure ImGui reorganization.

src/dxvk/rtx_render/rtx_fork_atmosphere.cpp — fork-owned change. Clouds node rebuilt as: Basic (open by default: Coverage, Cloud Type, Density, Altitude, Depth, Color) · Shaping ▸ {Variation, Detail & Edges, Columns} · Lighting · Wind · Layer 2 · Performance · Horizon & Haze (Curvature moved here from "Look"). The Color control is a ColorEdit3 swatch/picker. Mode-inert controls are greyed via ImGui::BeginDisabled: the six Columns sliders when cloudColumnShapingEnable is off, Bottom Darkening when it is on (Underside Shading supersedes it), and the Layer 2 body when cloudLayer2Enable is off. The "Vertical Stretch" slider (cloudVerticalStretch) is removed from the menu — the option stays conf-only (experimental / superseded by Columns). All tooltips preserved.

Workstream — Remove legacy cloud paths + realistic sun-gated underside darkening (fork — 2026-06-19)

Two dead cloud code paths were removed and the column model made the only path: (1) the legacy analytical evalClouds view-march in atmosphere_sky.slangh (already compile-gated out behind ATMOSPHERE_CLOUD_VIEW_MARCH, used as a runtime A/B fallback that was never the validated default — clouds come from the cloud_render compute pass on primary rays and the secondary dome LUT on indirect/PSR/reflection rays), and (2) the legacy global-slab shaping (cloudColumnShapingEnable == false): the trapezoid+anvil height-profile curve family, the gate-off branches in the density/bake samplers, and the constant bottom-darkening gradient. With the legacy gradient gone, bottom darkening was reworked into the column model as a physically-motivated underside light field driven by cloud density and sun position: the analytic downwelling exp(-cloudUndersideLightSigma × downTau) (density / overlying-water term, unchanged shape) now scales by cloudBottomDarkening (overall strength) × smoothstep(0, 0.35, sunElev) so the darkening is strongest at high sun and fades out toward the horizon — the low sun rakes under the deck and the warm ambient / silver-lining terms light the bases (classic lit-orange sunset undersides). cloudBottomDarkening default 0.55 → 1.0 so the high-sun look matches the validated rev-3 column underside bit-for-bit; only low-sun behavior is new. The legacy ambient exp(-D_ambient) tap is gone (the analytic verticalLight replaces it), so sampleDAmbient is no longer called from evalNubisCubedSample.

src/dxvk/shaders/rtx/pass/atmosphere/atmosphere_sky.slangh — fork-owned change. Deleted evalClouds (the whole #ifdef ATMOSPHERE_CLOUD_VIEW_MARCH block) and its only consumer sampleCloudSunOpticalDepth. The cloud-source else branch in evalSkyRadiance (both RT + secondary-LUT off) is now an unconditional transparent vec4(0) fallback, no #ifdef. Stale evalClouds comment references updated.
src/dxvk/shaders/rtx/pass/atmosphere/cloud_height_lut_baker.comp.slang — fork-owned change. Deleted the legacy densityEnvelope (trapezoid+anvil) R-curve and the cloudColumnShapingEnable mode select; densityEnvelopeColumn is now used unconditionally for both R and the B integral. coverageThresholdScale (G) unchanged.
src/dxvk/shaders/rtx/pass/atmosphere/atmosphere_common.slangh — fork-owned changes. Dropped the cloudColumnShapingEnable gate in sampleCloudDensityTextured and the bake-mirror sampler (column derivation now unconditional). evalNubisCubedSample: verticalLight reworked from the column/legacy if-else to the single realistic sun-gated form ( $\text{mix}(1, \text{exp}(-\text{sigma}·\text{downTau}), \text{cloudBottomDarkening} \times \text{smoothstep}(0, \text{kUndersideSunFadeElev}, \text{sunElev}))$ , kUndersideSunFadeElev = 0.35); ambient term is now ambient_shape × verticalLight unconditionally; sampleDAmbient call removed (unused).
src/dxvk/shaders/rtx/pass/atmosphere/atmosphere_args.h — fork-owned change. cloudColumnShapingEnable → padCloudLook2 and cloudBottomDarkeningHeight → pad_cloudVoxel1 (both reverted to pad placeholders). No CB layout change.
src/dxvk/rtx_render/rtx_options.h — fork-owned change. Removed cloudColumnShapingEnable and cloudBottomDarkeningHeight RTX_OPTIONs. cloudBottomDarkening default 0.55 → 1.0, description reworked (sun-gated strength). cloudUndersideLightSigma description reworked (no more legacy-gradient / column-mode wording).
src/dxvk/rtx_render/rtx_atmosphere.cpp / .h — fork-owned changes. Removed the m_cachedHeightLutColumnMode member + the height-LUT flip-rebake in computeLuts (the LUT bakes one curve family once at init). Dropped the cloudColumnShapingEnable / cloudBottomDarkeningHeight args fills. Updated the bake doc comments.
src/dxvk/rtx_render/rtx_fork_atmosphere.cpp — fork-owned change. Removed the "Volumetric Cloud Columns" checkbox and the BeginDisabled/EndDisabled gating around the Columns sliders and Bottom Darkening (all always enabled now). Reworked the Bottom Darkening + Underside Shading tooltips (strength × shape; sunset fade).
RtxOptions.md — hand-edited to match (removed the two options, updated cloudBottomDarkening / cloudUndersideLightSigma); regenerate in-app for canonical form.

Workstream — Cloud sky-dome ambient fill (clouds reflect the sky) (fork — 2026-06-19)

First half of the sky↔cloud color-coupling work. Symptom: under a bright blue daytime sky the cloud undersides read dark / gloomy, and clouds never take on the surrounding sky's color, because the cloud ambient samples the sky-view LUT in only TWO directions — toward the sun (warm) and the anti-sun horizon (cool) — and buildCloudShadeContext deliberately skips the overhead/zenith sky (the old comment feared zenith would blue-tint sunset cumulus). So the large bright sky dome that lights real cloud bottoms from below/around is never sampled. Fix: add a zenith sky-view sample and a new ambient term, the "sky-dome fill," that the underside picks up WITHOUT the bottom-darkening attenuation (that skylight reaches the base from below/around, not filtered down through the overlying water). It reads the zenith color, so a bright daytime dome lifts gloomy undersides and tints them with the actual sky; at sunset the zenith sample is dim so the term self-fades and the warm top-down ambient carries the look (the validated sunset is preserved bit-for-bit at fill 0, and nearly so at the 0.5 default since zenith→0 there). The companion sky←clouds half (cloud radiance bleeding into the visible sky) is still pending.

src/dxvk/shaders/rtx/pass/atmosphere/cloud_march_common.slangh — fork-owned changes. CloudShadeContext gains skyRadianceZenith; buildCloudShadeContext samples sampleSkyAmbientForVolume(vec3(0,1,0), …) * dayFactor (same cloud-occlusion + day gate as warm/cool) and assigns it; the evalNubisCubedSample call passes it through.
src/dxvk/shaders/rtx/pass/atmosphere/atmosphere_common.slangh — fork-owned change. evalNubisCubedSample gains a skyRadianceZenith param; ambient split into topAmbient = ambient_shape × verticalLight × skyRadiance (legacy, attenuated) + domeFill = ambient_shape × cloudSkyAmbientFill × skyRadianceZenith (NOT attenuated by verticalLight); result.ambient = topAmbient + domeFill.
src/dxvk/shaders/rtx/pass/atmosphere/atmosphere_args.h — fork-owned change. pad_cloudVoxel1 (the freed cloudBottomDarkeningHeight slot) → cloudSkyAmbientFill. No CB layout change.
src/dxvk/rtx_render/rtx_options.h / rtx_atmosphere.cpp — fork-owned changes. New RTX_OPTION(float, cloudSkyAmbientFill, 0.5f, …); getAtmosphereArgs fills args.cloudSkyAmbientFill.
src/dxvk/rtx_render/rtx_fork_atmosphere.cpp — fork-owned change. "Sky Fill" DragFloat added to the Clouds → Lighting tree after Bottom Darkening, with tooltip.
RtxOptions.md — hand-edited to add cloudSkyAmbientFill; regenerate in-app for canonical form.

Workstream — Sky-reflects-clouds bleed (sky picks up cloud color) (fork — 2026-06-19)

Second half of the sky↔cloud color coupling. Symptom: the visible sky and the cloud layer are composited independently (radiance = mix(sky, cloud.rgb, cloud.a) in evalSkyRadiance), so a colored deck never tints the clear sky — leaving vivid blue gaps between orange sunset clouds (and a too-clean sky around grey overcast). Fix: before the cloud composite, add a fraction of the LOCAL cloud radiance to radiance as colored inscatter. The source is the secondary cloud dome LUT (AtmosphereCloudSecondaryLut) sampled in the view direction — it is low-res (256×128), hence inherently smooth, giving a neighborhood-averaged cloud color per direction (bright/colored next to a cloud, ~0 in open sky far from any) with no extra blur pass. Because it is added BEFORE the composite, the composite's own (1 - cloudAlpha) factor keeps the bleed in the visible-sky fraction and out of opaque cloud cores. Gated on the secondary LUT being baked (default on) and cloudSkyBleedStrength > 0. Pairs with the clouds←sky "sky-dome fill" (Sky Fill); together the two close the loop so clouds and sky share color in both directions at all times of day.

src/dxvk/shaders/rtx/pass/atmosphere/atmosphere_sky.slangh — fork-owned change. In evalSkyRadiance, after the cloud-source selection and before the temporal-smoothing/composite: sample the secondary dome LUT at cloudDomeDirToUv(viewDirYUp) and radiance += cloudSkyBleedStrength * bleedCloud.rgb.
src/dxvk/shaders/rtx/pass/atmosphere/atmosphere_args.h — fork-owned change. padCloudLook2 (the freed cloudColumnShapingEnable slot) → cloudSkyBleedStrength. No CB layout change.
src/dxvk/rtx_render/rtx_options.h / rtx_atmosphere.cpp — fork-owned changes. New RTX_OPTION(float, cloudSkyBleedStrength, 0.3f, …); getAtmosphereArgs fills args.cloudSkyBleedStrength.
src/dxvk/rtx_render/rtx_fork_atmosphere.cpp — fork-owned change. "Sky Cloud Bleed" DragFloat added to the Clouds → Lighting tree after "Sky Fill", with tooltip.
RtxOptions.md — hand-edited to add cloudSkyBleedStrength; regenerate in-app for canonical form.

Workstream — Sky color correctness: physical multiscatter default (fork — 2026-06-19)

The clear sky read wrong vs reality at all times of day: over-saturated/wrong blue hue, no warm horizon, and a flat zenith→horizon gradient. Two Opus research passes converged on the cause: the DEFAULT multiscatter path was the analytical inline fit (multiScatterPhysicalStrength defaulted to 0), which is a flat, isotropic (no phase, no transmittance), strongly blue-biased fill added uniformly in every direction — it raises the floor everywhere (flattening the gradient) and desaturates the warm horizon. The sunsetSaturation knob had been added purely to band-aid this (a desaturation mask is itself the tell the base output was wrong). Fix: make the physical Hillaire multiscatter LUT the default and retire the band-aid. Also fixed a real dimensional bug in the analytical path (kept for A/B): its computeGroundReflectionAnalytical / computeAirMultiscatteringAnalytical baked sunIlluminance in, then the caller multiplied by sunIlluminance again — sunIlluminance² (≈225× at the default 15) — whereas the LUT-baker twins correctly omit it. Separately, the user's per-game config had non-physical coefficient overrides (Mie 4× low, Rayleigh ~35% low, Mie g 0.99); the code defaults are already the canonical Bruneton/Hillaire values, so those were corrected in the game's rtx.conf (not in-repo).

src/dxvk/rtx_render/rtx_options.h — fork-owned changes. multiScatterPhysicalStrength default 0.0 → 1.0 (physical LUT is now the default multiscatter); sunsetSaturation default 1.0 → (briefly 0.5) → 1.0 (band-aid retired now that the base is correct). Descriptions rewritten.
src/dxvk/shaders/rtx/pass/atmosphere/atmosphere_common.slangh — fork-owned change. computeGroundReflectionAnalytical / computeAirMultiscatteringAnalytical no longer bake in sunIlluminance (the caller applies it once, matching the physical LUT twins and contribLut); fixes the sunIlluminance² double-count on the analytical A/B path. Call-site comment added so it isn't re-introduced.

Workstream — Sky-reflects-clouds bleed: mip-blur fix (fork — 2026-06-19)

The first cut of the sky<-clouds bleed sampled the 256x128 secondary cloud dome LUT at mip 0 and added it to the sky. Result: cloud EDGES read pixelated/faceted (a single bilinear tap of a near-binary low-res signal steps across the LUT's coarse texels at silhouettes, and the coarse dome silhouette is misaligned with the sharp screen-space cloud RT), and the sky "barely changed color" (a true gap samples ~0 cloud in its exact direction — no neighborhood spread). An Opus debug pass identified both and prescribed a pre-blurred coarse-mip source. Fix: give the secondary LUT a mip chain, regenerate it (Gaussian) each frame after the bake, and sample a coarse mip (mip 4 = 16x8) for the bleed — one tap = a wide neighborhood blur, which removes the facets AND spreads cloud color smoothly into the gaps next to clouds. Re-enabled by default (cloudSkyBleedStrength 0 -> 0.15).

src/dxvk/rtx_render/rtx_atmosphere.h / rtx_atmosphere.cpp — fork-owned changes. m_cloudSecondaryLut becomes an RtxMipmap::Resource (6 mips, 256x128 -> 8x4) via RtxMipmap::createResource; the bake binds .views[0] (mip 0) for the storage write; after the dispatch a write->read barrier + RtxMipmap::updateMipmap(..., Gaussian) fills mips 1..5 (ctx is the RtxContext, cast from the DxvkContext param). Added #include "rtx_mipmap.h".
src/dxvk/rtx_render/rtx_fork_atmosphere.cpp — fork-owned change. The shared sky-view sampler gains mipmapLodMax = VK_LOD_CLAMP_NONE so explicit-LOD mip sampling works (harmless for the mip-less sky-view LUT / cloud RT).
src/dxvk/shaders/rtx/pass/atmosphere/atmosphere_sky.slangh — fork-owned change. The bleed samples AtmosphereCloudSecondaryLut at kBleedMip = 4.0 instead of 0.
src/dxvk/shaders/rtx/pass/atmosphere/atmosphere_common.slangh — fork-owned change. cloudDomeDirToUv wraps u with frac() so dome sampling no longer depends on sampler address mode (the mip-gen uses a CLAMP sampler).
src/dxvk/rtx_render/rtx_options.h — cloudSkyBleedStrength default 0.0 -> 0.15 (re-enabled).

Workstream — Cloud direct energy conservation + multiscatter sunset gather fix (fork — 2026-06-19)

Two physically-motivated lighting corrections, both addressing "lit things out-bright/mis-color the sky they composite against." WIP: shipped + deployed for in-game look validation, NOT yet eyeballed — values/approach may change after the look-check.

(1) Cloud direct dual-lobe energy conservation. The Nubis direct term summed two full-amplitude phase lobes (L_direct = T_primary*HG1 + M*HG2); HG1 and HG2 each integrate to 1 over the sphere, so the in-scatter phase integrated to up to (1+M) ≈ 2 — the cloud scattered up to ~2× the energy one scattering event can redistribute, worst at the sunlit edge where both lobes fire at once, so lit clouds out-brightened the physical sky LUT. Reformulated as an energy-conserving convex blend (phase integral 1), lerped from the legacy additive sum by a strength knob for in-game A/B. Brings the sun path onto the bounded-energy footing the moon path already had.

(2) Multiscatter sunset gather fix. The MS LUT gather in computeMultiscattering swept zenith 0..π at a SINGLE azimuth (rayDir = vec3(sinZenith, cosZenith, 0), x-y plane), but the sun lies in the y-z plane (x=0), so every gather ray sat 90° from the sun and the integral never sampled the sun-ward sky. At sunset it gathered only the blue sky orthogonal to the sun and missed the warm reddened horizon → pale-blue MS fill (sky too blue, warm band couldn't climb, clouds inherited the cold ambient via the sky-view LUT). Fixed to a proper 2D sphere quadrature (8 zenith × 8 azimuth, same 64 samples); the sinZenith Jacobian + 2π/N normalization are kept so a uniform integrand changes <1% — corrects the integral's direction/color without shifting the tuned brightness. Knob-free. (A first attempt — a multiScatterReddening tint knob — was reverted as a hack once the gather bug was found.)

src/dxvk/shaders/rtx/pass/atmosphere/atmosphere_common.slangh — fork-owned change. In evalNubisCubedSample, the direct term T_primary*HG1 + M*HG2 becomes A*(1 - s*w) + B*(1 - s*(1-w)) with A = T_primary*HG1, B = M*HG2, s = cloudEnergyConserve, w = cloudMsLobeWeight. s=0 reproduces the legacy additive sum byte-for-byte.
src/dxvk/shaders/rtx/pass/atmosphere/multiscattering_lut.comp.slang — fork-owned change. computeMultiscattering gather loop: single-azimuth 64× zenith sweep → nested 8 zenith × 8 azimuth full-sphere quadrature; rayDir gains an azimuth term. Jacobian + 2π/N normalization unchanged (uniform-integrand magnitude preserved to <1%).
src/dxvk/shaders/rtx/pass/atmosphere/atmosphere_args.h — fork-owned change. pad_artistic1 / pad_artistic2 (trailing slots of the cloudShadowFactorStrength 16-byte row) → cloudEnergyConserve / cloudMsLobeWeight. No CB layout change.
src/dxvk/rtx_render/rtx_options.h / rtx_atmosphere.cpp — fork-owned changes. New RTX_OPTION(float, cloudEnergyConserve, 1.0f, …) and RTX_OPTION(float, cloudMsLobeWeight, 0.5f, …) after cloudPhaseG2; getAtmosphereArgs fills both args.*.
RtxOptions.md — PENDING: cloudEnergyConserve / cloudMsLobeWeight not yet added; regenerate in-app for canonical form.

Workstream — Sunset "neon cloud" fixes: cloud sun-transmittance Mie + dayFactor rolloff (fork — 2026-06-20)

Final two corrections that fixed the right-before-sunset "neon orange clouds" (flat, too bright, too saturated; worst at sun elevation below sun.y ~0.2, recovering once the sun is well below the horizon). Both validated in-game.

(1) Mie/aerosol extinction added to the cloud + moon sun-direction transmittance. getAtmosphericTransmittanceForDir (the analytical Kasten-Young model that lights clouds + moons; the sky body uses the transmittance LUT) was composing extinction as Rayleigh + ozone only — the Mie term was MISSING, while the LUT (transmittance_lut.comp.slang) correctly includes Rayleigh + Mie + ozone. At a low sun the omitted grey aerosol extinction left the cloud beam under-extinguished, so cloud direct light stayed a bright, fully-saturated red while the LUT-lit sky looked correct. Added the Mie slant optical depth (mieScaleHeight * airMass), mirroring the LUT's extinction composition — physically the aerosol extinction that dims a hazy setting sun, and it realigns the cloud/moon sun color with the sky.

(2) Cloud daytime-lighting rolloff widened through the sunset approach. buildCloudShadeContext's dayFactor = smoothstep(-0.05, 0.02, sun.y) held the cloud's sun + sky-ambient lighting at FULL brightness until the geometric horizon, then crashed to night — so through the whole low-sun band (where airTrans + the sky-view LUT are most saturated) the clouds were lit at full strength = full-bright, full-saturated, flat orange. Widened the upper bound to 0.20 (~11.5°) so the daytime cloud lighting dims PROGRESSIVELY as the sun descends (physically: the scene darkens through sunset); the reddest light is no longer the brightest. Above 0.20 daytime is unchanged (golden hour untouched); horizon dayFactor ~0.2, reaching 0 at -0.05 where the moon/night terms take over.

src/dxvk/shaders/rtx/pass/atmosphere/atmosphere_common.slangh — fork-owned change. getAtmosphericTransmittanceForDir: add - args.mieScattering * (args.mieScaleHeight * airMass) to the transmittance exponent (Rayleigh + Mie + ozone, matching the LUT bake). Stale "mirrors lines 308-328" comment corrected.
src/dxvk/shaders/rtx/pass/atmosphere/cloud_march_common.slangh — fork-owned change. buildCloudShadeContext: dayFactor smoothstep upper bound 0.02 → 0.20 for a progressive sunset-approach dim of all sun-derived cloud lighting.

Workstream — Layer-2 "echo deck" rework (fork — 2026-06-21)

Replaced the old layer-2 path (a fully-independent second marchCloudSlab pass that inherited layer 1's ~32-step floor and tapped the layer-1-only D_sun grid for a decorrelated cloudscape, so every deck sample read stale residual) with a lean echo deck: the same per-column cloud-slab density model at a higher, gapped altitude, marched cheaply. Perf comes from a low, user-adjustable step budget (the dominant lever — deck cost is ~linear in step count) and an analytic sun-shadow proxy in place of the stale D_sun grid tap. Variety ("varied echo") comes from the deck's own coverage/type means + cloudLayer2NoiseSeed, which decorrelates its control field from layer 1 while sharing the same field machinery. The column-placement model is deliberately KEPT — it provides the per-sample early-out that refunds the placement tap, contrary to the earlier (now-corrected) analysis. The deck DOES take moonlight (same Beer-Lambert shadow + Wrenninge phase as layer 1, deck-aware self-shadow tap) so the two decks read consistently at night. 3 cloudLayer2* options were added (2 step-budget + 1 color) and 1 removed (cloudLayer2CoverageSpread); the rest are unchanged. Validated in-game 2026-06-21 (works after the CB-alignment fix below); same-day follow-ups: density/thickness/type decoupled from the main layer, the deck's coverage "tiling" fixed (coverage spread forced uniform + option removed), and an independent deck color.

Same-day follow-ups (2026-06-21):

Density/thickness/type decoupling. The deck ran on a copy of the main args, so main-layer sliders (most visibly cloudDensity) bled into the deck. marchCloudLayers now overrides on args2: cloudDensity → fixed kEchoDeckExtinction = 1.8 (= cloudDensity's default, so the deck decouples from the live slider with no look change at default), cloudThickness → cloudLayer2Thickness, cloudTypeMean → cloudLayer2TypeMean (the shared lighting — sampleCloudSdf, sampleDimProfile, underside down-integral — reads these directly rather than via the deck slab params). marchEchoDeck also rebuilds ctx.moonBaseSigma from the decoupled density. Coverage was already deck-specific. Audit conclusion: all OTHER look knobs (phase, multi-scatter, detail, bottom-darkening, edge, anvil, etc.) stay shared by "same slab" design — only color was split out (below).
Coverage "tiling" fix. The single-octave Worley coverage field, amplified by the spread slider, read as a fine tiled/stippled texture at the deck's distance (~3 km cells subtend a tiny angle). A de-tile attempt (domain-warp + extra octave) made it worse (small/smudgy), so it was reverted; instead marchCloudLayers forces the deck's cloudCoverageSpread to 0 — coverage collapses to a uniform cloudLayer2CoverageMean sheet, no field, no tiling. cloudLayer2CoverageSpread was removed as an option (its arg slot kept as pad_cloudLayer2CoverageSpread to preserve CB layout). Type spread is still available.
Independent deck color. New cloudLayer2Color (Vector3, defaults to cloudColor's near-white). marchCloudLayers overrides args2.cloudColor. The one look knob split from layer 1, per user request.
GPU-hang backstop. marchEchoDeck hard-caps the step count at kEchoStepHardCap = 256 so a future CB misread can't escalate to a device hang.
src/dxvk/shaders/rtx/pass/atmosphere/atmosphere_common.slangh — fork-owned change. evalNubisCubedSample gains a trailing float dSunOverride param: < 0 samples the D_sun voxel grid (production path, byte-identical to before); >= 0 is used verbatim and the grid tap is skipped. Only the echo-deck march passes >= 0.
src/dxvk/shaders/rtx/pass/atmosphere/cloud_march_common.slangh — fork-owned change. New marchEchoDeck (lean second-deck march: cloudLayer2StepFloor/cloudLayer2StepMax step budget — default 8/16 — honoring cloudViewStepKm; analytic $\text{dSunProxy} = (1 - \text{shapeHf}) \times \text{thickness} \times \text{densityScale}$ ; otherwise identical density model + control fields + per-sample moon block + Beer-Lambert/aerial as marchCloudSlab). The moon shadow uses a new slab-parametric sampleCloudSunOpticalDepth_localSlab so the deck self-shadows against ITS altitude band (the args-default helper keys on layer 1's slab → zero density at the deck = unshadowed). marchCloudLayers' layer-2 branch calls marchEchoDeck instead of a second marchCloudSlab (the old moon-coefficient zeroing is gone — the deck wants moonlight). marchCloudSlab's evalNubisCubedSample call passes dSunOverride = -1.0 (grid path). Both consumers (view RT + secondary dome LUT) inherit via marchCloudLayers.
src/dxvk/shaders/rtx/pass/atmosphere/atmosphere_args.h — fork-owned change. Adds a 16-byte (vec4) tail block: cloudLayer2StepFloor, cloudLayer2StepMax, + pad_cloudLayer2Step0/1. No free pad slots remained, so the CB grows — but it MUST grow by a whole vec4 or sizeof(AtmosphereArgs) stops being 16-byte aligned and the whole-struct updateBuffer corrupts the cbuffer (the new tail fields read garbage → marchEchoDeck's step count blows up → GPU hang → solid black when layer 2 is on). A first draft appended bare scalars (8 bytes) and hit exactly that; the two pad words fix it. Future additions should consume the pads first.
src/dxvk/rtx_render/rtx_options.h — fork-owned change. Added cloudLayer2StepFloor (uint32, 8) and cloudLayer2StepMax (uint32, 16). Rewrote cloudLayer2Enable / cloudLayer2Altitude / cloudLayer2Thickness / cloudLayer2DensityScale help strings to describe the echo deck + inter-deck gap; corrected three stale default-value mentions (7.5 km / 0.5 km / 0.30 vs the actual 5.5 / 2.0 / 0.65). No option removed or renamed.
src/dxvk/rtx_render/rtx_atmosphere.cpp — fork-owned change. Packs args.cloudLayer2StepFloor / cloudLayer2StepMax from the new RtxOptions alongside the existing cloudLayer2* packing.
src/dxvk/rtx_render/rtx_fork_atmosphere.cpp — fork-owned change. Adds "Layer 2 Step Floor" / "Layer 2 Max Steps" DragInt widgets to the Layer 2 imgui node; refreshed the Layer 2 Density tooltip off the old cirrus framing.
RtxOptions.md — regenerated; reflects the 3 new options (cloudLayer2StepFloor, cloudLayer2StepMax, cloudLayer2Color) and the removed cloudLayer2CoverageSpread.

(The atmosphere_args.h / rtx_options.h / rtx_atmosphere.cpp / rtx_fork_atmosphere.cpp bullets above also cover the follow-up cloudLayer2Color field/option/packing/picker; the args color field is a second vec4-aligned tail block — vec3 cloudLayer2Color + pad_cloudLayer2Color0. The de-tile attempt was reverted: the deck's coverage tiling is instead handled by forcing cloudCoverageSpread to 0 and removing the cloudLayer2CoverageSpread option, see below.)

Workstream — Dead-code removal + stale-doc cleanup (fork — 2026-06-21)

Audit-driven cleanup of the sky/cloud system; no behavior change. Removed two unused shader helpers — evalSunDisk (atmosphere_sky.slangh) and uvToSkyViewLutParams (atmosphere_common.slangh); removed the orphaned rtx.atmosphere.sunDisc option (never consumed — the sun disc renders via the sun-as-distant-light / NEE path) along with its RemixSkyAPI.md row and a RemixApiChangelog.md "Removed" entry. Demoted the three dead cloudVoxelGrid* dirty/offset args fields to pad_* reserves (CPU zero-writes dropped) — kept as pads so the CB layout is unchanged. Corrected ~30 stale comments + 3 user-facing RTX_OPTION descriptions that referenced the long-removed analytical evalClouds path / "no consumer in this commit" voxel grids as if current; each description was re-verified against its consumer before rewording. RtxOptions.md regenerated.

src/dxvk/rtx_render/rtx_options.h — fork-owned change. Removed the sunDisc RTX_OPTION; reworded cloudRenderRTEnable / cloudSecondaryLutEnable / cloudHeightLutEnable descriptions to drop the removed-evalClouds references (off = cloudless; secondary rays get clouds from the dome LUT). The "original 17 atmosphere options" historical list earlier in this file still names sunDisc — left as history.
src/dxvk/shaders/rtx/pass/atmosphere/atmosphere_sky.slangh, atmosphere_common.slangh — fork-owned change. Deleted evalSunDisk / uvToSkyViewLutParams; comment fixes for removed-evalClouds references.
src/dxvk/shaders/rtx/pass/atmosphere/atmosphere_args.h — fork-owned change. cloudVoxelGridFrameOffset / …SunDirty / …AmbientDirty → pad_*; comment fixes.
Comment-only doc-rot fixes in cloud_render.comp.slang, cloud_secondary_lut.comp.slang, cloud_march_common.slangh, common_bindings.slangh, common_binding_indices.h, debug_view.comp.slang, rtx_atmosphere.cpp/.h, rtx_debug_view.cpp. No functional change.
RtxOptions.md / docs/RemixSkyAPI.md / docs/RemixApiChangelog.md — regenerated / sunDisc row removed / [0.4.3] Removed entry added.

Workstream — Cloud drift overhaul: field evolution (fork — 2026-06-21)

Replaces the artificial-looking cloud "drift." The old motion had two pieces: rigid wind advection (fine) and a separate global weather-parameter drift (rtx_fork_weather.cpp) that modulated 9 cloud scalars with a sum-of-sines whose fast layer had base period exactly 30 s — its dominant inner sine was a clean 30 s beat, and all 9 fields shared one phase clock, so the whole sky visibly "breathed" in lockstep every ~30 s. The fix moves shape change into the field (differential advection, the Nubis/Decima trick) and demotes the global scalar drift to genuinely slow weather-scale change. Two new view-path levers, both reverting to legacy rigid behavior at speed 0: morph (a slow 3D scroll of the base noise sample position, Y-dominant so it samples a continuously different, tile-wrapping slice → formations form/dissolve in place, also breaking the 12 km wind tile-repeat) and edge boil (an independent faster scroll on the edge-detail tap → billows churn at the silhouette). No constant-buffer growth — four reserve pad slots are reused.

src/dxvk/shaders/rtx/pass/atmosphere/atmosphere_args.h — fork-owned change. Reuses the three contiguous pad_cloudVoxelGrid{FrameOffset,SunDirty,AmbientDirty} slots (from the 2026-06-21 dead-code pass) as cloudEvolutionOffsetX/Y/Z (float; the two former uint slots retype to float), and the pad_cloudLayer2CoverageSpread slot as cloudBoilPhase. CB layout byte-identical.
src/dxvk/shaders/rtx/pass/atmosphere/atmosphere_common.slangh — fork-owned change. In sampleCloudDensityTextured (view path only): builds noiseSamplePos = perturbed + cloudEvolutionOffset and a parallel noiseTexcoord, used for the base 3D tap (incl. hex variants) and the step-3b footprint slice ONLY — placement map, hex lattice, column model, and height fraction stay on the real perturbed/texcoord so cluster location and altitude are unaffected. Step 4b detail tap adds cloudBoilPhase * kBoilDir (fixed off-axis unit dir) to its sample position. The shadow sampler (sampleCloudDensityForShadow) is intentionally NOT evolved in v1 (baked grids keep describing the bulk field).
src/dxvk/rtx_render/rtx_atmosphere.cpp — fork-owned change. In getAtmosphereArgs(): accumulates the morph offset from cloudEvolutionSpeed / cloudEvolutionVerticalBias (Y-dominant + diagonal XZ remainder) and the boil phase from cloudBoilSpeed, both × timeSeconds (mirrors the existing cloudWindOffset pattern). Zeros the four new per-frame fields in normalizeForSkyLutCache so the sky-LUT memcmp gate doesn't fire every frame.
src/dxvk/rtx_render/rtx_options.h — fork-owned change. Adds cloudEvolutionSpeed (0.0015), cloudBoilSpeed (0.004), cloudEvolutionVerticalBias (0.8) to the rtx.atmosphere cluster.
src/dxvk/rtx_render/rtx_fork_atmosphere.cpp — fork-owned change. Adds an "Evolution" subtree (Morph Speed / Morph Vertical Bias / Edge Boil Speed) to Atmosphere → Clouds, next to "Wind". Named distinctly from the weather "Cloud Drift" subtree to avoid confusion.
src/dxvk/rtx_render/rtx_fork_weather.cpp — fork-owned change. De-pulse: driftOffsetForField drops the fast 30 s layer (slow-only now); removed kDriftFastPeriodSec. kDriftTable trimmed 9 → 3 (kept cloudCoverageMean, cloudWindSpeed, cloudWindDirection — the weather-scale fields field evolution does not reproduce; dropped the shape-ish density/thickness/type/anvil/coverage-spread entries now owned by field evolution). static_assert updated to 3.
RtxOptions.md — REGEN PENDING (run Remix with DXVK_DOCUMENTATION_WRITE_RTX_OPTIONS_MD=1): will add the 3 new options.

Workstream — Cloud motion unification + advection integrator fix (fork — 2026-06-21)

Folds the three cloud-motion sources (wind advection, field-evolution morph/boil, and the slow weather-parameter drift) into ONE per-frame integrator, fixing a latent bug. The wind/evolution/boil offsets were computed stateless as instantaneousSpeed * timeSeconds inside the const getAtmosphereArgs (called ~12×/frame, hence stateless). That is only correct while wind is constant — but the weather blender writes drift-modulated cloudWindSpeed/cloudWindDirection into the Derived config layer, so each frame the entire accumulated offset re-scaled (speed change) or rotated about the origin (direction change), snapping the whole field. Replaced with persistent accumulators advanced once per frame by offset += velocity * dt, so a varying wind eases the field instead of jumping. Rates stay independent (no cross-coupling). No CB/shader change — only how the CPU fills the existing cloudWindOffset / cloudEvolutionOffset* / cloudBoilPhase fields. Also: merged the "Wind" + "Evolution" dev-menu subtrees into one "Cloud Motion" tree, and bumped cloudLayer2StepMax 16 → 32.

src/dxvk/rtx_render/rtx_atmosphere.h — fork-owned change. Adds void advanceCloudMotion(float dt) (public) and three persistent accumulator members m_cloudAdvectOffset (Vector2) / m_cloudEvolutionOffset (Vector3) / m_cloudBoilPhase (float), next to the other per-frame-pushed members.
src/dxvk/rtx_render/rtx_atmosphere.cpp — fork-owned change. Implements advanceCloudMotion (integrates wind from drift-modulated cloudWindSpeed/cloudWindDirection, plus morph + boil, += vel*dt, with a dt<=0 pause guard). getAtmosphereArgs now reads the three members instead of computing speed*timeSeconds. normalizeForSkyLutCache already zeros those per-frame fields (unchanged).
src/dxvk/rtx_render/rtx_fork_atmosphere.cpp — fork-owned change. Calls ctx.m_atmosphere->advanceCloudMotion(GlobalTime::get().deltaTime()) once per frame in updateAtmosphereConstants (Numos block, alongside the camera-basis push); adds #include "../util/util_global_time.h". Merges the Wind + Evolution UI subtrees into one "Cloud Motion" tree with a pointer to the Weather → Cloud Drift slow modulator.
src/dxvk/rtx_render/rtx_options.h — fork-owned change. cloudLayer2StepMax default 16 → 32 (per request — layer-2 echo-deck step budget to match layer 1's cloudViewSamples = 32).
RtxOptions.md — REGEN PENDING: reflects the cloudLayer2StepMax default change.

Workstream — Rename weather "Cloud Drift" → "Weather Variation" (fork — 2026-06-21, UI clarity)

Pure dev-menu/label rename to stop the word collision with the new Atmosphere → Clouds → "Cloud Motion" tree. The weather-parameter wander (slow, preset-driven modulation of coverage + wind) is a different layer from the per-frame field motion; calling both "drift"/"motion" in adjacent panels was confusing. The __weather.drift_* GameStateStore API keys and the internal m_drift* members are UNCHANGED — only user-facing strings moved.

src/dxvk/rtx_render/rtx_fork_weather.cpp — fork-owned change. Subtree "Cloud Drift" → "Weather Variation"; slider/button labels "Drift …" → "Variation …"; tooltips note the unchanged __weather.drift_* keys; added a one-line pointer to Cloud Motion.
src/dxvk/rtx_render/rtx_fork_atmosphere.cpp — fork-owned change. Cloud Motion tree's cross-reference text updated to point at "Weather → Weather Variation".
docs/CloudSystem.md, docs/integrators/weather-presets.md — updated the dev-menu sub-tree name references (keys unchanged).

Workstream — Remove the bespoke atmosphere sun/moon NEE (fork — 2026-06-21)

Graduates the "sun/moon as real distant lights" work from an experiment-with-fallback to the sole sun/moon path in Numos, deleting the now-redundant bespoke atmosphere NEE that was previously only runtime-gated off. Supersedes the earlier evalAtmosphere*NEE* touchpoint entries above. Runtime-neutral at default settings (the bespoke path was already gated off whenever useDirectionalLights was on, the default). Cloud-on-terrain shadows are unaffected — they fold per-pixel onto the real sun distant light in the integrators (gated on atmosphereCloudShadowed + cloudVoxelShadowsEnable); the volumetric fold via sampleAtmosphereSunLightVolume is unchanged.

src/dxvk/shaders/rtx/algorithm/integrator_direct.slangh — fork-owned change. Deleted evalAtmosphereSunNEE + evalAtmosphereMoonNEE and their gated call block; the direct path now lights the atmosphere sun/moon purely via the standard NEE on the injected distant lights.
src/dxvk/shaders/rtx/algorithm/integrator_indirect.slangh — fork-owned change. Deleted evalAtmosphereSunNEESecondary + evalAtmosphereMoonNEESecondary and their gated secondary call block.
src/dxvk/shaders/rtx/pass/atmosphere/atmosphere_common.slangh — fork-owned change. Deleted the now-orphaned surface samplers sampleAtmosphereSunLight / sampleAtmosphereMoonLight and their private pickMoon helper (only the bespoke NEE called them). The volumetric samplers (sampleAtmosphereSunLightVolume, …) are retained — still used by the RTXDI volume integrator.
src/dxvk/rtx_render/rtx_options.h — fork-owned change. Removed the useDirectionalLights RTX_OPTION (injection is now unconditional in Numos) and the dead debugEnableAtmosphereNee diagnostic; reworded directionalLightRadianceScale (no longer references the toggle).
src/dxvk/rtx_render/rtx_atmosphere.cpp — fork-owned change. debugSkyBisectFlags now packs only bit 1 (debugEnableSkyMissShading); bit 0 (atmosphere NEE) and bit 2 (directional-lights skip) retired with the bespoke NEE.
src/dxvk/rtx_render/rtx_fork_atmosphere.cpp — fork-owned change. fhSyncAtmosphereDistantLights now gates on skyMode == Numos only (dropped the useDirectionalLights condition); comments de-experimentalised.
docs/CloudSystem.md — updated the cloud-on-terrain shadow section to the real-distant-light fold; corrected cloudShadowStrength default (0 → 0.5).
RtxOptions.md — REGEN PENDING (drops useDirectionalLights + debugEnableAtmosphereNee).

Workstream — Diffuse-indirect sky radiance scale (fork — 2026-06-28)

Adds the rtx.atmosphere.skyIndirectRadianceScale knob (default 1.0 = physical). Since the sun/moon graduated to real distant lights, the distant-light sun out-radiates the sky and indirect lighting reads dull; raising this multiplies sky radiance gathered by diffuse indirect bounces only. The scale is applied per-ray inside evalSkyRadiance after the sky-view LUT sample, so it never feeds any LUT bake. Eligibility is classified at the lobe-sample site (where the material type is known): only an opaque diffuse reflection/transmission gather opts in — sky seen via specular reflection, refraction, alpha-cutout (opacity transmission), translucent/glass lobes, PSR, or the primary view stays at physical brightness, so reflections keep matching the visible sky.

src/dxvk/rtx_render/rtx_options.h — fork-owned addition. Adds RTX_OPTION_ARGS(rtx.atmosphere, float, skyIndirectRadianceScale, 1.0f, …, args.minValue = 0.0f) after sunsetSaturation.
src/dxvk/shaders/rtx/pass/atmosphere/atmosphere_args.h — fork-owned change. Repurposes the dead pad_cloudAnisotropy slot as skyIndirectRadianceScale; CB layout / sizeof(AtmosphereArgs) unchanged.
src/dxvk/rtx_render/rtx_atmosphere.cpp — fork-owned change. getAtmosphereArgs populates the field (clamped >= 0). normalizeForSkyLutCache zeroes it in the sky-LUT memcmp key — it is applied post-LUT, so dragging the slider must not trigger a rebake.
src/dxvk/shaders/rtx/pass/atmosphere/atmosphere_sky.slangh — fork-owned change. Adds the trailing applySkyIndirectRadianceScale parameter (default false) to evalSkyRadiance and multiplies the final radiance by args.skyIndirectRadianceScale only when it is set.
src/dxvk/shaders/rtx/algorithm/path_state.slangh — fork-owned change. Adds _skyGatherEligible (uint8) + a skyGatherEligible bool property so the indirect integrator knows whether the in-flight ray came from a diffuse sky gather (_flags is bit-saturated, hence its own byte).
src/dxvk/shaders/rtx/algorithm/integrator_indirect.slangh — fork-owned change. Initializes skyGatherEligible (false in pathStateCreateEmpty; !firstSampledLobeIsSpecular at first bounce); reclassifies it after each continuation sample using materialType + the opaque lobe enum; passes it as applySkyIndirectRadianceScale at the physical-atmosphere sky miss.
src/dxvk/rtx_render/rtx_fork_atmosphere.cpp — fork-owned change. Adds the "Sky Indirect Scale" ImGui slider (0–20) + tooltip in the atmosphere UI.
RtxOptions.md — REGEN PENDING (adds rtx.atmosphere.skyIndirectRadianceScale).

Workstream — Cloud detail-shading pass (fork — 2026-07-14)

Attacks the "blobby clouds" read at its root: the edge-detail field previously only wobbled the coverage threshold (silhouette), while every lighting input (D_sun grid, dim profile, SDF proxy) is km-scale smooth — so threshold-grown billows were invisible inside the cloud body. The same detail signal now also SHADES: billow micro-AO (grown knuckles brighten, carved crevices darken, gated by SDF surface proximity so cores stay clean), a Schneider powder term (thin sun-facing wisps darken; faded toward the sun so silver linings survive), a wispy-base/billowy-top character flip over the bottom quarter of each column, and a height-fading horizontal shear on the detail tap. View path only; the cheap shadow sampler and the validated self-shadow bakes are untouched; every knob at 0 is bit-identical legacy.

src/dxvk/shaders/rtx/pass/atmosphere/atmosphere_common.slangh — fork-owned change. sampleCloudDensityTextured grows an out float detailSigned (zero-mean character-shaped detail signal; convenience overload forwards a dummy) and step 4b gains the character flip + base shear; the tap now also runs when only micro-AO wants it. evalNubisCubedSample gains the detailSigned param plus the micro-AO (ambient + MS body lobe only, clamp [0.15, 1.25]) and powder blocks.
src/dxvk/shaders/rtx/pass/atmosphere/cloud_march_common.slangh — fork-owned change. Both march loops (layer 1 + echo deck) thread detailSigned from the density sampler into the lighting evaluator.
src/dxvk/shaders/rtx/pass/atmosphere/atmosphere_args.h — fork-owned change. Reclaims the pad_cloudShadowTint (vec3) + pad_cloudShadowTintStrength row as cloudMicroAoStrength / cloudPowderStrength / cloudDetailHeightCharacter / cloudDetailBaseShearKm; CB layout unchanged.
src/dxvk/rtx_render/rtx_options.h — fork-owned additions. 4 RTX_OPTIONs in the rtx.atmosphere cluster after cloudDetailScale: cloudMicroAoStrength (0.6), cloudPowderStrength (0.5), cloudDetailHeightCharacter (0.7), cloudDetailBaseShearKm (0.2).
src/dxvk/rtx_render/rtx_atmosphere.cpp — fork-owned change. getAtmosphereArgs populates the four fields.
src/dxvk/rtx_render/rtx_fork_atmosphere.cpp — fork-owned change. 4 sliders (Detail Shading / Powder Darkening / Base Wispiness / Base Wisp Shear) in the Detail & Edges tree.
RtxOptions.md — REGEN PENDING (4 new options).

Workstream — Cloud dramatic shading (fork — 2026-07-14)

Adds the contrast axis the Nubis ambient lacked (reference: CoD4 iw3xo daynight cumulus). The sky-ambient fill (topAmbient) was the one light term with no response to D_sun, so it reflooded sun-shadowed bulk with bright daytime sky light and put a high flat floor under the shading — clouds read soft/flat however the direct lobes were tuned. evalNubisCubedSample now attenuates the top-down ambient by exp(-0.6 * D_sun) (internal diffusion-flavored sigma, well below the beam sigma), lerped in by the O(1) cloudAmbientShadowStrength knob: sunlit faces / silver linings (D_sun ~ 0) keep full ambient, shaded cores plunge dark. The sky-dome underside fill (cloudSkyAmbientFill) is deliberately exempt — it models open-sky light arriving from below/around (not through) the cloud, and stays the tunable underside floor. Echo deck inherits via its analytic dSunProxy; secondary-ray LUT re-bakes on slider change (field not zeroed in the cache-key normalizers). 0 = bit-identical legacy. Also fixes the INVERTED cloudMsScale doc/tooltip (higher = darker shadowed bulk, not brighter).

src/dxvk/shaders/rtx/pass/atmosphere/atmosphere_common.slangh — fork-owned change. evalNubisCubedSample: ambientShadow block ahead of the ambient composite; multiplies topAmbient only.
src/dxvk/shaders/rtx/pass/atmosphere/atmosphere_args.h — fork-owned change. Reclaims pad_cloudMultiScatterStrength as cloudAmbientShadowStrength; CB layout unchanged.
src/dxvk/rtx_render/rtx_options.h — fork-owned addition. cloudAmbientShadowStrength (default 0.6) after cloudSkyAmbientFill; cloudMsScale doc direction fixed.
src/dxvk/rtx_render/rtx_atmosphere.cpp — fork-owned change. getAtmosphereArgs populates the field (next to cloudMsScale).
src/dxvk/rtx_render/rtx_fork_atmosphere.cpp — fork-owned change. "Ambient Shadowing" slider in the Lighting tree (between Bottom Darkening and Sky Fill); Multi-Scatter tooltip direction fixed.
RtxOptions.md — REGEN PENDING (1 new option + cloudMsScale doc fix).

Workstream — Lightning (fork — 2026-07-14)

In-cloud lightning flashes plus a synchronized transient scene light, driven by a CPU strike scheduler. Strikes fire via a per-frame Bernoulli draw at lightningStrikesPerMinute (memoryless Poisson — adapts instantly to the weather blender's continuous rate ramp), place themselves in a 1 km–lightningRangeKm annulus around the camera at cloud-base height, and run a ~70 ms-decay envelope with 0–2 restrike pulses. The in-cloud emissive scales with local density via the Beer-Lambert accumulation, so a strike where the column model placed no cloud lights nothing. The flash is compile-gated (CLOUD_MARCH_LIGHTNING) to the screen cloud pass only — the persistent secondary-ray cloud LUT must never bake a transient flash. lightningEnable (default TRUE) is a master mute; the rate (default 0) is the real switch, raised by storm weather presets (thunderstorm 12/min, rainstorm 4/min) via the new 53rd preset field lightningStrikesPerMinute.

src/dxvk/shaders/rtx/pass/atmosphere/atmosphere_args.h — fork-owned addition. Two new vec4 rows at the struct end: (lightningStrikePosKm, lightningFlashIntensity) and (lightningColor, lightningEnvelope — the raw envelope, so the scene light calibrates independently).
src/dxvk/shaders/rtx/pass/atmosphere/cloud_march_common.slangh — fork-owned addition. CLOUD_MARCH_LIGHTNING gate (default 0) + evalLightningFlash (inverse-square, 0.5 km soft core, 0.35/km extinction reach) added to both march loops' in-scatter source.
src/dxvk/shaders/rtx/pass/atmosphere/cloud_render.comp.slang — fork-owned addition. Defines CLOUD_MARCH_LIGHTNING 1 before including the march header (sole flash consumer).
src/dxvk/rtx_render/rtx_atmosphere.h / rtx_atmosphere.cpp — fork-owned additions. advanceLightning(dt) scheduler (xorshift32, envelope decay, restrikes, annulus placement) + requestLightningStrike() static latch for the Test Strike button; getAtmosphereArgs publishes the lightning CB fields; normalizeForSkyLutCache zeroes them (per-frame animated, never feeds a LUT bake).
src/dxvk/rtx_render/rtx_fork_atmosphere.cpp — fork-owned additions. advanceLightning tick after the camera-position push; transient RtSphereLight (150 m emitter, persistent handle, zero-radiance between strikes, NOT cloudShadowed) in fhSyncAtmosphereDistantLights; "Lightning" ImGui tree (enable + Test Strike + rate/intensities/range/color).
src/dxvk/rtx_render/rtx_options.h — fork-owned additions. 6 RTX_OPTIONs: lightningEnable (true), lightningStrikesPerMinute (0), lightningFlashIntensity (50), lightningSceneLightIntensity (2000), lightningRangeKm (10), lightningColor (blue-white).
src/dxvk/rtx_render/rtx_fork_weather.h / rtx_fork_weather.cpp — fork-owned additions. lightningStrikesPerMinute as weather-preset field 53 (WK_Scalar, Clouds → Lightning): FIELD_LIST row + all 12 preset value macros + the four hand-listed sites (tooltip map, snapshotRenderer, WVARIES, gated writeBlendedToDerivedLayer).
RtxOptions.md — REGEN PENDING (6 lightning options + 12 preset fields).

Workstream — Weather preset retune (fork — 2026-07-14)

Look-check driven ("thunderstorm looks like a normal day — the sky is blue, there's a lot of light"). Root cause: every preset inherited the clear-day Rayleigh spectrum and only dimmed the sun. The moody presets now retune three axes — sky COLOR (rayleighScattering flattened toward grey; sandstorm inverts the spectrum for an orange-brown sky), MURK (aerosolDensity up), and LIGHT (sunIlluminance cut hard, skyIndirectRadianceScale reduced). cloudShadowStrength is 1.0 in ALL presets (per request). Thunderstorm/rainstorm also tighten cloudAerialFadePerKm so horizon blue can't bleed through the deck. Clear is deliberately untouched. Because rayleighScattering and skyIndirectRadianceScale now vary across presets, their weatherVaries gates fire and both get written during a blend (comment updated at the write site).

src/dxvk/rtx_render/rtx_fork_weather.h — fork-owned change. Per-preset values only (no structural change): shadow 1.0 ×12; graded sky/murk/light retunes for overcast, drizzle, rainstorm, thunderstorm, snow, blizzard, foggy; sandstorm orange sky + warm night sky; smoggy brown-grey; hazy milky; partlyCloudy coverage 0.35 / type 0.65.
src/dxvk/rtx_render/rtx_fork_weather.cpp — fork-owned change. Stale "neutral in every preset" gate comment updated in writeBlendedToDerivedLayer.

Workstream — Lightning ghost suppression (fork — 2026-07-14)

A lightning flash embedded into the cloud temporal smoother's ~1 s EMA (fixed 0.92 history weight in evalSkyRadiance) outlived itself 3–10x; on camera move the reprojected history dragged a sharp-edged "old frame" imprint of the flash-lit deck across the sky. The strike scheduler now tracks a ghost-suppression signal — 1 while a flash is live, decaying over ~0.25 s after — and the temporal blend collapses its history weight by x(1 - 0.8 * fade) (0.92 → ~0.18 at full fade, ~2-frame convergence), so the flash never embeds. The jitter the EMA normally hides is masked by the flash itself; inert at fade 0. Validated in-game.

src/dxvk/shaders/rtx/pass/atmosphere/atmosphere_args.h — fork-owned change. Reclaims the pad_cloudSunsetWarmth slot as lightningHistoryFade; CB layout unchanged.
src/dxvk/rtx_render/rtx_atmosphere.h / rtx_atmosphere.cpp — fork-owned change. m_lightningHistoryFade tracked at the end of advanceLightning (max of envelope and a tau-0.25 s decay); published by getAtmosphereArgs; zeroed in normalizeForSkyLutCache.
src/dxvk/shaders/rtx/pass/atmosphere/atmosphere_sky.slangh — fork-owned change. Temporal-smoothing block scales kHistoryWeight by the fade before the history lerp.

Workstream - Nubis3 conversion Phase A: cloud NVDF SDF bake (fork - 2026-07-16)

Foundation for converting the Numos cloud MODELING to the Nubis3 (SIGGRAPH 2023) voxel/SDF architecture. The procedural cloud BODY (placement map + column model - our "Frankencloudscape") is voxelized into a tile-periodic 256x64x256 occupancy grid and distance-transformed by a wrap-aware jump- flooding (JFA) chain into a REAL signed distance field (R16F, raw km, negative inside; texture y = VERTICAL - see cloud_nvdf.h). Double-buffered publish-swap; full synchronous chain at init, amortized 2-JFA-passes-per- frame state machine for runtime re-bakes (dirty keys: cell/tile size, column shape knobs, quantized thickness, quantized nominal coverage). Detail noise never enters the SDF (sample-time erosion per Nubis3). Phase A ships the bake infra + a debug slice view only - no visual change; the density-model swap (profile-from-SDF behind rtx.atmosphere.nubis3ModelEnable) is Phase B and SDF sphere-trace stepping is Phase C.

src/dxvk/shaders/rtx/pass/atmosphere/cloud_nvdf.h - NEW fork-owned file. Shared CPU/GPU dims (256x64x256), pass-local binding maps for the three NVDF passes, CloudNvdfJfaArgs push-constant struct.
src/dxvk/shaders/rtx/pass/atmosphere/cloud_nvdf_common.slangh - NEW fork-owned file. Seed pack/unpack (R32_UINT, 0xFFFFFFFF invalid), wrap-aware anisotropic voxel-center distance in km, voxel -> tile-UV / height-fraction mapping.
src/dxvk/shaders/rtx/pass/atmosphere/cloud_nvdf_occupancy.comp.slang - NEW fork-owned file. Voxelizes computeCloudColumn at the baked NOMINAL coverage (hexOn=false, no noise taps) into R8 occupancy.
src/dxvk/shaders/rtx/pass/atmosphere/cloud_nvdf_jfa.comp.slang - NEW fork-owned file. Push-constant mode 0 = boundary seed init (occupied voxel with an empty 6-neighbor; Y out-of-slab counts empty), mode 1 = $3^{3}$ jump pass at jumpSizeVoxels comparing wrapped physical-km distances.
src/dxvk/shaders/rtx/pass/atmosphere/cloud_nvdf_resolve.comp.slang - NEW fork-owned file. Distance to closest boundary seed, sign from occupancy, +100 km when no seeds exist; writes the BACK SDF buffer.
src/dxvk/rtx_render/rtx_atmosphere.h / rtx_atmosphere.cpp - fork-owned change. Resources (occupancy, JFA ping-pong x2, SDF front/back x2, ~54 MB), jump schedule {128..1,1}, dirty-key struct + amortized state machine (stepCloudNvdfBake in computeLuts after the placement-rebake block; placement rebake clears the NVDF key), full init chain in initialize(), three dispatch fns + 4 ManagedShader classes, getCloudNvdfSdf() front-buffer getter, args.nvdfNominalCoverage fill (auto = live coverage quantized to 0.25 steps, option pins).
src/dxvk/shaders/rtx/pass/atmosphere/atmosphere_args.h - fork-owned change. Batched pad consumption (CB layout unchanged, one edit for the whole phase): pad3->nubis3SharpenStrength, padMilkyWay0->nvdfStepScale, pad_sunShadowMaxSamples->nubis3ModelEnable, pad_moonShadowMaxSamples->nvdfNominalCoverage, pad_cloudLayer2Step0/1->nvdfProfileDepthKm/nvdfCoverageOffsetKm, pad_cloudLayer2Color0->nubis3ErosionStrength. Only nvdfNominalCoverage is consumed in Phase A; the rest are zero-filled until their phase.
src/dxvk/rtx_render/rtx_options.h - fork-touchpoint inline tweak. Adds rtx.atmosphere.nvdfNominalCoverage (0 = auto-track quantized weather coverage).
src/dxvk/rtx_render/rtx_context.h - fork-touchpoint inline tweak. fork_hooks::getCloudNvdfSdf forward declaration + friend line (mirrors getCloudDSun).
src/dxvk/rtx_render/rtx_fork_atmosphere.cpp - fork-owned addition. getCloudNvdfSdf lazy-init accessor.
src/dxvk/shaders/rtx/utility/debug_view_indices.h - index-only, fork. Adds DEBUG_VIEW_CLOUD_NVDF_SDF = 879.
src/dxvk/shaders/rtx/pass/debug_view/debug_view_binding_indices.h - index-only, fork. Adds DEBUG_VIEW_BINDING_CLOUD_NVDF_SDF_INPUT = 43 (extends the fork cloud-debug block to 39-43).
src/dxvk/shaders/rtx/pass/debug_view/debug_view.comp.slang - fork-owned addition. Texture3D<float> DebugViewCloudNvdfSdf + case 879: horizontal slice at height fraction 0.25, warm-inside / cool-outside ramps, green zero-crossing band.
src/dxvk/rtx_render/rtx_debug_view.cpp - fork-owned addition. TEXTURE3D(43) parameter entry, bind via fork_hooks::getCloudNvdfSdf, selector entry with validation gates (smooth blobby cells, clean iso-line, seamless tile wrap, live re-bake on cell-size drag).
RtxOptions.md - REGEN PENDING (nvdfNominalCoverage).

Workstream - Nubis3 conversion Phase B: SDF density model + evaluator move (fork - 2026-07-16)

The density-model swap, behind rtx.atmosphere.nubis3ModelEnable (default OFF - legacy path bit-identical for in-game A/B). Cloud shape becomes the Nubis3 recipe: dimensional profile = sample-time remap of the (Phase A) body SDF with live coverage as a level-set offset (zero-rebake coverage dynamics), eroded by a new wispy/billowy detail volume via the Schneider ValueErosion remap + page-123 pow sharpen. ONE sampler serves the view march AND the shadow/OD bakes (iso parity by construction; also erases the legacy view-0.15 vs shadow-0.25 gate-softness parity gap). The Nubis Cubed lighting evaluator MOVED out of atmosphere_common.slangh into cloud_march_common.slangh (its only callers) and split into a parameterized core (real profile / live SDF meters / D_ambient-derived downTau) + a bit-identical legacy wrapper - lighting edits no longer recompile the path tracer. Echo deck stays on the legacy sampler (deferred to Phase E).

src/dxvk/shaders/rtx/pass/atmosphere/cloud_detail_noise_baker.comp.slang - NEW fork-owned file. $128^{3}$ RGBA8 volume: R/G = low/high-freq wispy (Perlin FBM), B/A = low/high-freq billowy (inverted Worley), integer cycles-per-volume periods; baked once at init (fixed pattern).
src/dxvk/shaders/rtx/pass/atmosphere/cloud_nubis3_common.slangh - NEW fork-owned file. valueErosion, sampleCloudDensityNubis3 (hex-de-tiled SDF tap -> coverage offset -> profile gate/early-out -> animated wispy/billowy erosion -> sharpen; outs profile/detailSigned/live-sdf-km), plus the MOVED D_sun/D_ambient bake integrals with a model-branched integrand.
src/dxvk/shaders/rtx/pass/atmosphere/atmosphere_common.slangh - fork-owned change (the phase's one batched edit). REMOVES the moved blocks: the two OD bake integrals and the whole Nubis lighting evaluator block (sampleDimProfile / sampleCloudSdf / hgPhaseNubis / NubisCubedLighting / evalNubisCubedSample); short pointers left in place. sampleDSun/sampleDAmbient + ground-shadow helpers stay (integrators need them).
src/dxvk/shaders/rtx/pass/atmosphere/cloud_march_common.slangh - fork-owned change. Receives the evaluator as evalNubisCubedSampleCore(dim_profile, cloud_sdf_m, downTau, ...) + legacy wrapper; includes cloud_nubis3_common; marchCloudSlab branches density (Nubis3 sampler vs legacy) and lighting (core with real SDF into the sigma_ms remap + sampleDAmbient x cloudDensity as downTau vs wrapper); layer-1 moon OD tap branches likewise. marchEchoDeck untouched (legacy).
src/dxvk/shaders/rtx/pass/atmosphere/cloud_render.comp.slang / cloud_secondary_lut.comp.slang - fork-owned change. Pass-local bindings 13 = AtmosphereCloudNvdfSdf (front buffer), 14 = AtmosphereCloudDetailNoise3D; sampled with the existing linear/REPEAT cloud sampler (slot 2).
src/dxvk/shaders/rtx/pass/atmosphere/cloud_sun_density_grid.comp.slang / cloud_ambient_density_grid.comp.slang - fork-owned change. Pass-local bindings 5 = NVDF SDF, 6 = detail volume; call the moved integrals with the extended signature. Terrain cloud shadows track the model switch with zero integrator changes.
src/dxvk/rtx_render/rtx_atmosphere.h / rtx_atmosphere.cpp - fork-owned change. m_cloudDetailNoise3D ( $128^{3}$ RGBA8, ~8 MB) + one-shot init bake + CloudDetailNoiseBakerShader; grid/render/secondary shader classes grow TEXTURE3D slots (5/6 and 13/14) with binds + read-tracking at all four dispatch sites (front SDF + detail); getAtmosphereArgs fills the five Nubis3 CB fields (pads consumed in Phase A).
src/dxvk/rtx_render/rtx_options.h - fork-touchpoint inline tweak. Adds rtx.atmosphere.nubis3ModelEnable (false), nvdfProfileDepthKm (1.0), nvdfCoverageOffsetKm (1.5), nubis3ErosionStrength (1.0), nubis3SharpenStrength (1.0).
src/dxvk/rtx_render/rtx_fork_atmosphere.cpp - fork-owned addition. ImGui Clouds -> "Nubis3 Model (SDF)" block: enable toggle + profile depth / coverage reach / erosion / sharpen / bake-nominal sliders.
RtxOptions.md - REGEN PENDING (5 new options + Phase A's nvdfNominalCoverage).

Workstream - Nubis3 anti-blobby pass + Phase C SDF stepping (fork - 2026-07-16)

Phase B validated ("more natural") but read as convex blobs. Root causes and fixes, each on its own runtime lever for in-game bisection: (1) the BODIES were convex by construction (placement blob x column band) - a tile-periodic 3D FBM now shifts the placement waterline per voxel in the occupancy bake, so columns bake in overhangs/notches/lumps (nvdfBodyErosionStrength, an NVDF re-bake dirty key; zero-mean so nominal-coverage semantics stay centered); (2) the detail volume's erosion/wobble content was far too coarse vs Nubis's own (~sub-100 m; ours was 700/350 m) - baker channels move to 6/16 cycles-per-volume with 4/3 octaves, and the silhouette wobble keys on a fixed half-mix of the high-freq channels instead of the profile-blend (which collapsed to pure low-freq exactly at the silhouette); (3) ported the Nubis p.125 near-camera HF detail (twice-folded high-freq channels replace a slice of the erosion composite near the camera; range scaled 0.2->2 km for our coarser cycles-per-km; nubis3HFDetailStrength, 1 = the paper's 10% max mix). Phase C: the sampler additionally returns a CONSERVATIVE step bound (MIN of the hex taps - a blended distance is not metric - minus coverage offset and max outward wobble), and marchCloudSlab jumps whole loop indices through provably-empty air (nvdfStepScale safety factor, 0 = off; index-quantized so samples stay on the jittered stratified lattice - no temporal/banding change).

src/dxvk/shaders/rtx/pass/atmosphere/cloud_nvdf_occupancy.comp.slang - fork-owned change. Body-erosion carve: tile-periodic fbmNoise3DPeriodic (6 XZ cycles, thickness-proportional Y cycles, 4 octaves) shifts placement.r before computeCloudColumn.
src/dxvk/shaders/rtx/pass/atmosphere/cloud_detail_noise_baker.comp.slang - fork-owned change. Channel frequencies 4/8 -> 6/16 cycles-per-volume, octaves 3 -> 4/3 (finest octaves pinned at the 128-texel Nyquist).
src/dxvk/shaders/rtx/pass/atmosphere/cloud_nubis3_common.slangh - fork-owned change. Full-overload signature grows cameraDistKm in + sdfStepKmOut out; min-of-taps tracking; p.125 HF detail block; wobble signal rev 2 (fixed high-freq half-mix). Convenience overload (bakes/OD taps) unchanged externally - passes 1e6 dist (HF detail must not make bakes camera-dependent).
src/dxvk/shaders/rtx/pass/atmosphere/cloud_march_common.slangh - fork-owned change. marchCloudSlab passes ray t as camera distance and, on empty samples, advances the loop index by floor(sdfStep x nvdfStepScale / stepLen) (4096 hard cap).
src/dxvk/shaders/rtx/pass/atmosphere/atmosphere_args.h - fork-touchpoint inline tweak. padMilkyWay1 -> nvdfBodyErosionStrength, padMilkyWay2 -> nubis3HFDetailStrength; CB layout unchanged. (All former reserve pads are now consumed - Phase D growth needs a new 16-byte block.)
src/dxvk/rtx_render/rtx_atmosphere.h / rtx_atmosphere.cpp - fork-owned change. CloudNvdfBakeKey.bodyErosion dirty key + cache; getAtmosphereArgs fills the two new pads + nvdfStepScale (CB field existed since Phase A, zero-filled until now).
src/dxvk/rtx_render/rtx_options.h - fork-touchpoint inline tweak. Adds rtx.atmosphere.nvdfBodyErosionStrength (0.6), nubis3HFDetailStrength (1.0), nvdfStepScale (0.8).
src/dxvk/rtx_render/rtx_fork_atmosphere.cpp - fork-owned addition. "Nubis3 Model (SDF)" block grows Body Erosion / HF Detail (Near) / SDF Step Scale sliders.
RtxOptions.md - REGEN PENDING (3 more options; 9 total outstanding for the Nubis3 workstreams).

Workstream - Cloud crispness: temporal-smoother weight knob (fork - 2026-07-16)

Follow-up to the anti-blobby pass: with the SHAPE fixed, the remaining "bad" read was smear - the cloud RT runs at 0.5x the DLSS-internal resolution and the temporal smoother blended with a HARDCODED 0.92 EMA weight (~0.5 s settle), so edges arrived pre-blurred and doubly upscaled. Render scale was already a live option; this makes the EMA weight one too. Consumes the first scalar of a NEW 16-byte CB block appended at the AtmosphereArgs tail (all former reserve pads were spent by the Nubis3 workstreams; 3 reserve pads remain for Phase D dynamics).

src/dxvk/shaders/rtx/pass/atmosphere/atmosphere_args.h - fork-touchpoint inline tweak. NEW tail block: cloudHistoryWeight + padReserve0/1/2 (16-byte aligned growth per the CB discipline).
src/dxvk/shaders/rtx/pass/atmosphere/atmosphere_sky.slangh - fork-owned change. evalSkyRadiance temporal blend reads args.cloudHistoryWeight (saturated) instead of the hardcoded kHistoryWeight = 0.92; lightning ghost-suppression modulation unchanged on top.
src/dxvk/rtx_render/rtx_atmosphere.cpp - fork-owned change. getAtmosphereArgs fills the new block (weight clamped [0..0.98], reserve pads zeroed); normalizeForSkyLutCache zeroes the weight (composite-only - slider drags must not invalidate the sky-LUT cache key).
src/dxvk/rtx_render/rtx_options.h - fork-touchpoint inline tweak. Adds rtx.atmosphere.cloudHistoryWeight (0.92 = previous hardcoded behavior; 0 = raw jittered march).
src/dxvk/rtx_render/rtx_fork_atmosphere.cpp - fork-owned addition. "Temporal Smoothing" slider next to Cloud Render Scale.
RtxOptions.md - REGEN PENDING (10 options now outstanding across the Nubis3 + crispness workstreams).

Workstream - Nubis3 interior texture + alligator noise (fork - 2026-07-16)

User verdict on the anti-blobby knobs: "barking up the wrong tree - the implementation is missing a core feature." Re-reading both references (Nubis Cubed pp.89-99; iw3xo ps_3_0_iw3xo_daynight.hlsl) confirmed it: (1) OUR DENSITY SATURATED TO A CONSTANT 1.0 inside the body (ValueErosion at profile 1 is 1 regardless of noise), so every face beyond the thin silhouette skirt rendered as a flat white mass. Both references keep density proportional to the noise EVERYWHERE inside: iw3xo multiplies its razor gate by the raw FBM (dens *= smoothstep(cov, cov+0.05, dens)), Nubis3 multiplies by its authored per-voxel Density Scale NVDF (uprezzed_density *= powered_density_scale). New sampler step 8 modulates the post-sharpen density by the type-blended raw detail channels (nubis3InteriorTexture, default 0.7, rides padReserve0; shared sampler, so the D_sun/D_ambient bakes track automatically). (2) The detail volume used the EXACT noises the talk rejects (p.98: inverted Worley = "packed spheres", Perlin "not wispy enough") - the baker now approximates their replacements: alligator3DPeriodic (sparse-bump, strongest-minus-runner-up = amplitude-layered lumps with creases) for billows, curl-warped inverted alligator ("Curly-Alligator", periodic curl of three Perlin FBM potentials) for wisps.

src/dxvk/shaders/rtx/pass/atmosphere/cloud_detail_noise_baker.comp.slang - fork-owned change. NEW alligator3DPeriodic / alligatorFbm3DPeriodic / curlNoise3DPeriodic helpers (local to the baker - one-shot init cost); channels rebuilt: R/G = curly-alligator wisps (pow-2 to recenter mean ~0.5 for the zero-mean wobble), B/A = alligator billows.
src/dxvk/shaders/rtx/pass/atmosphere/cloud_nubis3_common.slangh - fork-owned change. Sampler step 8: interior density modulation, range [0.25, 1.75] around a roughly density-neutral mean, applied post-sharpen pre-saturate.
src/dxvk/shaders/rtx/pass/atmosphere/atmosphere_args.h - fork-touchpoint inline tweak. padReserve0 -> nubis3InteriorTexture; CB layout unchanged (2 reserve pads left).
src/dxvk/rtx_render/rtx_atmosphere.cpp - fork-owned change. getAtmosphereArgs fills it (clamped [0..1]).
src/dxvk/rtx_render/rtx_options.h - fork-touchpoint inline tweak. Adds rtx.atmosphere.nubis3InteriorTexture (0.7).
src/dxvk/rtx_render/rtx_fork_atmosphere.cpp - fork-owned addition. "Interior Texture" slider at the top of the Nubis3 block.
RtxOptions.md - REGEN PENDING (11 options outstanding).

Workstream - Nubis3 edge wisp cut (fork - 2026-07-16)

User feedback after validating the interior-texture round ("helped a lot"): "the erosion should cut out wisps and stuff." Root cause: the base erosion signal tops out ~0.4 at liked settings (type-blended composite mean ~0.35 x strength 0.6), so ValueErosion only nibbles the outer ~40% of the profile - it can never cut THROUGH the shell and detach a wisp shape. Fix: (1) sampler step 7b adds erosion shaped by the WISPY channel alone, weighted (1-profile)^2 - strand-shaped cuts reach full depth at the silhouette and fade by mid-shell, so billowy cores keep rounded edges at any type (nubis3EdgeErosion, default 1.0, rides padReserve1 - ONE reserve pad left); (2) the baker's wispy channels are now anisotropic (2x vertical cycle count = features twice as wide as tall; integer per-axis periods keep exact tiling), so the cuts read as sheared trailing streaks instead of round webs.

src/dxvk/shaders/rtx/pass/atmosphere/cloud_nubis3_common.slangh - fork-owned change. Step 7b edge-wisp erosion added into the ValueErosion input.
src/dxvk/shaders/rtx/pass/atmosphere/cloud_detail_noise_baker.comp.slang - fork-owned change. kWispSqueeze (1,2,1) anisotropy on the wispy channels' sample coords, curl field, and periods.
src/dxvk/shaders/rtx/pass/atmosphere/atmosphere_args.h - fork-touchpoint inline tweak. padReserve1 -> nubis3EdgeErosion; CB layout unchanged.
src/dxvk/rtx_render/rtx_atmosphere.cpp - fork-owned change. getAtmosphereArgs fills it (clamped [0..3]).
src/dxvk/rtx_render/rtx_options.h - fork-touchpoint inline tweak. Adds rtx.atmosphere.nubis3EdgeErosion (1.0).
src/dxvk/rtx_render/rtx_fork_atmosphere.cpp - fork-owned addition. "Edge Wisp Cut" slider after Erosion Strength.
RtxOptions.md - REGEN PENDING (12 options outstanding).

Same-day follow-up (user look checks: "alligator noise is really intense / looks fake", "especially looking straight up"):

cloud_detail_noise_baker.comp.slang - alligator tempering: kernel exponent 2.0 -> 1.6 (softer shoulders), amplitude range widened 0.35..1 -> 0.25..1 (lump-size variety), contrast gain 1.8 -> 1.25 (the hard gain saturated lumps into plateaus with razor creases).
cloud_nubis3_common.slangh - (1) TWO-SCALE DE-TILE on the detail tap: the volume repeat (~2.8 km) was NOT hex-de-tiled and reads as a grid when a face is seen flat-on (worst at the zenith); a second tap at 0.531x scale blended 40/60 makes the pattern effectively non-repeating, variance re-expanded 1.35x around channel means. (2) Nubis p.125 NEAR-CAMERA DENSITY SOFTEN ported (pow 0.5->1.0, gain 0.666->1.0 over the HF range) — the HF fold without its companion soften read harsh in fly-through. (3) D_SUN SUNSET BANDING FIX in sampleCloudSunOpticalDepthAtWorld (user report: "squares"/"bands" cast into clouds at sunset, both models — the grid predates Nubis3): the fixed 8 taps spread over an uncapped near-horizontal 20-50 km sun path landed kilometers apart, so neighbor voxels integrated random cloud subsets (squares = grid texels) and tap-distance quantization made onion shells perpendicular to the sun (arcs). Fix: path capped at 12 km (integral saturated beyond), adaptive 8..24 taps (~0.6 km step), deterministic world-anchored per-voxel tap-phase jitter (residual error becomes bilinear-absorbed noise, stable across frames).

Workstream - GT7 cross-audit bug fixes (fork - 2026-07-16)

A Fable-subagent audit of our sky/cloud shaders against the GDC23 GT7 talk (report: Fable_5_testing/gt7-applicability-report.md) surfaced three verified BUGS, landed here. (Its LUT verdict: our sky-view LUT does NOT have GT7's misaligned-horizon arc failure — per-0.1-degree re-bake + v^2 horizon pin — so the in-cloud sunset artifacts trace to the D_sun grid instead.)

src/dxvk/shaders/rtx/pass/atmosphere/atmosphere_sky.slangh - fork-owned change. PREMULTIPLY DOUBLE-APPLY: both cloud sources store premultiplied rgb, but the temporal-smoother input did rgb * a again and the no-smoothing path composited with mix(radiance, rgb, a) (another x a). Opaque cores hid it; THIN cloud radiance scaled quadratically with alpha — wisps and edge skirts everywhere rendered darker than intended. Fixed to the plain premultiplied over in both paths.
src/dxvk/rtx_render/rtx_atmosphere.h + cloud_sun_density_grid.comp.slang / cloud_ambient_density_grid.comp.slang - fork-owned change. D_SUN/D_AMBIENT GRID AXIS SWAP: the UVW mapping puts VERTICAL on uvw.y and world Z on uvw.z, but allocation was Y=256/Z=32 — world-Z shadow texels were 375 m wide (the blocky "squares" cast into the deck) while vertical wasted 256 texels on ~12 m. Now X=256, Y(vertical)=32, Z=256; dispatch math derives from the constants, samplers use normalized uvw — constants-only change.
src/dxvk/rtx_render/rtx_atmosphere.cpp - fork-owned change. NIGHT RE-BAKE STORM: normalizeForSkyViewLutKey's comment claimed the star / Milky Way fields were zeroed out of the cache key; no zeroing existed anywhere, so the game-driven per-frame starRotation re-baked the transmittance -> multiscatter -> sky-view cascade AND the voxel grids every frame at night. Zeroing added to normalizeForSkyLutCache (base) so all derived keys inherit.

REV 2 (same day, after the user's sunset test — three regressions/misses):

cloud_nubis3_common.slangh - (a) HF fold + soften range 0.2-2 km -> 0.1-0.6 km: the deck base sits ~1.3 km overhead, so the 2 km range repainted the whole zenith view with folded HF noise + flattened contrast ("noise directly above is bad", part of the shadow-wash read). (b) EDGE WISP CUT rev 2: rev 1 never produced cutouts — its (1-profile)^2 band decayed before the erosion could EXCEED the profile (the requirement for ValueErosion to return a hard 0 that pow-sharpen cannot lift back); now pow(wispy,2) x linear band x 1.5 gain — strand cores cut through at profile ~0.4 where the optical edge lives. (c) D_sun rev 2: the 12 km cap had REMOVED real sunset occlusion ("internal shadow casting gone") — added a 4-tap coarse jittered far tail over [12..40 km]; and the intersect-fail path + a ~1.5-deg grazing ramp now return/blend to fully-occluded OD 60 instead of flipping the whole grid to 0 at sun elevation 0 (the "flashes bright orange then vanishes" discontinuity). Tap jitter hash scale 37.1 -> 173.3 (hash31 int-truncates; neighbor voxels now always decorrelate).

Workstream - Nubis3 detail round: sqrt-adaptive march + spectrum/mean fixes (fork - 2026-07-16)

"How do the papers get cloud detail" round — three pillars from the two reference decks, in priority order. (1) SQRT-ADAPTIVE HYBRID MARCH (Nubis p.172/174): the fixed 100 m step length was the detail CEILING — the detail volume carries sub-200 m content the lattice could never resolve. marchCloudSlab's per-sample body is extracted into integrateCloudSample (the fixed-step loop is preserved bit-identical for legacy / knob-off), and a new variable-step loop marches step = max(SDF x nvdfStepScale, clamp(cloudViewStepKm x sqrt(t / 12 km), floor, 4 x cloudViewStepKm)) — the existing knob keeps its meaning at the tile distance (100 m at 12 km), ~30 m near the deck base, floor (default 25 m) for fly-through. The SDF fold subsumes the Phase C index-skip in this mode. Nubis jitter policy: animated hash < 250 m, static per-pixel hash (FAST-noise frame slice 0) beyond — no distant shimmer; the cursor stays unjittered (p.174). Iterations hard-capped by cloudViewSamplesMax. (2) ZERO-MEAN BAKER NORMALIZATION (GT7 p.228): the detail channels' raw means were MEASURED (numpy Monte-Carlo port of the alligator kernel; curl warp skipped = measure-preserving): wispy 0.7224/0.7267, billowy 0.1555/0.1542 — NOT the ~0.5 / 0.3 the sampler consumers assumed. The baker now recenters every channel to mean 0.5 (pure bias + saturate), making the "zero-mean" wobble actually zero-mean (was pushing wispy silhouettes OUT ~+0.22 x amplitude, billowy IN ~-0.10) and the interior-texture factor exactly density-neutral (was hiding up to ~52% density cut on billowy cores). kChannelMean in the two-scale de-tile collapses to vec4(0.5). (3) MID-OCTAVE SILHOUETTE LOBES (GT7 pp.229-230): the silhouette spectrum was 87-875 m against km-scale smooth bodies — the empty 1-2.5 km octave is GT7's "reads synthetic" gap. A third tap of the same volume at 0.193x base frequency (low channels only, ~2.4 km lobes) folds into the wobble at 60/40; the combined signal stays in [-0.5, 0.5] so the step-5 early-out and conservative step bound hold. (4) ALTITUDE OCTAVE WEIGHTS (GT7 pp.236-237 approx): typeShaped shifts the wispy<->billowy blend by +-0.25 across the slab height band (smoothstep 0.1..0.8) — wispy bases, billowy tops — and feeds every character consumer (erosion composite, HF fold, wobble + mid, interior).

src/dxvk/shaders/rtx/pass/atmosphere/cloud_march_common.slangh - fork-owned change. Per-sample shading extracted to integrateCloudSample (verbatim; returns sampled?/sdfStepKm); marchCloudSlab + marchCloudLayers grow pixelJitterStatic; new adaptive loop gated on nubis3AdaptiveStepKm > 0 (FP stall guard at stepKm <= 1e-5).
src/dxvk/shaders/rtx/pass/atmosphere/cloud_render.comp.slang - fork-owned change. Passes fastJitter(pixelCoord, 0) as the static far-field hash.
src/dxvk/shaders/rtx/pass/atmosphere/cloud_secondary_lut.comp.slang - fork-owned change. Passes its frame-constant jitter for both hashes.
src/dxvk/shaders/rtx/pass/atmosphere/cloud_detail_noise_baker.comp.slang - fork-owned change. kRawChannelMean = vec4(0.7224, 0.7267, 0.1555, 0.1542) bias-recenter to 0.5 (measured via scratchpad measure_channel_means.py — RE-MEASURE if kernel gain/exponent/amp/octaves/transforms change).
src/dxvk/shaders/rtx/pass/atmosphere/cloud_nubis3_common.slangh - fork-owned change. kChannelMean -> vec4(0.5); step 6d mid-octave wobble tap (0.193x, 60/40); typeShaped altitude keying replaces saturate(typeLocal) at all four character consumers.
src/dxvk/shaders/rtx/pass/atmosphere/atmosphere_args.h - fork-touchpoint inline tweak. padReserve2 -> nubis3AdaptiveStepKm. THE LAST reserve pad is now consumed — any further CB growth needs a new full 16-byte row.
src/dxvk/rtx_render/rtx_atmosphere.cpp - fork-owned change. getAtmosphereArgs fills nubis3AdaptiveStepKm (clamped [0..0.2]; stays in the LUT cache keys — same class as nvdfStepScale).
src/dxvk/rtx_render/rtx_options.h - fork-touchpoint inline tweak. Adds rtx.atmosphere.nubis3AdaptiveStepKm (0.025; 0 = legacy fixed-length stepping).
src/dxvk/rtx_render/rtx_fork_atmosphere.cpp - fork-owned addition. "Adaptive Step Floor" slider next to SDF Step Scale.
RtxOptions.md - REGEN PENDING (11 options now outstanding across the Nubis3 workstreams).

Workstream - Legacy cloud model retirement (fork - 2026-07-16)

With the Nubis3 detail round user-validated, the pre-Nubis3 density model is DELETED end to end - the A/B toggle, the legacy noise-threshold view sampler and its whole helper web, the $256^{3}$ Perlin/Worley noise volume + baker, the 64x128 cloud height LUT + baker, four legacy-only look knobs, and their ImGui sliders ("they're making it very hard to tune"). The ECHO DECK (layer 2) was the last legacy consumer - it is CONVERTED to the Nubis3 sampler: the deck now samples the same NVDF/detail model as layer 1 with the density-FIELD position offset by a seed-keyed XZ shift (a decorrelated-but-related copy of the layer-1 bodies at the deck altitude - the original Phase E design), lit by the evaluator core with the analytic dSunProxy and downTau=0. The COLUMN MODEL and placement map SURVIVE - they author the NVDF occupancy bodies - but their view-pass bindings are gone (the placement map is now purely a bake input). Weather-preset ABI: unaffected (audited - no preset field mapped to a deleted option).

DELETED shader files: rtx_cloud_noise_baker.comp.slang, cloud_height_lut_baker.comp.slang. DELETED functions (atmosphere_common.slangh): sampleCloudDensityTextured x3, sampleCloudDensityForShadow x2, sampleCloudPlacement, sampleCloudNoiseHexBlend, sampleTricubicBSpline, cloudHeightProfileFull x2, sampleCloudHeightLUT, cloudHeightProfile, cloudHeightProfileDownIntegral x2, cloudTypeProfile, cloudTypeProfileIntegralFromTop; (cloud_march_common.slangh): sampleDimProfile, sampleCloudSdf, evalNubisCubedSample legacy wrapper. DELETED options: cloudAnvilBias, cloudVerticalStretch, cloudEdgeSoftness, cloudDetailHeightCharacter, cloudWorleyCarveStrength, cloudWorleyFrequency, cloudWorleyOctaves, cloudNoiseBaseFreqScale, cloudHeightLutEnable, nubis3ModelEnable (their CB fields became padRetired0-9 - zero-filled reserve, free for Phase D). RETIRED binding indices: 203 (cloud noise 3D), 216 (view-pass placement map)

annotated do-not-reuse in common_binding_indices.h; BINDING_ATMOSPHERE_MAX repointed to 215.
src/dxvk/shaders/rtx/pass/atmosphere/cloud_march_common.slangh - fork-owned change. marchEchoDeck + sampleCloudSunOpticalDepth_localSlab converted to the Nubis3 sampler (fieldOffsetKm seed shift; SDF index-skip added to the deck loop); integrateCloudSample + sampleCloudSunOpticalDepth_local unbranched; adaptive/skip gates drop the model term.
src/dxvk/shaders/rtx/pass/atmosphere/cloud_nubis3_common.slangh - fork-owned change. sampleCloudBakeDensity unbranched; bake-integral signatures drop the legacy noise/placement params (grid bakers updated).
src/dxvk/shaders/rtx/pass/atmosphere/atmosphere_common.slangh - fork-owned change (deletions above).
src/dxvk/shaders/rtx/pass/atmosphere/cloud_render.comp.slang / cloud_secondary_lut.comp.slang - fork-owned change. Bindings 1/10/11/12 (noise, height LUT + sampler, placement) removed; ATMOSPHERE_CLOUD_HEIGHT_LUT_AVAILABLE gone.
src/dxvk/shaders/rtx/pass/atmosphere/cloud_sun_density_grid.comp.slang / cloud_ambient_density_grid.comp.slang - fork-owned change. Bindings 2/4 removed.
src/dxvk/shaders/rtx/pass/common_bindings.slangh / common_binding_indices.h / atmosphere/atmosphere_bindings.slangh - fork-touchpoint inline tweak. Noise/placement decls + layout-macro rows removed; retired indices annotated.
src/dxvk/shaders/rtx/pass/atmosphere/atmosphere_args.h - fork-touchpoint inline tweak. 10 fields -> padRetired0-9 (layout unchanged).
src/dxvk/rtx_render/rtx_atmosphere.h / rtx_atmosphere.cpp - fork-owned change. m_cloudNoise3D + m_cloudHeightLut resources, bakers, dispatch fns, re-bake gate, cached inputs, binds/tracks: all removed; retired pads zero-filled in getAtmosphereArgs.
src/dxvk/rtx_render/rtx_options.h - fork-touchpoint inline tweak (10 options removed).
src/dxvk/rtx_render/rtx_fork_atmosphere.cpp - fork-owned change. Nubis3 enable checkbox + Anvil Spread / Noise Frequency / Base Wispiness / Edge Softness sliders removed; common-bind of noise/placement removed.
RtxOptions.md - REGEN PENDING (10 options removed + prior additions outstanding).

Workstream - Fine-frequency detail band (fork - 2026-07-16, detail round follow-up)

External feedback on the GT7 low/mid comparison slides: "you might need another higher-frequency noise texture." Spectrum audit agreed - after the detail round the sqrt-adaptive march resolves 25-30 m near the camera but the detail volume's finest content is ~87 m, so the TEXTURE became the detail ceiling within a few km. GT7-style fix (reuse, zero new memory): a THIRD incommensurate tap of the same $128^{3}$ volume at 2.63x (content ~177..33 m), folded into the erosion composite and the interior-texture signal via mean-matched lerps (all channels are mean-0.5 by the bake normalization, so coverage is unchanged). Distance-gated full-inside-2km -> gone-by-8km: sub-pixel at range (would only feed the temporal EMA noise) and keeps the OD bakes camera-independent for free (bake taps pass cameraDistKm = 1e6). No wobble term - the step-5 early-out and conservative step bound are untouched. Placed after the step-7 profile gate so gated-out samples never pay the tap.

src/dxvk/shaders/rtx/pass/atmosphere/cloud_nubis3_common.slangh - fork-owned change. Sampler step 7a (fine tap + erosion fold) + step-8 interior fold.
src/dxvk/shaders/rtx/pass/atmosphere/atmosphere_args.h - fork-touchpoint inline tweak. padRetired1 -> nubis3FineDetailStrength (first reuse of the retired legacy pads).
src/dxvk/rtx_render/rtx_atmosphere.cpp - fork-owned change (args fill, clamp [0..2]).
src/dxvk/rtx_render/rtx_options.h - fork-touchpoint inline tweak. Adds rtx.atmosphere.nubis3FineDetailStrength (1.0; 0 = two-band spectrum).
src/dxvk/rtx_render/rtx_fork_atmosphere.cpp - fork-owned addition. "Fine Detail" slider after HF Detail (Near).
RtxOptions.md - REGEN PENDING.

Workstream - Mid-band shape variety (fork - 2026-07-17)

User reframing of the GT7 low-vs-mid comparison: the mid band belongs in the SHAPE system, not edges/shading - "actually change the SHAPE of all of the clouds ... more varied cloud shapes instead of nice round singular blobs." The existing 2.4 km mid tap gets a DEDICATED level-set displacement on top of the wobble mix: +-0.5 x nubis3ShapeVarietyKm of iso-surface push/pull, deep enough to lobe and fully split km-scale bodies. Zero-mean signal -> coverage-neutral on average; step-5 early-out + conservative step bound budget the extra excursion. Live, no rebake. (Entry backfilled - shipped in c3ba707d.)

src/dxvk/shaders/rtx/pass/atmosphere/cloud_nubis3_common.slangh - fork-owned change. Sampler step 6e (level-set displacement) + widened step-5 excursion bound.
src/dxvk/shaders/rtx/pass/atmosphere/atmosphere_args.h - fork-touchpoint inline tweak. padRetired2 -> nubis3ShapeVarietyKm (second reuse of the retired pads).
src/dxvk/rtx_render/rtx_atmosphere.cpp - fork-owned change (args fill, clamp [0..1.5]).
src/dxvk/rtx_render/rtx_options.h - fork-touchpoint inline tweak. Adds rtx.atmosphere.nubis3ShapeVarietyKm.
src/dxvk/rtx_render/rtx_fork_atmosphere.cpp - fork-owned addition. "Shape Variety" slider above Body Erosion.
RtxOptions.md - REGEN PENDING.

Workstream - Near-field live sun occlusion + mid interior fold (fork - 2026-07-17)

Shape-variety look check: mid-band lobes still read as silhouette recutting ("cloud is a giant blob that's just a slightly different shape now") - no toward/away-camera depth. Root cause: both mid mechanisms move only the SURFACE, and every lighting input is smooth at lobe scale - the D_sun grid bake integrates with ~0.6 km jittered taps + bilinear filtering (lobe-scale self-shadowing is low-passed away) and micro-AO is a nondirectional local-value emboss. Fix = the Nubis Cubed p.129 split ("first light samples live, grid for the far field"), the top open item on the porting-gap list: per lit march sample (density >= 0.05, sun up), 2 live density taps of the displaced field over nubis3SunNearFieldKm + a D_sun grid tap at the range-end offset for the far field (no double count), plumbed through the existing dSunOverride evaluator param - every D_sun consumer (T_primary, body lobe, ambient shadow, warm/cool blend) sharpens together, so each lobe gets its own sunlit face and shadowed crevice. Companion: the step-6d mid tap now also folds into the interior-texture factor (0.35 x saturate(shapeVarietyKm x 2), mean-matched = density-neutral) so lobes differ in optical thickness too - 2.4 km structure is well above the ~0.56 km mean free path, so unlike the fine band it survives to transmittance. Echo deck untouched (its analytic dSunOverride short-circuits the new block).

src/dxvk/shaders/rtx/pass/atmosphere/cloud_march_common.slangh - fork-owned change. integrateCloudSample: near-field live-tap block ahead of the evaluator call.
src/dxvk/shaders/rtx/pass/atmosphere/cloud_nubis3_common.slangh - fork-owned change. Step-8 mid-band interior fold.
src/dxvk/shaders/rtx/pass/atmosphere/atmosphere_args.h - fork-touchpoint inline tweak. padRetired3 -> nubis3SunNearFieldKm (third reuse of the retired pads).
src/dxvk/rtx_render/rtx_atmosphere.cpp - fork-owned change (args fill, clamp [0..3]; padRetired3 zero-fill dropped).
src/dxvk/rtx_render/rtx_options.h - fork-touchpoint inline tweak. Adds rtx.atmosphere.nubis3SunNearFieldKm (1.2 km; 0 = grid only).
src/dxvk/rtx_render/rtx_fork_atmosphere.cpp - fork-owned addition. "Sun Shadow (Near)" slider after Shape Variety.
RtxOptions.md - REGEN PENDING.

Workstream - Nubis3 shipping defaults from the FNV reference tuning (fork - 2026-07-17)

The maintainer's in-game-validated FNV look becomes the shipping defaults: 27 rtx.atmosphere option defaults re-pinned to the reference rtx.conf. Highlights of the new look: the shaped mechanisms carry the detail (shape variety 1.11 km, edge wisp cut 2.28, body erosion 1.5, near-field sun occlusion 3 km) while the uniform ones are off (erosion 0.42, interior texture 0, fine detail 0, silhouette wobble 0); denser water (cloudDensity 4), broken coverage (mean 0.29, spread 0), full-strength ambient shadow + cloud ground shadow, full-res cloud RT (resolution scale 1), echo deck off by default, nvdfNominalCoverage pinned at 0.65 (bodies bake fat, the live level-set offset at 0.2 km/unit separates them - replaces auto-track as the default). sunElevation deliberately NOT pinned (live scene state driven by game wrappers).

src/dxvk/rtx_render/rtx_options.h - fork-touchpoint inline tweak (default values only, no signature/description changes).
RtxOptions.md - REGEN PENDING (defaults column).

Workstream — FSR 3.1 upscaler + frame generation (fork — 2026-07-21)

Adds AMD FSR 3.1 as a new upscaler backend (UpscalerType::FSR), a new frame generation backend (FrameGenerationType::FSR, alongside the existing DLSS-G path), and a standalone RCAS sharpening pass shared by the upscalers that have no sharpener of their own. Links against the prebuilt signed amd_fidelityfx_vk.dll via the submodules/FidelityFX-SDK pin (tag v1.1.4; import lib + ffx-api headers only, no source build).

The FidelityFX DLL is delay-loaded, not statically imported. Every FFX entry point sits behind fork_hooks::fsrRuntimeAvailable(), a latched LoadLibrary("amd_fidelityfx_vk.dll") probe. Without this, an install that ships d3d9.dll but not the 9.3 MB FidelityFX DLL would fail to load d3d9.dll at all and the game would not start; with it, FSR simply reports unsupported. The same probe gates the two Vulkan extensions and the two extra device queues the FFX backend needs, so an install without the DLL gets upstream's instance, device and queue set byte-for-byte.

FSR Frame Generation needs its own swapchain presenter (DxvkFSRFGPresenter, fork-owned, alongside DxvkDLFGPresenter), which required exposing surface-transfer plumbing on the base Vulkan presenter so that switching presenter types at runtime does not destroy and re-create the native window surface.

Fork-owned files (not upstream touches)

src/dxvk/rtx_render/rtx_fork_fsr.cpp/.h — DxvkFSR upscaler backend, plus the fsrRuntimeAvailable / isFsrUpscalerActive / setFsrDownscaleExtent / dispatchFsrUpscale / fsrUpscalingMipBias / fsrJitterSequenceLength hooks.
src/dxvk/rtx_render/rtx_fork_fsr_framegen.cpp/.h — DxvkFSRFrameGen backend and DxvkFSRFGPresenter, plus the isFsrFrameGenEnabled / dispatchFsrFrameGeneration hooks.
src/dxvk/rtx_render/rtx_fork_rcas.cpp/.h — standalone RCAS sharpening pass and the dispatchRcasSharpening hook. Owns the shared sharpening option at rtx.sharpening.sharpness (deliberately not under rtx.fsr: it drives the RCAS pass for DLSS / DLSS-RR / XeSS / TAA-U as well as FSR's internal sharpener, and an FSR-namespaced key silently owning four non-FSR paths would be a trap for anyone later reworking FSR).
src/dxvk/rtx_render/rtx_fork_upscaler_ui.cpp — the FSR upscaler panel, the shared sharpness slider, and the frame-generation backend selector. Kept out of dxvk_imgui.cpp so that ImGUI::showDLFGOptions stays byte-identical to upstream: the selector decides whether the upstream DLSS-G panel is drawn rather than that panel being rewritten to host a second backend.
src/dxvk/shaders/rtx/pass/post_fx/fork_rcas.comp.slang / fork_rcas.h — the RCAS compute pass. fork_-prefixed because post_fx/ is an upstream directory; an unprefixed rcas.* there would collide if NVIDIA ever ships its own RCAS pass (same convention as tonemap/fork_tonemap_operators.slangh).

Upstream files touched

.gitmodules — inline tweak. Adds the submodules/FidelityFX-SDK submodule (GPUOpen-LibrariesAndSDKs/FidelityFX-SDK, pinned to tag v1.1.4). Note this adds ~411 MB to a git clone --recursive.
meson.build (root) — inline tweak. Adds fsr3_dll_path and an install_data rule deploying the prebuilt amd_fidelityfx_vk.dll next to the other upscaler DLLs (NIS/XeSS/DLSS).
src/dxvk/meson.build — inline tweak. Registers the rtx_fork_fsr* / rtx_fork_rcas* / rtx_fork_upscaler_ui.cpp sources inside the existing contiguous rtx_fork_* block; declares fsr3_dep (import library + ffx-api includes) and fsr3_delay_load_args. fsr3_dep is appended to dxvk_deps on its own dxvk_deps += [ fsr3_dep ] line rather than inlined into the dxvk_deps array literal — upstream rewrites that literal (it is currently moving xess_dep under a Windows-on-ARM guard), and an inlined entry turns every such edit into a merge conflict.
src/d3d9/meson.build — inline tweak. Adds a copy_fsr3_dll custom target placing amd_fidelityfx_vk.dll next to d3d9.dll, and adds fsr3_delay_load_args to the d3d9 link args so the DLL is delay-loaded.
src/dxvk/dxvk_adapter.cpp / dxvk_adapter.h — inline tweak. Adds imageAcquire and fsrPresent queue-family/queue-info fields alongside the existing DLFG present queue. The family-selection logic in findQueues and the VK_KHR_GET_MEMORY_REQUIREMENTS_2 request are both gated on fork_hooks::fsrRuntimeAvailable(), so a build/install without the FidelityFX DLL allocates exactly upstream's queue set and asks for exactly upstream's extensions.
src/dxvk/dxvk_instance.cpp — inline tweak. Requests VK_EXT_DEBUG_UTILS (used by the FFX Vulkan backend for object naming), gated on fork_hooks::fsrRuntimeAvailable().
src/dxvk/dxvk_context.cpp / dxvk_context.h — hook. DxvkContext::isFSRFGEnabled() — the DXVK-side mirror of isDLFGEnabled(), consumed by d3d9_swapchain.cpp to decide which presenter to create — is a one-line delegation to fork_hooks::isFsrFrameGenEnabled(m_device.ptr()). The predicate folds in device support, so callers never need a separate "supported" check.
src/dxvk/dxvk_device.cpp / dxvk_device.h — inline tweak. Adds the imageAcquire / fsrPresent device queues mirroring the adapter-level fields; constructs the m_fsr, m_fsrFrameGen, m_rcas Active<> members alongside the existing upscaler members and tears down m_fsrFrameGen in onDestroy.
src/dxvk/dxvk_objects.h — inline tweak. Adds DxvkFSR / DxvkFSRFrameGen / DxvkRCAS forward declarations and metaFSR() / metaFSRFrameGen() / metaRCAS() accessors + backing Active<> members, following the existing metaXeSS() pattern. Forward declarations only — no FFX headers are pulled in here, so the macro-heavy ffx_api chain does not reach every translation unit that includes dxvk_objects.h.
src/dxvk/imgui/dxvk_imgui.cpp — hook (+19 / -3 LOC). An FSR entry in the two upscaler-type combos and in the RtxFramePassStage name map (one line each); the V-Sync guard's DxvkDLFG::enable() test replaced by fork_hooks::anyFrameGenerationEnabled() so it covers both backends; the upscaler switch gains an FSR case that calls fork_hooks::showFsrUpscalerSettings(ctx); fork_hooks::showSharedSharpnessSlider() after the switch; and the whole frame-generation section becomes one fork_hooks::showFrameGenerationOptions(ctx, dlfgSupported) call. dxvk_imgui.h has zero fork delta. Note: ImGUI::showDLFGOptions is left byte-identical to upstream but is no longer called from either menu. Its body is built around a per-backend "Enable DLSS Frame Generation" checkbox, and this fork's frame-generation UI has no such checkbox — the technology dropdown is itself the enable control. Leaving the function untouched-and-unused keeps it at zero conflict surface forever; the tradeoff is that DLFG UI upstream adds to it will not appear here, so it is worth a glance on each NVIDIA rebase.
src/dxvk/imgui/rtx_user_menu.cpp — hook (+11 / -2 LOC). Same shape as the dev menu: an UpscalerType::FSR case delegating to fork_hooks::showFsrUpscalerSettings, fork_hooks::showSharedSharpnessSlider(), and a "Frame Generation Settings" section gated on fork_hooks::anyFrameGenerationSupported(ctx, dlfgSupported) — so it appears on non-DLSS GPUs that can still do FSR FG — whose body is the same single fork_hooks::showFrameGenerationOptions call.
src/dxvk/rtx_render/rtx_camera.cpp — hook. calcPixelJitter also jitters when FSR is active, and takes its sequence length from fork_hooks::fsrJitterSequenceLength() (which returns 0 — "keep your own" — unless FSR is the active upscaler). The formula itself, 8 * upscaleFactor² matching ffxFsr3UpscalerGetJitterPhaseCount, lives in the fork file.
src/dxvk/rtx_render/rtx_context.cpp / rtx_context.h — hook (+11 LOC in the .cpp). Adds InternalUpscaler::FSR and four one-line dispatches: fork_hooks::setFsrDownscaleExtent, isFsrUpscalerActive, dispatchFsrUpscale, dispatchRcasSharpening (called unconditionally after the main upscale; no-ops unless sharpening is non-zero and the active upscaler has no internal sharpener) and dispatchFsrFrameGeneration (called from the same site as dispatchDLFG). The header adds the four matching friend declarations to the existing fork-hook friend block. No FSR type or FFX header reaches rtx_context.cpp.
src/dxvk/rtx_render/rtx_options.h — inline tweak. Adds UpscalerType::FSR, the FrameGenerationType enum (None/DLSS/FSR) + its RTX_OPTION (rtx.frameGenerationType, env DXVK_FRAMEGEN_TYPE), and RtxOptions::isFSREnabled() alongside the existing isXeSSEnabled()-style helpers. Also forward-declares fork_hooks::applyFrameGenerationType (rather than including rtx_fork_hooks.h, which would be circular) and wires it as the option's onChangeCallback: rtx.frameGenerationType is the single enable control for frame generation, and that handler drives rtx.dlfg.enable / rtx.fsrfg.enable from it. Putting the invariant on the option rather than in the menu means it holds for config files, the environment variable and SetConfigVariable, not only while the settings window is open.
src/dxvk/rtx_render/rtx_resources.cpp — hook. The sampler mip-bias helper adds fork_hooks::fsrUpscalingMipBias(m_device), which returns 0 unless FSR is the active upscaler.
src/dxvk/rtx_render/rtx_scene_manager.cpp — hook. getTotalMipBias() / getCalculatedUpscalingMipBias() gain an FSR branch delegating to the same fork_hooks::fsrUpscalingMipBias helper.
src/dxvk/rtx_render/rtx_types.h — inline tweak. Adds RtxFramePassStage::FSR to the frame-pass-stage enum.
src/vulkan/vulkan_presenter.cpp / vulkan_presenter.h — inline tweak. Adds a protected Presenter(..., VkSurfaceKHR existingSurface) constructor overload plus a releaseSurface() accessor, so DxvkFSRFGPresenter can be constructed from an already-live surface — avoiding VK_ERROR_NATIVE_WINDOW_IN_USE_KHR when the swapchain switches presenter type at runtime. Deliberately minimal: this is the deepest layer the fork touches (src/vulkan/ is base DXVK, not NVIDIA-remix code), and there is no way to express a protected-constructor overload as a one-line hook, so every member added here is permanent hand-merge surface. takeSurfaceFrom(), getSurface() and getSwapChain() were part of the original cherry-pick and are not kept: nothing called them.
src/d3d9/d3d9_swapchain.cpp / d3d9_swapchain.h — inline tweak. m_fsrfgPresenter (a third, mutually-exclusive presenter slot alongside m_presenter / m_dlfgPresenter) is created/torn down in CreatePresenter() based on isFSRFGEnabled(); NeedRecreatePresenter() and GetPresenter() consider all three slots. Surface ownership transfers to the new presenter via releaseSurface() plus the protected Presenter constructor (see the vulkan_presenter.* entry); the single destroy-the-old-surface path sits above the presenter if/else rather than being duplicated in two of its three branches. Adds GetActivePresenterOrNull(), a null-tolerant sibling of GetPresenter(), because CreatePresenter() legitimately runs before any presenter exists and GetPresenter() asserts on null — calling it there tripped that assert on startup in debug and debugoptimized builds.

Workstream - sRGB texture linearization fix (fork - 2026-07-21)

FO4 albedo/emissive were double-linearized: the game's diffuse/glow maps arrive in sRGB formats (the sampler decodes them on read) AND the opaque surface shader applied its own gammaToLinear (pow 2.2), so the value was gamma-corrected twice -> washed-out albedo. Fix is a per-material "source texture is sRGB" flag: the shader skips its software gamma correction for a channel whose source texture uses an sRGB VkFormat (the sampler already linearized it), while constants and UNORM textures still get the software conversion. Detection reads the resolved albedo/emissive image-view format in createSurfaceMaterial and self-heals across async texture loads because the material cache key already folds isImageEmpty(). Gated by rtx.linearizeSrgbTextures (default on), folded into preCreationHash so toggling the menu checkbox rebuilds cached materials live. Separately, gammaToLinear now uses the exact sRGB piecewise EOTF instead of x^2.2, so software-linearized inputs (notably albedo the FO4 plugin submits as UNORM) match the hardware sampler's sRGB decode; linearToGamma (output encode) is deliberately left as x^(1/2.2).

src/dxvk/shaders/rtx/utility/shared_constants.h - fork-touchpoint inline tweak. Adds OPAQUE_SURFACE_MATERIAL_FLAG_ALBEDO/EMISSIVE_TEXTURE_IS_SRGB at TYPE_OFFSET(6/7). (Originally 5/6; renumbered on the 2026-07-29 upstream sync — upstream took TYPE_OFFSET(5) for OPAQUE_SURFACE_MATERIAL_FLAG_IS_HAIR_CARD. These offsets are a shared, silently-colliding namespace: check for a free slot on every sync.)
src/dxvk/rtx_render/rtx_texture.h - fork-touchpoint inline tweak. Adds TextureUtils::isSRGB(VkFormat).
src/dxvk/shaders/rtx/concept/surface_material/opaque_surface_material_interaction.slangh - fork-owned change. Skips gammaToLinear for albedo/emissive when the loaded source texture is sRGB.
src/dxvk/rtx_render/rtx_materials.h - fork-owned change. RtOpaqueSurfaceMaterial carries albedoTextureIsSrgb/emissiveTextureIsSrgb (ctor/members/hash/writeGPUData flags); hash static_assert 120 -> 128.
src/dxvk/rtx_render/rtx_scene_manager.cpp - fork-owned change. createSurfaceMaterial detects the sRGB image-view format and folds rtx.linearizeSrgbTextures into preCreationHash for live A/B.
src/dxvk/rtx_render/rtx_options.h - fork-touchpoint inline tweak. Adds rtx.linearizeSrgbTextures (default true).
src/dxvk/imgui/dxvk_imgui.cpp - fork-owned addition. "Linearize sRGB Textures" checkbox under Material Filtering.
src/dxvk/shaders/rtx/utility/color.slangh - fork-owned change. gammaToLinear -> exact sRGB piecewise EOTF (linearToGamma unchanged).
RtxOptions.md - REGEN PENDING (new rtx.linearizeSrgbTextures option).

Workstream - weather precipitation particle system (fork - 2026-07-25)

Rain / snow / blowing sand as a property of a Numos weather preset. Ported from the rain effect in xoxor4d's NFS Carbon RTX mod (https://github.com/xoxor4d/nfsc-rtx, src/comp/modules/rain.cpp), which builds its rain GAME-side: CreateMaterial + CreateMesh for a quad "spawner", then one DrawInstance per frame carrying an InstanceInfoParticleSystemEXT anchored to the camera. This workstream does the same thing from INSIDE the runtime against the same RtxParticleSystemManager, so no game integration has to re-implement it and the values blend with the rest of the weather.

Ten new preset fields (group "Precipitation": intensity, fall speed, wind response, turbulence, drag, streak, drop width/length, opacity, colour) ride the existing WEATHER_PRESET_FIELD_LIST machinery, so per-preset options, blending, the preset editor UI, conf export and tooltips are all generated. They write the live rtx.weather.precipitation.* options that PrecipitationSystem reads. The 12 preset archetypes are authored: dry for clear/partlyCloudy/overcast/hazy/smoggy, mist for foggy, escalating rain for drizzle/rainstorm/thunderstorm, tumbling flakes for snow/blizzard (no streak, high drag + turbulence), and tinted near-horizontal grit for sandstorm.

Implementation notes worth keeping:

No shipped assets. The drop sprite is generated on the CPU at init (64x64 RGBA8 soft radial falloff + box-filtered mip chain, RGB pinned white so the UNORM-vs-sRGB question is moot). With motion trails the runtime stretches only the sprite's centre and preserves its edges, so the same round drop is a capped rain streak with trails on and a snowflake with them off.
Emitter winding is load-bearing. particle_system_spawn.comp.slang derives the emission direction as cross(normalize(p1-p0), normalize(p2-p0)) and uses it UNNORMALIZED to scale initial velocity, so both quad triangles must be wound off perpendicular unit-length edges (indices {0,1,2} and {3,2,1}) or the two halves emit at different speeds.
Drag/gravity are coupled deliberately. The evolve shader integrates v = (v + up*gravity*dt) * (1 - drag*dt), so terminal velocity is gravity/drag. Setting gravity = -fallSpeed * drag makes the spawn velocity also the terminal velocity: turbulence kicks decay back to the fall speed instead of the particles stalling or accelerating. That is what separates snow (drag + turbulence: flutters, then resumes falling) from rain (no drag: dead straight).
Descriptor stability is mandatory, not an optimisation. RtxParticleSystemManager::fetchParticleSystem keys systems on materialHash ^ desc.calcHash() and allocates full maxNumParticles-sized buffers per system. A descriptor recomputed from continuously-blending weather values every frame would therefore allocate a whole particle system per frame, with orphans living until maxTimeToLive. The descriptor is adopted on a dwell timer (descUpdateIntervalMs, default 750 ms) instead, which bounds a transition to a handful of systems; the orphans stop spawning, their particles finish their lifetime, and the manager evicts them - so the step is a crossfade rather than a pop. The material is kept constant (white albedo, colour/opacity ride on the per-particle colour that submitDrawState modulates in via VertexColor0) so it never forks the system identity.
Nothing to reset on scene clear. SceneManager::clear() does not touch the asset replacer's external mesh/material tables, and the material holds its TextureRef (and therefore the Rc<DxvkImageView>) directly rather than resolving through the texture table, so it does not hit the API-uploaded texture failure mode documented in rtx_fork_api_entry.cpp.
src/dxvk/rtx_render/rtx_fork_precipitation.h - fork-owned addition. PrecipitationSystem: the 10 blend-target options + the global budget/spawn-volume/collision options.
src/dxvk/rtx_render/rtx_fork_precipitation.cpp - fork-owned addition. Procedural drop texture, emitter quad registration, parameter resolution, dwell-gated descriptor, per-frame external draw, global ImGui panel, and the three fork_hooks bodies.
src/dxvk/rtx_render/rtx_fork_hooks.h - fork-owned addition. Declares submitPrecipitation and showPrecipitationUI.
src/dxvk/rtx_render/rtx_types.h - fork-touchpoint inline tweak - 2 blocks. Forward-declares fork_hooks::precipitationEmitterDrawCall next to the existing externalDrawTextureCategories decl, and friends it inside DrawCallState. Required because DrawCallState::transformData / ::materialData are private and the emitter builds its draw call from scratch - the same access RemixAPIPrivateAccessor::toRtDrawState has for the API's external draws.
src/dxvk/rtx_render/rtx_context.cpp - fork-touchpoint inline tweak - 4 LOC in injectRTX. Calls fork_hooks::submitPrecipitation(*this) immediately before getSceneManager().prepareSceneData(...). Ordering is required: prepareSceneData is where RtxParticleSystemManager::simulate consumes the frame's spawn contexts.
src/dxvk/rtx_render/rtx_fork_weather.h - fork-owned change. 10 rows appended to WEATHER_PRESET_FIELD_LIST + per-preset values in all 12 WEATHER_PRESET_VALUES_* macros (53 -> 63 fields, 636 -> 756 options).
src/dxvk/rtx_render/rtx_fork_weather.cpp - fork-owned change. snapshotRenderer reads, writeBlendedToDerivedLayer writes (gated as a block on weatherVaries_precipitationIntensity), tooltips mirrored from the live options, and the global panel hosted under the weather tree.
src/dxvk/meson.build - fork-touchpoint inline tweak. Adds the two new source entries.
RtxOptions.md - REGEN PENDING (new rtx.weather.precipitation.* options + 120 new per-preset options).

Fix - precipitation appeared to follow the camera (fork - 2026-07-25)

User report against the initial precipitation drop: "when I turn the camera from left to right, the rain changes the direction it's falling... from left to right. Same with any other camera angle." Three separate camera couplings stacked; the first is the one that made it read as camera-locked at every heading rather than varying with the wind direction.

Billboard type. FaceCamera_Spherical puts the billboard plane in the CAMERA plane (basisRight/basisUp = camera right/up in particle_system_generate_geometry.comp.slang), and the motion trail stretches along the world velocity projected into that plane. Any horizontal component of the fall therefore gets re-projected as the camera yaws, sweeping the apparent fall direction. Streaked precipitation now uses FaceCamera_UpAxisLocked, which pins basisUp to world up and only yaws to face the viewer, so a falling streak stays locked to the world vertical. Un-streaked snow keeps spherical - a round sprite has no orientation to betray and spherical silhouettes better when looking up/down.
Emitter placement read camera orientation. The spawn plane was biased along the camera's flattened forward vector to push the volume into view. That made panning swing the whole volume around the player, and because the spawn shader smears new particles between the emitter's previous and current transform (lerp(worldPosition, prevWorldPosition, randSeed)), a turn dragged freshly spawned drops along the swing. The bias is GONE and the option with it: only the camera's position may move the emitter. A 20 m radius disc centred on the viewer already covers the view in every direction.
Wind coupling was ~3x too strong. windResponse is a fraction of cloudWindSpeed, which defaults to 0.02 km/s = 20 m/s - an ALTITUDE wind. At 0.15-0.30 that produced ~21 deg of slant on rain. The slant is correctly fixed in world space, which is exactly why it re-projects on screen as the camera turns; big enough to notice reads as "the rain changes direction". Retuned so rain drifts 6-9 deg and only blizzard (42 deg) / sandstorm (63 deg) slant hard.

Also added a "Live values" tree to the global panel exposing the per-preset fields directly, so a look can be A/B'd in-game without the blender running.

src/dxvk/rtx_render/rtx_fork_precipitation.cpp - fork-owned change. Billboard selection, orientation-free emitter transform, live-value ImGui tree.
src/dxvk/rtx_render/rtx_fork_precipitation.h - fork-owned change. Dropped forwardOffsetMeters (with a note on why no view-bias option may exist); windResponse default 0.15 -> 0.04 and reworded.
src/dxvk/rtx_render/rtx_fork_weather.h - fork-owned change. Per-preset precipitationWindResponse retuned across all 12 presets.
docs/integrators/weather-presets.md, weather-presets-reference.md - fork-owned change. Orientation-independence + wind-tuning guidance; stale "the plugin owns all particles" / field-count claims corrected.

Fix - precipitation review pass (fork - 2026-07-25, adversarial review)

Verification pass over the precipitation drop against the upstream particle manager/shaders. Confirmed sound: staging-buffer lifetime in ensureTexture (DxvkContext::copyBufferToImage calls trackResource<Read> on the source, so the command list owns the staging buffer until the GPU is done), emitter quad winding (both triangles yield exactly unit local +Z off unit-length edges, and the transform's third column is the unit fall direction), the sceneScale unit conventions, orphan crossfade/eviction (prepareForNextFrame erases a system maxTimeToLive after its last spawn; simulate keeps stepping spawn-less systems), and survival of the external mesh/material tables across SceneManager::clear(). Five real defects found and fixed:

fp16 timeToLive stalls at high refresh rates (upstream bug, exposed by slow snow). GpuParticle::timeToLive was a half; fp16 spacing in [16,32) is 1/64 s, so the evolve shader's timeToLive -= dt rounds back to the same value whenever dt < 1/128 s. At >128 fps every particle with >=16 s of remaining life is IMMORTAL: the conservative counter never decrements, the capacity gate blocks all further spawns, and after the initial population falls out of view precipitation stops entirely. The snow preset's 1.1 m/s fall over a 22 m volume needs ~20 s lifetimes, exactly in the broken range. Promoted to a float by absorbing the two pad halves (struct stays 48 bytes); sentinel + whole-buffer clear word are now float +inf.
Constant-population mode trap. spawnParticles flips into "re-init every particle every frame" when spawnRatePerSecond >= maxNumParticles; the intensity-driven budget/ttl rate crosses that whenever ttl < 1 s (fast fall
- small volume) and would freeze precipitation into a static sheet at the spawn plane. Rate now capped at 0.9x budget.
Orphan pile-up during blends with long lifetimes. Orphaned systems live for their remaining particle lifetime, so a flat 750 ms dwell blending the snow preset (~20 s ttl, ~7.5 MB buffers/system at default budget) stacks ~27 systems (~200 MB transient + 27 simulate groups/frame). Effective dwell now floored at maxTimeToLive/6 (~7 systems worst case); descUpdateIntervalMs = 0 still bypasses everything for debugging.
roughness / metallic were write-only after first init. The material registered once; later edits silently did nothing. refreshDesc now re-registers the material (destroy + register under the same handle) when they change, paced by the same dwell — submitExternalDraw stamps the external material hash onto the emitter draw call via setHashOverride, and that is the hash the particle manager keys systems on, so a material edit forks the system identity exactly like a desc edit (crossfade; dwell prevents per-frame forks while a slider drags).
Weather write gate missed constant-nonzero authoring. Gate was "intensity varies across presets"; a game authoring the SAME nonzero intensity into all 12 presets would never get its live options written and would never rain. Gate is now varies-OR-nonzero (all-zero remains the only dormant case, preserving hand-authored configs).

Also hardened: the singleton now tracks the DxvkDevice* its resources were built on and drops/rebuilds everything if the device changes (previously the registered-flags claimed "done" while a recreated device's asset replacer had never seen the handles -> per-frame "External mesh has no submeshes" forever); stale "cleared by onSceneClear" comment (no such hook exists) replaced with the real lifetime story. Deliberately NOT changed: frame-time-dependent motion-trail length (speed * multiplier * dt is upstream's motion-blur semantic — per-frame streaks tile temporally into continuous lines; compensating would gap them), and no interior suppression runtime-side (no renderer signal for "indoors" exists; documented as the integration's job in weather-presets.md).

src/dxvk/shaders/rtx/pass/particles/particle_system_binding_indices.h - fork-touchpoint inline tweak. GpuParticle.timeToLive half -> float (absorbs pads, 48 bytes preserved); +inf sentinel/clear word.
src/dxvk/rtx_render/rtx_fork_precipitation.cpp - fork-owned change. Spawn-rate cap, lifetime-scaled dwell floor, dwell-paced material re-registration, device-change reset, staging-lifetime note.
src/dxvk/rtx_render/rtx_fork_precipitation.h - fork-owned change. refreshDesc signature (+ctx), m_resourceDevice / applied-material members, corrected lifetime comments + tooltips.
src/dxvk/rtx_render/rtx_fork_weather.cpp - fork-owned change. Precipitation write gate: varies-OR-nonzero.
docs/integrators/weather-presets.md - fork-owned change. Interiors-are-integration-responsibility section, fast-movement note, global-knob table additions.
docs/integrators/weather-presets-reference.md - fork-owned change. Write-gate semantics note.

Workstream - ray-traced spawn occlusion for particles (fork - 2026-07-25)

Precipitation now stops under roofs, bridges and overhangs by asking the actual scene geometry, not the screen. Supersedes the previous "interiors are the integration's responsibility" position in docs/integrators/weather-presets.md.

Motivation: upstream's only particle collision (enableCollisionDetection -> collideParticleWithScene in particle_system_evolve.comp.slang) reprojects the particle into the PREVIOUS FRAME'S G-BUFFER and early-outs if it lands outside [0,1] UV. That is a screen-space test, so it cannot see anything off-screen, behind the camera, or outside the frustum - i.e. it cannot see the roof over the player's head, which is exactly the geometry that decides whether it should be raining on them. This is a ray tracer; the TLAS is right there.

Everything needed was already wired and simply unused:

all three particle passes already declare COMMON_RAYTRACING_BINDINGS (which includes BINDING_ACCELERATION_STRUCTURE), and particle_system_bindings.slangh already includes common_bindings.slangh, so topLevelAS was already visible to the shaders;
khrRayQueryFeatures.rayQuery is already an enabled device feature and RayQuery<> is used elsewhere (visibility.slangh);
the descriptor was never populated for these dispatches only because bindCommonRayTracingResources runs later in the frame with the raytracing work.

Frame ordering makes this cheap rather than awkward: particle simulation runs from SceneManager::prepareSceneData, BEFORE AccelManager::prepareSceneData builds and swaps this frame's TLAS - so getTLAS(Opaque).accelStructure at that point is still last frame's fully-built structure. One frame of staleness is irrelevant for "is there a roof above this raindrop". It is null for the first frames of a scene, hence the sceneTlasValid gate.

Design (reworked 2026-07-25 after the first in-game test): TWO rays per SPAWNED particle (hundreds per frame at default settings, not thousands).

Shelter probe - upward, against the fall direction, 1 km of authored distance (100000 cm * sceneScale). A hit means opaque cover hangs somewhere overhead, so the drop would have been stopped up there and must not exist: it is killed at birth (timeToLive = 0 -> born sleeping, never rendered, retired through the conservative counter exactly like upstream's Kill collision mode).
Landing trace - along the initial velocity for the particle's whole ballistic path, shortening timeToLive so it dies at the surface. Skipped when the shelter probe already killed the drop.

For rain (no drag, no turbulence) the path IS a straight line so the landing trace is exact; heavy turbulence/drag curves it and makes it an approximation. Opaque-only (OBJECT_MASK_OPAQUE + RAY_FLAG_CULL_NON_OPAQUE), so glass and foliage cards do not stop precipitation and traversal needs no any-hit - a verified-safe combination here because fully opaque draws get VK_GEOMETRY_OPAQUE_BIT_KHR + OBJECT_MASK_OPAQUE (rtx_instance_manager.cpp), meaning solid world geometry is never culled by the flag and a single Proceed() completes traversal. Strictly opt-in: the new descriptor bit defaults to 0, so the remixapi, USD and global-preset particle paths are byte-identical to before.

POST-MORTEM of the first version (single downward ray), which shipped and failed in-game ("rain still falls under roofs"): every link was verified correct from source afterwards - the C++/Slang bitfield layouts are byte-identical (proven by SPIR-V disassembly: the desc bit reads byte 95 bit 1 on both sides), the compiled shader contains the full ray-query sequence with flags 0x84 and mask 0x8, the TLAS binding/ordering/lifetime all hold, and FO4 world geometry really is opaque-flagged with the OPAQUE mask. The actual defect was geometric: the ray only looked DOWN from the spawn plane, so any cover ABOVE the spawn plane could never occlude - and in FO4 that is nearly all cover, because the integration runs rtx.sceneScale = 0.1 (meterToWorldUnitScale = 10 units/m) while the game world is ~70 units per real meter, which shrinks the authored 14 m spawn-plane height to ~2 REAL meters above the camera - beneath canopies, overpasses and ceilings. The upward shelter probe closes the hole at any scale (it is also the correct model with an accurate scale: a bridge 30 m up should shelter the street even though the spawn plane hangs at 14 m). Diagnostics were added at the same time so the next failure is a measurement, not archaeology.

Diagnostics (read-only, dev menu -> Weather Presets -> Precipitation (global), visible while occludeUnderCover is on): scene-TLAS availability, per-frame spawn-ray / shelter-kill / landing-hit counters (GPU atomics in the spawn kernel, 10-frame readback ring mirroring ConservativeCounter), and the spawn-plane height in WORLD UNITS next to the sceneScale that produced it - the line that would have exposed this bug on day one. Interpretation: rays = 0 means the trace is not running; rays > 0 with zero kills/hits under a roof means the roof is not reachable in the TLAS under OBJECT_MASK_OPAQUE; shelter kills > 0 under cover means the feature is working.

src/dxvk/shaders/rtx/pass/particles/particle_system_common.h - fork-touchpoint inline tweak. Adds GpuParticleSystemDesc::traceSpawnOcclusion : 1 (default 0 in the ctor) and repurposes the trailing ParticleSystemConstants::pad1 as sceneTlasValid.
src/dxvk/shaders/rtx/pass/particles/particle_system_spawn.comp.slang - fork-owned change. Shelter probe + landing trace at spawn; includes instance_definitions.h for OBJECT_MASK_OPAQUE; SpawnTraceStats diagnostic atomics.
src/dxvk/shaders/rtx/pass/particles/particle_system_binding_indices.h - fork-touchpoint inline tweak. Adds PARTICLE_SYSTEM_BINDING_SPAWN_TRACE_STATS_OUTPUT (63).
src/dxvk/rtx_render/rtx_particle_system.h / .cpp - fork-touchpoint inline tweaks. Binds BINDING_ACCELERATION_STRUCTURE (last frame's Opaque TLAS) in simulate, sets constants.sceneTlasValid in setupConstants, and owns the spawn-trace stats buffer + 10-frame readback ring + getSpawnTraceDiagnostics().
src/dxvk/rtx_render/rtx_fork_precipitation.h / .cpp - fork-owned change. rtx.weather.precipitation.occludeUnderCover (default on) driving the new bit, the dev-menu toggle, and the read-only occlusion diagnostics block.
docs/integrators/weather-presets.md - fork-owned change. "Interiors and cover" replaces "Interiors are the integration's responsibility".
RtxOptions.md - REGEN PENDING.

Fix - sheltered spawns relocate instead of dying (fork - 2026-07-25)

User report against the spawn-occlusion drop above: "when I go into a sheltered area, the rain entirely stops" - including the rain that should stay visible outside, through the doorway. Everything else about occlusion was confirmed working (rain no longer falls through roofs).

Root cause is the interaction of two facts, each fine on its own:

The spawn volume is a camera-centred disc. With the integration's understated scene scale (rtx.sceneScale = 0.1 => 10 units per authored meter, in a world that is really ~70 units/m) the authored 20 m radius / 14 m height is a ~3 REAL meter bubble around the player's head.
The shelter probe killed sheltered drops at birth.

Step under any porch, canopy or awning bigger than the bubble and EVERY spawn candidate is (correctly) judged sheltered, every drop dies at birth, and there is no rain left ANYWHERE - the rain "visible outside" only ever existed inside that same bubble. A pure scale correction (worldUnitsPerMeter = 70) was tried earlier, did not resolve the user-visible symptom, and was reverted at user request; this fix removes the starvation mechanism itself and works at any scene scale.

The fix, in particle_system_spawn.comp.slang: a sheltered spawn candidate REROLLS instead of dying. Spawn-point sampling was factored into sampleSpawnPoint() (triangle pick + barycentric + emitter-motion smear + cone direction, i.e. everything random about a candidate), and the shelter probe loops: draw a candidate, probe upward, and on a hit draw a completely fresh candidate - up to kShelterRelocationAttempts (8) in all - keeping the first one with open sky. Only when every attempt is covered (a genuinely enclosed interior) is the drop killed at birth as before. The landing trace runs on the accepted candidate. Per-attempt RNG seed bases stride by maxNumParticles so a retry can never replay a neighbouring particle's seed window (which would spawn coincident drops).

Consequences, deliberate:

The particle budget migrates to wherever the sky is visible: standing under a roof, rain continues outside the doorway; a 95%-covered volume still lands ~1/3 of its budget in the open 5% (1 - 0. $95^{8}$ ).
Density conservation is traded away: the surviving spawn rate concentrates in the uncovered fraction, so rain framed in a doorway can read up to ~8x denser than open-field rain. That errs on the visible side, which is the point; killing proportionally is exactly the starvation this replaces.
Cost: worst case 9 rays per spawned particle (8 shelter probes + landing), and only in enclosed spaces; the common outdoor case is unchanged at 2.

Diagnostics: a fourth counter, "relocated" (SpawnTraceStats[3]), counts spawns rescued by the reroll. Interpretation shift: under PARTIAL cover (doorway, porch) expect relocated > 0 with shelter kills near ZERO now - shelter kills climbing there means the retries all land under the same cover (volume too small relative to the roof); shelter kills > 0 with relocated == 0 is the signature of a genuinely enclosed space, where killing is correct.

src/dxvk/shaders/rtx/pass/particles/particle_system_spawn.comp.slang - fork-owned change. SpawnSample / sampleSpawnPoint() factoring, shelter relocation loop, kShelterRelocationAttempts, stats[3].
src/dxvk/rtx_render/rtx_particle_system.h / .cpp - fork-touchpoint inline tweaks. kSpawnTraceStatsCount 3 -> 4, s_spawnTraceRelocations, extended SpawnTraceDiagnostics.
src/dxvk/rtx_render/rtx_fork_precipitation.h / .cpp - fork-owned change. occludeUnderCover description now states the relocation semantics; the dev-panel diagnostics show "relocated" and the interpretation comment covers the new counter.

USER-VERIFIED 2026-07-26 ("you fixed it").

Workstream - precipitation appearance overrides (fork - 2026-07-26)

User request: adjustable rain appearance - "not just the roughness or whatever, but the color, length, etc."

The per-preset look fields (color, opacity, drop width/length, streak) have existed since the first drop, but they are OWNED by the WeatherBlender whenever a preset is active: writeBlendedToDerivedLayer rewrites them every frame, so dragging them in the dev menu never sticks. The only look knobs that held were the material constants (roughness / metallic), which ride the registered material rather than the blended options - exactly the user's observation.

Fix: a global appearance-override layer, rtx.weather.precipitation. appearance.*, applied multiplicatively ON TOP of the blended per-preset values inside buildDesc. The blender never touches these, so they stick under any weather, persist across presets and transitions, and save like any other rtx option. All defaults are identity => untouched configs produce bit-identical descriptors (and therefore identical particle-system hashes).

Options: tintColor (multiplies preset color), opacityScale, widthScale, lengthScale, streakScale (0 suppresses trails entirely, also dropping velocity alignment like snow), plus two variance knobs: sizeVariance and opacityVariance (0..0.9). Variance costs nothing: the animation-data texture already carries a min row and a max row per property and the GPU samples between them with each particle's stable randSeed (computeDataRow, randomizeAcrossTwoRows) - upstream plumbing that the precipitation descriptor had been collapsing by writing identical min/max curves. Spreading the curves to [1-v, 1+v] gives every drop its own size / opacity, fixed for its lifetime. Per-drop variation is what breaks up the "sheet of identical sprites" look, especially for snow.

Edits go through the existing descriptor dwell (~750 ms + lifetime floor), so slider drags fork a bounded number of crossfading systems - same contract as weather blends. UI: a new "Appearance overrides (stack on presets)" tree in the Precipitation (global) panel, between Live values and the budget knobs; the Live values note now points at it.

src/dxvk/rtx_render/rtx_fork_precipitation.h - fork-owned change. Seven rtx.weather.precipitation.appearance.* options with rationale block.
src/dxvk/rtx_render/rtx_fork_precipitation.cpp - fork-owned change. buildDesc applies the overrides (streak, size + variance, tint + opacity + variance); Appearance overrides ImGui tree; Live values note.
RtxOptions.md - REGEN PENDING.

Fix - rain always renders dark (fork - 2026-07-26)

User report (with screenshot, FNV overcast): rain streaks read as dark silhouettes against the bright cloud deck, "no matter what tint or color or opacity settings I choose", and the user asked why - "aren't these part of the remix particle system? I thought they'd be able to react to light."

They DO react to light - that is precisely the problem. The drops are fully path-traced lit geometry, and the renderer shades them correctly for what they are: opaque diffuse sprites. Under an overcast preset that kills the sun, a diffuse card receives very little irradiance and legitimately renders darker than the bright sky behind it; albedo, tint and opacity are all MULTIPLIERS on received light, so no setting can lift a surface above the light falling on it. A real raindrop is transparent water that transmits and refracts the bright sky behind and around it (strong forward scattering) - that transmission is why real rain reads as bright streaks against clouds, and it is exactly the term a flat sprite cannot have. Every production rain system substitutes a small self-lit component for it.

Fix: two more appearance options, rtx.weather.precipitation.appearance.glow (emissive intensity, default 2.0, 0 restores the previous scene-lit-only behavior) and .glowColor (default the same cool white-blue as the default drop color). Wired as material emission (setEnableEmission / setEmissiveIntensity / setEmissiveColorConstant - the same mechanism the instance manager uses for emissive-blend particles), so scene lighting still applies ON TOP: sunlit rain still catches real highlights, the glow only stands in for the missing sky transmission. Rides the existing dwell-paced material re-registration (change detection extended from roughness/metallic to glow/glowColor), so live drags fork crossfading systems at the bounded rate. The emissive term is uniform across the sprite; the drop texture's alpha falloff still shapes it at composite time because blending weights the whole contribution by opacity.

Note: emissive intensity is scene-referred radiance, so the right value depends on exposure/time-of-day - it is a taste dial, surfaced in the Appearance overrides tree (Glow Intensity / Glow Color).

REVISED SAME DAY - the constant glow failed its first night test ("rain just glows during the night", user) and was replaced by a sun-elevation-gated version, which the user also rejected in favor of researching the real mechanism. BOTH glow variants are now REMOVED - superseded by the sky-lit particle workstream below, which fixes the actual missing term.

Workstream - sky-lit particles: real skylight for precipitation (fork - 2026-07-26)

The definitive dark-rain fix, replacing the glow/emissive stand-ins after an in-depth trace of how the renderer actually shades the drops.

MECHANISM (all source-verified):

The particle manager stamps InstanceCategories::Particle on its generated geometry -> surface.isParticle.
The primary resolve DIVERTS isParticle surfaces away from normal lit shading into the "opacity lighting approximation" (RESOLVE_OPACITY_LIGHTING_APPROXIMATION, resolve.slangh): emission += albedo x evalVolumetricNEE(froxel radiance cache) - the same treatment as upstream's dust motes.
The froxel radiance cache is filled ONLY by NEE over the analytic light pool. volume_integrate.slangh has NO sky/ambient/environment term: skylight exists in Remix only via path-traced miss rays, which never feed froxels.
Under a Numos overcast preset the sun distant light is heavily dimmed, so froxels in open air are nearly black -> rain = albedo x ~0 = dark silhouettes against a bright sky. No albedo-side knob (tint / opacity / roughness / emissive hacks) can fix a lighting-side hole.

FIX: add the missing skylight term at the same place the froxel term lives. evaluateOpaqueApproximations takes a new defaulted skyLitParticle flag (passed at all three call sites from a new opaque-material flag bit); when set - and only in Numos mode with a nonzero scale - the resolver samples the atmosphere's own sky-view LUT via sampleSkyAmbientForVolume (which folds in the cloud-transmittance LUT: real storm decks dim it, sunsets tint it, night zeroes it) and adds albedo x skyAmbient x cb.particleSkyAmbientScale. Two LUT taps per rain pixel: the view direction lifted above the horizon (forward transmission - the sky behind the drop) and the zenith (ambient), averaged. The froxel term stays, so local lights still light rain at night.

WHY A MATERIAL FLAG AND NOT isParticle: generic game particles (indoor smoke) must not glow sky-blue - sky visibility is unknowable for them. It IS known for precipitation: the spawn-time shelter probe with relocation guarantees every living drop saw open sky at birth. The flag rides the opaque material (sky_lit_particle), set only by the precipitation drop material. Surface-flag bits (data2.w) are full; opaque material flags had bit TYPE_OFFSET(7) free.

COMPILATION SCOPING: the sky term needs the atmosphere helpers + LUT bindings, which the geometry-resolver TUs include before resolve.slangh but other resolve consumers (integrators, visibility) may not. So geometry_resolver.slangh defines RESOLVE_SKY_LIT_PARTICLES right before including resolve.slangh, and the term compiles out everywhere else. This covers the primary view (including ray-query mode, whose variant directive defines ATMOSPHERE_AVAILABLE); indirect bounces keep froxel-only particle lighting, which is fine - the effect matters where the rain is looked at.

The glow options (glow, glowColor) and skyGlowFactor() are REMOVED, replaced by one dial: rtx.weather.precipitation.appearance.skyLightScale (default 1 = physical, 0 = old froxel-only look; a per-frame constant, so it applies immediately with no descriptor/material rebuild). Stale glow keys in saved configs are harmless.

src/dxvk/shaders/rtx/utility/shared_constants.h - fork-touchpoint inline tweak. OPAQUE_SURFACE_MATERIAL_FLAG_SKY_LIT_PARTICLE (TYPE_OFFSET(8)). (Originally 7; renumbered on the 2026-07-29 upstream sync, which pushed the sRGB pair up to 6/7.)
src/dxvk/rtx_render/rtx_material_data.h - fork-touchpoint inline tweak. X(SkyLitParticle, sky_lit_particle, bool, ...) in LIST_OPAQUE_MATERIAL_CONSTANTS.
src/dxvk/rtx_render/rtx_materials.h - fork-touchpoint inline tweaks. RtOpaqueSurfaceMaterial: defaulted ctor param, member, flags pack, hash-struct entry.
src/dxvk/rtx_render/rtx_scene_manager.cpp - fork-touchpoint inline tweak. Passes getSkyLitParticle() into the GPU material.
src/dxvk/shaders/rtx/algorithm/geometry_resolver.slangh - fork-touchpoint inline tweak. Defines RESOLVE_SKY_LIT_PARTICLES before including resolve.slangh.
src/dxvk/shaders/rtx/algorithm/resolve.slangh - fork-touchpoint inline tweaks. skyLitParticle param + sky-ambient term in the opacity lighting approximation; flag passed at all three call sites.
src/dxvk/shaders/rtx/pass/raytrace_args.h - fork-touchpoint inline tweak. particleSkyAmbientScale appended at the END of RaytraceArgs (no existing offsets move).
src/dxvk/rtx_render/rtx_fork_atmosphere.cpp - fork-owned change. Fills constants.particleSkyAmbientScale in updateAtmosphereConstants.
src/dxvk/rtx_render/rtx_fork_precipitation.h / .cpp - fork-owned change. Glow machinery removed; setSkyLitParticle(true) on the drop material.
RtxOptions.md - REGEN PENDING.

ALIGNMENT PASS (same day, user request - "in line with the other Numos features"): the sky-light dial was folded into the weather preset table as an 11TH precipitation field, precipitationSkyLight (group Precipitation -> Look, 0..10, default 1 in all 12 archetypes), driving a per-preset live option rtx.weather.precipitation.skyLight exactly like the other ten - so a blizzard can author different skylight than a drizzle, and the preset editor generates its slider from the same table as everything else. The short-lived global appearance.skyLightScale option was removed. The "Live values (per-preset fields)" tree in Precipitation (global) was also RETIRED in favor of the Weather Preset Editor (a pointer note remains) - preset fields now surface in exactly one place, matching the house pattern. The Appearance overrides tree (tint/scales/variance) stays: it is the one deliberate pattern deviation, kept because blender-owned fields cannot be live-edited mid-storm.

src/dxvk/rtx_render/rtx_fork_weather.h - fork-owned change. precipitationSkyLight field row + 12 preset archetype rows; field counts 10 -> 11.
src/dxvk/rtx_render/rtx_fork_weather.cpp - fork-owned change. Description mirror, snapshot read, derived-layer write for the new field.
src/dxvk/rtx_render/rtx_fork_precipitation.h / .cpp - fork-owned change. skyLight per-preset option replaces appearance.skyLightScale; Live values tree retired.
src/dxvk/rtx_render/rtx_fork_atmosphere.cpp - fork-owned change. cb fill reads skyLight().
src/dxvk/rtx_render/rtx_fork_precipitation.h - fork-owned change. glow / glowColor options; m_materialGlow / m_materialGlowColor registered-state members.
src/dxvk/rtx_render/rtx_fork_precipitation.cpp - fork-owned change. ensureMaterial sets emission; refreshDesc change detection extended; Glow widgets in the Appearance overrides tree.
RtxOptions.md - REGEN PENDING.

Workstream - cloud march hot-path hoists (fork - 2026-07-30, perf)

Pure perf pass over the Nubis Cubed cloud raymarch, prompted by an external review noting the inner loops likely repeat work that belongs outside the hot path. No look change is intended: every edit is either an exact hoist of a loop-invariant subexpression or a provably-conservative skip of dead work. Full audit + the ranked backlog of what is NOT yet done lives outside the repo at Fable_5_testing/cloud-perf-optimization-plan.md.

Four changes:

Ray/frame-constant lighting hoisted out of the per-sample evaluator. CloudShadeContext now has two halves - the original frame-constant fields, plus a ray-constant half filled once per ray by the new cloudShadeContextBeginRay. Moved out of evalNubisCubedSampleCore and the two per-sample moon blocks: the sun dot product, both Henyey-Greenstein lobe evaluations, the sigma_ms sun-dot term, the powder view fade, the underside darkening amount, the sunset ambient ramp, both energy-conserving lobe weights, the micro-AO gain, the ambient-shadow amount, and - the largest - the Wrenninge moon phase-octave loop, a 3-iteration loop over hgPhase() that produced an identical value at every one of the march's 64-96 samples. Two dead exp()s are now also branch-skipped: the sunset cool-blend reach curve above cloudSunsetAmbientRampHighSun (every midday frame) and the underside skyDown when darkening is zero.
Density-sampler tap reorder. The step-6d mid-frequency tap moved ahead of the two-scale base taps so a conservative surface gate can run on it alone. The base pair (two volume fetches plus the erosion composite) used to be paid by every sample in the maxOutwardKm skirt band before the step-7 gate discarded the result - and the SDF sphere-trace parks samples in exactly that band. The gate bounds the unknown base-tap wobble by its construction range [-0.5, 0.5], so it is exact at the shipping default cloudDetailStrength = 0 and strictly conservative above it; returned density is unchanged either way.
cloudPlanetRadius hoisted. It evaluates an exp() and is frame-constant, but computeCloudHeightFraction called it up to five times per march sample (the density tap, two near-field sun taps, two moon taps). New computeCloudHeightFractionR takes the radius as a parameter; the radius lives in CloudShadeContext.cloudRadiusKm.
D_ambient tap skipped when it cannot matter. downTau's only consumer is verticalLight, which is exactly 1 whenever darkenStrength is 0 - the whole night / low-sun band, and any config with cloudBottomDarkening = 0. Saves one 3D texture tap plus an exp() per dense sample there.

src/dxvk/shaders/rtx/pass/atmosphere/cloud_march_common.slangh - fork-owned change. CloudShadeContext moved above the evaluator and split frame-constant / ray-constant; new cloudShadeContextBeginRay called once per ray from marchCloudLayers; evalNubisCubedSampleCore signature takes the context in place of the six view/sun/sky vectors it used to receive; both moon blocks drop their per-sample octave loop; sampleCloudSunOpticalDepth_local / _localSlab take the cloud radius; D_ambient tap gated on ctx.darkenStrength.
src/dxvk/shaders/rtx/pass/atmosphere/cloud_nubis3_common.slangh - fork-owned change. Step-6d mid tap + wobbleMid moved ahead of the two-scale base taps, with the conservative surface gate between them.
src/dxvk/shaders/rtx/pass/atmosphere/atmosphere_common.slangh - fork-owned addition. computeCloudHeightFractionR radius-parameterized core; both existing overloads now delegate to it, unchanged.

No RTX_OPTION, CB, or binding changes - RtxOptions.md regen not required.

Workstream - clouds-off gate on the cloud voxel-grid bakes (fork - 2026-07-30, perf + correctness)

The D_sun / D_ambient voxel grids (256x256x32 voxels, 8 and 6 density taps per voxel respectively) were dispatched every frame regardless of cloudEnabled - measured at ~0.5 ms in the 2026-06-11 sky-perf bisect, which flagged the missing gate as a follow-on that was never done. With clouds off nothing consumes them, so the cost was pure waste, and it silently inflated every "what does the sky alone cost" measurement.

Gating the bake alone would have been a bug: the terrain cloud-shadow consumer was gated only on cloudVoxelShadowsEnable, so a closed gate would have left it sampling whatever the last bake left in the grid. That path turns out to have had a pre-existing correctness bug of the same shape - with cloudEnabled off and cloudVoxelShadowsEnable on, terrain was already being shadowed by clouds that render nothing, while the analytical twin evalCloudGroundShadow had always early-outed on cloudEnabled. Both shadow paths now agree.

The consumer gate lives in the shared _impl rather than at the call sites: there are several consumers (surface NEE, volumetric NEE, the SSS fold, the debug view) and the enum-875 debug view exists specifically to compare the production and debug forms, so they must not diverge.

src/dxvk/rtx_render/rtx_atmosphere.cpp - fork-owned change. computeLuts takes a cloudsEnabled local; the D_sun / D_ambient dispatch condition gains it. While the gate is closed the cached voxel-grid key is zeroed each frame so the frame clouds return forces a fresh bake instead of trusting a key that went stale behind the gate.
src/dxvk/shaders/rtx/pass/atmosphere/atmosphere_common.slangh - fork-owned change. sampleCloudGroundShadow_OptionB_impl early-returns 1.0 when args.cloudEnabled < 0.5, mirroring evalCloudGroundShadow's long-standing gate.

No RTX_OPTION, CB, or binding changes - RtxOptions.md regen not required.

Workstream - contribution-weighted cloud lighting LOD (fork - 2026-07-30, perf)

Ablation measured the near-field live sun refinement (nubis3SunNearFieldKm, shipped at its 3 km cap) at ~1 ms of the 3.3 ms cloud pass - dragging it to 0 recovers the full millisecond but loses the lobe self-shadowing that makes clouds read as volumes. It costs two full density-sampler calls per dense lit sample, roughly 9 of the ~16 3D texture taps.

It was gated on density >= 0.05 alone - nothing about visibility. With cloudDensity = 4 the view transmittance decays geometrically, so a large fraction of dense samples sit deep inside the cloud paying full price while contributing almost nothing to the pixel.

New cloudLightingLodThreshold gates both the near-field sun refinement and the moon shadow march on the sample's ACTUAL contribution weight - view transmittance x aerial haze x its own opacity, which is exactly the factor its color is multiplied by in the composite. Folding aerial haze into the weight also retires distance-dimmed samples automatically, with an error bound, instead of needing a hand-tuned distance cutoff.

Both fallbacks are the ones the existing thin-sample density gates already use (the D_sun grid, and unshadowed moonlight), so this only ever coarsens lighting that was designed to degrade that way - it never removes cloud material. There is no held value or refresh quantum: the weight is a deterministic function of already-jittered quantities, so the switch point moves with the jitter like every other threshold in the march, unlike the caching patterns that caused the posterization family.

Enabling it required hoisting the Beer-Lambert and aerial-perspective factors above the lighting block - an exact reorder, since they depend only on density, segment length and camera distance, all final at that point.

src/dxvk/shaders/rtx/pass/atmosphere/cloud_march_common.slangh - fork-owned change. integrateCloudSample computes effectiveDensity / stepTransmittance / aerialT_haze / aerialT_fade before the lighting and derives sampleWeight + refineLighting from them; the near-field sun gate and the moon shadow gate both take refineLighting. The accumulation at the end reuses the hoisted values.
src/dxvk/shaders/rtx/pass/atmosphere/atmosphere_args.h - fork-owned change. cloudLightingLodThreshold rides the former padRetired6 slot; CB layout unchanged.
src/dxvk/rtx_render/rtx_options.h - fork-owned change. New RTX_OPTION, default 0.02.
src/dxvk/rtx_render/rtx_atmosphere.cpp - fork-owned change. CB fill replaces the padRetired6 zero-fill.
src/dxvk/rtx_render/rtx_fork_atmosphere.cpp - fork-owned change. "Lighting LOD" slider under Clouds > Shape, plus "Sun Shadow (Near)" restored to the panel (the curation pass had demoted it to conf-only as "pinned 3 km"; it is the primary live perf-ablation lever and those have to be ImGui widgets).

RtxOptions.md REGEN PENDING (new rtx.atmosphere.cloudLightingLodThreshold).

Workstream - de-jittered D_sun near section (fork - 2026-07-30, perf)

Attacks the largest measured cost in the cloud pass: the near-field LIVE sun refinement is ~1 ms of the 3.3 ms (nubis3SunNearFieldKm 3 -> 0 recovers exactly that), costing two full density-sampler calls per dense lit sample.

The refinement's own comment justifies it as "the grid bake integrates with ~0.6 km jittered taps and bilinear filtering smooths what survives, so lobe-scale self-shadowing is low-passed away". That is arithmetically false:

the bake already spends 5-8 taps inside the first 3 km (8 on a short sun path, 5 at 0.6 km steps on a long one) - DENSER along the sun ray than the live refinement's 2 taps at 0.75 / 2.25 km;
grid texels are ~47 m horizontal (12 km / 256) and ~95 m vertical, far above Nyquist for the 1-2.5 km lobes in question; trilinear blur is ~1 texel.

The lobe signal is representable in the grid and is not being low-passed away. What differs is estimator COHERENCE. The tap phase is hash31(worldPosYUpKm * 173.3) - decorrelated per voxel, world-anchored, frozen. Sampling 87-875 m detail content at ~0.5 km strata leaves per-voxel OD sigma ~0.15-0.25, and through exp(-sigma * OD) with sigma = 4 that is up to a ~2x STATIC shading swing at 1-2 texel (50-150 m) granularity - worst on lit faces, where low OD makes the exponential most sensitive. The lobe gradients are present but buried under frozen mottle. The live refinement wins by being a DETERMINISTIC estimator anchored to the shade point: its (larger) quadrature bias varies continuously and reads as shading structure, not noise.

Fix: the near section becomes a deterministic fixed-count quadrature (8 taps, phase 0.5, over min(tExit, 3 km)); the far tail keeps its jitter, where undersampling genuinely needs decorrelating and the contribution sits deep enough in the exponential that mottle is invisible. A FIXED count is required - the historical "concentric arcs" artifact came from ceil(tExit / 0.6) jumping between neighbouring voxels, not from determinism, so a fixed near count cannot reintroduce it.

The far section is RE-BASED to partition [nearEnd, tExit] rather than having taps skipped out of the old [0, tExit] lattice. Skipping would leave a jittered partial cell at the handoff whose width varies per voxel - reintroducing exactly the per-voxel variance the near section removes, right at the boundary. Re-based, the tap weights sum to the path length exactly for every voxel (verified for tExit = 1.5 / 3 / 6 / 12 km). Bake tap count goes 8/10/20 -> 8/13/23.

Zero shade-path changes: nubis3SunNearFieldKm = 0 already routes the evaluator to the pure grid tap, so the restored "Sun Shadow (Near)" slider IS the in-game A/B between live and baked near-field. If the grid-only look holds, the knob ships at 0 and the ~1 ms is recovered.

Also fixes the same frozen mottle in terrain cloud shadows, which read the same grid.

src/dxvk/shaders/rtx/pass/atmosphere/cloud_nubis3_common.slangh - fork-owned change. sampleCloudSunOpticalDepthAtWorld splits into a deterministic 8-tap near section and the original jittered scheme re-based onto the remaining path. The old single samples / stepLen lattice is gone. Bake integrand still uses the FULL sampler (coarseOd = false), so render/grid iso-parity is untouched.

OPEN, only if the knob ships at 0: with cloudVoxelShadowsEnable = false the grid falls back to the granularity key, which zeroes the boil/evolution animation fields - so the grid would go stale against animated detail. Either force per-frame bakes when nubis3SunNearFieldKm == 0, or fold a quantized boil phase into the voxel key. (Not a problem in the shipping config: cloudVoxelShadowsEnable defaults true, which makes voxelGridsDirty unconditionally true and the grid rebake every frame.)

Note - cloud march micro-optimizations (fork - 2026-07-30, perf)

Two small changes shipped alongside the cloud perf work above, plus one negative result worth recording so it is not re-attempted:

[unroll] on the moon precompute loop (cloud_march_common.slangh). Its two sibling MAX_MOONS loops were already unrolled; this one was not. Predicted to eliminate the shader's only local-memory arrays (a 256 B MoonParams[4] plus three vec3[4]) by making every index constant. MEASURED: it did not. Re-parsing the rebuilt SPIR-V gave byte-identical accounting (1 function, 309 Function-storage vars, 2,048 B, still one 256 B array[4]). A cross-shader correlation showed the array tracks reading moons[] at all — it is absent from every atmosphere shader that never touches moons, including ones that also copy AtmosphereArgs by value, so it is caused neither by the dynamic index nor by the CB copy. Standing lesson: a Function-storage array in SPIR-V does NOT imply scratch in the final ISA — constant-indexed local arrays are routinely promoted by the driver's own SROA, so SPIR-V local accounting cannot settle occupancy questions. The [unroll] was kept (bit-identical, consistent with its siblings) but is not a win.
frac() dropped on the SDF taps only (cloud_nubis3_common.slangh). tileUV is already fracced and the hex offsets are in [0,1), so those operands are bounded to [0,2) and the sampler's REPEAT mode wraps them identically. Deliberately KEPT on the detail taps: boilPos carries wind/boil/evolution offsets that accumulate without bound over a session, so boilPos * detailFreq can reach the hundreds, and the frac is what keeps the coordinate small before the texture unit converts it to limited-precision fixed point. A comment marks this at the site.

Workstream - retire the live near-field sun path (fork - 2026-07-30, perf)

Follows the de-jittered D_sun near section above. With the bake's near-field estimator made coherent, the grid-only look was compared against the live-refined look in an 11-person vote and PREFERRED - on top of recovering ~1 ms of the 3.3 ms cloud pass. The live path is therefore removed outright rather than defaulted off.

Verified at the IR level: cloud_render.spv static SampleLevel count drops 85 -> 65 (SDF 40 -> 30, detail volume 32 -> 24, D_sun 5 -> 3), exactly the two inlined near-field sampler bodies (fixed-step + adaptive march) plus their two range-end grid taps. Runtime DLL shrinks ~56 KB.

A latent bug had to be fixed first. normalizeForSkyLutCache zeroes cloudBoilPhase / cloudEvolutionOffset* on the stated grounds that they "feed only the view-path cloud taps, not any LUT bake". That holds for the sky LUTs but NOT for the D_sun / D_ambient bakes, whose integrand is the shared density sampler and so reads the animated detail field through boilPos. The voxel-grid cache key therefore never noticed clouds evolving. This was harmless while the live near-field taps re-sampled the animated field every frame; with them gone the grid is the sole source of sun occlusion, and any config running cloudVoxelShadowsEnable = false (which is the only path that consults the granularity key at all) would have lit animating clouds with a frozen shadow field - shadows visibly lagging the clouds they belong to. The animation fields are now restored into the key, quantized on the same km granularity as wind and camera, so staleness stays bounded by one step instead of forcing an unconditional per-frame bake.

Kept deliberately: the coarse OD sampler path (both moon shadow marches still use it) and cloudLightingLodThreshold (now gates only the moon march; still defaults to 0 and is still documented as tested-and-rejected).

src/dxvk/shaders/rtx/pass/atmosphere/cloud_march_common.slangh - fork-owned change. Near-field block deleted from integrateCloudSample; dSunOverride stays -1 so the evaluator takes the pure grid tap. A note at the site records why the path existed, why the bake fix supersedes it, and the warp-uniformity lesson - a contribution-weighted gate on this exact block measured a recovery of precisely zero, because the pass is tap-latency-bound and warp-synchronized, so per-lane conditional work does not convert to time.
src/dxvk/rtx_render/rtx_atmosphere.cpp - fork-owned change. normalizeForVoxelGridKey captures and re-quantizes cloudBoilPhase / cloudEvolutionOffset* after the base normalizer zeroes them. CB fill for the retired option replaced by a pad zero-fill.
src/dxvk/shaders/rtx/pass/atmosphere/atmosphere_args.h - fork-owned change. nubis3SunNearFieldKm -> padRetired12; CB layout unchanged.
src/dxvk/rtx_render/rtx_options.h - fork-owned change. RTX_OPTION removed.
src/dxvk/rtx_render/rtx_fork_atmosphere.cpp - fork-owned change. "Sun Shadow (Near)" slider removed; stale conf-only comment corrected.
src/dxvk/shaders/rtx/pass/atmosphere/atmosphere_common.slangh - fork-owned change. Second half of the voxel terrain-shadow gate: || args.cloudCoverageMean < 0.01f. User-reported: snapping coverage 0.8 -> 0 left terrain shadows behind. Coverage is a live LEVEL-SET OFFSET, so 0 only shrinks bodies by ~130 m at defaults - the density field still contains clouds. What removes them from screen is the render pass's separate hard early-out at effectiveCoverageMean < 0.01, so the D_sun bake kept integrating a field the renderer had stopped drawing. The analytical twin evalCloudGroundShadow always had both halves of this condition; the earlier cloudEnabled fix ported only one. Plain cloudCoverageMean (not the render pass's layer-2 max) is correct here because the D_sun grid is baked over layer 1 only. Distinct from the animation-key bug above: this one fires in ANY config, whereas the staleness bug needs cloudVoxelShadowsEnable = false.

Note: user rtx.conf files carrying rtx.atmosphere.nubis3SunNearFieldKm will log a harmless unknown-option warning.