d4r for Windows / RDNA4
A Windows fork of countervolts/d4r that runs DLSS Super Resolution on AMD Radeon RX 9070 XT / gfx1201 through D3D12, OptiScaler, ZLUDA and HIP. The target models are DLSS 4 preset K and DLSS 4.5 preset M. The runtime runs directly on Windows 11.
Status: development / prerelease. K and M run on the tested RX 9070 XT, including 4K in Silent Hill 2. Improving K performance is the current priority.
Download and run
Prebuilt releases include the Windows shim, patched ZLUDA, isolated HIP runtime, patched OptiScaler, 11 native K kernels, 5 native M kernels and the corresponding committed sources.
You need Windows 11 x64, an RX 9070 XT and a D3D12 game without anti-cheat.
The tested AMD driver is 32.0.31041.1004. Other GPUs and drivers have not
received the same validation.
Supply your own local _nvngx.dll and nvngx_dlss.dll. The tested pair is
NGX 32.0.16.1714 and DLSS 310.9.1.0. The native manifest checks the
DLSS DLL's SHA256. NVIDIA binaries are excluded from the release.
Extract the ZIP and run this command from its directory, replacing the paths with your game and local DLL locations:
.\windows-game.ps1 -GameExe "D:\Games\SILENT HILL 2\SHProto\Binaries\Win64\SHProto-Win64-Shipping.exe" -NgxCore "C:\Users\Administrator\d4r\_nvngx.dll" -DlssDll "C:\Users\Administrator\d4r\nvngx_dlss.dll" -Preset 11 -AsyncInterop
Select an upscaler intercepted by OptiScaler in the game. -Preset 11 selects
K; -Preset 13 selects M. Start M testing without -AsyncInterop: the async
path is currently being tested primarily with K. The first launch may take
time to compile PTX; subsequent launches reuse the cache.
The script installs dxgi.dll, an OptiScaler configuration and a d4r
directory beside the game EXE, backing up replaced files first. The current
frame's inputs and output stay in VRAM. Restore the original files with:
.\windows-game.ps1 -Action restore -GameExe "D:\Games\SILENT HILL 2\SHProto\Binaries\Win64\SHProto-Win64-Shipping.exe"
See docs/windows-game.md for installation options, DLL requirements and restoration details.
Validation
| Check | Result on RX 9070 XT |
|---|---|
| GPU detection | HIP automatically detects gfx1201; D3D12/HIP LUID is checked |
| HIP and CUDA through ZLUDA | Native HIP kernel and CUDA Driver API / PTX tests pass |
| D3D12 ↔ HIP | Shared VRAM buffers and fences; import/release lifetime test passes |
| K transformer | All 11 native layers, NumPy/PTX/replay validation |
| M transformer | All 5 native layers, 40 temporal captures, exact baseline |
| Async K | Queued frames and Release/CreateFeature match synchronous RGB exactly |
| Silent Hill 2 / K / 4K | 16,941 frames checked for NaN/Inf, 203,292 native launches, 0 backend errors |
| Windows hardware tests | All 16 CTest gates pass |
Exact comparisons refer to this project's validated reference/control implementations. Comparison with DLSS on a physical RTX has not been performed. Testing one game does not establish compatibility with every D3D12 game.
Performance and limitations
In the local 4K scene, the user reports approximately 62 FPS / 63% GPU with async interop, compared with 49–51 FPS / 53% in the earlier synchronous run. These measurements also use locally built and validated K output-store kernels. Those objects contain NVIDIA-derived code and are excluded from the public ZIP. Local build and validation commands are in docs/windows-performance.md.
K performance remains under development. These readings are not a controlled
FSR4 comparison. -AsyncInterop is opt-in and requires one D3D12 command queue.
Feature release and resource reconfiguration wait for GPU consumers to finish;
ordinary frame submission does not hold the CUDA mutex.
M uses the validated FP16-equivalent baseline. Native RDNA4 FP8 is separately validated but has not demonstrated a full-network speedup and is disabled by default. Windows preset L is not validated. Frame Generation and DLSS 5 are outside the current scope.
The package uses TheRock 10.2.0a20260929 / HIP 7.17.26386. Stable HIP 7.2
reproduces a mapped external-memory release leak, so the validated package
pins a separate TheRock runtime. GPU architecture is never overridden.
Build from source
git clone https://github.com/xdfnx-dev/d4r.git
cd d4r
Pinned toolchain and source-patch setup is documented in
docs/windows-rdna4-port.md. You need CMake + Ninja,
LLVM/MinGW, a Rust GNU toolchain, TheRock HIP and, for OptiScaler, MSVC v143
and the Windows SDK. Setup scripts use .tools; source dependencies use external.
After preparing the dependencies, run from the repository root:
.\scripts\windows\build-zluda-windows.ps1
.\scripts\windows\build-windows-rdna4.ps1 -RuntimeProfile therock -ZludaRoot "$PWD\dist\zluda-windows-final" -InstallDirectory "$PWD\dist\windows-rdna4-command-list"
.\scripts\windows\stage-native-k.ps1 -DlssDll "$PWD\nvngx_dlss.dll" -PackageRoot "$PWD\dist\windows-rdna4-command-list"
.\scripts\windows\stage-native-m.ps1 -DlssDll "$PWD\nvngx_dlss.dll" -PackageRoot "$PWD\dist\windows-rdna4-command-list"
.\scripts\windows\build-optiscaler-windows.ps1
.\scripts\windows\package-windows-game.ps1 -ArchivePath "$PWD\dist\windows-rdna4-game.zip"
Creating a distributable ZIP requires committed source. The package includes a hash manifest, dependency licenses, source patches and a snapshot of that commit.
Tests and diagnostics
After building, run the hardware gates:
& .\.tools\python\cmake\data\bin\ctest.exe --test-dir build/windows-rdna4-therock --output-on-failure
Run the async K regression with your local NVIDIA DLLs at the repository root:
.\scripts\windows\test-async-k.ps1 -ZludaRoot "$PWD\dist\zluda-windows-final"
Success requires PASS, exact RGB, finite output and passed: true in
validation.json. For a bounded game test, add -RunSeconds 120 -ValidateOutput
to the launch command. The script saves one ZIP containing stdout/stderr,
runtime/driver information, native/translated kernel records and OptiScaler logs.
Use -CaptureExceptions when investigating a crash.
-ProfileStages measures host stages; -ProfileCudaApi measures existing CUDA
API waits; -ProfileGpuBoundary with -AsyncInterop measures D3D12 GPU intervals
around interop. -ProfileKernels synchronizes HIP events and changes scheduling.
Leave profiling and output scans disabled when measuring ordinary FPS.
Documentation and licenses
- Architecture, dependencies, milestones and results.
- Game installation and restoration.
- K/M and FP8 measurements.
- OptiScaler source patches.
- ZLUDA source patches.
This project builds on countervolts' work; Linux upstream code and documentation are retained for compatibility. The fork is not affiliated with NVIDIA or AMD. d4r's license is in LICENSE; dependency licenses are included in the package.