An open ML temporal upscaler for AMD RDNA2 and RDNA3 GPUs, in the class of FSR 4 and DLSS: jittered low-resolution colour, depth and motion vectors in, anti-aliased high-resolution frames out. The runtime is DirectX 12 compute with INT8 networks (DP4a on RDNA2, wave-matrix on RDNA3) and exposes the FidelityFX API so it can drop into the FSR sample or be injected into games.
The network is Arm Neural Super Sampling
(148K params, ~2.8K MAC per output pixel at 2×), chosen after the review in
docs/Model_Candidates.md. The earlier RDG-based model (CNN with
G-buffer guidance, trained here on GameIR/M3VIR/vKITTI2/Sintel) is kept under RDG/, scripts/
and runtime/models/ as the reference baseline.
- NSS backbone exported to the engine's bundle format with per-layer goldens (fp32 and INT8 QAT); the harness graph for it is written and awaits GPU validation.
- The RDG network runs end to end on the GPU in the harness (bit-exact INT8, WMMA 1×1 conv).
- Not yet: the NSS pre/post passes in HLSL, the DX12 backend for Arm's NSS component, sample integration. Order of work: docs/NSS_Runtime_Plan.md.
RDNU/
├── docs/ # plans and reviews (start with NSS_Runtime_Plan.md)
├── ml/
│ ├── model-gym/ # Arm Neural Graphics Model Gym (train / fine-tune / QAT / export) [submodule]
│ ├── capture-unreal/ # UE 5.5 dataset capture plugin [submodule]
│ ├── models/nss, nfru # released weights (Git LFS) [submodules]
│ └── data/ # datasets, opt-in submodules (see docs/Datasets.md)
├── runtime/
│ ├── harness/ # standalone DX12 engine + kernels, validated layer by layer vs golden
│ ├── tools/ # nss_export.py (NSS bundles), dump_golden.py (RDG bundles)
│ ├── src/, shaders/ # FSR-sample backend and its placeholder shaders (to be replaced)
│ └── external/ # vcpkg, DirectX-Headers, FidelityFX SDK fork, Arm Neural Graphics SDK
├── RDG/, scripts/ # RDG baseline: training code, configs, utilities
└── CMakeLists.txt, CMakePresets.json
git clone --recurse-submodules git@github.com:Adam-AtlasSoftware/RDNU.git # FidelityFX fork needs SSH
cd RDNU
git -C ml/models/nss lfs pull # 15 MB of weightsDatasets are opt-in submodules; fetch instructions and licences: docs/Datasets.md. Windows build (VS Code + CMake, DXC): docs/build-windows.md.
pip install torch numpy # CPU is fine
python runtime/tools/nss_export.py # -> runtime/tools/golden/nss_backbone{,_int8}.rdnut
cmake --build build --target run_engine # runs every bundle through the DX12 enginerun_engine prints max|err| per bundle and per checkpoint; fp32 must be within 1e-3, INT8
bit-exact against the integer-path reference.
- NSS_Runtime_Plan.md: architecture, passes, network contract, INT8 scheme, work list
- NSS_Quality_Plan.md: reaching FSR4-class detail; RDNA2 and RDNA3 tiers
- Model_Candidates.md: the model review that led here
- Datasets.md: sources, licences, fetching
- RDNU_Runtime_Plan.md: engine design (tensor layout, WMMA conv), RDG parts superseded
RDNU is GPL-3.0 (LICENSE). Arm NSS weights: Arm AI Model Community License v1.0 (commercial use permitted; acceptable-use pass-through, notices, marked modifications). Model gym and capture plugin: Apache-2.0. Arm Neural Graphics SDK: MIT. RDG: Apache-2.0. Datasets: see docs/Datasets.md.
Arm (Neural Super Sampling, Model Gym, SDK), the RDG authors, and the creators of the GameIR, M3VIR, MPI-Sintel and Virtual KITTI 2 datasets.