SamG-Coder/OpenDLSS-NR-WebCuda

CUDA reconstruction of OpenDLSS-NR's native implementation, compiled into WebGPU compute shaders by cuda-webshader.
The source repository has no .cu files at the pinned revision. This project reconstructs CUDA from its native GLSL kernels, PTX generators, C++ graph, weight loader, numerical reference, and specifications. No code from ports/browser-webgpu is used.
All inference and frame-processing kernels are authored in kernels/.cu with shared numerical code in kernels/numeric.cuh. JavaScript handles model decoding, graph scheduling, GPU resources, and the browser interface. WGSL is generated, never hand-maintained.
## Status
Performance work parked as of 2026-09-22. The current browser renderer is the accepted baseline. Compiling this project's CUDA kernels with NVCC produced bit-identical network heads and reduced measured time from about 195 to 158 ms at 720p and 417 to 334 ms at 1080p on the tested RTX 5080; browser wall-clock and native GPU timing scopes differ. The remaining large gap to upstream's tensor-core implementation is not explained by WebCuda translation alone. The exact integer-dot prototype was about 14% slower and remains isolated from production. See the native CUDA comparison and integer-dot experiment. Further small tuning gains are not being pursued; revisit for materially better browser matrix-compute capabilities or a demonstrated algorithmic breakthrough. This is a project decision, not a proven hardware ceiling or a measured 1% limit.
Implemented: the full 71-block graph (including block 39's transition-only operation), dense and expert FFNs, 512-channel branch FFNs, global ViT attention, shifted windows, encoder/decoder transitions, preprocessing, motion-based history reprojection, temporal composition, model loading with SHA-256 verification, and a native-fixture parity runner.
This is a correctness-focused backend, with 529 graph dispatches per frame and fused local attention and shape-selected GEMM for packed FP8 matrices. It emulates Ada's grouped F13/F24 accumulation, preserves half and FP8 publication points, and stores published activations as packed FP8/FP16. Prepared execution plans, packed model caching, and working-buffer reuse reduce overhead; rendering is still not realtime. Large resolutions can exceed the device's buffer limit and are rejected before inference.
Real-model browser inference has passed with a locally supplied nvngxdlssnr.dll version 310.8.0.0: all 153 tensors loaded, and all 529 dispatches completed at source resolution on an RTX 5080 in Edge, including 720p and 1080p. Original-capture parity remains unverified; successful inference does not establish image equivalence to NVIDIA's implementation. Model weights and original captures are not distributed.
The browser workspace accepts images or local GLB/glTF objects. Objects are rendered with Three.js into an sRGB display proxy, then processed by the same CUDA-derived NR network. This is a still-frame workflow with an interactive camera; it does not reproduce the original Filament renderer, its HDR display pipeline, or its custom postprocessing styles. Temporal users supply motion vectors; this project does not generate scene motion vectors.