OpenDLSS: A Vulkan Reimplementation of Nvidia's DLSS 5 Neural Rendering Network

A Vulkan reimplementation of DLSS 5's neural rendering net that matches the original byte for byte

Graphics or inference-infrastructure developers who run transformer inference directly on GPUs or care about FP8 kernel optimization

Byte-for-byte match with the original
- It implements the same 71-block Swin/ViT network as DLSS-NR build 310.8.0.
- All 75 block-boundary intermediates match byte for byte, not just the final image.
A re-rendering network, not an upscaler
- It re-renders the frame the engine already drew, generating detail from injected noise and adjusting tone, structure and skin per a style setting.
- Input and output resolutions are the same.
Inputs: frame, noise, previous frame, scalars
- Inputs are an LDR proxy of the frame, three lanes of Gaussian noise, the reprojected previous output and five conditioning scalars.
- Output is four f32 channels per pixel: an RGB residual and one temporal-blend logit.
1080p takes 7.8ms on an RTX 4070 SUPER
- This is the minimum over 40 frames for the whole network, with 241 dispatches at every resolution.
- 768x768 takes 2.8ms, 1440p 12.6ms and 2160p 29.3ms.

Adopt the fixture-based verification pattern
- When building your own inference kernels, save reference outputs at each block boundary and compare them byte for byte.
- This repo's parity command is an example, and docs/numerics.md describes the exactness contract.
Toggle optimizations off one by one to check exactness
- Use env vars like DLSS5VK_CHAIN=0 or DLSS5VK_UNFUSED=1 to disable optimizations and compare outputs.
- Every switch is designed to keep the output byte-identical.

bit-exactThe result is identical byte for byte, not just approximately equal. It is like a copy whose hash matches the original.

FP8 (E4M3)An 8-bit floating-point format that trades precision for speed and memory savings.

Swin / ViTTransformers that process an image in local windows (Swin) or apply attention across the whole image (ViT).

cooperative matrixA Vulkan extension that lets shaders run matrix multiplication on tensor cores.

PTXAn assembly-like intermediate language for NVIDIA GPUs. This project generates its fast-route kernels with Python.