A C++20 / DirectX 12 game engine with GPU liquid simulation, path-traced lighting and spectral caustics on NVIDIA RTX GPUs.
NVMatrixEngine couples APIC/FLIP fluid simulation with DirectX Raytracing (DXR): a reconstructed liquid surface participates directly in path-traced reflection, refraction and light-side spectral photon mapping. It includes rigid-body gameplay, an interactive settings UI, audio, a staggered MAC pressure grid, anisotropic particle surface reconstruction, sparse procedural water geometry, DLSS Ray Reconstruction, and optional CUDA and ReSTIR PT modes.
Explore Ocean Island, a 256 × 256 metre ocean with broad Hamiltonian swells, a powered boat, an island and pier, and local 3D water simulation around interactions. The indoor Water Lab lets you open a wall inlet, roll or dive as a glass ball, and watch animated water redirect light onto the grid floor. Source, a portable Windows demo, real engine captures and implementation notes are included.
Download v0.1.5 · Engine development · Algorithms · Build · Roadmap
Captured from the v0.1.5 engine at 1600 × 900 with DLSS RR Quality and frame generation off. Capture details.
This independently developed game engine is in preview. The source includes the native engine and the shared code needed by its demos. NVIDIA does not sponsor or endorse this project.
| System | Implemented capability |
|---|---|
| Native rendering | Custom C++20 / DX12 / DXR renderer; camera and photon DispatchRays passes; capability-gated Shader Execution Reordering |
| Light transport | Camera path tracing plus light-side spectral photon tracing; reflection, refraction, Fresnel, total internal reflection, dispersion, and texture-space caustics |
| Water simulation | Affine Particle-In-Cell (APIC), optional Fluid-Implicit-Particle / Particle-In-Cell (FLIP/PIC), staggered marker-and-cell (MAC) velocities, pressure projection, moving solid boundaries, viscosity and surface-tension models |
| Ocean waves and local flow | Nonlinear Hamiltonian spectral waves coupled in both directions to local APIC/FLIP regions; activity fades and unneeded regions retreat as the wake settles |
| Continuous water geometry | Anisotropic particle reconstruction into a sparse brick scalar field; procedural AABB BLAS/TLAS; cell-wise root-refined DXR intersections |
| Water optics | Animated dielectric boundaries, depth-dependent Beer–Lambert absorption, underwater views, and laser single scattering |
| Reconstruction / presentation | DLSS Ray Reconstruction with super resolution, optional DLSS Frame Generation / Multi Frame Generation, and Reflex integration |
| Interaction | Bullet rigid bodies, glass ball sink/float density, GPU-sampled buoyancy, powered boat, wall emitter, bounded foam/bubbles/spray |
| Interface | RmlUi settings, lighting presets, flashlight, lens/view modes, simulation sliders, audio, debug overlays and GPU timings |
| Experimental modes | CUDA/DX12 interop, graph replay, mixed-resolution pressure, live narrow-band particle/grid ownership, adaptive optical sampling, and RTXDI-based ReSTIR PT |
The renderer traces from both the camera and the lights, combining eye-path illumination with light-side photons that have passed through glass or water. This makes difficult refractive caustics practical without asking random camera paths to discover every focused light path.
That is a hybrid, two-sided transport architecture—not a claim that general bidirectional path tracing with vertex connection/MIS, or ReSTIR BDPT, is complete. The optional ReSTIR PT mode is implemented for multi-bounce diffuse indirect lighting from opaque primary surfaces. It does not yet reuse GI behind refractive camera prefixes. Algorithm details and scope.
Engine-wide rendering defaults: fast RIS light sampling, one base camera path, two full light samples, four imported-light candidates, four surface bounces, and Russian roulette. Adjust them under Esc → Path tracing. Settings and prior measurements.
Ocean Island: open Play Extra Large Water Lab.cmd for broad, slow swells, a beach island, palms, pier and a buoyant motorboat. Press E near the boat to board, W/S for thrust, A/D to steer, and Y for day/night. Version 0.1.5 raises the wave-height scale to 1.8 m and removes wavelengths below 48 m from the starting spectrum. Particle contact now accounts for clearance beyond the hull's baked collision volume, avoiding collisions with empty parts of its bounding box. Ocean controls and limits.
Neon Night: launch engine/Play Neon Night.cmd or NVMatrixFluidLab.exe --scene=neon-night for a playable wet neon alley with imported CC0 props, PBR materials, the rolling ball, FPS display and Esc settings. Scene and credits.
The engine integrates real-time liquid rendering, GPU particle-grid simulation, refractive caustics, and interactive rigid-body gameplay. Water participates in reflections, underwater views and light transport. These entry points explain how the engine's simulation and renderer work together.
| Engine subsystem | Start here |
|---|---|
| How do I load glTF, GLB or OBJ objects with PBR textures? | Model loading and material support |
| How are ocean swells coupled to boat wakes and local 3D flow? | Hamiltonian architecture, Ocean Island preset |
| Where is the neon alley scene? | Neon Night: launch, assets and rendering |
| How do APIC/FLIP transfers and incompressible pressure projection map to GPU compute? | Fluid solver architecture, MAC implementation, HLSL kernels |
| How can a ray tracer intersect a particle-reconstructed liquid without meshing? | Anisotropic reconstruction, procedural DXR intersection, scalar-field representation |
| How are dynamic water and glass caustics computed? | Spectral photon transport, photon sampling, water optics |
| Where are path reuse and adaptive ray allocation implemented? | RTXDI ReSTIR PT integration, adaptive optics, current reuse limits |
| What is implemented for CUDA interop and narrow-band FLIP? | CUDA solver, DX12 interoperability, live particle/grid ownership |
| How can results be reproduced or compared? | Build and test instructions, release validation, optimization acceptance criteria |
The sparse rendering field is a reconstructed scalar field, not a guaranteed exact signed-distance field (SDF). The baseline simulation uses a bounded dense grid; sparse surface bricks do not imply a fully sparse fluid solver. CUDA, adaptive pressure and narrow-band modes are experiments, not advertised speedups. The current integrator is camera path tracing plus photon caustics—not a completed ReSTIR BDPT implementation. See the architecture and limitations before using this as a comparison baseline.
To discuss a result or report a numerical/rendering issue, open an issue with the commit/release, GPU, driver, launch arguments and bounded-run JSON. Public visibility is for review; reuse requires the permissions described in COPYRIGHT.md.
Recorded from the running engine at 720p / 30 fps with DLSS RR Quality and frame generation off. The main showcase uses a solid sinking ball, ordinary roll/jump inputs and a following camera. These are visual demonstrations, not FPS benchmarks. Capture setup.
Water room — sinking glass ball rolling, jumping and disturbing the water:
NVMatrixEngine-water-room.mp4
Underwater — first-person view beneath the animated dielectric surface:
NVMatrixEngine-underwater.mp4
Wall inlet — animated fluid refraction and surface interaction:
NVMatrixEngine-wall-inlet.mp4
Water is simulated in three dimensions, not as a displaced plane. Simulation particles carry motion; a MAC grid enforces approximate incompressibility; a reconstructed continuous field is the geometry seen by camera rays, shadow rays, and caustic photons. There is no screen-space water surface or marching-cubes mesh in the primary rendering path.
The default room starts with approximately 100,000 particles. Settings reserve up to 1 million particles, expose MAC cell spacing and simulation frequency, and keep filling bounded: the inlet closes at capacity instead of silently deleting water. Higher capacity is not the same as higher active particle count.
The main release launcher uses the DX12 uniform-grid baseline. CUDA and adaptive modes are opt-in because their additional machinery is not yet a demonstrated end-to-end speedup in this scene. The Large and Ocean launchers use Hamiltonian waves with fixed fluid cell sizes and local 3D activity regions. Optional CUDA/deep-pool experiments remain in the source and are excluded from this portable package.
Deep pool — a larger body of water with calmer interior regions:
NVMatrixEngine-deep-pool.mp4
- Download the ZIP from the v0.1.5 preview release.
- Extract the entire folder to a writable location; do not run inside the ZIP.
- Open Play Extra Large Water Lab.cmd for the ocean boat scene, or Play Water Lab.cmd for the indoor pool. Double-clicking
NVMatrixFluidLab.exeopens the indoor water room. - Press Esc for DLSS RR Quality/Balanced/Performance, frame generation, lighting, lens, buoyancy, water quality, and music settings.
No installer, administrator access, CUDA Toolkit, Visual Studio, or SDK downloads are required to run the packaged demo. A current compatible NVIDIA graphics driver is required. This preview is unsigned; Windows may display a reputation warning. Only use release assets from this repository and verify the accompanying SHA-256 checksum if needed.
| Controls | Action |
|---|---|
| WASD / Space | Move / jump; hold Space underwater to swim upward |
| Right-drag / wheel | Orbit or look / zoom |
| Tab / Ctrl | First-person view / dive |
| T / I | Toggle the inlet / inlet view or ocean overview |
| Y | Toggle ocean day/night |
| E | Interact, grab, or board/exit the boat near it |
| W/S, A/D aboard | Thrust and steering |
| P / period / B | Pause water / single step / reset water |
| Esc / U | Settings / additional HUD detail |
| F8 | Cycle supported frame-generation modes |
The camera defaults to a 90° diagonal FOV, adjustable in settings for both the normal lens and optional equisolid fisheye. The fisheye models a lens projection, not a full compound lens assembly. Both lens modes support frame generation; fisheye supplies DLSS with a cached bidirectional distortion map.
In the indoor labs, the glass ball starts as solid glass in Sink mode (2,500 kg/m³) from physics construction onward. The ocean preset starts with a lighter hollow ball. Esc → Ball buoyancy switches to the lighter hollow Float mode without resetting the scene. The selected mode survives chamber resets; a new session uses its preset’s default.
- Windows 11 x64 and a high-end NVIDIA RTX GPU are the intended platform.
- Development runs use an RTX 5090. The v0.1.5 executable and affected shaders were rebuilt, and the ocean screenshot was captured from the updated engine. Release tests remain skipped at the maintainer's request. Other GPUs are not certified by this preview.
- DLSS feature availability is queried at runtime. Frame generation depends on GPU, driver, OS configuration, and supported presentation mode; it does not accelerate simulation or raw rendering.
- Both normal and fisheye lenses support FG. The HUD reports actual presentations so you can see whether extra frames are being generated. See validation scope.
- This portable release uses DX12 fluid simulation. Optional CUDA experiments remain available in source builds.
- Resolution, active particles, fluid depth, inlet activity, photon budget, and experimental solvers materially affect frame time. There is no blanket 60 FPS guarantee.
The engine reports raw rendering/simulation timings separately from generated presentation frames. See release validation and measurement scope. The project prioritizes reproducible comparisons over multiplying an FPS counter by the frame-generation factor.
The HUD shows Render FPS and DLSS output FPS continuously. Output includes actual generated presentations reported by DLSS; the status says when FG is off, paused, generating, or reporting no extra frames. RR upscaling alone does not add frames. Esc → DLSS & Performance changes the upscaling mode immediately and shows the actual input/output resolution; water keeps its current state.
engine/src/renderer.* DX12 resources, DXR pipelines, frame orchestration
engine/src/fluid/ GPU liquid subsystem, surface, pressure, CUDA, buoyancy
engine/shaders/ HLSL transport, caustics, reconstruction and fluid kernels
engine/src/streamline.* DLSS RR / FG and Reflex integration
engine/src/gameplay.* Demo interaction and rigid-body integration
engine/ui/ RmlUi markup and styling
engine/tests/ Native numerical, gameplay and GPU fixtures
shared/ Minimal common audio, UI, camera and rigid-body support
docs/ Current architecture, build guide, captures and roadmap
release/ Portable launchers and packaging inputs
See BUILDING.md to build the DX12-only or optional CUDA configuration. Algorithm experiments and their evidence are retained in the engine notes; older checkpoint statements describe their date, not necessarily the current integrated feature set.
The next major work is path-tracing optimization and water-simulation optimization: reduce procedural intersection and photon cost, improve stable reuse through refraction, bound adaptive topology work, and make coarse bulk ownership save actual frame time. Full ReSTIR BDPT, fully sparse multiresolution simulation, and general multiple-scattering liquids remain future work, not shipping claims. Detailed roadmap and acceptance criteria.
For a paper, presentation or benchmark discussion, cite the software and identify the exact version/commit and configuration you evaluated. CITATION.cff provides machine-readable metadata for GitHub's citation tools. This is a software preview, not a peer-reviewed publication or reproduction of every referenced paper. Citation does not grant additional source or asset rights.
Built around original engine integration with NVIDIA Streamline/DLSS, NVAPI, RTXDI PT, the DirectX 12 Agility SDK, Bullet, RmlUi, FreeType, and miniaudio. Spectral conversion uses the CIE 1931 standard-observer dataset. The demo includes a custom soundtrack; startup selection is randomized and volume defaults low.
Public source is available for portfolio review; an open-source license has not been granted. See COPYRIGHT.md and third-party notices. Third-party components and the CIE dataset retain their own licenses. SDK caches and development credentials are not included in this repository.
The Large Water Lab keeps the Water Lab scene with four times the floor area and 1.5 m default depth. Launch Play Large Water Lab.cmd.
For the outdoor scene, launch Play Extra Large Water Lab.cmd: a 256 × 256 m Hamiltonian ocean with 6 m offshore depth, a beach island, palms, pier, boat, and HDRI day/night lighting (Y). Controls, physical assumptions and limits.