A production-capable spectral Monte Carlo renderer written in Mojo, designed for high-end light transport simulation. Gonzales renders complex scenes — including Disney's Moana Island, all 32 Bitterli benchmark scenes and a 65-scene pbrt/Bitterli corpus — with path tracing, SPPM, BDPT and VCM on the CPU and on NVIDIA and Intel GPUs, plus an à-trous wavelet denoiser.
📖 Read the Gonzales Book for detailed documentation with annotated source code.
The whole Moana island, interactive and fullscreen on an RTX 3060 with 12 GB (video).
Benedikt Bitterli's scenes:
The pbrt-v4 scenes:
All rendered on an Intel Arc B580. The pbrt-v4 scenes and Bitterli's bathrooms, bedroom, cars, classroom and coffee maker: 64 samples per pixel, denoised with Intel Open Image Denoise. The other Bitterli scenes: 64 samples per pixel, or 32 SPPM passes for the two caustic scenes, denoised with Gonzales' own à-trous denoiser.
The renderer is written entirely in Mojo (~54,600 lines in src/gonzales/, plus
~12,700 lines of unit tests) and organized into focused groups of modules:
| Group | Lines | Responsibility |
|---|---|---|
| Shading and BxDFs | 12,700 | Diffuse, coated (pbrt's LayeredBxDF), conductor, dielectric, hair, measured, BSSRDF; spectral upsampling and sensors |
| Path tracer | 9,900 | CPU per-tile wavefront and GPU wavefront kernels, Z-Sobol sampler, film and outputs |
| Scene parsing | 9,400 | pbrt-v4 and Mitsuba parsers, PLY loader (threaded prefetch), materials, lights |
| BDPT / VCM | 6,900 | Bidirectional path tracing and vertex connection and merging (CPU and GPU) |
| Acceleration | 4,600 | SAH BVH2 (GPU) and BVH4 (CPU), instancing, native curve primitives |
| SPPM | 3,700 | Stochastic progressive photon mapping (CPU and GPU) |
| ReSTIR and guiding | 1,900 | Reservoir resampling (DI, GI, volumes, SMS) and path guiding |
| MNEE / SMS | 1,800 | Manifold next-event estimation and specular manifold sampling for glass caustics |
| Media | 1,800 | Homogeneous, uniform-grid, NanoVDB and cloud media |
| Denoising | 650 | GPU à-trous and CPU joint bilateral denoisers |
| Vulkan RT, RT cores, viewer | 460 | Second GPU backend, --rt-hardware, interactive viewer |
External C/C++ libraries (OpenImageIO, Ptex, Vulkan) are called via Mojo's C FFI — not reimplemented.
- Spectral transport — Four hero wavelengths per path in every integrator, pbrt-style RGB-to-spectrum upsampling and sensor models
- Wavefront GPU path tracing — GPU kernels in Mojo with live-path compaction, compile-time kernel variants, deferred software shadow rays and an optional ray-tracing hardware backend (
--rt-hardware): on NVIDIA with its own acceleration-structure builders (no Vulkan or OptiX in that path), on Intel through Level Zero's ray-tracing extension and ray queries - Fast CPU renderer — Per-tile wavefront, 4-wide BVH with triangle records, vectorised shading
- Four light-transport algorithms — Path tracing, SPPM (
--sppm), BDPT/VCM (--vcm) and ReSTIR variants, sharing materials, lights and MIS code - Specular caustics — MNEE and SMS for glass and mirror caustics, with path-identity MIS so the result matches pbrt on the test cubes
- Veach-style MIS — Power heuristic balancing NEE and BSDF sampling; Russian roulette
- Full material set — Coated diffuse/conductor, rough and smooth dielectrics, hair, subsurface, measured BRDFs, alpha cutouts, bump and normal maps, Ptex
- Volumes — Heterogeneous media, NanoVDB, clouds, chromatic media
- Scene formats — PBRT-v4 (including object instancing and curves) and Mitsuba
- Denoising — Variance-adaptive à-trous GPU denoiser (Dammertz 2010) with albedo and normal guides
- Interactive viewer — Vulkan viewer with progressive refinement
make unittest runs ~60 unit-test files concurrently; make smoketest runs an
integrator × feature matrix, and make analytictest checks renders against closed-form
furnace and plane solutions.
| Version | Resolution | SPP | Time | Notes |
|---|---|---|---|---|
| v0.0 (2021) | 2048×858 | 64 | 26h | GCE 8 CPU, 64 GB |
| v0.1 (2023) | 1920×800 | 64 | 78 min | Threadripper 1920X, with Embree |
| v0.2 (2026) | — | — | — | ARC cleanup, Embree removed |
| v0.3 (2026) | — | — | — | Release Notes |
| v0.4 (2026) | 1280×720 | 1/frame | ~7 fps | Whole island on a 12 GB RTX 3060, RT cores, interactive. Release Notes |
Benchmark: Bitterli bathroom scene, 1024×1024, 64 spp, no denoiser. Hardware: AMD Ryzen Threadripper 1920X (12 cores, 24 threads), NVIDIA RTX 3060 12 GB.
| Renderer | Mode | Wall time | Notes |
|---|---|---|---|
| Gonzales | GPU | 4.9s | Wavefront path tracing, software BVH traversal (2.9 s of it rendering) |
| Gonzales | GPU (--rt-hardware) |
6.6s | Own RT-core structures; 3.3 s rendering, the rest is building them |
| pbrt-v4 | GPU (OptiX) | 4.9s | Hardware RT cores |
| Embree pathtracer | CPU | 10.3s / 15.0s | Bare geometry / approximate materials, no textures (earlier measurement) |
| Gonzales | CPU | 24.2s | Full materials and textures (earlier measurement, not re-run) |
| pbrt-v4 | CPU | 76.2s | Full materials and textures |
GPU rows measured on 2026-10-12 (median of three runs). Gonzales GPU is on par with pbrt GPU on this scene. The CPU rows were not all measured on the same day, so they are not directly comparable, and Embree's tutorial path tracer has no textures and only approximate materials.
| Project | Lines (own code) |
|---|---|
| Gonzales | ~66,800 (Mojo, C/C++ and CUDA under src/) |
| pbrt-v4 | ~84,000 (excluding bundled data tables and third-party libs) |
| Embree kernel | ~96,000 (BVH/traversal only, no rendering) |
| Dependency | Description | Install (Arch) |
|---|---|---|
Mojo (via uv) |
Compiler; GPU target requires --target-accelerator sm_XX |
uv run mojo |
| OpenImageIO | EXR/HDR image I/O | yay openimageio |
| Ptex | Per-face texture mapping (Disney) | yay ptex |
| Vulkan | Interactive viewer | yay vulkan-icd-loader |
| Loupe | EXR image viewer (for make view_release) |
yay loupe |
For GPU rendering, set --target-accelerator to match your GPU's compute
capability (e.g. sm_86 for RTX 3060, sm_89 for RTX 4090).
make debug # debug build
make release # optimized release build- Download scenes from Bitterli (PBRT-v4 format) or pbrt-v4-scenes
- Quick test — render and view a Cornell Box:
make view_release
- Or render any scene directly:
.build/release/gonzales path/to/scene.pbrt - Environment switches (caches, memory budgets, logs):
gonzales --list-envlists them; the table issrc/gonzales/core/env.mojo.
Physically Based Rendering: From Theory to Implementation has been an inspiration since the project was called lrt.
© Andreas Wendleder 2019–2026