RayD

RayD is a minimalist differentiable ray tracing package wrapping OptiX ray tracing with Dr.Jit autodiff.

pip install rayd

RayD is not a full renderer. It is a thin wrapper around Dr.Jit and OptiX for building your own renderers and simulators.

The goal is simple: expose differentiable ray-mesh intersection on the GPU without bringing in a full graphics framework.

RayD provides three frontends:

• Dr.Jit (Native) — direct Dr.Jit array API, maximum control
• PyTorch — rayd.torch module, CUDA torch.Tensor in/out, integrates with torch.autograd
• Slang — C++ POD/handle bridge for Slang cpp target interop

Why RayD?

RayD is for users who want OptiX acceleration and autodiff, but do not want a full renderer.

Why not Mitsuba? Mitsuba is excellent for graphics rendering, but often too high-level for RF, acoustics, sonar, or custom wave simulation. In those settings, direct access to ray-scene queries and geometry gradients is usually more useful than a full material-light-integrator stack.

RayD keeps only the geometric core:

• differentiable ray-mesh intersection
• scene-level GPU acceleration through OptiX
• edge acceleration structures for nearest-edge queries
• primary-edge sampling support for edge-based gradient terms

For intersection workloads, RayD targets Mitsuba-level performance and matching results with a much smaller API surface.

What RayD Provides

• Mesh: triangle geometry, transforms, UVs, and edge topology
• Scene: a container of meshes plus OptiX acceleration
• scene.intersect(ray): differentiable ray-mesh intersection
• scene.shadow_test(ray): occlusion testing
• scene.nearest_edge(query): nearest-edge queries for points and rays, returning shape_id, mesh-local edge_id, and scene-global global_edge_id
• scene.set_edge_mask(mask) / scene.edge_mask(): scene-global filtering for the secondary-edge BVH used by nearest_edge(...)
• edge acceleration data that is useful for edge sampling and edge diffraction methods

Performance

The chart below was generated on March 25, 2026 on an NVIDIA GeForce RTX 5080 and AMD Ryzen 7 9800X3D, comparing RayD (0.1.2) against Mitsuba 3.8.0 with the cuda_ad_rgb variant.

Raw benchmark data is stored in docs/performance_benchmark.json.

• RayD is consistently faster on static forward and static gradient workloads across all three scene sizes.
• Dynamic reduced forward reaches parity or better from the medium scene onward, and dynamic full is effectively tied on the largest case.
• On the largest 192x192 mesh / 384x384 ray benchmark, RayD vs Mitsuba average latency in milliseconds is: static full 0.162 vs 0.190, static reduced 0.124 vs 0.224, dynamic full 0.741 vs 0.740, dynamic reduced 0.689 vs 0.714, gradient static 0.411 vs 0.757, gradient dynamic 1.324 vs 1.413.
• Correctness stayed aligned throughout the sweep: forward mismatch counts remained 0, and the largest static gradient discrepancy was 9.54e-7.

Quick Examples

If you only want to see the package in action, start here:

• examples/basics/ray_mesh_intersection.py: custom rays against a mesh
• examples/basics/nearest_edge_query.py: nearest-edge queries
• examples/basics/camera_edge_sampling_gradient.py: camera-driven edge-sampling gradients
• docs/slang_interop.md: Slang cpp target interop for host-side RayD scene queries

Differentiable Cornell Box with Edge Sampling

GPU path tracing + interior AD + edge sampling (Li et al.) in ~180 lines (examples/renderer/cornell_box.py):

Differentiable Radio Frequnecy Wave Propagation

(Mini-Differentiable-RF-Digital-Twin):

Minimal Differentiable Ray Tracing Example

The example below traces a single ray against one triangle and backpropagates the hit distance to the vertex positions.

import rayd as rd
import drjit as dr


mesh = rd.Mesh(
    dr.cuda.Array3f([0.0, 1.0, 0.0],
                    [0.0, 0.0, 1.0],
                    [0.0, 0.0, 0.0]),
    dr.cuda.Array3i([0], [1], [2]),
)

verts = dr.cuda.ad.Array3f(
    [0.0, 1.0, 0.0],
    [0.0, 0.0, 1.0],
    [0.0, 0.0, 0.0],
)
dr.enable_grad(verts)

mesh.vertex_positions = verts

scene = rd.Scene()
scene.add_mesh(mesh)
scene.build()

ray = rd.Ray(
    dr.cuda.ad.Array3f([0.25], [0.25], [-1.0]),
    dr.cuda.ad.Array3f([0.0], [0.0], [1.0]),
)

its = scene.intersect(ray)
loss = dr.sum(its.t)
dr.backward(loss)

print("t =", its.t)
print("grad z =", dr.grad(verts)[2])

This is the core RayD workflow. Replace the single ray with your own batched rays, RF paths, acoustic paths, or edge-based objectives.

PyTorch Frontend

rayd.torch is an optional Python-level wrapper that mirrors the native API using CUDA torch.Tensor inputs and outputs. AD mode is inferred automatically from requires_grad.

import rayd.torch as rt

verts = torch.tensor([...], device="cuda", requires_grad=True)
mesh = rt.Mesh(verts, faces)
scene = rt.Scene()
scene.add_mesh(mesh)
scene.build()

its = scene.intersect(rt.Ray(origins, directions))
loss = (its.t - target).pow(2).mean()
loss.backward()  # gradients flow to verts

Key conventions:

• vectors use shape (N, 3) or (N, 2); (3,) and (2,) are accepted as batch size 1
• index tensors use shape (F, 3); images use shape (H, W); transforms use shape (4, 4)
• CPU tensors are rejected; rayd.torch does not do implicit device transfers

The native Dr.Jit API remains unchanged and does not depend on PyTorch.

Device Selection

RayD follows Dr.Jit's current-thread CUDA device selection. If you need to choose a GPU explicitly, do it before constructing any RayD resources:

import rayd as rd

rd.set_device(0)  # also initializes OptiX on that device by default

rd.set_device() / rayd.torch.set_device() are intended for selecting the device up front. Existing RayD scenes, OptiX pipelines, and BVHs should not be reused across device switches in the same process.

Slang Frontend

RayD ships a Slang interop layer for Slang's cpp target. Slang code can import rayd_slang; and call RayD scene queries directly.

NearestPointEdge and NearestRayEdge returned through Slang include global_edge_id in the same scene-global index space as scene.edge_info().global_edge_id. The Slang bridge also exposes scene edge-mask helpers for host code: sceneEdgeCount(scene), sceneEdgeMaskValue(scene, index), and sceneSetEdgeMask(scene, maskPtr, count).

Minimal Slang Example

import rayd_slang;

export float traceRayT(uint64_t sceneHandle,
                       float ox, float oy, float oz,
                       float dx, float dy, float dz)
{
    SceneHandle scene = makeSceneHandle(sceneHandle);
    Ray ray = makeRay(float3(ox, oy, oz), float3(dx, dy, dz));
    Intersection hit = sceneIntersect(scene, ray);
    return itsT(hit);  // use accessor, not hit.t
}

Load and call from Python:

import rayd as rd
import rayd.slang as rs

m = rs.load_module("my_shader.slang")  # use rayd.slang.load_module, not slangtorch.loadModule

scene = rd.Scene()
scene.add_mesh(mesh)
scene.build()

t = m.traceRayT(scene.slang_handle, 0.25, 0.25, -1.0, 0.0, 0.0, 1.0)

Differentiable Slang Example

sceneIntersectAD returns an IntersectionAD with analytic gradients dt_do (∂t/∂origin) and dt_dd (∂t/∂direction):

import rayd_slang;

export IntersectionAD traceAD(uint64_t sceneHandle,
                              float ox, float oy, float oz,
                              float dx, float dy, float dz)
{
    SceneHandle scene = makeSceneHandle(sceneHandle);
    Ray ray = makeRay(float3(ox, oy, oz), float3(dx, dy, dz));
    return sceneIntersectAD(scene, ray);
}

Use it from Python with torch.autograd:

import torch
import rayd as rd
import rayd.slang as rs

m = rs.load_module("my_shader.slang")
scene = rd.Scene()
scene.add_mesh(mesh)
scene.build()

class DiffTrace(torch.autograd.Function):
    @staticmethod
    def forward(ctx, oz):
        ctx.save_for_backward(oz)
        hit = m.traceAD(scene.slang_handle, 0.25, 0.25, oz.item(), 0, 0, 1)
        return torch.tensor(hit.t, device=oz.device)

    @staticmethod
    def backward(ctx, g):
        oz, = ctx.saved_tensors
        hit = m.traceAD(scene.slang_handle, 0.25, 0.25, oz.item(), 0, 0, 1)
        return torch.tensor(hit.dt_do.z * g.item(), device=oz.device)

oz = torch.tensor(-1.0, device="cuda", requires_grad=True)
t = DiffTrace.apply(oz)
t.backward()
print(f"t={t.item()}, dt/doz={oz.grad.item()}")  # t=1.0, dt/doz=-1.0

load_module() runs slangc -target cpp, auto-generates pybind11 bindings, and links against rayd_core. See docs/slang_interop.md for the full compilation pipeline, API reference, and known workarounds.

Edge Acceleration Structure

RayD also provides a scene-level edge acceleration structure.

This is useful for:

• edge sampling
• nearest-edge queries
• visibility-boundary terms
• geometric edge diffraction models

Scene.set_edge_mask(mask) filters this secondary-edge BVH in scene-global edge index space. It does not modify scene.edge_info(), scene.edge_topology(), scene.mesh_edge_offsets(), or primary-edge camera sampling.

In other words, RayD is not limited to triangle hits. It also gives you direct access to edge-level geometry queries, which are important in many non-graphics simulators.

Compiling Locally

RayD is a Python package with a C++/CUDA extension.

You need Python >=3.10, CUDA Toolkit >=11.0, CMake, a C++17 compiler, drjit>=1.2.0, nanobind==2.9.2, and scikit-build-core.

On Windows, use Visual Studio 2022 with Desktop C++ tools. On Linux, use GCC or Clang with C++17 support.

Recommended environment

conda create -n myenv python=3.10 -y
conda activate myenv
python -m pip install -U pip setuptools wheel
python -m pip install cmake scikit-build-core nanobind==2.9.2
python -m pip install "drjit>=1.2.0"

Install

conda activate myenv
python -m pip install .

Dependencies

RayD depends on:

• Python 3.10+
• Dr.Jit 1.2.0+
• OptiX 8+

RayD does not include:

• BSDFs
• emitters
• integrators
• scene loaders
• image I/O
• path tracing infrastructure

That is by design.

Repository Layout

• rayd/: Python package (flat layout)
• rayd/torch/: PyTorch frontend
• rayd/slang/: Slang / slangtorch interop utilities
• include/rayd/: public C++ headers
• src/: C++ and CUDA implementation
• src/rayd.cpp: Python bindings
• include/rayd/slang/interop.h: C++ POD/handle bridge for Slang
• include/rayd/slang/rayd.slang: Slang declarations for the C++ interop layer
• examples/: basic, renderer, and Slang examples
• tests/drjit/: Dr.Jit native geometry tests
• tests/torch/: PyTorch frontend tests
• tests/slang/: Slang interop and gradient tests
• docs/api_reference.md: Python API reference
• docs/slang_interop.md: Slang interop notes and examples

Testing

python -m unittest tests.drjit.test_geometry -v

Optional PyTorch wrapper tests:

python -m unittest tests.torch.test_geometry -v

Optional Slang interop and gradient tests (requires slangtorch):

python -m unittest tests.slang.test_slang -v

Credits

RayD is developed with reference to:

• psdr-jit
• redner
• mitsuba3

Citation

@inproceedings{chen2026rfdt,
  title     = {Physically Accurate Differentiable Inverse Rendering
               for Radio Frequency Digital Twin},
  author    = {Chen, Xingyu and Zhang, Xinyu and Zheng, Kai and
               Fang, Xinmin and Li, Tzu-Mao and Lu, Chris Xiaoxuan
               and Li, Zhengxiong},
  booktitle = {Proceedings of the 32nd Annual International Conference
               on Mobile Computing and Networking (MobiCom)},
  year      = {2026},
  doi       = {10.1145/3795866.3796686},
  publisher = {ACM},
  address   = {Austin, TX, USA},
}

License

BSD 3-Clause. See LICENSE.