rayd-drjit 0.6.0

RayD: Dr.Jit-native GPU primitives for differentiable ray geometry and multipath simulation.

RayD Dr.Jit

RayD is a Dr.Jit-native GPU library for differentiable ray geometry and multipath simulation primitives built on OptiX.

pip install rayd-drjit

RayD is not a full renderer. It exposes low-level scene, ray, edge, visibility, and reflection-path queries for building custom renderers, RF simulators, acoustic tools, and inverse-design systems without adopting a material-light-integrator framework.

Release compatibility

Release wheels cover Linux x86-64 and Windows x86-64 on CPython 3.10-3.14. They are built with CUDA 12.8 and contain native code for sm_70, sm_75, sm_80, sm_86, sm_89, sm_90, sm_100, sm_101, and sm_120, plus compute_120 PTX for forward compatibility. This spans RTX 2080-class Turing GPUs through current data-center and GeForce/RTX PRO Blackwell families.

RayD is Dr.Jit-native and does not depend on PyTorch. Because its nanobind extension uses the CPython ABI, CI builds one wheel per Python version rather than one Python-independent wheel. The complete build matrix and release configuration are documented in CI_BUILD_MATRIX.md.

Scope

RayD focuses on geometry and wave-propagation primitives:

• differentiable ray-mesh intersection
• scene-level GPU acceleration through OptiX
• nearest-edge queries through a scene-global edge BVH
• primary-edge sampling support for edge-based gradient terms
• segment visibility and multipath reflection primitives
• Dr.Jit arrays and Dr.Jit autodiff as the public Python API

RayD intentionally does not provide:

• a material or BSDF system
• emitters
• integrators
• scene loading
• image I/O
• a complete path-tracing runtime
• alternate Python frontend wrappers

Why RayD?

RayD is for users who need fast differentiable geometry queries, but do not want a full rendering framework.

Mitsuba is excellent for physically based rendering, but it can be too high-level when the main workload is RF propagation, acoustics, sonar, visibility analysis, or custom wave simulation. In those settings, direct control over ray-scene queries, edge queries, reflection paths, and geometry gradients is often more useful than a complete renderer.

RayD keeps the API surface small: meshes, scenes, rays, intersections, edges, and multipath query results.

Core API

• Mesh: triangle geometry, transforms, UVs, and edge topology
• Scene: a container of meshes plus OptiX acceleration structures
• scene.intersect(ray): differentiable ray-mesh intersection
• scene.shadow_test(ray): occlusion testing
• scene.nearest_edge(query): nearest-edge queries for points and rays
• scene.set_edge_mask(mask) / scene.edge_mask(): scene-global filtering for secondary-edge queries
• scene.trace_reflections(...): specular reflection-path tracing
• scene.visible(...): batched segment visibility queries
• scene.trace_refl_epc(...): equivalent-path correction primitives for reflection paths
• scene.accum_dfr_direct(...) / scene.accum_dfr(...): native diffraction grid accumulation
• scene.trace_dfr_paths(...): compact native first-order diffraction path export

Feature Overview

Each core query, what it computes, its input/output, and how it behaves under Dr.Jit autodiff (AD):

API	What it computes	Input -> Output	AD
scene.intersect(ray)	Closest-hit ray-mesh intersection	rays -> Intersection (t, point, normal, uv, ids)	AD
scene.shadow_test(ray)	Any-hit occlusion test	rays -> boolean mask	Boolean
scene.nearest_edge(point)	Nearest scene edge to each point	points -> NearestPointEdge (distance, points, edge id)	AD
scene.nearest_edge(ray)	Nearest scene edge to each ray (segment on [0, tmax] when tmax is finite)	rays -> NearestRayEdge	AD
scene.nearest_edges(point, k)	k nearest scene edges per point (k <= 16)	points -> NearestEdgesTopK	AD
scene.visible(start, end)	Mutual visibility between two segment endpoints	endpoints -> SegmentVisibility	Boolean
scene.visible_pair / visible_chain / visible_edge	Shared-origin pair, polyline-chain, and edge-sample visibility	segments -> segment/chain/edge visibility	Boolean
scene.trace_reflections(ray, max_bounces)	Specular reflection paths with image sources	rays -> ReflectionChain (hit points, normals, image sources, ids)	AD
scene.trace_refl_epc(ray, receiver, ...)	Equivalent-path-correction reflection toward a receiver	rays + receiver -> ReflEpc	Detached
scene.trace_refl_epc_field(tx, receiver, ...)	EPC trace returning the complex reflected field	tx position + receiver + ReflEpcFieldOptions(AD) -> ReflEpcField (complex E-field)	Native AD
scene.accumulate_reflections(...)	Accumulate reflected field/power onto a grid	rays + grid + material -> AccumResult	Native AD
scene.accum_dfr_direct(...)	Native direct/Keller/suffix diffraction accumulation onto a grid	DfrStates + DfrGrid + DfrMaterial -> DfrAccum	Native AD for direct, Keller, and suffix-reflection order-1, including suffix reflector mesh-vertex gradients; detached for unsupported AD strategies
scene.accum_dfr(...)	Native order-2/3 diffraction-chain accumulation onto a grid	initial/recursive DfrStates + grid + material -> DfrAccum	Native AD for direct, Keller, and suffix-reflection order-2/3 chains, including suffix reflector mesh-vertex gradients
scene.trace_dfr_paths(...)	Compact first-order diffraction path export	tx/rx + DfrStates + DfrMaterial -> DfrPaths	AD

AD legend:

• AD - differentiable geometry: geometric outputs carry Dr.Jit gradients with respect to mesh vertices and transforms; the discrete hit/edge/path selection runs detached.
• Native AD - the native RayD multipath implementation preserves Dr.Jit gradients for continuous inputs passed as AD arrays and supplies RayD-side derivative code. Diffraction grid accumulation uses CUDA custom-op JVP/VJP over a fixed OptiX forward tape for direct, Keller, and suffix-reflection order-1 and order-2/3 chain accumulation. Discrete visibility/path-selection decisions remain detached.
• Boolean - returns occlusion/visibility booleans and is not differentiable.
• Detached - native fast path: runs detached only and rejects AD inputs.

Trace vs. accum:

• trace_* APIs enumerate geometric or field paths and return per-ray/per-path records. They do not reduce contributions into receiver grids.
• accum_* APIs launch native accumulation kernels that reduce contributions into aggregate outputs, usually grid cells, with atomic or tiled writes. They are the high-throughput path for radiomap-style workloads when individual path records are not needed.

Minimal Example

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

import drjit as dr
import rayd.drjit as rd


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.RayAD(
    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])

Edge Queries

RayD provides a scene-level edge acceleration structure for nearest-edge and edge-sampling workloads.

This is useful for:

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

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

Multipath Queries

RayD includes low-level reflection and visibility primitives for custom wave simulators:

• trace_reflections(...) for specular reflection chains
• visible(...), visible_pair(...), visible_chain(...), and visible_edge(...) for segment and polyline-chain visibility
• trace_refl_epc(...) and trace_refl_epc_field(...) for equivalent-path correction workflows
• accumulate_reflections(...) for reflection grid accumulation workloads
• accum_dfr_direct(...), accum_dfr(...), trace_dfr_paths(...) for native diffraction kernels

These APIs expose primitives, not a complete propagation simulator. Callers own the source model, receiver model, material policy, objective, and optimization loop.

Naming rule: use Dfr for diffraction (DfrStates, DfrAccum, accum_dfr_direct), Refl for reflection-specific short names, keep Epc in equivalent-path-correction APIs, and reserve AD for automatic differentiation. See API_NAMING_STANDARD.md.

Examples

• examples/basics/ray_mesh_intersection.py: custom rays against a mesh
• examples/basics/nearest_edge_query.py: nearest-edge queries
• examples/renderer/cornell_box.py: a compact renderer built on RayD primitives

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.

Device Selection

RayD follows Dr.Jit's current-thread CUDA device selection. Choose a GPU before constructing RayD resources:

import rayd.drjit as rd

rd.set_device(0)

Existing RayD scenes, OptiX pipelines, and BVHs should not be reused across device switches in the same process.

Building 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==1.3.1, 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==1.3.1

Install from the repository root:

python -m pip install .

For editable development builds:

python -m pip install --no-build-isolation -ve .

Repository Layout

• rayd/: Python package and native extension loader
• include/rayd/: public C++ headers
• include/rayd/edge/: edge-query data structures
• include/rayd/multipath/: multipath query result and option types
• src/: C++ and CUDA implementation
• src/edge/: edge BVH and edge-query implementation
• src/multipath/: reflection, visibility, EPC, and accumulation kernels
• src/rayd.cpp: Python bindings
• examples/: small usage examples
• tests/drjit/: Dr.Jit-native geometry, edge, visibility, and multipath tests
• tests/test_project_metadata.py: packaging and repository-boundary checks

Testing

python -m unittest tests.drjit.test_geometry -v
python -m unittest tests.drjit.test_visibility_topk -v
python -m unittest tests.drjit.test_reflection_epc -v
python -m unittest tests.drjit.test_reflection_accumulation -v
python -m unittest tests.test_project_metadata -v

The default development environment used by this repository is:

conda activate witwin3

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.