PyRXMesh 0.2.1rc1

Python bindings for RXMesh, a CUDA/C++ library for mesh processing on the GPU.

PyRXMesh

Python bindings for RXMesh, a CUDA/C++ library for GPU mesh processing.

PyRXMesh allows you to keep mesh workflows in Python while using RXMesh data structures and GPU kernels underneath. It supports exchanging arrays and attributes with NumPy and PyTorch, building sparse matrices, and solving linear systems. We use a Python plugin to allow CUDA kernels to be written, compiled, and used directly in Python. Check out CUSTOM_CUDA_PLUGINS.md for more details.

Installation

Install PyRXMesh with:

python -m pip install PyRXMesh

PyRXMesh is a CUDA package. You need an NVIDIA driver compatible with the CUDA version used by the wheel. For source builds and development setup, see DEVELOPING.md.

Simple Example

import pyrxmesh as rx

#Use GPU 0 
rx.init(0) 

mesh = rx.RXMeshStatic("mesh.obj")
print(mesh.num_vertices, mesh.num_edges, mesh.num_faces)

rx.show()

The repository includes a small viewer example:

python examples/load_and_show.py --input mesh.obj

Mesh Attributes

Attributes are data attached to vertices, edges, or faces

import pyrxmesh as rx

mesh = rx.RXMeshStatic("mesh.obj")

coords = mesh.input_vertex_coordinates()
velocity = mesh.add_vertex_attribute("velocity", dtype="float32",
                                     dim=3, location=rx.Location.ALL)

velocity.reset(0.0, rx.Location.DEVICE)

host_coords = coords.to_numpy_copy(source=rx.Location.HOST)
print(host_coords.shape)

NumPy And Torch Interop

Unsuffixed to_* methods return zero-copy views when possible. Methods ending in _copy make an explicit copy.

import pyrxmesh as rx

mesh = rx.RXMeshStatic("mesh.obj")
attr = mesh.add_vertex_attribute("temperature", dtype="float32",
                                 dim=1, location=rx.Location.ALL)

attr.reset(1.0, rx.Location.ALL)

view = attr.to_numpy(rx.Location.HOST)
view[0, 0] = 42.0

owned = attr.to_numpy_copy(source=rx.Location.HOST)
print(owned[0, 0])

Torch views use DLPack:

tensor = attr.to_torch(rx.Location.DEVICE)
tensor += 2.0

Dense And Sparse Matrices

PyRXMesh exposes RXMesh dense matrices and CSR sparse matrices.

import numpy as np
import pyrxmesh as rx

matrix = rx.DenseMatrix(4, 3, dtype="float32", location=rx.Location.ALL)
values = np.arange(12, dtype=np.float32).reshape(4, 3)

matrix.from_numpy_copy(values, target=rx.Location.ALL)
print(matrix.norm2())

Sparse matrices are CSR-only:

mesh = rx.RXMeshStatic("mesh.obj")
laplace_like = mesh.sparse_matrix(rx.Op.VV, dtype="float32")

row_ptr, col_idx, values = laplace_like.to_numpy_copy()
print(laplace_like.shape, laplace_like.nnz)

You can multiply sparse matrices by dense vectors:

x = np.ones((laplace_like.cols, 1), dtype=np.float32)
y = laplace_like.multiply_vector(x)
print(y.to_numpy_copy(source=rx.Location.HOST))

Solvers

RXMesh solver bindings work with SparseMatrix and DenseMatrix objects:

import numpy as np
import pyrxmesh as rx

row_ptr = np.array([0, 2, 5, 7], dtype=np.int32)
col_idx = np.array([0, 1, 0, 1, 2, 1, 2], dtype=np.int32)
values = np.array([4, 1, 1, 3, 1, 1, 2], dtype=np.float32)

A = rx.SparseMatrix.from_numpy_copy(row_ptr, col_idx, values,
                                    shape=(3, 3), dtype="float32")

b = rx.DenseMatrix(3, 1, dtype="float32", location=rx.Location.ALL)
b.from_numpy_copy(np.array([[1], [2], [3]], dtype=np.float32))

solver = rx.CGSolver(A, unknown_dim=1)
x = solver.solve(b)

print(x.to_numpy_copy(source=rx.Location.HOST))

Documentation

• RXMeshDocs: RXMesh CUDA/C++ documentation website
• DEVELOPING.md: build PyRXMesh from source and work against RXMesh forks, tags, or local checkouts.
• CUSTOM_CUDA_PLUGINS.md: write small compiled CUDA plugins that operate on PyRXMesh meshes and attributes.