Installation

Prerequisites

isoext requires:

• Python 3.10 – 3.12

• PyTorch with CUDA support

• CUDA Toolkit 12.4 or newer. Older nvcc versions fail on the glibc headers of recent Linux distributions (Ubuntu 24.04 and later) with errors like "__builtin_dynamic_object_size" is undefined. The build checks for this and stops with instructions if it finds an old nvcc.

• A C++ compiler (GCC on Linux, Visual Studio on Windows)

Install from PyPI

The simplest way to install:

pip install isoext

This will compile the CUDA extension during installation.

Note

On Windows, you may encounter errors due to path length limits (260 characters). Enable long paths by following this guide.

Install from Source

For development or to get the latest changes:

git clone https://github.com/GuangyanCai/isoext
cd isoext
pip install -e .

Verify Installation

import isoext

# Create a small test grid
grid = isoext.UniformGrid([8, 8, 8])
print(f"Grid has {grid.get_num_cells()} cells")

# Run marching cubes (should return empty mesh for default values)
v, f = isoext.marching_cubes(grid)
print(f"Extracted {len(f)} triangles")

Troubleshooting

CUDA not found

Make sure PyTorch is installed with CUDA support:

import torch
print(torch.cuda.is_available())  # Should print True
print(torch.version.cuda)         # Should print your CUDA version

Compilation errors

Ensure your CUDA toolkit version matches PyTorch’s CUDA version. Check with:

nvcc --version

Slow first call

The extension ships GPU code as PTX. The driver compiles it for your GPU the first time each algorithm runs, which takes a few seconds in total. The result is cached on disk, so this happens once per machine.

Import errors

If you see ImportError: PyTorch is required, install PyTorch first:

pip install torch --index-url https://download.pytorch.org/whl/cu121

Replace cu121 with your CUDA version (e.g., cu118, cu124).