Torchbp

Fast C++ Pytorch extension for differentiable synthetic aperture radar image formation and autofocus library on GPU.

Only Nvidia GPUs are supported. Currently, only some of the operations are supported on CPU.

On RTX 3090 Ti backprojection on polar grid achieves 370 billion backprojections/s and fast factorized backprojection (ffbp) is ten times faster on moderate size images.

Installation

Tested with CUDA version 12.9.

From source

Install PyTorch first, then build torchbp against that same torch with build isolation disabled:

pip install torch
git clone https://github.com/Ttl/torchbp.git
cd torchbp
pip install --no-build-isolation -e .

--no-build-isolation is required. The extension links libtorch C++ symbols that are not ABI-stable across torch versions, so it must be compiled against the exact torch you run at import time. Without the flag, pip builds in an isolated environment with a different (freshly downloaded) torch, producing a .so that fails to import with an undefined symbol error. For the same reason, rebuild torchbp whenever you change your torch version.

Tests

Install test dependencies, then run pytest:

pip install --no-build-isolation -e .[test]
pytest tests/

On a CPU-only machine the CUDA tests are skipped rather than failing.

Documentation

Latest documentation can be viewed at: https://ttl.github.io/torchbp/

API documentation and examples can be built with sphinx.

pip install --no-build-isolation .[docs]
cd docs
make html

Open docs/build/html/index.html.