My attempt at trying to outperform Andy Zeng's PyCuda TSDF fusion implementation with PyTorch.
To build the PyTorch C++ integration function on macOS use:
CXX=clang python setup.py install
To run the benchmark, use:
python benchmark.py py/cpp/jit
| Method | FPS | Avg Integration Time |
|---|---|---|
| PyCuda | 64.45 | 0.006 |
| Vanilla PyTorch | 24.77 | 0.031 |
| C++ PyTorch | 24.71 | 0.031 |
| JIT PyTorch | 26.01 | 0.029 |