Kai Zhao presents TorchMorph, a lightweight PyTorch extension implementing 22 morphological operators as fused CUDA kernels that operate directly on GPU tensors with up to eight spatial dimensions, deliberately mirroring scipy.ndimage’s argument conventions for drop-in integration. The library covers binary morphology, greyscale morphology, distance transforms, and optimal transport operations. Benchmarks show throughput improvements reaching 1,100x for greyscale morphology and 350x for Euclidean distance transforms relative to SciPy, with binary operators producing identical results and float-valued operators maintaining accuracy within 1.8e-6 absolute error.
