Welcome to Minuet’s Documentation#
Minuet🎶 is a social dance for two people, usually in 3/4 time a library that efficiently implements sparse convolutions (SC) for point clouds on GPUs. Minuet use sorted tables and binary search for building kernel maps, which results in on average \(15.8\times\) (up to \(26.8\times\)) speedup compared to hash table implementations in existing libraries.
Installation#
pip3 install "torch~=2.1" "packaging~=23.2"
CMAKE_BUILD_PARALLEL_LEVEL=$(nproc) pip3 install .
APIs#
Please refer to here for detailed APIs.
Examples#
Coming soon.
Code Structure and Call Stack#
Minuet is implemented as a torch extension. Every functionality of Minuet is implemented with a Python interface with compiled C++ extension binary which consists of three layers of function calls:
PyTorch Wrappers: This layer interacts with the calls from Python and is responsible for calling the corresponding C++ wrappers.
C++ Function Wrappers: We use a C++ function wrapper to avoid including the PyTorch’s header to reduce the compile time during development. This layer is also responsible for launching CUDA kernels.
CUDA Kernels: This layer implements the core functionality of Minuet on GPUs.
The following figure demonstrates the call stack of the kernel map building process.
