Bit-Sliced Vector (BSV)
Source Code
Bit-Sliced Vectors (BSV) are a compact representation for numeric and categorical data that enables fast analytical queries using bitwise operations. This project provides two complementary implementations:
bsiCPP is a C++ library for building, compressing, and querying bit-sliced vectors over large datasets. It supports hybrid compressed and non-compressed bit-vector operations and is designed for scalable data management workloads.
bsiPytorch extends BSV into the deep learning stack as a PyTorch + CUDA extension. It supports GPU-accelerated BSI quantization and compression experiments, tensor-core dot-product kernels, and end-to-end LLM evaluation benchmarks (e.g., OPT models on LAMBADA) to study the trade-offs between model accuracy, compression, and inference performance.
Contributors
To be added.
Publications
To be added.