TACO is a compiler-based library for sparse and dense linear and tensor algebra, intended for applications in data analytics, scientific computing, and machine learning. It supports computations ranging from sparse matrix-vector multiplication to MTTKRP on higher-order sparse tensors. Users can work with storage formats including CSR and COO, as well as specialized formats such as CSF. Its central approach is to generate kernels tailored to the requested computation, rather than relying on a fixed set of operations.
The project presents TACO as versatile, fast, and accessible: its website claims performance can exceed MATLAB’s Tensor Toolbox by several orders of magnitude and can compete with libraries such as Eigen, Intel oneMKL, and SPLATT. C++ and Python APIs let users express computations in a style similar to Eigen or SciPy, while command-line and browser-based tools support generating custom kernels. The code is available on GitHub, with documentation and examples for getting started. Users can report bugs, suggest improvements, or contribute code through issues and pull requests. Development is shared by research groups at Stanford and MIT and other contributors.
Summary generated by AI from the linked article. hn.today is not affiliated with Hacker News or Y Combinator.