• UST decouples tensor sparsity from memory representation, enabling flexible storage formats. • Developers describe storage via a DSL, focusing solely on sparsity patterns. • Compile-time/runtime dispatch selects optimized libraries or auto-generates sparse code. • Supports common formats (COO, CSR, DIA) and custom, application-specific layouts. • Integrates with SciPy, CuPy, PyTorch, and leverages MLIR sparse compiler tech. • Aims to reduce manual effort, errors, and improve scalability across architectures.

Article Summaries:

  • The article introduces the Universal Sparse Tensor (UST), a framework that separates a tensor’s sparsity pattern from its memory layout. By employing a domain‑specific language (DSL), developers can describe how logical tensor dimensions map to physical storage levels-specifying properties such as dense, compressed, singleton, or range formats. UST automatically selects the most efficient implementation, dispatching to hand‑tuned libraries when available or generating sparse code otherwise. It already maps common formats (COO, CSR, DIA) to the DSL, while also enabling the definition of custom sparsity formats without manual coding. The approach aims to simplify sparse tensor handling across scientific computing, signal processing, and deep learning, improving scalability and portability.

Sources: