The Sparse Abstract Machine: A Model for Sparse Tensor Algebra on Dataflow Accelerators

Explore the Sparse Abstract Machine (SAM), an innovative abstract machine model designed for targeting sparse tensor algebra to reconfigurable and fixed-function spatial dataflow accelerators. Delve into SAM's streaming dataflow abstraction with sparse primitives, which encompasses a wide range of scheduled tensor algebra expressions. Discover how SAM dataflow graphs effectively separate tensor formats from algorithms and accommodate various iteration orderings and hardware-specific optimizations. Learn about Custard, a compiler that demonstrates SAM's potential as an intermediate representation by translating from a high-level language to SAM. Examine the automatic binding process from SAM to a streaming dataflow simulator. Gain insights into SAM's evaluation as a comprehensive system for sparse tensor algebra, its role in design-space exploration for sparse accelerator performance, and its ability to model dataflow hardware implementations.

[CTSTA'23] The Sparse Abstract Machine