Mapping Sparse Applications to Accelerated Computing Systems
summary
Describes a compiler stack that matches sparse tensor algebra to accelerators: Mosaic external function interface finds and offloads matching subexpressions (calls Taco, Intel MKL, custom accelerators) but its automatic search is functionality-only, not performance-aware. Introduces the Sam primitives streaming IR that models data flows and reproduces prior extensor cycles, and a Custard front end that lowers schedules to Sam. Fabricated Onyx CGRA implements Sam primitives in RTL and shows 4.4× to over 500× energy‑delay improvements across eight workloads.
AI-written summary · about
chapters (6)
- 0:00Sparse application mapping
- 2:28Why sparse accelerators
- 4:52Mosaic: Offloading accelerators
- 9:06Sam: Fine-grained programming
- 17:30Onyx: Sparse CGRA accelerator
- 20:46Summary & Q&A
related talks
28mFunctional Sparse Tensor Compilation
19mCompiler Support for Sparse Tensor Convolutions
17mA Sparse Iteration Space Transformation Framework for Sparse Tensor Algebra
19mMosaic: An Interoperable Compiler for Tensor Algebra
15mCompilation of Sparse Array Programming Models
25mFormat Abstraction for Sparse Tensor Algebra Compilers
20mVectorizing Sparse Coiteration for Two-finger Loop Structure (Extended Abstract)
19mCompiling Recurrences over Dense and Sparse Arrays