Stardust: Compiling Sparse Tensor Algebra to a Reconfigurable Dataflow Architecture

Abstract

We introduce Stardust, a compiler from sparse tensor algebra languages to a sparse reconfigurable dataflow architecture via a parallel-patterns programming model. Stardust lets performance engineers specify the placement of data into memories separately from the placement of computation onto compute units. Users first schedule data placement onto an abstract memory model, and then Stardust binds that data to complex, on-chip physical memories. With guidance from user schedules, Stardust binds computation using these on-chip data structures to the appropriate parallel patterns. Through cycle-accurate simulation, we show that Stardust generates nine more tensor algebra kernels than the original Capstan sparse RDA work. The generated kernels perform, on average, 138x better than generated CPU kernels and 41x better than generated GPU kernels.

Article

BibTeX

@inproceedings{hsu2025stardust, 
  author = {Hsu, Olivia and Rucker, Alexander and Zhao, Tian and Desai, 
  Varun and Olukotun, Kunle and Kjolstad, Fredrik}, 
  title = {Stardust: Compiling Sparse Tensor Algebra to a Reconfigurable Dataflow Architecture}, 
  year = {2025}, 
  isbn = {9798400712753}, 
  publisher = {Association for Computing Machinery}, 
  address = {New York, NY, USA}, 
  url = {https://doi.org/10.1145/3696443.3708918}, 
  doi = {10.1145/3696443.3708918}, 
  abstract = {We introduce Stardust, a compiler from sparse tensor algebra 
  languages to a sparse reconfigurable dataflow architecture via a 
  parallel-patterns programming model. Stardust lets performance engineers 
  specify the placement of data into memories separately from the placement of 
  computation onto compute units. Users first schedule data placement onto an 
  abstract memory model, and then Stardust binds that data to complex, on-chip 
  physical memories. With guidance from user schedules, Stardust binds 
  computation using these on-chip data structures to the appropriate parallel 
  patterns. Through cycle-accurate simulation, we show that Stardust generates 
  nine more tensor algebra kernels than the original Capstan sparse RDA work. The 
  generated kernels perform, on average, 138x better than generated 
  CPU kernels and 41x better than generated GPU kernels.}, 
  booktitle = {Proceedings of the 23rd ACM/IEEE International Symposium on Code Generation and Optimization}, 
  pages = {628–643}, 
  numpages = {16}, 
  keywords = {DSLs, compilers, dataflow, parallel patterns, reconfigurable architectures, sparse tensor algebra}, 
  location = {Las Vegas, NV, USA}, 
  series = {CGO '25} 
  }