Abstract

With the slowing of Moore’s law, computer architects have turned to domain-speciic hardware specialization to continue improving the performance and eiciency of computing systems. However, specialization typically entails signiicant modii- cations to the software stack to properly leverage the updated hardware. The lack of a structured approach for updating both the compiler and the accelerator in tandem has impeded many attempts to systematize this procedure. We propose a new approach to enable lexible and evolvable domain-speciic hardware specialization based on coarse-grained reconigurable arrays (CGRAs). Our agile methodology employs a combination of new programming languages and formal methods to automatically generate the accelerator hardware and its compiler from a single source of truth. This enables the creation of design-space exploration frameworks that automatically generate accelerator architectures that approach the eiciencies of hand-designed accelerators, with a signiicantly lower design efort for both hardware and compiler generation. Our current system accelerates dense linear algebra applications, but is modular and can be extended to support other domains. Our methodology has the potential to signiicantly improve the productivity of hardware-software engineering teams and enable quicker customization and deployment of complex accelerator-rich computing systems.

Article

pdf

BibTeX

@article{koul2023,
  title={AHA: An agile approach to the design of coarse-grained reconfigurable accelerators and compilers},
  author={Kalhan Koul and Jackson Melchert and Kavya Sreedhar and Leonard Truong and Gedeon Nyengele and Keyi Zhang and Qiaoyi Liu and Jeff Setter and Po-Han Chen and Yuchen Mei and Maxwell Strange and Ross Daly and Caleb Donovick and Alex Carsello and Taeyoung Kong and Kathleen Feng and Dillon Huff and Ankita Nayak and Raj Setaluri and James Thomas and Nikhil Bhagdikar and David Durst and Zachary Myers and Nestan Tsiskaridze and Stephen Richardson and Rick Bahr and Kayvon Fatahalian and Pat Hanrahan and Clark Barrett and Mark Horowitz and Christopher Torng and Fredrik Kjolstad and Priyanka Raina},
  journal={ACM Transactions on Embedded Computing Systems},
  year={2023},
  month={January}
}