Gamma @ ASPLOS 2021
Table of Contents
Paper link: https://people.csail.mit.edu/sanchez/papers/2021.gamma.asplos.pdf
Motivation
- SpGeMM is bottlenecked by memory traffic and data movement.
- Accelerators must minimize data movement.
- SpGeMM has rich algorithmic diversity.
- Wide range of dataflows with different tradeoffs.
- Gustavson's algorithm is more efficient than inner/outer product.
- more irregular reuse across data structures, requires specialized storage organization.
Gustavson's algorithm
The difference between inner product/outer product and Gustavson's algorithm is the order of loops.
