Unifying Graph Compiler Framework for Novel Architectures @ ISCA 2021
This paper propose a GraphIR + GraphVM design that compiles GraphIt (very similar to Gunrock) programs on different hardware (include Swarm and Hammerblade (1000+ cores, each one carries scratchpad memory, connected w/ NOC) ).
They also support some scheduling language such as parallelize by push/pull/edge mode.
For Hammerblade architecture, the optimization they propose is three-fold:
- atomics
- an atomic pass to determine whether atomics are necessary.
- block access optimization
- alignment-base partitioning