slyubomirsky / flexmatch

Flexible Matching: Explore and reveal accelerator-supported workloads in Deep Learning Models

flexmatch

Flexible Matching via Equality Saturation

Workflow (Deprecated)

1. Having patterns defined in Relay
2. Run EqSat on each pattern and get the corresponding EGraph
3. Pattern match on a given model using EGraphMatcher
4. Annotate matched region with BYOC-style function annotations for the codegen

Utilities

• flexmatch: Rust interfaces that call glenside and apis in (my fork of) egg
• tests/run_eqsat
  • Takes a relay source file, an output file name and config(s) (under configs/)
  • Run equality saturation on the given model with respect to rewrite rules in configs
  • Example: python3 run_eqsat.py models/resmlp.relay resmlp im2col-rewrites linear-rewrites
• tests/compile_model
  • Takes the relay source file, the output model and eclass analysis data json from EqSat
  • Compiles the rewritten model back to a relay executable model and saves to a file
  • Optional argument: --debug; instead of inserting accelerator calls, if this argument is passed, the equivalent relay function will be generated
  • Example: python3 compile_model.py models/resmlp.relay resmlp-rewritten.relay resmlp-rewritten.json resmlp-data.json linear-rewrites im2col-rewrites --debug