Skip to content

README.md

Latest commit

 

History

History
102 lines (71 loc) · 2.79 KB

README.md

File metadata and controls

102 lines (71 loc) · 2.79 KB
Raw

end-to-end C harness

This directory complements the strict .loop -> .loop regression suite with a small whole-C harness workflow.

The purpose is to test a more realistic source-to-source story:

C wrapper
-> loop fragment
-> theorem-aligned polopt route
-> optimized loop
-> transpile/splice back into C
-> compile and run

The current first slice is intentionally small:

  • one ordinary tiled case: matmul
  • one ISS-positive case: reverse_iss
  • heavier perf-oriented siblings: matmul_perf, reverse_iss_perf

These cases are not yet in default CI. They are a strengthening track for the artifact rather than part of the minimal regression gate.

For the broader generated whole-C campaign over the 62-case regression corpus, see ../end-to-end-generated. That path synthesizes a complete C harness from each input.loop / optimized.loop pair instead of relying on handwritten wrappers.

Layout

  • cases/<name>/meta.json
    • points to the source .loop
    • records the polopt flags for the case
  • cases/<name>/wrapper.c.in
    • C wrapper template with /* POLCERT_KERNEL */ splice marker

Untracked generated output goes to:

  • out/<name>/

Run

Build polopt, then run:

opam exec -- make test-end-to-end-c-smoke

For the heavier perf-oriented pair, run:

opam exec -- make test-end-to-end-c-perf

For the generated 62-case suite, first materialize the polopt outputs and then run:

opam exec -- make test-end-to-end-generated

Or run a single case manually:

python3 tools/end_to_end_c/run_case.py tests/end-to-end-c/cases/matmul --polopt ./polopt

Current boundary

This harness currently compares:

  • baseline wrapper + transpiled input .loop
  • optimized wrapper + transpiled polopt output

It does not yet compare against a Pluto-generated full C output. That is the next natural extension once the basic splice workflow is stable.

It now records both:

  • exact stdout equality
  • numeric drift summaries (max_abs_diff, max_rel_diff)

So if a future case uses tolerances, the numeric difference is still reported instead of being silently hidden.

Generated suite boundary

The generated suite under tests/polopt-generated/cases/* does not reuse the original benchmark C sources. Some of those sources are whole programs, but others are only #pragma scop fragments. To cover the entire generated corpus, the suite instead:

  • parses each .loop
  • synthesizes declarations and deterministic initialization
  • emits a checksum-based whole-program summary
  • compares baseline and optimized executables on that synthesized wrapper

This makes it possible to cover the full generated corpus with a uniform end-to-end benchmark path, at the cost of using generated wrappers rather than the original benchmark driver code.