Datasets:

---

Mascinissa

/

LOOPerSet

like 2

Summary

This update introduces LOOPerSet v2, expanding program metadata and simplifying schedule representation, while also adding compact dataset files for faster loading.

Key changes

Added raw generator code to each program

• New top-level field: Tiramisu_cpp containing the raw Tiramisu C++ generator source before any transformations.
• Added a companion archive: data/source/looperset_v2_generators.tar.gz with ~220k generator files named <program_name>_generator.cpp.

Redesigned schedule annotation schema

• Each schedule entry now follows a unified structure:
  • transformations_list (list of dict transformation steps)
  • schedule_str (new unified, pipe-separated schedule string)
  • legacy_schedule_str (legacy string preserved for backward reference)
  • ISL_AST, Halide_IR, Tiramisu_transform_commands
  • execution_times unchanged in meaning (raw timings list)

Added “Compact” variants for speed / convenience

• New configs/files: full_compact and pact25_split_compact.
• Compact files remove the following fields: Tiramisu_cpp, ISL_AST, Halide_IR, Tiramisu_transform_commands, exploration_trace.

Repo structure + README updated

• New filenames include the v2 prefix (e.g., data/full/looperset_v2_full.jsonl.gz).
• Updated dataset card to document new configs, schedule format, and source archive.

Backward compatibility

• This is a schema-breaking release (v1 parsers will not work without updates).
• legacy_schedule_str is kept to help map v1 schedule to v2.