TheRock/rocm-libraries/projects/hipblaslt/scripts

hipBLASLt scripts

For full build prerequisites and installation, see the project README and Building and installing hipBLASLt.

run_tensile_logic_check.py

Runs TensileLogic --check-all on the library logic YAMLs (same check as the pre-build gate). This file is cross-platform (Windows and Unix).

How to run

From the hipblaslt project root (where library/ and tensilelite/ live):

python scripts/run_tensile_logic_check.py

Windows: python scripts\run_tensile_logic_check.py (or use py if you have the launcher).

That’s it. Use whatever type of Python you have (system, store, or a venv). If that Python is missing dependencies (e.g. joblib) and the project has a .venv in the repo root, the script will re-run itself with .venv\Scripts\python.exe (Windows) or .venv/bin/python (Unix), so you don’t have to remember to activate.

One-time setup

1. Build once so rocisa is present under build/tensilelite/rocisa/ (nanobind _rocisa next to the rocisa package) or the legacy build/tensilelite/rocisa/lib layout.
2. Python deps – either:
   • Use the project .venv: python -m venv .venv then .venv\Scripts\pip install -r tensilelite/requirements.txt (Windows) or .venv/bin/pip install -r tensilelite/requirements.txt (Unix), or
   • Install into your current Python: pip install -r tensilelite/requirements.txt.

Optional: check a single directory

python scripts/run_tensile_logic_check.py library/src/amd_detail/rocblaslt/src/Tensile/Logic/asm_full/navi33/GridBased

Known-bugs list (ROCM-7144 / validation exceptions)

If tensilelite/Tensile/TensileLogic/known_bugs.yaml exists, this script passes --known-bugs to TensileLogic automatically so specific (logic file path, SolutionIndex) pairs are skipped (same as the CMake pre-build gate). Paths in that YAML are relative to the library logic root (library/), with optional # comments and an optional ticket: field for Jira keys. Override or disable this list by passing your own --known-bugs path on the command line.

Exit code

• 0 – All solutions passed (Reject = 0).
• 1 – One or more failed; errors are printed.

The full tree (~2246 files) can take several minutes, so passing a subdirectory is useful for quick checks. To tune parallelism: python scripts/run_tensile_logic_check.py -j 16 (default is 48 workers, capped by CPU count).

run_tensile_logic_check.sh (Unix only)

Thin wrapper that runs the script with .venv/bin/python, if it's present, and with python3 otherwise. Use the .py script directly on Windows.