Skip to content

CONTRIBUTING.md

Latest commit

 

History

History
42 lines (30 loc) · 1.81 KB

CONTRIBUTING.md

File metadata and controls

42 lines (30 loc) · 1.81 KB
Raw

Contributing

Bug reports and pull requests are welcome. A few things make them easy to take.

Building and testing

Requires the .NET 10 SDK (global.json pins the version).

dotnet build ReedSolomonFast.sln -c Release
dotnet test src/ReedSolomonFast.Tests -c Release

# The suite on one kernel tier (CI runs Scalar, Ssse3, Avx2, the best tier, Windows and ARM64)
REEDSOLOMONFAST_KERNEL=Avx2 dotnet test src/ReedSolomonFast.Tests -c Release

Every kernel tier must produce exactly what the shift-and-add reference in Gf256Reference produces, and parity must stay byte-identical to the Backblaze ports (InteropTests). A change that makes a test fail on any tier is a bug in the change until proven otherwise.

Performance changes

Measure, do not reason. The benchmark project compares the working tree against a frozen copy of the library (src/ReedSolomonFast.Baseline) in one process and prints an IMPROVED, REGRESSED or NO RESULT verdict per case:

dotnet run --project src/ReedSolomonFast.Benchmarks -c Release

A performance change lands on IMPROVED verdicts in two separate runs, with no REGRESSED case and a flat control case (one that does not run the changed code). Absolute times from different sessions are not comparable; only ratios inside one run are. Please include the verdict lines in the pull request, and say which machine and kernel tier produced them.

Code

  • Public API changes need tests for every overload shape (arrays, memories, contiguous stripe).
  • Kernels are generic over IGfVector<T>; a change to the engine applies to every tier, so run the suite with REEDSOLOMONFAST_KERNEL set to each tier your machine supports.
  • Comments say why, and cite the measurement when the reason is a number.
  • No new runtime dependencies. Build-time packages are fine with PrivateAssets="all".