Correctness Policy

Performance work is only accepted when correctness still holds against the repository's oracle.

Oracle and baseline

The project uses cuBLAS SGEMM as the reference implementation for runtime verification. Kernel outputs are compared against that reference rather than against another project-local kernel.

This matters because optimization work frequently changes launch geometry, staging, and mixed-precision behavior. A local baseline is not strong enough for those checks.

Tolerance policy

The verifier uses a NumPy-style allclose rule:

|test - ref| <= atol + rtol * |ref|

Standard FP32 kernels

• rtol = 1e-3
• atol = 1e-4

These thresholds apply to the Naive, Tiled, Bank Conflict Free, and Double Buffer paths.

Tensor Core / mixed precision path

• rtol = 5e-2
• atol = 1e-2

The Tensor Core path uses a relaxed tolerance because it includes mixed-precision behavior.

Shape coverage expectations

The test suite covers more than one friendly square case.

• standard dimensions include tiny, square, rectangular, and irregular shapes
• Tensor Core fast-path dimensions include 16-aligned cases
• Tensor Core fallback dimensions include deliberately unaligned cases such as 15x15x15, 17x19x23, and 511x513x1025

A performance change is not considered trustworthy if it only works on the friendly subset while silently weakening the fallback path.

CI versus local responsibility

Hosted CI can prove structure, documentation workflows, and governance alignment. It does not prove runtime correctness for CUDA kernels.

Local GPU execution must run:

ctest --test-dir build

Only after that passes should a benchmark-based performance statement be considered reviewable.