kernel-loop

GPU kernel optimization prevents regressions by enforcing an evidence-gated loop requiring correctness verification and Nsight Compute profiling before accepting any single-variable code change. This disciplined approach mandates KBS-backed hypotheses between versions to ensure traceability.

Evidence-gated iteration applies exactly one independent performance variable per version transition, then requires correctness validation, Nsight Compute artifacts, and KBS evidence to pass a gate script before creating the next kernel version. This prevents bundled optimization tactics and preserves prior work.

Yes, the optimization loop orchestrates iterative performance tuning for CUDA, CUTLASS, CuTe DSL, and Triton GPU kernels. It preserves dimensions and measurement settings throughout the benchmarking and Nsight Compute profiling process across all supported frameworks.

To start iterative GPU kernel benchmarking, you need a baseline kernel as version zero, a reference Python file, the implementation, dimensions, and optional GPU and iteration count specifications. These inputs enable the system to iteratively validate, profile, and generate final benchmark artifacts.

One-variable-at-a-time optimization forces each version transition to apply exactly one independent performance variable, preventing bundled tactics that obscure which change caused a performance shift. This ensures traceability and reduces the risk of introducing hidden regressions during GPU kernel tuning.