kernel-loop

What problem does it solve?

It reduces the risk of performance regressions by turning GPU kernel optimization into a disciplined, evidence-backed iteration loop with correctness and Nsight Compute (NCU) gates between changes.

Core Features & Use Cases

• Measured one-change iterations: Forces each version transition to apply exactly one independent performance variable, preserving prior work and preventing bundled tactics.
• Guardrails and traceability: Requires correctness, NCU artifacts, KBS evidence, and a structured hypothesis before allowing the next version to be created.
• Deterministic end-to-end loop orchestration: Coordinates the expected artifact layout, iteration gating script, and final reporting/benchmark step sequencing across existing profiling, KBS, and benchmarking skills.

Quick Start

Tell your system to run the kernel-loop using your baseline kernel as v0 and provide the required inputs (kernel, ref.py, implementation, dims, and optional GPU and iteration count) so it can iteratively validate, profile, gather KBS evidence, apply one-variable hypotheses, and produce final benchmark and report artifacts.