ako4all

You can automate GPU kernel optimization by running an agentic loop that iteratively profiles, edits, and benchmarks your code. The workflow establishes a workspace to continuously evaluate and improve kernel runtime speedups automatically.

Yes, the automated optimization workflow supports CUDA, Triton, TileLang, CuTe DSL, and Python kernels. It applies its optimize-benchmark-improve loop iteratively across these multiple languages and backends.

To start an automated GPU kernel benchmark loop, you need your kernel file, a bench script to drive evaluations, and optional reference data, inputs, or hints. Open your directory in Claude Code to initiate the optimization process.

Iterative kernel optimization handles profiling and correctness by running a repeatable loop. It uses ncu for profiling, performs correctness checks against reference data, and logs evaluation results directly to an ITERATIONS.md file.

Yes, you can resume automated GPU kernel optimization across sessions. The workflow commits optimization progress as a traceable artifact via git, enabling resumable iterations and continuous improvement over multiple separate runs.

The best way to evaluate GPU kernel speedups is using a plug-in benchmark like KernelBench. It drives the evaluation loop, checks correctness, and logs iterative profiling data to track performance improvements accurately.