Product
AutoPerf — AI‑driven HPC benchmarking & optimization platform
•3 min read
Why I’m building this
Performance is a feedback loop. The shorter the distance from code to tests to measurement, the faster meaningful wins land in production. AutoPerf tightens that loop and keeps it honest: unit tests guard correctness (GoogleTest) and microbenchmarks report speed with Google Benchmark.
How it works
AutoPerf asks a language model to propose variants of a C++ kernel, compiles them, runs the test suite, and—only if everything passes—benchmarks the candidate and compares it to the baseline. The repository ships a handful of reference kernels (axpy, matvec, matmul, reduce, search) plus a template for custom ones, and a small Python orchestrator that drives the compile → test → benchmark cycle end to end. In practice, this already surfaces meaningful speedups while staying reproducible.
Get the source: github.com/sbstndb/AutoPerf
What’s next
I’m expanding the kernel set (including CUDA/OpenMP variants), making the “thinking” modes smarter and better traced, and packaging cleaner artifacts—diffs, concise notes, and a tiny UI. If this work resonates, follow along: I’ll keep sharing results and engineering notes as the project evolves.