Optimization Playbook

A practical diagnosis loop for SGEMM performance bottlenecks

End-to-end optimization loop

Use one loop per hypothesis. If you change five things at once, you learn nothing.

Experiment templates

Template A: isolate one shape

./build/bin/sgemm_benchmark --dims 1024 1024 1024

Use this when you want to remove shape noise and focus on one bottleneck.

Template B: sweep shape diversity

./build/bin/sgemm_benchmark -a

Use this to detect regressions that only appear on awkward dimensions.

Template C: longer measurements

./build/bin/sgemm_benchmark -a --warmup 10 --benchmark 50

Use this before making a final claim in docs or PRs.

Quality gate before claiming a speedup

• Re-run ctest --test-dir build and keep cuBLAS comparison clean.
• Compare both a canonical shape (1024 x 1024 x 1024) and at least one irregular shape.
• Report whether the number is end-to-end or compute-only.
• Keep the kernel launcher contract unchanged unless there is a strong reason.

---