Capstone kernel

Three artifacts, one kernel

The capstone is a kernel plus three claims. Correctness: every benchmarked shape and dtype compared against a reference, with the tolerance justified by accumulation order and input dtype. Speed: a named baseline, named shapes, and <ms> and <speedup> with the band from one interleaved session. Bottleneck: a roofline prediction and the counter that confirms it.

Predict the counter before you profile

Write the prediction down before ncu runs. A kernel that is bandwidth-limited at <GB/s> should show dram__throughput.avg.pct_of_peak_sustained_elapsed high and sm__throughput.avg.pct_of_peak_sustained_elapsed low, and dram__bytes.sum should match the traffic the shape implies. If the counters disagree, the model was wrong, and the honest move is to fix the model and then the kernel. A session that only confirms what you already believed is decoration.

What survives

A negative result is a result. If the fused kernel loses at short sequence lengths because launch overhead dominates, say so and show the gap in the nsys timeline, where the kernel time is a sliver between gaps. If the win disappears against a baseline that is already fused, that is the finding, and it is worth more in an interview than a ratio against a straw man.

Things that break

Choosing the one shape where the kernel wins and presenting it as the kernel’s performance. A correctness check that runs on a single square case while the benchmark runs on six shapes. A bottleneck story contradicted by the counters that stays in the writeup because the roofline predicted it, which is the failure this node exists to prevent. And a speedup measured while tuning runs inside the timed region, so the number includes compilation. The harness uses Triton 3.9 do_bench, which returns milliseconds; confirm the return units against your installed version before the number enters the table.