[jowens]

The last item on the "move it into core" list: benchmarking a primitive, and setting up a device in Node.

Why

• Every timing script rewrote the same steps: warm up, time several executes, work out which per-kernel time belongs to which kernel. They were copied across regression_node.mjs, the MultisplitPerm and histogram benchmark scripts, the sort diagnostic and the paper's bench. The sort diagnostic left out the warm-up, and the V100's idle clock-down made it look 10× slow.
• Every Node script also set up its own device, with its own set of pitfalls:
  • keep create()'s result referenced, or Dawn tears the device down;
  • Dawn's timestamp rounding;
  • requesting features and limits.

Core

• primitive.kernelLabels(): the kernels' labels, in the order of getTimingResult()'s GPU times, nested primitives included.

• primitive.warmUp(args, warmMs = 200): runs the primitive back to back, doubling trials until one execute keeps the GPU busy for 20 ms.

• primitive.benchmark({ ...executeArgs, warmMs, repeats, trials }): warms up, times repeats executes, and returns medians per trial, per kernel with GPU timestamps (or CPU timing without them):

  const r = await primitive.benchmark({ inputBuffer, outputBuffer });
  // { ns, timing: "gpu", trials, repeats, kernels: [{ label, ns }, ...] }

Node (misc/, not part of the library: src/ also runs in browsers)

• misc/node_device.mjs: nodeGPU() and createNodeDevice(). They give full-resolution timestamps, keep the instance referenced, request subgroups and timestamps when available along with the adapter's largest limits, and set WebGPU's globals.
• misc/benchmark_node.mjs [primitives] [log2 sizes] [--kernels]: benchmark() over DLDFScan, HierarchicalScan, OneSweep, Histogram and MultisplitPerm. It replaces the one-off benchmark scripts.
• Existing scripts: regression_node.mjs now uses primitive.warmUp(), whose executes aren't timed, and nodeGPU(). test_wgsl_functions_node.mjs and list_adapters.mjs use nodeGPU().

Checked (lavapipe)

• Full regression suite with timing and large tests: everything passes except the existing sort failures at 8 lanes, and all 135 timings are recorded.
• WGSL function tests: the same three subgroup failures as on main.
• list_adapters and benchmark_node.mjs: both work. Every primitive is benchmarked, with per-kernel labels that include nested kernels (HierarchicalScan's partials scan, MultisplitPerm's DLDFScan).
• Package test: passes.

Independent of #84 and #85.

---

Generated by Claude Code