// Experiment: keep the GPU frequency governor boosted by running a small
// continuous compute load (one workgroup) from a separate device, to
// quantify how much of the multi-dispatch backend's time is the 300 MHz
// idle clock rather than the work.
import { installNodeWebGpu, requestDevice } from './gpu.mjs';

async function main() {
  await installNodeWebGpu();
  const { device } = await requestDevice();
  const code = `
@group(0) @binding(0) var<storage, read_write> st: array<f32>;
@compute @workgroup_size(64)
fn main(@builtin(local_invocation_index) lid: u32) {
  var x: f32 = st[lid];
  for (var i: u32 = 0u; i < 15000u; i = i + 1u) {
    x = x * 1.0000001 + 1e-8;
  }
  st[lid] = x;
}`;
  const module = device.createShaderModule({ code });
  const pipeline = device.createComputePipeline({ layout: 'auto', compute: { module, entryPoint: 'main' } });
  const buf = device.createBuffer({ size: 256, usage: GPUBufferUsage.STORAGE });
  const bg = device.createBindGroup({
    layout: pipeline.getBindGroupLayout(0),
    entries: [{ binding: 0, resource: { buffer: buf } }],
  });
  const secs = Number(process.argv[2] ?? 30);
  const t0 = performance.now();
  while (performance.now() - t0 < secs * 1000) {
    const enc = device.createCommandEncoder();
    const pass = enc.beginComputePass();
    pass.setPipeline(pipeline);
    pass.setBindGroup(0, bg);
    pass.dispatchWorkgroups(1);
    pass.end();
    device.queue.submit([enc.finish()]);
    await device.queue.onSubmittedWorkDone();
  }
}

main().then(() => process.exit(0)).catch((e) => { console.error(e); process.exit(1); });