Implement write instruction compilation

[reczkok]

extends #808

closes #801

[codesandbox-ci]

This pull request is automatically built and testable in CodeSandbox.

To see build info of the built libraries, click here or the icon next to each commit SHA.

[reczkok]

Looks promising 🧑‍🍳 Need to sort out the actual benchmarks as they don't seem to be accurate for the new method (this results are from a temporary test example)

Here is the code (not exactly a benchmark but just for reference)

async () => {
  // warmup
  for (let i = 0; i < 1000; i++) {
    buffer.write({ x: 0, y: 0, z: 0, index: i });
  }
  root.unwrap(buffer);

  const t0 = performance.now();
  for (let i = 0; i < 100000; i++) {
    buffer.write({ x: 0, y: 0, z: 0, index: i });
  }
  const t1 = performance.now();

  console.log(`100000 writes took ${t1 - t0}ms`);

  const t2 = performance.now();
  for (let i = 0; i < 100000; i++) {
    const hostBuffer = new ArrayBuffer(16);
    const view = new DataView(hostBuffer);
    view.setFloat32(0, 0);
    view.setFloat32(4, 0);
    view.setFloat32(8, 0);
    view.setUint32(12, i);
    device.queue.writeBuffer(root.unwrap(buffer), 0, hostBuffer, 0, 16);
  }
  const t3 = performance.now();

  console.log(`100000 writes (manual ref) took ${t3 - t2}ms`);
},

[iwoplaza]

Suggested change

	const compiledWriters = new Map<
	const compiledWriters = new WeakMap<

[reczkok]

Turns out a lot of our overhead is the creation of buffers per write that triggers the garbage collector

After changing from buffer per write to buffer per TgpuBuffer we are almost at vanilla performacne