demo · v137

Per-frame Copies

The reason the spec authors added the terse overload: tight engine loops that issue dozens to hundreds of full-buffer copies per frame. Each five-argument call goes through additional bounds validation that the two-argument overload can skip. The benchmark below records both shapes back-to-back.

probing…

old: copyBufferToBuffer(src, 0, dst, 0, size)

total time
—
per-call avg
—

new: copyBufferToBuffer(src, dst)

total time
—
per-call avg
—
delta
—

the code

// Tight loop, common in compute simulations / particle systems:
function frame() {
  const enc = device.createCommandEncoder();
  for (let i = 0; i < copies; i++) {
    enc.copyBufferToBuffer(src[i], dst[i]);     // new shape
    // enc.copyBufferToBuffer(src[i], 0, dst[i], 0, byteLen);  // old shape
  }
  device.queue.submit([enc.finish()]);
}

// At >60 calls/frame the savings on validation add up.

see also