pub const MIN_ELEMS: usize = _; // 524_288usizeExpand description
Least elements worth a dispatch.
Below this the round trip — two submissions, a queue wait, a readback —
costs more than the whole pass does on the CPU, whatever the arithmetic
per element is. Measured on a Radeon Pro 560 against the 8-thread CPU
path, the crossover for the simplest chain (+/ w * x) is around a
million elements; the threshold is set an octave below that so that a
heavier chain, which crosses over sooner, is not kept off the device.