write linear

Coalesced grid-stride loop: adjacent threads store adjacent computed vec4<f32>s into a large buffer with no buffer reads. Write-bandwidth-bound.

Measured in B/s. Higher values mean better performance. Benchmark your device.