u32 variable shift

Eight independent u32 lanes, unrolled 4x, even lanes shifting left and odd lanes right by an amount taken from the loop counter, so every lane cycles through all 32 shift amounts and the number is the average across them. Counted as 1 op per lane step (the shift only; the XOR that refills shifted-out bits is not counted).

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