Function sum_with_per_step_rounding 

pub fn sum_with_per_step_rounding(values: &[Bf16]) -> Bf16

Sum a slice, rounding to bfloat16 after every step.

This is the weaker contract, and it is the one to avoid when the accumulation is promised in f32: on a slice of three hundred ones it stalls at 256.0, because from there on bfloat16 spacing above one is 2.0 and adding one rounds back down.

Examples

use tenferro_bf16_proof::{reduction::sum_with_per_step_rounding, Bf16};

let ones = vec![Bf16::from_f32(1.0); 300];
assert_eq!(sum_with_per_step_rounding(&ones).to_f32(), 256.0);