triton-ascend-kahan-precision-fix

Replace simple accumulation with Kahan compensated summation: maintain a compensation variable that captures the low-order bits lost in each addition and adds them back in the next iteration. This reduces accumulation error from O(K × eps) to O(eps).

Kahan summation is an algorithm that tracks rounding error in floating-point accumulation using a compensation variable. Each iteration computes y = partial - comp, t = acc + y, then comp = (t - acc) - y, recovering the precision lost when adding a small value to a large accumulator.

Floating-point addition is not associative, so error grows with the number of accumulation steps. At K ≥ 4096, simple FP32 accumulation loses enough low-order bits that the Triton sequential path diverges from the NPU Cube engine path, producing hard_fail results.

Skip Kahan for matmul with K < 4096, where simple accumulation is accurate enough, and for elementwise kernels with no reduction, where no accumulation error exists. The extra operations add overhead without benefit in these cases.

No. Kahan only addresses accumulation rounding error. If hard_fail persists between NPU runs after applying Kahan, the kernel likely has a logic bug rather than a precision issue and needs functional debugging.