Experiment: with one K-segment (K<=8192, the common case) each dst element is written exactly once (N-segments are disjoint), so the full-MxN memset and the read-modify-write accumulate are dead work. Write each element once; keep the +=/memset path for multi-K. Byte-identical output. Throughput ~neutral (NPU-bound, so CPU-side traffic saving is marginal): E1 pp300/pp512/pp700/tg 37.01/40.25/40.03/2.61 -> E2 35.8/41.1/40.3/2.57 (pp512/pp700 +1-2%, pp300 within run variance). Kept for traffic reduction and as prerequisite for E5 NEON collect.
mfritsche
c5a47492ff