Reuse each widened weight chunk across up to four activation rows. Use a scalar path when the CPU lacks AVX2 or FMA. The blocked kernel reduces scoring time for both locked models. Keep a focused test for row blocks, input tails, and optional bias.
luojiyin