* Unroll multiply and add instructions in dotProd_32f - 35% faster.

* Eliminate unnecessary v_reduce_sum instructions.