Two products, one sum
You will learn
How a GF-T16 multiply-accumulate computes a1 * b1 + a2 * b2, the kernel of a matrix multiply.
Every layer of a network is a stack of dot products, and a dot product is products added up. gft_dot2.t27 is the smallest one: y = a1 * b1 + a2 * b2 in GF-T16, a 16-bit magnitude of a 7-bit offset and a 9-bit mantissa with bias 40. In that layout 1.0 is code 20480 and 2.0 is code 20992. The spec has two tests: 1 * 1 = 1, and 1 * 1 + 1 * 1 = 2.