Let's say I have two arrays a and b both with dtype np.complex128 and I want to compute C = np.matmul(a, b).real . That is, I don't care about the imaginary part, only the real part. Is there a better way to do this? It seems like such a method would provide a decent speedup considering that only around half of the floating point multiplications are actually required. As a bonus question, what I actually want to do is compute the real part of the einsum np.einsum("ab,bc,cd,da->a", ...).real with rather large operands, so hopefully methods for the basic question are generalizable to einsum. Continue reading...
