The reverse AD implementation of mac(A,b,c) requires the transpose of A. (Surprisingly), computing the transpose of A can require a significant number of operations and there is no need to perform these when A is constant since the transpose can be performed off-line. I will propose an implementation that performs these computations off-line.

Note that the overhead for computing the transpose of A (in my not so rigorous benchmark) is in the same order of mtimes() while the transpose of A should be O(nnz) ~= O(n^2) and mtimes ~ O(n^3). Chances are that the memory allocation required by the transpose is much more expensive.