Performance Tips

The following tips are meant to decrease the derivative computation's runtime complexity, especially when derivatives of the same function are needed repeatedly. There are two major modifications for all kinds of derivatives:

1. Use the derivative! driver, and work with allocator as explained in the guide Working with C++ Memory
2. Reuse the tape as often as possible. derivative! (derivative) supply the flag reuse_tape, which if set to true suppresses the creation of the tape, in addition the identifiyer of an existing tape must be provided as the parameter tape_id. More details can be found here.

An example could look like this:

using ADOLC # hide

# problem setup
f(x) = (x[1] - x[2])^2
x = [3.0, 7.5]
dir = [1/3, 1/7]
m = 1
n = 2
mode = :jac_vec
num_dir = size(dir, 2)[1]
num_weights = 0
tape_id = 1
num_iters = 100

# pre-allocation
cxx_res = allocator(m, n, mode, num_dir, num_weights)

derivative!(cxx_res, f, m, n, x, mode, dir=dir, tape_id=tape_id)

# conversion

# do computations ....

for i in 1:num_iters
    # update x ...
    derivative!(cxx_res, f, m, n, x, mode, dir=dir, tape_id=tape_id, reuse_tape=true)
    # do computations ...
end

Moreover, for higher-order derivatives you might consider the generation of a seed. However, if you do not pass a seed to the derivative! (derivative) driver, the partial identity is created as the seed automatically (see here).