Initial conditions don’t cause a recompile. It’s just the first time you solve with a method in a new Julia session. You can change initial conditions, parameters, etc (run a parameter estimation). The only thing that matters is that every time a new code is generated (new types are used), then you have to recompile.
That is indeed a plus point. I will try that with random initial conditions when I get time.
BTW, it looks like you’re doing Verner robust and efficient? How’re you seeing the robust ones doing? I did a bunch of tests with a tableau-based implementation awhile back but never did too many benchmarks after optimizing everything else. I’ve wanted to revisit them.
Actually, I used efficient version of the coefficients for Verner6 and robust coefficients for Verner9 for no particular reason. I have not compared the same algorithms with two different sets of coefficients to see the difference. Even though adding new coefficients in FLINT is fairly easy. Some day…