Speed up sol[UnaryExpr] via C-level math and refactor UnaryExpr

[Zeroto521]

This branch is 1.45x faster than master.

• master branch: 2.9803s
• this branch: 4.3398s

from timeit import timeit

from pyscipopt import Model, sqrt

m = Model()

n = 1000
x = m.addMatrixVar(n, lb=0, ub=1, vtype="C", name="x")
e = sqrt(x)
obj = m.addVar(lb=0, ub=None, vtype="C", name="obj")
m.addCons(obj == e.sum())

m.setObjective(obj, "maximize")
m.hideOutput()
m.optimize()
sol = m.getBestSol()

number = 10_000
cost = timeit(lambda: sol[e], number=number)
print(f"Cost of squaring an expression over {number:,} runs: {cost:.4f} seconds")

[Joao-Dionisio]

By the way, @Zeroto521 , out of curiosity. Are you also using PySCIPOpt for your work, or are you just speeding things up for fun?

[Zeroto521]

80% of the OR problems I encounter use commercial solvers or our closure algorithms for large-scale, speed-critical cases, while the rest use open-source ones for small tasks and research.

[Zeroto521]

ExprLike.__abs__ can replace UnaryExpr(Operator.fabs, self)

[Joao-Dionisio]

I see @Zeroto521 , thank you! We're actually in the process of collecting users of SCIP "in the wild", particularly in industry. Would it be possible for us to reference the work you're doing? We can talk via email, if you prefer: joao.goncalves.dionisio@gmail.com

[Joao-Dionisio]

Now, things like log of negative numbers will return NaN instead of a ValueError (and log(0) returns -inf). This should probably go into the changelog.