A better way is to use nested functions. Long-lived state goes in the outer function's variables, so you do not have to repeatedly declare them. Ephemeral variables go in the inner functions. Sub-tasks get encapsulated with zero boilerplate.
If your language does not have nested functions, use a worker class. The algorithm-wide state goes in member variables, initialized from constructor parameters. The sub-tasks are methods. A benefit is that sub-tasks from different jobs can be batched to better use the instruction cache and branch prediction information: a.do_step_1(); b.do_step_1(); a.do_step_2(); b.do_step_2(); ...
Nested functions are good for this kind of code, even when they are only used once and do not abstract --- because you have less variables in the innermost scope. (And when you are going to mutate variables at all, then strongly prefer to do it to variables in the innermost scope.)
Splitting the routine into 20 functions, one for each step, doesn't necessarily make it any more understandable: it reminds me of a DailyWTF in which a programmer was told to split a large routine into functions and just did something like this:
It's just as effective, if not better, to comment what each section of the large routine does. One advantage of keeping it in one routine is if there are dozens of values which are global to the routine and are not changed throughout it: it would create an ugly mess to pass those all around to each of the subroutines.