I agree with what you're saying, and I would phrase it as "write the abstractions that reflect the essential complexity". The whole program should minimally reflect the essential complexity of the problem. Of course actually doing that isn't easy, but the result is obviously a simple solution for a given problem. It becomes another challenge to maintain and refactor: the question of changing problem constraints and being able to minimally change a program to match.
Why are ADTs like stacks, queues, and hashmaps so popular? Why are languages like C or Forth so highly praised for a high ceiling for performance and efficiency? Because they are usually "about as good as it gets" to solve a problem, "what you would've more or less done anyways". Maybe on a GPU, a language like C isn't quite fit, because the problem has changed. Make tools (e.g. CUDA) that reflect that distinct complexity.
Why are ADTs like stacks, queues, and hashmaps so popular? Why are languages like C or Forth so highly praised for a high ceiling for performance and efficiency? Because they are usually "about as good as it gets" to solve a problem, "what you would've more or less done anyways". Maybe on a GPU, a language like C isn't quite fit, because the problem has changed. Make tools (e.g. CUDA) that reflect that distinct complexity.