New Selectors in Fireplace
Over the last couple of weeks, at Jerome’s suggestion, I rewrote all the selector code. The following summarizes what changed, and some of the motivation behind the rewrite and the new implementation.
Selectors are part of the card definition DSL that pick out subsets of game entities. As opposed to Actions, they do not mutate the game state at all. The lack of state mutation makes selectors much easier to reason about, and was a good reason to explicitly separate them out from the rest of the DSL.
For example, the card definition for Lightning Storm is simply:
play = Hit(ENEMY_MINIONS, RandomNumber(2, 3))
ENEMY_MINIONS is a selector that takes all the entities in the games and
picks out the enemy minions, which the
Hit() action then takes as argument.
This way, if Lightning Storm ever changed for whatever reason (say, if it hit the
hero as well), then we could simply change
Selectors can be combined and composed with one another; for example,
is equivalent to
IN_PLAY + ENEMY + MINION, where
+ means set intersection of results.
The previous selector implementation had some interesting features:
- It defined a smaller DSL that looked a lot like Forth, with all variables and commands on an explicit stack
- Each selector was itself defined in terms of this Forth-like sub-language
- Upon selector evaluation time, we iterate through the stack and execute any commands we find
- When we reach the end of the stack, we pop the last value on the stack and return it as the selector’s result
If you’ve ever used an HP-12C calculator’s reverse Polish Notation syntax, the selector implementation’s sub-language should be immediately familiar.
The existing implementation was very flexible, but as we started profiling the performance of Fireplace, a few issues became apparent:
- Selectors were around 35-40% of total runtime for the existing implementation; almost every card in the game uses selectors, and they need to iterate through ~60-odd entities every time they’re run
- Selectors were slow in part because deeply composed selectors
FRIENDLY + IN_PLAY) had long Forth-like programs, and there was significant overhead in building up the explicit stack and tearing it down again
- Reduced readability: To understand a selector’s implementation, you need to understand
both the DSL and the selector sub-language
- Representing a selector like
DRAGON | (FRIENDLY + IN_PLAY)essentially meant we took the card DSL’s AST, linearized it, then implicitly rebuilt the tree again when we evaluated the selector
- The reader needs to mentally switch between infix and postfix notation on the fly
- Representing a selector like
The solution for the new implementation was to represent the DSL directly using Python selector objects. To show what I mean, here’s part of the implementatin for the root class:
class Selector: ... def eval(self, entities: List[BaseEntity], source: BaseEntity) -> List[BaseEntity]: return entities def __add__(self, other: SelectorLike) -> "Selector": return SetOpSelector(operator.and_, self, other) def __or__(self, other: SelectorLike) -> "Selector": return SetOpSelector(operator.or_, self, other) ...
Instead of defining an extra stack and another low-level DSL, the logic of selectors
is embedded directly in their implementations. For example, a
simply takes two other selectors, evaluates them, and returns the merged result:
def eval(self, entities, source): left_children = self.left.eval(entities, source) right_children = self.right.eval(entities, source) result_entity_ids = self.op(self._entity_id_set(left_children), self._entity_id_set(right_children)) # Preserve input ordering and multiplicity return [e for e in entities if e.entity_id in result_entity_ids]
By doing things this way, we use the Python interpreter’s program counter and stack in place of the sub-language’s PC and stack. This ends up being much faster, because the interpreter’s own PC and stack are in highly optimized C. Profiling determined that new selectors only accounted for 15% of runtime, as opposed to 35-40% before, so we more than doubled selector performance. Some of the gains are from using the interpreter’s stack, and some of them are in removing some redundant entity checks directly.
Readability is also enhanced a bit, as the new selector logic is written in
straightforward Python. The mapping from DSL to selectors is straightforward;
SELECTOR1 + SELECTOR 2 | SELECTOR3 is a tree with 3 leaf
nodes and two
SetOpSelector nodes. All in all, the new implementation
is about 100 lines of code shorter.
PEP 0484 or: how I Learned to Stop Worrying and Love Gradual Typing
This change also marks the debut of PEP 0484 type hints into the codebase. These were introduced in Python 3.5, and I’m personally a huge fan type hinting and static type checking in general. While still in the early stages, they have a lot of potential to aid readability, help catch bugs, and most importantly in Python, are completely optional and work hand in hand with duck typing. The last point is especially important, as Fireplace currently uses dynamic dispatch and duck typing widely.
For anyone interested in type theory or where Python is going or just writing cleaner code, I highly recommend seeing Guido’s PyCon talk on the topic here: