Skip to content

Commit

Permalink
[MIRROR] Adds pathmaps, refactors pathfinding a bit [MDB IGNORE] (#24…
Browse files Browse the repository at this point in the history
…414)

* Adds pathmaps, refactors pathfinding a bit (#78684)

## About The Pull Request

Implements /datum/pathfind/sssp, which generates /datum/path_map

/datum/path_maps allow us to very efficently generate paths to any turf
they contain from their central point.

We're effectively running the single source shortest paths algorithm.
We expand from the center turf, adding turfs as they're found, and then
processing them in order of addition.
As we go, we remember what turf "found" us first. Reversing this chain
gives us the shortest possible path from the center turf to any turf in
its range (or the inverse).

This isn't all that useful on its own, outside of a few niche cases
(Like if we wanted to get the farthest reachable turf from the center)
but if we could reuse the map more then once, we'd be able to swarm
to/from a point very easily.

Reuse is a bit troublesome, reqiures a timeout system and a way to
compare different movables trying to get paths.
I've implemented it tho. I've refactored CanAStarPass to take a datum,
/datum/can_pass_info. This is built from a movable and a list of access,
and copies all the properties that would impact pathfinding over onto
itself.

There is one case where we don't do this, pathing over openspace
requires checking if we'd fall through the openspace, and the proc for
that takes an atom.
So instead we use the weakref to the owner that we hold onto, and hold
copies of all the values that would impact the check on the datum.

When someone requests a swarmed path their pass info is compared with
the pass info of all other path_maps centered on their target turf. If
it matches and their requested timeout isn't too short, we just reuse
the map.

Timeout is a tricky thing because the longer a map exists the more out
of date it gets.
I've added a few age defines that let you modulate your level of risk
here. We default to only allowing maps that are currently
being generated, or finished generating in our tick.
Hopefully this prevents falling into trouble, but consumers will need to
allow "failed" movements.

As a part of this datumized pass info, I've refactored pathfinding to
use access lists, rather then id cards directly. This also avoids some
dumbass harddel oppertunities, and prevents an idcard from changing mid
path.

Did a few things to the zPass procs, they took args that they did NOT
need, and I thought it'd be better to yeet em.

If you'd all like I could undo the caching/can_pass_info stuff if you'd
all like. I think it's useful generally because it avoids stuff changing
mid pathfind attempt, but if it's too clunky I could nuke it.

Oh also I added optional args to jps that constricts how it handles
diagonals. I've used this to fix bot paths.

## Why It's Good For The Game

Much of this is redundant currently. I'm adding it because it could have
saved hugglebippers, and because I get the feeling it'll be useful for
"grouping" mobs like bees and such.
We're doing more basic mob work currently and I want to provide extra
tools for that work.

https://github.com/tgstation/tgstation/assets/58055496/66aca1f9-c6e7-4173-9c38-c40516d6d853

## Changelog
🆑
add: Adds swarmed pathfinding, trading accuracy for potential
optimization of used correctly
fix: Bots will no longer take diagonal paths, preventing weirdo looking
path visuals
refactor: Refactored bits of pathfinding code, hopefully easier to add
new pathfinding strategies now
/🆑

* Adds pathmaps, refactors pathfinding a bit

---------

Co-authored-by: LemonInTheDark <[email protected]>
  • Loading branch information
2 people authored and FFMirrorBot committed Oct 18, 2023
1 parent 0b37520 commit 144774a
Show file tree
Hide file tree
Showing 43 changed files with 1,302 additions and 592 deletions.
20 changes: 20 additions & 0 deletions code/__DEFINES/path.dm
Original file line number Diff line number Diff line change
Expand Up @@ -3,3 +3,23 @@
#define CANASTARPASS_DENSITY 0
/// If this is set, we bypass density checks and always call the proc
#define CANASTARPASS_ALWAYS_PROC 1

/**
* A helper macro to see if it's possible to step from the first turf into the second one, minding things like door access and directional windows.
* If you really want to optimize things, optimize this, cuz this gets called a lot.
* We do early next.density check despite it being already checked in LinkBlockedWithAccess for short-circuit performance
*/
#define CAN_STEP(cur_turf, next, simulated_only, pass_info, avoid) (next && !next.density && !(simulated_only && SSpathfinder.space_type_cache[next.type]) && !cur_turf.LinkBlockedWithAccess(next, pass_info) && (next != avoid))

#define DIAGONAL_DO_NOTHING NONE
#define DIAGONAL_REMOVE_ALL 1
#define DIAGONAL_REMOVE_CLUNKY 2

// Set of delays for path_map reuse
// The longer you go, the higher the risk of invalid paths
#define MAP_REUSE_INSTANT (0)
#define MAP_REUSE_SNAPPY (0.5 SECONDS)
#define MAP_REUSE_FAST (2 SECONDS)
#define MAP_REUSE_SLOW (20 SECONDS)
// Longest delay, so any maps older then this will be discarded from the subsystem cache
#define MAP_REUSE_SLOWEST (60 SECONDS)
463 changes: 0 additions & 463 deletions code/__HELPERS/path.dm

This file was deleted.

306 changes: 306 additions & 0 deletions code/__HELPERS/paths/jps.dm
Original file line number Diff line number Diff line change
@@ -0,0 +1,306 @@
/**
* This file contains the stuff you need for using JPS (Jump Point Search) pathing, an alternative to A* that skips
* over large numbers of uninteresting tiles resulting in much quicker pathfinding solutions. Mind that diagonals
* cost the same as cardinal moves currently, so paths may look a bit strange, but should still be optimal.
*/

/// A helper macro for JPS, for telling when a node has forced neighbors that need expanding
/// Only usable in the context of the jps datum because of the datum vars it relies on
#define STEP_NOT_HERE_BUT_THERE(cur_turf, dirA, dirB) ((!CAN_STEP(cur_turf, get_step(cur_turf, dirA), simulated_only, pass_info, avoid) && CAN_STEP(cur_turf, get_step(cur_turf, dirB), simulated_only, pass_info, avoid)))

/// The JPS Node datum represents a turf that we find interesting enough to add to the open list and possibly search for new tiles from
/datum/jps_node
/// The turf associated with this node
var/turf/tile
/// The node we just came from
var/datum/jps_node/previous_node
/// The A* node weight (f_value = number_of_tiles + heuristic)
var/f_value
/// The A* node heuristic (a rough estimate of how far we are from the goal)
var/heuristic
/// How many steps it's taken to get here from the start (currently pulling double duty as steps taken & cost to get here, since all moves incl diagonals cost 1 rn)
var/number_tiles
/// How many steps it took to get here from the last node
var/jumps
/// Nodes store the endgoal so they can process their heuristic without a reference to the pathfind datum
var/turf/node_goal

/datum/jps_node/New(turf/our_tile, datum/jps_node/incoming_previous_node, jumps_taken, turf/incoming_goal)
tile = our_tile
jumps = jumps_taken
if(incoming_goal) // if we have the goal argument, this must be the first/starting node
node_goal = incoming_goal
else if(incoming_previous_node) // if we have the parent, this is from a direct lateral/diagonal scan, we can fill it all out now
previous_node = incoming_previous_node
number_tiles = previous_node.number_tiles + jumps
node_goal = previous_node.node_goal
heuristic = get_dist(tile, node_goal)
f_value = number_tiles + heuristic
// otherwise, no parent node means this is from a subscan lateral scan, so we just need the tile for now until we call [datum/jps/proc/update_parent] on it

/datum/jps_node/Destroy(force, ...)
previous_node = null
return ..()

/datum/jps_node/proc/update_parent(datum/jps_node/new_parent)
previous_node = new_parent
node_goal = previous_node.node_goal
jumps = get_dist(tile, previous_node.tile)
number_tiles = previous_node.number_tiles + jumps
heuristic = get_dist(tile, node_goal)
f_value = number_tiles + heuristic

/// TODO: Macro this to reduce proc overhead
/proc/HeapPathWeightCompare(datum/jps_node/a, datum/jps_node/b)
return b.f_value - a.f_value

/datum/pathfind/jps
/// The movable we are pathing
var/atom/movable/caller
/// The turf we're trying to path to (note that this won't track a moving target)
var/turf/end
/// The open list/stack we pop nodes out from (TODO: make this a normal list and macro-ize the heap operations to reduce proc overhead)
var/datum/heap/open
/// The list we compile at the end if successful to pass back
var/list/path
///An assoc list that serves as the closed list. Key is the turf, points to true if we've seen it before
var/list/found_turfs

/// How far away we have to get to the end target before we can call it quits
var/mintargetdist = 0
/// If we should delete the first step in the path or not. Used often because it is just the starting tile
var/skip_first = FALSE
///Defines how we handle diagonal moves. See __DEFINES/path.dm
var/diagonal_handling = DIAGONAL_REMOVE_CLUNKY

/datum/pathfind/jps/proc/setup(atom/movable/caller, list/access, max_distance, simulated_only, avoid, list/datum/callback/on_finish, atom/goal, mintargetdist, skip_first, diagonal_handling)
src.caller = caller
src.pass_info = new(caller, access)
src.max_distance = max_distance
src.simulated_only = simulated_only
src.avoid = avoid
src.on_finish = on_finish
src.mintargetdist = mintargetdist
src.skip_first = skip_first
src.diagonal_handling = diagonal_handling
end = get_turf(goal)
open = new /datum/heap(/proc/HeapPathWeightCompare)
found_turfs = list()

/datum/pathfind/jps/Destroy(force)
. = ..()
caller = null
end = null
open = null

/datum/pathfind/jps/start()
start = start || get_turf(caller)
. = ..()
if(!.)
return .

if(!get_turf(end))
stack_trace("Invalid JPS destination")
return FALSE
if(start.z != end.z || start == end ) //no pathfinding between z levels
return FALSE
if(max_distance && (max_distance < get_dist(start, end))) //if start turf is farther than max_distance from end turf, no need to do anything
return FALSE

var/datum/jps_node/current_processed_node = new (start, -1, 0, end)
open.insert(current_processed_node)
found_turfs[start] = TRUE // i'm sure this is fine
return TRUE

/datum/pathfind/jps/search_step()
. = ..()
if(!.)
return .
if(QDELETED(caller))
return FALSE

while(!open.is_empty() && !path)
var/datum/jps_node/current_processed_node = open.pop() //get the lower f_value turf in the open list
if(max_distance && (current_processed_node.number_tiles > max_distance))//if too many steps, don't process that path
continue

var/turf/current_turf = current_processed_node.tile
for(var/scan_direction in list(EAST, WEST, NORTH, SOUTH))
lateral_scan_spec(current_turf, scan_direction, current_processed_node)

for(var/scan_direction in list(NORTHEAST, SOUTHEAST, NORTHWEST, SOUTHWEST))
diag_scan_spec(current_turf, scan_direction, current_processed_node)

// Stable, we'll just be back later
if(TICK_CHECK)
return TRUE
return TRUE

/datum/pathfind/jps/finished()
//we're done! turn our reversed path (end to start) into a path (start to end)
found_turfs = null
QDEL_NULL(open)

var/list/path = src.path || list()
path = reverseList(path)
switch(diagonal_handling)
if(DIAGONAL_REMOVE_CLUNKY)
path = remove_clunky_diagonals(path, pass_info, simulated_only, avoid)
if(DIAGONAL_REMOVE_ALL)
path = remove_diagonals(path, pass_info, simulated_only, avoid)
if(skip_first && length(path) > 0)
path.Cut(1,2)
hand_back(path)
return ..()

/// Called when we've hit the goal with the node that represents the last tile, then sets the path var to that path so it can be returned by [datum/pathfind/proc/search]
/datum/pathfind/jps/proc/unwind_path(datum/jps_node/unwind_node)
path = new()
var/turf/iter_turf = unwind_node.tile
path.Add(iter_turf)

while(unwind_node.previous_node)
var/dir_goal = get_dir(iter_turf, unwind_node.previous_node.tile)
for(var/i in 1 to unwind_node.jumps)
iter_turf = get_step(iter_turf,dir_goal)
path.Add(iter_turf)
unwind_node = unwind_node.previous_node

/**
* For performing lateral scans from a given starting turf.
*
* These scans are called from both the main search loop, as well as subscans for diagonal scans, and they treat finding interesting turfs slightly differently.
* If we're doing a normal lateral scan, we already have a parent node supplied, so we just create the new node and immediately insert it into the heap, ezpz.
* If we're part of a subscan, we still need for the diagonal scan to generate a parent node, so we return a node datum with just the turf and let the diag scan
* proc handle transferring the values and inserting them into the heap.
*
* Arguments:
* * original_turf: What turf did we start this scan at?
* * heading: What direction are we going in? Obviously, should be cardinal
* * parent_node: Only given for normal lateral scans, if we don't have one, we're a diagonal subscan.
*/
/datum/pathfind/jps/proc/lateral_scan_spec(turf/original_turf, heading, datum/jps_node/parent_node)
var/steps_taken = 0

var/turf/current_turf = original_turf
var/turf/lag_turf = original_turf
var/datum/can_pass_info/pass_info = src.pass_info

while(TRUE)
if(path)
return
lag_turf = current_turf
current_turf = get_step(current_turf, heading)
steps_taken++
if(!CAN_STEP(lag_turf, current_turf, simulated_only, pass_info, avoid))
return

if(current_turf == end || (mintargetdist && (get_dist(current_turf, end) <= mintargetdist)))
var/datum/jps_node/final_node = new(current_turf, parent_node, steps_taken)
found_turfs[current_turf] = TRUE
if(parent_node) // if this is a direct lateral scan we can wrap up, if it's a subscan from a diag, we need to let the diag make their node first, then finish
unwind_path(final_node)
return final_node
else if(found_turfs[current_turf]) // already visited, essentially in the closed list
return
else
found_turfs[current_turf] = TRUE

if(parent_node && parent_node.number_tiles + steps_taken > max_distance)
return

var/interesting = FALSE // have we found a forced neighbor that would make us add this turf to the open list?

switch(heading)
if(NORTH)
if(STEP_NOT_HERE_BUT_THERE(current_turf, WEST, NORTHWEST) || STEP_NOT_HERE_BUT_THERE(current_turf, EAST, NORTHEAST))
interesting = TRUE
if(SOUTH)
if(STEP_NOT_HERE_BUT_THERE(current_turf, WEST, SOUTHWEST) || STEP_NOT_HERE_BUT_THERE(current_turf, EAST, SOUTHEAST))
interesting = TRUE
if(EAST)
if(STEP_NOT_HERE_BUT_THERE(current_turf, NORTH, NORTHEAST) || STEP_NOT_HERE_BUT_THERE(current_turf, SOUTH, SOUTHEAST))
interesting = TRUE
if(WEST)
if(STEP_NOT_HERE_BUT_THERE(current_turf, NORTH, NORTHWEST) || STEP_NOT_HERE_BUT_THERE(current_turf, SOUTH, SOUTHWEST))
interesting = TRUE

if(interesting)
var/datum/jps_node/newnode = new(current_turf, parent_node, steps_taken)
if(parent_node) // if we're a diagonal subscan, we'll handle adding ourselves to the heap in the diag
open.insert(newnode)
return newnode

/**
* For performing diagonal scans from a given starting turf.
*
* Unlike lateral scans, these only are called from the main search loop, so we don't need to worry about returning anything,
* though we do need to handle the return values of our lateral subscans of course.
*
* Arguments:
* * original_turf: What turf did we start this scan at?
* * heading: What direction are we going in? Obviously, should be diagonal
* * parent_node: We should always have a parent node for diagonals
*/
/datum/pathfind/jps/proc/diag_scan_spec(turf/original_turf, heading, datum/jps_node/parent_node)
var/steps_taken = 0
var/turf/current_turf = original_turf
var/turf/lag_turf = original_turf
var/datum/can_pass_info/pass_info = src.pass_info

while(TRUE)
if(path)
return
lag_turf = current_turf
current_turf = get_step(current_turf, heading)
steps_taken++
if(!CAN_STEP(lag_turf, current_turf, simulated_only, pass_info, avoid))
return

if(current_turf == end || (mintargetdist && (get_dist(current_turf, end) <= mintargetdist)))
var/datum/jps_node/final_node = new(current_turf, parent_node, steps_taken)
found_turfs[current_turf] = TRUE
unwind_path(final_node)
return
else if(found_turfs[current_turf]) // already visited, essentially in the closed list
return
else
found_turfs[current_turf] = TRUE

if(parent_node.number_tiles + steps_taken > max_distance)
return

var/interesting = FALSE // have we found a forced neighbor that would make us add this turf to the open list?
var/datum/jps_node/possible_child_node // otherwise, did one of our lateral subscans turn up something?

switch(heading)
if(NORTHWEST)
if(STEP_NOT_HERE_BUT_THERE(current_turf, EAST, NORTHEAST) || STEP_NOT_HERE_BUT_THERE(current_turf, SOUTH, SOUTHWEST))
interesting = TRUE
else
possible_child_node = (lateral_scan_spec(current_turf, WEST) || lateral_scan_spec(current_turf, NORTH))
if(NORTHEAST)
if(STEP_NOT_HERE_BUT_THERE(current_turf, WEST, NORTHWEST) || STEP_NOT_HERE_BUT_THERE(current_turf, SOUTH, SOUTHEAST))
interesting = TRUE
else
possible_child_node = (lateral_scan_spec(current_turf, EAST) || lateral_scan_spec(current_turf, NORTH))
if(SOUTHWEST)
if(STEP_NOT_HERE_BUT_THERE(current_turf, EAST, SOUTHEAST) || STEP_NOT_HERE_BUT_THERE(current_turf, NORTH, NORTHWEST))
interesting = TRUE
else
possible_child_node = (lateral_scan_spec(current_turf, SOUTH) || lateral_scan_spec(current_turf, WEST))
if(SOUTHEAST)
if(STEP_NOT_HERE_BUT_THERE(current_turf, WEST, SOUTHWEST) || STEP_NOT_HERE_BUT_THERE(current_turf, NORTH, NORTHEAST))
interesting = TRUE
else
possible_child_node = (lateral_scan_spec(current_turf, SOUTH) || lateral_scan_spec(current_turf, EAST))

if(interesting || possible_child_node)
var/datum/jps_node/newnode = new(current_turf, parent_node, steps_taken)
open.insert(newnode)
if(possible_child_node)
possible_child_node.update_parent(newnode)
open.insert(possible_child_node)
if(possible_child_node.tile == end || (mintargetdist && (get_dist(possible_child_node.tile, end) <= mintargetdist)))
unwind_path(possible_child_node)
return
Loading

0 comments on commit 144774a

Please sign in to comment.