RTS Devlog #1 – Back to native & Editor work

In 2018 I started experimenting with Unitys ‘DOTS’ (data orientated tech stack). I was failing to get decent performance in standard MonoBehaviour Unity for a experimental RTS project. Fast forward to 2021 with a few years experience of DOTs I’ve decided to switch to native C++. Previous posts on the blog were using Unity DOTs. The following post details my experience going back to a simple & not overly complex subset of C++ (ie. I not using all ‘modern’ C++ features if there’s no good reason too).

Where ‘native’ is used in this post it is referring to writing in a native language, in this case C/C++, as opposed to Unity C#.

Why switch?

Some of the reasons for switching are not really Unity specific as such, but more just general 3rd party engine/middleware things. As previous blog posts showed, I found myself rewriting the navigation (wanted a deterministic, higher performing dynamic 2d planar navmesh), animation (to cope with thousands of characters), map editor (needs to work in standalone and be usable by players) and eventually even the maths and physics (determinism and other reasons).

The breaking point came writing a delaunay triangulation implementation and having to hack quite a bit of C# code with unsafe regions & just being limited in general by the DOTS container types. I did it, and got it working (see previous blog post), but it was a total square peg in a round hole. I didn’t really feel comfortable with the state of the code and it’s maintanability into the future. This really made me re-evaluate what I was doing. Why go to all this effort if I can’t run it in 1 year without even more work, let alone 3 or 6 years later. My gut feeling is that code written in C/C++ is going to require less catch-up work in 2, 5 or 20 years into the future compared with a 3rd party middleware specific language, like C# DOTs (DOTs is effectively like using another language compared to normal C#). I like not having to worry about the future for DOTs or Unity anymore, I like being able to implement features myself more, and most of all I like not worring about middleware going the way of Microsofts XNA (dead).

Experience going back to C/C++

Despite all this, my initial plan was to convert only the simulation code to C++. I would create a plugin and then continue to leverage a 3rd party engine for graphics & other generalist things. However, I started enjoying programming in a native project again so much that I decided to convert all the way, graphics and all. I’m not a C/C++ super-fan, but it’s still a very viable language for game dev. I like straight-forward code that compiles fast and runs quickly. This can be achieved with C/C++ by keeping code simple, well organised and not over-using certain features of the language.
My iteration time with Unity from a single line change in the average gameplay implementation file to the app running was 13 seconds. In the C++ project it’s 0.5-2 seconds. This alone makes working on the project so much more enjoyable and dramatically cuts down on distractions. Obviously the iteration time may vary as the project continues but I don’t see it changing that drastically at this point. Having said all that, I can still see the pros of working with a shared general engine in a larger team or if the engine is a better fit to the project, but it’s not the use-case for this project at the moment.

Native app task list

There was suddenly quite a few things to code. Fortunately I’ve done most of these things before at somepoint so it wasn’t long before I had a equivalent app in C++. And eventually I passed the features of the Unity version (not Unity the engine, obviously! But the things I need in this project). Here is a limited list of the things I implemented in approximate chronological order:

  • App input, window setup (using GLFW)
  • Basic primitive debug drawing API (OpenGL for now), Camera
  • Custom memory allocation setup, persistent or frame/scratch allocators
  • Custom container classes (Stretchy array, hashmap, multihashmap) that can take custom allocators & other things to avoid std lib
  • Initial heightmap terrain work
  • Hierarchical transform system
  • Import box2d & glm (I’d like to replace glm later with a leaner, smaller maths lib but for now it’ll do)
  • Fix point maths & convert box2d to fixedpoint math
  • Delaunay navmesh
  • All gameplay systems (pathfinding, boids, spatial partitioning, attacking and lots lots more)
  • Tried entt (ECS) then wrote my own simpler & basic entity manager with pools
  • Import ImGui, Notifications and setup debug UI
  • fbx loading, custom binary model format, instanced mesh rendering, skeletal animation
  • Various shaders for terrain, screenspace, skinned mesh & static mesh
  • Basic depth-buffer shadows to add some depth to objects in the scene
  • Minimalist in-app profiler
  • Custom level file format
  • Editor work, Translation & rotation gizmos, map save/load, Terrain brushes, prefab palette
  • Generic config variable file
  • Hot-reloading for the config var file & shaders
  • GameDatabase/ResourceManager & very basic prefab workflow stuff

I had done variations of most of these things before, so it was quite quick to get it all done. It was nice to have a firmer target this time round. A mistake I’d made in the past was over-engineering by trying to write a generalist engine, when I didn’t really need one. I focussed on actual requirements and semantically compressed code or tidied up as appropriate afterwards. With all the above in the project I am still able to have an iterative build time of ~1 second. Full rebuild or changing something in a precompiled header can take lots longer, but it’s very rare I need to do that.

Procedural obstacle problems

In my previous post: RTS Pathfinding 2, I spoke about generating polygon bounds from chasms in the terrain by comparing height deltas of adjacent heightmap cells. Initially, I planned to expand this to other terrain features such as mountains, but when implementing in practice it was quite hard to solve robustly. The images below show problems that come up when I tried to scan height deltas for hills that had ramps leading up to them. In the end I opted for a simpler solution of simply manually placing collider nodes.

Cliff painting

As mentioned above, procedural object generation of terrain features didn’t seem like a good choice anymore. It’s hard to precisely configure the obstacle shapes. There was a lot of noice generated from just using height deltas to draw polygons around mountains/chasms which ultimately meant awkward navmesh/rigid body obstacles were produced. I wanted to be able to control obstacle shapes very well so that I could focus on setting up gameplay/map design just how I wanted it. I didn’t want to worry if a polygon wrapped around a terrain feature correctly.

Lots of RTS games have ramps and cliffs. Cliffs allow for easily setting up mazes and directing the flow of units, & thus tactics across the map. Blizzard RTS games tend to have well defined cliffs. These are essential in a way for the Terran race in Starcraft2 which depends on ‘turtling’ up with strong base defences. Total Annihilation or Supreme Commander style RTS games tend to have lots more open maps but with quite interesting features. I wanted a bit of both but I thought the way Blizzard RTS games like Warcraft3/Starcraft2 maps were setup might provide for more interesting tactics & gameplay, which is why I added a cliff painting tool.

The cliffs are baked into the terrain with a preset height gradient brush. Each cliff covers 5×5 height map cells with strengths applied between a lower and upper cliff height. Ramps can also be painted to allow paths between cliff levels. Ramps are done in the same way but with a smoother elevation gradient. Rigid bodies are automatically generated at cliff boundaries. These will act as physical barriers to units at or below the cliff level, and will also be inserted into the navmesh as obstacles.

A more complex cliff base ramp with multiple entry ramps & visualised in the navmesh.

The cliff obstacles are 100% predictable allowing for the designer to focus on crafting the map. However, I didn’t want the game world to look too uniform and grid-like. I still wanted more varied terrain features, if only for background visuals, so I added a manual obstacle painting tool to the editor.

Terrain brushes

I added a terrain brush to the editor that allowed basic height altering operations such as stamp, average & invert and then sourced a few textures for it. This allowed me to very quickly paint down complex and interesting terrain around the map with a few clicks.

I painted down some mountainous terrain near the map boundaries and next to base cliffs. This made the cliffs appear to be merged into more natural terrain formations. I was quite happy with the end result which allows endless room for visual variation in the terrain with simple, predictable & 100% controllable obstacles for gameplay.

The map after painting mountains and chasms into the terrain.

Gameplay ready map

The focus now was to get a map ready that could be used for unit gameplay experimentation. This really just required obstacles for cliffs, mountains & chasms (navmesh only obstacle with no physical barrier). But it was nice to add some basic graphics in the form of terrain brushes to see these things better. I went ahead and threw together a very basic splatmap terrain shader too, so that I could visualise things a bit better.

I wanted one more map feature before starting with gameplay prototyping, which was prefab obstacles, such as rocks. In a Jonathan Blow stream I had previously seen a very neat looking prefab palette tool. I added a similar tool to my editor which lets me visually select objects to place on the map. I added a few rock meshes to the prefab database and scattered a few around a part of the map.

The map now feels ready to use as a base for some gameplay experimentation. The embedded video for this post shows more of the general C++ tech progress and map editing.

RTS Pathfinding 2: Dynamic Navmesh with Constrained Delaunay Triangles

Unity NavMesh uses Recast navigation under-the-hood which is a very good general solution for navigation. However, for the particular needs of the type of lockstep multiplayer RTS I’m interested in, I ran into a couple of issues.

Firstly, dynamic updates of the Unity NavMesh are slow. Unitys default NavMesh was taking ~30-50ms to add a single box obstacle into a ~50 vert mesh. This is after fiddling with the tiling & various settings, a few experiments & posting on the forum for help. Recast constructs a ‘voxel mould’ which makes it very adaptable to different shapes and meshes. I imagine it probably scales quite well in the worst case with more complex meshes, but that is not really a problem that needed solving for a low-poly RTS navmesh.

Secondly, AFAIK it’s not cross-platform deterministic. This would be required for a lockstep multiplayer RTS where all clients need to rebuild the navmesh locally.

Why not use a grid? Well, I could. But it’s slow in comparison to a navmesh for path searches, and a bit cumbersome with its low resolution model of the world. Although it’s worth mentioning that for beginners a grid is far easier to code & makes dynamic updates straight-forward and quick.

StarCraft 2 Navmesh using constrained delaunay triangulation

Edges of a Terrain chasm

The terrain can be shaped to influence the pathing, so I needed a solution that could generate a navmesh on terrain, much like Unitys built-in solution. This needn’t be especially fast as the terrain isn’t going to change shape whilst the game is been played. I carved out a deep chasm in the map editor as a test sample for the navmesh generation.

First, I detected an outline of the chasm by scanning the heightmap for changes in height from cell to cell over a certain threshold.

The most obvious potential problem here is that the heightmap outline has quite a few points and consists of several severe square corners. I don’t need this many points and ideally I would have far less so that the navmesh is simpler when building & using for navigation. The outline should be as simple as possible without being ‘wrong’, in so much as giving a jarring or unexpected experience to the player when a unit paths around it.

I saw a very quick and easy way to immediately get rid of some points, which was to replace corners consisting of 2 edges of 1 cell length with a single horizontal edge.

Corners with edges of length 1 collapsed to single horizontal edge

The next thing would help navigation by giving units a bit of extra room around the terrain chasm was a polygon line expansion. I started by offsetting all polygon line segments by their perpendicular vector * the expansion radius, and then calculating the ray intersection points of neighbour lines (4).

This left some special cases where lines went inwards of the polygon outline as can be seen in (4). To solve this I check all line segments for intersections against each other. If two lines intersect, then the vertices between the 2 lines are removed and replaced with the intersection point, the result of this step is displayed in (5). The vertex count was now 52.

Following corner removal & the outline expansion I applied a Douglas Peucker line simplification to further reduce the vertex count to 30. The Delaunay triangulation now had less processing to do and yet the outline of the chasm looked pretty reasonable. The expansion distance from the original height cells outline is 2 units giving it a fairly wide birth of the drop on most edges.

Why Dynamic?

One of the other main reasons for embarking on a custom navmesh was to have a faster solution for dynamic updates at runtime. I had seen that StarCraft2 achieved this by building a ‘base’ navmesh with all the static obstacles features in the level, and then each time a new dynamic obstacle is added to the navmesh they would rebuild the entire mesh by resetting to the base navmesh and inserting all dynamic obstacles. At least that’s my understanding of what they did based on listening to James Anhalt’s GDC talk AI-Navigation Its Not a solved problem.

There seemed room for improvement here. If my interpretation is correct SC2 was re-calculating several duplicate parts of the navmesh each time they add/remove a dynamic obstacle, when only a small subset need be updated. My plan was therefore to limit my navmesh updates to this smaller triangle subset around the obstacle that is been added/removed. After finally getting dynamic updates of the navmesh subsets working; I can definitely see why SC2 team accepted the presumably slightly slower but simpler solution of rebuilding the navmesh completely from an intermediate base state. Updating a subset of the navmesh dynamically, including the ability to remove as well as add obstacles is quite a hard problem to solve robustly and is fraught with edge cases and math precision issues (using a Q48.16 fixed point number type certainly didn’t help here, either). I easily spent more time on just the dynamic subset updating than I did getting a single navmesh to generate across the map. I’ll elaborate more on that later.

Delaunay Triangulation

The first goal was to get a constrained delaunay triangulation building across the entire map. I won’t go into details here as there is lots of easily searchable information available already, eg. here’s a great introduction to constrained delaunay triangulation in C# by Eric Nordeus, alternatively, here is another implementation by QThund.

Constrained Delaunay triangulation. Note the constraints around rocks & chasms in black & the crystals in magenta.

Sometimes there were overlapping obstacles in the map. When this happened I would first process them using boolean operations to get the union of the polygons. I won’t talk further about this either as there is information available elsewhere including open source implementations such as Clipper lib.

Inserting obstacles

My idea was to find the triangles in the existing mesh that intersect with an obstacle to be inserted, and then generate a delaunay triangulation for this subset of triangles with the addition of the new obstacle. Next, the outline of the subset would be carved out of the base mesh & the updated subset merged into the base mesh.

Append the regenerated mesh subsection to the original mesh, minus the cut-out of the old subset. Merge duplicate vertices and edges.

Removing obstacles

To remove an obstacle I would find all the triangles with centre points that intersected the obstacle, create a set of all the vertices in these triangles, and then additionally get all triangles that use any of these vertices. This extended triangle set defined the subset of the mesh that would need to be regenerated when the obstacle was removed.

Regenerate the hull outline without the obstacle constraints, then replace the subset back into the original mesh.

Edge cases & additional complexities

The basic ideas described above worked quite well but there were additional complexities I encountered, such as ensuring that other obstacles that overlapped the extended triangle set continued to have their edge constraints respected in the new mesh generation.

After getting obstacle insertion on a subset of the navmesh working, I was slightly disheartened to see a less than ideal triangulation that included long & very thin strips. Thankfully I realised this was only because the delaunay triangulation hadn’t been run across the mesh as a whole. Running the edge flipping algorithm across the entire mesh isn’t too expensive if ~98% of the edges are already in the correct state so after obstacle insertion/removal I would run the delaunay algorithm across the entire mesh once more to help keep the mesh triangulation in a better state.

After getting dynamic delaunay working I expanded the testing include a dynamic spinning obstacle rect. To simulate a dynamic obstacle I would insert the obstacle at a different transformation each frame (and destroy the previous frames version of the obstacle), eg. frame1: {add rect1}, frame2: {add rect2, remove rect1}. The rect would be rotated slightly each iteration and positioned at the latest cursor location, but otherwise the same size. This turned out to be a good basic test as it showed up lots of edge cases and bugs.

Only the blue region around the newly inserted obstacle is regenerated

Dynamic updating of the navmesh is a challenging engineering task. There were a lot of edge-cases, many math precision issues and compromises to overcome. It required a lot of careful step by step debugging using verbose visualisation to stay on top of what was happening. This in turn made it one of the most rewarding projects to work on once it was finally working in a robust manner.

Marking Obstacles

Having got a pretty robust fixed-point dynamic navmesh working I was happy to move onto the easier task; using the data in a practical way for pathfinding. To begin with, I needed to flag triangles that were underneath obstacles. This is done with a simple intersection test between the triangle centre & an obstacle polygon in updated subsets of the navmesh.

Obstacles in red

Triangle Spatial Partition

Before a path can be found we need to know which triangles the start and end positions of the path are in. There could be thousands of triangles across the map, and we don’t want to do an intersection test for everyone of them, so a cell based spatial partition map is used to keep this reasonably cheap to query. Each corner of a triangles AABB bounds are added to a multi-map. For large triangles, where the corners don’t include all the grid cells that the triangle area covers, the bounds of the triangle are iterated over by the grid cell width. I imagine there are better spatial partitioning solutions than this for triangles, but it was still an order of magnitude performance increase, easy to write-up and can always be looked into later if required.

The start & end positions are displayed by two red crosses. The dashed squares are the grid cells in the spatial partition that the search positions hash into. The start position need only check against 2 triangles (in white), where as the goal position checks ~20 triangles (in yellow).

Triangle A* Search

With the start and end triangles found, the triangle navmesh data could now be leveraged to get neighbours and write a basic heuristic for an A* algorithm (I used triangle centre distance for now). The A* search & triangle spatial partition were blessedely easy & quick to code up compared to the dynamic constrained delaunay code. The result of this stage gave me a nice path of triangles from the start position to the goal.

The starting position/triangle in white, the goal in yellow & the path of portal edges in blue.

Simple Stupid Funnel Algorithm

The above image shows a path of 14 nodes (triangles) to navigate around two obstacles across a direct 55m distance. This is a decent gain compared to a grid which would require possibly ~55 nodes or more (and that is just the path, not to mention the exponential increase in size of the A* open list), assuming each cell is 1 square metre. Also, as previously mentioned, the other benefit with triangles is that they allow for far more precision relative to the obstacle outlines, compared to a grid.

However, you can’t path based on triangles alone. A path of positions is required for units to be able to follow. This is where the simple stupid funnel algorithm, a type of ‘string pulling’ algorithm, comes in. A list of portal edges is collected by finding the shared edge between each adjacent triangle in the list and then SSF algorithm is used to find the corners in the path.

Edge portals in purple. Direct path in black.

Variable agent size

One of the nice properties of using constrained triangle edges for the portals is that the maximum agent size that can pass through is known by the length of the edge portal. I added an extra parameter to the A* search so that it would ignore edges with shorter lengths than the agent diameter.

The edges on path were now always long enough for an agent to pass through, however, the path positions can’t be right on the boundary of an obstacle, this would only work for zero size agents. To fix this I narrowed the portal edges by the radius of the agent on both ends so that the path corners were adjusted to be away from obstacles.

Agent radius 1. Small agents can hug closely around obstacles for a more efficient path.
Agent radius 4. The portal edges have narrowed so that the corners are further from the obstacles. This leads to a different & wider path near the rocks.
Agent radius 12. The larger agent has to take an even longer route around the crystals to reach its destination.

It should be noted I’m not sure this will work perfectly on all corners of the navmesh. I can already see a corner when the agent has a radius of 4 where it’s not quite far enough away from an obstacle between path nodes. My hope is that by allowing a little extra buffer on the radius size that this isn’t really an issue in practice.

MapLocation & RayCast

It’s often very useful to be able to get the nearest location available on a navmesh One use-case I had for this was when moving a group of units to a destination, their individual goal positions may vary slightly to within a ‘soft’ / group radius of the final path waypoint – I need to make sure these spread positions aren’t inside an obstacle. The MapLocation function does this by looking up triangles intersecting a circle of specified radius around the target position. It uses the spatial grid to test against only nearby circles. If the target position is inside an obstacle it will map to the nearest valid (not intersecting an obstacle) navmesh location.

Navmesh raycast can be useful for a few things too, including LOS (line of sight) queries. Agents periodically do LOS checks with waypoints further along the path to see if they can skip positions. In addition if an agent has LOS with a destination from the start then a path query may be avoided altogether.

MapLocation: The input position maps to the nearest valid navmesh location. The outer circle denotes the area within which triangle intersection tests are done.

Soft & hard waypoint radius

For groups of agents moving together along the same path it was useful to introduce the concept of a ‘soft’ radius. The soft radius is the maximum radius around a waypoint that may be compressed down to a hard radius depending on obstacles around a waypoint. The hard radius is the minimum radius that must exist along a waypoint on a path request. The hard radius limits the edges that can be traversed in a path request and is set to the largest units radius. The soft radius is set depending on the group size and can be useful to stop a large number of smaller units all competing for the exact same waypoint position.

The soft radius expands so groups of units don’t all compete for a single waypoint position. The hard radius defines the minimum space required on the path.

The soft radius extents are set based on the portal edge length. This method is working fairly well for now. Previously I was setting the soft radius extents after the SSFD string pulling and raycasting either way into navmesh obstacles to determine the extents, however, I found this to be far more error prone, particularly in a maze like structure, compared to using the path portal edges.

Closing Notes

There’s quite a bit of work in the form of maths & problem solving that goes into making a custom navmesh. However, it’s very useful (and probably necessary for some types of games) to be able to have full control of the inner workings when compared to a black-box solution. There is still probably lots of low hanging optimisation that could be done but even in the Unity editor with non-bursted C#, the navmesh updates ~50-100* than Unitys built-in solution for a single obstacle insertion. It’s worth saying again that this doesn’t mean the constrained delaunay navmesh is better in all situations, just that it’s a more well-suited tool for this particular task.

RTS Pathfinding 1 – Flowfields

Since Unity’s data-orientated tech stack (DOTS) appeared there have been quite a few tech demos testing it out. Writing flowfield implementations seems to have recently become popular in the niche crowd of early adopter Unity game tech enthusiasts. Flowfields were first used in games such as Supreme Command 2 and then later Planetary Annihilation. They make sense when you have very large numbers of agents path-finding. The initial cost for a single agent is going to be higher than using something like a NavMesh but as you get into the hundreds or thousands of agents they may scale better.

For an RTS pathfinding prototype I initially used Unity’s built in NavMesh with the hybrid-DOTs navmesh query API. It was very fast for path queries but struggled performance wise when updating the NavMesh with dynamic obstacles. When weighing up the pros/cons of various methods I thought flowfields could be well suited since my problem-case involves large numbers of units. The grid data-structure inherent with flowfields can also be useful for storing other map information such as additional movement costs, heights, island IDs (an island is a single isolated area of the the grid that can’t path to another island). The grid data structure is also quite quick to look-up information in. Finally, for an RTS style project, pathfinding is one of the most critical bits of tech and having full control & understanding over it with no blackboxes was desirable (which wasn’t the case with using Unitys out-of-the-box implementation) and I had some ambitious highly specific goals that I wasn’t sure would be accommodated by Unitys built-in solution.

Making a basic dijkstra flowfield is one thing but a robust RTS game can be a lot more involved, and may require a fair few additional items such as:

  • Flowfield tiles
  • Hierarchical A* or some other high level pathing between tile portals
  • Caching of paths & flowfield tiles
  • Eikonal solver for smoother cost field generation
  • Greater than 8 directions for each cell for smoother movement
  • Line of sight (LOS) flags for smoother movement
  • Parallelised generation of paths and flowfields

There’s lots of media on the internet of people doing a basic implementation but it’s hard to find many that are using it in more complex situations such as an RTS game. The basics of flowfields are already well explained in numerous blogs and youtube videos such as this great one at HowToRTS https://howtorts.github.io/2014/01/04/basic-flow-fields.html. Although I might touch on some basics below, in general the rest of this post will assume a knowledge of basic flowfields. The following is more a collection of notes & workings I collected recently whilst trying to implement flowfields in a prototype RTS, rather than a robust tutorial.

Dijkstra vs Eikonal cost spread

The Dijkstra flood filling algorithm is used to expand the cells of the flowfield. Given a starting cell, the path goal, it will spread to all valid neighbour cells.

Flowfield spread visualised with different iteration counts. Cost increases as colour changes outwards from centre.

Whilst running the Dijkstra algorithm you can calculate cost in a basic way by simply incrementally increasing as it spreads (like image 2 below), however, this can lead to diamond like patterns forming in the cost field. This may in turn lead to slightly geometric looking path movement.

Image 2. Dijkstra cost field with a very low cost colour wavelength, basically the colour changes more frequently as the cost changes. This can help to see cost-spread patterns better.

Smoother cost spread can be generated by using using an Eikonal solver as shown in image 3. This can lead to more optimal paths, especially if there are up to 16 possible directions per a cell available.

Image 3. Eikonal equation generating a smoother more spherical cost field.

A sample Eikonal equation implementation is available here: https://github.com/SCIInstitute/StructuredEikonal/blob/master/src/cuda_fim_kernel.cu

Flowfield Tiles

To create flowfields across the entire map area would be very costly so an initial optimisation is to only generate the flowfield across a high-level (or hierarchical) path. To accomplish this the map is divided up into tiles. Flowfields will be generated for the tiles that are on the hierarchical path. After some basic optimisation a single 50*50 flowfield area generates in ~0.3ms. The map itself is ten times this size. The map is represented by a 500*500 grid, which is divided into 100 50*50 tiles.

The underlying map grid. Green cells are passable, red cells are obstacles.

When a path is required, hierarchical A* is run through tile portals (explained more below). The result of this is a list of tiles that the path runs through. Flowfields are requested for these tiles and the units read the direction value on their cell to get to the destination.

Flowfield tiles along a path

Each of these map tiles has portals along the edges that act as hierarchical A* path nodes. These portal nodes are generated at initialisation or even offline entirely and loaded from disk. A better description of the portals is explained by Eiljah who did the SupCom2 pathfinding here: http://www.gameaipro.com/GameAIPro/GameAIPro_Chapter23_Crowd_Pathfinding_and_Steering_Using_Flow_Field_Tiles.pdf

A map tile with portal cells in pink and paths in green between the portals. The length of these paths is the cost of the edge in the hierarchical A* graph.
Green lines visualising the hierarchical graph edges

Other tips

Debug visualisation is very important obviously with this sort of work. It’s really worth putting some time in here to have a robust visualisation system to keep you sane & discover bugs quickly.

Using a branchless math.select in GetNeighbours to replace if branches gave a crazy 10* speed-up compared to the if else branching I was doing previously. Although this will vary on the compiler it’s probably worth giving some serious consideration to, profiling before & after.

eg. here is a snippet of the branching version of some code to find neighbour indices on a grid:

if (x > 0)
	int neighbourIndex = ((x - 1) + (gridWidth * z));
	neighbours [neighboursCount ++] = neighbourIndex;
	if (z > 0)
		neighbourIndex = ((x - 1) + (z - 1) * gridWidth);
		neighbours [neighboursCount ++] = neighbourIndex;

	if (z < (gridHeight - 1))
		neighbourIndex = ((x - 1) + (z + 1) * gridWidth);
		neighbours [neighboursCount ++] = neighbourIndex;

And the branchless version:

bool leftValid  = (x > 0);
bool rightValid = (x < (gridWidth - 1));
bool upValid    = (z < (gridHeight - 1));
bool downValid  = (z > 0);

neighbours [neighboursCount] = math.select (neighbours [neighboursCount], (x - 1) + (gridWidth * z), leftValid );
neighboursCount = math.select (neighboursCount, neighboursCount + 1, leftValid );

neighbours [neighboursCount] = math.select (neighbours [neighboursCount], (x - 1) + (z - 1) * gridWidth, leftValid && downValid );
neighboursCount = math.select (neighboursCount, neighboursCount + 1, leftValid && downValid );

neighbours [neighboursCount] = math.select (neighbours [neighboursCount], ((x - 1) + (z + 1) * gridWidth), leftValid && upValid );
neighboursCount = math.select (neighboursCount, neighboursCount + 1, leftValid && upValid );

Making sure to access large grid/tile based arrays in a cache aware way was very important. eg. if you are iterating over large arrays (such as flowfield tile data) make sure you are going in the right column/row major direction to avoid jumping about in memory, or just use 1 rather than 2 dimension loop statements. This was another thing I got wrong first but fixed after profiling and checking the code.

Writing the flowfield implementation started with the need to have more control over the pathfinding implementation but it ended up being a decent revision for writing faster data-orientated code.

Closing Notes – movement behaviours

Having got reasonably well optimised and usable flowfields working I soon found out I was still someway from having large group movement work without any issues. A single unit could get to its destination fine. A small group was ok. But when 100 or so units all had to move around a corner to a common destination they tended to cluster quickly and end up forming into a single/double file line, like a line of ants. This was not what I had in mind. My original goal was to achieve something like Starcraft2 which has with a very smooth & responsive level of control for group movement. Specifically, when going around a corner, I wanted the group to move around a centralised centre as one, more or less maintaining relative offsets to each other. A lot of this would be dramatically helped by adding some boids style localised forces. However, even after some initial work in that area I couldn’t quite get it to fix the problem entirely with large groups of units when using flowfields.

Formations forces applied in addition to the flowfield are a potential idea to solve this problem, but I’ve not been able to get this to work yet in practice. The group movement just didn’t look well suited to the situation. Although it could be more appropriate when trying to model an actual swarm of insects or liquid like things. I might revisit flowfields for swarm style units or multiple goal based paths (such as close combat fighting with groups) in the future.

I shelved the flowfield work whilst keeping it as an option to switch back to if required and started on a new approach. I went back to using a more traditional pathing solution involving waypoints at path corners which you can then apply offsets to and assign individual positions per unit. This method would also need a lot of boid forces interplay but the path destinations would hopefully be more precise at least.

Resisting the urge to go back to Unity NavMesh and its slow rebuild/update times, I started using my existing ‘NavGrid’ with plain old grid based A* searches. Although the path search cost for a grid is higher than I’d expect from a good NavMesh implementation the insertion or rebuild time for a grid based graph is trivial, and I had ambitious ideas that involved regular updates of the graph. I tweaked the heuristics and added a tie-breaker http://theory.stanford.edu/~amitp/GameProgramming/Heuristics.html to improve the efficiency. Then I calculated the path ‘corners’ or waypoints along the path.

The path corner positions could then be used to calculate position offsets of varying length depending on the group size allowing for better formation movement.

This will be continued in RTS pathfinding part 2.