Check if we have enough space for num new links before starting the complex calculations. This optimization saved 10+ seconds on the firing range and had no effect on the output results; the navmesh still hashed to the same value as the one build before this patch.
Split traverse link creation and updating code from the editor, only alloc traverse table when we are creating them or when we load navmeshes that have them, don't alloc them before the navmesh is built. Also moved the disjoint set data to the editor class so we can cache the results and reuse them later after editing the navmesh post build. Traverse link algorithm now also ignores polys marked unlinked.
* Properly calculate how much we need to offset from the navmesh's edge to account for the ledge span
* Properly check if the edge we are connecting to overhangs our lower link position.
* Check if our ledge span offset causes the link to clip into geometry, and skip linking if so.
* Removed an improper intersection test in 'Editor::connectTileTraverseLinks', which caused many false positives (all issues this fixes, have been fixed with the improved 'traverseLinkInLOS'.
* Removed extraneous 'CanOverlapPoly' check; all links can now overlap poly's without issues if they pass through the 'traverseLinkInLOS' test.
* Only perform offsetting calculation if we have an amount to offset.
* Split 'dtCalcLinkDistance' into 2 functions, one that calculates it and 'dtQuantLinkDistance' that performs the quantization.
* Removed incorrect and experimental code from 'dtNavMesh::connectExtOffMeshLinks', this was to test off-mesh links for MSET 5, but this wasn't the correct approach.
Now connect to neighbor tiles (still work in progress). Also moved poly overlap and raycast logic to separate functions. The raycast checks are now performed as a very last step as previously we would raycast, and then determine we can't establish the link because the traverse type was NULL or the distance was too great, which both could be checked before the raycast and are also a lot cheaper to check. Changing the order of checks significantly improved the performance.
Also fixed a bug in dtCalcLinkDistance, if the number exceeds 2550.f, it would overflow and cause links to be a lot larger. If the value exceeds the number now, it would return 0 and no link will be established.
Traversing now works in-game, but there's still work needed to polish the algorithm and creating a proper lookup for jump types as currently only 1 jump type is supported.
Allow to compile the editor for separate versions of the navmesh set. NavMesh cells are now also partially reversed and properly read out. Creating and writing them will happen in the future.
Titanfall navmeshes have the dtOffMeshConnection::userId field changed from unsigned into to unsigned short, and the lower half seems to be used for the jump types and (possibly) some flag. Work in progress implementation of the jump system on off-mesh links.
The dtNavMesh structure members were marked as public, but they should be private. Made them private. Also, dtNavMesh::addTile was reconstructing all links, even when we add tiles with pre-existing links. The link building logic has been moved to a new member function 'dtNavMesh::connectTile'. This should only be called once the navmesh tile has been added succesfully during build. If we load a navmesh now, the existing links will be used from now on. The dtNavMesh::m_tileCount member is now also correctly incremented and decremented in addTile/removetile.
Some data that is in between the poly's and links weren't skipped properly. The engine skips this data in dtNavMesh::addTile, we didn't. This is now skipped properly by advancing over it before going over the links. The 2nd and 3rd bug was that the poly's end pointer, and off-mesh connections end pointer was set to the data size of the poly's and off-mesh connections rather than the poly and off-mesh con count. This small oversight has been fixed and the end pointers are now set correctly. The function dtCreateNavMeshData has also been updated to take this data into account correctly, it was using the Titanfall 2 NavMesh (NavMesh Set Version 5), structure sizes but this has been changed in Apex (NavMesh Set Version 7 and 8). These fixes did reduce the NavMesh's file size by a few kilobytes per megabyte.
There was an issue where Recast's common math library was less complete than that of Detour. Some common math operations were also isolated in specific translation units causing copies everywhere. All common math operations have been moved to the Shared library ultimately fixing all the above issues.
Everything 'dtAlloc', 'rcAlloc', 'dtAlloc', 'dtFree', got renamed into 'rdAlloc', 'rdFree'. There were a lot of object allocators that used these suffixes which were not accounted for causing those to be renamed. Everything has been renamed back to their original names (excluding the actual rdAlloc/rdFree functions as these were supposed to be renamed).
No code logic was changed in this commit.
The accidental renaming was caused in commit fa8d89d287752782ebdd5d9563f04fa72ef0bee9
- Separate disjoint set and traversal table building code
- Use game types and data to determine what to build into the traversal tables (e.g., 5 traversal tables for _small and only 1 for the rest).
- Automatically initialize the editor using the _small navmesh parameters instead of the broken defaults.
The static pathing code is split into 2 elements, the disjoint sets (poly groups), and the traversal table (which links poly groups, visually known as poly islands) together through jump or offmesh links. Reachability covers more than just traversal in the context of this game engine, therefore it has been renamed to avoid future confusion.
A more consistent name with what has bee reversed so far. Also added documentation on what has been reversed entirely and as much information that i have about stuff that has been partially figured out.
All recast/detour allocation and assertion code were identical, with the exception of their names (rc* for recast, dt* for detour). We want to use Recast's rcVectorBase class in Detour code, as there is no Detour equivalent, but copying it in its whole isn't good practice (especially considering there is more boilerplate code we want to get rid of in the future). Moved these to Shared so Detour could use it as well under the name rdVectorBase (rd stands for Recast Detour). No changes to the logic of the code were made in this patch.
Instead of building the static pathing data when the NavMesh is getting saved, build it right after the NavMesh has been build. also, parse this data out from NavMesh files loaded from the disk. This patch also documented some extra features and designs. In the future, static pathing logic will also be implemented in the detour path query code.
Used by the engine for AI wall running, but also other AI logic like move direction. This probably is some cached ref pos + yaw to save on computation in the runtime, and probably used to lerp the AI to the correct direction before jumping. Current implementation seems very close to original navs now, though, the original navs appear to have some yaw angles being rotated a bit, probably manual adjustments to make the AI face a wall that isn't perpendicular.
Also implemented debug drawing for the new ref positions, and made the start circle of the off-mesh connection red, and the end position green to make it easy what is what in bidirectional connections.
