Skip to main content

January

Monday 5th​

  • December notes and 2025 summary.
  • Planning long term work
  • Some optimisation work on orbital interpolation
  • Cleaning up issue tracker
    • Closed previous Epic (joust)
    • New epic (combat alpha)
  • Adding components to represent engine thrust and mass flow
    • Fixing everything that interacted with engine burns to use these

Tuesday 6th​

  • Investigating per-chunk dirty bits for Myriad.ECS
    • Need a counting bloom filter (to un-set dirty bits) if using bloom filter
    • Clearing a dirty bit that's already clear must be a no-op, so a counting bloom filter isn't good enough
    • If dirty bits aren't components this becomes a new dimension the safety system must handle
    • This seems like a lot of effort for marginal gain. Just doing chunk-shared components would probably be better.
  • Optimising bloom filter for fast query pre-filtering
    • Expanded from 384 buts to 512 bits
    • Sped up SIMD path by reducing copying
  • Added implicit cast from Entity to EntityId
  • Adding unique IDs to kepler bodies, so they can be identified across the network
    • Just use Wikidata IDs, parsed as uint32
  • Starting work on orbit "events"
    • Calculating "major" body and distance to it in integrator
    • Checking for change in major body (SOI change event)
    • Checking for change in sign of distance delta (periapsis/apoapsis)
  • Implementing equality for EntityId by hand instead of using record - compatible with Burst compiler

Wednesday 7th​

  • Storing generated events in rail pages
  • Updating inspector to show events
  • Investigating why major body in Luna orbit changes between Earth and Sol!
    • Could use hill sphere generated as part of planet asset. SOI isn't truly spherical though
    • Using GM/r3GM/r^3 instead of GM/r2GM/r^2 produces better results
  • Encoding rail page events into network package
    • Adding support for writing interleaved sequences of compressed doubles to HandySerialization.

Thursday 8th​

  • Testing multiplayer
    • Seems to be allocating an enormous amount
      • Bug in Myriad.ECS "Collect" queries (infinitely growing capacity)
      • Chasing down allocations in UI (TextMeshPro)
        • Editor only allocations of strings!
    • Bullets aren't properly synced?
      • They appear on clients, but look awful
    • Investigating poor performance
      • Reducing triangle counts on all planets (probably not actually important, but an easy win)
      • Bullet stuff is too expensive:
        • RailIntegrator
          • Looks like there are "urgent" entities blocking the queue
        • SetWorldPositionFromRail
          • This one is always expensive
        • SetScenePositionFromRelativeFuturePosition
          • Even with no entities matched!
          • Skipping more work if matched entity count is zero
          • Optimised FindKeplerObjects (span lists)
        • GPU bullet rendering
          • Early out if there are no bullet edges
        • Collisions are expensive
          • Got octree and SAP being setup - octree is used for rendering
  • Refactored rail events to store Kepler body ID instead of entity
    • Significantly simplifies networking!

Monday 12th​

  • Debugging why bullet rails are being marked as urgent
    • They integrate 0.5s ahead, but urgency is if they're below min length
    • Set min length to just a tiny bit below the pre-integrate time
  • Fixed name getting set every frame in the bullet test scene, which was ruining performance
  • Investigating cost of setting up bullet sweep & prune
    • Largest cost seems to be setting up the job, copying data into input buffers
    • Possibly due to checking entity structured
    • Checking the entire chunk instead of each entity individually (for aspect)
    • Created a "chunk aspect" to retrieve entity structure once per chunk too
  • Adding a way to check entity structure (HasComponent) for chunk handle to Myriad.ECS
  • Tested with 1300 active spaceships, top systems:
    • SetWorldPositionFromRail: 3ms
      • Use cheaper sampling modes for far away bodies?
        • Would need to be careful about collision detection quality. Not just distance from camera, but also proximity to other bodies
    • RailIntegrator: 1.1ms
    • CopyScenePositionToUnityTransform: 0.75ms, 0.6ms (2 different layers)
      • Don't do this work for things that don't need it (e.g. not in camera view)
    • CreateRailImpactSymbol: 0.8ms
      • Opt some entities out of this checking
        • Or specifically opt into allowing a rail end symbol
        • We don't actually want this for bullets
  • Fixed memory leak of orbital rail event collection if rail is modified while job is running

Tuesday 13th​

  • Sharing the array of kepler data between all jobs in integrator, instead of making a copy per job
  • Pre-integrating bullets for a bit more time, and shortening the min-rail length. Reduce the chance of bullets becoming urgent.
  • Opted bullets out of automatic creation of impact symbols
    • Saves time
    • Don't want these symbols for bullets anyway!
  • Tested in multiplayer again
    • Perf is better
    • Events aren't synced?
      • Setting events into page
      • Fixed syncing of null events
    • Possible memory leak in encoding/decoding
      • False positive as far as I can tell

Wednesday 14th​

  • Investigating pooling for tree view
    • Needs to be added in to the treeview library itself
  • Adding orbit picking panel to multiplayer game scene
    • Added gizmo creation systems
    • Changing orbit with gizmo works in multiplayer!
  • Upgrading to Unity 6.3
    • Installing editor
    • Upgraded project
    • Name collision with Entity/EntityId since they've been moved into UnityEngine
    • Reimport
      • Surprisingly fast!
    • Everything seems to work

Thursday 15th​

  • Testing in Unity 6.3 properly
    • Disabled old input system
    • Replaced some deprecated method calls
    • Updating packages
  • Added narrowphase colliders to bullet test scene
  • Investigating using the same sweep and prune collection for rendering as collision
    • Not really practical, S&P stores edges not points
  • Caching some more data about the next chain link entity
    • Using this cache to accelerate sweep and prune filling (don't need to follow the entity relation, it's all stored locally)
    • Speeds up S&P setup work by about ~30%, since it no longer has to query entity structure at all
  • Moved all structural checks to the QueryBuilder, instead of doing it dynamically within the query
    • About ~30% faster
    • Added helpers in Myriad.ECS to do this more cleanly for future queries

Friday 16th​

  • Doing some planning for the combat alpha
  • Investigating damage model system (optimising)
    • Using TryGetComponentRef
    • Adding measurements of cache hit rates
  • Adding TryGetComponentRef to Myriad.ECS - merges together several checks into one method
  • Disabled legacy Unity input system (it's being deprecated)
    • This broke some FUI stuff
      • Fixing usage in FUI framework

Monday 19th​

  • More fixing FUI framework to not use legacy Input system
  • Tweaking system list to show time relative to total update time (16ms)
  • Noticed potential resource leak in Job based queries
    • Added OnSystemDisabled callback to Myriad.ECS system groups and system group items
    • Use this in demo/test job based query systems
      • These aren't in use in the main game yet, need to be cautious when I do use them
  • Investigating how to dismiss orbit gizmo when clicking on empty space
    • There's already "backboards" at the camera far plane to capture clicks on nothing
      • Clicks on backboard are inconsistent?
        • Camera frustum in scene view is misaligned with the actual ray!?

Tuesday 20th​

  • Investigating weird camera frustum issue
    • Mouse.current.position.value returns different value when called in OnDrawGizmos!
    • Physics.Raycast returns different (correct) results compared to the camera PhysicsRaycaster!
    • PhysicsRaycaster calculates the distance to the far plane incorrectly, so it misses some collisions when the camera is rotated.
    • Unity discussion topic: https://discussions.unity.com/t/physicsraycaster-incorrect-ray-length/1704971

Wednesday 21st​

  • Working around physics raycaster bug
    • Adding a child camera, disabled, with far plane set to 1.5x parent camera. Physics raycaster is attached to this.
    • Created a script to manage this process
    • Producing a reproduction project to submit a Unity bug report
    • Applied workaround/hack with that extra camera to the camera compositing stack prefab
    • Clicking in empty space dismisses current selection. Finally!

Thursday 22nd​

  • Investigated adding "high watermark" to system time, so systems which spike leave a visible marker
    • Results are too noisy to be useful
  • Started creating system to create, update and destroy icons for orbit metadata
  • Added new TypedReference<T> component to Myriad.ECS
    • This helps with the frequent problem of following entity relationships being costly (due to multiple repeated checks) in two ways:
      • Checks are all done at once, combining some of the work
      • Check results are cached, so next time the link is cheaper to follow

Friday 23rd​

  • Refactoring TypedReference<T> into TypedEntityReference<T>
    • Code generation for multiple types, from 1 to 16
  • Refactoring some existing systems to use TypedEntityReference
    • Attach to planet surface
    • Attach to rail at future time
  • Noticed that orbit gizmo is in wrong position for some time (1 frame?) when first created
    • Looks like this has been an issue at least for a couple of weeks, so it wasn't broken by the TypedEntityReference changes
    • Moved prefab out to xyz=1000000, so it's not visible until the system places it in the right place one frame later

Monday 26th​

  • Working on systems for creating and maintaining rail metadata icons
  • Created a general purpose "zipper" to keep two rails in sync
    • This same pattern is used other places, could port there
      • Rail -> RelativeRail
      • Rail -> NetworkRail
    • Used this to create new "icon rail" that is zipped with the main orbital rail to create icons for all metadata events
    • Need to remove icons which are in the past

Tuesday 27th​

  • Removing rail icons which are in the past
  • Profiling relative rail attachment system
    • Adding hinting for rail sampling (caching some work)
    • Storing kepler hierarchy offset (caching more work)
    • Investigating why cache is refreshing more often than expected
      • It's fine, this is from picking results (mouse moving, busting cache)
    • Sharing CommandBuffer used for deletion with other systems
  • Using new TypedReference and TryGetComponentRef in more places:
    • SetWorldPositionFromSurfaceCoordinate
    • DrawBulletsWithGpuCompute
    • NativeEcsOctree
    • CopyWorldPositionBetweenEntities
    • CreateRailImpactSymbol
    • SetScenePositionFromWorldPosition
    • FindKeplerObjects
  • Fixing issue in Myriad.ECS that caused TryGetReference to not work the same as HasComponent+GetComponent

Wednesday 28th​

  • Investigating why adding orbit gizmo sometimes cuts line short but then doesn't show gizmo
    • Seems to be when gizmo is placed at a time past the line max time length (soft cap, so it can be exceeded in some cases)
    • Extending the line past this time makes the gizmo appear
      • It's out at xyz=1000000, i.e. position it not being set
    • Modified adding a burn to a rail to synchronously integrate up to the start of the burn right away: ensures rail isn't too short
    • Modified rail integrator to integrate up to max length or last_burn+extra_time whichever is later. This means there's always some rail showing what the last burn does
  • Using Fused-Multiply-Add in RKF45 integrator, faster and more accurate
  • Adding some extension methods to collections to do more work in jobs
    • Adding [BurstCompile] to these new jobs
    • Reviewing other jobs in codebase to burst compile where possible
  • Researching GPU driven text rendering
    • I can extract glyph info from text mesh pro
    • Instanced render quads with this glyph info
    • Potentially call TMP pixel shader, unclear if it can be imported?

Thursday 29th​

  • Updating Myriad version in ECS benchmark: https://github.com/friflo/ECS.CSharp.Benchmark-common-use-cases/pull/13
  • Cleaning up GetRefTuple in Myriad, pushed the core logic up into Chunk
  • Profiling SolarSystem scene
    • Removed allocation in ResizableComputeBuffer (using native memory)
    • Fixed closure allocation in ConvertPagedRailToIconRail
    • Investigating Cinemachine ResolveAndReadInputAction allocation - seems to be a profiler false positive
  • Building standalone player (more accurate profiling)
    • Investigating usage of double in NBodyLine shader (not supported on Vulkan)
    • Building normally works, it's only building a development build with "auto connect profiler" that does not!

Friday 30th​

  • Building development build (no profiler)
    • Fails with some Burst/AVX problems
      • Only from one particular AVX optimised method? Deleted it for now.
    • Noticed some allocations:
      • CinemachineInputAxisController.Update
        • Investigated this one before and been unable to resolve it
      • EventSystem.Update
        • Setting MaxRayIntersections on PhysicsRaycaster to fix this
      • BaseSimulationHost.Update
        • This is due to Unity not properly taking an allocation free path in struct allocation when generics are involved
    • OrbitPicking is very expensive
      • This needs to be jobbified or something
  • Experimenting with using uint2 instead of double
    • Writing a function to convert to single precision
  • Added loop unrolling to Myriad.ECS queries

Sunday 1st​

  • Bonus round!
  • Added a primitive "soft float" system just for doubles on Vulkan
    • Supported ops:
      • Convert to float using only integer maths
      • Compare 2 "soft doubles" to see which is greater