branchrust_reboot/main/parallel_validatemovecancel
90 Commits over 92 Days - 0.04cph!
Optim: GamePhysics.HandleIgnoreTerrain - replace const seting managed loop with a Burst job
Saves 0.1ms on 10k test
Tests: ran unit test
Optim: allocate smaller buffers for WaterCollision.GetIgnoreIndirect
- Added a couple profiling scopes
Tests: unit tests
Bugfix: when we're missing TerrainCollision or WaterCollision, run all of the physics checks for trigger volumes
- Added UtilityJobs namespace, moved FillJob<T> to it
- Added a utility extension to NativeList<T> to copy from NativeArray<T>.ReadOnly (surprised it's missing)
Tests: all relevant unit tests
Clean: remove GamePhysics.OverlapSphere(batch) and GamePhysics.OverlapCapsule(batch)
- These have been inlined into CheckSphere<T> and CheckCapsule<T>
- Also updated CoarseQueryGrid.CheckJobIndirect(capsule) (returns JobHandle, works on a NativeList) since I missed it in previous commit
Tests: ran all unit tests
Update: CoarseQueryGrid - add non-indirect CheckJob API
- Also changed the API to return a JobHandle and accept a NativeList
Helps avoid extra busy work when we run direct work
Tests: ran unit tests
Bugfix: prevent WaterCollision.GetIgnore from returning garbage values
Another case that's the same as previous
Tests: none, trivial change
Clean: move all CoarseQueryGridJobs to it's own namespace
- rename the jobs to avoid duplication with namespace
Tests: compiles in editor
Bugfix: make sure TerrainCollision.GetIgnore doesn't return garbage values
Same bug as in WaterCollision that I fixed last friday.
Tests: none, trivial change
Bugfix: WaterCollision.GetIgnoreIndirect properly fills entire result set
Tests: ran unit tests + staging demo playback
Tests: adding TestWaterIgnoreConsistency to further narrow down what is going wrong
Tests: ran new unit test, it does fail as begrudgingly expected
Optim: FinalizeTickParallel - use a burst job to gather players to validate for tick history
- Cleaned up a bunch of code by switching to NativeList from NativeArray
- more TickInterpolatorCache usage
Tests: staging demo playback
Update: adding HandleIgnoreTerrain that works on a batch of positions
- also ninja-fixing a bug in batch overload of HandleIgnoreCollisions that screwed layermasks - good thing it wasn't used anywhere
Tests: ran unit tests
Merge: from parallel_validatemove/stable_playercache
Promoting stable player cache version as main dev direction
Tests: none, no conflicts
Bugfix: fix transform cache rebuild loop
Previously threw away dirty indices too early, missing on some player cache changes
Tests: played back the staging demo - no more error logs about transfomr cache inconsistency
Clean: fixing broken whitespaces that I spotted with the merge
Tests: none, trivial change
Merge: from main
Tests: none, no conflicts
Bugfix: PlayerCache now tracks players that are removed from the server
Original impl left a space where it could miss some of the players, leading to rogue nulls.
Tests: played back staging demo - had same numbers
Bugfix: one of validation checks was doing nothing
Tests: none, trivial change
Clean: removing BasePlayer OnDisable/-Destroy logs for UsePlayerUpdateJobs
Got confirmation that it's not Unity nuking the objects, so something else must be afoot.
Tests: built server locally
Merge: from main
Tests: none, no conflicts
Clean: fixing broken formatting that snuck in on main
Tests: none, trivial changes
Merge: from main
Tests: none, no conflicts
Debug: adding sanity debug error logs to baseplayer OnDisable and OnDestroy
Hoping to catch where it's coming from
Tests: local build tests
Bugfix: ServerDemoPlayer - handle player reconnecting multiple times
Not 100% sure it's the correct way, but I think it works for now.
Tests: played back new staging demo 3 times
Tests: replace Assert.AreEqual with Assert.IsTrue
- Brings TickIntrpolatorCache test from 30s+ down to 6s
Turns out AreEqual is slow and inflates the test times by quite a bit.
Tests: ran unit test
Optim: TickInterpolatorCache reduce number of segments being copied when growing
Unit test is still slow, need to dig a bit more
Tests: ran unit tests
Clean: fixing whitespace issues after auto-merge
Tests: none, trivial changes
Merge: from main
Tests: none, no conflicts
Debug: sprinkling additional validation checks in ServerUpdateParallel
Trying to narrow down at which point this null sneaks in
Tests: none, trivial change
New: TickInterpolatorCache - a sparse, bulk TickInterpolator array
- Comes with it's own stress tests (they pass, but need to investigate perf)
- Depends on PlayerCache, but I need to modify it to provide more stability
Building block towards jobifying tick history processing.
Tests: ran unit tests
Optim: replace couple managed loops with a burst job
Tests: none, trivial changes
Buildfix: move ValidateTransformCache to SERVER region
- also fixed missing Profiler.EndSample()
Tests: built client and server in editor
Merge: from main
Tests: none, trivial merge
Clean: promote server var to a const
- no codegen since I didn't do one when I added this
It's temp code, but it makes things safer while I investigate, so no reason to disable it at runtime
Tests: none, trivial change
Bugfix: purge player cache when player update jobs has emergency shutdown
Allows to restart player update jobs cleanly in the same session
Tests: in editor with debugger forced an emergency scenario, confirmed the cache was empty and rebuilt
Update: move player transform cache updates to be first step of parallel player processing
- Bugfix - using Steam networking backend can have temporary outtages, which can cause gaps in processing and player cache desyncing
- also enables us to compose parallel flows better in the future
Tests: local editor session on craggy
Update: add temp emergency disable of player job processing
- Only active on staging servers while I investigate the crashes
Tests: forced an error via debugger - confirmed fallback is working
Tests: move the PlayerCache stress test to it's relvant tests file and clean up
- commented out the expected-to-fail case
Tests: ran the tests
Update: adding throw away tests to investigate how I cuased a native crash yesterday
So far everything points to BasePlayer being removed after we cache all transforms for the burst jobs and before the first RecacheTransform invoke, but I haven't tracked where it's coming from.
Tests: ran the hacky unit test
