2,334 Commits over 580 Days - 0.17cph!
Merge: from optim_jobs2_builddepchains
- Optim: UsePlayerUpdateJobs2 - BuildDependencyChains is now 40%+ faster and scales better
Tests: unit tests, loaded into Craggy, loaded into 2.5k procgen
Update(tests): adding Perf_BuildSnapshotDependencyChains
Basic test churning throuhg 1 snapshot queue of 1k entities. Mostly shallow hierarchies are fast (20% parented -> 0.15ms), but dense hierarchies are slower(200% parented -> 0.6ms).
Tests: ran perf test
Update(tests): add BuildSnapshotDependencyChains unit tests
- DummyServer now automatically cleans up spawned entities in CleanTestState
Tests: ran unit tests
Clean: refactor some of unit tests
- take out DummyServer and move bot, player creation logic into it
- move out Measuring utils to it's own file
Tests: ran occlusiongroup tests, all pass
Optim: Jobs 2 - rewrite BuildSnapshotDependencyChains
- fix a bug that caused more snapshots to be sent than necessary (causing massive inflation)
- add a couple fast paths, rewrote it to be a single-pass algo
- reduced number of pooled lists usage
Need to write unit tests for this
Tests: with useplayerupdatejobs 2 set from start, loaded craggy and 2.5k procgen world in editor, visited ferry terminal - didn't see anything missing, npcs had weapons and were shooting
Merge: from serverprofiler_disable_memstates
- Update: added extra safety try-finally
Tests: none, trivial change
Update: missed an extra try-finally in perfsnapshot_stream
Tests: none, trivial change
Merge: from serverprofiler_disable_memstates
- Bugfix for running multiple capture commands without waiting for completion of initial
- Bugfix for memstats parsing exception (disabled memstats recording)
Tests: couple 128mb perfsnapshot_stream in editor
Bugfix: perfsnapshot/_stream - set exportDone to false when starting recording
- wrapped inner task logic into a try-finally to ensure exportDone latch is released
Well this is a derp up, the safety latch was in place for a year but never worked because I never engaged it. Whoops.
Tests: done perfsnapshot_stream 128mb as another one is exporting - no crash, got message
Update: ServerProfiler.Core - disable system memory tracking
- release binaries built from 25d14ef0
Its too unstable and overengineered, will rework later
Tests: perfsnapshot_stream 128mb of data - no system memory tracks in json
Merge: from tracespheres_optim
- Bugfix for exceptions when shooting
Tests: shot the ak on Craggy in editor
Bugfix: VerifyRays/Spheres - early out if passed empty command set
Fixes exceptions when shooting
Tests: shot ak on Craggy in editor
Merge: from fix_sortdeferred_race
- Bugfix for sorting of hits potentially running in a race with previously scheduled jobs
Tests: unit tests
Optim: merge count and sort jobs into one
Tests: unit tests
Bugfix: GamePhysics.SortDeferred - respect dependsOn handle param
Tests: ran unit tests - all pass
Merge: from spectate_stay_after_dc
- Update: when spectating a player that disconnects, the client starts orbiting their last position. Third-Person is the only valid camera mode in that situation
Tests: spectated a player that disconnected - was able to look around.
Update: client now hovers last spectated player's position in third-person mode when spectated player disconnects/gets nuked
Tests: spectated player that disconnected - started orbiting. spectated as only player on server - orbiting original position (as before)
Update: when spectated player disconnects, stay spectating above their player
Don't like how the logic is tied to the player/sleeper, need to improve on the clientside
Tests: spectated player, then that player disconnected - stayed in an area
Merge: from tracespheres_optim
- Optim: GamePhysics.TraceSpheres and TraceSpheresUnordered are now faster at 32 or more spherecasts (previously started scaling at slightly more than 128)
Tests: unit tests
Merge: from tracespheres_optim
- Optim: GamePhysics.TraceSpheres and TraceSpheresUnordered are now faster at 32 or more spherecasts (previously started scaling at slightly more than 128)
Tests: unit tests
Clean: remove hacky testing code
Tests: none, trivial change
Optim: GamePhysics.Trace - replace sort with a serial scan
- added unit tests and perf tests
In existing perf tests scene complexity is not enough, but expecting to be a minor improvement in real world.
Tests: unit tests
Optim: replicate TraceRays optims to TraceSpheres/-Unordered
TraceSpheres now scales better and earlier than it's serial counterpart:
* TraceSphereUnordered vs TraceSpheresUnordered: 128 - 0.57ms vs 0.3ms; 1024 - 5.1ms vs 1.38ms; 8196 - 43ms vs 9.9ms
* TraceSphere vs TraceSpheres: 128 - 0.64ms vs 0.44ms; 1024 - 5.6ms vs 2.2ms; 2048 - 11ms vs 4ms
Tests: unit tests
Bugfix: veeeeery rare out of bounds access when Unity returns invalid hit
Tests: none, don't have a unit test for that
Update: GamePhysics.TraceSpheres - add ignore entity support
Tests: ran unit tests
Update(tests): add TraceSpheres and TraceSpheresUnordered unit tests
- get rid of old TestSpheres test, didn't spawn base entities
Tests: ran new unit tests, some are failing as expected
Update: expanded GamePhysics.Perf tests with spheres (batched and serial)
- Fixed a perf bug of TestTraceRay causing it to do too much sorting
- Further refactored & generalized code
At 128 spheres, batched and serial seem to cost the same - going to continue moving ray optims across after I add unit tests
Tests: ran perf tests, numbers within expectation
Clean: more refactoring of unit tests in Test.GamePhysics
Tests: ran perf tests
Clean: take out perf test measuring code into helper Utils class
Tests: ran unit tests
Merge: from remove_old_occlusion_group_logic
- Clean: more code removal in unit tests to silence Rider
Tests: ran unit tests
Clean: remove more dead code/checks
Missed some leftover if(true) style code
Tests: ran unit tests
Merge: from main (yeah, weird one - redo of CL 143929)
When I did a server merge from extend_perf_telem, plastic warned me that another change made it to main, but commit bot reported mine merged. In reality it wasn't fully merged, just "attached". This should fix it.
Tests: compiler simulator in editor
Clean: leave a note why we're not running water traces and raycs concurrently
Tests: none, trivial change
Update: remove TestLineOfSightBatched unit test as we never implemented the batched version (I think?)
Tests: ran all GamePhysics.UnitTests - now they all green
Merge: from expand_perf_telem
- Update: object work queues now send their budget time + an aggregate record
- Update: frame_profiling now contains extra fields about invokes and a pre_lateupdate timing record
Needs a bit of backend work to fully finish this.
Tests: uploaded to backend, inspected what arrived
Update: couple analytics fixes:
- new TimeSpanExt.FromMicroseconds utility to get sub ms timespans
- changed CSV data pushing to object_work_queue_2 and frame_profiling_2
- moved BudgetTime to be last field in the CSV blob for object_work_queue_2
Tests: inspected csvs on backend
Merge: from remove_old_occlusion_group_logic
- Clean: ServerOcclusion now uses new occlusion group logic by default, old logic ripped out
Tests: ran unit tests
Clean: remove old occlusion group code
- removed -enable-new-server-occlusion-groups command line
- cleaned tests
Tests: ran occlusion group unit tests, all 144 pass
Merge: from delete_obsolete_pool_api
Tests: compiler simulator in editor
Clean: remove deprecated Pool api - they've stayed 1 year longer than supposed to
- replaced internal usages
* Pool.GetList<T> -> use Pool.Get<List<T>> instead
* Pool.FreeList<T> -> Pool.FreeUnmanaged
* Pool.FreeListAndItems<T> -> Pool.Free(ref obj, true)
* Pool.ClearList<T> -> write your own
* Pool.FreeMemoryStream -> Pool.FreeUnmanaged
Tests: compile simulator in editor
