Inserting, updating, and removing entries from a VisibilityScene was made thread safe in a previous change, and now both the MeshFeatureProcessor and DiffuseProbeGridFeatureProcessor update entries at the same time from multiple threads. This leads to an assert in the Culling concurrency_checker, even though this is now valid behavior.
However, we still don't want to be adding, removing, or updating cullables between BeginCulling and EndCulling, which could cause a mismatch between the result of OctreeScene::GetEntryCount and the actual number of cullables in the scene.
-Added soft_lock_shared/soft_unlock_shared to the concurrency checker to allow multiple threads to acquire a lock when that is the desired behavior, while still asserting that nothing tries to acquire a shared lock when the concurrency checker is already locked.
-Update Culling.cpp to use the new soft_lock_shared when adding, updating, or removing cullables
-Added unit tests for the concurrency_checker
-Updated ArrayView unit test to use AZ_TEST_START_TRACE_SUPPRESSION instead of manually checking the assertion count, so that test now passes in release builds which do not assert.
- Enable the previously failing network unit tests for TcpTransportTests for Linux: TestSingleClient and TestMultipleClients
- Switch Linux to use SELECT implementation of the TCP Socket Manager instead of the EPOLL version (root cause of the failure)
While trying to process all slices and levels in Automated Testing, a few bugs came up that needed to be addressed:
- [LYN-3832] TransformComponent had a field removed without updating the version number and converter, which caused a lot of excessive warnings
- SliceComponent would crash in debug builds on instantiation failures due to a null dereference that was guarded against in most but not all places
- SliceConverter now detects when nested slices exist and gracefully warns about it.
- InstanceUpdateExecutor / TemplateInstanceMapper will now immediately remove instances that are unregistered, so that any in the queue don't get processed on a subsequent tick. This was causing crashes when the instance was destroyed before the processing occurred. It also has a side benefit of preventing the same instance from executing multiple times.
- Minor logic bugfix to the pack close warning, the boolean check was flipped.
Also added an early-out on SetTemplateId, since this was causing some unnecessary instance queue entries.
puvvadar
daimini
lumberyard-employee-dm
amzn-sj
karlberg
sconel
Vincent Liu
sconel
amzn-sean
sphrose
greerdv
AMZN-puvvadar
srikappa-amzn
SergeyAMZN
Chris Galvan
Tommy Walton
bosnichd
srikappa
Aaron Ruiz Mora
Tom Hulton-Harrop
pereslav
Benjamin Jillich
Steve Pham
Mike Balfour
hultonha