* Change AtomTressFX passes to write to exsiting DepthLinear buffer instead of creating a new one as output.
Signed-off-by: Qing Tao <55564570+VickyAtAZ@users.noreply.github.com>
* Added new api in RPI to apply render pipeline changes from FP
Signed-off-by: Qing Tao <55564570+VickyAtAZ@users.noreply.github.com>
* Update AtomTressFX gem to create hair parent pass at runtime
Signed-off-by: Qing Tao <55564570+VickyAtAZ@users.noreply.github.com>
* Change Scene::ApplyRenderPipelineChange() to TryApplyRenderPIpelineChanges().
Have TryApplyRenderPIpelineChanges() called automatically when a render pipeline is added to a Scene.
Re-apply the render pipeline change when RenderPipeline got recreated (Pass hot-reloading support)
Add AddPassBefore() and AddPassAfter() function to RenderPIpeline class.
Signed-off-by: Qing Tao <55564570+VickyAtAZ@users.noreply.github.com>
GCC doesn't accept accessing member variables inside the noexcept expression of a member function.
GCC also requires that template specializations for an inner template be outside of all classes.
Signed-off-by: lumberyard-employee-dm <56135373+lumberyard-employee-dm@users.noreply.github.com>
- Use `unordred_map` (speed is about the same but providing a hash is slightly nicer than operator<)
- Do better arg forwarding where applicable
- Swap EXPECT_ test ordering
Signed-off-by: Nicholas Van Sickle <nvsickle@amazon.com>
* Up the hard coded size for the HpAllocator buffer to 18KiB
Added a static assert in the HphaSchema.cpp file to validate that the
HpAllocator aligned storage buffer is at least the size of the
HpAllocator class.
Signed-off-by: lumberyard-employee-dm <56135373+lumberyard-employee-dm@users.noreply.github.com>
* Removed comment blocks with username from HphaSchema.h
Signed-off-by: lumberyard-employee-dm <56135373+lumberyard-employee-dm@users.noreply.github.com>
Previously, with no level loaded in the main editor, and no assets loaded
in the EMotionFX editor, the AnimGraph viewport would only update at 15fps,
even on a 64 core machine.
Digging into this with Pix, I found that EMotionFX's UI would only get
updated for every other call to the ComponentApplication tick.
The Editor's QApplication instance controls how the
`ComponentApplication::OnTick` method is called. This is done in the
`maybeProcessIdle()` method. Generally, I found callstacks like this:
```
EditorQtApplication::maybeProcessIdle()
CCryEditApp::OnIdle()
CCryEditApp::IdleProcessing()
CGameEngine::Update() # ~2.2ms
PhysX::Update() # ~1.8ms
EMotionFX::Update() # ~0.2ms
calls QWidget::update() on all the widgets that change per
frame, only puts update events on the event queue, doesn't
paint anything
AZ::ComponentApplication::TickSystem() # ~25ms
renders the frame with Atom # ~24ms
```
The `maybeProcessIdle()` method is invoked by a QTimer, with a timeout
that changes depending on the application state. If the Editor is in
game mode, it used a timeout of 0, which essentially forced the game to
run at as high an fps as possible. If the Editor application has focus,
it used a timeout of 1ms, asking for the idle processing to happen at
1000 fps. Otherwise, it used a timeout of 10ms, asking for idle
processing at 100 fps.
Those fps targets are not realistic. What happened in this case is that
while the PhysX system was being updated, while the previous
`maybeProcessIdle()` call was still processing, the idle processing
timer would timeout again, and place a timer event on Qt's event queue,
before anything else had a chance to do anything. Only afterward would
EMotionFX's MainWindow::OnTick would be invoked, placing paint events on
Qt's event queue.
The fix for this is to use a single shot timer at the end of each
`maybeProcessIdle()` call. This ensures that any events that are
enqueued during the `maybeProcessIdle()` call are processed prior to the
next call to `maybeProcessIdle()`.
Signed-off-by: Chris Burel <burelc@amazon.com>
Fixed crash in Allocators Benchmark multithreaded test due to the HPHA schema not having proper multithread protection around the `mFreeTree` member in `tree_get_unused_memory` function.
The `mFreeTree intrusive set is able to modified on multiple threads.
Replaced the custom intrusive_list implementation n HPHA schema with AZStd::intrusive_list
Added a `ScopedAllocatorBenchmarkEnvironment` class to provide an RAI mechanism for initializing the SystemAllocator in Benchmark Test
Rermoved the `AzCoreBenchmarkEnvironment` in lieu of the `ScopedAllocatorBenchmarkEnvironment` class
Fixed assert when running Allocator Benchmarks in debug due to mismatch PauseTiming/ResumeTiming in Allocator Benchmark Fixtures
Added `ScopedRegisterBenchmarkEnvironment` RAII class to provide lifetime guarantees on BenchmarkEnvironments registered via the `AZ_UNIT_TEST_HOOK`
Initialized the intrusive_multiset_node members to nullptr in all build configurations instead of only debug as the cost negligible and it is useful for debugging.
fixes LYN-10210
Signed-off-by: lumberyard-employee-dm <56135373+lumberyard-employee-dm@users.noreply.github.com>
* Added Async APIs for the various Process*FromList terrain functions.
Please note that we are currently defaulting the number of worker threads to one, because splitting the work over multiple threads causes contention when locking various mutexes, resulting in slower overall wall time for async requests split over multiple threads vs one where all the work is done on a single thread. The latter is still preferable over a regular synchronous call because it is just as quick and prevents the main thread from blocking. This should be changed once the mutex contention issues have been addressed, so that async calls automatically split the work between available job manager worker threads, unless the ProcessAsyncParams specify a different desired number of jobs.
Signed-off-by: bosnichd <bosnichd@amazon.com>
* Fix Linux builds by adding missing #include
Signed-off-by: bosnichd <bosnichd@amazon.com>
* Added a test for cancellation of terrain async requests, and fix it so that it works.
Note that the benchmarks show this implementation to be slightly slower than the previous one, which I presume is because we're now calling a 'perSurfacePointFunction' in the inner loop; this can probably be addressed, but will result in a lot of code duplication, and I think efforts will be better spent on removing the mutex contention to enable running multiple terrain async jobs at the same time.
Signed-off-by: bosnichd <bosnichd@amazon.com>
* Added Async versions for all Process*Region terrain API functions, along with benchmarks.
Signed-off-by: bosnichd <bosnichd@amazon.com>
* Fix the newly added terrain async request benchmarks to actually use the async APIs.
Signed-off-by: bosnichd <bosnichd@amazon.com>
* Revert to the original version which just calls the synchronous API from the job function, along with some other updates in response to review feedback.
Signed-off-by: bosnichd <bosnichd@amazon.com>
* Change the TerrainWorldDebugger to use the async API, along with the following changes:
- TerrainJobContext no longer uses a JobCancelGroup so we can guarantee the completion callbacks of associated jobs will be invoked even if it is cancelled.
- As a result of the above change, the ProcessAsyncCompleteCallback function signature again accepts the associated TerrainJobContext as a param.
- The TerrainProcessAsyncCancellation test has been resurrected and simplified by using binary semaphores instead of condition variables.
- All the async related TerrainSystemBenchmark functions have been simplified by using binary semaphores instead of condition variables.
- Global cancellation of all terrain jobs on deactivation of the TerrainSystem has been reintroduced, but in a different way than before.
- Other miscellaneous changes/fixes made while testing and based on earlier PR feedback.
Signed-off-by: bosnichd <bosnichd@amazon.com>
* Updates based on review feedback:
- Go back to using a vector instead of an array (fixed the original problem by adding custom copy/assignment constructors/operators to the WireframeSector struct).
- When calling WireframeSector::Reset, block until any associated in flight has completed.
- Added the concept of a minimum number of positions per terrain job.
Signed-off-by: bosnichd <bosnichd@amazon.com>
* Use semaphore instead of binary_semaphore in a bunch of places to account for the race condition where a completion callback fires before we started waiting for it.
Signed-off-by: bosnichd <bosnichd@amazon.com>
This work has been on hold for a while. Revisiting removal of thumbnail contexts now while debugging related lockups exiting tools from script. Thumbnail contexts are an unused layer of indirection that makes debugging the system more difficult. All systems were only referencing the default context.
Signed-off-by: Guthrie Adams <guthadam@amazon.com>
Chris Galvan
amzn-mike
Allen Jackson
puvvadar
Qing Tao
John Jones-Steele
Scott Romero
Nicholas Van Sickle
lumberyard-employee-dm
Mike Balfour
Ronald Koppers
Danilo Aimini
Chris Burel
Olex Lozitskiy
Guthrie Adams
bosnichd
Tom Hulton-Harrop
greerdv
Guthrie Adams