* utils: add LRU cache to RefCountedMap
This change introduces a new data structure LruCache and uses it in
RefCountedMap to keep a fixed number of cache entries alive after their
reference count has dropped to zero in the main map.
* utils: address LRU cache comments
Instead of computing the fog "inline", in the forward pass, we can
instead compute it as post-process pass that is applied with a
simple fullscreen quad blending. On tilers, the operation entirely
stays in the tile, on desktop GPU it is a blending operation.
This works only for opaque materials.
The benefit is that fog will become immune to overdraw, and the forward
pass shader will be simplified, hopefully leading to less register
pressure. Overall performance should be improved.
Another benefit is that it will allow us to free the "fog" texture
slot from all opaque materials.
Transparent materials are unchanged.
This feature is currently DISABLED, and still work in progress; but it
should be mostly functional.
To test it:
```
env material.enable_fog_as_postprocess=true ./out/samples/gltf_viewer
```
This change refactor the fog code, but shouldn't have any impact on the
current behavior.
* engine: add program cache
This is another chunky change.
The core of this change is to cache programs in MaterialCache according to a
"specialization" (ProgramSpecialization) which is defined as the program cache
ID (the same key used for the OpenGL binary blob cache), the variant, and the
set of spec constants.
As part of this change, a lot of the implementation details of shader
compilation were refactored from Material to MaterialDefinition. The resulting
flow is a lot cleaner and easier to reason about, since shader compilation is
now a pure function of the MaterialDefinition + ProgramSpecialization.
Since the global cache program lookups might take a bit of time to compute
hashes, etc, I left the set of cached programs in Material as well, which kind
of acts like an L1 cache. The effect is that prepareProgram() and getProgram()
should be no slower than HEAD, even with the more complex caching requirements.
I'm planning on writing a document about this (and all changes up until this
point), but I'm being asked to work on higher priority things and I wanted to
have this PR out for review in the meantime so it doesn't bitrot.
* engine: fix unit tests
* engine: fix spec constants intern pool memory leak
* engine: address program cache comments
* engine: address more program cache comments
* engine: matdbg support for program cache
* engine: reinstate descriptorLayout calls
* engine: address bitrot
* engine: add feature flag to disable program cache
* engine: use material CRC32 for program cache
The "cache ID" of a material is supposed to uniquely identify a shader program
and all its variants. This is true to a certain extent, but does not account for
the code generation that happens at runtime. Two materials may have "identical"
shader programs, but due to each material's differing unique metadata, the final
compiled programs may end up very different. Unfortunately, this means we cannot
rely on the "cache ID" alone to determine a shader program's reusability.
Ideally, we should hash this "cache ID" with the exact set of changes to each
shader program so that we could reuse programs across materials. Instead, as a
stopgap solution, use the material's CRC32 instead.
* engine: fix double-free in program cache
* engine: address comments
* engine: assert_invariant empty material cache
PerformanceHint manager needs a java thread during initialization,
so we need to attach a jvm to the thread that's going to be used.
That thread is the filament backend thread, not necessarily the thread
the platform is created on.
So we make sure to do that from the backend thread.
FIXES=[427945768]
* feat(engine): Add automatic frame skipping to manage CPU/GPU latency
Introduces a new feature, disabled by default, that allows the engine to automatically skip frames when the CPU gets too far ahead of the display's refresh rate. This helps to reduce overall latency by preventing a backlog of frames from building up in the driver queue.
The feature can be enabled with the "engine.skip_frame_when_cpu_ahead_of_display" property.
To implement this, the compositor timing mechanism has been refactored. Instead of reporting an absolute `expectedPresentTime`, the backend now provides an `expectedPresentLatency` relative to vsync. This is more robust against synchronization issues with platform callbacks (like Android's Choreographer), as the latency is generally a constant value.
BUGS=[474599530]
* split FeatureFlagManager out of FEngine
- Separate the Feature flags management from FEngine.
- Make FeatureFlagManager available to backends through DriverConfig
- add a way to override a feature flag value at runtime using
environment variables or system properties (on Android).
e.g.:
```
env "feature.name=true" gltf_viewer
```
or
```
adb shell setprop debug.feature.name true
```
Co-authored-by: Powei Feng <powei@google.com>
* HandleAllocator::deallocate() was unsafe
It needs to know the concrete type to call the proper destructor, so
if it was given a base type handle (e.g. Handle<HwFoo>) it would not
destroy it properly.
* add AsyncJobQueue::cancelAll()
* Fix a race condition when tearing down FrameInfo
It is actually invalid to destroy a Handle<HwFence> while inside
fenceWait().
Updated the HwFence implementations so that they don't pretend they can
handle being destroyed during fenceWait(), they can't.
FrameInfo now cancels all the pending callbacks and waits for the
currently executing one to terminate, *before* destroying the
handles.
Reenable the gpuFrameComplete metric, as it should be working now.
* fix a possible deadlock in AsyncJobQueue
drainAndExit() could get stuck because it waited for the job
queue to be empty, but that was not signaled.
in fact, drainAndExit() didn't need to do that.
* Improvements to FrameInfo
- return the GPU Complete timestamp
- return the app VSYNC timestamp
- works TimerQueries are not supported
The VSYNC time is just a convenience as it is the same value
provided by the application during Renderer::beginFrame() or via
Renderer::setVsyncTime().
* Use utils::Status in MaterialParser
* Use utils::sstream instead of std::stringstream
* Remove remaining std::cerr and dep; update MaterialParser::reflectParameters
* make error message in utils::Status more generic
---------
Co-authored-by: Powei Feng <powei@google.com>
ImmutableCString is a string class similar to CString except it's
immutable. ImmutableCString occupies 16 bytes instead of 8 for CString.
However, ImmutableCString is able to avoid memory allocation when
constructed from a string literal, and in that way it us similar
to StaticString.
ImmutableCString can be auto converted from StaticString.
The backend tag tracking is updated to use ImmutableCString and
the FrameGraph resource manager us updated to use StaticString.
Together these changes significantly cut down heap allocations due to
internal tagging.
We also add optional tracking to {Immutable}CString.
* Minor changes in utils::Status
- << operator doesn't have to be friend
- simplify getErrorMessage to not use strlen internally
* Replace std::ostream to utils::io::ostream
* new utility AsyncJobQueue
this is a very simple job queue, it spawns a thread and runs the jobs
pushed to the queue in sequence.
* use AsyncJobQueue in OpenGLTimerQuery
- libbackend (except webgpu)
- libfilament
- libutils
std::string generates a lot of code bloat, we use CString instead.
We also update CString to be more compatible with std::string's api.
We link the behavior of certain precondition/postcondition checks to feature flag states.
- If provided *flag* is true, then this macro will assert when the *condition* is false.
- If provided *flag* is fase, then this macro will output a warning when the condition is false.
- If the condition is true, then neither of the above will happen
These two macros will enable us to provide less restrictive behavior for certain correctness assertions, allowing clients to modify their before enabling these assertions by default.
The addition JobSystem.cpp allows for defining
FILAMENT_TRACING_ENABLED across targets.
Addingin FILAMENT_TRACING_ENABLED to the #if in Tracing.h prevents
perfetto from being included.
* re-do "Use the Perfetto SDK instead of ATRACE" (#8701)
This time we create an entirely new private header: Tracing.h which
uses the perfetto SDK instead of ATRACE. The old Systrace.h is
unchanged to presever backward compatibility but is essentially
deprecated and no longer used within the filament repo.
The new TRACING_ macros use an explicit CATEGORY parameter, which is
declared in Tracing.h.
Moreover, tracing can be compiled out by defining
FILAMENT_TRACING_ENABLED to false before including Tracing.h
iOS tracing is still supported and still controlled via
FILAMENT_APPLE_SYSTRACE.
There are three perfetto categories defined:
- "filament/filament"
- "filament/jobsystem"
- "filament/gltfio"
The "filament/jobsystem" category is compiled out by default.
And they can be enabled in AGI / perfetto by adding:
```
data_sources {
config {
name: "track_event"
track_event_config {
disabled_categories: "*"
enabled_categories: "filament/filament"
enabled_categories: "filament/jobsystem"
enabled_categories: "filament/gltfio"
}
}
}
```
* Update libs/utils/include/private/utils/Tracing.h
Co-authored-by: Powei Feng <powei@google.com>
* Update libs/utils/src/android/Tracing.cpp
Co-authored-by: Powei Feng <powei@google.com>
* remove all references to SYSTRACE_TAG
---------
Co-authored-by: Powei Feng <powei@google.com>
We add rvalue version of append/insert/replace so that calling
those on a temporary yields to a temporary.
Also add string literal versions of those so that we can
append/insert/replace a literal without allocation, e.g.:
`CString foo = CString{ bar }.append("baz");`
This will not end-up creating a temporary CString for "baz".
Perfetto has significantly less overhead. The User facing API is
mostly unchanged:
Here are the differences:
- SYSTRACE_ENABLE() does nothing on ANDROID, initializes systraces on darwin.
- SYSTRACE_DISABLE() is removed.
- A new "gltfio" tag is added.
- SYSTRACE_TAG *must* be defined before including `utils/Systrace.h`
- `utils/Systrace.h` should not be used from a public header
- the new SYSTRACE_TAG_DISABLE disables systrace at compile time
For android a data source MUST be created in the perfetto config:
```
data_sources {
config {
name: "track_event"
track_event_config {
enabled_categories: ["filament", "jobsystem", "gltfio"]
disabled_categories: "*"
}
}
}
```
This can for example be added to AGI's custom/advanced config.
FIXES=[407572663]
* Remove redundant qualifiers in filament public headers
* remove redundant qualifiers in filament implementation
* remove redundant qualifiers in libutils public headers
* remove redundant qualifier for libutils implementation
* remove redundant qualifiers for libmath
* use is_same_v<> instead of is_same<>
* bring back Builder::name()
we keep Builder::name() on all object, and forward to the MixIn class
that does the implementation, so that we have correct documentation, and
better IDE completion.
* add missing const parameters in filament's implementation
* various source cleanup
- missing includes
- missing const
- C cast style
- superfluous inline keyword
* improve parallel_for a bit
We get about 40% performance increase. The gain comes from not having
to copy the JobData structure each time we create a job, by using a
new emplaceJob() method, we can create the structure directly into
its destination.
* avoid calling wakeAll() when possible
wakeAll() is very expensive and not always needed when a job finishes
because there may not be anyone waiting on that job.
We now maintain a waiter count per job, and use that to determine if
we need to notify or not.
And now that the JobSystem overhead is lower, we can decrease the size
of the jobs, which improves the load balancing.
* mActiveJobs fixes
some comments claimed mActiveJobs needed to be modified before or after
accessing the WorkQueue; this couldn't be correct because there were no
guaranteed global ordering with the workQueue.
- reduce the number of calls to notify_one() and notify_all().
notify_one() is not only called when running a new job, and
notify_all() only when a job finishes.
- don't hold the condition lock while calling notify_*(), as it is not
strictly needed, and because notify_*() can be very slow, there can
be a lot of contention on this lock as a result; blocking the whole
jobsystem thread pool.
- add a new version of run() that takes an opaque thread id that can
be retrieved from a job's execute function; this is especially
intended to be used by parallel_for(); it's just a more efficient
version of run() that avoids a hashmap lookup.
Overall these change yield a significant performance boost:
- running + waiting a job: +200%
- running many jobs: +150%
- running many jobs in parallel: +50%
going forward, instead of using the printf style syntax for panics
we use the c++ stream syntax
The new macros that replace ASSERT_*CONDITON are
FILAMENT_CHECK_PRECONDITON
FILAMENT_CHECK_POSTCONDITION
FILAMENT_CHECK_ARITIHMETIC
Example usage:
FILAMENT_CHECK_PRECONDITON(condition) << "Message";
It's also now possible to define FILAMENT_PANIC_USES_ABSL=1 to redirect
all these calls to Abseil's CHECK() macro.
using thread affinity naively on big.little architectures is very flaky,
for now it's better to simplify and not use it at all, let the kernel
figure things out.
BUGS=[333582569]
