this change shouldn't have any impact on ARM, however, according
to cppreference it's not safe to mix seq_cst with other memory
orders:
"as soon as atomic operations that are not tagged memory_order_seq_cst
enter the picture, the sequential consistency guarantee for the program
is lost"
* 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.
- each handle now has a 4-bits "age", meaning that handles are recycled
only after 16 alloc/free cycles.
This is used to detect double-free and use-after free.
This should also allow us to compare handles, because freeing and
reallocating an object, won't produce the same Handle (at least
for 16 rounds).
- removed "type safety" checks because it's almost impossible to
get it wrong thanks to our compile time type safety checks. This
didn't provide a useful value added.
- This feature is built on top of being able to set/get a 8 bits tag
associated with the memory block returned by the pool allocator. We
use the "extra" parameter of the allocator to allocate a "hidden"
structure containing the age of that memory block.
- Also we don't allow to compare Handle<> of different types
- update the pools sizes for metal and vulkan, which were very outdated.
- add debug code on all backends to print the size of each handle
(with a compile time switch)
The most important change is that now the 3 pools of HandleAllocator
are sized so that each can accommodate about the same amount of handles.
This makes it easier to reason about. The total amount of handles is
three times that, since there are 3 pools.
We also try to allocate the buckets so that handles are evenly
distributed, however, that's very hand wavy.
With the current setup the number of handles per pool is as follows:
- GL : 3240 / pool / MiB
- VK : 1820 / pool / MiB
- MTL: 1310 / pool / MiB
* Automatically flush CommandStream
When generating commands, we now automatically flush the CommandStream,
so that we're guaranteed to not overrun the circular buffer.
* clenaup CircularBuffer implementation and API
Also fix a bug in DEBUG mode that could corrupt the CircularBuffer, it
was due to a wrong debugging code attempting to clear the unused
area of the buffer (this was wrong because in "ashmem" mode, there are
no guaranteed unused areas).
* Fix a couple threading vs. allocations
- prepareVisibleLights was run on a dedicated thread (via JobSystem),
but was using its own local ArenaScope. This is wrong because it
could reset the root arena at any later point. This is fixed by
just not using a local ArenaScope.
- related to the above, the root Arena (LinearAllocatorArena) didn't
use a locked policy, which cause also cause problems since some
allocations are done off the main thread. We now pre-allocate the one
buffer we need.
This PR also renames some variable and types to improve readability.
* Rework RenderPass to improve allocations and API
RenderPass now is a fully immutable object that gets constructed with a
RenderPassBuilder. RenderPassBuilder can be passed around and doesn't
do any (major) allocations.
All RenderPass allocations and heavy lifting is done in
RenderPassBuilder::Build().
Additionally, RenderPass cannot be copied anymore.
Where allocations happen is now much clearer.
* new LinearAllocatorWithFallback
LinearAllocatorWithFallback is a linear allocator that can fall back
to the heap allocator. We use it for the high level command buffer to
avoid crashing when running out of memory.
FIXES=[277115740]
* Update filament/src/RenderPass.h
Co-authored-by: Powei Feng <powei@google.com>
* Update libs/utils/include/utils/Allocator.h
Co-authored-by: Powei Feng <powei@google.com>
---------
Co-authored-by: Powei Feng <powei@google.com>
This has caused issues and over time we have reduced the use of
spinlocks, it was only used in few places and we still have evidence
that it's causing ANRs.
We use utils::Mutex instead which is a low overhead mutex implementation
on Linux systems.
FIXES=[321101014]
Material constants (a.k.a: specialization constants) can only be set
during Material creation through Material::Builder.
This change somewhat relaxes that limitation by allowing constants to
be set at runtime on Material directly.
Currently this new API is still private and only supported on FMaterial.
This feature works by invalidating the HwProgram cache of the concerned
Material, causing a shader recompile per variant; so this API is costly
and should be used only for debugging or during app/game configuration.
The TAA material is modified to use constants instead of #defines for
various settings and those are exposed in TaaOptions as well is in
ViewerGui. So with this change all aspects of the TAA material can
be changed at runtime.
* prevent public classes from being created on the stack
- we used to to this by deleting operator delete, but this prevented
the internal "F" classes from being virtual; which can be useful
when using EntityManger::Listener.
now we just make the destructor protected in each class.
- EntityManger::Listener now has a virtual destructor so that
objects could be correctly destroyed from Listener*
* improve EntityManger and Component managers
- all component managers now have the same "base" API
- getComponentCount()
- empty()
- getEntity()
- getEntities()
- Scene now has getEntityCount()
- EntityManager now has getEntityCount()
- all component manager implement gc() the same way, by calling destroy()
- SingleInstanceComponentManager::gc() that calls removeComponent() has
been removed because it's dangerous. removeComponent() is often
not enough, some additional cleanup might be needed.
- don't rely on it being 32-bits
- update the jni code to store SamplerParams in a long (64 bits)
instead of a int. This gives us some future-proofing of the java side.
It supports KHR_parallel_shader_compile as well as a
thread pool of GL contexts.
- we have a new 2-priorities queue for shader compilation
- use this feature in gltfio in the ubershader case
