In the case where we have 2 cores, we would spawn only one thread in
the thread pool. If that thread got to try to steal() from another
thread before the main thread was adopted, it would end-up always
trying to steal from itself and enter an infinite loop.
This seems to happen during windows builds.
this is because the JobSystem's queue works as a LIFO, by creating
jobs in reverse (memory) order, we attempt to help streaming to
the d-cache on that threads -- until the point where
jobs are stolen.
we also execute the last job immediately instead of creating a job
for it -- since we're already in a job.
- parallel_for doesn't use recursion anymore to create the "leaf"
jobs, this is now done linearly on N thread (one thread per CPU).
This uses less stack space, and reduces miss-predicted branches.
- remove almost all SYSTRACE calls because they have a huge impact
on things like parallel_for() and are misleading. They can be
enabled again by setting HEAVY_SYSTRACE to true.
- we simplify the waiting code by using only a single
condition variable instead of two.
- wait() now behaves just like a looper, it will process jobs until
the one it's waiting for finishes -- before it could just sit there
(the idea was that the job would finish quickly, but that's not always
the case).
- we also make sure to never call notify_n() when it's not needed.
We track how many waiters we have and use that to decide if we need
to notify().
notify is pretty slow on all architectures, even on linux it's always
a syscall, so it's better to avoid it.
- don't use stand-alone fences, makes things ugly for no real benefit
- refactored the code a bit, hopefully it's more clear.
The Work-stealing dequeue indices could wrap around after ~2 billion
calls to steal(). This could probably be achieved in a few hours.
By using 64-bits indices, we avoid the problem entirely.
For jobs with that do very little work, the jobsystem can introduce
a lot of overhead, we mitigate this by:
- don't wake-up worker threads when scheduling several very small jobs,
like when scheduling the per-face jobs.
- don't wait for per-face jobs to finish -- we only did that to avoid
a copy of the job's data.
- don't use multi-threading at all if the job has too little work. We
evaluate the work using the scanline length and number of samples.
* Make CMake 3.10 the minimum version, add LTO option
* Install a newer CMake on Linux CI builds
* Update LLVM and Cmake on Windows CI
* Update build/windows/ci-common.bat
Co-Authored-By: Ben Doherty <benjdoherty15@gmail.com>
* Update formatting
* Apply suggestions from code review
* Update build/windows/ci-common.bat
* Update CMake
* Switch Android projects back to CMake 3.6
Even when hex modifier is used, 'char' should be printed as characters,
this is particularly relevant with 0.
e.g. out << (char)0, should write a nul terminator, not "0".
Instead of passing SamplerInterfaceBlocks and a SamplerBindingMap to
Program, we now set a 'sampler group' per binding point:
Program::addSamplerGroup(...)
A sampler group here consists of a list of N 'Sampler' and a 'Sampler'
is just a unique name (unifrom name in the shader) and binding point in
the shader.
That's all the driver layer needs.
With this change we get rid of the code that re-created the uniform
names in the driver -- this should never have been done there. And we
also remove the hash-map lookups in vulkan and metal drivers.
Program::shader() was taking a string before which didn't make sense
for spirv. Now it's just a blob, in the case of GL/Metal, the blob
must be a null terminated c-string, and the size must include the
terminating null character.
This fixes an out-of-bound access in ShaderBuilder::getShader() (which
doesn't exist anymore), because it was creating a CString passing
a size that included the null terminating char, which is not was CString
expects. CString can now assert() in that case.
driver::Program now uses a std::vector<> for storage, which we should
fix at some point (b/c it's a public header). CString was not suited to
store binary blobs.
This reverts a JobSystem optimization that attempted to avoid signaling
a condition when there was no waiters. Unfortunately, there was a
race that caused the the signaling thread to miss that the waiter flag
was set, thus not signaling.
Fixes a header usage problem for MSVC, which complains about StaticString not having a constexpr constructor thus making the "make" method not constexpr.
- only signal waitAndRelease() when the corresponding job finishes and
only if there is waitAndRelease() active -- instead of signaling
every time a job ends.
- don't surrender time slice when attempting to steal a job and it fails
as long as some queue has jobs.
- check that we have to wait, because taking the lock
- add a benchmark
This change more than doubles the amount of jobs we can handle per
second (~965,000 jobs/s on Pixel3)
JobSystem::waitAndRelease used to spin to wait for the job to finish,
usually this wasn't a problem because the spinning thread was
able to handle other jobs. However, in cases where no job was
available it would actually spin in burn cpu cycles.
we now use a (separate) condition variable to handle that case.
In a lot of case the StaticString hash can be computed a compile time,
so we now take advantage of that.
Removed StaticString(const char*) ctor, and replaced it with a
StaticString::make() method.
Fixed a couple wrong uses of the old StaticString(const char*) ctor.
We simply don't emit unwind tables, which are not needed anyways since
we're compiling without exceptions. the combined saving for all four
targets we support is about 120K.
This seems to improve .aar's compression, for a total gain of 152 KiB.
We also disable stack-protector in the jni code, since it wasn't
enabled in libfilament.a anyways. However, we now compile all debug
builds with -fstack-protector
- make Profiler::readCounters() not inline as it didn't need to be,
it's not performance critical and it's sufficiently large.
- don't inline hasExtensions(), same reason.
We used to only wake up a job-queue if there was already some jobs
running, the idea was that the current thread would handle the new job
as soon as calling wait(). However, there is no guarantee that wait()
will be called anytime soon.
cv.signal() is not very expensive on Android/Linux, as we're using
a custom implementation.
