The idea is to accumulate all new tasks in a thread local queue
first without doing any thread synchronization (aka, locks and
conditional variables) and move those tasks to a scheduler queue
once they are all ready. This way we avoid per-task-pool lock
and only have one lock per bunch of tasks.
This is particularly handy when scheduling new dependency graph
node children. Brings FPS of cached simulation from the linked
below file from ~30 to ~50.
See documentation for BLI_task_pool_delayed_push_{begin, end}
and for TaskThreadLocalStorage::do_delayed_push.
Fixes T50027: Rigidbody playback and simulation performance regression with new depsgraph
Thanks Bastien for the review!
We can not re-use anything for such pools, because we will know nothing about whether
the main thread is sleeping or not. So we identify such threads as 0, but we don't
use main thread's TLS.
This fixes dead-locks and crashes reported by Luca when doing playblasts.
Suspended pools allows to push huge amount of initial tasks
without any threading synchronization and hence overhead.
This gives ~50% speedup of cached rigid body with file from
T50027 and seems to have no negative affect in other scenes
here.
The idea is to allow some amount of tasks to be pushed from working
thread to it's local queue, so we can acquire some work without doing
whole mutex lock.
This should allow us to remove some hacks from depsgraph which was
added there to keep threads alive.
This allows us to avoid TLS stored in pool which gives us advantage of
using pre-allocated tasks pool for the pools created from non-main thread.
Even on systems with slow pthread TLS it should not be a problem because
we access it once at a pool construction time. If we want to use this more
often (for example, to get rid of push_from_thread) we'll have to do much
more accurate benchmark.
Basically move all thread-specific data (currently it's only task
memory pool) from a dedicated array of taskScheduler to TaskThread.
This way we can add more thread-specific data in the future with
less of a hassle.
This feature was adding extra complexity to task scheduling
which required yet extra variables to be worried about to be
modified in atomic manner, which resulted in following issues:
- More complex code to maintain, which increases risks of
something going wrong when we modify the code.
- Extra barriers and/or locks during task scheduling, which
causes extra threading overhead.
- Unable to use some other implementation (such as TBB) even for
the comparison tests.
Notes about other changes.
There are two places where we really had to use that limit.
One of them is the single threaded dependency graph. This will
now construct a single-threaded scheduler at evaluation time.
This shouldn't be a problem because it only happens when using
debugging command line arguments and the code simply don't
run in regular Blender operation.
The code seems a bit duplicated here across old and new
depsgraph, but think it's OK since the old depsgraph is already
gone in 2.8 branch and i don't see where else we might want
to use such a single-threaded scheduler.
When/if we'll want to do so, we can move it to a centralized
single-threaded scheduler in threads.c.
OpenGL render was a bit more tricky to port, but basically we
are using conditional variables to wait background thread to
do all the job.
Comments said that function was supposed to 'stop worker threads', but
it absolutely did not do anything like that, was merely wiping out TODO
queue of tasks from given pool (kind of subset of what
`BLI_task_pool_cancel()` does).
Misleading, and currently useless, we can always add it back if we need
it some day, but for now we try to simplify that area.
Freeing pool was calling `BLI_task_pool_stop()`, which only clears
pool's tasks that are in TODO queue, whithout ensuring no more tasks
from that pool are being processed in worker threads.
This could lead to use-after-free random (and seldom) crashes.
Now use instead `BLI_task_pool_cancel()`, which does waits for all tasks
being processed to finish, before returning.
Not really happy of per-pool threads limit, need to find better
approach to that. But at least it's possible to get rid of half
of the nastyness here by removing getter which was only used in
an assert statement.
That piece of code was already well-tested and this code becomes
obsolete in the new depsgraph and does no longer exists in blender
2.8 branch.
This was only used for progress report, and it's wrong because:
- Pool might in theory be re-used by different tasks
- We should not make any decision based on scheduling stats
Proper way is to take care of progress by the task itself.
We were checking for number of tasks from given pool already active, and
then atomically increasing it if allowed - this is not correct, number
could be increased by another thread between check and atomic op!
Atomic primitives are nice, but you must be very careful with *how* you
use them... Now we atomically increase counter, check result, and if we
end up over max value, abort and decrease counter again.
Spotted by Sergey, thanks!
This is a bug in the multithreaded task manager in negative value range.
The problem here is that if previter is unsigned, the comparison in the
return statement is unsigned, and works incorrectly if stop < 0 &&
iter >= 0. This in turn can happen if stop is close to 0, because this
code is designed to overrun the stop by chunk_size*num_threads as
the threads terminate.
This probably should go into 2.78 as it prevents a crash.
Together with the extended loop callback and userdata_chunk, this allows to perform
cumulative tasks (like aggregation) in a lockfree way using local userdata_chunk to store temp data,
and once all workers have finished, to merge those userdata_chunks in the finalize callback
(from calling thread, so no need to lock here either).
Note that this changes how userdata_chunk is handled (now fully from 'main' thread,
which means a given worker thread will always get the same userdata_chunk, without
being re-initialized anymore to init value at start of each iter chunk).
New code is actually much, much better than first version, using 'fetch_and_add' atomic op
here allows us to get rid of the loop etc.
The broken CAS issue remains on windows, to be investigated...
There are some serious issues under windows, causing deadlocks somehow (not reproducible under linux so far).
Until further investigation over why this happens, better to revert to previous
spin-locked behavior.
This reverts commits a83bc4f597 and 98123ae916.
Reading the shared state->iter value after storing it in the 'reference' var could in theory
lead to a race condition setting state->iter value above state->stop, which would be 'deadly'.
This **may** be the cause of T48422, though I was not able to reproduce that issue so far.
This commit makes use of new taskpool feature (instead of allocating own tasks),
and removes the spinlock used to generate chunks (using atomic ops instead).
In best cases (dynamic scheduled loop with light processing func callback), we
get a few percents of speedup, in most cases there is no sensible enhancement.
Appears mutex was guarateeing number of tasks is not modified at moments
when it's not expected. Removing those mutexes resulted in some hard-to-catch
locks where worker thread were waiting for work by all the tasks were already
done.
This reverts commit a1d8fe052c.
Brain melt here, intention was to reduce number of tasks in case we have not much chunks of data to loop over,
not to increase it!
Note that this only affected dynamic scheduling.
This commit implements new function BLI_task_pool_push_from_thread()
who's main goal is to have less parasitic load on the CPU bu avoiding
memory allocations as much as possible, making taks pushing cheaper.
This function expects thread ID, which must be 0 for the thread from
which pool is created from (and from which wait_work() is called) and
for other threads it mush be the ID which was sent to the thread working
function.
This reduces allocations quite a bit in the new dependency graph,
hopefully gaining some visible speedup on a fewzillion core machines
(on my own machine can only see benefit in profiler, which shows
significant reduce of time wasted in the memory allocation).
Based on usages so far:
- Split callback worker func in two, 'basic' and 'extended' versions. The former goes back
to the simplest verion, while the later keeps the 'userdata_chunk', and gets the thread_id too.
- Add use_threading to simple BLI_task_parallel_range(), turns out we need this pretty much systematically,
and allows to get rid of most usages of BLI_task_parallel_range_ex().
- Now BLI_task_parallel_range() expects 'basic' version of callback, while BLI_task_parallel_range_ex()
expectes 'extended' version of the callback.
All in all, this should make common usage of BLI_task_parallel_range simpler (less verbose), and add
access to advanced callback to thread id, which is mandatory in some (future) cases.
This can happen quite often in forloops, and would be annoying to have to check for this
in caller code! So now, just return without doing anything in this case.
From recent experience, turns out we often do want to use something else than basic
range of parallelized forloop as control parameter over threads usage, so now BLI func
only takes a boolean, and caller defines best check for its own case.
When called with very small range, `BLI_task_parallel_range_ex()` would generate a zero `chunk_size`,
leading to some infinite looping in `parallel_range_func` due to `parallel_range_next_iter_get` returning
true without actually increasing the counter!
So now, we ensure `chunk_size` and `num_tasks` are always at least 1 (and avoid generating too much tasks too).
This mimics OpenMP's 'firstprivate' feature. It is sometimes handy to have some persistent local data during a whole chunk.
Reviewers: sergey
Reviewed By: sergey
Subscribers: campbellbarton
Differential Revision: https://developer.blender.org/D1635
With current code, in single-threaded context, a pool of task may never be executed
until one calls BLI_task_pool_work_and_wait() on it, this is not acceptable for
asynchronous tasks where you never want to actually lock the main thread.
This commits adds an extra thread in single-threaded case, and a new 'type' of pool,
such that one can create real background pools of tasks. See code for details.
Review: D1565
