When TSAN is enabled we use standard mutexes instead of futexes. With
futexes the fence->signalled is read using an atomic operation, to best
mimic this let's protect the read with a locked mutex.
This avoids TSAN reporting a race condition (false positive with
futexes) with Zink when accessing the pipeline cache.
Signed-off-by: Gert Wollny <gert.wollny@collabora.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/28650>
This flag allow to create a single thread initially, but set
max_thread to the request thread count.
If the queue is full and num_threads is lower than max_threads,
we spawn a new thread to help process the queue faster.
This avoid creating N threads at queue creation time.
Reviewed-by: Marek Olšák <marek.olsak@amd.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/11296>
When both UTIL_QUEUE_INIT_RESIZE_IF_FULL and
UTIL_QUEUE_INIT_USE_MINIMUM_PRIORITY are set, we can get into a
situation where the queue never executes and grows to a huge size
due to all other threads being busy.
This is the case with the shader cache when attempting to compile a
huge number of shaders up front. If all threads are busy compiling
shaders the cache queues memory use can climb into the many GBs
very fast.
The use of these two flags with the shader cache is intended to
allow shaders compiled at runtime to be compiled as fast as possible.
To avoid huge memory use but still allow the queue to perform
optimally in the run time compilation case, we now add the ability
to track memory consumed by the jobs in the queue and limit it to
a hardcoded 256MB which should be more than enough.
Reviewed-by: Marek Olšák <marek.olsak@amd.com>
Schedule one job for every thread, and wait on a barrier inside the job
execution function.
v2: avoid alloca (fixes Windows build error)
Reviewed-by: Marek Olšák <marek.olsak@amd.com> (v1)
Fences are now 4 bytes instead of 96 bytes (on my 64-bit system).
Signaling a fence is a single atomic operation in the fast case plus a
syscall in the slow case.
Testing if a fence is signaled is the same as before (a simple comparison),
but waiting on a fence is now no more expensive than just testing it in
the fast (already signaled) case.
v2:
- style fixes
- use p_atomic_xxx macros with the right barriers
Acked-by: Marek Olšák <marek.olsak@amd.com>
Consider the following situation:
mtx_lock(mutex);
do_something();
util_queue_add_job(...);
mtx_unlock(mutex);
If the queue is full, util_queue_add_job will wait for a free slot.
If the job which is currently being executed tries to lock the mutex,
it will be stuck forever, because util_queue_add_job is stuck.
The deadlock can be trivially resolved by increasing the queue size
(reallocating the queue) in util_queue_add_job if the queue is full.
Then util_queue_add_job becomes wait-free.
radeonsi will use it.
Reviewed-by: Nicolai Hähnle <nicolai.haehnle@amd.com>
for HUD integration in following commits. This valuable profiling data
will allow us to see on the HUD how well glthread is able to utilize
parallelism. This is better than benchmarking, because you can see
exactly what's happening and you don't have to be CPU-bound.
u_threaded_context has the same counters.
Reviewed-by: Timothy Arceri <tarceri@itsqueeze.com>
Sil Vilerino