| Age | Commit message (Collapse) | Author | Files | Lines |
|---|
|
Fixes non-deterministic failures in
dEQP-EGL.functional.sharing.gles2.multithread.simple_egl_sync.images.texture_source.teximage2d_render
and others in dEQP-EGL.functional.sharing.gles2.multithread.*
Reviewed-by: Marek Olšák <marek.olsak@amd.com>
|
|
The current value was introduced in commit a27180d0d8666, which claims
that it represents ~1.11 years. However, it is interpreted in nanoseconds,
so it actually only represents ~9.8 hours. That seems a bit short.
Use the largest value consistent with both int32 and int64. It
corresponds to ~292 years in nanoseconds.
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
Reviewed-by: Marek Olšák <marek.olsak@amd.com>
|
|
TODO: It looks like si_fence_server_sync still isn't strict enough
for unflushed fences.
|
|
st_flush should flush state tracker-internal state and the pipe, but
not mesa/main state. Of the four callers:
- glFlush/glFinish already call FLUSH_{VERTICES,STATE}.
- st_vdpau doesn't need to call them.
- st_manager will now call them explicitly.
Reviewed-by: Marek Olšák <marek.olsak@amd.com>
|
|
There may be pending operations (e.g. vertices) that need to be flushed
by the state tracker.
Found by inspection.
Reviewed-by: Marek Olšák <marek.olsak@amd.com>
|
|
Reviewed-by: Marek Olšák <marek.olsak@amd.com>
|
|
Reviewed-by: Marek Olšák <marek.olsak@amd.com>
|
|
Transfer commands can have associated GPU operations.
Enabled by passing GALLIUM_DDEBUG=transfers.
Reviewed-by: Marek Olšák <marek.olsak@amd.com>
|
|
Reviewed-by: Marek Olšák <marek.olsak@amd.com>
|
|
Reviewed-by: Marek Olšák <marek.olsak@amd.com>
|
|
This patch has multiple goals:
1. Off-load the writing of records in 'always' mode to another thread
for performance.
2. Allow using ddebug with threaded contexts. This really forces us to
move some of the "after_draw" handling into another thread.
3. Simplify the different modes of ddebug, both in the code and in
the user interface, i.e. GALLIUM_DDEBUG. In particular, there's
no 'pipelined' anymore, since we're always pipelined; and 'noflush'
is replaced by 'flush', since we no longer flush by default.
4. Fix the fences in pipelining mode. They previously relied on writes
via pipe_context::clear_buffer. However, on radeonsi, those could
(quite reasonably) end up in the SDMA buffer. So we use the newly
added PIPE_FLUSH_{TOP,BOTTOM}_OF_PIPE fences instead.
5. Improve pipelined mode overall, using the finer grained information
provided by the new fences.
Overall, the result is that pipelined mode should be more useful, and
using ddebug in default mode is much less invasive, in the sense that
it changes the overall driver behavior less (which is kind of crucial
for a driver debugging tool).
An example of the new hang debug output:
Gallium debugger active.
Hang detection timeout is 1000ms.
GPU hang detected, collecting information...
Draw # driver prev BOP TOP BOP dump file
-------------------------------------------------------------
2 YES YES YES NO /home/nha/ddebug_dumps/shader_runner_19919_00000000
3 YES NO YES NO /home/nha/ddebug_dumps/shader_runner_19919_00000001
4 YES NO YES NO /home/nha/ddebug_dumps/shader_runner_19919_00000002
5 YES NO YES NO /home/nha/ddebug_dumps/shader_runner_19919_00000003
Done.
We can see that there were almost certainly 4 draws in flight when
the hang happened: the top-of-pipe fence was signaled for all 4 draws,
the bottom-of-pipe fence for none of them. In virtually all cases,
we'd expect the first draw in the list to be at fault, but due to the
GPU parallelism, it's possible (though highly unlikely) that one of
the later draws causes a component to get stuck in a way that prevents
the earlier draws from making progress as well.
(In the above example, there were actually only 3 draws truly in flight:
the last draw is a blit that waits for the earlier draws; however, its
top-of-pipe fence is emitted before the cache flush and wait, and so
the fact that the draw hasn't truly started yet can only be seen from a
closer inspection of GPU state.)
Acked-by: Marek Olšák <marek.olsak@amd.com>
|
|
Reviewed-by: Marek Olšák <marek.olsak@amd.com>
|
|
Reviewed-by: Marek Olšák <marek.olsak@amd.com>
|
|
Reviewed-by: Marek Olšák <marek.olsak@amd.com>
|
|
Reviewed-by: Marek Olšák <marek.olsak@amd.com>
|
|
Change format to %p while we're at it.
Reviewed-by: Marek Olšák <marek.olsak@amd.com>
|
|
v2: use uncached system memory for the fence, and use the CPU to
clear it so we never read garbage when checking the fence
|
|
v2: remove the change to si_fence_server_sync, we'll handle that more
robustly
Reviewed-by: Marek Olšák <marek.olsak@amd.com> (v1)
|
|
For running post-draw operations inside the driver thread. ddebug will
use it.
Reviewed-by: Marek Olšák <marek.olsak@amd.com>
|
|
Reviewed-by: Marek Olšák <marek.olsak@amd.com>
|
|
Queries should still get marked as flushed when flushes are executed
asynchronously in the driver thread.
To this end, the management of the unflushed_queries list is moved into
the driver thread.
Reviewed-by: Marek Olšák <marek.olsak@amd.com>
|
|
This requires out-of-band creation of fences, and will be signaled to
the pipe_context::flush implementation by a special TC_FLUSH_ASYNC flag.
v2:
- remove an incorrect assertion
- handle fence_server_sync for unsubmitted fences by
relying on the improved cs_add_fence_dependency
- only implement asynchronous flushes on amdgpu
|
|
The driver uses (and must use) the flushed flag of queries as a hint that
it does not have to check for synchronization with currently queued up
commands. Deferred flushes do not actually flush queued up commands, so
we must not set the flushed flag for them.
Found by inspection.
Reviewed-by: Marek Olšák <marek.olsak@amd.com>
|
|
Reviewed-by: Marek Olšák <marek.olsak@amd.com>
|
|
The idea is to fix the following interleaving of operations
that can arise from deferred fences:
Thread 1 / Context 1 Thread 2 / Context 2
-------------------- --------------------
f = deferred flush
<------- application-side synchronization ------->
fence_server_sync(f)
...
flush()
flush()
We will now stall in fence_server_sync until the flush of context 1
has completed.
This scenario was unlikely to occur previously, because applications
seem to be doing
Thread 1 / Context 1 Thread 2 / Context 2
-------------------- --------------------
f = glFenceSync()
glFlush()
<------- application-side synchronization ------->
glWaitSync(f)
... and indeed they probably *have* to use this ordering to avoid
deadlocks in the GLX model, where all GL operations conceptually
go through a single connection to the X server. However, it's less
clear whether applications have to do this with other WSI (i.e. EGL).
Besides, even this sequence of GL commands can be translated into
the Gallium-level sequence outlined above when Gallium threading
and asynchronous flushes are used. So it makes sense to be more
robust.
As a side effect, we no longer busy-wait on submission_in_progress.
We won't enable asynchronous flushes on radeon, but add a
cs_add_fence_dependency stub anyway to document the potential
issue.
|
|
These bits are intended to be used by the ddebug hang detection and are
named in analogy to the Vulkan stage bits (and the corresponding Radeon
pipeline event).
Hang detection needs fences on the granularity of individual commands,
which nothing else really covers. The closest alternative would have
been PIPE_QUERY_GPU_FINISHED, but (a) queries are a per-context object
and we really want a per-screen object, (b) queries don't offer a
wait with timeout, and (c) in any case, PIPE_QUERY_GPU_FINISHED is
meant to imply that GPU caches are flushed, which the new bits
explicitly aren't.
Reviewed-by: Marek Olšák <marek.olsak@amd.com>
|
|
Also document some subtleties of pipe_context::flush.
Reviewed-by: Marek Olšák <marek.olsak@amd.com>
|
|
v2:
- style fixes
- fix missing timeout handling in futex path
Reviewed-by: Marek Olšák <marek.olsak@amd.com> (v1)
|
|
C11 threads were changed to use struct timespec instead of xtime, and
thrd_sleep got a second argument.
See http://www.open-std.org/jtc1/sc22/wg14/www/docs/n1554.htm and
http://en.cppreference.com/w/c/thread/{thrd_sleep,cnd_timedwait,mtx_timedlock}
Note that cnd_timedwait is spec'd to be relative to TIME_UTC / CLOCK_REALTIME.
v2: Fix Windows build errors. Tested with a default Appveyor config
that uses Visual Studio 2013. Judging from Brian's email and
random internet sources, Visual Studio 2015 does have timespec
and timespec_get, hence the _MSC_VER-based guard which I have
not tested.
Cc: Jose Fonseca <jfonseca@vmware.com>
Cc: Brian Paul <brianp@vmware.com>
Reviewed-by: Marek Olšák <marek.olsak@amd.com> (v1)
|
|
Cc: Jose Fonseca <jfonseca@vmware.com>
Reviewed-by: Marek Olšák <marek.olsak@amd.com>
|
|
Reviewed-by: Marek Olšák <marek.olsak@amd.com>
|
|
With Gallium threaded contexts, creating shader/compute states is
effectively a screen operation, so we should not use context state.
In particular, this allows us to avoid using the context's LLVM
TargetMachine.
This isn't an issue yet because u_threaded_context filters out non-async
debug callbacks, and we disable threaded contexts for debug contexts.
However, we may want to change that in the future.
Reviewed-by: Marek Olšák <marek.olsak@amd.com>
|
|
Found by inspection.
Reviewed-by: Marek Olšák <marek.olsak@amd.com>
|
|
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)
|
|
The #if guard is probably not 100% equivalent to the previous PIPE_OS
check, but if anything it should be an over-approximation (are there
pthread implementations without barriers?), so people will get either
a good implementation or compile errors that are easy to fix.
Reviewed-by: Marek Olšák <marek.olsak@amd.com>
|
|
v2: use util_vasprintf for Windows portability
Reviewed-by: Marek Olšák <marek.olsak@amd.com> (v1)
|
|
Some locking is unfortunately required, because well-formed GL programs
can have multiple threads racing to access the same texture, e.g.: two
threads/contexts rendering from the same texture, or one thread destroying
a context while the other is rendering from or modifying a texture.
Since even the simple mutex caused noticable slowdowns in the piglit
drawoverhead micro-benchmark, this patch uses a slightly more involved
approach to keep locks out of the fast path:
- the initial lookup of sampler views happens without taking a lock
- a per-texture lock is only taken when we have to modify the sampler
view(s)
- since each thread mostly operates only on the entry corresponding to
its context, the main issue is re-allocation of the sampler view array
when it needs to be grown, but the old copy is not freed
Old copies of the sampler views array are kept around in a linked list
until the entire texture object is deleted. The total memory wasted
in this way is roughly equal to the size of the current sampler views
array.
Fixes non-deterministic memory corruption in some
dEQP-EGL.functional.sharing.gles2.multithread.* tests, e.g.
dEQP-EGL.functional.sharing.gles2.multithread.simple.images.texture_source.create_texture_render
Reviewed-by: Marek Olšák <marek.olsak@amd.com>
|
|
Move the early-out for surface-based textures earlier. This narrows the
scope of the locking added in a follow-up commit.
Fix one remaining case of initializing a surface-based texture
without properly finalizing it.
Reviewed-by: Marek Olšák <marek.olsak@amd.com>
|
|
r600 expects the context that created the sampler view to still be alive
(there is a per-context list of sampler views).
svga currently bails when the context of destruction is not the same as
creation.
The GL state tracker, which is the only one that runs into the
multi-context subtleties (due to share groups), already guarantees that
sampler views are destroyed before their context of creation is destroyed.
Most drivers are context-agnostic, so the warning message in
pipe_sampler_view_release doesn't really make sense.
Reviewed-by: Marek Olšák <marek.olsak@amd.com>
|
|
We only need the lock to guard changes in the variant linked list. The
actual compilation can happen outside the lock, since we use the ready
fence as a guard.
v2: fix double-unlock
Reviewed-by: Marek Olšák <marek.olsak@amd.com>
|
|
There's a race condition between si_shader_select_with_key and
si_bind_XX_shader:
Thread 1 Thread 2
-------- --------
si_shader_select_with_key
begin compiling the first
variant
(guarded by sel->mutex)
si_bind_XX_shader
select first_variant by default
as state->current
si_shader_select_with_key
match state->current and early-out
Since thread 2 never takes sel->mutex, it may go on rendering without a
PM4 for that shader, for example.
The solution taken by this patch is to broaden the scope of
shader->optimized_ready to a fence shader->ready that applies to
all shaders. This does not hurt the fast path (if anything it makes
it faster, because we don't explicitly check is_optimized).
It will also allow reducing the scope of sel->mutex locks, but this is
deferred to a later commit for better bisectability.
Fixes dEQP-EGL.functional.sharing.gles2.multithread.simple.buffers.bufferdata_render
Reviewed-by: Marek Olšák <marek.olsak@amd.com>
|
|
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
|
|
Reviewed-by: Marek Olšák <marek.olsak@amd.com>
|
|
Reviewed-by: Marek Olšák <marek.olsak@amd.com>
|
|
Reviewed-by: Marek Olšák <marek.olsak@amd.com>
|
|
The closest to it in the old-style gcc builtins is __sync_lock_test_and_set,
however, that is only guaranteed to work with values 0 and 1 and only
provides an acquire barrier. I also don't know about other OSes, so we
provide a simple & stupid emulation via p_atomic_cmpxchg.
|
|
v2: style fixes
Reviewed-by: Marek Olšák <marek.olsak@amd.com> (v1)
|
|
|
|
While modern pthread mutexes are very fast, they still incur a call to an
external DSO and overhead of the generality and features of pthread mutexes.
Most mutexes in mesa only needs lock/unlock, and the idea here is that we can
inline the atomic operation and make the fast case just two intructions.
Mutexes are subtle and finicky to implement, so we carefully copy the
implementation from Ulrich Dreppers well-written and well-reviewed paper:
"Futexes Are Tricky"
http://www.akkadia.org/drepper/futex.pdf
We implement "mutex3", which gives us a mutex that has no syscalls on
uncontended lock or unlock. Further, the uncontended case boils down to a
cmpxchg and an untaken branch and the uncontended unlock is just a locked decr
and an untaken branch. We use __builtin_expect() to indicate that contention
is unlikely so that gcc will put the contention code out of the main code
flow.
A fast mutex only supports lock/unlock, can't be recursive or used with
condition variables. We keep the pthread mutex implementation around as
full_mtx_t for the few places where we use condition variables or recursive
locking. For platforms or compilers where futex and atomics aren't available,
mtx_t falls back to the pthread mutex.
The pthread mutex lock/unlock overhead shows up on benchmarks for CPU bound
applications. Most CPU bound cases are helped and some of our internal
bind_buffer_object heavy benchmarks gain up to 10%.
Signed-off-by: Kristian Høgsberg <krh@bitplanet.net>
Signed-off-by: Timothy Arceri <tarceri@itsqueeze.com>
|
