It's unnecessary because we can just check if the timestamp
is to different to the default value when a pool is created
or resetted. Instead of waiting for the availability bit to
be 1, we have to emit a not equal WAIT_REG_MEM for checking
if the timestamp is ready.
Signed-off-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>
Reviewed-by: Dave Airlie <airlied@redhat.com>
In environments where we cannot cache, e.g. Android (no homedir),
ChromeOS (readonly rootfs) or sandboxes (cannot open cache), the
startup cost of creating a device in radv is rather high, due
to compiling all possible built-in pipelines up front. This meant
depending on the CPU a 1-4 sec cost of creating a Device.
For CTS this cost is unacceptable, and likely for starting random
apps too.
So if there is no cache, with this patch radv will compile shaders
on demand. Once there is a cache from the first run, even if
incomplete, the driver knows that it can likely write the cache
and precompiles everything.
Note that I did not switch the buffer and itob/btoi compute pipelines
to on-demand, since you cannot really do anything in Vulkan without
them and there are only a few.
This reduces the CTS runtime for the no caches scenario on my
threadripper from 32 minutes to 8 minutes.
Reviewed-by: Dave Airlie <airlied@redhat.com>
There are no fixed sized array arguments in C, those are simply pointers
to unsized arrays and as the size is passed in anyway, just rely on that.
where possible calls are replaced by nir_channel and nir_channels.
Reviewed-by: Jason Ekstrand <jason@jlekstrand.net>
Signed-off-by: Karol Herbst <kherbst@redhat.com>
We mostly use the same priority for all buffer objects, so
I don't think that matter much. This should reduce CPU
overhead a little bit.
Signed-off-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>
Reviewed-by: Bas Nieuwenhuizen <bas@basnieuwenhuizen.nl>
A ZPASS_DONE or PIXEL_STAT_DUMP_EVENT (of the DB occlusion
counters) must immediately precede every timestamp event to
prevent a GPU hang on GFX9.
Cc: 18.1 <mesa-stable@lists.freedesktop.org>
Signed-off-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>
Reviewed-by: Bas Nieuwenhuizen <bas@basnieuwenhuizen.nl>
Ported from RadeonSI.
This appears to fix some random fails with:
dEQP-VK.query_pool.statistics_query.*
Signed-off-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>
Reviewed-by: Bas Nieuwenhuizen <bas@basnieuwenhuizen.nl>
Errors are not that common of a case so we can eat a slight perf
hit in having to call a function and do a runtime check.
In turn this makes debugging random errors happening for end users
easier, because they don't have to have a debug build on hand.
Reviewed-by: Timothy Arceri <tarceri@itsqueeze.com>
Reviewed-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>
This moves the extra queries to after the main query ended, instead
of doing it after the begin and hence doing nesting.
We also emit only (view count - 1) extra queries, as the main query
is already there for the first view.
This fixes the CTS occasionally getting stuck in
dEQP-VK.multiview.queries* waiting on results.
Fixes: 32b4f3c38d "radv/query: handle multiview queries properly. (v3)"
CC: 18.1 <mesa-stable@lists.freedesktop.org>
Reviewed-by: Dave Airlie <airlied@redhat.com>
Reviewed-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>
This unnecessary when the precision bit flag is not set, and this
might hurt performance. The Vulkan explains that not setting
VK_QUERY_CONTROL_PRECISE_BIT might be more efficient on some
implementations.
Signed-off-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>
Reviewed-by: Bas Nieuwenhuizen <bas@basnieuwenhuizen.nl>
For multiview we need to emit a number of sequential queries
depending on the view mask.
This avoids dEQP-VK.multiview.queries.15 waiting forever
on the CPU for query results that are never coming.
We only really want to emit one query,
and the rest should be blank (amdvlk does the same),
so we emit begin/end pairs for all the others except
the first query.
v2: fix tests
v3: split out patch.
Fixes: dEQP-VK.multiview.queries*
Reviewed-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>
This just splits out the begin/end query hw emissions,
it makes it easier to add multiview support for queries.
Reviewed-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>
This is an optimization which reduces the number of flushes for
small pool buffers.
Signed-off-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>
Reviewed-by: Bas Nieuwenhuizen <bas@basnieuwenhuizen.nl>
If the query pool has been previously resetted using the compute
shader path.
Fixes: a41e2e9cf5 ("radv: allow to use a compute shader for resetting the query pool")
Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=105292
Cc: "18.0" <mesa-stable@lists.freedesktop.org>
Signed-off-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>
Reviewed-by: Bas Nieuwenhuizen <bas@basnieuwenhuizen.nl>
It's legal to a pipeline stat query on a compute queue,
but we'd emit the wrong packet here. This should fix it to emit
the correct packet.
Noticed while inspecting the mpv hang.
Fixes: ad61eac250 (radv: factor out eop event writing code. (v2))
Reviewed-by: Bas Nieuwenhuizen <bas@basnieuwenhuizen.nl>
Signed-off-by: Dave Airlie <airlied@redhat.com>
As per the spec, the query identified by queryPool and query
must currently be active. Applications have to call vkCmdBeginQuery()
before, and thus the query pool BO will already be in the list.
Signed-off-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>
Reviewed-by: Dave Airlie <airlied@redhat.com>
For consistency and it might help for debugging purposes.
Signed-off-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>
Reviewed-by: Bas Nieuwenhuizen <bas@basnieuwenhuizen.nl>
The next patch will try and avoid calling the indirect function.
v2: add a missing conversion.
Reviewed-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>
Signed-off-by: Dave Airlie <airlied@redhat.com>
Serious Sam Fusion 2017 uses a huge number of occlusion queries,
and the allocated query pool buffer is greater than 4096 bytes.
This slightly improves performance (tested in Ultra) from
117.2 FPS to 119.7 FPS (~+2%) on my RX480.
This also improves Talos, from 69 FPS to 72/73 FPS (~+5%).
Signed-off-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>
Reviewed-by: Bas Nieuwenhuizen <bas@basnieuwenhuizen.nl>
This uses the new kernel interfaces for reduced cs overhead,
We only set the local flag for memory allocations that don't have
a dedicated allocation and ones that aren't imports.
v2: add to all the internal buffer creation paths.
v3: missed some command submission paths, handle 0/empty bo lists.
Reviewed-by: Bas Nieuwenhuizen <bas@basnieuwenhuizen.nl>
Signed-off-by: Dave Airlie <airlied@redhat.com>
We are pushing 16-bytes of constants, so we have to save/restore
the same amount of data to avoid data corruption.
Cc: 17.2 <mesa-stable@lists.freedesktop.org>
Signed-off-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>
Reviewed-by: Dave Airlie <airlied@redhat.com>
We are really not going to use a winsys which does not need to store
the va, so might as well store it in a standard field.
Not sure this helps perf much though, as most of the cost is in the
cache miss accessing the bo anyway, which we stil need to do.
Reviewed-by: Dave Airlie <airlied@redhat.com>
Reviewed-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>
This doesn't get used yet, it just adds support to various PKT3
emissions to enable it later.
Reviewed-by: Bas Nieuwenhuizen <bas@basnieuwenhuizen.nl>
Signed-off-by: Dave Airlie <airlied@redhat.com>
In prep for GFX9 refactor some of the eop event writing code
out.
This changes behaviour, but aligns with what radeonsi does,
it does double emits on CIK/VI, whereas previously it only
did this on CIK.
v2: bump the size checks.
Reviewed-by: Bas Nieuwenhuizen <bas@basnieuwenhuizen.nl>
Signed-off-by: Dave Airlie <airlied@redhat.com>
This code was in a few places, consolidate into one.
Reviewed-by: Bas Nieuwenhuizen <bas@basnieuwenhuizen.nl>
Signed-off-by: Dave Airlie <airlied@redhat.com>
valgrind reports them as leaked, and I could not find anything making a
copy of the nir pointer. Also, radv_device_init_meta_blit_color() is
already freeing them unconditionally like this.
Signed-off-by: Grazvydas Ignotas <notasas@gmail.com>
Reviewed-by: Bas Nieuwenhuizen <bas@basnieuwenhuizen.nl>
Commit e1af20f18a changed the shader_info
from being embedded into being just a pointer. The idea was that
sharing the shader_info between NIR and GLSL would be easier if it were
a pointer pointing to the same shader_info struct. This, however, has
caused a few problems:
1) There are many things which generate NIR without GLSL. This means
we have to support both NIR shaders which come from GLSL and ones
that don't and need to have an info elsewhere.
2) The solution to (1) raises all sorts of ownership issues which have
to be resolved with ralloc_parent checks.
3) Ever since 00620782c9, we've been
using nir_gather_info to fill out the final shader_info. Thanks to
cloning and the above ownership issues, the nir_shader::info may not
point back to the gl_shader anymore and so we have to do a copy of
the shader_info from NIR back to GLSL anyway.
All of these issues go away if we just embed the shader_info in the
nir_shader. There's a little downside of having to copy it back after
calling nir_gather_info but, as explained above, we have to do that
anyway.
Acked-by: Timothy Arceri <tarceri@itsqueeze.com>
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
This isn't necessary yet but I'd like to use the range in
some future patches.
[airlied: add new resolve pass]
Reviewed-by: Bas Nieuwenhuizen <bas@basnieuwenhuizen.nl>
Signed-off-by: Dave Airlie <airlied@redhat.com>
Set the bit in the same stage as the timestamp, instead always at top of pipe.
Signed-off-by: Bas Nieuwenhuizen <basni@google.com>
Reviewed-by: Dave Airlie <airlied@redhat.com>
Tested-by: Grazvydas Ignotas <notasas@gmail.com>
Use push descriptors instead of temp descriptor sets.
Signed-off-by: Fredrik Höglund <fredrik@kde.org>
Reviewed-by: Bas Nieuwenhuizen <bas@basnieuwenhuizen.nl>
The devil is in the shader again, otherwise this is
fairly straightforward.
The CTS contains no pipeline statistics copy to buffer
testcases, so I did a basic smoketest.
Signed-off-by: Bas Nieuwenhuizen <basni@google.com>
Reviewed-by: Dave Airlie <airlied@redhat.com>
For using them with both occlusion and pipeline statistics queries.
Signed-off-by: Bas Nieuwenhuizen <basni@google.com>
Reviewed-by: Dave Airlie <airlied@redhat.com>
The buffer sizes are specified just a few lines earlier, so don't
repeat ourselves.
Signed-off-by: Bas Nieuwenhuizen <basni@google.com>
Reviewed-by: Dave Airlie <airlied@redhat.com>
Use the new occlusion query copy shader.
We don't use the shader for the waiting as a polling loop ineracts badly
with having caching enabled. I noticed on my GPU (Tonga) that the values
are written out in order, so I just use a WAIT_REG_MEM on the last value.
If it turns out other chips don't do that we may need to look a bit more
into this. Having 8 WAIT_REG_MEM packets per query doesn't sound ideal.
This also restricts the availability word in the pool to timestamp queries
only, as occlusion queries don't use it, and pipeline statistic queries
likely won't either.
Signed-off-by: Bas Nieuwenhuizen <basni@google.com>
Reviewed-by: Dave Airlie <airlied@redhat.com>