Instead of computing an index at the end which we hope maps to the
number of things written, just count the number of things as we go.
Reviewed-by: Lionel Landwerlin <lionel.g.landwerlin@intel.com>
and _mesa_bitcount_64 with util_bitcount_64. This fixes a build problem
in nir for platforms that don't have popcount or popcountll, such as
32bit msvc.
v2: - Fix additional uses of _mesa_bitcount added after this was
originally written
Acked-by: Eric Engestrom <eric.engestrom@intel.com> (v1)
Acked-by: Eric Anholt <eric@anholt.net>
Reviewed-by: Ian Romanick <ian.d.romanick@intel.com>
From the Vulkan spec with KHX extensions:
"If queries are used while executing a render pass instance that has
multiview enabled, the query uses N consecutive query indices
in the query pool (starting at query) where N is the number of bits
set in the view mask in the subpass the query is used in.
How the numerical results of the query are distributed among the
queries is implementation-dependent. For example, some implementations
may write each view's results to a distinct query, while other
implementations may write the total result to the first query and write
zero to the other queries. However, the sum of the results in all the
queries must accurately reflect the total result of the query summed
over all views. Applications can sum the results from all the queries to
compute the total result."
In our case we only really emit a single query (in the first query index)
that stores the aggregated result for all views, but we still need to manage
availability for all the other query indices involved, even if we don't
actually use them.
This is relevant when clients call vkGetQueryPoolResults and pass all N
queries to retrieve the results. In that scenario, without this patch,
we will never see queries other than the first being available since we
never emit them.
v2: we need the same treatment for timestamp queries.
v3 (Jason):
- Better an if instead of an early return.
- We can't write to this memory in the CPU, we should use
MI_STORE_DATA_IMM and emit_query_availability (Jason).
v4 (Jason):
- No need to take the value to write as parameter, just hard code it to 0.
Fixes test failures in some work-in-progress CTS multiview+query tests.
Reviewed-by: Jason Ekstrand <jason@jlekstrand.net>
Unless you have data, the compiler knows better than you whether a
function should be inlined.
No difference in the resulting binary with gcc-6.3.0 or clang-4.0.
Reviewed-by: Emil Velikov <emil.velikov@collabora.com>
Reviewed-by: Eric Engestrom <eric.engestrom@imgtec.com>
The idea behind doing this was to make it easier to set various flags.
However, we have enough custom flag settings floating around the driver
that this is more of a nuisance than a help. This commit has the
following functional changes:
1) The workaround_bo created in anv_CreateDevice loses both flags.
This shouldn't matter because it's very small and entirely internal
to the driver.
2) The bo created in anv_CreateDmaBufImageINTEL loses the
EXEC_OBJECT_ASYNC flag. In retrospect, it never should have gotten
EXEC_OBJECT_ASYNC in the first place.
Reviewed-by: Nanley Chery <nanley.g.chery@intel.com>
Cc: "17.1" <mesa-stable@lists.freedesktop.org>
Commit b2c97bc789 which made us start
using a busy-wait for individual query results also messed up cache
flushing on !LLC platforms. For one thing, I forgot the mfence after
the clflush so memory access wasn't properly getting fenced. More
importantly, however, was that we were clflushing the whole query range
and then waiting for individual queries and then trying to read the
results without clflushing again. Getting the clflushing both correct
and efficient is very subtle and painful. Instead, let's side-step the
problem by just snooping.
Reviewed-by: Chris Wilson <chris@chris-wilson.co.uk>
Before, we were just looking at whether or not the user wanted us to
wait and waiting on the BO. Some clients, such as the Serious engine,
use a single query pool for hundreds of individual query results where
the writes for those queries may be split across several command
buffers. In this scenario, the individual query we're looking for may
become available long before the BO is idle so waiting on the query pool
BO to be finished is wasteful. This commit makes us instead busy-loop on
each query until it's available.
This significantly reduces pipeline bubbles and improves performance of
The Talos Principle on medium settings (where the GPU isn't overloaded
with drawing) by around 20% on my SkyLake gt4.
Reviewed-by: Chris Wilson <chris@chris-wilson.co.uk>
Tested-by: Eero Tamminen <eero.t.tamminen@intel.com>
Tested-by: Grazvydas Ignotas <notasas@gmail.com>
When a client causes a GPU hang (or experiences issues due to a hang in
another client) we want to let it know as soon as possible. In
particular, if it submits work with a fence and calls vkWaitForFences or
vkQueueQaitIdle and it returns VK_SUCCESS, then the client should be
able to trust the results of that rendering. In order to provide this
guarantee, we have to ask the kernel for context status in a few key
locations.
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
We don't need to make the caller (CmdCopyQueryPoolResults) aware of the
problem since compute_query_result() only emits state. The caller is also
expected to hit OOM in this scenario right after calling this function, but
it is already handling it safely.
Fixes:
dEQP-VK.api.out_of_host_memory.cmd_copy_query_pool_results
Reviewed-by: Topi Pohjolainen <topi.pohjolainen@intel.com>
If we know the device has been lost we should return this error code for
any command that can report it before we attempt to do anything with the
device.
Reviewed-by: Jason Ekstrand <jason@jlekstrand.net>
In the end, pipeline statistics queries look a lot like occlusion
queries only with between 1 and 11 begin/end pairs being generated
instead of just the one.
Reviewed-By: Lionel Landwerlin <lionel.g.landwerlin@intel.com>
The new version is a nice GPU parallel to cpu_write_query_result and it
nicely handles things like dealing with 32 vs. 64-bit offsets in the
destination buffer.
Reviewed-By: Lionel Landwerlin <lionel.g.landwerlin@intel.com>
Not all queries are the same. Even the two queries we support today
require a different amount of data per slot. Once we introduce pipeline
statistics queries, the size will vary wildly.
Reviewed-By: Lionel Landwerlin <lionel.g.landwerlin@intel.com>
We're about to make slots variable-length and always having the
available bits at the front makes certain operations substantially
easier once we do that.
Reviewed-By: Lionel Landwerlin <lionel.g.landwerlin@intel.com>
From the Vulkan 1.0.39 Specification:
"If VK_QUERY_RESULT_64_BIT is not set and the result overflows a
32-bit value, the value may either wrap or saturate."
So we can either clamp or wrap. Wrapping is both easier and what the
user gets if they use vkCmdCopyQueryPoolResults and we should be
consistent. We could make vkCmdCopyQueryPoolResults clamp but it's
annoying and ends up burning extra batch for something the spec clearly
doesn't require.
Reviewed-By: Lionel Landwerlin <lionel.g.landwerlin@intel.com>
The Vulkan spec is fairly clear about when we should and should not
write query pool results. We're also supposed to return VK_NOT_READY if
VK_QUERY_RESULT_PARTIAL_BIT is not set and we come across any queries
which are not yet finished. This fixes rendering corruptions on The
Talos Principle where geometry flickers in and out due to bogus query
results being returned by the driver. These issues are most noticable
on Sky Lake GT4 2hen running on "ultra" settings.
Reviewed-By: Lionel Landwerlin <lionel.g.landwerlin@intel.com>
Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=100182
Cc: "17.0 13.0" <mesa-stable@lists.freedesktop.org>
