According to RadeonSI, this seems to be required by the hardware
to avoid GPU hangs. I think I just forgot to set that bit when I
implemented VK_EXT_transform_feedback.
This fixes a GPU hang with Space Engineers and DXVK.
Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=110291
Fixes: b4eb029062 ("radv: implement VK_EXT_transform_feedback")
Signed-off-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>
Reviewed-by: Bas Nieuwenhuizen <bas@basnieuwenhuizen.nl>
This parameter is actually useless as the immediate value
can always be zero.
Signed-off-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>
Reviewed-by: Bas Nieuwenhuizen <bas@basnieuwenhuizen.nl>
If the driver used a compute shader for resetting a query pool,
it should be completed when caches are flushed.
This might reduce the number of stalls if operations are done
between vkCmdResetQueryPool() and vkCmdBeginQuery()
(or vkCmdWriteTimestamp()).
Signed-off-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>
Reviewed-by: Bas Nieuwenhuizen <bas@basnieuwenhuizen.nl>
Reviewed-by: Alex Smith <asmith@feralinteractive.com>
This implementation should work and potential bugs can be
fixed during the release candidates window anyway.
Signed-off-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>
Reviewed-by: Dave Airlie <airlied@redhat.com>
We already track if the DMA engine is busy/idle with a flag,
and we emit a packet that waits for all CP DMA operations
to be complete. This is done at end of command buffer because
the kernel doesn't wait for them, and also when emitting
barriers, so it should be safe.
This improves small copies for both aligned and unaligned sizes.
Aligned sizes:
BEFORE:
1 KB: 59.840000 ms
2 KB: 71.200000 ms
AFTER:
1 KB: 31.200000 ms
2 KB: 31.040000 ms
Unaligned sizes:
BEFORE:
2 KB: 68.3200 ms
3 KB: 79.3600 ms
5 KB: 76.6400 ms
9 KB: 90.8800 ms
17 KB: 116.0000 ms
AFTER:
2 KB: 31.0400 ms
3 KB: 32.0000 ms
5 KB: 30.8800 ms
9 KB: 30.5600 ms
17 KB: 29.6000 ms
Signed-off-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>
Reviewed-by: Bas Nieuwenhuizen <bas@basnieuwenhuizen.nl>
A ZPASS_DONE packet doesn't make sense for the compute queue. It will
result in a gpu hang.
This change resolves a gpu hang for SteamVR+Vega.
Cc: mesa-stable@lists.freedesktop.org
Fixes: 1f616a840e "radv: emit a dummy ..."
Signed-off-by: Andres Rodriguez <andresx7@gmail.com>
Reviewed-by: Dave Airlie <airlied@redhat.com>
It shouldn't be needed to emit the initial graphics or compute
state when beginning a new command buffer. Emitting them in
the preamble should be enough and this will reduce IB sizes.
Signed-off-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>
Reviewed-by: Bas Nieuwenhuizen <bas@basnieuwenhuizen.nl>
The last parameter of radeon_set_sh_reg_seq() is the number of
dwords to emit. We were lucky because WAVES_PER_SH(0x3) is 3 but
it was initialized to 0.
COMPUTE_RESOURCE_LIMITS is correctly set when generating
compute pipelines, so we don't need to initialize it.
Signed-off-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>
Reviewed-by: Bas Nieuwenhuizen <bas@basnieuwenhuizen.nl>
By default, our internal rendering commands are discarded
only if the predicate is non-zero (ie. DRAW_VISIBLE). But
VK_EXT_conditional_rendering also allows to discard commands
when the predicate is zero, which means we have to use a
different flag.
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>
Ported from RadeonSI.
Not sure why this is needed but AMDVLK does something similar.
Signed-off-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>
Reviewed-by: Bas Nieuwenhuizen <bas@basnieuwenhuizen.nl>
AMDVLK also always uses CACHE_FLUSH_AND_INV_TS_EVENT. The other
workaround is to flush DB metadata after emitting the framebuffer,
but that seems slower.
Signed-off-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>
Reviewed-by: Bas Nieuwenhuizen <bas@basnieuwenhuizen.nl>
This follows what radeonsi does.
Ported from radeonsi:
radeonsi: emit PA_SC_RASTER_CONFIG_1 only once
Reviewed-by: Bas Nieuwenhuizen <bas@basnieuwenhuizen.nl>
Reviewed-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>
Enable it directly in the preamble, but do not enable line
on Polaris10/11/12 because there is a hw bug.
There is possibly an issue when MSAA is off, but this doesn't
regress any CTS and AMDVLK doesn't have a workaround as well.
Signed-off-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>
Reviewed-by: Bas Nieuwenhuizen <bas@basnieuwenhuizen.nl>
This can lead to a situation where cache flushes could get conditionally
disabled while still clearing the flush_bits, and thus flushes due to
application pipeline barriers may never get executed.
Fixes: a6c2001ace (radv: add support for cmd predication.)
Signed-off-by: Dave Airlie <airlied@redhat.com>
Also moved everything in a struct and then return the struct from
the helper function, so it is clear in the caller what part of the
pipeline gets modified.
Reviewed-by: Dave Airlie <airlied@redhat.com>
Reviewed-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>
Copied from radeonsi.
Putting in the correct metadata flush commands for eventually not
flushing L2 on CB/DB switch.
Does not remove the need for V_028A90_CACHE_FLUSH_AND_INV_TS_EVENT
at the moment.
Reviewed-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>
When rasterization is disabled we can have that few.
Fixes: 76603aa90b "radv: Drop the default viewport when 0 viewports are given."
Reviewed-by: Dave Airlie <airlied@redhat.com>
Overall it does not really help or hurt. The deferred demo gets 1%
improvement and some games a 3% decrease, so I don't think this
should be enabled by default.
But with the code upstream it is easier to experiment with it.
v2: Remove initializing the registers from si_emit_config.
Reviewed-by: Dave Airlie <airlied@redhat.com>
amdvlk is probably more subtle than this but it never uses
the inv cb/db variants, we fail some CTS tests without this.
Fixes:
dEQP-VK.renderpass.dedicated_allocation.formats.d32_sfloat_s8_uint.input*.
Fixes: c2fbeb7ca0 (radv: add GFX9 cache flushing support.)
Reviewed-by: Bas Nieuwenhuizen <bas@basnieuwenhuizen.nl> (for now :-)
Signed-off-by: Dave Airlie <airlied@redhat.com>
Fix a bunch of labels indicating when registers were added/removed
and normalize the SI-class GRBM_GFX_INDEX.
Reviewed-by: Marek Olšák <marek.olsak@amd.com>
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>
This patch helps lower high priority compute latency. Found by
bisecting a perf regression on computeparticles with high priority
compute queues enabled.
Reverting this micro-optimization doesn't seem to have any negative
effect on performance on Dota2 or ssao.
Signed-off-by: Andres Rodriguez <andresx7@gmail.com>
Reviewed-by: Bas Nieuwenhuizen <bas@basnieuwenhuizen.nl>
When WAVE_LIMIT is set, a submission will opt-in for SPI based resource
scheduling. Because this mechanism is cooperative, we must ensure that
all submissions have this field set, otherwise they will bypass resource
arbitration.
We always hardcode the field to its maximum value, instead of attempting
to calculate an approximate usage. In testing, there were no benefits to
using anything other than the maximum.
Signed-off-by: Andres Rodriguez <andresx7@gmail.com>
Reviewed-by: Bas Nieuwenhuizen <bas@basnieuwenhuizen.nl>
