Maintaining two different paths is annoying but this gets
rid of the performance regression introduced by the global
BO list.
We might find a better solution in the future, but for now
just keeps two paths.
Signed-off-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>
Reviewed-by: Bas Nieuwenhuizen <bas@basnieuwenhuizen.nl>
In order to reduce a performance regression introduced by
4b13fe55a4 ("radv: Keep a global BO list for VkMemory."),
we are going to maintain two different paths.
One when VK_EXT_descriptor_indexing is enabled by the
application because we need to have a global BO list, and
one (the old one) when it's not enabled.
With Talos on Polaris, the global BO list reduces performance
by 10% which is too much for me.
This reverts commit ab6cadd3ec.
Reviewed-by: Bas Nieuwenhuizen <bas@basnieuwenhuizen.nl>
I have seen a few applications and games do the dynamic buffer bounds incorrectly, this
make it easier to work around, e.g. for debugging.
Reviewed-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>
DCC implies a fast-clear eliminate, so I think this sounds
reasonable.
Signed-off-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>
Reviewed-by: Bas Niuwenhuizen <bas@basnieuwenhuizen.nl>
Into radv_handle_color_image_transition().
Signed-off-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>
Reviewed-by: Bas Niuwenhuizen <bas@basnieuwenhuizen.nl>
In order to separate initialization from decompression. In the
future, that will allow us to init DCC/FMASK/CMASK in one shot.
Signed-off-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>
Reviewed-by: Bas Niuwenhuizen <bas@basnieuwenhuizen.nl>
To handle CMASK, FMASK and DCC transitions in the same place.
Signed-off-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>
Reviewed-by: Bas Niuwenhuizen <bas@basnieuwenhuizen.nl>
Mostly because DCC implies a fast-clear eliminate and we
should be able to skip some DCC decompressions by setting
a predicate like for CMASK and FMASK.
Signed-off-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>
Reviewed-by: Bas Niuwenhuizen <bas@basnieuwenhuizen.nl>
According to Marek, not enabling it on Stoney has a significant
negative performance impact. (And I guess this might impact
performance on Raven as well)
The register settings are pretty much copied from radeonsi. I did
not put this in the pipeline as that would make the pipeline more
dependent on the format which mean we would have to have more
pipelines for the meta shaders.
v2: Don't clear RB+ regs if not enabled as the CLEAR_STATE packet
does already.
Reviewed-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>
Otherwise, the shader BOs are not added to the list on SI because
prefetching isn't supported. Calling radv_cs_add_buffer() in the
prefetch codepath was a bad idea.
Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=105952
Fixes: 4ad7595f35 ("radv: rename radv_emit_prefetch() to radv_emit_prefetch_L2")
Signed-off-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>
Tested-by: Turo Lamminen <turo@alternativegames.net>
Reviewed-by: Bas Nieuwenhuizen <bas@basnieuwenhuizen.nl>
As we sometimes reset them to -1, -1 does not mean that they are
not written by the secondary command buffer.
Fixes: ad11fc3571 "radv: don't emit unneeded vertex state."
Reviewed-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>
The packet can sometimes be skipped, but we still think the change takes effect.
This just makes the packet always take effect.
Fixes: ad11fc3571 "radv: don't emit unneeded vertex state."
Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=105942
Reviewed-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>
Forgot one check... Too many mistakes for a simple change.
Fixes: f1d7c16e85 ("radv: fix prefetching compute shaders on CIK and older chips")
Signed-off-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>
Reviewed-by: Bas Nieuwenhuizen <bas@basnieuwenhuizen.nl>
They will help for DCC MSAA textures and if we support mipmaps
in the future.
Signed-off-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>
Reviewed-by: Bas Nieuwenhuizen <bas@basnieuwenhuizen.nl>
This is unnecessary because we record an error which should
be returned by vkEndCommandBuffer(), and the app shouldn't
submit a command buffer when this happens.
Signed-off-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>
Reviewed-by: Bas Nieuwenhuizen <bas@basnieuwenhuizen.nl>
Because the check was moved to radv_emit_prefetch_L2().
Fixes: 4ad7595f35 ("radv: rename radv_emit_prefetch() to radv_emit_prefetch_L2()")
Signed-off-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>
Reviewed-by: Bas Nieuwenhuizen <bas@basnieuwenhuizen.nl>
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>
This allows to start draws as soon as possible.
Signed-off-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>
Reviewed-by: Bas Nieuwenhuizen <bas@basnieuwenhuizen.nl>
Tested-by: Dieter Nützel <Dieter@nuetzel-hh.de>
Disabled by default for now, it can be enabled with
RADV_PERFTEST=outoforder.
No CTS regressions on Polaris, and all Vulkan games I tested
look good as well.
Expect small performance improvements for applications where
out-of-order rasterization can be enabled by the driver.
Loosely based on RadeonSI.
Signed-off-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>
Reviewed-by: Bas Nieuwenhuizen <bas@basnieuwenhuizen.nl>
This is just useless for two reasons:
1) flush_bits is not set accordingly, so nothing will be flushed
in BeginQuery().
2) we always flush caches in EndCommandBuffer(), so if a reset
is done in a previous command buffer we are safe.
Cc: "18.0" <mesa-stable@lists.freedesktop.org>
Signed-off-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>
Reviewed-by: Alex Smith <asmith@feralinteractive.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>
When transitioning to an htile compressed depth format, Set the full
depth range, so later rasterization can pass HiZ. Previously, for depth
only formats, the depth range was set to 0 to 0. This caused unwanted
HiZ rejections with a VK_FORMAT_D16_UNORM depth buffer
(VK_FORMAT_D32_SFLOAT was not affected somehow).
These values are derived from PAL [0], since I can't find the
specification describing the htile values.
[0] 5cba4ecbda/src/core/hw/gfxip/gfx9/gfx9MaskRam.cpp (L1500)
CC: Dave Airlie <airlied@redhat.com>
CC: Bas Nieuwenhuizen <bas@basnieuwenhuizen.nl>
CC: mesa-stable@lists.freedesktop.org
Reviewed-by: Bas Nieuwenhuizen <bas@basnieuwenhuizen.nl>
Tested-by: Grazvydas Ignotas <notasas@gmail.com>
Fixes: 5158603182 "radv: Use correct HTILE expanded words."
The Vulkan spec says:
"pipelineBindPoint is a VkPipelineBindPoint indicating whether
the descriptors will be used by graphics pipelines or compute
pipelines. There is a separate set of bind points for each of
graphics and compute, so binding one does not disturb the other."
Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=104732
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>
This gives about 2% performance improvement on dota2 for me.
This is mostly a mechanical copy and replacement, but at bind time
we still do:
1) Some stuff that is only based on num_samples changes.
2) Some command buffer state setting.
Reviewed-by: Dave Airlie <airlied@redhat.com>
Reviewed-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>
The GPU hangs when the driver forces a PS_PARTIAL_FLUSH after
a dispatch call (and vice versa for graphics). Something has
changed in the kernel driver because it used to work.
Signed-off-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>
Reviewed-by: Bas Nieuwenhuizen <bas@basnieuwenhuizen.nl>
We need to enable the pos float location 2 mode anytime we have
persample not just when forced by the frag shader.
This fixes:
dEQP-VK.pipeline.multisample.min_sample_shading*
Fixes: 58c97a079 (radv: enable location at sample when persample is forced.)
Reviewed-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>
Signed-off-by: Dave Airlie <airlied@redhat.com>
The user SGPR location can change between pipelines, so we need to
emit it again to the pottentially changed SGPR index.
Reviewed-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>
Looks like the decompress does not handle invalid encodings well,
which happens with random memory. Of course apps should not use it
with random memory, but they are allowed to ....
Fixes: 44fcf58744 "radv: Disable DCC for GENERAL layout and compute transfer dest."
Reviewed-by: Dave Airlie <airlied@redhat.com>
If the number of instances hasn't changed and we've already
emitted it, don't emit it again.
If the vertex shader is the same and the first_instance, vertex_offset
haven't changed don't emit them again.
This increases the fps in GL_vs_VK -t 1 -m -api vk from around 40
to around 60 here, it may not impact anything else.
Dieter also reported smoketest going from 1060->1200 fps.
Reviewed-by: Bas Nieuwenhuizen <bas@basnieuwenhuizen.nl>
Reviewed-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>
Tested-by: Dieter Nützel <Dieter@nuetzel-hh.de>
Signed-off-by: Dave Airlie <airlied@redhat.com>
Tested with a modified deferred demo and no regressions in a 1.0.2
mustpass run.
Reviewed-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>
Reviewed-by: Dave Airlie <airlied@redhat.com>
The EXT values are really large, e.g.
VK_DYNAMIC_STATE_DISCARD_RECTANGLE_EXT = 1000099000, so 1 << value
is not going to fit into a 32-bit mask.
Reviewed-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>
Reviewed-by: Dave Airlie <airlied@redhat.com>
For Vega10 and Raven that need a special workaround for the
scissor bug.
This seems to give a minor boost for Talos and Dota 2, at least.
To reduce the cost of memcmp, the driver checks if it's
really useful to do the comparison.
Signed-off-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>
Reviewed-by: Bas Nieuwenhuizen <bas@basnieuwenhuizen.nl>
These are just shaders reads, so we need to invalidate L1.
Fixes: 6dbb0eaccc "radv: handle subpass cache flushes"
Reviewed-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>
Seems like users are actually hitting 0xFFFFFFFF actually making
things broken for them, and the mad max regression is fixed, so
lets put this in once more.
v2: Use 0xf for depth-only htile. (Dave)
Fixes: af2844116f "radv: Revert HTILE reset word to 0xFFFFFFFF."
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>
Apps can use this for render feedback loops, where things are
defined if they render each pixel only once. However, DCC fails
here, as the level of coherence is a block not a pixel, so disable it.
This is also going to help implementing other stuff.
Even if we optimize this later to only happen if there actually is
a loop (if possible at all ...), then the machinery is still useful
to exclude images accessible by the SDMA queue when that is implemented.
Reviewed-by: Dave Airlie <airlied@redhat.com>
Tested-by: Dieter Nützel <Dieter@nuetzel-hh.de>
The event emission wasn't sending the correct packet for gfx8 compute
queues, which explains why it works on vega fine.
This fixes the mpv vulkan 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>
This fixes vmfaults seen on vega with:
dEQP-VK.pipeline.multisample_interpolation.sample_interpolate_at_single_sample_.128_128_1.samples_1
These were caused by the don't allocate cmask but it was just accidental.
The actual problem was the shader was trying to get the sample positions from
a buffer, but the buffer was never getting configured to contain them, as the
previous shader never needed them.
Reviewed-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>
Reviewed-by: Bas Nieuwenhuizen <bas@basnieuwenhuizen.nl>
Fixes: 1171b304f3 (radv: overhaul fragment shader sample positions.)
Signed-off-by: Dave Airlie <airlied@redhat.com>
pipeline->layout can't be NULL now.
Signed-off-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>
Reviewed-by: Bas Nieuwenhuizen <bas@basnieuwenhuizen.nl>
The number of grid components is always 3 when gl_NumWorkGroups
is declared, because it relies on the number of components of
nir_instrinsic_load_num_work_groups.
Signed-off-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>
Reviewed-by: Bas Nieuwenhuizen <bas@basnieuwenhuizen.nl>
dota2 binds a ton of index buffers but the type is always 16-bit.
Note that we have to invalidate the type when switching from
indexed draws to normal draws.
Signed-off-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>
Reviewed-by: Dave Airlie <airlied@redhat.com>
dota2 always calls vkResetCommandBuffer() before
vkBeginCommandBuffer() which is quite useless.
Signed-off-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>
Reviewed-by: Dave Airlie <airlied@redhat.com>
RADV_CMD_BUFFER_STATUS_INVALID is not used for now, but I think
it makes sense to declare it. Could be used later with better
command buffer error handling.
Signed-off-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>
Reviewed-by: Dave Airlie <airlied@redhat.com>
"vkAllocateCommandBuffers can be used to create multiple command
buffers. If the creation of any of those command buffers fails, the
implementation must destroy all successfully created command buffer
objects from this command, set all entries of the pCommandBuffers
array to NULL and return the error."
This has been suggested by gabriel@system.is.
Signed-off-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>
Reviewed-by: Bas Nieuwenhuizen <bas@basnieuwenhuizen.nl>
This fixes two CTS regressions:
- dEQP-VK.api.object_management.alloc_callback_fail_multiple.command_buffer_primary
- dEQP-VK.api.object_management.alloc_callback_fail_multiple.command_buffer_secondary
These two tests are part the mustpass lists, so presumably they
are correct and my change was wrong.
This reverts commit 0f68208f1d.
Signed-off-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>
Reviewed-by: Bas Nieuwenhuizen <bas@basnieuwenhuizen.nl>
Similar to how the driver sets the depth clear regs after a
fast depth clear. Most of the time, this will copy a 32-bit reg
instead of a 64-bit reg.
Signed-off-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>
Reviewed-by: Bas Nieuwenhuizen <bas@basnieuwenhuizen.nl>
For the fast path, radv_fill_buffer() ensures that the BO is
already in the list. For the slow path, the depth surface is
part of the framebuffer which means the BO is added to the list
when the framebuffer is emitted.
Signed-off-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>
Reviewed-by: Bas Nieuwenhuizen <bas@basnieuwenhuizen.nl>
aspects can't be zero and there is an assertion that ensures
it's not in emit_clear().
Signed-off-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>
Reviewed-by: Bas Nieuwenhuizen <bas@basnieuwenhuizen.nl>
This should reduce the overhead of adding a BO to the current
list, especially when the list is huge. Also, when a new pipeline
is bound, we only need to update the descriptor, the buffer objects
should already be in the list.
Signed-off-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>
Reviewed-by: Bas Nieuwenhuizen <bas@basnieuwenhuizen.nl>
radv_fill_buffer() ensures that the image BO is added to the list.
Signed-off-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>
Reviewed-by: Bas Nieuwenhuizen <bas@basnieuwenhuizen.nl>
radv_fill_buffer() ensures that the image BO is added to the list.
Signed-off-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>
Reviewed-by: Bas Nieuwenhuizen <bas@basnieuwenhuizen.nl>
This should not be needed, if the allocation fails an error is
returned and the host should handle it.
Signed-off-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>
Reviewed-by: Bas Nieuwenhuizen <bas@basnieuwenhuizen.nl>
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>
Just after the vertex shader.
This seems to give a minor boost for, at least, Serious Sam
Fusion 2017 and Dawn of War 3. I don't see any real impacts
with The Talos Principle.
Signed-off-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>
Reviewed-by: Bas Nieuwenhuizen <bas@basnieuwenhuizen.nl>
Will be used for VBO descriptors prefetching.
Signed-off-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>
Reviewed-by: Bas Nieuwenhuizen <bas@basnieuwenhuizen.nl>
For consistency because this function will also prefetch VBO
descriptors.
Signed-off-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>
Reviewed-by: Bas Nieuwenhuizen <bas@basnieuwenhuizen.nl>
This register is the same on all gpus so far, so emit it in one
place and also for the pre-gfx9 gpus set the value in the pipeline
creation.
Reviewed-by: Bas Nieuwenhuizen <bas@basnieuwenhuizen.nl>
Signed-off-by: Dave Airlie <airlied@redhat.com>
This moves some calculations of register values into the pipeline
construction, it saves looking at outinfo in the cmd buffer emit.
Reviewed-by: Bas Nieuwenhuizen <bas@basnieuwenhuizen.nl>
Signed-off-by: Dave Airlie <airlied@redhat.com>
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>
The function that calls us has just added the buffer to the
list already, no need to try and add it again.
Reviewed-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>
Signed-off-by: Dave Airlie <airlied@redhat.com>
There's no point recalculating these the whole time on descriptor
emission, just store them at pipeline creation.
Reviewed-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>
Signed-off-by: Dave Airlie <airlied@redhat.com>
Instead of storing all the pointers and zeroing them all out,
just store a valid bitmask in the state. This also moves
the CmdBindPipeline path down the cpu usage path for the
multithreading demo as it no longer has to traverse MAX_SETS
to find the active descriptor sets.
Reviewed-by: Bas Nieuwenhuizen <bas@basnieuwenhuizen.nl>
Signed-off-by: Dave Airlie <airlied@redhat.com>
Grazvydas Ignotas
Samuel Pitoiset
Bas Nieuwenhuizen
James Legg
Matthew Nicholls
Dave Airlie