Emitting render pass clears while we're in the process of starting
a render pass overrides dsv layout tracking info.
Signed-off-by: Philip Rebohle <philip.rebohle@tu-dortmund.de>
D3D12 validation layer errors out, so unless we can prove that specific
behavior is relied upon, we should be okay to just ignore.
Signed-off-by: Hans-Kristian Arntzen <post@arntzen-software.no>
Goal here is to avoid unnecessary image layout transitions when render
passes toggle depth-stencil PSO states. Since we cannot know which
states a resource is in, we have to be conservative, and assume that
shader reads *could* happen.
The best effort we can do is to detect when writes happen to a DSV
resource. In this scenario, we can deduce that the aspect cannot be
read, since DEPTH_WRITE | RESOURCE state is not allowed.
To make the tracking somewhat sane, we only promote to OPTIMAL if an
entire image's worth of subresources for a given aspect is transitioned.
The common case for depth-stencil images is 1 mip / 1 layer anyways.
Some other changes are required here:
- Instead of common_layout for the depth image, we need to consult the
command list, which might promote the layout to optimal.
- We make use of render pass compatibility rules which state that we can
change attachment reference layouts as well as initial/finalLayout.
To make this change, a pipeline will fill in a
vkd3d_render_pass_compat struct.
- A command list has a dsv_plane_optimal_mask which keeps track
of the plane aspects we have promoted to OPTIMAL, and we know cannot
be read by shaders.
The desired optimal mask is (existing optimal | PSO write).
The initial existing optimal is inherited from the command list's
tracker.
- RTV/DSV/views no longer keep track of VkImageLayout. This is
unnecessary since we always deduce image layout based on context.
Overall, this shows a massive gain in HZD benchmark (RADV, 1440p ultimate, ~16% FPS on RX 6800).
Signed-off-by: Hans-Kristian Arntzen <post@arntzen-software.no>
Idea is to keep track of scenarios where we know a resource's aspect is
known to be in a OPTIMAL state. Based on this, we can override the image
layout from the common_layout in order to avoid unnecessary full
barriers.
Signed-off-by: Hans-Kristian Arntzen <post@arntzen-software.no>
For copies, we can always use the intended aspects, since we have
separate DS layouts now.
Signed-off-by: Hans-Kristian Arntzen <post@arntzen-software.no>
When clearing a DSV, we must get aliasing guarantees, so we must
transition away from UNDEFINED. This is only possible when using
separate_ds_layouts and for render pass clears we need to use
renderpass2 mechanisms to do this.
Signed-off-by: Hans-Kristian Arntzen <post@arntzen-software.no>
GravityBench ends up using ClearView with too large dimensions.
This is a validation error in Vulkan, so just clamp the extents.
To make full rect detection a bit more robust, do a range check instead
of memcmp().
Signed-off-by: Hans-Kristian Arntzen <post@arntzen-software.no>
If we're doing a layout transition of depth-stencil aspects, we need to ensure all potential
accesses are made visible.
Signed-off-by: Hans-Kristian Arntzen <post@arntzen-software.no>
- Honor resource barriers for resource states which cannot automatically
decay or promote. This includes COLOR_ATTACHMENT, UNORDERED_ACCESS and
VRS image. If SIMULTANEOUS_ACCESS is used, we can still promote, and
we handle that by setting common layout to GENERAL for these resources.
- Avoid redundant barriers in render passes since normal resource
barriers will always make sure we are already in
COLOR_ATTACHMENT_OPTIMAL.
- Do not force GENERAL layout if resource has UNORDERED_ACCESS flag set.
As this is not a promotable state, we have to explicitly transition
into it. I tested this on validation layers, where even COMMON state
refuses to promote to UAV state. The exception here of course is
SIMULTANOUS_ACCESS, but we handle that properly now.
- Verify that UAV or SIMULTANEOUS access is not used together with DSV
state. This is explicitly banned in the API docs.
- Actually emit image barriers. Batch the image transitions as that's
what D3D12 docs encourage app developers to do, and it also expects
that drivers can optimize this. Ensure that we respect the in-order
resource barrier rules by splitting batches if there are overlaps in
the transitions.
- Ensure that correct image layout is used when clearing a suspended
render pass attachment.
Signed-off-by: Hans-Kristian Arntzen <post@arntzen-software.no>
Avoid using the separate layouts if we're only using formats with one
aspects. This makes it more likely to match layouts with common layout,
and we can avoid awkward transition barriers.
Signed-off-by: Hans-Kristian Arntzen <post@arntzen-software.no>
Fixes test TODOs. Apparently Vulkan drivers can saturate here, which
caused the TODO to appear, at least on AMD Windows.
Signed-off-by: Hans-Kristian Arntzen <post@arntzen-software.no>
Needed so we can switch between having a VRS and non-VRS attachment on the fly.
Extensible enough for this to work for other things down the line also.
Signed-off-by: Joshua Ashton <joshua@froggi.es>
This can happen if the fence thread starts with a delay and
the queue gets destroyed shortly after being created.
Signed-off-by: Philip Rebohle <philip.rebohle@tu-dortmund.de>
There isn't much of a reason why we should have to do this. The original
implementation was more of a hack if anything.
Signed-off-by: Hans-Kristian Arntzen <post@arntzen-software.no>
Increment physical value one by one, find the exact timeline value we're
supposed to signal and perform the update.
Select lowest physical timeline value correctly.
Array can be reordered now, so lowest value isn't necessarily first.
Fixes some super weird hangs in Control DXR.
Signed-off-by: Hans-Kristian Arntzen <post@arntzen-software.no>
Rather than one per device. This solves issues with D3D12 fences
being signalled too late because the fence worker is waiting on
a different set of semaphores while the fence is being enqueued.
Greatly increases performance in Horizon Zero Dawn and Death
Stranding with multi-queue mode enabled.
Signed-off-by: Philip Rebohle <philip.rebohle@tu-dortmund.de>
This will be necessary once we introduce fence workers per
command queue, since we cannot reliably store pointers to
queues.
Signed-off-by: Philip Rebohle <philip.rebohle@tu-dortmund.de>
Replaces d3d12_device_get_vkd3d_queue when mapping D3D12
command queues to Vulkan device queues.
Signed-off-by: Philip Rebohle <philip.rebohle@tu-dortmund.de>
Compacts ranges and only issues one bind for buffer ranges and
full subresource updates, rather than one bind per tile.
Signed-off-by: Philip Rebohle <philip.rebohle@tu-dortmund.de>
Bindless CBV is *pretty* bad on NVIDIA, so add a code path which can
promote descriptor table CBVs into push descriptors.
We can safely do this with Root Signature 1.1 STATIC or
the somewhat obscure STATIC_KEEPING_BUFFER_BOUNDS_CHECKS.
With VOLATILE, which basically all titles are using,
we can still force this behavior through a config flag,
but this is an incorrect speed hack. It works in most
titles however, since bindless CBV is exceptionally rare.
We only hoist descriptors when the root signature range has 1 descriptor
anyway, so we should avoid any reasonable bindless scenario.
Signed-off-by: Hans-Kristian Arntzen <post@arntzen-software.no>
We might have to emit to different bind point than our binding entry
suggests due to DXR, so pass down information explicitly to leaf
functions.
Signed-off-by: Hans-Kristian Arntzen <post@arntzen-software.no>
Refactor push constant invalidation to SetPipelineState,
it is technically more correct to only invalidate when actually pushing
constants, but we need to do full state invalidation when transitioning
between RT pipelines and non-RT pipelines due to bind point aliasing
shenanigans in D3D12, so it makes more sense to invalidate state based
on active bind point there.
Signed-off-by: Hans-Kristian Arntzen <post@arntzen-software.no>
Gives a massive boost on NVIDIA for some reason.
RADV defers push constant update, so ALL_STAGES doesn't have
that much of a perf hit.
~20% uplift in RE2, ~5% uplift in CP77 from some quick and dirty testing.
Seems to be heavily content dependent either way.
Also a bug fix, since we would clobber graphics push constants from
compute and vice versa if both graphics and compute used the same root
signature.
Signed-off-by: Hans-Kristian Arntzen <post@arntzen-software.no>
As per MSDN, SetName is just a wrapper around SetPrivateData and a specific GUID.
Some apps and tools will use this to retrieve their name back.
So instead, just forward the name to Vulkan in the SetPrivateData call.
Signed-off-by: Joshua Ashton <joshua@froggi.es>
Can only support a subset in Vulkan without extra heroics. The DXR API
lets you query things that you technically should know apriori in the
application. We might need to allocate some side-channel buffers on
demand, but let's defer that until actually needed ... :\
DXR is also very awkward in that we have a query which is resolved in
UNORDERED_ACCESS state instead of COPY_DEST state, so we'll have to
ping-pong through some barriers redundantly.
Signed-off-by: Hans-Kristian Arntzen <post@arntzen-software.no>
Allows local root signatures to work correctly and is also a good
optimization since we no longer need to dereference memory (potentially
cold cache lines) to figure out heap offset in command buffer.
Signed-off-by: Hans-Kristian Arntzen <post@arntzen-software.no>
If we're signalling and waiting on same physical queue (always true for
current SINGLE_QUEUE define), we can rely on submission boundary
synchronization which doesn't require any extra submissions to resolve.
Avoids awkward GPU driver bubbles with back to back signal -> wait pairs
with timeline.
Observed 2% GPU uplift on RE2 on AMD.
Signed-off-by: Hans-Kristian Arntzen <post@arntzen-software.no>
Need it here since local root signatures need to know
the physical layout of the record buffer up front.
Signed-off-by: Hans-Kristian Arntzen <post@arntzen-software.no>
Fixes a validation error. With VK_QUERY_RESULT_64_BIT we need
to use 8-byte alignment, but ssbo_alignment may be less.
Signed-off-by: Philip Rebohle <philip.rebohle@tu-dortmund.de>
No longer requires BDA support since it's easier now to work
around buffer alignment issues.
Signed-off-by: Philip Rebohle <philip.rebohle@tu-dortmund.de>
The first range will store the byte offset, the second one will
be the typed buffer range. Typed descriptors should write both.
Signed-off-by: Philip Rebohle <philip.rebohle@tu-dortmund.de>
Co-authored-by: Hans-Kristian Arntzen <post@arntzen-software.no>
We currently never reset occlusion queries. For some reason,
validation layers do not report this.
Signed-off-by: Philip Rebohle <philip.rebohle@tu-dortmund.de>
By resetting query pools in advance, we can reduce the number of
stalls between draw calls in passes with occlusion queries, which
is currently causing serious performance issues in some games.
Signed-off-by: Philip Rebohle <philip.rebohle@tu-dortmund.de>
Since we'll be inserting lots of single queries, we want to
avoid having to resize the range array since that is an O(n)
operation at worst.
Signed-off-by: Philip Rebohle <philip.rebohle@tu-dortmund.de>
Official AMD drivers do not support VK_EXT_conditional_rendering,
so we'll use indirect draws instead to emulate the feature.
This also handles 64-bit predicates in combination with the
Vulkan extension, which was not possible previously.
Signed-off-by: Philip Rebohle <philip.rebohle@tu-dortmund.de>
Potentially avoids some unnecessary host memory access. Use BDA for
the compute shader so that we can ignore alignment restrictions on
some GPU architectures.
Signed-off-by: Philip Rebohle <philip.rebohle@tu-dortmund.de>
Command lists may need to allocate temporary device memory for
certain operations. In order to avoid frequent alloc/free calls,
we'll recycle these scratch buffers until a certain threshold.
Signed-off-by: Philip Rebohle <philip.rebohle@tu-dortmund.de>
Realign VBO strides and offsets if we have to, for sake of
robustness. Violating these rules is against D3D12 spec, but it does not
cause crashes on native drivers. On RDNA we can hit hangs with unaligned
vertex attributes. It appears that native drivers apply some kind of
fixup here to avoid the crash, even if the result is not what we expect.
Signed-off-by: Hans-Kristian Arntzen <post@arntzen-software.no>
Ensures that queries are always available and initialized
in the correct order on the GPU timeline.
Signed-off-by: Philip Rebohle <philip.rebohle@tu-dortmund.de>
We have observed a lot of large GPU bubbles when using back-to-back
timeline semaphores to synchronize GPU submissions. Use prebaked
pipeline barrier command buffers instead.
To resolve queue sparse serialization, use two binary semaphore pairs to
resolve this. There is no need to use timeline semaphores in this case.
Signed-off-by: Hans-Kristian Arntzen <post@arntzen-software.no>
USE_PUSH_DESCRIPTORS may be misleading since it would be set even when
we're not using push descriptors at all due to root descriptors being
passed in via VAs. Instead, make the flag represent whether or not we
use a regular descriptor set for root parameters.
Signed-off-by: Philip Rebohle <philip.rebohle@tu-dortmund.de>
The packed descriptor index is no longer needed, and causes issues in
case a game sets a root signature, then binds a root descriptor, and
then sets a different root signature which maps the given root parameter
index to a different descriptor since we may now read undefined data
when updating push descriptors.
Fixes#366.
Signed-off-by: Philip Rebohle <philip.rebohle@tu-dortmund.de>
The struct definitions were identical anyway, and unifying
these will prevent unnecessary code duplication.
Signed-off-by: Philip Rebohle <philip.rebohle@tu-dortmund.de>
The only currently known use case for this requires us to actually
perform the dispatch operation. Executing more than one indirect
dispatch command is not meaningful, however there might be
differences in behaviour in case the indirect count is zero.
Signed-off-by: Philip Rebohle <philip.rebohle@tu-dortmund.de>
This logic has to be the same as in d3d12_command_list_update_descriptor_table_offsets,
since not all active descriptor tables are necessarily used by the root signature.
Fixes an assert in the StarsX IrradianceMap demo (Github issue #347).
Signed-off-by: Philip Rebohle <philip.rebohle@tu-dortmund.de>
We will not have offset information for root descriptors, so
we can still only use them with four-byte aligned SSBOs.
Signed-off-by: Philip Rebohle <philip.rebohle@tu-dortmund.de>
We cannot rely on alignment analysis since games are buggy and screw up
RAW vs structured on occasion.
Signed-off-by: Hans-Kristian Arntzen <post@arntzen-software.no>
Relevant for swapchain since a swapchain resource can be presented right
away without ever having been touched by an API call.
Signed-off-by: Hans-Kristian Arntzen <post@arntzen-software.no>
It is broken by design and won't be needed by a swapchain
implementation which uses user buffers.
Signed-off-by: Hans-Kristian Arntzen <post@arntzen-software.no>
Hans-Kristian Arntzen
Philip Rebohle
rochaudhari
David McCloskey
Joshua Ashton
Georg Lehmann
Alexander Gabello