Since vkCmdClearAttachments executes inside a render pass, we would
benefit from converting it to a draw within the current subpass job to
improve batching and avoid expensive tile load/store operations.
This can dramatically improve performance for applications using this
command, however, we can only use this if we are clearing the base
layers of framebuffer attachments, since otherwise we would need to
use layered rendering, which we don't support yet.
This improves vkQuake3 performance dramatically (almost 100%
performance improvement at 1080p), which calls this twice per frame.
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/6766>
This gets vkQuake to render correctly, which creates render passes with
stencil load operations even when the depth/stencil attachment format
doesn't have a stencil aspect. While this is a bit weird, it seems to
be allowed by the spec:
"If the format has depth and/or stencil components, loadOp and storeOp
apply only to the depth data, while stencilLoadOp and stencilStoreOp
define how the stencil data is handled."
In our case we were not ignoring it and this was causing that we emitted a
Z buffer load that seemed to clobber the Z clear, preventing all draw calls
from passing the depth test.
While we are at it, also change the depth/stencil store operation (which
was already handling this scenario correctly) to use the format of the
render pass attachment description rather than the underlying image
format.
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/6766>
So far we have been getting away with finishing the current job in the
presence of a pipeline barrier and relying on the RCL serialization,
but of course this is not always enough.
This patch addresses synchronization across different GPU units
(i.e. draw indirect after compute), as well as cases where we need to
sync before binning.
Fixes CTS failures in:
dEQP-VK.synchronization.op.single_queue.barrier.*
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/6766>
The order in which we emit the attributes is relevant, since
GL_SHADER_STATE_ATTRIBUTE_RECORD packets don't include an explicit
attribute index. This means that we need to emit them in driver
location order, since the compiler uses that location to compute
attribute offsets in the VPM.
Fixes ~1300 CTS tests in:
dEQP-VK.pipeline.vertex_input.multiple_attributes.out_of_order.*
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/6766>
For some reason, CTS expects E5B9G9R9 and B10G11R11 with
transparent black border clamping produce alpha 1 instead of 0.
Since border color takes precedence over the texture state swizzle,
the only way to fix this is to lower the texture swizzle in the shader
to set alpha to 1.
Fixes:
dEQP-VK.pipeline.sampler.view_type.*b10g11r11*clamp_to_border_transparent_black
dEQP-VK.pipeline.sampler.view_type.*e5b9g9r9*.clamp_to_border_transparent_black
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/6766>
It seems that we only want to set the texture state's depth to the
number of 2D layers divided by 6 when sampling, not wen doing
load/store.
This means that we need to generate two different states and choose
the one to use depending on the descriptor.
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/6766>
In OpenGL, unnormalized coordinates are implicit based on the sampler
type (rectangle textures), so the compiler can set the flag when needed.
In Vulkan, however, this is configured explicitly in the sampler object,
so the compiler won't set it and we need to do it manually when we are
writing the P1 uniform.
Fixes:
dEQP-VK.pipeline.sampler.exact_sampling.*.unnormalized_coords
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/6766>
Instead of asserting that users don't try to create images that
would require 4GB+ of memory, error out with the corresponding
OOM error when the user tries to actually allocate the memory
for the image.
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/6766>
If the meta operation did not change descriptor state then we should keep it,
not reset it.
Fixes:
dEQP-VK.fragment_operations.early_fragment.early_fragment_tests_stencil
dEQP-VK.fragment_operations.early_fragment.no_early_fragment_tests_stencil
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/6766>
Only free the underlying BO when the exported memory object is freed
to avoid multiple frees of the same memory.
The only exception is winsys BOs where we import a BO created in the
display device into the render device. In this case, we only have one
memory object referencing the BO and we want to destroy it with that
memory object.
Fixes:
dEQP-VK.api.external.memory.dma_buf.*
dEQP-VK.api.external.memory.opaque_fd.*
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/6766>
The hardware can't do this, so we need to record a CPU job that will
map the indirect buffer at queue submission time, read the dispatch
parameters and then submit a regular dispatch.
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/6766>
New jobs need to re-emit all state. Typically, this is achieved
by resetting all dirty state flags when we start a new job, but
for index buffers we were not using a dirty bit because we always
emit them immediately. This patch adds the bit and only tries
to skip index buffer state if the bit is not dirty, which will
ensure that we will always emit it for new jobs.
This fixes a regression in the shadowmapping demo from Sascha Willems
introduced with "v3dv: try harder to skip emission of redundant state".
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/6766>
V3DV_MAX_BO_CACHE_SIZE can be used to configure it.
So one way to disable the bo cache is setting V3DV_MAX_BO_CACHE_SIZE
to zero. This would still run all the bo_cache size checks, but having
another envvar just to ensure that anything related to the cache is
used seemed like an overkill.
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/6766>
Heavily based on the already existing for the v3d OpenGL driver, but
without references, and with some extra OOM checks (Vulkan CTS has
several OOM tests).
With this commit v3dv_bo_alloc and v3dv_bo_free became frontends to
the bo_cache. The former tries to get a BO from the cache if possible,
and the latter stores the BO on the cache if possible. The former also
adds a new parameter to point if the BO to allocate is private.
As v3d we are only caching private BOs, those created by the driver
for internal use (like CLs, tile_alloc, etc). They are the ones with
the highest change of being reused (for example, CL BOs are always
4KB, so they can always be reused). User-created BOs can have any
size, including some very large ones for buffers and images, which
makes them far less likely to be reused and would add a lot of memory
pressure if we decided to cache them.
In any case, in practice, we found that we could get a performance
improvement by caching also user-created BOs, but that would need more
care and an analysis to decide which ones makes sense. Would also
require to change how the cached BOs are stored by size. Right now
there are an array of list_head, that doesn't work well with big
BOs. If done, that would be handled on a separate commit.
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/6766>
The blit shader path for buffer to image copies is pretty bad,
since it needs to produce a tiled image from the linear buffer
prior to emitting the blit copy.
This patch adds a new preferential path where we implement the
copy using the CPU, similar to what the GL driver does for
texture uploads. This makes vkQuake2 at least 4x faster when
dynamic lights are enabled (which triggers dynamic texture
updates).
We also tested a GPU path where we use a shader that takes the
linear buffer as a UBO and copies directly from it. This also
shows a clear performance gain, but still worse than the CPU
implementation.
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/6766>
This allows us to remove some individual bos for the image and
sampler, used to store the SAMPLER_STATE and TEXTURE_SHADER_STATE. Now
they are prepacked on static memory as part of the vulkan object
struct.
This commit introduces small descriptor structs, used to define what
the bo subregion would contain. It is used mostly to compute offsets
to that specific data, and define the size needed. Having said so, it
would be possible to replace them with some kind of flag (like anv) or
just compute the offset based on the context.
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/6766>
So far we were saving all the descriptor info on the host memory. With
this commit we do the equivalent that other mesa vulkan drivers (Anvil
and Turnip) and create a bo on the descriptor pool that would be
suballocated for each descriptor.
This would allow to clean up individual bos from some vulkan objects,
reducing device memory fragmentation, and allowing to avoid to alloc
bos for that info. After all, pre-allocating needed memory is one of
the purposes of the descriptor pool.
This commit introduces all the infrastructure, but doesn't use it for
any descriptor yet, as if no descriptor needed data uploaded to a bo.
The idea to decide which info goes to the descriptor pool bo is info
that we would need to upload to a bo in any case, as it is referenced
as an address by any packet.
We could be more aggressive with that general rule, but that would be
enough for now. If in the future we support
VK_EXT_descriptor_indexing, we probably would need to store more info,
as under that extension, descriptors can be updated after being bound.
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/6766>
We were assuming that if the command buffer state doesn't have any
attachments (as per the attachment count) the attachment state array
should not be allocated, however, during meta operations it is
possible that the attachment state grows (since meta operations can
emit render passes of their own). In that case, we would grow the
state for the meta operation but then pop the previous attachment
count and we would leak the state.
An example of that is a secondary command buffer which has no
attachment state by default since it doesn't execute a render pass
begin, but that executes one in a meta operation (for
vkCmdClearAttachments for example).
Fix this by making the attachment count an allocation count instead
and not popping it once we finish a meta operation. Also, always free
the state so long as there is a valid pointer, and assert that the
allocated count is not zero in that case.
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/6766>
The main change we are introducing here is that now we allow secondary
command buffers that execute in a render pass to have a job list with
more than one job.
The main issue with vkCmdClearAttachments is that we currently need
this to spawn multiple jobs to clear multilayered framebuffers, as we
need to setup a different 2D framebuffer for each layer to clear and
therefore emit a different RCL for each. We could avoid this
completely by used layered rendering with the "clear rect" path to
redirect the clear rects to appropriate layers of the primary
framebuffer, however, our hardware only supports layered rendering
with geometry shaders, which we don't support at present.
Because vkCmdClearAttachments relies on having framebuffer state
available (something we would not need if we used the geometry shader
implementation), if this is not available in the secondary we need to
postpone emission of the command until the secondary is executed
inside a primary. We do this by using a new CPU job
V3DV_JOB_TYPE_CPU_CLEAR_ATTACHMENTS that is processed during
vkCmdExecuteCommands by calling vkCmdClearAttachments directly in the
primary.
As a consequence of these changes, it is now possible that a secondary
command buffer that runs inside a render pass have any kind of job in
its job list, including partial CLs that need to be branched to and
full CLs that need to be submitted to the GPU as is, so we introduced
a new GPU job type V3DV_JOB_TYPE_GPU_CL_SECONDARY to identify partial
CLs.
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/6766>
Event waits can be safely moved before a render pass start, since
event setting and resetting commands cannot happen inside one. We
don't need to go that far, but we can use this to record the wait
in its own separate job and then execute this job before the
binning commands recorded in the secondary command buffer when
we execute the secondary into a primary.
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/6766>
There are basically two types of scenarios to consider:
- Secondary command buffers that run inside a render pass.
- Secondary command buffers that run outside a render pass.
For the former we want to record their commands into a binning command
list that we can branch to when executed into a primary command
buffer. This means this kind of command buffers don't spawn new jobs,
just the default one where they record the binning commands which
won't include the frame setup, which will be provided by the primary
they will be executed in.
For the latter we don't require anything special, we just record as
many jobs as we need as usual and link that job list from the primary
job list when executed.
This handles most scenarios except:
- vkCmdWaitForEvents
- VkCmdClearAttachments
Both of these can spawn new jobs inside a render pass, which is not
what we want for secondary command buffers. We will address this is
follow-up patches.
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/6766>
In order to properly check (and possibly compile) shader variants we
need a pipeline and a compatible descriptor set. So far we were trying
to do that check as early as possible, so we were trying to do it at
CmdBindPipeline or CmdBindDescriptorSets, and a combination of dirty
flags. This showed to not cover all the corners cases, and made the
code complex, as needed to handle cases where the descriptors were not
yet available, and return early. The latter also meant that we were
running several checks that failed in the middle.
This commit moves the variant check to CmdDraw, when we should have a
pipeline and compatible descriptor sets, and simplifies and makes more
strict the existing code.
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/6766>
This reverts a previous half-attempt at an implementation of events
using a BO to hold the event state, and provides a full
implementation. V3D doesn't have any built-in GPU functionality to
wait on any kind of events, so we need to implement this in the driver
an therefore we no longer need to use a BO for the event state.
Instead, we implement GPU waits by using a CPU job for the wait
operation that spawns a wait thread if the wait operation doesn't have
all its events signaled by the time it is processed. To implement the
semantics of the wait correctly, any jobs in the same command buffer
that come after the wait will not be emitted until the wait thread
completes.
If a submit spawns any wait threads for a command buffer we can't
signal any semaphores for it until all the wait threads complete and
we know that all the jobs for those command buffers have been
submitted. The same applies to the submit fence, if present.
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/6766>
This is generally very difficult to handle properly everywhere, but
at least this is good enough to make the few CTS tests for this happy.
Fixes (on Rpi4):
dEQP-VK.wsi.xlib.swapchain.simulate_oom.*
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/6766>
We were declaring the destroy callback function as taking a pointer for the
vulkan object handle and relying on an implicit conversion to the Vulkan
handle type, however that would be incorrect on 32-bit platforms, where
non-dispatchable Vulkan objects (the kind that we may allocate privately during
command buffer recording), are defined as uint64_t, so the signature of the
destry callback type doesn't match the signature of the actual Vulkan
function, leading to bogus results. Fix that by using uint64_t instead.
This fixes compilation warnings and also crashes in some tests when
compiling and executing natively in Rpi4.
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/6766>
We were pre-packing the constants from the pipeline state and then
always emitting that at draw time, ignoring dynamic state. This makes
it so we don't prepack at pipeline creation time and we always emit
the correct constants directly the command buffer dynamic state.
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/6766>
We can now implement all depth/stencil blits as compatible color blits,
so let's just have the blit shader interface simply convert any blit with
a depth/stencil format to a compatible color blit (like we were already
doing for s8d24) and get rid of the depth blitting path. This also allows
us to ignore the blit aspect in the blit pipeline cache key.
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/6766>
We implement this by blitting the requested region to a linear image
setup to use the buffer memory store at the requested offset.
Because we can't store linear depth/stencil images, we implement
copies of depth/stencil aspects by using a compatible color blit.
To do this, we also need to account for the fact that when we are
copying depth from a d24 format we need to copy them from the MSB
24-bits of each word as provided by the hardware and store them
in the LSB 24-bit of the buffer (as per Vulkan requirements). This
is achieved by expanding our blit interface to also accept a swizzle
to apply to the source texture.
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/6766>
To do this we just implement the stencil blit as a masked color bit
with uint8 format. This allows us to support blitting on combined
depth/stencil formats, and therefore, also partial image copies
for these formats.
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/6766>
So far we were just asserting or aborting if any of the internal
method used during the pipeline creation failed.
We needed to change the return value of several methods, in order to
bubble up the proper memory allocation error.
Note that as the pipeline creation is complex and splitted in several
methods, if an error happens in the middle, it returns back, and rely
on the higher level to call PipelineDestroy. This method needs to take
into account that some of the resources could have not been allocated
when freeing it.
Also note that v3dv_get_shader_variant is used during the pipeline
bind, as with the new resources bound, we need to check if we need to
recompile a new variant. At that moment we are not creating a new
vulkan object so we can really return a OOM error. For now we just
assert on that case.
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/6766>
This uses the framework to register private commmand buffer objects
that get freed automatically when the command buffer is destroyed by
the application.
This change also moves the descriptor set pool that the meta blit path
uses to allocate descriptors for the blit source textures, from the
device to the command buffer, so we can have a descriptor pool per
command buffer. This is necessary to ensure correct behavior when
doing multi-threaded command buffer recording (alternatively, we would
have to lock around the descriptor set allocation code, which would be
undesirable).
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/6766>
This allows the driver to register private Vulkan objects it creates as part
of command buffer recording (usually for meta operations) in the command
buffer, so they can be destroyed together with it.
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/6766>
While very limited in scope, this might be the most efficient way to blit
when applicable. In fact, we might also want to use this for the image copy
commands when possible instead of the TLB.
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/6766>
The design for queries in Vulkan requires that some commands execute
in the GPU as part of a command buffer. Unfortunately, V3D doesn't
really have supprt for this, which means that we need to execute them
in the CPU but we still need to make it look as if they happened
inside the comamnd buffer from the point of view of the user, which
adds certain hassle.
The above means that in some cases we need to do CPU waits for certain
parts of the command buffer to execute so we can then run the CPU
code. For exmaple, we need to wait before executing a query resets
just in case the GPU is using them, and we have to do a CPU wait wait
for previous GPU jobs to complete before copying query results if the
user has asked us to do that. In the future, we may want to have
submission thread instead so we don't block the main thread in these
scenarios.
Because we now need to execute some tasks in the CPU as part of a
command buffer, this introduces the concept of job types, there is one
type for all GPU jobs, and then we have one type for each kind of job
that needs to execute in the CPU. CPU jobs are executed by the queue
in order just like GPU jobs, only that they are exclusively CPU tasks.
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/6766>
We were enabling VK_FORMAT_FEATURE_SAMPLED_IMAGE_FILTER_LINEAR_BIT for
any format valid for texturing, but for example, right now we don't
support linear filtering on any depth format.
This is needed to get some hundreds of tests like this:
dEQP-VK.pipeline.sampler.view_type.1d.format.r32g32_sfloat.mag_filter.linear
properly skipped (those were all Crashes with the simulator, and
almost all Fails with the real device).
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/6766>
Iago Toral Quiroga
Alejandro Piñeiro