And add our own abstraction to amortize the cost of heap
allocation/switch.

While at it, we delay the RTV/DSV descriptor emission until those
are actually needed (when a command needing it is issued), and
we keep a hash table at the command buffer level to avoid re-issuing
RTV/DSV that were already issued in this command buffer level.

Signed-off-by: Boris Brezillon <boris.brezillon@collabora.com>

Instead of allocating/releasing the descriptor heaps at set
allocation/destruction.

The defragmentation logic is probably not optimal, but it's simple,
and can be improved if needed.

Signed-off-by: Boris Brezillon <boris.brezillon@collabora.com>

UAVs are incompatible with MSAA resources.

Signed-off-by: Boris Brezillon <boris.brezillon@collabora.com>

Signed-off-by: Boris Brezillon <boris.brezillon@collabora.com>

Now that event set/waits don't force a cmdlist split, we can get
rid of the dzn_batch concept and all the associated complexity.

Signed-off-by: Boris Brezillon <boris.brezillon@collabora.com>

We still wait/signal event fences at the cmdlist boundary, to support
host -> device signaling and inter-cmdlist signaling, but we no longer
split the cmdlist. That means that the event might take a bit longer
to expose its new state to the host, but it shouldn't impact
inter-cmdlist dependencies, since those are serialized at the
ExecuteCommandList() level anyway.

Signed-off-by: Boris Brezillon <boris.brezillon@collabora.com>

Signed-off-by: Boris Brezillon <boris.brezillon@collabora.com>

When vkCmdPipelineBarrier() is not passed any memory barriers, it's
assumed to be an execution barrier.

Signed-off-by: Boris Brezillon <boris.brezillon@collabora.com>

And transition to the generic queue_submit() infrastructure.

Signed-off-by: Boris Brezillon <boris.brezillon@collabora.com>

Signed-off-by: Boris Brezillon <boris.brezillon@collabora.com>

We need SetEventOnMultipleFenceCompletion() to implement
vk_sync.wait_many(), and this method is part of the ID3D12Device1
interface.

Signed-off-by: Boris Brezillon <boris.brezillon@collabora.com>

We map it to DXGI_FORMAT_B4G4R4A4_UNORM and patch the component
mapping at SRV creation time.

Signed-off-by: Boris Brezillon <boris.brezillon@collabora.com>

DXIL doesn't like when SV_Coverage (AKA SampleMask in DXIL) is a
signed integer. Fix the type in the NIR domain.

Signed-off-by: Boris Brezillon <boris.brezillon@collabora.com>

Signed-off-by: Boris Brezillon <boris.brezillon@collabora.com>

nir_intrinsic_get_var() can return a NULL value.

Signed-off-by: Boris Brezillon <boris.brezillon@collabora.com>

pCreateInfo->subresourceRange.levelCount can map to the magic
ALL_LEVELS value. Use the level_count local variable which
contains the actual number of levels.

Signed-off-by: Boris Brezillon <boris.brezillon@collabora.com>

We should disable rasterization when rasterizerDiscardEnable is true
or when front+back face culling is requested.

Signed-off-by: Boris Brezillon <boris.brezillon@collabora.com>

The pipeline layout is not guaranteed to stay around after the pipeline
has been created. We need to copy all the relevant info to the pipeline
object at creation time.

Signed-off-by: Boris Brezillon <boris.brezillon@collabora.com>

Signed-off-by: Boris Brezillon <boris.brezillon@collabora.com>

vk_command_buffer already has it.

Signed-off-by: Boris Brezillon <boris.brezillon@collabora.com>

Louis-Francis Ratté-Boulianne authored 3 years ago and Boris Brezillon committed 3 years ago

This is paving the road for generic secondary command buffer support,
where commands are simply recorded in a software queue and replayed
on the primary command buffer when vkCmdExecuteCommands() is called.

Louis-Francis Ratté-Boulianne authored 3 years ago and Boris Brezillon committed 3 years ago

Nothing fancy here, we just auto-generate a bunch of functions doing:

   if (cmd_buffer->level == PRIMARY)
      return <primary_prefix>_CmdXXX(...);
   else
      return <secondary_prefix>_CmdXXX(...);

to avoid duplicating the same boilerplate in drivers using this sort
of dispatching.

Louis-Francis Ratté-Boulianne authored 3 years ago and Boris Brezillon committed 3 years ago

If the secondary command buffer has a software command queue, we can
auto-generate the CmdExecuteCommands() function and replay all commands
recorded to this software queue on the primary command buffer.

Louis-Francis Ratté-Boulianne authored 3 years ago and Boris Brezillon committed 3 years ago

That's basically what lavapipe does, except this time it's autogenerated
and can be re-used by other drivers if they don't have a smarter way
of supporting secondary command buffers.

Louis-Francis Ratté-Boulianne authored 3 years ago and Boris Brezillon committed 3 years ago

This should simplify auto-generation of CmdExecuteCommands().

Louis-Francis Ratté-Boulianne authored 3 years ago and Boris Brezillon committed 3 years ago

Looks like 3 implementations already have that field in their private
command_buffer struct, and having it at the vk_command_buffer opens the
door for generic (but suboptimal) secondary command buffer support.

We will need it to implement CmdBindDescriptorSets() on secondary command
buffers.

Signed-off-by: Boris Brezillon <boris.brezillon@collabora.com>

This field should only be accessed by dzn_descriptor_set methods.

Signed-off-by: Boris Brezillon <boris.brezillon@collabora.com>

The spec says:
"
  VkDescriptorSetLayout objects may be accessed by commands that operate
  on descriptor sets allocated using that layout, and those descriptor
  sets must not be updated with vkUpdateDescriptorSets after the
  descriptor set layout has been destroyed. Otherwise, a
  VkDescriptorSetLayout object passed as a parameter to create another
  object is not further accessed by that object after the duration of
  the command it is passed into.
"

We need to make sure all set info are copied to the pipeline layout
and avoid accessing those pointers outside the vkUpdateDescriptorSets
path.

Signed-off-by: Boris Brezillon <boris.brezillon@collabora.com>

We don't support secondary command buffers natively anyway, so let's
just assume the dzn_cmd_buffer always created direct command lists.

Signed-off-by: Boris Brezillon <boris.brezillon@collabora.com>

Signed-off-by: Boris Brezillon <boris.brezillon@collabora.com>

Signed-off-by: Boris Brezillon <boris.brezillon@collabora.com>

Signed-off-by: Boris Brezillon <boris.brezillon@collabora.com>

Signed-off-by: Boris Brezillon <boris.brezillon@collabora.com>

Signed-off-by: Boris Brezillon <boris.brezillon@collabora.com>

Signed-off-by: Boris Brezillon <boris.brezillon@collabora.com>

VkClearRect can cover more than one layer and can't be translated
directly to a D3D12_RECT. We need to clear one rectangle at a time to
be safe.

Signed-off-by: Boris Brezillon <boris.brezillon@collabora.com>

Signed-off-by: Boris Brezillon <boris.brezillon@collabora.com>