This is the first step to optimizing buffer updates for applications
that frequently invalidate buffers. The goal is to reduce the number
of buffer allocations per frame and reduce the cost of invalidation.
Improves performance in applications that use UpdateSubresources
to frequently update constant buffers, such as Unigine Heaven,
Unigine Valley and Homefront.
Support for extensions can now be queried from the device
object in an efficient way. This will allow the backend to
use optional extensions for the purpose of optimization.
Checks whether all input slots consumed by the vertex shader
are provided by the input layout, and disables rendering in
case the state validation fails. This should hopefully fix
GPU lockups in Nier:Automata.
The shader compiler now queries whether a constant buffer or texture is
bound before trying to access it for reading. This is not yet implemented
for image fetch operations, atomic operations, or buffer load/store ops.
In order to emulate D3D11 behaviour when a resource is not bound,
we use specialization constants. This requires further changes in
the shader compiler.
If an application binds vertex buffers before changing the input layout,
it might happen that the vertex buffers are not bound correctly to the
Vulkan command buffer. This issue is now resolved.
- When a D3D11 texture was created with a mip level count of zero,
DXVK would compute the correct number of mip levels, but not write
them back to the description field.
- Framebuffers would always use the size of the top level, not the
mip levels that were actually attached to the framebuffer.
- Several image copy operations did not handle mip levels other
than the top mip level correctly.
Image mapping now returns the map pointer of a separate
buffer, rather than the the image itself. This fixes
issues with applications that ignore the RowPitch
and/or DepthPitch fields of the MappedSubresource struct.
Unlike VK_KHR_maintenance1, not all D3D11 applications rely on
features provided by this extension. Keeping this optional helps
with RenderDoc, which does not expose the extension.
Four-byte alignment is required in order to avoid illegal
vkCmdCopyBufferToImage calls. We align staging buffer slices to
a full cache line in order to improve performance.
This fixes an issue where the wrong resource would be used for a
descriptor set update if multiple resources have been bound to a
single resource slot.
