This commit does a couple of things in one go.
* The layer squasher now dispatches once per view.
* Reducing the number of samplers needed.
* Allows for per view target.
* Refactor the setup code to comp_util.
* The layout transition part now happens in new helper code.
On my system this change is very marginally slower,
around 0.05ms and 0.1ms slower on average.
Includes a workaround for D3D12 and Vulkan interop issues in NVidia
hardware. We want to directly allocate in D3D12 to be able to use all
resource flags during creation, such as multi-gpu ones.
Add a D3D12 allocator and helper code to copy from shadow images,
the copy is needed to work around a interop issue on NVIDIA hardware.
Co-authored-by: Ryan Pavlik <ryan.pavlik@collabora.com>
Work done by lots of different authors that did various pieces of the library.
Co-authored-by: Korcan Hussein <korcan.hussein@collabora.com>
Co-authored-by: Jakob Bornecrantz <jakob@collabora.com>
Co-authored-by: Ryan Pavlik <ryan.pavlik@collabora.com>
Relation history was being accessed guarded by an external lock in a couple of
places. This is redundant since its implementation already defines an internal
lock.
Relation history was being accessed guarded by an external lock in a couple of
places. This is redundant since its implementation already defines an internal
lock.
Relation history was being accessed guarded by an external lock in a couple of
places. This is redundant since its implementation already defines an internal
lock.
This fixes a bug on NVIDIA Jetson. Note this isn't so much the NVIDIA Jetson
fault, while the code was working on desktop, Monado did something wrong.
What happned was that Monado would select a mode with one size, while then
creating a VkSurface/VkSwapchain of a different size. This would work on
hardware with scalers/panning modes. The NVIDIA Jetson apparently doesn't have
support for that so failed when presenting. This patch makes sure that the
VkSurface/VkSwapchain extents match the mode for all direct mode targets.
