Use none shader when XRT_DISTORTION_MODEL_NONE is set.
Fall back to none distortion if mesh generation fails or unsupported
distortion model is set.
comp_render init updates comp_settings::distortion_model.
u_compute_distortion_mesh: Don't touch xdev::distortion::preferred,
just add XRT_DISTORTION_MODEL_MESHUV mesh to xdev::distortion::models.
Each HMD driver now has to implement compute_distortion() which will be called
by the compositor implementation to generate a mesh (usually).
u_distortion_mesh contains implementations for the defaults (panotools, OpenHMD, vive).
Also adds compute_distortion function for Vive distortion
There are differences between OpenHMD and Panotools values, main differences for now:
* psvr has 5 pano coefficients, ohmd has 3
* psvr uses viewport size and lens center in pixels for distortion calculation, ohmd in meter
* psvr uses different distortion scaling than ohmd
It was left in as a debug measure, but is more confusing than useful,
especially with northstar directly generating a mesh and vive with its own shader.
For now add only the depth formats mandated by OpenGL to maximize the
chances of the Vulkan driver supporting a reasonable set of usage flags
for the formats.
This adds some Android support in composition clients,
and fixes the breakage from 2 commits ago.
Thanks to Jakob for finding my error in an earlier version.
As before, on the service side the GPU index the compositor runs on can be selected with
* XRT_COMPOSITOR_FORCE_GPU_INDEX=INDEX1
By default xrGetVulkanGraphicsDevice() will suggest the same GPU the compositor runs on.
It is also possible to override the GPU index suggested to applications with
* XRT_COMPOSITOR_FORCE_CLIENT_GPU_INDEX=INDEX2
The reason this is both done on the service side is that if compositor and client run
on different GPUs, the swapchains use linear tiling instead of optimal tiling.
To make chosen GPUs comparable across the compositor's and the client's vulkan instance,
VkPhysicalDeviceIDProperties.deviceUUID is used.
This patch passes the offset and extent properties to the layer shader
by extending the uniform. The fragment shader stage now also receives
the transformation uniform, which contains a has_sub boolean to
distinguish if the properties are set, so between projection and quad layers.
To avoid color bleeding the subimage sampling happens on a global pixel
coordinates basis as ivec2 using the GLSL texelFetch function.
Projection layers will be sampled as before.
Many functions returned void and were assumed to always succeed, and some functions
returned only a bool to indicate vague success/failure.
Now that these functions get piped over IPC all of them have to be able to indicate
an IPC failure like for example an unreachable service.
With the xrt_result_t return type they now have the opportunity to report various
types of failures.
Instead of maintaining this chunk of code and build options, the Vulkan
loader can be used to load up validation. This has the advantage that no
layer name needs to be hard coded inside Monado, which was subject of
change recently.
Instead of using our own environment variable we can easily set the one
from the loader, e.g. `VK_INSTANCE_LAYERS=VK_LAYER_KHRONOS_validation`.
This fixes a regression introduced by the IPC and layer rendering
patchsets and passes the array_index to the correct places, so the
correct VkImageView is used instead of the one with index 0.
Make the renderer and IPC aware of multiple projection and quad layers
using the layer renderer.
Remove redundant code related to idle images and imported buffers and
command buffer rebuild, since we now always just display the layer
renderer's frame buffer.
Get view and projection properties from xrt_device.
Do wait on device idle when destroying the swap chain.
Don't wait on device idle multiple times in resize.
Validation on NVIDA and AMD is happy when only keeping
the two idles when resizing and destroying the swapchain.
Use the same counter for frame and command buffers, as their number is
equal, which is retrieved from swapchain.image_count.
Use the struct member num_buffers in favor of passing it as parameter.
Reuse some command buffer functions.
Create code file for common direct mode code.
Make common functions take generic parameters.
Use common code in randr and nvidia back ends.
Remove redunant includes.
In favor of own struct to determine best available mode.
The loop now also starts as one, as we do not need compare the 0th mode
it itself and returns if there is only one mode.
The code was already essentialy C, just needed to replace the pairs with
structs and the vector with an array, it was fixed length anyway.
The code feels more natural in C.
In order to save redundancy with initializing empty fields with NULL / 0
values and improve readability of the Vulkan code.
This patch also uses struct initializers where possible.
This essentially reverts 1eae45212e.
Since there currently is no Vulkan extension that takes XCB handles to
aqcuire the display Xlib interop needs to be used.
Before this patch Monado was opening connections for both APIs, which
introduced overhead.
Even though all XCB handles can be casted to Xlib ones, this cannot be
done with the main xcb_connection_t / Display. In it's design the
interop between both APIs can create a xcb_connection_t from a XCB
handle, but not the other way round. So in an interop case the Xlib
connection is the main one, since it's on a higher level.
More information on this can be found here:
https://xcb.freedesktop.org/MixingCalls/
Unfortunately the clean solution for this would be to specify a Vulkan
extension that takes XCB handles. This would make sense since Vulkan
is aware of XCB in other parts of the API as well. In Mesa the Xlib
structs will be ultimately casted to XCB.
According the XCB documentation the reply structs must be freed. I found
this when running a minimal example related to the direct mode code in
valgrind.
The variable should be set to the index in the enumeration of a modes according to VK_KHR_display.
Monado can print a list of available modes with their indices with the env var XRT_COMPOSITOR_PRINT_MODES=1.
