With the NTT code path in the guest we might end up with
the generic array outputs also with GS and TES.
Signed-off-by: Gert Wollny <gert.wollny@collabora.com>
Reviewed-by: Corentin Noël <corentin.noel@collabora.com>
Make sure that the condition for which we are emitting the uniform
corresponds to the one in the iter_declaration function.
Signed-off-by: Corentin Noël <corentin.noel@collabora.com>
Reviewed-by: Gert Wollny <gert.wollny@collabora.com>
Allows virglrenderer-venus to passthrough the VK_EXT_4444_formats
extension to the venus client.
Signed-off-by: Igor Torrente <igor.torrente@collabora.com>
When we first convert a tgsi shader into TGSL, we fill the shader key
with a value of `(gs|tcs|tes)_present` based on the currently bound
shaders. But since a shader is always going to be bound if it's being
used, we should already assume that it is going to be present in the
shader key, saving a recompilation.
Signed-off-by: Italo Nicola <italonicola@collabora.com>
Reviewed-by: Corentin Noël <corentin.noel@collabora.com>
Reviewed-by: Gert Wollny <gert.wollny@collabora.com>
It is not valid to link a program that has a TCS but no TES, therefore
we shouldn't attempt to pre-link this combination of shaders.
Signed-off-by: Italo Nicola <italonicola@collabora.com>
Reviewed-by: Corentin Noël <corentin.noel@collabora.com>
Reviewed-by: Gert Wollny <gert.wollny@collabora.com>
Currently, we always use the `tfout` variables when writing TF output
for geometry shaders, but only declare them under specific conditions,
causing GLSL errors in tests like ``
Fixes: 9157dcbca0
Signed-off-by: Italo Nicola <italonicola@collabora.com>
Reviewed-by: Corentin Noël <corentin.noel@collabora.com>
Reviewed-by: Gert Wollny <gert.wollny@collabora.com>
It's better done in proxy than in vkr.
Signed-off-by: Yiwei Zhang <zzyiwei@chromium.org>
Reviewed-by: Chia-I Wu <olvaffe@gmail.com>
Reviewed-by: Ryan Neph <ryanneph@google.com>
This is required for support of untyped resources on virtio
devices which create their own egl context.
Currently untyped resources support is reported to mesa only in case
virglrenderer creates it's own egl context from scratch. The problem
that this context is not "sharable" with device context and if it's
created vrend uses it instead of context callbacks implemented by device.
Thus such approach is not suitable for devices with it's own egl context.
The only thing that is required for untyped resources support is
initialized eglDisplay handle that is requried for eglCreateImageKHR,
thus add callback for device to share it.
Untyped resources feature support is required for sharing vkr resources
with vrend.
Signed-off-by: Oleksandr.Gabrylchuk <Oleksandr.Gabrylchuk@opensynergy.com>
Signed-off-by: Andrii Pauk <Andrii.Pauk@opensynergy.com>
Reviewed-by: Chia-I Wu <olvaffe@gmail.com>
This is mostly a naming change, along with more docs
Signed-off-by: Yiwei Zhang <zzyiwei@chromium.org>
Reviewed-by: Chia-I Wu <olvaffe@gmail.com>
Reviewed-by: Ryan Neph <ryanneph@google.com>
The ring cmd can come earlier than render server receiving the shmem
after removing redundant guest side roundtrips.
Signed-off-by: Yiwei Zhang <zzyiwei@chromium.org>
Reviewed-by: Chia-I Wu <olvaffe@gmail.com>
Reviewed-by: Ryan Neph <ryanneph@google.com>
server needs to persist the resource and attach to workaround a guest
kernel out-of-order map and attach cmds.
proxy needs to track and de-dup the attach request for the attached
resources.
Signed-off-by: Yiwei Zhang <zzyiwei@chromium.org>
Reviewed-by: Chia-I Wu <olvaffe@gmail.com>
Reviewed-by: Ryan Neph <ryanneph@google.com>
The same res_id is also used for later resource attach of the same.
Signed-off-by: Yiwei Zhang <zzyiwei@chromium.org>
Reviewed-by: Chia-I Wu <olvaffe@gmail.com>
Reviewed-by: Ryan Neph <ryanneph@google.com>
It will be used by the proxy context.
Signed-off-by: Yiwei Zhang <zzyiwei@chromium.org>
Reviewed-by: Chia-I Wu <olvaffe@gmail.com>
Reviewed-by: Ryan Neph <ryanneph@google.com>
This is dedicated heap memory allocations approach.
This type of blob should be used in the following way:
1) Guest virtio-gpu driver on start checks for dedicated memory region.
2) On create_blob drm ioctl guest driver reserves chunk of required memory and
send it to host using sg list. Heap is managed on guest.
3) Device creates dmabuf fd from this sg entry and sends it to virglrenderer on
create_blob virtio-gpu command. Blob is created using
virgl_renderer_resource_import_blob call.
4) On next vkAllocateMemory call from mesa, virglrenderer will allocate vk memory
handle from this dmabuf fd. It will receive resource id in
VK_STRUCTURE_TYPE_IMPORT_MEMORY_RESOURCE_INFO_MESA vkAllocateMemory structure.
The flow is opposite to the way it's done for HOST_3D types of blob, where
vkAllocateMemory is called first, vk memory is allocated from random host place.
Then create_blob is called, where dmabuf fd is exported from VkDeviceMemory.
Signed-off-by: Andrii Pauk <Andrii.Pauk@opensynergy.com>
Reviewed-by: Chia-I Wu <olvaffe@gmail.com>
Improves the error handling code that deals with resource
deallocation.
Reviewed-by: Chia-I Wu <olvaffe@gmail.com>
Signed-off-by: Igor Torrente <igor.torrente@collabora.com>
This only updates venus-protocol. There is no visible functional
difference.
Reviewed-by: Yiwei Zhang <zzyiwei@chromium.org>
Reviewed-by: Ryan Neph <ryanneph@google.com>
vkr_extension_table allows us to fine-control which extensions are
advertised.
To support an extension, we need venus-protocol support and vkr support.
One is generated and one is open-coded. We don't want to advertise an
extension automatically whenever an updated venus-protocol includes it.
Reviewed-by: Yiwei Zhang <zzyiwei@chromium.org>
Reviewed-by: Ryan Neph <ryanneph@google.com>
Temporarily disable the old CI configuration file and replace it with
the one belonging to the reworked CI infrastructure.
The project CI/CD settings should be updated in order to switch
'CI/CD configuration file' from 'ci/.gitlab-ci.yml' to the default
'.gitlab-ci.yml', for consistency with Mesa CI.
Note the 'ci/' folder becomes obsolete in favor of '.gitlab-ci/' and
could be removed once all needed functionality has been migrated to the
new setup.
Signed-off-by: Cristian Ciocaltea <cristian.ciocaltea@collabora.com>
Reviewed-by: Tomeu Vizoso <tomeu.vizoso@collabora.com>
Reviewed-by: Rohan Garg <rohan.garg@intel.com>
Add initial support for reusing the Mesa CI infrastructure.
Note this doesn't conflict with the old CI since it doesn't share the
'ci/' folder, but instead uses '.gitlab-ci/', which makes the new setup
consistent with Mesa CI.
Signed-off-by: Cristian Ciocaltea <cristian.ciocaltea@collabora.com>
Reviewed-by: Tomeu Vizoso <tomeu.vizoso@collabora.com>
Reviewed-by: Rohan Garg <rohan.garg@intel.com>
It is to avoid integer overflows and to catch bogus allocations (e.g.,
the guest driver encodes an uninitialized value).
Signed-off-by: Chia-I Wu <olvaffe@gmail.com>
Reviewed-by: Ryan Neph <ryanneph@google.com>
Reviewed-by: Yiwei Zhang <zzyiwei@chromium.org>
Using a pointer for fs_info in the shader key is not really that useful
because the contents of the structure can change even though the pointer
remains the same, and the pointer can be different when the contents are
the same.
Fixes 5f488ed00d
Signed-off-by: Italo Nicola <italonicola@collabora.com>
Tested-by: Cristian Ciocaltea <cristian.ciocaltea@collabora.com>
Reviewed-by: Gert Wollny <gert.wollny@collabora.com>
With GLES we always set the correct binding if dual source blend
is disabled, with OpenGL we still have to do this to handle outputs
that might be optimized away.
v2: Fix comment (Yiwei Zhang)
Signed-off-by: Gert Wollny <gert.wollny@collabora.com>
Reviewed-by: Yiwei Zhang <zzyiwei@chromium.org>
This is a fix for segfault issue, because of usage of uninitialized
vkQueue handle.
Spec [1] states that vkGetDeviceQueue must only be used to get queues
that were created with the flags parameter of VkDeviceQueueCreateInfo
set to zero. To get queues that were created with a non-zero flags
parameter use vkGetDeviceQueue2.
The problem that in case queue was created with flags set to zero following
vkGetDeviceQueue2 call will fail. By failing I mean that pQueue set to NULL,
which later leads to segfault.
This problem was also reported by vulkan validation layer:
vkr: vkGetDeviceQueue2: value of pQueueInfo->flags must not be 0. The Vulkan
spec states: flags must not be 0
(https://www.khronos.org/registry/vulkan/specs/1.1-extensions/html/vkspec.html#VUID-VkDeviceQueueInfo2-flags-requiredbitmask)
[1]: https://www.khronos.org/registry/vulkan/specs/1.2-extensions/man/html/vkGetDeviceQueue.html
Signed-off-by: Andrii Pauk <Andrii.Pauk@opensynergy.com>
Reviewed-by: Chia-I Wu <olvaffe@gmail.com>
Clean up the code a bit and only try to read the primitive type for
point mode if the prev shader really exists.
This fixes a VM crash when running the GLES 3.1 cts.
Signed-off-by: Gert Wollny <gert.wollny@collabora.com>
Reviewed-by: John Bates <jbates@chromium.org>
If we always emit the glip distance in GS we might end up
having too many outputs, which will result in failures.
Hence, for GS only emit the clip distance evaluation code
when the clip planes are enabled.
Fixes: 072f30955b
shader: Always write code to toggle clip plane
Closes: #254
Signed-off-by: Gert Wollny <gert.wollny@collabora.com>
Reviewed-by: Italo Nicola <italonicola@collabora.com>
If we receive a link command without a VS or FS, just early exit to
avoid doing extra work.
Signed-off-by: Italo Nicola <italonicola@collabora.com>
Reviewed-by: Gert Wollny <gert.wollny@collabora.com>
Now that render_worker_jail tracks workers, the apis are simplified such
that
- render_worker_create forks a subprocess
- render_worker_destroy kills a subprocess
- render_worker_jail_reap_workers reaps all exited subprocesses
Also add render_worker_jail_detach_workers that can be called by a
forked subprocess to "detach" (free without killing/reaping) workers.
v2: replace some checks for worker->{destroyed,reaped} by asserts (Ryan)
Signed-off-by: Chia-I Wu <olvaffe@gmail.com>
Reviewed-by: Yiwei Zhang <zzyiwei@chromium.org> (v1)
Reviewed-by: Ryan Neph <ryanneph@google.com> (v1)
It contains only struct minijail, but we plan to add more fields to it.
Signed-off-by: Chia-I Wu <olvaffe@gmail.com>
Reviewed-by: Yiwei Zhang <zzyiwei@chromium.org>
Reviewed-by: Ryan Neph <ryanneph@google.com>
Add render_client_detach_all_records to "detach" context records from
the child process before returning from render_client_create_context.
"Detaching" means destroying records without killing nor reaping the
workers.
This makes it clearer how context records are destroyed in subprocesses.
Signed-off-by: Chia-I Wu <olvaffe@gmail.com>
Reviewed-by: Yiwei Zhang <zzyiwei@chromium.org>
Reviewed-by: Ryan Neph <ryanneph@google.com>
Gert Wollny
Corentin Noël
Igor Torrente
Italo Nicola
Yiwei Zhang
Ryan Neph
Andrii Pauk
Omar Akkila
Chia-I Wu
Cristian Ciocaltea