Virtual method set_config() can be called several times, and if the number of
profiles counter isn't reset the pool will reach an error state.
The purpose of number of profiles is to check the number of valid vulkan video
profiles (two in the case of transcoding use-case, for example) so it's local to
set_config() virtual method.
Part-of: <https://gitlab.freedesktop.org/gstreamer/gstreamer/-/merge_requests/7247>
This is to avoid a regression in validation layer (introduced by commit
916c4e70cd) when using vulkandownload
VUID-VkImageMemoryBarrier2-srcAccessMask-03914 .. vkCmdPipelineBarrier2():
pDependencyInfo->pImageMemoryBarriers[1].srcAccessMask (VK_ACCESS_TRANSFER_READ_BIT)
is not supported by stage mask (VK_PIPELINE_STAGE_2_VIDEO_DECODE_BIT_KHR)
since vulkandownload set DPB memories' access mask to
VK_ACCESS_TRANSFER_READ_BIT, while they are retain by the DPB queue, so when
they are used as DPB after been shown, this validation error is raised.
Must of the barrier values are set ignoring the previous state of the vulkan
images.
Part-of: <https://gitlab.freedesktop.org/gstreamer/gstreamer/-/merge_requests/7211>
The access flags are kept around the operations, but when the buffer is
released, the access flag should be reset to its original value, since queue
transfers can be done along the pipeline and, when reusing the buffer, the new
queue might not support the latest access flag.
Part-of: <https://gitlab.freedesktop.org/gstreamer/gstreamer/-/merge_requests/7165>
Instead of dragging the last destination pipeline stage as current barrier
source pipeline stage (which isn't a valid semantic) this patch adds a parameter
to gst_vulkan_operation_add_frame_barrier() to set the source pipeline stage to
define the barrier.
The previous logic brought problems particularly with queue transfers, when the
new queue doesn't support the stage set during a previous operation in a
different queue.
Now the operation API is closer to Vulkan semantics.
Part-of: <https://gitlab.freedesktop.org/gstreamer/gstreamer/-/merge_requests/7165>
null event NT handle to ID3D12Fence::SetEventOnCompletion()
will block the calling CPU thread already, thus it has no point that
creating an event NT handle in order to immediate wait for fence at CPU-side.
Note that passing a valid event NT handle to the fence API might be useful
when we need to wait for the fence value later (or timeout is required),
or want to wait for multiple fences at once via WaitForMultipleObjects().
But it's not a considered use case for now.
Part-of: <https://gitlab.freedesktop.org/gstreamer/gstreamer/-/merge_requests/7176>
Define a default usage and use it instead of repeating the same bitwise
addition.
Therefore, when usage is defined as zero, the usage is defined with the
format's supported usage and the default usage, now without the storage
bit, but with color and input attachment bits.
Part-of: <https://gitlab.freedesktop.org/gstreamer/gstreamer/-/merge_requests/6798>
With clang on macOS:
```
error: format specifies type 'long' but the argument has type 'uint64_t' (aka 'unsigned long long')
...
error: format specifies type 'unsigned long' but the argument has type 'VkImageView' (aka 'unsigned long long')
```
Part-of: <https://gitlab.freedesktop.org/gstreamer/gstreamer/-/merge_requests/7091>
When building for iOS in Cerbero, as of MoltenVK SDK 1.3.283, we have
to statically link to libMoltenVK since it no longer ships a dylib.
This requires linking to libc++, so we find the dep with the objc++
compiler to ensure that meson uses the right linker.
Part-of: <https://gitlab.freedesktop.org/gstreamer/gstreamer/-/merge_requests/7091>
Schedule (semi-)static resource upload at converter creation time.
And use single resource for all vertex, index, and constant
buffers, since separate resources will waste GPU memory.
Note that size and address of a committed resource are 64K aligned
even if requested buffer size is small.
Part-of: <https://gitlab.freedesktop.org/gstreamer/gstreamer/-/merge_requests/7081>
As NVIDIA Amperium. In this case the each output buffer is also a DPB,
but using a different view layer.
Still pending a validation layer issue:
VUID-VkVideoBeginCodingInfoKHR-flags-07244
Co-authored-by: Victor Jaquez <vjaquez@igalia.com>
Part-of: <https://gitlab.freedesktop.org/gstreamer/gstreamer/-/merge_requests/6954>
query_result_status can be optional so we should not create
the query pool if the queue does not support it,
ie, AMD does not support VK_QUERY_TYPE_RESULT_STATUS_ONLY_KHR
In other use case such as VK_QUERY_TYPE_VIDEO_ENCODE_FEEDBACK_KHR, the
query pool must be created.
Part-of: <https://gitlab.freedesktop.org/gstreamer/gstreamer/-/merge_requests/7043>
As only queryResultStatusSupport can be optional,
the variable name should be more specific.
queryResultStatusSupport reports VK_TRUE if query type
VK_QUERY_TYPE_RESULT_STATUS_ONLY_KHR and use of
VK_QUERY_RESULT_WITH_STATUS_BIT_KHR are supported.
Part-of: <https://gitlab.freedesktop.org/gstreamer/gstreamer/-/merge_requests/7043>
According to the SPEC:
The frame id numbers (represented in display_frame_id, current_frame_id,
and RefFrameId[ i ]) are not needed by the decoding process, but allow
decoders to spot when frames have been missed and take an appropriate action.
So we should just print out warning and should not return error in parser when
mismatching. The decoder itself is already robust to handle the reference missing.
Fixes#3622
Part-of: <https://gitlab.freedesktop.org/gstreamer/gstreamer/-/merge_requests/7047>
A large refactoring commit for adding features and improve performance
* Reuse internal converter and overlay compositor:
Converter can be reused as long as input and display formats are not
changed. Also overlay compositor reconstruction is required only if
display format is changed
* Don't wait for full GPU flush on resize or close:
D3D12 swapchain requires GPU idle in order to resize backbuffer.
Thus CPU side waiting is required for swapchain related commands
to be finished. However, don't need to wait for full GPU flushing.
* Support multiple sink on a single external window
Keep installed subclass window procedure even if there's no associated
our internal HWND. This will make window procedure hooking less racy.
Then parent HWND's message will be transferred to our internal HWNDs
if needed.
* Adding support for window handle update
Application can change target HWND even when videosink is playing or
paused state. So, users can call gst_video_overlay_set_window_handle()
against d3d12videosink anytime. The videosink will be able to update
internal state and setup resource upon requested.
Part-of: <https://gitlab.freedesktop.org/gstreamer/gstreamer/-/merge_requests/7013>
Observed Intel GPU driver crash when multiple decoders are
configured in a process. It might be because of frequent
command queue alloc/free or too many in-flight decoding commands.
In order to make command queue persistent and limit the number of
in-flight command lists, holds global decoding command queue.
Part-of: <https://gitlab.freedesktop.org/gstreamer/gstreamer/-/merge_requests/7019>
Not all GPUs can support arbitrary offset of
D3D12_PLACED_SUBRESOURCE_FOOTPRINT when copying GPU memory between
texture and buffer. Instead of calculating size/offset per plane,
calculate the entire size and offsets at once.
Part-of: <https://gitlab.freedesktop.org/gstreamer/gstreamer/-/merge_requests/6967>
drm_fourcc.h should be picked up via the pkgconfig include, not the
system includedir directly.
Also consolidate the libdrm usage in va and msdk.
All this allows it to be picked up consistently (via the subproject,
for example).
Part-of: <https://gitlab.freedesktop.org/gstreamer/gstreamer/-/merge_requests/6932>
From the spec (chapter 34, v1.3.283):
````
UNORM: the components are unsigned normalized values in the range [0, 1]
SRGB: the R, G and B components are unsigned normalized value that represent
values using sRGB nonlinear encoding, while the A component (if one
exists) is a regular unsigned normalized value
```
The difference is the storage encoding, the first one is aimed for image
transfers, while the second is for shaders, mostly in the swapchain stage in the
pipeline, and it's done automatically if needed [1].
As far as I have checked, other frameworks (FFmpeg, GTK+), when import or export
images from/to Vulkan, use exclusively UNORM formats, while SRGB formats are
ignored.
My conclusion is that Vulkan formats are related on how bits are stored in
memory rather their transfer functions (colorimetry).
This patch does two interrelated changes:
1. It swaps certain color format maps to try first, in both
gst_vulkan_format_from_video_info() and gst_vulkan_format_from_video_info_2(),
the UNORM formats, when comparing its usage, and later check for SRGB.
2. It removes the code that check for colorimetry in
gst_vulkan_format_from_video_info_2(), since it not storage related.
1. https://community.khronos.org/t/noob-difference-between-unorm-and-srgb/106132/7
Part-of: <https://gitlab.freedesktop.org/gstreamer/gstreamer/-/merge_requests/6797>
Doing it in gst_play_new() means that bindings that directly call
g_object_new() with the GType wouldn't end up initializing both.
This affects at least the Python and GJS bindings.
gst_init() is nonetheless only called from gst_play_new() once because
calling it from class_init would likely lead to problems as that's
called from somewhere in the middle of GObject.
Part-of: <https://gitlab.freedesktop.org/gstreamer/gstreamer/-/merge_requests/6801>
Use of VK_PIPELINE_STAGE_2_ALL_COMMANDS_BIT instead of
specific VK_PIPELINE_STAGE_2_VIDEO_DECODE_BIT_KHR
Fix for VUID-vkCmdPipelineBarrier2-srcStageMask-03849
pDependencyInfo->pImageMemoryBarriers[0].srcStageMask
(VK_PIPELINE_STAGE_2_VIDEO_DECODE_BIT_KHR) is not compatible with
the queue family properties
(VK_QUEUE_GRAPHICS_BIT|VK_QUEUE_COMPUTE_BIT|VK_QUEUE_TRANSFER_BIT|
VK_QUEUE_SPARSE_BINDING_BIT|VK_QUEUE_PROTECTED_BIT) of this
command buffer. The Vulkan spec states: The srcStageMask member
of any element of the pMemoryBarriers, pBufferMemoryBarriers, or
pImageMemoryBarriers members of pDependencyInfo must only
include pipeline stages valid for the queue family that was
used to create the command pool that commandBuffer was allocated
from (
https://www.khronos.org/registry/vulkan/specs/1.3-extensions/
html/vkspec.html#VUID-vkCmdPipelineBarrier2-srcStageMask-03849)
The frame barrier should use a compatible srcStageMask for all
the queues.
Remove reset_pipeline_stage_mask as it is redundant
Part-of: <https://gitlab.freedesktop.org/gstreamer/gstreamer/-/merge_requests/6780>
Ideally, GPU waiting should be scheduled just before executing command list.
But handling the case outside of converter is a bit complicated.
Under an assumption that constructed command list will be executed
immediately, schedules GPU-side waiting inside of conversion method
to simplify the flow.
Part-of: <https://gitlab.freedesktop.org/gstreamer/gstreamer/-/merge_requests/6749>
Since we don't want to expose video decoding API outside of GStreamer, the
header is removed from installation and both source files are renamed as
-private.
The header must remain in gst-libs because is referred by GstVulkanQueue,
which's the decoder factory.
Part-of: <https://gitlab.freedesktop.org/gstreamer/gstreamer/-/merge_requests/6723>
First it derived mapping was disabled for P010 formats, but also there's an
issue with interlaced frames.
It would be possible to disable derived mapping only for interlaced (H.264
decoder and vadeinterlace) but it would spread the hacks along the code. It's
simpler and contained to disable derived completely for Mesa <23.3
Fixes: #3450
Part-of: <https://gitlab.freedesktop.org/gstreamer/gstreamer/-/merge_requests/6729>
Instead of duplicating the GStreamer format to DRM fourcc mapping, this patch
uses the GstVideo library helpers. This duplicates the big O of looking for,
since the two lists are traversed, but it's less error prone.
Partially reverts commit 547f3e8622.
Fixes: #3354
Part-of: <https://gitlab.freedesktop.org/gstreamer/gstreamer/-/merge_requests/6731>
If waylandsink received buffer rate is high which causes frame
drop, the cached staged buffer will be replaced when next buffer
needs to be rendered and be freed after redraw. But there is
chance to get memory leak if ended without redraw. So need to
free staged buffer when do gst_wl_window_finalize().
Part-of: <https://gitlab.freedesktop.org/gstreamer/gstreamer/-/merge_requests/6670>
In addition to device removed status monitoring in gst_d3d11_result()
method, if ID3D11Device4 interface is available,
an event handle will be used for device removed status update.
And "device-removed" signal is removed since applications can monitor
the device removed status via gobject notify
Part-of: <https://gitlab.freedesktop.org/gstreamer/gstreamer/-/merge_requests/6699>
Adding new property in order to notify users of device removed status.
Once device removed status is detected, application should release
all ID3D12Device objects corresponding to the adapter, including
GstD3D12Device object. Otherwise D3D12CreateDevice() call for the
adapter will fail.
Part-of: <https://gitlab.freedesktop.org/gstreamer/gstreamer/-/merge_requests/6699>
It seems that when D3D11CreateDevice collides in time
with other D3D11 calls, in particular the proccess of
creating a shader, it can corrupt the memory in the driver.
D3D11 spec doesn't seem to require any thread safety from
D3D11CreateDevice. Following MSDN, it is supposed to be called
in the beginning of the proccess, while GStreamer calls it with each
new pipeline.
Such crashes in the driver were frequently reproducing on the
Intel UHD 630 machine.
Part-of: <https://gitlab.freedesktop.org/gstreamer/gstreamer/-/merge_requests/6686>
Fix
gstanalyticsmeta.c:134: Warning: GstAnalytics: "@instance"
parameter unexpected at this location
warning (caused by the extraneous empty line in the doc chunk)
and align function arguments with documentation and header file
(handle -> instance).
Part-of: <https://gitlab.freedesktop.org/gstreamer/gstreamer/-/merge_requests/6585>
