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>
Adding a property to control error reporting behavior when output
window is closed in playing or paused state. This can be useful
for apps where an app wants to close window even if it's playing
a stream, and the closed window is expected.
Part-of: <https://gitlab.freedesktop.org/gstreamer/gstreamer/-/merge_requests/6939>
Conceptually identical to the present signal of d3d11videosink.
This signal will be emitted with current render target
(i.e., swapchain backbuffer) and command queue. Signal handler
can record GPU commands for an overlay image or to blend
an image to the render target.
In addition to d3d12 resources, videosink will send
d3d11 and d2d resources depending on "overlay-mode"
property, so that signal handler can render by using
preferred/required DirectX API.
Part-of: <https://gitlab.freedesktop.org/gstreamer/gstreamer/-/merge_requests/6838>
The idea of using separate command queue per videosink was that
swapchain is bound to a command queue and we need to flush the
command queue when window size is changed. But the separate
queue does not seem to improve performance a lot.
Part-of: <https://gitlab.freedesktop.org/gstreamer/gstreamer/-/merge_requests/6838>
Address below message reported by SDK debug layer.
ID3D12Device::CheckFeatureSupport: Unsupported Decode Profile Specified.
Use ID3D12VideoDevice::CheckFeatureSupport with D3D12_FEATURE_VIDEO_DECODE_PROFILES
to retrieve a list of supported profiles
Part-of: <https://gitlab.freedesktop.org/gstreamer/gstreamer/-/merge_requests/6839>
Fixing deadlock in below case
* GC lock is taken by background thread, and the background thread calls
gst_d3d12_ipc_client_release_imported_data() which takes ipc lock
* ipc lock is already taken in ipc thread and trying to pushing GC data
via gst_d3d12_command_queue_set_notify()
* gst_d3d12_command_queue_set_notify() is trying to take GC lock
but it's already taken by background thread
Part-of: <https://gitlab.freedesktop.org/gstreamer/gstreamer/-/merge_requests/6749>
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>
A DPB buffer held by codec picture object may not be writable
at the moment, then gst_buffer_make_writable() will unref passed buffer.
Specifically, the use after free or double free can happen if:
* Crop meta of buffer copy is required because of non-zero
top-left crop position
* zero-copy is possible with crop meta
* A picture was duplicated, interlaced h264 stream for example
Interlaced h264 stream with non-zero top-left crop position
is not very common but it's possible configuration in theory.
Thus gst_buffer_make_writable() should be called with
GstVideoCodecFrame.output_buffer directly.
Part-of: <https://gitlab.freedesktop.org/gstreamer/gstreamer/-/merge_requests/6706>
Because ID3D12Device objects are singletons per adapter,
GstD3D12Device was following the API design, that is, keep track
of global GstD3D12Device objects and reuses it.
That means ID3D12Device object can be released at the time
when GstD3D12Device is destroyed.
But exetrnal APIs such as NVENC does not seem to be happy
with the released ID3D12Device, that could be a driver bug though.
Let's hold already opened ID3D12Device permanently without releasing
it for now.
Part-of: <https://gitlab.freedesktop.org/gstreamer/gstreamer/-/merge_requests/6395>
Do not chain up to parent's GstBufferPool::start() which will do
preallocation. We don't want it to be preallocated
since there are various cases where negotiated downstream buffer pool is
not used at all (e.g., zero-copy decoding, IPC elements).
Part-of: <https://gitlab.freedesktop.org/gstreamer/gstreamer/-/merge_requests/6326>
Gets being released memory back to queue even if allocator is flushing
in order to count the number of outstanding memory objects.
Also, clear queue if there's no outstanding memory object and
allocator is flushing
Part-of: <https://gitlab.freedesktop.org/gstreamer/gstreamer/-/merge_requests/6240>
According to recommendation from MS, IDXGIOutputDuplication::ReleaseFrame()
needs to be called just before IDXGIOutputDuplication::AcquireNextFrame()
for performance reasons, so that driver can accumulate dirty rects
and update texture at once. But it seems to cause choppy output.
Do release acquired frame immediately once processing done,
like d3d11 implementation does.
Part-of: <https://gitlab.freedesktop.org/gstreamer/gstreamer/-/merge_requests/6092>
Fence data could hold GstD3D12Device directly or indirectly.
Then if it's holding last refcount, the device object will
be released from the device object's internal thread,
and will try join self thread.
Delegates it to other global background thread to avoid
self thread joining.
Part-of: <https://gitlab.freedesktop.org/gstreamer/gstreamer/-/merge_requests/6042>
Fixing below debug layer report
ID3D12Device::CreateCommittedResource: Ignoring InitialState D3D12_RESOURCE_STATE_COPY_DEST.
Buffers are effectively created in state D3D12_RESOURCE_STATE_COMMON.
Buffer resource will be automatically promoted to D3D12_RESOURCE_STATE_COPY_DEST
at the very first COPY operation time.
Part-of: <https://gitlab.freedesktop.org/gstreamer/gstreamer/-/merge_requests/5895>
Since DXGI desktop duplication API does not work with Direct3D12 device,
this element will use Direct3D11 device to acquire frame.
Then other rendering operations (e.g., texture copy, render pipeline) will
happen using Direct3D12 API
Part-of: <https://gitlab.freedesktop.org/gstreamer/gstreamer/-/merge_requests/5883>
In case of tier 1 decoder, always use reference-only picture to avoid
fixed-size pool limitation and output decoded picture without
copy even for negative rate. Also do not use copy queue for GPU to GPU
copy. Copy queue is specialized for upload/download and may occupy
PCIE bandwidth. Use direct queue as recommended by vendors.
Part-of: <https://gitlab.freedesktop.org/gstreamer/gstreamer/-/merge_requests/5877>
On flush event, baseclass will discard all pictures from DPB
but there can be still in-flight commands not finished yet.
Use our command queue, allocator and fence data helper objects
to keep resource available during command execution.
Part-of: <https://gitlab.freedesktop.org/gstreamer/gstreamer/-/merge_requests/5877>
Shader visible descriptors occupy GPU resource and there are hardware
limits. Thus, in order to minimize the amount of shader visible heaps,
only non shader visible descriptor heap (staging) will be held by d3d12memory.
Then converter will copy the staging descriptor to shader visible
descriptor heap per draw.
Part-of: <https://gitlab.freedesktop.org/gstreamer/gstreamer/-/merge_requests/5875>
Zero initialization would have overhead and it's not required
most cases except for textures. Use CREATE_NOT_ZEROED flag
in case of buffer resource or if a texture will be rendered without any
prior read operation.
Part-of: <https://gitlab.freedesktop.org/gstreamer/gstreamer/-/merge_requests/5875>
Conversion will happen when constructed command list is executed,
not by converter element. Thus this object should not map output buffer
with write flag which will result in error if multiple threads
are building commands for the same output target frame.
Part-of: <https://gitlab.freedesktop.org/gstreamer/gstreamer/-/merge_requests/5875>
