Introducing CUDA buffer pool with generic CUDA memory support.
Likewise GL memory, any elements which are able to access CUDA device
memory directly can map this CUDA memory without upload/download
overhead via the "GST_MAP_CUDA" map flag.
Also usual GstMemory map/unmap is also possible with internal staging memory.
For staging, CUDA Host allocated memory is used (see CuMemAllocHost API).
The memory is allowing system access but has lower overhead
during GPU upload/download than normal system memory.
Part-of: <https://gitlab.freedesktop.org/gstreamer/gst-plugins-bad/-/merge_requests/1633>
Fix a bug in _create_other_pool(). The old way of checking the
base->other_pool make that other_pool never be changed until the
gst_va_base_dec_stop() to stop the current decoding context.
But in some stream, the resolution may change during the decoding
process, and we need to re-negotiate the buffer pool. Then, the
old other_pool can not be clean correctly and the new correct one
can not be created.
Part-of: <https://gitlab.freedesktop.org/gstreamer/gst-plugins-bad/-/merge_requests/1692>
The current bufferpool wastes all pre-allocate buffers when the
buffer pool is actived.
The pool->priv->size is 0 for va buffer pool. And every time, the
reset_buffer() will clean all mem and make the buffer size 0, that
can cache the gst_buffer in the buffer pool.
But when the buffer pool is activing, the default_start() just
allocate the buffer and release_buffer() immediately, all the pre
allocated buffers and surfaces are destroyed because of
gst_buffer_get_size (buffer) != pool->priv->size.
We need to use release_buffer() to do the clean job at the pool
start time.
Part-of: <https://gitlab.freedesktop.org/gstreamer/gst-plugins-bad/-/merge_requests/1686>
Since allocators keep an available memory queue to reuse, video format and usage
hint are now persistant while allocator's memories are around.
This patch adds _set_format() and _get_format() for both VA allocators.
_set_format() validates if given format can be used or reused. If no allocated
surface previously it creates a dummy one to fetch its offsets and
strides. Updated info is returned to callee.
GstVaPool uses _set_format() at config to verify the allocator capacity and to
get the surfaces offsets and strides, which are going to be used by the video
meta.
Allocator extracted caps are compared with caps from config and if they have
different strides or offsets, force_videometa is set.
A new bufferpool method gst_va_pool_requires_video_meta() is added return the
value of force_videometa. This value is checked in order to know if decoders
need to copy the surface if downstream doesn't announce video meta support.
Part-of: <https://gitlab.freedesktop.org/gstreamer/gst-plugins-bad/-/merge_requests/1667>
Without preallocating buffers and memories a deadlock in pool allocator is
highly probably since it might hit the case were buffer is returned to the pool
but their memories are still hold by a copy downstream, without other
preallocated buffers available.
This kind of a hack, where buffer_reset() follow the normal path if it's called
from start().
Part-of: <https://gitlab.freedesktop.org/gstreamer/gst-plugins-bad/-/merge_requests/1667>
Every time a new surface is created the counter increases by one, and when it is
destroyed (or will be destroyed in case of GstVaAllocator), the counter is
decreased. Then, at allocator dispose, it is warning if the counter is not zero.
This counter will be also used to check if the allocator can change its
configuration if the counter is zero or can not.
Part-of: <https://gitlab.freedesktop.org/gstreamer/gst-plugins-bad/-/merge_requests/1667>
Updated Decklink SDK to version 11.2 in order to support newer cards like the Decklink 8K Pro.
This required to replace the duplex property by a profile property.
Profile values can be the following:
- bmdProfileOneSubDeviceFullDuplex
- bmdProfileOneSubDeviceHalfDuplex
- bmdProfileTwoSubDevicesFullDuplex
- bmdProfileTwoSubDevicesHalfDuplex
- bmdProfileFourSubDevicesHalfDuplex
Fixes#987
Part-of: <https://gitlab.freedesktop.org/gstreamer/gst-plugins-bad/-/merge_requests/1665>
1. Allocators don't implement memory free() methods since all the memories will
implement dispose() returning FALSE
2. Memory/miniobject dispose() will act as memory release, enqueueing the
release memory
3. A new allocator's method prepare_buffer() which queries the released memory
queue and will add the requiered memories to the buffer.
4. Allocators added a GCond to synchronize dispose() and prepare_buffer()
5. A new allocator's method flush() which will free for real the memories.
While the bufferpool will
1. Remove all the memories at reset_buffer()
2. Implement acquire_buffer() calling allocator's prepare_buffer()
3. Implement flush_start() calling allocator's flush()
4. start() is disabled since it pre-allocs buffers but also calls
our reset_buffer() which will drop the memories and later the
buffers are ditched, something we don't want. This approach avoids
buffer pre-allocation.
Part-of: <https://gitlab.freedesktop.org/gstreamer/gst-plugins-bad/-/merge_requests/1626>
Renamed the first variable member of GstVaMemory from parent to mem in
order to avoid confusion with GstMemory's parent.
When freeing the structure, memory's parent is check in order to
decide if surfaces has to be destroyed or not, since only the parent
class have to destroy it.
Removed GST_MEMORY_FLAG_NO_SHARE in memory initialization, since it is
deprecated.
Implemented allocator's share virtual method which creates a new
shallow GstVaMemory structure based on the passed one which will be
it's parent.
Part-of: <https://gitlab.freedesktop.org/gstreamer/gst-plugins-bad/-/merge_requests/1626>
Staging texture is used for memory transfer between system and
gpu memory. Apart from d3d11{upload,download} elements, however,
it should happen very rarely.
Before this commit, d3d11bufferpool was allocating at least one
staging texture in order to calculate cpu accessible memory size,
and it wasn't freed for later use of the texture unconditionally.
But it will increase system memory usage. Although GstD3D11memory
object is implemented so that support CPU access, most memory
transfer will happen in d3d11{upload,download} elements.
By this commit, the initial staging texture will be freed immediately
once cpu accessible memory size is calculated.
Part-of: <https://gitlab.freedesktop.org/gstreamer/gst-plugins-bad/-/merge_requests/1627>
