When importing dmabuf from downstream, we want the allocator to be in
OTHER_POOL mode despite output_mode being DMABUF.
So check first if other_pool is set before checking for pool's
output_mode.
One big restriction of the OMX buffer pool has always been
that the GstMemory objects were flagged with NO_SHARE.
This was because the buffer pool needed to be sure that when
a buffer returned to the pool, it would be safe to release the
OMX buffer back to OpenMAX.
With this change, this is no longer a restriction. What this
commit introduces is a new allocator that allows us to track
the GstMemory objects independently. Now, when a buffer returns
to the pool, it is not necessary for the memory to be released
as well. We simply track the memory's ref count in the allocator
and we return the OMX buffer back when the memory's ref count
drops to 0.
The reason for doing this is to allow implementing zero-copy
transfers in situations where we may need to copy or map a
certain region of the buffer. For instance, omxh264enc ! h264parse
should be possible to be zero-copy by using an OMX buffer pool
between them.
gst_memory_map() is already adding the offset to the mapped pointer.
Doing it in the memory implementation was resulting in the offset being
accounted twice.
It doesn't matter yet as we are only creating memory without offset for
now but it will once we'll start sharing OMX memories.
This was the single place where this category was used in gst-omx so
most users, including me, are generally not turning it and were missing this
important information from logs.
The copying code uses gst_video_frame_copy() which is already logging
with CAT_PERFORMANCE so we can still have this information when using
only this debug category.
If buffers were released from the pool while
gst_omx_video_enc_handle_frame() was waiting for new buffers,
gst_omx_port_acquire_buffer() was never awaken as the buffers weren't
released through OMX's messaging system.
GQueue isn't thread safe so also protect it with the lock mutex.
We used to track the 'allocating' status on the pool. It is used while
allocating so output buffers aren't passed right away to OMX and input
ones are not re-added to the pending queue.
This was causing a bug when exporting buffers to v4l2src. On start
v4l2src acquires a buffer, read its stride and release it right away.
As no buffer was received by the encoder element at this point, 'allocating'
was still on TRUE and so the the buffer wasn't put back to the pending
queue and, as result, no longer available to the pool.
Fix this by checking the active status of the pool instead of manually
tracking it down. The pool is considered as active at the very end of
the activation process so we're good when buffers are released during
the activation.
Propose pool upstream so input buffers can be allocated by the port and
exported as dmabuf.
The actual OMX buffers are allocated when the pool is activated, so we
don't end up doing useless allocations if the pool isn't used.
https://bugzilla.gnome.org/show_bug.cgi?id=796918
Now that the pool is responsible of freeing the OMX buffers, we need to
ensure that the OMX component stay alive while the pool is as we rely on
the component to free the buffers.
The GstOMXPort is owned by the component so no need to ref this one.
https://bugzilla.gnome.org/show_bug.cgi?id=796918
The pool is stopped when all the buffers have been released. Deallocate
when stopping so we are sure that the buffers aren't still used by
another element.
https://bugzilla.gnome.org/show_bug.cgi?id=796918
I spent quiet some time figuring out why performance of my pipeline were
terrible. Turned out it was because of output frames being copied
because of stride/offset mismatch.
Add a PERFORMANCE DEBUG message to make it easier to spot and debug from logs.
https://bugzilla.gnome.org/show_bug.cgi?id=793637
The zynqultrascaleplus OMX implementation has a custom extension
allowing decoders to output dmabuf and so avoid buffers copy between OMX
and GStreamer.
Make use of this extension when built on the zynqultrascaleplus. The
buffer pool code should be re-usable for other platforms as well.
https://bugzilla.gnome.org/show_bug.cgi?id=784847
We have to return the buffers back to the pool in when stopping to
not mess with the GstBufferPool accounting.
The OMX buffers will be freed when those won't be in charge of the
pool in the chained up call to 'stop'.
Fixes segfaults on finalize and pool not being properly deactivated.
https://bugzilla.gnome.org/show_bug.cgi?id=726337
When using GstOMXBufferPool on an output port, it internally uses
a GPtrArray to manage the GstBuffers instead of the default queue
from the GstBufferPool base class.
In this case GstBufferPool::default_free_buffer is not called when
the pool is stopped. Because the queue is empty. So explicitely
call gst_omx_buffer_pool_free_buffer on each buffer contained in
the GPtrArray.
https://bugzilla.gnome.org/show_bug.cgi?id=726337
