Previously the frame map counter increased independently if the map succeeded
or not. This leaded to critical messages and crashes if the frame was unable
to be mapped, but the counter increased.
This patch increases the map counter only if the map operation occurred.
https://bugzilla.gnome.org/show_bug.cgi?id=773939
If the allocator was configured to use direct upload or rendering, the
generated derived image created at mapping needs to be destroyed after
unmapping, because, in order to process the surface, it should not be marked
as "busy" by the driver.
The way to experiment with the direct rendering is through and internal
compiler pre-processor flag.
The current change set enables a way to specified at run-time, as a flag
passed to the allocator at instanciation time.
In order to clarify the use of flag as input parameter, it is renamed to
surface_alloc_flag, since it is used when creating a VA surface with certain
properties.
https://bugzilla.gnome.org/show_bug.cgi?id=773497
When calling gst_video_frame_map() with GST_MAP_WRITE flag, it doesn't call
ensure_image_is_current(), which means it doesn't guarentee VAImage is valid
in this case.
https://bugzilla.gnome.org/show_bug.cgi?id=766978
Add the method gst_allocator_get_vaapi_image_size() for the
GstVaapiVideoAllocator, which gets the size of the allocated images with the
current video info.
This method replaces the direct call to the allocator's image info when the
pool is configured.
Originally vaapivideobufferpool instantiates its own allocator regardless the
received configuration, and it relies in custom configuration options to
choose which kind of allocator instantiate.
This patch transfers the responsibility of the allocator instantiate to
vaapipluginbase and pass it to the vaapivideobufferpool through its
configuration.
* gst/vaapi/gstvaapipluginbase.c
+ set_dmabuf_allocator(): inserts a dmabuf allocator in the bufferpool
+ ensure_sinkpad_buffer_pool(): set a normal vaapi video allocator in
bufferpool configuration
+ gst_vaapi_plugin_base_propose_allocation(): call set_dmabuf_allocator() if
needed.
+ gst_vaapi_plugin_base_decide_allocation(): set a normal vaapi video
allocator in bufferpool configuration
* gst/vaapi/gstvaapivideobufferpool.c
+ gst_vaapi_video_buffer_pool_set_config(): instead of instantiate the
allocator, process the received one through its configuration.
* gst/vaapi/gstvaapivideobufferpool.h: removed
GST_BUFFER_POOL_OPTION_DMABUF_MEMORY since it is not used anymore.
* gst/vaapi/gstvaapivideomemory.c
+ gst_vaapi_is_dmabuf_allocator(): new helper function to identify a dmabuf
allocator with the vaapi qdata.
https://bugzilla.gnome.org/show_bug.cgi?id=765435
Instead of a dummy alloc() vmethod, the allocator instance set the flag
GST_ALLOCATOR_FLAG_CUSTOM_ALLOC, which is used by the framework to avoid call
gst_allocator_alloc() on the allocator.
As gst-indent generated ugly code in these cases, this patch changes the used
idiomatic into other one.
No functional changes were introduced.
Signed-off-by: Víctor Manuel Jáquez Leal <victorx.jaquez@intel.com>
If USE_NATIVE_FORMATS is defined we bail out before configuring the surface
info based on the derived image configuration.
Signed-off-by: Víctor Manuel Jáquez Leal <victorx.jaquez@intel.com>
https://bugzilla.gnome.org/show_bug.cgi?id=744042
First added the function gst_vaapi_video_format_get_best_native(), which
returns the best native format that matches a particular chroma type:
YUV 4:2:0 -> NV12, YUV 4:2:2 -> YUY2, YUV 4:0:0 -> Y800
RGB32 chroma and encoded format map to NV12 too.
That format is used to configure, initially, the surface's pool for the
allocator.
Signed-off-by: Víctor Manuel Jáquez Leal <victorx.jaquez@intel.com>
https://bugzilla.gnome.org/show_bug.cgi?id=744042
The purpose of gstcompat.h is to couple the API differences among
gstreamer-1.0 and gstreamer-0.10. Since gstreamer-0.10 is obsolete, the code
in this compatibility layer shall be removed.
Nevertheless, the gstcompat.h header should be kept, if new incompatibilites
appear in the future, but it shall live in gst/vaapi, not in gst-libs.
This patch removes the crumbs defined gstcompat.h and moves it to gst/vaapi.
In order to avoid layer violations, gstcompat.h includes sysdeps.h and all
the includes in gst/vaapi of sysdeps.h are replaced with gstcompat.h
https://bugzilla.gnome.org/show_bug.cgi?id=745728
Signed-off-by: Gwenole Beauchesne <gwenole.beauchesne@intel.com>
Signed-off-by: Sreerenj Balachandran <sreerenj.balachandran@intel.com>
The dmabuf allocator would close the DMABUF handle passed in the init
function gst_dmabuf_allocator_alloc(). So, we need to dup() it so that
to avoid a double close, ultimately in the underlying driver that owns
the DMABUF handle.
Allow v4l2src element to connected to vaapipostproc or vaapisink when
"io-mode" is set to "dmabuf-import". In practice, this is a more likely
operational mode with uvcvideo. Supporting v4lsrc with "io-mode" set
to "dmabuf" could work, but with more demanding driver or kernel reqs.
Note: with GStreamer 1.4, v4l2src (gst-plugins-good) needs to be built
with --without-libv4l2.
https://bugzilla.gnome.org/show_bug.cgi?id=743635
Reword surface pool allocation helpers so that to allow for a simple
form, e.g. gst_vaapi_surface_pool_new(format, width, height); and a
somewhat more elaborated/flexible form with optional allocation flags
and precise GstVideoInfo specification.
This is an API/ABI change, and SONAME version needs to be bumped.
In some occasions, a buffer pool is created for pre-initialization
purposes regardless of whether a valid image size is available or
not. However, during actual decode stage, the vaapidecode element
is expected to update the srcpad caps with the new dimensions, thus
also triggering a reset of the underlying bufferpool.
The ensure_surface() and ensure_image() functions shall only relate
to the underlying backing store. The actual current flags are to be
updated only through ensure_{surface,image}_is_current() or very other
particular cases in GstMemory hooks.
If the surface proxy is updated into the GstVaapiVideoMemory, then
it is assumed it is the most current representation of the current
video frame. Likewise, make a few more arrangements to have the
"current " flags set more consistently.
When interacting with SW elements, the buffers and underlying video
memory could be mapped as read/write. However, we need to use those
buffers again as plain VA surfaces, we have to make sure the VA image
is thus committed back to VA surface memory.
This fixes pipelines involving avdec_* and vaapi{postproc,sink}.
Hence vaapisink can display buffers decoded by gst-libav, or HW decoded
buffers can be further processed in-place, e.g. with a textoverlay.
https://bugzilla.gnome.org/show_bug.cgi?id=704078
[ported to current git master branch, amended commit message]
Signed-off-by: Gwenole Beauchesne <gwenole.beauchesne@intel.com>
Add GST_VAAPI_VIDEO_BUFFER_POOL_ACQUIRE_FLAG_NO_ALLOC params flag that
can be used to disable early allocations of vaapi video metas on buffers,
thus delagating that to the bufferpool user.
Make sure to always prefer native internal formats for the VA surfaces
that get allocated. Also disable "direct-rendering" mode in this case.
This is needed so that to make sure that anything that gets out of the
decoder, or anything that gets into the encoder, is in native format
for the hardware, and thus the driver doesn't need to perform implicit
conversions in there. Interop with SW elements is still available with
fast implementations of VA imaging APIs.
Forbid shares of GstMemory instances, and rather make copy of it. This
effectively copies the GstMemory structure and enclosed metadata, but
this does not copy the VA surface contents itself. It should though.
This fixes preroll and makes sure to not download garbage for the first
frame when a SW rendering sink is used.
Use an image pool to hold VA images to be used for downloads/uploads
of contents for the associated surface.
This is an optmization for size. So, instead of creating as many VA
images as there are buffers (then VA surfaces) allocated, we only
maintain a minimal set of live VA images, thus preserving memory
resources.
Disable read-write mappings if "direct-rendering" is not supported.
Since the ordering of read and write operations is not specified,
this would require to always download the VA surface on _map(), then
commit the temporary VA image back to the VA surface on _unmap().
Some SW decoding plug-in elements still use R/W mappings though.
https://bugzilla.gnome.org/show_bug.cgi?id=733242
Fix error messages introduced in the previous commit for the _map()
imaplementation. Also use the new get_image_data() helper function
to determine the base pixels data buffer from a GstVaapiImage when
updating the video info structure from it.
Allow raw pixels of the whole frame to be mapped read-only. i.e. in
cases where the buffer pool is allocated without VideoMeta API, thus
individual planes cannot be mapped.
This is initial support for Firefox >= 30.
https://bugzilla.gnome.org/show_bug.cgi?id=731886
While creating the vaapi video allocator, make sure the associated
surface pool has correct format instead of defaulting to NV12 video
format even though there is no direct rendering support.
https://bugzilla.gnome.org/show_bug.cgi?id=732691
Make sure to always update the VA surface pointer whenever the proxy
changes. This used to only work when the VA surface is written to, in
interop with SW element ("upload" feature), and this now fixes cases
when the VA surface is needed for reading, in interop with SW element
("download" feature).
Always expose I420 format by default when the VA surface could be
mapped for interoperability with non harware accelerated elements.
However, the default behaviour remains the auto-plugging of vaapi
elements, down to the sink.
Side effect: "direct-rendering" mode is also disabled most of the
times as plain memcpy() from uncached speculative write combining
memory is not going to be efficient enough.
Fix support for VA surface download capability in vaapidecode element
for GStreamer >= 1.2. This is a fix to supporting libva-vdpau-driver,
but also the libva-intel-driver while performing hardware accelerated
conversions from the native VA surface format (NV12) to the desired
output VA image format.
For instance, this fixes pipelines involving vaapidecode ! xvimagesink.
https://bugzilla.gnome.org/show_bug.cgi?id=733243
When GstVideoMeta maps were used, the supporting functions incorrectly
used gst_buffer_get_memory() instead of gst_buffer_peek_memory(), thus
always increasing the associated GstMemory reference count and giving
zero chance to actually release that, and subsequently the VA display.
