Previously, vaapi_recorder_frame() would wait until the encoded
contents for a frame is written to the output file descriptor. This
delayed the repainting of the next frame, and affected frame rate
when capturing with high resolutions. Instead, wait only if there is
and attempted to encode two frames at the same time.
Increases framerate from 30 to 60 fps when capturing at 1920x1200 on
my SandryBridge system, although there are periodic slowdowns due to
disk writes.
it uses the Android fbdev HAL[1] (through libhybris[2])
and the libhybris implementation of wayland-egl.
Configure flags:
cairo:
--enable-glesv2=yes --enable-egl=yes
weston:
--with-cairo-glesv2 --enable-fbdev-compositor
hybris:
--enable-wayland --enable-arch=x86
--with-android-headers=<android-headers> --enable-alinker=jb
The android headers are extracted from an AOSP tree,
using hybris/utils/extract-headers.sh
[1]:
https://github.com/android/platform_hardware_libhardware/blob/master/include/hardware/fb.h
[2]: https://github.com/libhybris/libhybris
Signed-off-by: Adrian Negreanu <adrian.m.negreanu@intel.com>
Hi Kristian,
Here's a new patch for ref counting weston_xkb_info, as suggested.
So a seat created with a NULL keymap will now point to the global xkb_info.
Typically we can write it immediately without blocking, so save the overhead
of setting up an fd watch and writing the data in a callback. For the
case where the immediate write doesn't write all data, we fallback and
set up the fd watch as usual.
This patch also consolidates setting up the async write a bit.
This makes the drag-and-drop code available to in-weston data sources,
similar to how we can set a selection data source internally. The
wl_data_device.start_drag entry point now calls this function after
validating protocol arguments.
Another silent regression from the wl_resource opaquify effort. This was
causing our pageflip-to-client-buffer and sprites optimizations to
not kick in.
We used to destroy the frame window and reparent the client window to
wm_window. That means that we lose the destroy_notify event when the
client window is destroyed later, since we don't select for
substructure_notify on wm_window.
Instead of destroying and reparenting, just unmap the frame window.
This patch adds a feature to the DRM backend that uses libva for
encoding the screen contents in H.264. Screen recording can be
activated by pressing mod-shift-space q. A file named capture.h264
will be created in the current directory, which can be muxed into
an MP4 file with gstreamer using
gst-launch filesrc location=capture.h264 ! h264parse ! mp4mux ! \
filesink location=file.mp4
This is limitted to the DRM compositor in order to avoid a copy when
submitting the front buffer to libva. The code in vaapi-recorder.c
takes a dma_buf fd referencing it, does a colorspace conversion using
the video post processing pipeline and then uses that as input to the
encoder.
I'm sending this now so I get comments, but this is not ready for
prime time yet. I have a somewhat consistent GPU hang when using
i915 with SandyBridge. Sometimes a page flip never completes. If you
want to try this anyway and your system get stuck, you might need to
run the following:
# echo 1 > /sys/kernel/debug/dri/0/i915_wedged
After that, alt-sysrq [rv] should work.
Once that's fixed it would also be good to make the parameters used by
the encoder more flexible. For now the QP parameter is hardcoded to 0
and we have only I and P frames (no B frames), which causes the
resulting files to be very large.
GL_UNSIGNED_INT is only supported when GL_OES_element_index_uint is
available (mesa implements that extension). We don't need 32-bit
indices, so just use GL_UNSIGNED_SHORT.
We don't always get both an X and an Y event in a SYN report, so we end
up transforming the coordinate we don't get twice. For example, if we
only receive an ABS_X event, we transform the already transformed
device->abs.y again in transform_absolute() when applying the calibration.
Now that we use AC_SYS_LARGEFILE, we need to pull in config.h at least
whereever we use mmap(). Fixes at least the test-suite and simple-shm
on 32 bit systems.
window->x/y is the coordinate of the top-level surface (whether that's
the frame window or an override-redirect window) and the wayland surface
should be placed there, without the t->margin offset.
The coordinate transformation was broken (worked for first output where
output->x/y was 0,0, broke on all other outputs). We can just use
surface->geometry.x/y directly. We can't use the full transformation,
the best we can do is to move the X window to the geometry.x/y location.
Get rid of the static old_sx/sy hack as well.
We only get configure notify for toplevel (frame or override-redirect window)
and those are the cases where we want to update window->x/y. The way the
code worked, we'd exit immeidately in those cases and window->x/y would
not be updated.
