This patch implements support in LibVNCClient for framebuffer updates
encoded as H.264 frames. Hardware accelerated decoding is performed
using VA API.
This is experimental support to let the community explore the possibilities
offered by the potential bandwidth and latency reductions that H.264 encoding
allows. This may be particularly useful for use cases such as online gaming,
hosted desktops, hosted set top boxes...
This patch only provides the client side support and is meant to be used
with corresponding server-side support, as provided by an upcoming patch for
qemu ui/vnc module (to view the display of a virtual machine executing under
QEMU).
With this H.264-based encoding, if multiple framebuffer update messages
are generated for a single server framebuffer modification, the H.264
frame data is sent only with the first update message. Subsequent update
framebuffer messages will contain only the coordinates and size of the
additional updated regions.
Instructions/Requirements:
* The patch should be applied on top of the previous patch I submitted with
minor enhancements to the gtkvncviewer application:
http://sourceforge.net/mailarchive/message.php?msg_id=30323804
* Currently only works with libva 1.0: use branch "v1.0-branch" for libva and
intel-driver. Those can be built as follows:
cd libva
git checkout v1.0-branch
./autogen.sh
make
sudo make install
cd ..
git clone git://anongit.freedesktop.org/vaapi/intel-driver
cd intel-driver
git checkout v1.0-branch
./autogen.sh
make
sudo make install
Signed-off-by: David Verbeiren <david.verbeiren@intel.com>
GCC < 4.6 failed to parse the declaration of ws_header_t correctly because
it did not accept anonymous structs and unions. [1]
Work around the bug by adding names to the unions and structs. Ugly, but
works.
[1] http://gcc.gnu.org/bugzilla/show_bug.cgi?id=4784
byteswap.h exists only on glibc, so building libvncserver with websockets
support was not possible in other systems.
Replace the inclusion of byteswap.h and the WS_* definitions with calls to
htobeNN, which should perform the same conversions, be more portable and
avoid the need to check for the platform's endianness.
libresolv is only present on systems which use glibc; platforms such as
FreeBSD have __b64_ntop as part of libc itself.
Improve the detection process and only link against libresolv if it exists
on the system, and remember to reset CMAKE_REQUIRED_LIBRARIES after
performing the necessary tests, since we do not always want to link against
libresolv.
The current definitions were mostly useful to glibc and followed its
feature_test_macros(3) documentation.
However, this means other platforms still had problems when building with
strict compilation flags. _BSD_SOURCE, for example, is only recognized by
glibc, and other platforms sometimes need _XOPEN_SOURCE instead, or even the
removal of some definitions (such as the outdate _POSIX_SOURCE one).
_POSIX_SOURCE also had to be conditionally defined in some places, as what
it enables or disables during compilation varies across systems.
Building with -ansi failed due to some code (as well as system
headers) using non-C89 features. Fix that by adding the usual
_POSIX_SOURCE and _BSD_SOURCE definitions already present in some
other files.
So far, libgcrypt was looked for but no targets linked against it
directly; this caused linking problems for the client and server
examples, as the symbols they needed were not passed to the linker.
The issue that the GnuTLS websockets code uses libgcrypt regardless of
whether it has been found or not has not been touched by this commit,
though.
TightPNG replaces the ZLIB stuff int Tight encoding with PNG. It still
uses JPEG rects as well. Theoretically, we could build TightPNG with only
libpng and libjpeg - without zlib - but libpng depends on zlib, so this is
kinda moot.
Joel Martin
Christian Beier
David Verbeiren
Raphael Kubo da Costa
Oliver Loch
Johannes Schindelin
Rostislav Lisovy
Kyle J. McKay
Luca Falavigna