Obviously, in the vast majority of cases, there's only one device
in the system, but doing this means we're more likely to get a
usable device in the multi-device case.
cuda would support decoding on one device and displaying on another
but the peer memory handling is not transparent and I have no way
to test it so I can't really write it.
The documentation around this stuff is poor, but I found an nvidia
sample that demonstrates how to use the interop API most efficiently.
(https://github.com/nvpro-samples/gl_cuda_interop_pingpong_st)
Key lessons are:
1) you can register the texture itself and have cuda write to it,
thereby skipping an additional copy through the PBO.
2) You don't have to be mapped when you do the copy - once you get a
mapped pointer, it remains valid. Magic!
This lets us throw out the PBOs as well as much of the explicit
alignment and stride handling.
CPU usage is slightly (~3%) lower for 4K content in one test case,
so it makes a detectable difference, and presumably saves memory.
On x11, you can change the fullscreen via the window manager and without
mpv's involvement. In these cases, the internal fullscreen flag has to
be updated.
The hack used for this didn't really work properly. Change it
accordingly. The important thing is that the shadow copy of the option
is updated. This is still not really ideal.
Fixes#3570.
The documentation claims that --video-unscaled will still perform
anamorphic adjustments, and it rightfully should. The current reality is
that it does not, because the video-unscaled size was based on the wrong
set of variables. (encoded width/height instead of nominal display
width/height)
When we rotate the inmage by 90° or 270°, chroma width and height need
to be swapped.
Fixes#3568.
But is the chroma sub location correct? Who the hell knows...
Extend the flag-based notification mechanism that was used via
M_OPT_TERM. Make the vo_opengl update mechanism use this (which, btw.,
also fixes compilation with OpenGL renderers forcibly disabled).
While this adds a 3rd mechanism and just seems to further the chaos, I'd
rather have a very simple mechanism now, than actually furthering the
mess by mixing old and new update mechanisms. In particular, we'll be
able to remove quite some property implementations, and replace them
with much simpler update handling. The new update mechanism can also
more easily refactored once we have a final mechanism that handles
everything in an uniform way.
There were multiple values under M_OPT_EXIT (M_OPT_EXIT-n for n>=0).
Somehow M_OPT_EXIT-n either meant error code n (with n==0 no error?), or
the number of option valus consumed (0 or 1). The latter is MPlayer
legacy, which left it to the option type parsers to determine whether an
option took a value or not. All of this was changed in mpv, by requiring
the user to use explicit syntax ("--opt=val" instead of "-opt val").
In any case, the n value wasn't even used (anymore), so rip this all
out. Now M_OPT_EXIT-1 doesn't mean anything, and could be used by a new
error code.
Negative height is used to signal a flipped framebuffer. There's
absolutely no reason to pass this down to overlay_adjust(), and only
requires implementers to deal with an additional special-case.
Instead of using input_ctx for waiting, use the dispatch queue directly.
One big change is that the dispatch queue will just process commands
that come in (e.g. from client API) without returning. This should
reduce unnecessary playloop excutions (which is good since the playloop
got a bit fat from rechecking a lot of conditions every iteration).
Since this doesn't force a new playloop iteration on every access, this
has to be enforced manually in some cases.
Normal input (via terminal or VO window) still wakes up the playloop
every time, though that's not too important. It makes testing this
harder, though. If there are missing wakeup calls, it will be noticed
only when using the client API in some form.
At this point we could probably use a normal lock instead of the
dispatch queue stuff.
Currently, calling mp_input_wakeup() will wake up the core thread (also
called the playloop). This seems odd, but currently the core indeed
calls mp_input_wait() when it has nothing more to do. It's done this way
because MPlayer used input_ctx as central "mainloop".
This is probably going to change. Remove direct calls to this function,
and replace it with mp_wakeup_core() calls. ao and vo are changed to use
opaque callbacks and not use input_ctx for this purpose. Other code
already uses opaque callbacks, or has legitimate reasons to use
input_ctx directly (such as sending actual user input).
'cuda-gl' isn't right - you can turn this on without any GL and
get some non-zero benefit (with the cuda-copy hwaccel). So
'cuda-hwaccel' seems more consistent with everything else.
This happened to break because the texture unit wasn't reset to 0, which
some code expects. The OSD code in particular set the OSD texture on the
wrong texture unit, with the result that OSD/OSC was not visible.
A minor cleanup that makes the code simpler, and guarantees that we
cleanup the GL state properly at any point.
We do this by reusing the uniform caching, and assigning each sampler
uniform its own texture unit by incrementing a counter. This has various
subtle consequences for the GL driver, which hopefully don't matter. For
example, it will bind fewer textures at a time, but also rebind them
more often.
For some reason we keep TEXUNIT_VIDEO_NUM, because it limits the number
of hook passes that can be bound at the same time.
OSD rendering is an exception: we do many passes with the same shader,
and rebinding the texture each pass. For now, this is handled in an
unclean way, and we make the shader cache reserve texture unit 0 for the
OSD texture. At a later point, we should allocate that one dynamically
too, and just pass the texture unit to the OSD rendering code. Right now
I feel like vo_rpi.c (may it rot in hell) is in the way.
The caller now has to call gl_sc_reset(), and _after_ rendering. This
way we can unset OpenGL state that was setup for rendering. This affects
the shader program, for example. The next commit uses this to
automatically manage texture units via the shader cache.
vo_rpi.c changes untested.
Stops Mesa from restricting us to OpenGL 3.0. It also tries to create
GLES 3 contexts for drivers which do not just return a higher context
when requesting GLES 2.
I don't know whether this code is a good or bad idea. A not-so-good
aspect is that we don't check for EGL 1.5 (or 1.4 extensions) for some
of the more advanced context attributes. But EGL implementations should
be able to tolerate it and return an error, and then we'd use the
fallback.
This used to be shared, but since vo_rpi is going to be removed,
untangle them. There was barely any actual code shared since the recent
changes anyway.
As a subtle change, we also stop opening libGLESv2.so explicitly in the
vo_opengl backend, and use RTLD_DEFAULT instead.
Minimal support just for testing.
Only the window surface creation (including size determination) is
really platform specific, so this could be some generic thing with
platform-specific support as some sort of sub-driver, but on the other
hand I don't see much of a need for such a thing.
While most of the fbdev usage is done by the EGL driver, using this
fbdev ioctl is apparently the only way to get the display resolution.
Add a function to egl_helpers.c for creating an EGL context and make
context_x11egl.c use it. This is meant to be generic, and should work
with other windowing APIs as well. The other EGL-using code in mpv can
be switched to it.
The wrong enum got copied here, so it was essentially using the transfer
characteristics as the primaries (instead of the primaries), which
accidentally worked fine most of the time (since the two usually
coincided), but broke on weird/mistagged files.
The consequence of this was that e.g. hardware decoding with VAAPI-EGL
could sometimes not work if the compiler didn't support C11. (Although I
found this one on RPI, which also uses this mechanism.)
If the shader fails to compile, and assertion could trigger in
gl_sc_gen_shader_and_reset() due to the code trying to recreate the
shader every time, and re-appending the uniforms every time. Just reset
the uniform array to fix this.
Some disturbed GL drivers might not return anything for glGetShaderiv()
if the GL state got "lost", so initialize variables just for additional
robustness.
Since vo_rpi is going to be deprecated, better port its features to the
vo_opengl backend.
The most tricky part is the fact that recreating dispmanx elements will
conflict with the GL context. Fortunately, RPI's EGL support is
reasonably compliant, and we can transplant the context to newly created
dispmanx elements, making this much easier. This means unlike vo_rpi,
the GL state will actually not be recreated.
This overlay support specifically skips the OpenGL rendering chain, and
uses GL rendering only for OSD/subtitles. This is for devices which
don't have performant GL support.
hwdec_rpi.c contains code ported from vo_rpi.c. vo_rpi.c is going to be
deprecated. I left in the code for uploading sw surfaces (as it might
be slightly more efficient for rendering sw decoded video), although
it's dead code for now.
The nvidia examples use the old (as in CUDA 3.x) interop API which
is deprecated, and I think not even functional on recent versions
of CUDA for windows. As I was following the examples, I used this
old API.
So, let's update to the new API, and hopefully, it'll start working
on windows too.
It was used to determine whether the VO supports VOCTRL_SET_PANSCAN.
With all those changes to property semantics this became unnecessary,
and its only use was dropped at some point.
Just another corner-caseish potential issue. Unlike unreffing the image
manually, unref_current_image() also takes care of properly unmapping
hwdec frames. (The corner-case part of this is that it's probably never
mapped at this point, but it's apparently not entirely guaranteed.)
The " || vimg->mpi" part virtually never seems to trigger, but on the
other hand could possibly create unintended corner cases (for example by
trying to upload a NULL image, which would then be marked as an error
and render a blue screen).
I guess it's a leftover from over times, where a NULL image meant
"redraw the current frame". This is now handled by actually passing
along the current frame.
Nvidia's "NvDecode" API (up until recently called "cuvid" is a cross
platform, but nvidia proprietary API that exposes their hardware
video decoding capabilities. It is analogous to their DXVA or VDPAU
support on Windows or Linux but without using platform specific API
calls.
As a rule, you'd rather use DXVA or VDPAU as these are more mature
and well supported APIs, but on Linux, VDPAU is falling behind the
hardware capabilities, and there's no sign that nvidia are making
the investments to update it.
Most concretely, this means that there is no VP8/9 or HEVC Main10
support in VDPAU. On the other hand, NvDecode does export vp8/9 and
partial support for HEVC Main10 (more on that below).
ffmpeg already has support in the form of the "cuvid" family of
decoders. Due to the design of the API, it is best exposed as a full
decoder rather than an hwaccel. As such, there are decoders like
h264_cuvid, hevc_cuvid, etc.
These decoders support two output paths today - in both cases, NV12
frames are returned, either in CUDA device memory or regular system
memory.
In the case of the system memory path, the decoders can be used
as-is in mpv today with a command line like:
mpv --vd=lavc:h264_cuvid foobar.mp4
Doing this will take advantage of hardware decoding, but the cost
of the memcpy to system memory adds up, especially for high
resolution video (4K etc).
To avoid that, we need an hwdec that takes advantage of CUDA's
OpenGL interop to copy from device memory into OpenGL textures.
That is what this change implements.
The process is relatively simple as only basic device context
aquisition needs to be done by us - the CUDA buffer pool is managed
by the decoder - thankfully.
The hwdec looks a bit like the vdpau interop one - the hwdec
maintains a single set of plane textures and each output frame
is repeatedly mapped into these textures to pass on.
The frames are always in NV12 format, at least until 10bit output
supports emerges.
The only slightly interesting part of the copying process is that
CUDA works by associating PBOs, so we need to define these for
each of the textures.
TODO Items:
* I need to add a download_image function for screenshots. This
would do the same copy to system memory that the decoder's
system memory output does.
* There are items to investigate on the ffmpeg side. There appears
to be a problem with timestamps for some content.
Final note: I mentioned HEVC Main10. While there is no 10bit output
support, NvDecode can return dithered 8bit NV12 so you can take
advantage of the hardware acceleration.
This particular mode requires compiling ffmpeg with a modified
header (or possibly the CUDA 8 RC) and is not upstream in ffmpeg
yet.
Usage:
You will need to specify vo=opengl and hwdec=cuda.
Note that hwdec=auto will probably not work as it will try to use
vdpau first.
mpv --hwdec=cuda --vo=opengl foobar.mp4
If you want to use filters that require frames in system memory,
just use the decoder directly without the hwdec, as documented
above.
