PresentNotifyMSC turns out to be not only redundant, but also harmful with
mesa-backed egl/glx/vulkan VOs because for all of them, mesa uses
PresentPixmap behind the scenes when DRI3 is available, which already
spawns a PresentCompleteNotify event when the buffer swap actually
finishes. This is important because without using the timing information
from these PresentCompleteKindPixmap events, there's no way for mpv to know
exactly when a frame becomes visible on the display.
By using PresentNotifyMSC in conjunction with DRI3-enabled mesa, two
problems are created:
1. mpv assumes that a vblank won't elapse (i.e., it assumes the current MSC
won't change) between the time when mesa enqueues the buffer swap and
the time when mpv calls PresentNotifyMSC to ask xorg for a notification
at the next MSC, relative to the current MSC at the time that xorg reads
it for the PresentNotifyMSC call. This means that mpv could get a
notification one or more vblanks later than it expects, since the
intention here is for mpv to get a notification at the MSC that the
buffer swap completes.
2. mpv assumes that a buffer swap always takes one vblank to complete,
which isn't always true. A buffer swap (i.e., a page flip) could take
longer than that depending on hardware conditions (if the GPU is running
slowly or needs to exit a low-power state), scheduling delays (under
heavy system or GPU load), or unfortunate timing (if the raster scan
line happens to be at one of the last few rows of pixels and a vblank
elapses just before the buffer swap is enqueued).
This causes mpv to have a faulty assumption of when frames become visible.
Since mpv already receives the PresentCompleteNotify events generated by
mesa's buffer swaps under the hood, the PresentNotifyMSC usage is unneeded
and just throws a wrench in mpv's vsync timing when xpresent is enabled.
Simply removing the PresentNotifyMSC usage from the egl, glx, and vulkan
VOs fixes the xpresent vsync timing.
This commit kind of mixes several related things together. The main
thing is to avoid calling any XPresent functions or internal functions
related to presentation when the feature is not auto-whitelisted or
enabled by the user. Internally rework this so it all works off of a
use_present bool (have_present is eliminated because having a non-zero
present_code covers exactly the same thing) and make sure it updates on
runtime. Finally, put some actual logging in here whenever XPresent is
enabled/disabled. Fixes#10326.
This builds off of present_sync which was introduced in a previous
commit to support xorg's present extension in all of the X11 backends
(sans vdpau) in mpv. It turns out there is an Xpresent library that
integrates the xorg present extention with Xlib (which barely anyone
seems to use), so this can be added without too much trouble. The
workflow is to first setup the event by telling Xorg we would like to
receive PresentCompleteNotify (there are others in the extension but
this is the only one we really care about). After that, just call
XPresentNotifyMSC after every buffer swap with a target_msc of 0. Xorg
then returns the last presentation through its usual event loop and we
go ahead and use that information to update mpv's values for vsync
timing purposes. One theoretical weakness of this approach is that the
present event is put on the same queue as the rest of the XEvents. It
would be nicer for it be placed somewhere else so we could just wait
on that queue without having to deal with other possible events in
there. In theory, xcb could do that with special events, but it doesn't
really matter in practice.
Unsurprisingly, this doesn't work on NVIDIA. Well NVIDIA does actually
receive presentation events, but for whatever the calculations used make
timings worse which defeats the purpose. This works perfectly fine on
Mesa however. Utilizing the previous commit that detects Xrandr
providers, we can enable this mechanism for users that have both Mesa
and not NVIDIA (to avoid messing up anyone that has a switchable
graphics system or such). Patches welcome if anyone figures out how to
fix this on NVIDIA.
Unlike the EGL/GLX sync extensions, the present extension works with any
graphics API (good for vulkan since its timing extension has been in
development hell). NVIDIA also happens to have zero support for the
EGL/GLX sync extensions, so we can just remove it with no loss. Only
Xorg ever used it and other backends already have their own present
methods. vo_vdpau VO is a special case that has its own fancying timing
code in its flip_page. This presumably works well, and I have no way of
testing it so just leave it as it is.
Because wayland is a special snowflake, mpv wound up incorporating a lot
of logic into its render loop where visibilty checks are performed
before rendering anything (in the name of efficiency of course). Only
wayland actually uses this, but there's no reason why other backends
(x11 in this commit) can't be smarter. It's far easier on xorg since we
can just query _NET_WM_STATE_HIDDEN directly and not have to do silly
callback dances.
The function, vo_x11_check_net_wm_state_change, already tracks net wm
changes, including _NET_WM_STATE_HIDDEN. There is an already existing
window_hidden variable but that is actually just for checking if the
window was mapped and has nothing to do with this particular atom. mpv
also currently assumes that a _NET_WM_STATE_HIDDEN is exactly the same
as being minimized but according to the spec, that's not neccesarily
true (in practice, it's likely that these are the same though). Anyways,
just keep track of this state in a new variable (hidden) and use that
for determing if mpv should render or not.
There is one catch though: this cannot work if a display sync mode is
used. This is why the previous commit is needed. The display sync modes
in mpv require a blocking vsync implementation since its render loop is
directly driven by vsync. In xorg, if nothing is actually rendered, then
there's nothing for eglSwapBuffers (or FIFO for vulkan) to block on so
it returns immediately. This, of course, results in completely broken
video. We just need to check to make sure that we aren't in a display
sync mode before trying to be smart about rendering. Display sync is
power inefficient anyways, so no one is really being hurt here. As an
aside, this happens to work in wayland because there's basically a
custom (and ugly) vsync blocking function + timeout but that's off
topic.
GLX_CONTEXT_PROFILE_MASK_ARB and related constants are provided by
GLX_ARB_create_context_profile but the check was for _create_context.
The former implies the latter (which we also need) so just replace
the checked extension.
Using a simple substring match for extension checks is considered bad practice
because it's incorrect when one extension is a prefix of another's name.
This will almost surely not make a difference in practice but do it for correctness anyway.
Drop the gl3 suffix from the function name - it's no longer needed, with
the _old function gone.
Push the mpgl_min_required_gl_versions[] looping within the function,
reducing the identical glXGetProcAddress/glXQueryExtensionsString calls
while making the code neater.
v2:
- tabs -> spaces indentation
- mpgl_preferred_gl_versions -> mpgl_min_required_gl_versions
- 320 -> 300 (in glx code path)
v3:
- legacy code path is gone \o/
Signed-off-by: Emil Velikov <emil.l.velikov@gmail.com>
The old/legacy code-path isn't really needed for mpv use-cases.
In particular:
All Mesa drivers (even GL 2.1 ones like lima/vc4) work fine with the
"non-legacy" path.
From proprietary/binary drivers - the vendor either does not support
desktop GL or the drivers/HW is not actively supported.
Looking at the Nvidia HW - anything GeForce 7 and older is GL 2.1 and
lacks decent video acceleration. The latest official drivers are from
2017.
All newer Nvidia HW is GL 3.3+ thus must have GLX_ARB_create_context as
in the non-legacy path, while also good acceleration albeit via VDPAU
in some cases.
With the old path gone, provide meaningful error message in the very
unlikely case that GLX_ARB_create_context is missing.
Signed-off-by: Emil Velikov <emil.l.velikov@gmail.com>
The ra_gl_ctx_test_version() helper is quite clunky, in that it pushes a
simple check too deep into the call chain. As such it makes it hard to
reason, let alone have the GLX and EGL code paths symmetrical.
Introduce a simple helper ra_gl_ctx_get_glesmode() which returns the
current glesmode, so the platforms can clearly reason about should and
should not be executed.
v2:
- mpgl_preferred_gl_versions -> mpgl_min_required_gl_versions
- 320 -> 300 (in glx code path)
Signed-off-by: Emil Velikov <emil.l.velikov@gmail.com>
Currently mpv requires a bare minimum of GL 2.1, although it tries to
use 3.2+ core contexts when possible.
The GLX and EGL spec effectively guarantee that the implementation will
give you the highest compatible version possible. In other words:
Requesting 3.2 core profile will always give you core profile and the
version will be in the 3.2 .. 4.6 range - as supported by the drivers.
Similarly for 2.1 - implementation will give you either:
- 2.1 .. 3.1, or
- 3.2 .. 4.6 compat profile
This has been verified against the Mesa drivers (i965, iris, swrast) and
Nvidia binary drivers.
As such, drop the list to 320, 210 and terminating 0.
v2:
- mpgl_preferred_gl_versions -> mpgl_min_required_gl_versions
- update ^^ comment
Signed-off-by: Emil Velikov <emil.l.velikov@gmail.com>
Surprisingly, we've managed to get this far without context_glx ever
adding the X11 display as a native resource. But with the recent change
to attempt to enable vdpau when using EGL, the hwdec now requires the
display to be added. So let's add it.
Originally, vo_gpu/vo_opengl considered the case of Nvidia proprietary
drivers, which required vdpau/GLX, and Intel open source drivers, which
require vaapi/EGL. Since window creation and GPU context creation are
inseparable in mpv's internal API, it had to pick the correct API very
early, or hardware decoding wouldn't work. "x11probe" was introduced for
this reason. It created a GLX context (without showing the window yet),
and checked whether vdpau was available. If yes, it used GLX, if not, it
continued probing x11/EGL. (Obviously it couldn't always fail on GLX
without vdpau, which is why it was a separate "probe" backend.)
Years passed, and now the situation is different. Vdpau is dead. Nvidia
drivers and libavcodec now provide CUDA interop, which requires EGL, and
fixes some of the vdpau problems. AMD drivers now provide vaapi, which
generally works better than vdpau. Intel didn't change.
In particular, vaapi provides working HEVC Main10 support. In theory, it
should work on vdpau too, with quality reduction (no 10 bit surfaces),
but I couldn't get it to work.
So always prefer EGL. And suddenly hardware decoding works. This is
actually rather important, because HEVC is unfortunately on the rise,
despite shitty encoders and unoptimized decoders. The latter may mean
that hardware decoding works better than libavcodec.
This should have been done a long, long time ago.
So the next commit can make EGL use it. EGL has a quite similar
function, that practically works the same. Although it's relatively
trivial, it's still tricky, and probably shouldn't end up as duplicated
code.
There are no functional changes, except initialization, and how failure
of the glXGetSyncValues call is handled. Also, some comments mention the
EGL extension.
Note that there's no intention for this code to handle anything else
than the very specific OML sync extension (and its EGL equivalent). This
is just too weirdly specific to the weird idiosyncrasies of the
extension, and it makes no sense to extend it to handle anything else.
(Such as Wayland or DXGI presentation feedback.)
I misunderstood how this extension works. If I understand it correctly
now, it's worse than I thought. They key thing is that the (ust, msc,
sbc) tripple is not for a single swap event. Instead, (ust, msc) run
independently from sbc. Assuming a CFR display/compositor, this means
you can at best know the vsync phase and frequency, but not the exact
time a sbc changed value.
There is GLX_INTEL_swap_event, which might work as expected, but it has
no EGL equivalent (while GLX_OML_sync_control does, in theory).
Redo the context_glx sync code. Now it's either more correct or less
correct. I wanted to add proper skip detection (if a vsync gets skipped
due to rendering taking too long and other problems), but it turned out
to be too complex, so only some unused fields in vo.h are left of it.
The "generic" skip detection has to do.
The vsync_duration field is also unused by vo.c.
Actually this seems to be an improvement. In cases where the flip call
timing is off, but the real driver-level timing apparently still works,
this will not report vsync skips or higher vsync jitter anymore. I could
observe this with screenshots and fullscreen switching. On the other
hand, maybe it just introduces an A/V offset or so.
Why the fuck can't there be a proper API for retrieving these
statistics? I'm not even asking for much.
Use the extension to compute the (hopefully correct) video delay and
vsync phase.
This is very fuzzy, because the latency will suddenly be applied after
some frames have already been shown. This means there _will_ be "jumps"
in the time accounting, which can lead to strange effects at start of
playback (such as making initial "dropped" etc. frames worse). The only
reasonable way to fix this would be running a few dummy frame swaps at
start of playback until the latency is known. The same happens when
unpausing.
This only affects display-sync mode.
Correct function was not confirmed. It only "looks right". I don't have
the equipment to make scientifically correct measurements.
A potentially bad thing is that we trust the timestamps we're receiving.
Out of bounds timestamps could wreak havoc. On the other hand, this will
probably cause the higher level code to panic and just disable DS.
As a further caveat, this makes a bunch of assumptions about UST
timestamps. If there are delayed frames (i.e. we skipped one or more
vsyncs), the latency logic is mostly reset. There is no attempt to make
the vo.c skipped vsync logic to use this. Also, the latency computation
determines a vsync duration, and there's no effort to reconcile or share
the vo.c logic for determining vsync duration.
This is done in several steps:
1. refactor MPGLContext -> struct ra_ctx
2. move GL-specific stuff in vo_opengl into opengl/context.c
3. generalize context creation to support other APIs, and add --gpu-api
4. rename all of the --opengl- options that are no longer opengl-specific
5. move all of the stuff from opengl/* that isn't GL-specific into gpu/
(note: opengl/gl_utils.h became opengl/utils.h)
6. rename vo_opengl to vo_gpu
7. to handle window screenshots, the short-term approach was to just add
it to ra_swchain_fns. Long term (and for vulkan) this has to be moved to
ra itself (and vo_gpu altered to compensate), but this was a stop-gap
measure to prevent this commit from getting too big
8. move ra->fns->flush to ra_gl_ctx instead
9. some other minor changes that I've probably already forgotten
Note: This is one half of a major refactor, the other half of which is
provided by rossy's following commit. This commit enables support for
all linux platforms, while his version enables support for all non-linux
platforms.
Note 2: vo_opengl_cb.c also re-uses ra_gl_ctx so it benefits from the
--opengl- options like --opengl-early-flush, --opengl-finish etc. Should
be a strict superset of the old functionality.
Disclaimer: Since I have no way of compiling mpv on all platforms, some
of these ports were done blindly. Specifically, the blind ports included
context_mali_fbdev.c and context_rpi.c. Since they're both based on
egl_helpers, the port should have gone smoothly without any major
changes required. But if somebody complains about a compile error on
those platforms (assuming anybody actually uses them), you know where to
complain.
