If the presentation time isn't known, e.g. if the monitor is virtual and
the actual presentation happens far away, the presentation time we
actually received tends to be the time a frame was presented to the next
layer, meaning practically immediately after painting.
When scheduling another update after that, don't assume that if the next
calculated update is not the immediate next update, schedule an update
sooner, as that will in such cases always be true, meaning we ended up
busy looping with constant frame updates being scheduled.
Fix this by only triggering that logic if the last presentation time was
actually vsync:ed.
Part-of: <https://gitlab.gnome.org/GNOME/mutter/-/merge_requests/3803>
Since commit e30eb78891 `ClutterFrameClock` assumes that a valid CPU time
implies timestamp query support, which is also checked in
`cogl_onscreen_egl_swap_buffers_with_damage()`.
Unconditionally setting the CPU time on direct scanout meant that the
compositing path would be stuck on the last (direct scanout optimized)
result on GL implementations without timestamp query support since.
be0aa2976e (clutter/frame-clock: Avoid rapidly toggling dynamic max render time)
Fix that by explicitly marking the gpu rendering duration as valid when
querying the GPU timestamps is supported and check for it ClutterFrameClock.
Fixes: 56580ea7c9 ("backends/native: Assume zero rendering time for direct scanout buffers")
Part-of: <https://gitlab.gnome.org/GNOME/mutter/-/merge_requests/3655>
Currently, ClutterFrameClock uses g_source_set_ready_time() to determine
the usec timing of the next frame. That translates into a poll() with a
millisecond timeout if no trigger occurs to break the poll() out early.
To avoid spinning the CPU, GLib always rounds *up* to the next millisecond
value unless a timeout of 0 was provided by a GSource.
This means that timeouts for the ClutterFrameClock can easily skew beyond
their expected time as the precision is too coarse.
This applies the same concept as GNOME/glib!3949 but just for the
ClutterFrameClock. That may be more ideal than adding a timerfd for every
GMainContext, but we'll see if that lands upstream. I wanted to provide
this here because it could easily be cherry-picked in the mean time if
this is found to be useful.
From a timer stability perspective, this improves things from erratically
jumping between 100s and 1000s off of the expected awake time to single
or low double digits.
Part-of: <https://gitlab.gnome.org/GNOME/mutter/-/merge_requests/3636>
A new variable scheduling mode is introduced which allows lower
priority updates to be scheduled on a timeout which represents a lower
refresh rate, while allowing high priority updates to be scheduled to
occur as soon as possible.
This mode will be used by following commits to implement
synchronization of page flips to the update rate of specifc surface
actors.
High priorty updates are either scheduled to occur "now" if they
arrive at a rate which is lower than the maximum refresh rate, or
according to the measured maximum render time if they arrive at a
rate which meets or exceeds the maximum refresh rate. This approach
allows achieving low input latency in both scenarios.
Seperate handling for low priority updates is needed to avoid visible
stutter in the content of the surface that drives the refresh rate. An
example for a low priority update is cursor movement when the KMS
deadline timer is disabled.
Part-of: <https://gitlab.gnome.org/GNOME/mutter/-/merge_requests/1154>
The new CLUTTER_FRAME_CLOCK_STATE_SCHEDULED_NOW state is almost
identical to CLUTTER_FRAME_CLOCK_STATE_SCHEDULED, with one important
difference being that it avoids updates from being repeatedly
rescheduled "now" when multiple calls to
clutter_frame_clock_schedule_update_now() are done before the source
is actually dispatched.
Such repeated calls to schedule an update "now" may actually postpone
the dispatch if the CPU is very busy and the source dispatch is
delayed, defeating the purpose of scheduling a frame "now".
It also allows rescheduling "now" when the frame clock is uninhibited
after being inhibited while an update was scheduled "now". This may
be important in cases where the frame clock is inhibited for very
short periods in which it would otherwise lose the state of being
scheduled "now".
Scenarios such as this would become more common with the introduction
of variable refresh rate since it makes scheduling "now" a commonplace
occurrence.
Part-of: <https://gitlab.gnome.org/GNOME/mutter/-/merge_requests/3561>
This removes an incorrect implicit assumption in
calculate_next_update_time_us() that a frame may only be scheduled
once in the duration of a refresh cycle. It accomplishes this by
setting last_next_presentation_time_us on presentation feedback
instead of calculating it every time an update is scheduled.
Specifically, it corrects the intended scheduling logic in scenarios
like the following, when all of the below occur in the context of a
single refresh cycle:
1. Frame update (1) is scheduled normally, and
"is_next_presentation_time_valid" is set to TRUE
2. Frame update (1) is dispatched but ends up being "empty" (no
presentation necessary)
3. Frame update (2) is scheduled "now" and
"is_next_presentation_time_valid" is set to FALSE
4. Frame update (2) is dispatched but ends up being "empty" (no
presentation necessary)
5. Frame update (3) is scheduled normally, and since
"is_next_presentation_time_valid" is set to FALSE, the
"early presented event" logic is unintentionally skipped in
calculate_next_update_time_us().
6. Frame update (3) is dispatched and ends up being a "non-empty"
update, but its update time was calculated incorrectly because
some logic was skipped.
Scenarios such as this would become more common with the introduction
of variable refresh rate since it makes scheduling "now" a commonplace
occurrence.
Part-of: <https://gitlab.gnome.org/GNOME/mutter/-/merge_requests/3560>
Calculate the frame deadline in ClutterFrameClock's
calculate_next_update_time_us() rather than in MetaWaylandCompositor's
on_after_update().
The specifics of the deadline calculation for a given frame should be
implementation detail of the frame clock and and remain internal to
allow extensibility.
This extensibility is specifically useful for scenarios where a
different deadline calculation is needed due to alternative frame
scheduling logic, such as for VRR.
No change in behavior.
Part-of: <https://gitlab.gnome.org/GNOME/mutter/-/merge_requests/3521>
In profilers with a timeline or flame graph views it is a very common
scenario that a span name must be displayed in an area too short to fit
it. In this case, profilers may implement automatic shortening to show
the most important part of the span name in the available area. This
makes it easier to tell what's going on without having to zoom all the
way in.
The current trace span names in Mutter don't really follow any system
and cannot really be shortened automatically.
The Tracy profiler shortens with C++ in mind. Consider an example C++
name:
SomeNamespace::SomeClass::some_method(args)
The method name is the most important part, and the arguments with the
class name will be cut if necessary in the order of importance.
This logic makes sence for other languages too, like Rust. I can see it
being implemented in other profilers like Sysprof, since it's generally
useful.
Hence, this commit adjusts our trace names to look like C++ and arrange
the parts of the name in the respective order of importance.
Part-of: <https://gitlab.gnome.org/GNOME/mutter/-/merge_requests/3402>
Scoped traces are less error prone, and they can still be ended
prematurely if needed (this commit makes that work). The only case this
doesn't support is starting a trace inside a scope but ending outside,
but this is pretty unusual, plus we have anchored traces for a limited
variation of that.
Part-of: <https://gitlab.gnome.org/GNOME/mutter/-/merge_requests/3396>
Group all the three config files from clutter/cogl/meta into one
and also remove unnused configurations and replace duplicated ones
This also fixes Cogl usage of HAS_X11/HAS_XLIB to match the expected
build options
Part-of: <https://gitlab.gnome.org/GNOME/mutter/-/merge_requests/3368>
For frame updates in response to sporadic user interaction, this results
in input → output latency somewhere between the minimum possible and the
minimum plus the length of one display refresh cycle (assuming the frame
update can complete within a refresh cycle).
Applying a max_render_time based deadline which corresponds to higher
than the minimum possible latency would result in higher effective
minimum latency for sporadic user interaction.
This was discovered by Ivan Molodetskikh, based on measurements
described in https://mastodon.online/@YaLTeR/110848066454900941 .
v2:
* Set min_render_time_allowed_us = 0 as well, to avoid unthrottled
frame events. (Robert Mader)
Part-of: <https://gitlab.gnome.org/GNOME/mutter/-/merge_requests/3174>
When more than one refresh interval has passed since
last_presentation_time_us.
I honestly can't tell if the previous calculation was correct or not,
but I'm confident the new one is, and it's simpler.
v2:
* ASCII art diagram didn't make sense anymore, try to improve
(Ivan Molodetskikh)
Part-of: <https://gitlab.gnome.org/GNOME/mutter/-/merge_requests/3330>
When CLUTTER_ENABLE_DEBUG is not defined, then CLUTTER_NOTE is defined
as an empty block of code. As a result of that, jitter_us is in that
situation unused, and a compiler warning about this unused variable
appears.
Part-of: <https://gitlab.gnome.org/GNOME/mutter/-/merge_requests/3275>
Measurements above 100% were originally allowed to show when frame skipping
was occurring so you didn't have to also check the frame rate. But that
also resulted in arbitrarily high jitter values being reported when
returning from idle. And those are frequent enough to look like a bug or
untrustworthy so let's not do that anymore.
Taking the remainder of a high jitter value is still a meaningful jitter
value.
Fixes: https://gitlab.gnome.org/GNOME/mutter/-/issues/2906
Part-of: <https://gitlab.gnome.org/GNOME/mutter/-/merge_requests/3123>
clutter_frame_clock_compute_max_render_time_us clamps to the refresh
interval anyway, so the only effect a higher recorded maximum can have
is to delay it falling below the refresh interval again.
Part-of: <https://gitlab.gnome.org/GNOME/mutter/-/merge_requests/3090>
Instead of separate pairs of short- and long-term maxima for each
measured step of the frame update process.
This should result in the render time estimate rising less often:
Previously it did whenever the measurement of any of at least 3 out of
4 steps reached a new maximum, even if that didn't result in a new
maximum for the whole update duration. Now it's only in the latter
case.
This should also result in a lower render time estimate (and thus
input→output latency) in general, since the variability of
measurements for each 3/4 steps doesn't always add up anymore. The flip
side of this is that it might result in missing a display refresh cycle
more often.
v2:
* Fix coding style in maybe_update_longterm_max_duration_us.
(Robert Mader)
Part-of: <https://gitlab.gnome.org/GNOME/mutter/-/merge_requests/3090>
This is expected for the common case of direct scanout of Wayland
buffers where transactions guarantee that all buffer fences are
signalled before a buffer is included in a frame.
Part-of: <https://gitlab.gnome.org/GNOME/mutter/-/merge_requests/3080>
Dispatch jitter is how much the dispatch interval has changed between
frames. It's a measure of sampling smoothness for events that are occurring
at a higher rate than the screen is refreshing:
* Mouse movement
* Clients rendering at swap interval zero
* Keyframe animation position
Zero jitter is ideal but will practically never happen, and a jitter value
of several thousand microseconds will be visible to the naked eye as stutter
even if you're maintaining a perfect frame rate.
To make the numbers easier to interpret we also log the jitter as a
percentage of the refresh interval.
Part-of: <https://gitlab.gnome.org/GNOME/mutter/-/merge_requests/3082>
This could happen when moving the cursor over GUIs that only redraw
in response to cursor movement. Mutter would experience alternating
cursor-only updates and page flips, and so the `max_render_time_allowed_us`
would jump between pessimised and optimised resulting in inconsistent
frame pacing.
Aside from fixing the smoothness problem this should also provide
lower latency cursor movement.
Fixes: https://launchpad.net/bugs/2023766
Part-of: <https://gitlab.gnome.org/GNOME/mutter/-/merge_requests/3074>
It will be used to schedule Wayland frame events independently from both
update and presentation time, as the former may happen multiple times
frame and the later not at all.
For frame events we want a timing that is just late enough to ensure
that a following commit by a Wayland client will not get included into
the current frame any more.
Part-of: <https://gitlab.gnome.org/GNOME/mutter/-/merge_requests/2823>
If we call schedule(), which will schedule an update some time in the
future, and then schedule_now(), we should reschedule the frame clock to
update immediately, and not some time in the future.
Part-of: <https://gitlab.gnome.org/GNOME/mutter/-/merge_requests/2854>
Calculating a timestamp from the past distorts the dispatch lateness
calculation, leading to an inflated max_render_time, which again
increases the likelyhood of next_update_time being in the past.
Fixes 99850f4645
Part-of: <https://gitlab.gnome.org/GNOME/mutter/-/merge_requests/2819>
Store only two values per kind of duration: The short term and long term
maximum.
The short term maximum is updated in each frame clock dispatch. The long
term maximum is updated at most once per second: If the short term
maximum is higher, the long term maximum is updated to match it.
Otherwise, a fraction of the delta between the two maxima is subtracted
from the long term maximum.
Compared to the previous algorithm:
* The calculcations are simpler.
* The calculated max render time has a slow exponential drop-off (by at
most a few milliseconds every second) instead of potentially abruptly
dropping after as few as 16 frames.
This should fix https://gitlab.gnome.org/GNOME/gnome-shell/-/issues/4830
since the short term maximum should always include a sample from the
clock's second tick.
v2:
* Use divisor 2 instead of 4.
Part-of: <https://gitlab.gnome.org/GNOME/mutter/-/merge_requests/2500>
Dispatch lateness is the difference between when we wanted frame clock
dispatch to run and when it actually started running. This can be up to
1ms even under normal circumstances due to process scheduling
granularity, or even higher under load.
This keeps track of dispatch lateness of the last 16 frame clock
dispatches, and incorporates the maximum into the dynamic render time
estimate.
Part-of: <https://gitlab.gnome.org/GNOME/mutter/-/merge_requests/2500>
A few calculations and assignments are done unnecessarily when the
last next presentation time is invalid. This increases the cognitive
complexity of the function for no reason.
No change in behavior.
Part-of: <https://gitlab.gnome.org/GNOME/mutter/-/merge_requests/2486>
In some hardware configurations, presentation timestamps could be
missing from some page flip events, leading to a temporary loss of
vblank synchronization.
This occurs at least with AMD GPUs for atomic commits that only update
the cursor plane. [0]
In those cases, it's better to calculate the next update time
according to the last valid presentation timestamp instead of relying
on the dispatch lateness.
[0] https://gitlab.freedesktop.org/drm/amd/-/issues/2030
Part-of: <https://gitlab.gnome.org/GNOME/mutter/-/merge_requests/2481>
This can happen with the native backend if the previous frame clock
dispatch didn't result in any KMS update, e.g. because it was triggered
by an input event, but the HW cursor didn't need updating on the stage
view. (This is likely to happen on some out of multiple stage views,
but might be possible even with a single stage view if the cursor isn't
visible)
We would previously delay next_presentation_time_us by one refresh
interval in this case, which could result in spuriously leaving one
refresh cycle unused.
Part-of: <https://gitlab.gnome.org/GNOME/mutter/-/merge_requests/2498>
In addition to the presented callback time, it shows the time to the
reported presentation time (which can be earlier or later than the
presented callback), as well as the GPU rendering duration.
Part-of: <https://gitlab.gnome.org/GNOME/mutter/-/merge_requests/1928>
This ensures they remain perfectly smooth regardless of how the
dispatch time has been adjusted/optimized/delayed/jittered.
Idea by Ivan Molodetskikh <yalterz@gmail.com>
For example, dragging a window on a 60Hz monitor:
BEFORE
delta(time_us) = 17014μs
delta(time_us) = 15998μs
delta(time_us) = 17006μs
delta(time_us) = 16975μs
delta(time_us) = 16001μs
delta(time_us) = 17002μs
delta(time_us) = 17006μs
delta(time_us) = 16004μs
AFTER
delta(time_us) = 16667μs
delta(time_us) = 16667μs
delta(time_us) = 16670μs
delta(time_us) = 16667μs
delta(time_us) = 16669μs
delta(time_us) = 16668μs
delta(time_us) = 16664μs
delta(time_us) = 16674μs
Caveat 1: Because we don't know a "next presentation time" on the first
frame, the interval between the first and second frame will usually be
different to the subsequent steady interval. So this change increases the
jitter of just frame 2, but eliminates jitter thereafter.
Caveat 2: `clutter_frame_clock_schedule_update_now` schedules updates
earlier than `clutter_frame_clock_schedule_update`. This means potentially
you could get multiple frames targeting the same "next presentation time".
That doesn't really change here though - we're dispatching at the same
times as we used to and just giving timelines a better vsync-aligned
timestamp now.
Closes: https://gitlab.gnome.org/GNOME/mutter/-/issues/25
Part-of: <https://gitlab.gnome.org/GNOME/mutter/-/merge_requests/2161>
