video().track() in the buffer format callback takes the libvlc input lock,
which stop() holds while it waits for the decoder thread running that
callback. Take the video track from the media info only.
* android, desktop: open group member profile without waiting for the core
Tapping a member avatar awaited apiGroupMemberInfo and apiGetGroupMemberCode
before the profile modal was created. Both are single-row queries, 1-2 ms when
the core is idle, but sendCmd is serialized against everything else the core is
doing, so while it is busy - startup, a batch of incoming events, a long database
operation - the tap produces nothing at all until the core drains.
The modal is now opened first and the two values arrive in state the tap handler
creates, so the profile opens at the speed of the UI. The card is shown from the
member that is already known at the tap, preferring the model's copy once it is
there, so the member no longer has to be written to the model before the modal
can be shown.
* plans: open member profile without waiting for the core
* android, desktop: keep member profile rows in place while they load
The security code and the connection stats arrive after the profile is shown, so
"Verify security code" and the Servers section appeared under the already
visible card and pushed the rows below them down.
They are now rendered from the first frame in their final positions, disabled,
and enabled in place when the data arrives - SectionItemView already drops the
clickable modifier when disabled, so nothing moves. Only what can be decided from
the member known at the tap is reserved: "Fix connection" is not, as it applies
only while a ratchet needs syncing and a placeholder for it would disappear on
nearly every open.
* android, desktop: reserve the network status row with the servers
The reserved Servers section had three rows where the loaded one has four, so
everything below it - "Block member" - moved down by one row when the stats
arrived.
Network status looks like agent state that cannot be predicted before loading,
but connSubStatus returns Just whenever the connection has receiving queues, and
that is the same list that decides whether "Receiving via" is rendered. The two
rows appear together, so the status row is now reserved with them.
---------
Co-authored-by: Evgeny Poberezkin <evgeny@poberezkin.com>
* ui: blur media by resampling the preview, not a per-frame effect
Modifier.blur put a BlurEffect on a display-sized graphics layer, so the
Gaussian was re-evaluated on every frame the media was drawn. It also
requires RenderEffect, which Android applies only from API 31, so below
Android 12 media was drawn unblurred while the setting read as on.
A blur and a downscale discard the same thing - detail finer than their
radius - so the preview is now resampled to about one pixel per radius and
stretched back. That runs once when the item composes, needs no
RenderEffect, and removes the detail irreversibly rather than convolving
it. The first resampling step bounds both sides, so no image, however
shaped, can produce a large intermediate.
While the blur hides the media the file is also left unread, so an image
scrolled past is no longer read, decoded at its full size and kept in the
image cache only to be hidden again. blurHidesMedia() is the single
definition that both the drawing and the loading decision use, so they
cannot disagree about whether the media is on screen.
* plans: blur media by resampling the preview
* ui: shorten comments in media blur
* ui: smooth the media blur and match its strength to the old one
Stretching the resampled preview straight to the screen showed its pixel grid. Double it back
up until its longest side is at least half the reference width before drawing, so the last
stretch is short and the tents compose into a bell. The reference width moves from 360 to 400:
fitted against Gaussians, 360 blurred 5-75% more than Modifier.blur did, 400 is within 10% at
every setting.
Skia's raster blur runs at full resolution, so each bar blurred width x (bar + 6 sigma) pixels every frame: about
80 ms per frame for the two bars of a chat at the default radius, which cut scrolling to 5-9 fps on the software
renderer. Scaling the layer cannot help, because a layer's filter is evaluated in device space and the scale is
applied to the sigma too. The bar is now drawn into an offscreen surface up to 8x narrower and blurred there, which
measured 5-6x cheaper and within 0.25/255 of the original.
Blurring a copy means the bar no longer follows the content on its own, so the scroll containers bump a version on
the app bar handler and the bar depends on it.
Co-authored-by: Evgeny Poberezkin <evgeny@poberezkin.com>
In channels members are usually connected via relays, so they are memberCurrent
but not memberActive, and #7255 widened the condition for requesting the member
security code to cover them. It updated GroupChatInfoView and
MemberSupportChatView, but not the same code in ChatView, so "Verify security
code" is not offered when the profile is opened by tapping member avatar in the
chat, while it is offered for the same member in the members list.
Alain Brenzikofer