renderCallback (fired on any libghostty wakeup) always called
drawFrame at the end, on top of the 125 FPS animation timer that
runs whenever custom-shader-animation is engaged. Under busy
terminal output the wakeup rate stacks on top of the timer rate,
multiplying the per-frame Wayland commit + GPU submit work on
both threads with no visible benefit (the timer already presents
within DRAW_INTERVAL of the cell rebuild).
When draw_active, the timer is the authoritative present clock —
let renderCallback do updateFrame (cells need rebuilding regardless)
but skip drawFrame and trust the next timer tick to pick up the
new cells. When draw_active is false (no shader animation), wakeups
are the only present trigger so the draw still has to happen here.
Cuts host-thread CPU substantially under high-output scenarios
(seq, yes, log tailing) — bounded total drawFrame rate at the
animation interval instead of summing wakeups + timer ticks.
Co-Authored-By: claude-flow <ruv@ruv.net>
renderCallback early-returns while !flags.visible to avoid the
cell rebuild for hidden surfaces (tab switch, minimize, etc.).
The .visible → true mailbox handler now runs updateFrame before
drawFrame so the first frame after re-show isn't stale.
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
The draw timer should only be activated in case a custom shader
is configured and the option custom-shader-animation is either always
or true. If the timer is kept alive CPU usage spiked enourmously.
Fixes#10667
We need this to get info about the padding, even if we do derive the
grid and screen size separately.
In the future this should possibly be changed to a message that only
sends the padding info and nothing else.
This commit is very large, representing about a month of work with many
interdependent changes that don't separate cleanly in to atomic commits.
The main change here is unifying the renderer logic to a single generic
renderer, implemented on top of an abstraction layer over OpenGL/Metal.
I'll write a more complete summary of the changes in the description of
the PR.
This commit changes a LOT of areas of the code to use decl literals
instead of redundantly referring to the type.
These changes were mostly driven by some regex searches and then manual
adjustment on a case-by-case basis.
I almost certainly missed quite a few places where decl literals could
be used, but this is a good first step in converting things, and other
instances can be addressed when they're discovered.
I tested GLFW+Metal and building the framework on macOS and tested a GTK
build on Linux, so I'm 99% sure I didn't introduce any syntax errors or
other problems with this. (fingers crossed)
Related to #3224
Previously, Ghostty used a static API for async event handling: io_uring
on Linux, kqueue on macOS. This commit changes the backend to be dynamic
on Linux so that epoll will be used if io_uring isn't available, or if
the user explicitly chooses it.
This introduces a new config `async-backend` (default "auto") which can
be set by the user to change the async backend in use. This is a
best-effort setting: if the user requests io_uring but it isn't
available, Ghostty will fall back to something that is and that choice
is up to us.
Basic benchmarking both in libxev and Ghostty (vtebench) show no
noticeable performance differences introducing the dynamic API, nor
choosing epoll over io_uring.
This sets the macOS QoS class of the renderer thread. Apple
recommends[1] that all threads should have a QoS class set, and there
are many benefits[2] to that, mainly around power management moreso than
performance I'd expect.
In this commit, I start by setting the QoS class of the renderer thread.
By default, the renderer thread is set to user interactive, because it
is a UI thread after all. But under some conditions we downgrade:
- If the surface is not visible at all (i.e. another window is fully
covering it or its minimized), we set the QoS class to utility. This
is lower than the default, previous QoS and should help macOS
unschedule the workload or move it to a different core.
- If the surface is visible but not focused, we set the QoS class to
user initiated. This is lower than user interactive but higher than
default. The renderer should remain responsive but not consume as
much time as it would if it was user interactive.
I'm unable to see any noticable difference in anything from these
changes. Unfortunately it doesn't seem like Apple provides good tools to
play around with this.
We should continue to apply QoS classes to our other threads on macOS.
[1]: https://developer.apple.com/documentation/apple-silicon/tuning-your-code-s-performance-for-apple-silicon?preferredLanguage=occl
[2]: https://blog.xoria.org/macos-tips-threading/
Before this fix, if vsync was on the GPU cells buffer could be cleared
for a frame while resizing the terminal down. This was due to the fact
that the surface sent messages for the resize to both the renderer and
the IO thread. If the renderer thread was processed first then the GPU
cells buffer(s) would be cleared and not rebuilt, because the terminal
state would be larger than the GPU cell buffers causing updateFrame to
bail out early, leaving empty cell buffers.
This fixes the problem by changing the origin of the renderer's resize
message to be the IO thread, only after properly updating the terminal
state, to avoid clearing the GPU cells buffers at a time they can't be
successfully rebuilt.
Nathan