It's here, the long-foretold and long-procrastinated renderer rework!
Hopefully this makes it easier to adapt and modify the renderer in the
future and ensures feature parity between Metal and OpenGL. Despite
having been a lot of work to write initially, with the abstraction layer
in place I feel like working on the renderer will be a much more
pleasant experience going forward.
## Key points
- CPU-side renderer logic is now mostly unified via a generic
`Renderer`.
- A graphics API abstraction layer over OpenGL and Metal has been
introduced.
- Minimum OpenGL version bumped to `4.3`, so can no longer be run on
macOS; I used the nix VM stuff for my testing during development. (Edit
by @mitchellh: Note for readers that Ghostty still works on macOS, but
the OpenGL backend doesn't, only the Metal one)
- The OpenGL backend now supports linear blending! Woohoo! The default
`alpha-blending` has been updated to `linear-corrected` since it's
essentially a strict improvement over `native`. The default on macOS is
still `native` though to match other mac apps in appearance, since macOS
users are more sensitive to text appearance.
- Custom shaders can now be hot reloaded.
- The background color is once again drawn by us, so custom shaders can
interact with it properly. In general, custom shaders should be a little
more robust.
## The abstraction layer
The general hierarchy of the abstraction layer is as such:
```
[ GraphicsAPI ] - Responsible for configuring the runtime surface
| | and providing render `Target`s that draw to it,
| | as well as `Frame`s and `Pipeline`s.
| V
| [ Target ] - Represents an abstract target for rendering, which
| could be a surface directly but is also used as an
| abstraction for off-screen frame buffers.
V
[ Frame ] - Represents the context for drawing a given frame,
| provides `RenderPass`es for issuing draw commands
| to, and reports the frame health when complete.
V
[ RenderPass ] - Represents a render pass in a frame, consisting of
: one or more `Step`s applied to the same target(s),
[ Step ] - - - - each describing the input buffers and textures and
: the vertex/fragment functions and geometry to use.
:_ _ _ _ _ _ _ _ _ _/
v
[ Pipeline ] - Describes a vertex and fragment function to be used
for a `Step`; the `GraphicsAPI` is responsible for
these and they should be constructed and cached
ahead of time.
[ Buffer ] - An abstraction over a GPU buffer.
[ Texture ] - An abstraction over a GPU texture.
```
More specific documentation can be found on the relevant structures.
## Miscellany
Renderers (which effectively just means the generic renderer) are now
expected to only touch GPU resources in `init`, certain lifecycle
functions such as the `displayRealized`/`displayUnrealized` callbacks
from GTK-- and `drawFrame`; and are also expected to be thread-safe.
This allows the renderer thread to build the CPU-side buffers
(`updateFrame`) even if we can only *draw* from the app thread.
Because of this change, we can draw synchronously from the main thread
on macOS when necessary to always have a frame of the correct size
during a resize animation. This was necessary to allow the background to
be drawn by our GPU code (instead of setting a background color on the
layer) while still avoiding holes during resize.
The OpenGL backend now theoretically has access to multi-buffering, but
it's disabled (by setting the buffer count to 1) because it
synchronously waits for frames to complete anyway which means that the
extra buffers were just a waste of memory.
## Validation
To validate that there are no significant or obvious problems, I
exercised both backends with a variety of configurations, and visually
inspected the results. Everything looks to be in order.
The images are available in a gist here:
https://gist.github.com/qwerasd205/c1bd3e4c694d888e41612e53c0560179
## Memory
Here's a comparison of memory usage for ReleaseFast builds on macOS,
between `main` and this branch.
Memory figures given are values from Activity Monitor measuring windows
of the same size, with two tabs with 3 splits each.
||Before|After|
|-:|-|-|
|**Memory**|247.9 MB|224.2 MB|
|**Real Memory**|174.4 MB|172.5 MB|
Happily, the rework has slightly *reduced* the memory footprint- likely
due to removing the overhead of `CAMetalLayer`. (The footprint could be
reduced much further if we got rid of multi-buffering and satisfied
ourselves with blocking for each frame, but that's a discussion for
another day.)
If someone could do a similar comparison for Linux, that'd be much
appreciated!
## Notes / future work
- There are a couple structures that *can* be unified using the
abstraction layer, but I haven't gotten around to unifying yet.
Specifically, in `renderer/(opengl|metal)/`, there's `cell.zig` and
`image.zig`, both of which are substantially identical between the two
backends. `shaders.zig` may also be a candidate for unification, but
that might be *overly* DRY.
- ~~I did not double-check the documentation for config options, which
may mention whether certain options can be hot-reloaded; if it does then
that will need to be updated.~~ Fixed: be5908f
- The `fragCoord` for custom shaders originates at the top left for
Metal, but *bottom* left for OpenGL; fixing this will be a bit annoying,
since the screen texture is likewise vertically flipped between the two.
Some shaders rely on the fragcoord for things like falling particles, so
this does need to be fixed.
- `Target` should be improved to support multiple types of targets right
now it only represents a framebuffer or iosurface, but it should also be
able to represent a texture; right now a kind of messy tagged union is
used so that steps can accept both.
- Custom shader cursor uniforms (#6912) and terminal background images
(#4226, #5233) should be much more straightforward to implement on top
of this rework, and I plan to make follow-up PRs for them once this is
merged.
- I *do* want to do a rework of the pipelines themselves, since the way
we're rendering stuff is a bit messy currently, but this is already a
huge enough PR as it is- so for now the renderer still uses the same
rendering passes that Metal did before.
- We should probably add a system requirements section to the README
where we can note the minimum required OpenGL version of `4.3`, any even
slightly modern Linux system will support this, but it would be good to
document it somewhere user-facing anyway.
# TODO BEFORE MERGE
- [x] Have multiple people test this on both macOS and linux.
- [ ] ~~Have someone with a better dev setup on linux check for memory
leaks and other problems.~~ (Skipped, will merge and let tip users
figure this out, someone should *specifically* look for memory leaks
before the next versioned release though.)
- [x] Address any code review feedback.
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.
If this was Swift code, we'd be using a strong reference, which would
retain the layer for us and release it when the object is deallocated,
but this is Zig land so we have to do that manually.
NOTE: We don't *have* to do this, but it fits much better with Zig idiom
and hopefully avoids potential future footguns. We should do this to any
autoreleased objects that we persist a reference to in a Zig struct.
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 wasn't working before but it just requires a restart of the machine
for the changes to take effect. The namespace runners have this prebuilt
so this should work now.
The other workaround was flaky for unknown reasons so I'd prefer to go
back to this.
We were depending on $GHOSTTY_RESOURCES_DIR for two reasons:
1. To locate our script-adjacent bash-preexec.sh script
2. To restrict our script's execution to environments in which
$GHOSTTY_RESOURCES_DIR is available (i.e. Ghostty-only shells)
For (1), we can instead determine our directory using $BASH_SOURCE[0].
This is slightly differently than our previous behavior, where we'd
always load bash-preexec.sh from the $GHOSTTY_RESOURCES_DIR hierarchy,
even if ghostty.bash from source from somewhere else on the file system
... but we never relied on that behavior, even in development.
For (2), there's no harm in source'ing this script outside of Ghostty,
and if that does become a concern, we can restore this condition or use
something more targeted based on those specific cases.
Historically, I believe (2) was in place to enable (1), so addressing
(1) removes the need for (2).
And lastly, none of the other shell integration scripts depend on
$GHOSTTY_RESOURCES_DIR.
