* Support sharing `Pipeline` state between `TextAtlas`
* Keep using `Vec` for pipeline cache
* Use `OnceCell` to keep `Pipeline` private
* Revert "Use `OnceCell` to keep `Pipeline` private"
This reverts commit 4112732b1734a3bb6b915d2103e699ef549b77c1.
* Rename `Pipeline` type to `Cache`
The latest `cosmic-text` release performs bucketing calculations after
layouting. This allows us to provide proper subpixel offsets, as well as
scale any buffer thanks to its linear layout properties.
See https://github.com/pop-os/cosmic-text/pull/143.
The `TextAtlas` will have an initial size of `256` (we could make this
configurable) and `try_allocate` will keep track of the glyphs in use
in the current frame, returning `None` when the LRU glyph is in use.
In that case, `TextRenderer::prepare` will return
`PrepareError::AtlasFull` with the `ContentType` of the atlas that is
full. The user of the library can then call `TextAtlas::grow` with the
provided `ContentType` to obtain a bigger atlas (by `256`).
A `TextAtlas::grow` call clears the whole atlas and, as a result, all of
the `prepare` calls need to be repeated in a frame until they all
succeed. Overall, the atlas will rarely need to grow and so the calls
will not need to be repated often.
Finally, the user needs to call `TextAtlas::trim` at the end of the
frame. This allows us to clear the glyphs in use collection in the atlas. Maybe
there is a better way to model this in an API that forces the user to
trim the atlas (e.g. make `trim` return a new type and changing `prepare` and `render` to take that type instead).
Instead, issue a `write_texture` for every glyph that isn't present in
the atlas. I think this should be fine, since cache eviction should be
rare and, therefore, uploads will be too.
The result here is lower memory usage for bigger caches.
I noticed that the `RecentlyUsed` map was actually a `RecentlyInserted`
map. This caused panics when trying to reduce the initial texture atlas size and
dynamically increase it (coming in another PR!).
