## About The Pull Request
This is a follow up patch for #92508, otherwise than said in coderbus,
calling the cleanup too early can lead to bad assets. We'll now keep
track of the count of assets in generation and only clean up after they
all have passed the generation.
<img width="3067" height="455" alt="grafik"
src="https://github.com/user-attachments/assets/0b65acf3-464f-436c-8a60-84c9472be6cd"
/>
## Why It's Good For The Game
## Changelog
There should be nothing player facing here as the asset job fails
straight up if the cache was cleared too early.
## About The Pull Request
The entire asset loading had multiple issues.
Firstly, we could hit a tick check after an asset was generated and
didn't remove the index, leading to the asset being generated twice.
Secondly, we had the issue that the icon forge batching uses UNTIL to
wait until generation finishes. This led to the situation that the
entire subsystem fire proc was paused outside of the internal tick check
pause.
## Why It's Good For The Game
Especially on larger sprite sheets or when there're many to be generated
this led to the situation that we looked up a rust job multiple times:
`[2025-08-08T18:42:11] Runtime in
code/modules/asset_cache/iconforge/batched_spritesheet.dm,200:
Spritesheet design UNKNOWN ERROR: NO SUCH JOB`
## Changelog
No player facing changes.
## About The Pull Request
Replaces the asset subsystem's spritesheet generator with a rust-based
implementation (https://github.com/tgstation/rust-g/pull/160).
This is a rough port of
https://github.com/BeeStation/BeeStation-Hornet/pull/10404, but it
includes fixes for some cases I didn't catch that apply on TG.
(FWIW we've been using this system on prod for over a year and
encountered no major issues.)
### TG MAINTAINER NOTE
### Batched Spritesheets
`/datum/asset/spritesheet_batched`: A version of the spritesheet system
that collects a list of `/datum/universal_icon`s and sends them off to
rustg asynchronously, and the generation also runs on another thread, so
the game doesn't block during realize_spritesheet. The rust generation
is about 10x faster when it comes to actual icon generation, but the
biggest perk of the batched spritesheets is the caching system.
This PR notably does not convert a few things to the new spritesheet
generator.
- Species and antagonist icons in the preferences view because they use
getFlatIcon ~~which can't be converted to universal icons~~.
- Yes, this is still a *massive* cost to init, unfortunately. On Bee, I
actually enabled the 'legacy' cache on prod and development, which you
can see in my PR. That's why I added the 'clear cache' verb and the
`unregister()` procs, because it can force a regeneration at runtime. I
decided not to port this, since I think it would be detrimental to the
large amount of contributors here.
- It is *technically* possible to port parts of this to the uni_icon
system by making a uni_icon version of getFlatIcon. However, some
overlays use runtime-generated icons which are ~~completely unparseable
to IconForge, since they're stored in the RSC and don't exist as files
anywhere~~. This is most noticeable with things like hair (which blend
additively with the hair mask on the server, thus making them invisible
to `get_flat_uni_icon`). It also doesn't help that species and antag
icons will still need to generate a bunch of dummies and delete them to
even verify cache validity.
- It is actually possible to write the RSC icons to the filesystem
(using fcopy) and reference them in IconForge. However, I'm going to
wait on doing this until I port my GAGS implementation because it
requires GAGS to exist on the filesystem as well.
#### Caching
IconForge generates a cache based on the set of icons used, all
transform operations applied, and the source DMIs of each icon used
within the spritesheet. It can compare the hashes and invalidate the
cache automatically if any of these change. This means we can enable
caching on development, and have absolutely no downsides, because if
anything changes, the cache invalidates itself.
The caching has a mean cost of ~5ms and saves a lot of time compared to
generating the spritesheet, even with rust's faster generation. The main
downside is that the cache still requires building the list of icons and
their transforms, then json encoding it to send to rustg.
Here's an abbreviated example of a cache JSON. All of these need to
match for the cache to be valid. `input_hash` contains the transform
definitions for all the sprites in the spritesheet, so if the input to
iconforge changes, that hash catches it. The `sizes` and `sprites` are
loaded into DM.
```json
{
"input_hash": "99f1bc67d590e000",
"dmi_hashes": {
"icons/ui/achievements/achievements.dmi": "771200c75da11c62"
},
"sizes": [
"76x76"
],
"sprites": {
"achievement-rustascend": {
"size_id": "76x76",
"position": 1
}
},
"rustg_version": "3.6.0",
"dm_version": 1
}
```
### Universal Icons
Universal icons are just a collection of DMI, Icon State, and any icon
transformation procs you apply (blends, crops, scales). They can be
convered to DM icons via `to_icon()`. I've included an implementation of
GAGS that produces universal icons, allowing GAGS items to be converted
into them. IconForge can read universal icons and add them to
spritesheets. It's basically just a wrapper that reimplements BYOND icon
procs.
### Other Stuff
Converts some uses of md5asfile within legacy spritesheets to use
rustg_hash_file instead, improving the performance of their generation.
Fixes lizard body markings not showing in previews, and re-adds eyes to
the ethereal color preview. This is a side effect of IconForge having
*much* better error handling than DM icon procs. Invalid stuff that gets
passed around will error instead of silently doing nothing.
Changes the CSS used in legacy spritesheet generation to split
`background: url(...) no-repeat` into separate props. This is necessary
for WebView2, as IE treats these properties differently - adding
`background-color` to an icon object (as seen in the R&D console) won't
work if you don't split these out.
Deletes unused spritesheets and their associated icons (condiments
spritesheet, old PDA spritesheet)
## Why It's Good For The Game
If you press "Character Setup", the 10-13sec of lag is now approximately
0.5-2 seconds.
Tracy profile showing the time spent on get_asset_datum. I pressed the
preferences button during init on both branches. Do note that this was
ran with a smart cache HIT, so no generation occurred.
Much lower worst-case for /datum/asset/New (which includes
`create_spritesheets()` and `register()`)
Here's a look at the internal costs from rustg - as you can see
`generate_spritesheet()` is very fast:
### Comparison for a single spritesheet - chat spritesheet:
**Before**
**After**
## Changelog
🆑
fix: Fixed lizard body markings and ethereal feature previews in the
preference menu missing some overlays.
refactor: Optimized spritesheet asset generation greatly using rustg
IconForge, greatly reducing post-initialization lag as well as reducing
init times and saving server computation.
config: Added 'smart' asset caching, for batched rustg IconForge
spritesheets. It is persistent and suitable for use on local, with
automatic invalidation.
add: Added admin verbs - Debug -> Clear Smart/Legacy Asset Cache for
spritesheets.
fix: Fixed R&D console icons breaking on WebView2/516
/🆑
* Adds lazyloading to the asset subsystems
This currently applies only to spritesheets because of how monumentally
expensive they are.
If an asset is requested it will immediately be fully loaded, but
otherwise we slowly load them in with a separate subsystem.
This allows us to not hold up initialize with hair stuff. Saves roughly
33% (16 seconds with LOW_MEMORY_MODE) of initialize on my machine
My target is something closer to the 9 second init that had back in
2019, this is a good first step. Lets see how much more we can do yeah
lads?
Co-authored-by: san7890 <the@san7890.com>
