Faster transform propagation release notes #2041
Conversation
release-content/0.16/release-notes/17840_Parallel_Transform_Propagation.md
Outdated
Show resolved
Hide resolved
release-content/0.16/release-notes/17840_Parallel_Transform_Propagation.md
Show resolved
Hide resolved
JMS55
left a comment
JMS55
left a comment
There was a problem hiding this comment.
I like the level of technical detail, I think it strikes a good balance between interesting and understandable.
release-content/0.16/release-notes/17840_Parallel_Transform_Propagation.md
Outdated
Show resolved
Hide resolved
| In many cases, this is the overwhelming majority of objects: level geometry and props are not typically moving around each frame! | ||
| We're now propagating a "dirty bit" up the hierarchy towards ancestors; allowing transform propagation can ignore entire subtrees of the hierarchy if they encounter an entity without the dirty bit. | ||
|
|
||
| The results speak for themselves: taken together, our testing on the incredibly beefy [Caldera Hotel] from Call of Duty: Warzone shows that transform propagation took 1.1 ms in 0.15, and 0.1 ms after these changes in 0.16. |
There was a problem hiding this comment.
We should note that Caldera is entirely static.
Tbh we should probably check how much perf impact we had on entirely dynamic scenes...
There was a problem hiding this comment.
Yeah I can test many_foxes or something.
There was a problem hiding this comment.
It is significantly faster thanks to parallel transform propagation.
There was a problem hiding this comment.
3-4x faster was observed on caldera in the parallel propagation PR - this is without static subtree optimizations, with all transforms being recomputed every frame.
3-10x faster was observed using the hierarchy stress tests: bevyengine/bevy#17840 (comment)
The static optimizations did regress some benchmarks, but they were still either on par or faster than 0.15.
release-content/0.16/release-notes/17840_Parallel_Transform_Propagation.md
Outdated
Show resolved
Hide resolved
release-content/0.16/release-notes/17840_Parallel_Transform_Propagation.md
Outdated
Show resolved
Hide resolved
release-content/0.16/release-notes/17840_Parallel_Transform_Propagation.md
Outdated
Show resolved
Hide resolved
release-content/0.16/release-notes/17840_Parallel_Transform_Propagation.md
Show resolved
Hide resolved
|
|
||
| The results speak for themselves: taken together, our testing on the incredibly beefy [Caldera Hotel] from Call of Duty: Warzone shows that transform propagation took 1.1 ms in 0.15, and 0.1 ms after these changes in 0.16. | ||
| While that's an impressive 11x performance improvement, the absolute magnitude of the time saved is the key metric. | ||
| With about 16 ms per frame at 60 FPS, that's 6% of your *entire* game's CPU budget saved, making huge open worlds or incredibly complex CAD assemblies more viable than ever before. |
There was a problem hiding this comment.
The numbers are more impressive when you consider that for most users, it was taking a much, much longer time, like, 4ms on 0.15. The M4 Max is just stupidly fast. In pcwalton's screenshots, it was taking up ~1/4 of the frame time, which is what really motivated the work.
There was a problem hiding this comment.
It might be nice to find a more representative profile comparison on caldera from 0.15, there were so many optimizations, and the aggregate difference is seriously impressive.
alice-i-cecile
force-pushed
the
faster-transform-prop
branch
from
11d070f to
0109b9f
Compare
3 participants
Please feel free to amend the technical descriptions liberally! I wasn't sure about the right level of detail to provide.
Fixes #1995.