-
Notifications
You must be signed in to change notification settings - Fork 2.4k
Description
Hi,
I’m currently studying Taichi’s sparse data structures (SNode: pointer/bitmasked/etc.)
and trying to evaluate their performance and semantic behavior on real workloads.
However, after going through:
- official examples
- Taichi benchmark suite
I noticed that most (if not all) of them are implemented with fully dense SNode layouts.
In practice, many of these examples cannot be converted to sparse layouts
by simply replacing dense with pointer/bitmasked,
because doing so may introduce semantic changes
(e.g. iteration domain changes, implicit zero values, activation/deactivation behavior).
So I would like to ask:
-
Are there any official sparse benchmarks or examples
that are designed to natively use sparse SNodes?
(i.e., sparsity is part of the original algorithm, not an afterthought) -
If not, is there a recommended subset of existing examples
where switching from dense to sparse is considered semantically equivalent?
I believe having even a small set of official sparse-native benchmarks
would be extremely helpful for users who want to study or optimize sparsity.
Thanks!
Metadata
Metadata
Assignees
Labels
Type
Projects
Status