Basic cpu iris search using HNSW #327
Conversation
| @@ -63,6 +63,32 @@ impl<N: NetworkTrait> IrisWorker<N> { | |||
| Ok(res) | |||
| } | |||
|
|
|||
| pub(crate) async fn rep3_dot_non_blocking( | |||
| &mut self, | |||
| a: &VecShare<u16>, | |||
There was a problem hiding this comment.
I think also the whole VecShare type can probably go here, since it is just a wrapper around Vec<Share<T>>. I think this was leftover from some experiments on holding the internal shares as two separate vectors of all a and all b sub-shares.
There was a problem hiding this comment.
Not sure what you mean here. There are some associated functions implemented for VecShare, which would make things a bit more difficult to change all VecShare to Vec<Share<_>>
Or do you mean passing a: VecShare<u16> instead of a: &VecShare<u16>?
There was a problem hiding this comment.
Would gladly change this to https://docs.rs/ndarray/latest/ndarray/index.html instead 😄 so we can get other vector operations for free hehe
| (*lock) | ||
| .rep3_lift_and_cross_mul(d1, t1 as u32, d2, t2 as u32) | ||
| .unwrap() |
There was a problem hiding this comment.
I think mixing async and sync stuff here is not what we want. I would spend some time and unify the networking stuff into only async calls and then not mix them.
There was a problem hiding this comment.
I agree, will refactor this once we merge this in and grab this one: #354
Should solve all our issues related to locking
|
So I guess the plan is to merge this and then build on it, improve the networking etc? |
|
That's correct 👍 |
Adds prototype of HNSW search using CPU.
Concretely:
IrisWorkerstruct. In the future I believe all these functions are going to be async - to avoid calling blocking code.