feat(gpu): implement fhe rand on gpu #1958
Conversation
guillermo-oyarzun
force-pushed
the
go/feature/implement-fherand-on-gpu
branch
from
349cfaa to
9088fb0
Compare
There was a problem hiding this comment.
Hey @guillermo-oyarzun! Thanks a lot for this PR, here comes my review. My main question is about the par_generate... entry points: at the moment the cuda calls are made on the same streams, which makes execution sequential. What we could do is use a different set of streams for each of the par_iter iteration maybe? Wdyt?
guillermo-oyarzun
force-pushed
the
go/feature/implement-fherand-on-gpu
branch
3 times, most recently
from
5d54d65 to
63eb9b8
Compare
guillermo-oyarzun
force-pushed
the
go/feature/implement-fherand-on-gpu
branch
2 times, most recently
from
cd28cff to
29dcb98
Compare
| .into_par_iter() | ||
| .enumerate() | ||
| .map(|(i, seed)| { | ||
| let stream_index = i; |
There was a problem hiding this comment.
Here if there are too many blocks this won't work, will it? We would need something like
let stream_index = i % streams.gpu_indexes.len()
instead
There was a problem hiding this comment.
It will work because before I generate a vector of streams with as many streams as blocks are. This is being executed in a single GPU but with many streams.
There was a problem hiding this comment.
Ah yes of course, thanks for the clarification!
guillermo-oyarzun
force-pushed
the
go/feature/implement-fherand-on-gpu
branch
from
29dcb98 to
17f5d05
Compare
closes: please link all relevant issues
PR content/description
Check-list: