Use raft::copy in the inner loop of the host-side raft::gather #2464
Conversation
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
achirkin
added
enhancement
achirkin
added
improvement
|
Converting this to a draft until we properly benchmark the change. |
tarang-jain
reviewed
Oct 8, 2024
tarang-jain
reviewed
Oct 8, 2024
| for (IdxT k = 0; k < buff.extent(1); k++) { | ||
| buff(i, k) = dataset(in_idx, k); | ||
| } | ||
| raft::copy(res, |
There was a problem hiding this comment.
So the resource is actually unused since the src and dest are both in host mem? I wonder if we can create a resource on the fly and supply it to raft::copy() rather than adding the resource to gather_buff's signature.
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
Use
raft::copyinstead of an assignment operator loop in the host-side gather operation. This should allow to leverage the logic ofraft::copyto use fast memcpy when possible (especially for high-dimensional datasets).As some very circumstantial evidence, this change is observed to reduce
raft::matrix::sample_rows(100000000, 128)time from approximately 5 to 4 minutes (bigann-100M int8 dataset in RAM, 16-core machine).Make the NVTX annotations a little bit more descriptive along the way.