[WIP] Batching haversine to improve concurrency #191
Conversation
cjnolet
added
enhancement
|
Tagging @miguelusque for awareness. The google test is passing. I'm going to do some profiling on your dataset before I mark this one for review. |
|
@miguelusque, after profiling this with better concurrency, it looks like the trigonometry here is definitely the bottleneck. I'm going to improve this further by pruning distance computations. I'm going to start with a modification of the random ball cover algorithm to k-neighbors (for exact search) since it is straightforward to implement on the GPU with our current brute force knn primitive and can be reused to speed up other algorithms which require an exhaustive brute-force on all data samples. |
|
@cjnolet thank you for the update, and for your effort to speed up this metric. Appreciate it! |
|
I've opened #213 as a continuation of this work, which investigates using the random ball cover to reduce the number of distance calculations. |
This updates the notebook link from https://github.com/rapidsai/cuvs/tree/HEAD/exmples (doesn't exist) to https://github.com/rapidsai/cuvs/tree/branch-24.08/notebooks Authors: - Ray Bell (https://github.com/raybellwaves) Approvers: - Corey J. Nolet (https://github.com/cjnolet) URL: rapidsai/cuvs#191
No description provided.