A toy for building a KNN graph on GPU.
1. Algorithm:
Based on the idea of NNDescent: A point's neighbor's neighbor is likely its neighbor.
a. Initalize a K-degree graph randomly
b. Repeat n times:
For every point in the graph
select K nearest points from neighbors and neighbors' neighbors.
2. Setup
$ git clone --recursive https://github.com/KinglittleQ/cuda-knng.git
$ mkdir build && cd build
$ cmake ..
$ make -j
3. Download SIFT1M dataset
See here http://corpus-texmex.irisa.fr/
4. Build a KNN graph and search on it
$ ./test_knng data_dir KG L iters
data_dir: directory of SIFT1M dataset
KG: 'K' in KNN graph, number of links per node
L: maximum length of the search path
iters: number of iterations
5. Performance
Time for building a KNN graph on SIFT1M with K = 128 (running on GPU): 861 s
Searching on the graph we built above with L = 200 (running on CPU with one thread):
- Speed: 4.233 ms/query
- Recall@100: 0.987
6. Core code
include/knng/distance.cuh
include/knng/priority_queue.cuh
include/knng/graph.cuh