Julian Miller

Enable vectorized global loads for the reduction algorithms

1

Vectorization is performance critical on SIMD architectures. This patch enables vectorization by unrolling vector size wide loop iterations on both coalesced (commutative algorithms) and consecutive (non-commutative algorithms) loads. Coalesced loads...

This reverts commit c02f915d42ac9bd1a3c4f2215d703b49b3a2173d since this Coverity issue is a false positive. `EventImpl` is always set when we pass it in the runtime, but the previous change suggests it might...

[TASK] Reuse `all_neighbors` APIs in CAGRA ACE build

2

The CAGRA augmented core extraction (ACE) build method introduced in PR #1404 supports building CAGRA indices on very large datasets that exceed GPU memory capacity. To this end, it partitions...

Multiple routines use `raft::matrix::sample_rows()` followed by a balanced `cuvs::cluster::kmeans::fit()` including `all_neighbors::get_centroids_on_data_subsample()`, `ivf_pq::build()`, `scann::build()`, and ACE introduced in #1404. Testings this PR with BIGANN 1B and 1% (10M) samples shows high...