Run the NN-descent per-iteration host update on the calling thread

[bdice]

GNND::build (both the dense and the BBQ overload) started and joined a std::thread on every iteration to run update_and_sample (update_graph + sample_graph) while the GPU ran add_reverse_edges + local_join. Each of those threads is a new OpenMP root thread, so its parallel regions ran on a second OpenMP team next to the calling thread's team. That puts two full teams on the same cores, and the idle workers of one team spin (libomp's default KMP_BLOCKTIME is 200 ms) while the other team works. For small and medium graphs (CAGRA and HNSW-ACE partition builds, HNSW upper layers, batched all-neighbors), each iteration was dominated by thread start-up and oversubscription rather than GPU work: in nsys traces of the tests, 2.3-5.6 ms per iteration against 0.3-1.1 ms of GPU time.

This PR changes only cpp/src/neighbors/detail/nn_descent.cuh. Both build overloads now enqueue the iteration's GPU work first and then call update_and_sample on the calling thread.

• The enqueue sequence is fully asynchronous, so the host work still overlaps the GPU work.
• The host and device halves of an iteration touch disjoint buffers, and the synchronization points are unchanged. The output is the same; run-to-run variation from local_join's lock-ordered inserts is unchanged.
• An exception thrown by local_join (for example a launch error or the __CUDA_ARCH__ < 700 check) while the helper thread was joinable used to end in std::terminate. It now propagates to the caller.

On the code before this change, fewer OpenMP threads did not help wall time (OMP_NUM_THREADS=8 was slightly slower than 36, 4 was 60% slower), but KMP_BLOCKTIME=0 was 12% faster with a quarter of the CPU time. That pointed at the two competing thread teams rather than the size of the host work.

Measurements

Single-process wall time of each test executable on an RTX 6000 Ada with a 36-thread host, before vs. after this change. Builds were run alternately, 2 or more repetitions each; mean of the runs.

executable	before	after	change
NEIGHBORS_ANN_NN_DESCENT_TEST	52.6 s	33.7 s	-36.0%
NEIGHBORS_ALL_NEIGHBORS_TEST	28.3 s	20.8 s	-26.4%
NEIGHBORS_ANN_CAGRA_FLOAT_UINT32_TEST	77.5 s	66.8 s	-13.8%
NEIGHBORS_ANN_CAGRA_HALF_UINT32_TEST	44.9 s	37.8 s	-15.7%
NEIGHBORS_ANN_CAGRA_INT8_UINT32_TEST	67.8 s	60.4 s	-10.9%
NEIGHBORS_ANN_CAGRA_UINT8_UINT32_TEST	70.8 s	63.0 s	-11.1%
NEIGHBORS_ANN_CAGRA_BBQ_UINT32_TEST	3.8 s	1.9 s	-51.2%
NEIGHBORS_ANN_CAGRA_FILTER_UDF_TEST	3.2 s	1.8 s	-41.8%
NEIGHBORS_ANN_CAGRA_MERGE_TEST	1.6 s	1.1 s	-33.4%
NEIGHBORS_ANN_CAGRA_TEST_BUGS	2.7 s	1.9 s	-28.2%
NEIGHBORS_ANN_HNSW_ACE_FLOAT_UINT32_TEST	6.7 s	4.3 s	-36.1%
NEIGHBORS_ANN_HNSW_ACE_HALF_UINT32_TEST	9.4 s	6.3 s	-32.6%
NEIGHBORS_ANN_HNSW_ACE_INT8_UINT32_TEST	7.3 s	3.9 s	-46.4%
NEIGHBORS_ANN_HNSW_ACE_UINT8_UINT32_TEST	7.1 s	4.2 s	-40.9%
total	383.9 s	308.1 s	-19.7%

All of these tests passed in every run, and the full ctest suite passes.

[copy-pr-bot]

Auto-sync is disabled for draft pull requests in this repository. Workflows must be run manually.

Contributors can view more details about this message here.

[bdice]

Closing in favor of #2754.