Function Description

The native IVFSP retrieval algorithm of Ascend uses the self-developed matrix approximation policy to compress feature vectors and save them to databases. In addition, an inverted chain policy is designed to select the database that is most likely to contain ground truth. Finally, the self-developed retrieval policy is used to retrieve databases after inverted chain filtering to obtain the top K vector results.

AscendIndexIVFSP supports only the standard-form scenario and Atlas inference product.

Multi-thread concurrent calling is supported. You need to set the environment variable MX_INDEX_MULTITHREAD to 1, that is, export MX_INDEX_MULTITHREAD=1. If this variable is set to another value or is not set, the multi-thread function is disabled. Currently, FeatureRetrieval uses the OMP for performance acceleration. The OMP cannot be used together with other multi-thread mechanisms. If new threads are repeatedly created to use the OMP, the memory usage keeps increasing. Therefore, you are advised to use fixed threads to run retrieval tasks.