Overview

The feature retrieval solution uses the TIK operators to implement the calculation logic on feature distance. The custom operators are as follows.

• Flat distance calculation operator: obtains the distance (L2/IP) between base data and vectors to be retrieved.
• SQ8 distance calculation operator: obtains the distance (L2/IP) between the SQ-quantized base data and the unquantized vectors to be retrieved.
• IVFSQ8 operator: operator required by the IVFSQ8 algorithm.
• INT8Flat distance calculation operator: obtains the distance (L2/COS) between the base data and the vectors to be retrieved, both of which are quantized by INT8.
• IVFSQT operator: distance operator required by the IVFSQT algorithm at three levels.
• FlatAT operator: reduces the time required by train and add operations in the IVF scenario. code_num is equal to nlist.
• FlatInt8AT operator: reduces the time required by train, add, and update operations of IVFSQT in the Atlas inference product.
• AICPU operator: schedules the CPU of the Ascend AI Processor for computing operations such as sorting, fully utilizing hardware performance.
• BinaryFlat operator: operator required by the binarization algorithm.
• Mask operator: mask operator required by the attribute filtering algorithm of spatiotemporal databases.
• IVFSP operator: service operator and AICPU operator required by the IVFSP algorithm, and the training operator required for generating and training the IVFSP codebook.
• VStar operator: service operator and AICPU operator required by the VStar algorithm.
• IVFFLAT: distance operators required by the level-1 and level-2 IVFFLAT algorithms.