SelB-k-NN: A Mini-Batch K-Nearest Neighbors Algorithm on AI Processors

The popularity of Artificial Intelligence (AI) motivates novel domain-specific hardware named AI processors. With a design trade-off, the AI processors feature incredible computation power for matrix multiplications and activations, while some leave other operations less powerful, e.g., scalar operations and vectorized comparisons & selections. For k-nearest neighbors (k-NN) algorithm, consisting of distance computation phase and k-selection phase, while the former is naturally accelerated, the previous efficient k-selection becomes problematic. Moreover, limited memory forces k-NN to adopt a mini-batch manner with tiling technique. As the distance computation's results are the kselection's inputs, the former's tiling shape determines that of the latter. Since the two phases execute on separate hardware units requiring different performance analyses, whether the former's tiling strategies benefit the latter and entire k-NN is doubtful. To address the new challenges brought by the AI processors, this paper proposes SelB-k-NN (Selection-Bitonic-k-NN), a minibatch algorithm inspired by selection sort and bitonic k-selection. SelB-k-NN avoids the expansion of the weakly-supported operations on the huge scale of datasets. To apply SelB-k-NN to various AI processors, we propose two algorithms to reduce the hardware support requirements. Since the matrix multiplication operates data with the specifically-designed memory hierarchy which kselection does not share, the tiling shape of the former cannot guarantee the best execution of the latter and vice versa. By quantifying the runtime workload variations of k-selection, we formulate an optimization problem to search for the optimal tiling shapes of both phases with an offline pruning method, which reduces the search space in the preprocessing stage. Evaluations show that on Huawei Ascend 310 AI processor, SelB-k-NN achieves 2.01x speedup of the bitonic k-selection, 23.93x of the heap approach, 78.52x of the CPU approach. For mini-batch SelB-k-NN, the optimal tiling shapes for two phases respectively achieve 1.48x acceleration compared with the matrix multiplication tiling shapes and 1.14x with the k-selection tiling shapes, with 72.80% of the search space pruned.