From 80x to 385x: A Best-Matching-Unit Search at the L2 Roof, Measured Against a Symmetrically Tuned Baseline

The researchers performed symmetry optimization on the sparse binarization algorithm SparseBin and its benchmark algorithm cuSPARSE through four optimization techniques: Tile size, tile-membership clustering, neuron-axis chunking, and vectorised loads. Within the range of map sizes from 32x32 to 512x512, the speed of optimal matching cell search increased by 5.6-10.1 times per iteration, and the performance gap compared to the MEDLINE graph improved from approximately 80 times to 385 times. The optimized cuSPARSE kernel achieved 77% of its peak performance under L2 bandwidth constraints, while other cells achieved 40-65%. Since any further optimization is limited by this bottleneck, the expected gain is only 1.3 times, and other untested optimization techniques were found to be ineffective or have reached their limits.