QuickFlow: An Efficient Local Search Method to Map Convolutions on Spatial Architectures

Efficiently running deep neural networks requires the hardware acceleration of convolutional kernels. Spatial Architectures (SAs) are a natural fit, employing multiple processing elements and a custom memory hierarchy to exploit parallelism and data reuse. In turn, SAs require a mapping to specify data movements and computation order. Thus, specialized mapping tools have been developed to explore the space of possible mappings and retrieve optimal ones, using analytical hardware models for performance feedback. However, for each SA-kernel pair, the mapping space is vast, with significant performance variations arising from subtle interactions between mapping decisions. Therefore, coordinating all problem aspects remains challenging for existing tools, often leading to long execution times. Yet, high-quality mappings must be promptly available to support downstream tasks, like runtime resource allocation and hardware design space exploration. To address this, we propose QuickFlow, a new mapping tool that efficiently finds near-optimal mappings for SAs and AI kernels. In QuickFlow, we build equivalences between mappings based on a detailed analysis of distinct data reuse opportunities. Then, we redesign the optimization paradigm accordingly, comparing tiling and parallelism decisions after quickly selecting their best dataflow up to equivalence, ultimately enabling a single greedy local search to effectively reach near-optimal mappings. QuickFlow also integrates an improved analytical model, supporting arbitrary convolutions through a novel, exact formula for tile sizes. Across our experiments on three SAs and twenty kernels, QuickFlow achieves a 1–2.1× better energy-delay product and up to 182× faster execution time compared to the best results from four state-of-the-art mapping tools.