Towards Efficient Multi-Scale Deformable Attention on NPU
This work addresses efficiency problems for visual tasks on NPUs, but it is incremental as it focuses on hardware-specific optimizations.
The paper tackled the optimization challenges of multi-scale deformable attention on NPUs by presenting a co-design approach for memory access and computation, achieving up to 7.3x speedup in end-to-end training over baselines.
Multi-scale deformable attention (MSDA) is a flexible and powerful feature extraction mechanism for visual tasks, but its random-access grid sampling strategy poses significant optimization challenges, especially on domain-specific accelerators such as NPUs. In this work, we present a co-design approach that systematically rethinks memory access and computation strategies for MSDA on the Ascend NPU architecture. With this co-design approach, our implementation supports both efficient forward and backward computation, is fully adapted for training workloads, and incorporates a suite of hardware-aware optimizations. Extensive experiments show that our solution achieves up to $5.9\times$ (forward), $8.9\times$ (backward), and $7.3\times$ (end-to-end training) speedup over the grid sample-based baseline, and $1.9\times$, $2.4\times$, and $2.0\times$ acceleration over the latest vendor library, respectively.