Efficient Deformable Attention Acceleration via Pruning-Assisted Grid-Sampling and Multi-Scale Parallel Processing

DEFA proposes algorithm-architecture co-design for efficient MSDeformAttn acceleration through pruning and parallel processing.