Tetris-SDK: Efficient Convolution Layer Mapping with Adaptive Windows for Fast In Memory Computing
DescriptionShifted-and-Duplicated-Kernel (SDK) mapping is gaining popularity for substantially accelerating CNN Networks in Compute-In-Memory (CIM) architectures compared to conventional image-to-column (im2col). However, the state-of-the-art SDK algorithm, i.e., Variable-Window-SDK (VW-SDK) suffers from lack of adaptability, leading to in-sufficient memory utilization and extra processing cycles. In this work, we propose Tetris-SDK, an enhanced strategy integrating a marginal-space mapping to increase CIM array utilization, an adjustable input channel partition to improve adaptation, and a square-inclined window slicing to decrease overall computing cycles. Compared with im2col, SDK and VM-SDK, Tetris-SDK speeds up a variety of CNN layers by up to 20x, 8x, and 1.29x, respectively.
TimeTuesday, July 11th6:00pm - 7:00pm PDT
LocationLevel 2 Lobby