VW-SDK: Efficient Convolutional Weight Mapping Using Variable Windows for Processing-In-Memory Architectures
This work addresses efficiency bottlenecks for CNN inference on PIM hardware, offering incremental improvements over existing mapping algorithms.
The paper tackles the problem of inefficient weight mapping for convolutional neural network inference on processing-in-memory (PIM) architectures by proposing a variable-window mapping algorithm (VW-SDK), which improves inference speed by 1.69x compared to prior methods in simulations with ResNet-18.
With their high energy efficiency, processing-in-memory (PIM) arrays are increasingly used for convolutional neural network (CNN) inference. In PIM-based CNN inference, the computational latency and energy are dependent on how the CNN weights are mapped to the PIM array. A recent study proposed shifted and duplicated kernel (SDK) mapping that reuses the input feature maps with a unit of a parallel window, which is convolved with duplicated kernels to obtain multiple output elements in parallel. However, the existing SDK-based mapping algorithm does not always result in the minimum computing cycles because it only maps a square-shaped parallel window with the entire channels. In this paper, we introduce a novel mapping algorithm called variable-window SDK (VW-SDK), which adaptively determines the shape of the parallel window that leads to the minimum computing cycles for a given convolutional layer and PIM array. By allowing rectangular-shaped windows with partial channels, VW-SDK utilizes the PIM array more efficiently, thereby further reduces the number of computing cycles. The simulation with a 512x512 PIM array and Resnet-18 shows that VW-SDK improves the inference speed by 1.69x compared to the existing SDK-based algorithm.