| Summary: | Deep Convolutional Neural Networks (CNN) have achieved state-of-the-art recognition accuracy in a wide range of computer vision applications like image classification, object detection, semantic segmentation etc. Applications based on CNN require millions of multiply-accumulate (MAC) operations to be performed between input pixels and kernel weights during inference. This work investigates a technique, which can be used to eliminate redundant multiplications for a subset of kernel weights in a CNN layer by utilizing identical and/or similar inter-kernel weights (IKW) across kernels. In this work, IKW technique is used to identify identical and/or similar inter-kernel weights in trained, unpruned/pruned, quantized CNN kernels before inference phase. After identification of identical and/or similar inter-kernel weights, a subset of kernel weights termed non-pivot kernel weights are made zero, the other subset called pivot kernel weights are left unchanged. The multiplication corresponding to non-pivot kernel weights are eliminated, thus reducing computations. The products corresponding to non-pivot kernel weights are supplied by multiplication operation of pivot kernel weights, and hence causing no degradation in inference accuracy. Through experiments on state-of-the-art CNNs, we demonstrate that application of IKW technique enhances kernel sparsity by 9-37% for 8-bit precision kernel weight and 18-43% for 4-bit precision kernel weight without degrading the recognition accuracy of the CNN model. Enhanced kernel sparsity can be used to save power by clock gating the compute unit, or increase execution performance by skipping computations pertaining to zero valued non-pivot kernel weights. In addition, power savings are achieved by eliminating redundant power expensive fixed-point multiplication operations. The practical utility of the IKW technique is demonstrated by mapping it to well-known state-of-the-art CNN accelerator architectures. Mapping of the IKW technique on existing CNN accelerator architectures shows reduction in power by at least 12% for 8-bit precision and 19% for 4-bit precision kernel weight. Improvement in execution performance by at least 2% for 8-bit precision and 13% for 4-bit precision kernel weight is observed.
|
|---|
