Appendix

Taking the gather-conv2 operation as an example, the input of this operation is the activation with the shape ofCin H W, where Cin is the number of input channels, andH and W are the resolution of the feature map. The tile shape has 4 dimensions. The estimation of the latency for data movement requires us to model thememory system ofahardwaredevice. The latencyof data movement to local memory is estimated by its bandwidthandefficiency. And due tothe small amount of data transmitted each time and the data israndomly distributed, the efficiencyof datamovementwillbelow.