Optimizing Memory Efficiency for Convolution Kernels on Kepler GPUs