Exploiting convolutional neural networks for point cloud  processing is quite challenging, due to the inherent irregular  distribution and discrete shape representation of point  clouds.  To address these problems, many handcrafted convolution  variants have sprung up in recent years.  Though  with elaborate design, these variants could be far from optimal  in sufficiently capturing diverse shapes formed by discrete  points.  In this paper, we propose PointSeaConv, i.e.,  a novel differential convolution search paradigm on point  clouds.  It can work in a purely data-driven manner and  thus is capable of auto-creating a group of suitable convolutions  for geometric shape modeling.  We also propose  a joint optimization framework for simultaneous search of  internal convolution and external architecture, and introduce  epsilon-greedy algorithm to alleviate the effect of discretization  error.  As a result, PointSeaNet, a deep network  that is sufficient to capture geometric shapes at both convolution  level and architecture level, can be searched out  for point cloud processing.  Extensive experiments strongly  evidence that our proposed PointSeaNet surpasses current  handcrafted deep models on challenging benchmarks  across multiple tasks with remarkable margins.