Sparsity-Specific Code Optimization using Expression Trees

We introduce a code generator that converts unoptimized C++ operating on sparse data into vectorized and parallel CPU or GPU kernels. Our approach unrolls the computation massive expression graph, performs redundant elimination, grouping, then generates an architecture-specific kernel to solve same problem, assuming sparsity pattern is fixed, which common scenario in many applications computer graphics scientific computing. show our scales large problems can achieve speedups of two orders magnitude CPUs three GPUs, compared set manually optimized baselines. To demonstrate practical applicability approach, we employ it optimize popular algorithms with physical simulation interactive mesh deformation.