Performance, portability, and productivity for data-parallel applications on multi- and many-core architectures

We present a novel approach to performance, portability, and productivity of data-parallel computations on multi- and many-core architectures. Our approach is based on Multi-Dimensional Homomorphisms (MDHs) -- a formally defined class of functions that cover important data-parallel computations, e.g., linear algebra routines (BLAS) and stencil computations. For MDHs, we present a high-level Domain-Specific Language (DSL) that contributes to high user productivity, and we propose a corresponding DSL compiler which automatically generates optimized (auto-tuned) OpenCL code, thereby providing high, portable performance, over different architectures and input sizes, for programs in our DSL. Our experimental results, on Intel CPU and NVIDIA GPU, demonstrate competitive and often significantly better performance of our approach as compared to state-of-practice approaches, e.g., Intel MKL/MKL-DNN and NVIDIA cuBLAS/cuDNN.