Intrepydd: performance, productivity, and portability for data science application kernels

Major simultaneous disruptions are currently under way in both hardware and software. In hardware, ``extreme heterogeneity'' has become critical to sustaining cost and performance improvements after Moore's Law, but poses productivity and portability challenges for developers. In software, the rise of large-scale data science is driven by developers who come from diverse backgrounds and, moreover, who demand the rapid prototyping and interactive-notebook capabilities of high-productivity languages like Python. We introduce the Intrepydd programming system, which enables data scientists to write application kernels with high performance, productivity, and portability on current and future hardware. Intrepydd is based on Python, though the approach can be applied to other base languages as well. To deliver high performance, the Intrepydd toolchain uses ahead-of-time (AOT) compilation and high-level compiler optimizations of Intrepydd kernels. Intrepydd achieves portability by its ability to compile kernels for execution on different hardware platforms, and for invocation from Python or C++ main programs. An empirical evaluation shows significant performance improvements relative to Python, and the suitability of Intrepydd for mapping on to post-Moore accelerators and architectures with relative ease. We believe that Intrepydd represents a new direction of ``Discipline-Aware Languages'' (DiALs), which brings us closer to the holy grail of obtaining productivity and portability with higher performance than current Python-like languages, and with more generality than current domain-specific languages and libraries.