Three Fingered Jack: Tackling Portability, Performance, and Productivity with Auto-Parallelized Python

We present Three Fingered Jack, a highly productive approach to writing high-performance software in Python. Our system applies traditional dependence analysis and reordering transformations to a restricted set of Python loop nests. It does this to uncover and exploit vector and thread parallelism. Then it generates native code for multiple platforms. On four benchmark kernels and two applications, our system achieves 0.4 — 13.5x  and 0.97 — 113.3x  of hand written but not highly optimized C++ performance on Intel and ARM CPUs, respectively.