AUTOMATING DATA-LAYOUT DECISIONS IN DOMAIN-SPECIFIC LANGUAGES

A long-standing challenge in High-Performance Computing (HPC) is the simultaneous achievement of programmer productivity and hardware computational efficiency. The challenge has been exacerbated by the onset of multi- and many-core CPUs and accelerators. Only a few expert programmers have been able to hand-code domain-specific data transformations and vectorization schemes needed to extract the best possible performance on such architectures. In this research, we examined the possibility of automating these methods by developing a Domain-Specific Language (DSL) framework. Our DSL approach extends C++14 by embedding into it a high-level data-parallel array language, and by using a domain-specific compiler to compile to hybrid-parallel code. We also implemented an array index-space transformation algebra within this high-level array language to manipulate array data-layouts and data-distributions. The compiler introduces a novel method for SIMD auto-vectorization based on array data-layouts. Our new auto-vectorization technique is shown to outperform the default auto-vectorization strategy by up to 40% for stencil computations. The compiler also automates distributed data movement with overlapping of local compute with remote data movement using polyhedral integer set analysis. Along with these main innovations, we developed a new technique using C++ template metaprogramming for developing embedded DSLs using C++. We also proposed a domain-specific compiler intermediate representation that simplifies data flow analysis of abstract DSL constructs. We evaluated our framework by constructing a DSL for the HPC grand-challenge domain of lattice quantum chromodynamics. Our DSL yielded performance gains of up to twice the flop rate over existing production C code for selected kernels. This gain in performance was obtained while using less than one-tenth the lines of code. The performance of this DSL was also competitive with the best hand-optimized and hand-vectorized code, and is an order of magnitude better than existing production DSLs.Doctor of Philosoph