Abstrakt
Programmers of high-performance applications face many challenging aspects of contemporary hardware architectures. One of the critical aspects is the efficiency of memory operations which is affected not only by the hardware parameters such as memory throughput or cache latency but also by the data-access patterns, which may influence the utilization of the hardware, such as re-usability of the cached data or coalesced data transactions. Therefore, a performance of an algorithm can be highly affected by the layout of its data structures or the order of data processing which may translate into a more or less optimal se- quence of memory operations. These effects are even more pronounced on highly-parallel platforms, such as GPUs, which often employ specific execution models (lock-step) or memory models (shared memory).
In this work, we propose a modern, astute approach for managing and implementing memory layouts with first-class structures that is very efficient and straightforward. This approach was implemented in Noarr, a GPU-ready portable C++ library that utilizes generic programming, functional design, and compile-time computations to allow the program- mer to specify and compose data structure layouts declaratively while minimizing the indexing and coding overhead. We describe the main principles on code examples and present a performance evaluation that verifies our claims regarding its efficiency.