Motivation

With the current trend of increasing parallelism, there is a need for a uniform way of making these parallel implementations available along with the default serial ones. Currently in PCL, this has been done by providing separate classes/functions having independent APIs.

With the integration of execution policies and executors (part of C++23 hopefully) libraries like PCL will be able to provide a unified API to switch between different implementations. To ensure there is no sudden breakage in the existing API, a transition phase is needed to slowly phase out the old API in favour of a unified API.

Implementation

This project aims to transition the existing API to forward-compatible unified API. The basic details are: Using tag dispatching to enable function overloading, allowing user to select between different available implementations. Lack of tags allows PCL to choose best possible implementations, this allows to maintain backward compatibility with current API. Missing implementation for a tag raises compile-time errors Providing SIMD/OpenMP/CUDA implementations of existing algorithms using the proposed unified API.

Organization

Student

Shrijit Singh

Mentors

  • Kunal Tyagi
  • Sérgio Agostinho
close

2020