New feature: per-Qureg parallelisation

[TysonRayJones]

contributors: Oliver Brown, Tyson Jones
release: v4

Parallelisation methods (i.e. multithreading, GPU-acceleration and distribution) can now be runtime enabled or disabled for individual Qureg instances. This permits some registers to be e.g. multithreaded, and others to be e.g. GPU-accelerated, as may be ideal for their respective sizes.

• developer benefits

  • Parameterised QuEST benchmarking (e.g. measuring speedup over modes) can be performed in a single program without recompilation.
  • Ergo checking the benefit/performance of different parallelisation strategies is very easy.

• user benefits

  • Users can disable deployments at run-time. For instance, although they may compile QuEST with GPU-acceleration enabled, they can switch off GPU-acceleration without recompilation in order to profile CPU simulation.
  • This means users and HPC staff can "install" QuEST. They can compile QuEST with all modes enabled (multithreaded, GPU-accelerated, distribution) and have users target any backend at run-time, without recompilation. QuEST can ergo be compiled once and provided as a module.
  • Users can always see how their simulations are actually being parallelised at runtime.
  • Users can switch off parallelisation techniques if their simulated systems are too small to overcome their overheads.
  • Users can simultaneously store multiple Qureg which are too small to distribute, or too big to fit into a single GPU.

[TysonRayJones]

Added in 372eedf, 8213557 and 6d59bf3