HySoP (Hybrid Simulation with Particles) is a library dedicated to high performance direct numerical simulation of fluid related problems based on semi-lagrangian particle methods, for hybrid architectures providing mul- tiple compute devices including CPUs, GPUs or MICs. The high level functionalities and the user interface are mainly written in Python using the object oriented programming model. This choice was made because of the large software integration benefits it can provide [Sanner 1999]. Moreover, the object oriented programming model offers a flexible framework to im- plement scientific libraries when compared to the imperative programming model [Arge et al. 1997][Cary et al. 1997]. It is also a good choice for the users as the Python language is easy to use for beginners while experienced programmers can pick it up very quickly [Oliphant 2007a]. The numerical solvers are mostly implemented using compiled languages such as Fortran, C/C++, or OpenCL for obvious performance reasons. It is also possible to implement numerical algorithms using directly Python, which is an interpreted language, hence slower for critical code paths under heavy arithmetic or memory load. The Python language support is however the key for rapid development cycles of experimental features. It also allows to easily implement routines that compute simulation statistics during runtime, relieving most of the user post-processing efforts and enabling live simulation monitoring.
HySoP (Hybrid Simulation with Particles) is a library dedicated to high performance direct numerical simulation of fluid related problems based on semi-lagrangian particle methods, for hybrid architectures providing multiple compute devices including CPUs, GPUs or MICs.
The high level functionalities and the user interface are mainly written in Python using the object oriented programming model. This choice was made because of the large software integration benefits it can provide [Sanner 1999]. Moreover, the object oriented programming model offers a flexible framework to im- plement scientific libraries when compared to the imperative programming model [Arge et al. 1997][Cary et al. 1997]. It is also a good choice for the users as the Python language is easy to use for beginners while experienced programmers can pick it up very quickly [Oliphant 2007a]. The numerical solvers are mostly implemented using compiled languages such as Fortran, C/C++, or OpenCL for obvious performance reasons. It is also possible to implement numerical algorithms using directly Python, which is an interpreted language, hence slower for critical code paths under heavy arithmetic or memory load. The Python language support is however the key for rapid development cycles of experimental features. It also allows to easily implement routines that compute simulation statistics during runtime, relieving most of the user post-processing efforts and enabling live simulation monitoring.
The compiled Fortran and C++ compute backends allow us to integrate a variety of external dependencies by connecting them to the main HySoP python module with interface wrappers such as F2PY or SWIG. Note that many scientific libraries already provide Python interfaces so that they can be directly used in python without needing the user to implement his own wrapper. In addition to the compiled languages, the library offers the possibility to compile generated code just-in-time during execution. This is the case for OpenCL, the language used to drive OpenCL-compatible accelerators, like GPUs, but also to translate python methods to fast machine code by using the Numba just-in-time compiler. Most of the dependencies that the HySoP library uses are described in appendix B.
This subsection provides implementation details of the HySoP library from its origin to the most recent developments (2012-2019). It aims to cover computational-backend agnostic design choices that allows us to discretize the variables and equations of a given fluid-related problem into subproblems, each being solved individually by some operator embedded into a directed acyclic graph. A first level of parallelism is explored here, namely task parallelism.
This subsection provides implementation details of the HySoP library from its origin to the most recent developments (2012-2019).
It aims to cover computational-backend agnostic design choices that allows us to discretize the variables and equations of a given fluid-related problem into subproblems, each being solved individually by some operator embedded into a directed acyclic graph. A first level of parallelism is explored here, namely task parallelism.
