Scilab on a hybrid platform

International audienceIn this article we show the work done to port Scilab on an heterogeneous platform used in the H4H project. The platform is made with parallel nodes composed of a GPU accelerator connected to a standard processor. Such platform offers a lot of performance optimization opportunities. The Scilab infrastructure is composed of a front-end parser, to process the input language, and a back-end which makes intensive use of multiple standard libraries, such as BLAS, to perform required operations. In the H4H project, we ported Scilab, which usually runs on general purpose processors, to a heterogeneous platform composed of general purpose processors and GPU accelerators. In summary, we adapted Scilab to use the GPU version of libraries such as cuBLAS from NVIDIA and also worked on a parallel version able to use MPI in the script language. More specifically, we integrated a "cross-JIT" (Just In Time Compiler) capability in Scilab. This "cross-JIT" has the ability to either use the host static code or generate optimized code for the GPU depending on the input set. Parameters for the GPU code are computed offline during an evaluation phase