Interaktive GPU Programmierung mit dem Cling-C++-Interpreter in Lehre und Wissenschaft

In the scope of this diploma thesis the Cling JIT compiler, which was developed by a team of the CERN, was extended by the function to compile CUDA C++ code. The Cling-CUDA allows for new concepts in developing and executing GPU applications within the area of HPC. The key finding of this progress is the elimination or reduction of the times for compiling, linking, job submitting and program start of a GPU application as soon as program flow changes. The important characteristic of C++ to generate high-performance code is retained. Furthermore, the possibility of reusing existing CUDA C++ code without changes allows a relatively simple migration into existing projects. The Cling was combined with Jupyter Notebook to provide a preferably wide range of users access to this new technology. This step has simplified the use of HPC resources on a remote system as well as the publication of reproducible scientific results in the form of program code, data and documentation. The widespread concept of Jupyter Notebook enables user groups who so far have been deterred by the traditional use of HPC systems to access HPC resources. Two use cases were developed to demonstrate the use and advantages of the newly developed Cling- CUDA. The first example describes the development and execution of a simulation and analysis using the Cling-CUDA and Jupyter notebook. For the second example, an exercise from the lecture "Hochparallele Programmierung von GPUs" of the computer science course of studies at the TU Dresden was converted into a notebook. This has been realized in cooperation with the respective lecturers. The two community projects ROOT (CERN) and PIConGPU (HZDR) are already interested in the results of the present research and evaluate feasibility and benefit of an Cling-CUDA integration. Especially the integration into ROOT could open up a large community of HPC users who could become strong drivers for the future development of Cling-CUDA