One of the objectives of the European EoCoE-III (https://www.eocoe.eu/) center of excellence in high performance computing is to provide a tool for thorough simulations of the non-equilibrium carrier and excitations dynamics in novel materials for energy conversion and storage. One of the key component for
processing such simulations is the solution of large linear systems.

In this context, we are seeking for candidates for a 14-months post-doc at the IRIT laboratory of Toulouse with a strong background in HPC numerical linear algebra.

The linear systems resulting from the problem under consideration have a particular block-sparse structure, for which few -- if any – general-purpose numerical linear algebra codes are available. The goal of the post-doctoral position will be to develop such a HPC computing code and assess it with the libNEGF library (https://github.com/libnegf/libnegf), a quantum transport code based on non-equilibrium Green's function method.

The broader objective of the project is to introduce further layers of parallelization and distribution, develop a flexible block-sparse structure scheme, consider the effect of data compression and accuracy, and design a new data flow strategy that maximizes the interleaving of computation with communication. In terms of new
algorithmic developments, the project plans to investigate tensor algorithms, tuning at run-time of loop over GEMM on the size of operands depending on the matrix representation used, and better exploit symmetries to minimize communication. Additionally, the plan includes the development of an abstraction layer that hides openACC and CUDA calls, enhances code design on various aspects such as middle- level parallelism, low-level parallelism, and modular implementation of algorithms. Finally, the project aims to achieve multi-level parallelism and interoperability between different platforms, such as NVIDIA, Intel, AMD, ARM, etc., by developing an automatic framework that combines compiler-based code generation with heuristic performance prediction tools that evaluate the generated code through modeling of the computation and communication patterns of the underlined algorithm. Within this scope, the post-doc will participate in one or more of the above-described research activities depending on his/her skills.

The post-doc will be based at the IRIT laboratory (Toulouse, France) and will be carried in tight collaboration with Inria (Bordeaux, France), Juelich Research Center (Juelich, Germany) and CNR (Rome, Italy).