Michele Martone

Software developers using high-level programming languages often find they need to call procedures from specialized libraries only offering a C or C++ API. While most languages expose mechanisms for calling C functions, writing the bindings for potentially hundreds of procedures remains menial work: tedious and error-prone.

Once a codebase gets large, refactoring for performance is difficult. Challenges include: the sheer time to carry out a possibly complicated transformation globally; the risk of introducing defects; poor acceptance by collaborators or code owners (if doing the HPC-oriented update for others); coping with diverging 'performance' paradigms (different parallelism models, APIs for GPU, etc).

Many science and engineering models reduce to problems involving huge sparse matrices -- matrices where most of the values are zeroes. Such computations are resource-intensive (time, memory, energy), and much research was devoted into data structures ("formats") and algorithms leading to fast sparse matrix operations.