IRTG Modern Inverse Problems (MIP)

Model-Based Generation of Linear Algebra Software

09. Oktober 2018 | von

Prof. Paolo Bientinesi, Ph.D.

This project tackles the automatic optimization of linear algebra operations that represent the computational bottleneck in scientific simulations and in data analyses.

The PI’s have extensive experience in parallel computing and in the development of linear algebra kernels, as well as performance models and automation; the combined expertise will make it possible to streamline the generation of high-performance algorithms and code, tailored towards specific applications, and targeting existing and upcoming computing architectures. The methodology builds on stochastic performance models, and formal derivation methods, and will be applied to selected operations, such as those arising in global optimization.
Concisely, the project has three intertwined aims.

Aim 1: Identification and development of a set of high-performance low-level kernels, sufficient to support the set of target operations.

Aim 2: Derivation of performance and scalability models for the building blocks, with quantification of the associated uncertainties.

Aim 3: “Performance model”-based decomposition of high-level operations in terms of the available building blocks.

In sharp contrast to the typical library development, our approach follows a reverse order, akin to an inverse problem: Given a target known functionality, the objective is to identify the composition of kernels that minimizes a cost function.

Kommentare sind geschlossen.