Value-Based Optimisation

Description:
Our value-based optimisation work focusses on the use of Pin to extract profiling information concerning load/store value locality and value-reuse at instruction level granularity on sparse and general purpose applications. A significant goal of our work is to examine how to use value locality information in order to compress the storage formats of sparse matrix based applications exploiting sparsity patterns, frequent values and knowledge of the floating-point value ranges of non-zero elements. To enable this aspect of the work we expect to collaborate with the cluster on "Using Adaptive Compilation to Produce High Performance Sparse Computations". We will investigate the potential for value-based optimisation of Sparse Matrix Vector Multiplication based applications using a wide variety of test matrices. The work will attempt to identify both software and hardware optimisation techniques, with a view to defining reconfigurable architectures for streaming compression/decompression and value sensitive specialisation of functional units.

Current Members:
* Andy Nisbet Dept. Computing & Mathematics, Manchester Metropolitan University, UK
* Mikel Lujan Department of Computer Science, University of Manchester, UK
* David Moloney Chief Technical Officer, Movidia, Ireland.
* Dermot Geraghty Dept. Mechanical & Manufacturing Engineering, Trinity College, Ireland.

The cluster welcomes new members interested in the general topics outlined above.