Transprecision techniques allow dynamic variation of precision to enhance performance. For memory bound applications as well as architectures which do not offer higher throughput or lower latency computation when using a lower precision, gains can be made by reducing the size of data for read/write operations. One technique to achieve that is mantissa segmentation where a high precision number is split into equallysized segments and for lower precision read/write, all segments which are not used are set to zero in the processor. There are further optimisations that target the structure of segments storage in memory (https://ieeexplore.ieee.org/document/8638359).
The consortium invites proposals for design of a library that allows easy definition of arrays of generic data types for segmented storage as well as automated conversion to the format required for computation in the processor. The approach may be validated by applying to various benchmarks (https://github.com/oprecomp) and evaluated across different platforms (PULP, POWER8, etc.).
For additional information please contact: Umar Minhas