Improving Code Generation for Reductions on Large Nested Data Structures.

What is it about?

Partial reductions such as summations of all rows of a matrix are notoriously challenging to generate efficient code for. The key problem lies in an efficient management of memory. While for summations this is trivial, in other cases, such as concatenations, more elaborate memory management is required. This paper proposes a compiler technique that generates memory management suitable for all cases while reusing memory when possible. We demonstrate the effectiveness of our approach by means of an empirical evaluation in the context of the programming language SaC and its attendant compiler tool chain.

Featured Image

Photo by White.Rainforest ™︎ ∙ 易雨白林. on Unsplash

Photo by White.Rainforest ™︎ ∙ 易雨白林. on Unsplash

Why is it important?

This paper creates the basis for a vectorisation of reduction operations on large multi-dimensional data and it enables transparent performance for such reductions irrespective of the way they are specified.