Section: New Results

Scheduling Trees

Participants : Laure Gonnord, Paul Iannetta.

As a first step to schedule non polyhedral computation kernels, we investigated the tree datastructure. A large bibliography on tree algorithmics and complexity leds us to chose to work on balanced binary trees, for which we have designed algorithms to change their memory layout into adjacent arrays. We rephrased the classical algorithms (construction, search, destruction ...) in this setting, and implemented them in C.

The conclusion of this study is unfortunately negative : the locality gain in transforming trees into linear structures is not contrabalanced by a better cache usage, all our codes have been slowed down in the process. Our experiments are still in progress, but our hypothesis is that our trees are too sparse to be more clever that the malloc implementation.

A research paper will be published early 2020. This work is done in collaboration with Lionel Morel (CEA Grenoble), coadvisor of the PhD of Paul Iannetta.