Texturizing PPCG: Supporting Texture Memory in a Polyhedral Compiler

Abhishek A Patwardhan and Ramakrishna Upadrasta

Poster-Presented in IEEE International Conference on High Performance Computing, 2016


In this paper, we discuss techniques to transform sequential programs to texture/surface memory optimized CUDA programs. We achieve this by using PPCG, an automatic parallelizing compiler based on the Polyhedral model. We implemented a static analysis in PPCG which validates the semantics of the texturized transformed program. Depending on the results of the analysis, our algorithm chooses to use texture and/or surface memory, and alters the Abstract Syntax Tree accordingly. We also modified the code-generation phase of PPCG to take care of various subtleties. We evaluated the texturization algorithm on the PolyBench (4.2.1 beta) benchmark and observed up to 1.6x speedup with a geometric mean of 1.103X. The title and at many places, the paper uses term Texture memory. But, the optimizations are for Texture and Surface memory.