Heuristics for Conversion Process of GPU's Kernels for Multiples Kernels with Concurrent Optimization Divergence

摘要

Graphics Processing Units have been created with the objective of accelerating the construction and processing of graphic images. In its historical evolution line, concerned with the large computational capacity inherent, these devices started to be used for general purposes. However, the design of the GPUs don't work well with divergent algorithms, mainly conditionals and repetitions. In this work we present a strategy for finding the divergence root of the kernels and try to deduce alternative solutions, decomposing them into concurrent kernels. We developed mechanisms for the user in order to easily readapt his code and take advantages of architectures that support concurrent kernels.