Adaptive GPU Scheduling for Efficient Numerical Computing and Computer Graphics, Markus Steinberger
Sprache des Titels:
Englisch
Original Kurzfassung:
In this talk, I will discuss our efforts towards dynamic GPU scheduling, which allow a larger class of algorithms to run efficiently on the GPU. Applying these ideas to well-known algorithms, we push the performance beyond state-of-the-art approaches in such contested fields like sparse matrix-vector multiplication while allowing for more flexibility. Applying these ideas to procedural geometry generation, we propose the fastest approach for generating buildings and large cities directly on the GPU. Analyzing the scheduling behavior, we present an autotuner framework that finds an optimal schedule for the execution of a specific procedural generator for a specific GPU.