GPUPerformance Prediction Using Parametrized Models
Date: 2011-06-30 Time: 12:00 Room: [[http://www.cs.uu.nl/docs/reach/bbl.php][BBL]] [[http://www.cs.uu.nl/info/plan/bbl.php]] ----+++++ Speaker: Andreas Resios ----+++++ Title: GPU Performance Prediction using Parametrized Models ----+++++ Abstract The recent developments in GP-GPU programming, have lead to the need to port sequential code to these parallel architectures. When parallelizing code for heterogeneous platforms it is important to estimate the benefits of the transformation. In order to address this problem we developed a GPU cost model which predicts the run-time and identifies parallelization bottlenecks of sequential programs. Using our model the user can quickly identify which parts of the program achieve speedup, thus increasing his productivity when parallelizing code.