A Cache-Aware Performance Prediction Framework for GPGPU Computations

We present a model for the automated prediction of average execution times for OpenCL-based computations on GPUs. The model encompasses the whole execution of the computation, including the transfer to and from the GPU, and the kernel execution itself. In contrast to existing static prediction frameworks, we incorporate the caches available on modern GPUs into our model. Using our benchmark suite, we show that memory access patterns can be grouped into five patterns that exhibit significantly different memory access performance. By extending our static analysis framework to differentiate the performance behavior of these memory access patterns, we improve on predictions made by existing GPU performance models. In order to evaluate the quality of our model, we compare cache-aware and cache-unaware predictions for a large set of randomly generated OpenCL kernels with their actual execution time.