Optimizing Stencil Codes Using Search

Stencil-based kernels constitute the core of many scientific applications on block-structured grids. Unfortunately, these codes achieve a low fraction of peak performance. In this paper, we demonstrate one method for automatically generating high-quality stencil code. First, we perform a search over various instruction-level optimizations to find the best platform-specific combination. These optimizations are then applied to three different cache-level algorithms. We then vary the grid size and iteration count to see where in the parameter space each algorithm performs best. In order for the system to predict the best algorithm for any point in this space, we pass the training data to a Support Vector Machine (SVM) and let it divide the space into different classes. By using this technique, we are able to choose the correct algorithm between 80% to 90% of the time.