A performance-portable generic component for 2D convolution computations on GPU-based systems

In this paper, we describe our work on providing a generic yet optimized GPU (CUDA/OpenCL) implementation for the 2D MapOverlap skeleton. We explain our implementation with the help of a 2D convolution application, implemented using the newly developed skeleton. The memory (constant and shared memory) and adaptive tiling optimizations are applied and their performance implications are evaluated on different classes of GPUs. We present two different metrics to calculate the optimal tiling factor dynamically in an automated way which helps in retaining best performance without manual tuning while moving to newGPU architectures. With our approach, we can achieve average speedups by a factor of 3.6, 2.3, and 2.4 over an otherwise optimized (without tiling) implementation on NVIDIA C2050, GTX280 and 8800 GT GPUs respectively. Above all, the performance portability is achieved without requiring any manual changes in the skeleton program or the skeleton implementation.