A Practical Performance Model for Compute and Memory Bound GPU Kernels