Transparent CPU-GPU Collaboration for Data-Parallel Kernels on Heterogeneous Systems