Merge or Separate? Multi-job Scheduling for OpenCL Kernels on CPU/GPU Platforms