https://hgpu.org/?p=28608
Improving the Efficiency of OpenCL Kernels through Pipes