Improve performance of OpenCL implementations (manual and PSyclone-generated) for the Xilinx FPGA

[sergisiso]

The OpenCL implementation significantly slower in the Xilinx FPGA than a single CPU, the main reason seems to be a serialized interface to access memory through a single memory bank.