Bridging the GPGPU-FPGA efficiency gap