An algorithm for trading off quantization error with hardware resources for MATLAB-based FPGA design

Most practical FPGA designs of digital signal processing (DSP) applications are limited to fixed-point arithmetic owing to the cost and complexity of floating-point hardware. While mapping DSP applications onto FPGAs, a DSP algorithm designer must determine the dynamic range and desired precision of input, intermediate, and output signals in a design implementation. The first step in a MATLAB-based hardware design flow is the conversion of the floating-point MATLAB code into a fixed-point version using "quantizers" from the filter design and analysis (FDA) toolbox for MATLAB. This paper describes an approach to automate the conversion of floating-point MATLAB programs into fixed-point MATLAB programs, for mapping to FPGAs by profiling the expected inputs to estimate errors. Our algorithm attempts to minimize the hardware resources while constraining the quantization error within a specified limit. Experimental results on five MATLAB benchmarks are reported for Xilinx Virtex II FPGAs.