Santa Cruz, Calif. Matlab users who want to speed verification or support system-level design with C-language models typically have to manually convert Matlab models to C. There's a better way, says DSP synthesis provider AccelChip Inc.: Let us generate the models.
AccelChip is rolling out the 2006.1 version of its AccelChip DSP Synthesis solution, which produces synthesizable RTL code from fixed-point Matlab models. New in this release is the M2C-Accelerator option, which automatically generates C++ verification models from the same Matlab models.
The new capability will appeal to two groups of users, said Tom Feist, AccelChip's vice president of marketing and sales. One is algorithm developers who want to speed verification for such applications as image processing. The other is system designers who want to pass a block back to top-level designers, who will simulate the entire design in C.
By automatically generating bit-true C++ models, M2C-Accelerator promises to speed fixed-point verification times by up to a thousandfold in C-based flows, or up to 150x in mixed Matlab-Simulink flows. The C++ code is available in source form for debugging, and all comments and variables in Matlab are transferred to the C++ model.
M2C-Accelerator also produces a testbench that ensures the C++ model correlates to the original fixed- and floating-point Matlab models. “We create a separate C++ main function that calls the model, and basically that main function is your testbench,” said Bradley Armstrong, vice president of engineering at AccelChip. “We compile it into an executable and you can run it on whatever platform you have.”
To use M2C-Accelerator, designers first use AccelChip DSP Synthesis to convert floating-point Matlab models to fixed-point representations. While the normal flow would produce fixed-point Matlab models, the M2C option produces fixed-point C++ models. In either case, AccelChip DSP Synthesis still retains the necessary information to produce synthesizable RTL code.
The C++ models generated by M2C-Accelerator can be used in Matlab, Simulink, Xilinx System Generator and standalone C verification environments. SystemC model generation is “on the road map,” said Feist.
Also new in the 2006.1 release is a feature called AccelProbe. Working with either the M2C option or the normal AccelChip flow, AccelProbe assists the floating-to-fixed-point conversion process by providing graphical feedback, including quantized signal-to-noise ratio and quantization histogram reports for any variable in the design, according to the company.
When users convert from floating point to fixed point, Feist noted, they have to choose the number of bits. AccelProbe can help users determine whether they can use fewer bits while still meeting the demands of their algorithm. “After I run my floating-to-fixed-point conversion, I can bring up a graph and see very easily what happened during quantization,” he said. “I can visually see the impact of cutting off an extra bit.”
AccelChip DSP Synthesis version 2006.1 is available now starting at $15,000. The M2C-Accelerator option starts at $5,000.