Algorithm development starts in MATLAB, Lab-view, C and JAVA which uses complex floating point arithmetic. These algorithms are then transferred to hardware implementation team. Then the hardware design team implements this abstract algorithm on the specific hardware that meets the system level specifications. However, it is very difficult to achieve system-level constraints because most of the algorithms contain complex data structure and loops which costs to power consumption and memory usage for particular application.
The point to be noted
Most of the algorithm developers usually ignore number representation e.g. algorithm from MATLAB to C using floating point which provides high precision and huge dynamic range. These floating point representations contain two types of data type, one is single precision and other is double precision. In single precision, significant data is represented in 23 bit and exponent is represented by 8 bit. However, double contains 52 bit of significant data and 11 bit for exponent.
Algorithm developers use floating point to get the more and more precise results. They don’t worry about significant bit left after whole operation, which may become critical for implantation engineer.
When any design is implemented using intensive numeric operation, it is hard to adapt algorithm for hardware implementation or takes a long time in development process. Sometimes, the results obtained are different from design specification.
So, before you start, it is necessary to be attentive about resources, cost, target device and risk at early stage of the design. A algorithm designer should take an account about the hardware specification given for a particular design, so, it is better to know whether given hardware supports floating point data types or not.
If there is a need to convert floating to fixed point data types, it requires considerably more time and some special skills. It is too intensive job when there is a need to convert the double precision to single precision as that of fixed point conversion.
HDL conversion for Hardware implementation:
When algorithm designer is done with the algorithm then it goes to the design implementation team. There, they read the specification and numeric operation used in given algorithm.
Implementation engineer firstly look after target hardware specification they had been given. If they found that data types (floating or fixed) used in this particular algorithm don’t support in target hardware they request for the revision of algorithm.
There are many automated tools which convert MATLAB and Simulink models into HDL like HDL coder from MATLAB itself converts system level algorithm into synthesizable code, System generator from Xilinx which is more suitable when you are targeting Xilinx hardware.
This is final and most important task in the journey of system level algorithm to real time hardware implementation. In real scenario there are many issues like speed, cost and power. When we are implementing our algorithm for floating point it may consume more power and may be low in speed relatively.
There are several techniques to optimize the design. Firstly, algorithm design team must write a simple algorithm avoiding intensive loop and large data types.
A well-defined pipelined architecture manages speed and power consumption.
Critical data path, custom data path width, signals and power gating also helps to get better performance at low cost.
While optimizing your design you should take care of initial results and results you get after optimizing the design. Bit error should be low.
Author - Varinder josan
(Ceo at Silicon Mentor)