Modeling hardware in Matlab/Simulink (delay, etc.)?

[moe]

Hello all,

I need some guidedance on how to go about modeling an algorithm before I
design it into an FPGA.

I'd like to use Matlab to run the simulation. Instead of just sanity
check, I also need to optimize the size (# of bits) of several
parameters. It'll be a balance act between # of bits (hence, FPGA
performance) and precision of the algorithm output. FPGA needs to run
150+Mhz, and the parameter sizes may end up being somewhat of an odd
size (48? 56?), perhaps up to 64. The algorithm is made up of mostly
adders and delay elements.

So far, it looks like I need to use Simulink with Fixed-point blockset.
Is that sufficient, or is there a better way to do this? I've been
reading previous posts in different places, and it sounds like modeling
delay isn't as easy as it should be. Is that true? I was able to use
'Triggered subsystem' to model a FF without much problem, but then it
was for a very simple single bit data path.

I'd appreciate any input, feedback, insight, suggestion. I'm sending
this to VHDL/Verilog newsgroups for a wider audience.

Please replace 'hard' with 'easy' in my email addr, if replying via
email.

-moe

 

[Tom Hawkins]

Hello all,

I need some guidedance on how to go about modeling an algorithm before I
design it into an FPGA.

I'd like to use Matlab to run the simulation. Instead of just sanity
check, I also need to optimize the size (# of bits) of several
parameters. It'll be a balance act between # of bits (hence, FPGA
performance) and precision of the algorithm output. FPGA needs to run
150+Mhz, and the parameter sizes may end up being somewhat of an odd
size (48? 56?), perhaps up to 64. The algorithm is made up of mostly
adders and delay elements.

You may want to consider Confluence (CF) for modeling your initial
design.
Like Simulink, CF is a structural dataflow language, i.e., you design
with systems and subsystems, wires and ports. But instead of a GUI,
CF is textual language.

The main strength of Confluence is it's extensive use of parameters;
every CF variable is a configuration parameter that controls how a
system is created. Therefore it's easy to add parameters to
experiment with design tradeoffs such as signal precision and system
architecture.

For simulation, Confluence compiles into cycle and bit accurate C and
Python models. I personally prefer Python as a high-level testbench
language, but compiled C is very fast. Also, it's straightforward to
integrate C models into Matlab via Simulink s-functions.

So far, it looks like I need to use Simulink with Fixed-point blockset.
Is that sufficient, or is there a better way to do this? I've been
reading previous posts in different places, and it sounds like modeling
delay isn't as easy as it should be. Is that true? I was able to use
'Triggered subsystem' to model a FF without much problem, but then it
was for a very simple single bit data path.

If you're targeting an FPGA, your delays (aka. registers) should be
synchronous, i.e., the output data only changes on the clock edge. If
you only have one clock domain, synchronous registers are modeled in
Simulink as a unit delay block (1/z). I would stay away from
triggered subsystems because their
semantics do not correspond to hardware. If you need a delay with a
synchronous
reset and enable, you can add a couple switches in front of the 1/z as
shown in the following pseudo code:

UnitDelayInput = Reset ? 0 : (Enable ? Input : UnitDelayOutput)
Output = UnitDelayOutput

In terms of delays, Confluence is a purely synchronous language; all
registers are synchronous. For example, gluing a chain of 7 delays
together is done like this:

{Delay, 7 DataIn, DataOut}

Note that 7 is a configuration parameter. Techincally speaking,
DataIn is too. One nice feature of the language is clocks, resets,
and enables are implicit. This allows you to clock, reset, or enable
entire subsystems, without having to drag these signals down to every
register in a hierarchy.

You can find Confluence at: http://www.launchbird.com/download.html

If you'd like to see a particular example, let me know.

Regards,
Tom