Define which loops are streamed in HDL coder

Hallo,

I am using the HDLcoder in Matlab2016a to implement a neural network on an FPGA. To control the resource consumption it would be perfect to individually control which loops are streamed and which ones are unrolled. Here is a short example:

function [ out ] = Layer( in, fak, neurons )
F = fimath('RoundingMethod','Floor','OverflowAction','Wrap','ProductMode','KeepMSB');
temp = fi(zeros(neurons,1),0,in.WordLength,in.WordLength/2,F);
for jj = 1:neurons
  for ii = 1:length(in)
      temp(jj) = fi(temp(jj)+in(ii)*fak(ii),0,in.WordLength,in.WordLength/2,F);
  end
end
out = temp;
end

Is it possible to stream the inner loop(index ii) and unroll the outer loop(index jj) in order to end up with 3 DSPs working parallel generating the output? Or is it possible to influence the arrangement of the hardware implementation in any way?

Thanks Roland

Answer 1

[Chun-Yu]

Hi Roland,

Yes, this is possible by using Coder pragmas. By using the coder.hdl.loopspec() pragma, you can control which loop(s) get streamed. You can combine it with the coder.unroll() pragma to get the desired results:

    function [ out ] = Layer( in, fak, neurons )
    F = fimath('RoundingMethod','Floor','OverflowAction','Wrap','ProductMode','KeepMSB');
    temp = fi(zeros(neurons,1),0,in.WordLength,in.WordLength/2,F);
    for jj = coder.unroll(1:neurons)
      coder.hdl.loopspec('stream');
      for ii = 1:length(in)
          temp(jj) = fi(temp(jj)+in(ii)*fak(ii),0,in.WordLength,in.WordLength/2,F);
      end
    end
    out = temp;
    end

Hope that helps,

Chun-Yu

[Roland]

Thank you very much, this is exactly what I was looking for. It works very well.

I was using the old documentation for HDLcoder2014 and these pragmas are not included there but I found them in the 2016 one.

Thanks,

Roland