Compilation step gets very slow with shift register (multiple unrolled shifts)

Altera_Forum · ‎02-09-2016

I just ran into this behaviour when trying to compile an unrolled loop over a shift register.

The code needs an incredible amount of time to get past the compilation step (Altera Offline Compiler on Windows).

-->The compilation step right before the logic - estimation output.

Issuing Code:

# pragma unroll
for(int i = 0; i < NUM_DISPARITIES; ++i)
{
    ushort sum = 0;    
   # pragma unroll
    for(int wx = 0; wx < BM_WINDOW; ++wx)
    {
         sum += abs((left) - (right));
    }
                
     //shift register
    # pragma unroll
      for (int s = (BM_WINDOW * IMAGE_WIDTH - 1); s > 0; --s) 
     {
          linSums = linSums;
     }
     linSums = sum;
}

'left' and 'right' are also shift registers, the problem was the same with using 'linSums' as a 1-Dimensional shift register.

With NUM_DISPARITIES = 2 the Code needs little more than expected.

At NUM_DISPARITIES = 16 the compilation step needed over 20 mins.

Maybe this isnt a bug, but bad code or misunderstand use of shift registers ^^

However I would appreciate any help!

Regards

Steffen