[Discuss-gnuradio] Half-Band Filter Module Output Clipping?

[Bruce McGuffin]

Hi

I've been looking at the USRP2 filter characteristics, to see if it will 
work for me. After going through the
half-band filter code (hb_dec.v from gnu-radio 3.2) I understand what 
the filter coefficients are, but I have
a question about the output from this module.

Near the end of the module the filtered data is clipped and rounded in 
lines 152-153:

143     wire [ACCWIDTH:0]     final_sum_unrounded;
...
149     add2_reg /* add2_and_round_reg */ #(.WIDTH(ACCWIDTH+1))
150          final_adder (.clk(clk), .in1({acc_out,1'b0}), 
.in2({data_even_signext,1'b0}), .sum(final_sum_unrounded));
151
152     round_reg #(.bits_in(ACCWIDTH-4),.bits_out(OWIDTH))
153          final_round 
(.clk(clk),.in(final_sum_unrounded[ACCWIDTH-5:0]),.out(final_sum));

where ACCWIDTH = 24
and     OWIDTH = 18

and the module round_reg just calls round and puts the result in a 
register, and round.v is
     module round
      #(parameter bits_in=0,                 // for this instantiation, 
bits_in =  ACCWIDTH-4 = 20
        parameter bits_out=0)                // for this instantiation, 
bits_out = OWIDTH = 18
        (input [bits_in-1:0] in,
         output [bits_out-1:0] out);

       assign out = in[bits_in-1:bits_in-bits_out] + (in[bits_in-1] & 
|in[bits_in-bits_out-1:0]);
 
    endmodule // round

It looks to me like final_sum_unrounded is 25-bits long, so the input to 
final_round is the
20 LSBs of final_sum_unrounded (final_sum_unrounded[19:0]), and the 
output final_sum is those 20
bits rounded to 18 bits in the usual way. But this data is two's 
complement, so clipping the 5 MSBs from
final_sum_unrounded can drasticaly change the data value, for example 
changing both the sign and
magnitude, even when the magnitude of final_sum_unrounded  <  2^17.

What am I missing here?

Thanks
Bruce