<html>
<head>
<meta content="text/html; charset=windows-1252"
http-equiv="Content-Type">
</head>
<body bgcolor="#FFFFFF" text="#000000">
<div class="moz-cite-prefix">On 9/3/2014 9:08 AM, Erik Anderson
wrote:<br>
<br>
</div>
<blockquote
cite="mid:CAAeZAZ6XTjBLuR07BsubZHBPOU0RjrFonXt3vVioLTa+qTC9Rw@mail.gmail.com"
type="cite"><i>I'd still like to know about how many taps people
end up using when doing convolution, mostly to know </i><i><br>
</i><i>whether there's any value in adapting the 16-tap
pure-register SSE implementation I found.</i></blockquote>
<br>
IMHO 16-tap FIR filters are quite on the low side, at least for a
bandpass filter ahead of the demodulator... <br>
as an example in Winrad I used a fast convolution filtering, whose
mask was computed starting <br>
from a 1527-tap kernel (1537 = 2048 - 512 + 1)<br>
<br>
In the firmware of the new Elad DUO, running on the Cortex ARM M4F,
I used a fast convolution<br>
whose mask started from a 769-tap kernel (769 = 1024 - 256 + 1).<br>
<br>
Both implementations have very good skirts, as shown by the Matlab
simulations done when computing <br>
the kernels.<br>
<br>
<div class="moz-signature">-- <br>
<i><b>73 Alberto I2PHD</b><br>
<small>Credo Ut Intelligam<br>
<br>
<br>
<br>
</small></i></div>
</body>
</html>