Gmsk strangeness

'lo all.

We are trying to use the USRP with the new gmsk demod block, and are
encountering some weirdness.
To generate our GMSK signal, we are using an Agilent signal generator.
BT=0.35, 125 ksymbols/sec, 2 samples/symbol, with a PN15 as the data
source.

We load the produced slicer.dat into Matlab and cross-correlate with the
known PN15 sequence.

Sometimes the output looks good: all our peaks are at 2^15 - 1 = 32767,
indicating that all bits are being demodded correctly,
as you can see here:
http://img472.imageshack.us/img472/2766/fullpeakscb3.jpg

Other times, we encounter “half-height” peaks, which go up to 2^14-1 =
16383, indicating that 25% of the bits are being demodded incorrectly.
You can see this here:
http://img441.imageshack.us/img441/7507/halfpeaksax7.jpg
If you zoom in on the previous image, you can see that each xcorr peak
actually consists of 2 high points, then a lull of 14 or so, then
another
high point. This pattern is always the same when the half-height
phenomena
is observed:
http://img168.imageshack.us/img168/7657/halfpeakszoomeduy7.jpg

Any idea what might be going on, or how to debug this?

-Steven

Other times, we encounter “half-height” peaks, which go up to 2^14-1 =
16383, indicating that 25% of the bits are being demodded incorrectly.

Can you get phase information from those samples? If you plot IQ in
your unit circle at 1 sample per symbol, do you get a correctly
equalized waveform?

Any idea what might be going on, or how to debug this?

Could you post your samples?

Brian