Conversation
pilot_gain and bottleneck-3 tanh were being applied to the frequency-domain symbol before the IDFT, but the Python set_eoo_bits applies them to the time-domain signal after IDFT and CP insertion. Move both operations to after rade_ofdm_insert_cp to match Python. Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
|
@tmiw - snap. We should probably decide who is going to work on bugs in future? 🙂 |
Heh, just saw this. I'm happy to close the other PR if this works better. I'll try both and see. (freedv-gui just finished recompiling, currently running ctest.) |
|
@tmiw - perhaps try running |
|
But FWIW both PRs seem to work for me with the extended reporting tests in freedv-gui. |
|
I'm getting some differences but in a section of the file away from the EOO symbols at the end: 
I used Claude to check both implementations and he says they have identical outputs 🤷 So the difference could be a numerical thing from different machines. So given your freedv-gui level tests check out I figure we're done 👍 I'm OK with using either PR. Good X-check if we both arrived at a similar solution I guess! |
|
I looked at both PRs again and there might be a theoretical performance benefit for mine, but TBH it likely doesn't matter much which one goes in. Your call :) |
2 participants
@peterbmarks @tmiw as per drowe67/freedv-gui#1251 (comment) the C EOO Tx magnitude was off. Following @tmiw hints I found the offending functions and set Claude to fix it. Ctests all pass.
The amplitude looks good now, you can see the well behaved EOO Tx data at the end of this plot:
@tmiw - suggest you try it out/review before we merge