demapper: NaN LLR values output for high SNR inputs.

[japm48]

Simple example to reproduce this: https://gist.github.com/japm48/41bc963efec39e506c17e419c3d1e13f

Basically, the demapper outputs NaN values when the SNR is too low, this can be seen that by reducing ebno_db this issue goes away.

The cause seems to be on the Demapper, where the exponents -squared_dist/no_eff passed to SymbolLogits2LLRs become -∞.

I'll create a PR with a simple fix.

Aside from that, I think that the "distance to all symbols" technique is computationally suboptimal in many common cases, e.g. on uniformly-distributed QPSK with Gray coding the exact LLRs should be proportional to the real and imag parts, no log nor exp needed (I will open a separate issue addressing this).

[SebastianCa]

Hi,
thanks for reporting this issue. We can reproduce the issue and will look into it.

[SebastianCa]

Fixed in Sionna v0.17