Skip to content

[Non BE] Added NONBE_1233_HQ_CLASSIFIER_DIV_BY_ZERO to address potential division by zero in HQ classifier

norvellrequested to merge
1233-division-by-zero-in-hq-hf_spectrum_sparseness into main

Reason why this change is needed

Possible division by zero in

https://forge.3gpp.org/rep/ivas-codec-pc/ivas-codec/-/blob/main/lib_enc/hq_classifier_enc.c#L615

Observed in BASOP when forcing HQ and the target MDCT spectrum is zero. Could potentially be solved by adding a small number to avoid the zero denominator. The BASOP seems to use Q9 for the inv_rms_fx variable, so adding 2^-9 before division?

Description of the change

The all-zero input is detected by checking if the maximum absolute value Amax is zero. In this case the crest = 1.0 and crest_mod = 1.0 and the calculation is skipped. None of the coefficients would be below the threshold, so low_count = 0. Note that the number of affected lines is larger since the existing code went into the else branch and was indented one step.

Affected operating points

  • Will affect HQ_HVQ/HQ_HARMONIC selection in HQ Core, when high band is all zero. This is a corner case which may not be triggered during normal operation. Was found when forcing the encoder to HQ, and the input segment was all zero.

Related to #1233 (closed)

Merge request reports