See #416. When the input is zero, the RMS AGC block computes 0 / (0 + 1e-20). The 1e-20 in the denominator is added to avoid 0 / 0, which gives nan. Nevertheless, it seems that on aarch64, 0 / 1e-20 still gives nan (at least when doing the complex division).
This replaces the small number 1e-20 by 1e-19, which works correctly on aarch64, and adds a unit test for the rms_agc and the rms_agc_f so that we can catch if this problem happens in another systems.
See #416. When the input is zero, the RMS AGC block computes 0 / (0 + 1e-20). The 1e-20 in the denominator is added to avoid 0 / 0, which gives nan. Nevertheless, it seems that on aarch64, 0 / 1e-20 still gives nan (at least when doing the complex division).
This replaces the small number 1e-20 by 1e-19, which works correctly on aarch64, and adds a unit test for the rms_agc and the rms_agc_f so that we can catch if this problem happens in another systems.