The NEON implementation of the volk_32fc_s32f_power_spectrum_32f kernel has a bit more error than the generic & sse3 implementations, because it uses an approximate logarithm calculation (_vlogq_f32). The test appears to begin passing reliably with a tolerance of around 1.2e6, so I increased that to 2e6 to leave some breathing room.
Fixes #663.
The NEON implementation of the
volk_32fc_s32f_power_spectrum_32fkernel has a bit more error than the generic & sse3 implementations, because it uses an approximate logarithm calculation (_vlogq_f32). The test appears to begin passing reliably with a tolerance of around 1.2e6, so I increased that to 2e6 to leave some breathing room.