This adds the cross talk filter to the noise covariance calculation (and corresponding correction factors). It turned out to involve a sizable matrix multiply that made the calculation take 15 minutes without any adjustments.
I tried different speedup schemes to ameliorate this including generating one filtration operator per spectral window, only using one of the streams, and decimating in frequency. I found that decimating in frequency by a factor of 10 improves the speed by an order of magnitude (just under two minutes to complete) with an error compared to the full calculation of one part in 10^4 at worst on the couple baselines that I checked.
This option is adjusted with the CORR_MATRIX_FREQ_DECIMATION global, defaulted to 10.
This adds the cross talk filter to the noise covariance calculation (and corresponding correction factors). It turned out to involve a sizable matrix multiply that made the calculation take 15 minutes without any adjustments.
I tried different speedup schemes to ameliorate this including generating one filtration operator per spectral window, only using one of the streams, and decimating in frequency. I found that decimating in frequency by a factor of 10 improves the speed by an order of magnitude (just under two minutes to complete) with an error compared to the full calculation of one part in 10^4 at worst on the couple baselines that I checked.
This option is adjusted with the CORR_MATRIX_FREQ_DECIMATION global, defaulted to 10.