When are double-precision samples needed?
- KVRAF
- 2637 posts since 3 Dec, 2008
Mystran, your filter is very good! It has a completely flat processing quantization error (ie residual inharmonic content after processing), and the result is very close to the desired cutoff and gain. It definitely beats the direct form 1 f32 filter, but still isn't quite as good as the svf I've proposed on the tests I've done so far. Here are some initial plots: www.cytomic.com/files/dsp/MystranPlots.pdf
The Glue, The Drop - www.cytomic.com
- KVRAF
- 7890 posts since 12 Feb, 2006 from Helsinki, Finland
Interesting, seems more or less what I expected (at least as far as the flat quantization goes). What's obvious in DF plots is how they get worse and worse with lower freqs and one of the things that triggered me to search for alternate structures was my frustration about the low-freq performance; I had some cases (eg multi-pole DC blockers at low cutoff, inside other filters with oversampling) where even double precision DFs wouldn't stay stable.andy_cytomic wrote:Mystran, your filter is very good! It has a completely flat processing quantization error (ie residual inharmonic content after processing), and the result is very close to the desired cutoff and gain. It definitely beats the direct form 1 f32 filter, but still isn't quite as good as the svf I've proposed on the tests I've done so far. Here are some initial plots: www.cytomic.com/files/dsp/MystranPlots.pdf
One thing I want to mention is that I didn't consider numerical precision as far as the actual order of calculations goes, so I have no idea if the order I posted is optimal.. but it works well enough for my purposes (and like I said, I generally run stuff at double precision anyway).