Hey everyone!

I build a well know Karlsen 24 db ladder filter in PD, and I really like the sound of it. It's in a stabil, working state, but it uses around 9-10 % of my older Macbook. I've tried a few things for optimisation but I am kind of stuck atm, I can't optimise it anymore, with the knowledge I have(hobby coder/Pd user).

I was wondering if anyone have suggestions for optimisation?

Besides that, merry christmas to everyone

Jakob!

The filter can be found here:
https://github.com/jaffasplaffa/Pure-data-patches/tree/master/Karlsen-Ladder-24-db

@jaffasplaffa why fexpr~ for a simple onepole filter.. when your patch already runs with blocksize 1? you can replace [fexpr~] with [lop~] or [rpole~] or some delays + math like in the original code.
I couldn't resit building the thing following this example: https://www.musicdsp.org/en/latest/Filters/240-karlsen-fast-ladder.html

spoiler.pd
!filter gets unstable at higher frequencies

also, do you notice any difference in sound with oversampling + anti aliasing filter..?
maybe make it optional and put the anti aliasing filter in a subpatch with [switch~]

Merry christmas to you too!

@solipp said:

why fexpr~ for a simple onepole filter.. when your patch already runs with blocksize 1? you can replace [fexpr~] with [lop~] or [rpole~] or some delays + math like in the original code.
I couldn't resit building the thing following this example: https://www.musicdsp.org/en/latest/Filters/240-karlsen-fast-ladder.html

spoiler.pd
!filter gets unstable at higher frequencies

also, do you notice any difference in sound with oversampling + anti aliasing filter..?
maybe make it optional and put the anti aliasing filter in a subpatch with [switch~]

Merry christmas to you too!

I tried removing the block~ 1 a while ago, for the reason you mention, that fexpr~ is already processing every sample, but I did not get the same result, So kind of scratched it, but I will look into it again.

I am still beginner with filters, but the reason I used fexpr~ was because of the "processes every sample", so I was kind og thinking this is what I need.

for the lop~ are you sure that these filters are made in the same way, the lop~ and the karlsen filter? There are a few filters that are made feom one pole filters on musicdsp, which are all a bit different. I am thinking the difference makes up different result?

For the oversampling usually oversampling makes sure that it doesn't blow up at higher frequencies. For example bob~, if you don't oversample it blows up too. With oversampling, it's much more stabil.

Ahh yes I should probably do the switch thing for the oversampling/butterworth filter. For now I just set the oversampling to 1, which is no oversampling, but yeah the filter is still running, you are right.

Anyway, thanks I will check your version and you gave me some food for thought here. Some things to check out

here are my changes: as @solipp says you can change the [fexpr~]s to [rpole~]s, that might help. There are also situations where you can use an argument to [*~ ] or [+~ ] to have it create a version where the 2nd inlet is control-rate, not signal rate (slightly more efficient). I also moved the audio calculations to control-rate where I could.
dunno if it actually helps that much (or if the [rpole~]s are really more efficient than the [fexpr~]s with the additional math required)
Karlsen-Ladder-24-db.pd

edit: (btw I recommend using some pd where signal and control in/outlets are distinguishable from each other, either Purr-Data or my mod https://github.com/sebshader/pdnext/releases)

edit2: maybe the 2 [*~ ]s to raise to the 4th power in the 4x1-pole subpatch should be replaced by [expr~ $v1*$v2*$v2*$v2*$v2], I never know which will be more efficient..

edit3: seems like all these changes make it a ~3-4 1/10 ms/s "cputime" more efficient here

edit4: I updated the patch so cutoff stays roughly the same when changing the upsampling factor

AHH I tried out both versions now.

Solipps versions seems a bit unstable, but I'll try oversampling it and see how that works out

Seb's version is by listening pretty much the same as my version, except it's "cheaper", mine uses around 8-9% cpu, Seb's uses only 4-5%. Thats a noticeable performance gain.

I was not really sure how those pole filters works, but will check them out in the future, instead of fexpr~.

For the oversampling and the filter/butterworth, I think the only thing it actually does in this filter is expanding the frequency range. And also affect the hipass filter. It actually sounds nice and does not blow up even at standard sample rate, within the range I've set for each param. Which is nice.

The plan was to use it for a bass like synth so I don't really need those high frequencies anyway

Anyway thanks for the tips

@jaffasplaffa btw when you change the upsampling you might want to change the coefficient to rpole~ (if you want the cutoff at a consistent point for all samplerates)

edit: changing the blocksize to account for the upsampling didn't work well when I tried, idk what to do but just be aware that the feedback time/cutoff are dependent on samplerate, not sure exactly how to fix it atm

@jaffasplaffa said:

I tried removing the block~ 1 a while ago, for the reason you mention, that fexpr~ is already processing every sample, but I did not get the same result, So kind of scratched it, but I will look into it again.

you need the block~ 1 for 1 sample feedback delay in a resonant filter

I am still beginner with filters, but the reason I used fexpr~ was because of the "processes every sample", so I was kind og thinking this is what I need.

for the lop~ are you sure that these filters are made in the same way, the lop~ and the karlsen filter?

yes, they are exactly the same. you can just stack 4 lop~ an feed back the output with one sample delay. or, if you want signal input for cutoff use rpole~ like this:

rpole is like [fexpr~ $x1 + $x2 * $y1] (see help) with the additional math it is what the fexpr~ in your patch does. same as lop~..

For the oversampling usually oversampling makes sure that it doesn't blow up at higher frequencies.

[clip~ 0.0001 0.8] prevents your patch from blowing up.

@solipp

Thanks for the info, more food for thought

Sounds terrific!!! Eats much CPU on my old machine but sounds wonderful!

@jaffasplaffa Two things which made the old moog ladder so good for basses which this misses are that the topology cancels even order distortion and the output stage loads filter which creates a second lowpass effect at a higher frequency and bends the waveform a bit. The distortion being primarily on the odd harmonics helped keep it from getting buzzy and gives a characteristic effect to the resonance. The filtering created by the loading of the ladder by the output stage varies with the cutoff of the main filter but at a higher frequency, this helps to tame the resonance and does filter the input signal some as well if it extends high enough. Later ladder filters have much less of this filtering effect, it is primarily a quirk of the moog mini/modular and the source of their nice warm bass sounds.

Either way, a great sounding filter as is and a definite keeper.

@jaffasplaffa I think the block size needs to stay @ 1, but you can divide the cutoff by the upsample factor to keep it approximately the same across samplerates

you might want to account for how many more times the feedback happens too, seems like at higher samplerate the resonance gets lower
edit: never mind I'm wrong the resonance is fine

Seb Harmonikar. your pd-abstraction saves a lot CPU-Time!
I'm really curious how much cpu savings you can get with the optimizations further.

@solipp

Checking your message again. I think I need to try the lop~ version too. So many options here, always great to learn new things

I always stayed away from filter design cause it seemed a bit overwhelming, as a hobbyist

@tungee

Try Sebs version, it's the lightest one for now, I think

@oid

Thanks for the info. Was actually wondering why/how the oscillators got the shapes it has on some analog synths. Need to look into that too then.

I have been working on a Model-D clone, not measuring anything, just copying the signal path, but might as well get as close as possible

First of all thank you so much for your filter, Jaffasplaffa. I have tried it, but with 4x Oversampling it gets also super heavy on my 2009 imac :D
I will try to integrate it into automatonism, if its ok for you?!

@tungee

Sure, you can use it for Automatism

Thinking about maybe getting a band pass filter out of it too and play around with the even harmonics saturation.

So maybe it will be updated, but it works pretty well as it is now, so if you can use it as it is now, go ahead

It coexists well in my Automatonism Fork

@tungee

Looks fine

Here is an image of the synth project I am working on. Need to fix a bunch of stuff, though. The oscillators I use screws up completely when modulating them in some cases. And it uses WAY to much cpu, so also working on that.

Anyway, here it is:

Looks delicious!
Are you usind Bandlimited Oscillators?!

@tungee

Yes and in some cases, if you use one of the oscillators as lfo and send the output back to itself(kind of fm feedback) it acts weird. I have seen it work in other projects, so I think I just need to try some other bandwidth limited oscs.