Making filters more efficient (CPU)?

[guyman]

Hellooooooo . I hope the quarantine treateth you kindly.


I am fooling with both RBJ shelves, and Martin's 1 pole matched shelf filters... I am going to have static settings on the filters, no modulation of slope,res, cutoff nor amplitude. I know that stages and hop can make static calculation save a huge load on cpu, but some of these are in assembly, and I've never really implemented stages as of yet...

Would anyone like to chime in on the best way to make filter's super efficient on cpu? I'd love to have dozens of these running at any given moment. How would I go about altering a filter that's already in assem? it's latin to me.

[tektoog]

hey,
Hope you're doing good yourself!
IMHO, you have 3 choices...
Get help, learn latin or ... learn ASM... Latin being a dead language you're better off learning ASM
Most reasonable one would be to wait for someone willing to help...
Most pragmatic one would be to learn the languages you wanna write with...
But first thing to do, is to post what you would like people to help you with...
I'm no good at ASM, I just copy and paste most of the ASM code I use, and then suit it to my needs...
But I'm pretty sure you can find the original filter which was probably first coded with FS code... that's the case most of the time, and then, it's ported to ASM... that's easier (at least for me) to read, understand and then to alter
I'm pretty sure some good soul(s) over here will give you a hand...
Take care

[guyman]

thanks for the response. my concern is static filter optimization, without a loss of quality.

[juha_tp]

If those are static filters then why calculate anything in runtime ... use static coefficients.

EDIT:
Dunno if this would be an efficient way but worth to try (one filter fc@1000Hz):

Code: Select all

switch x
  case { 44100, 88200, 176400, 352800 }
    b(1) = x * ((1.71962425129703E-11 - 2.60013231898188E-17 * x) * x - 3.53741686355636E-06) + 1.67970352331451;
    b(2) = x * ((1.7196242512970E-11 - 2.6001323189818E-17 * x) * x - 3.5374168635563E-06) - 1.145371565931;
    a(1) = x * ((8.6185372138594E-12 - 1.3031531240154E-17 * x) * x - 1.7729081720323E-6) + 1.5464377224026;
    a(2) = x * ((8.6185372138593E-12 - 1.3031531240154E-17 * x) * x - 1.7729081720323E-6) - 1.27863736684291;

  case { 48000, 96000, 192000, 384000 }
    b(1) = x * ((1.33279187011394E-11 - 1.85142995570431E-17 * x) * x - 2.98433575868698E-6) + 1.65790207111673;
    b(2) = x * ((1.3327918701139E-11 - 1.8514299557042E-17 * x) * x - 2.9843357586869E-6) - 1.16717301812879;
    a(1) = x * ((6.67978270383326E-12 - 9.27913057754111E-18 * x) * x - 1.49571098311128E-6) + 1.53551111288653;
    a(2) = x * ((6.67978270383326E-12 - 9.27913057754111E-18 * x) * x - 1.49571098311128E-6) - 1.28956397635897;
  otherwise
    b = [1 0 ];
    a = [1 0 ];
  end


This calculation supports those commonly used sample rates but, you would need to make separate function for each filter.

[guyman]

I think I got it... I'll report back

[martinvicanek]

When I post filters I often include a static and a dynamic version. In the static version the filter coefficients are evayluated in green (or Ruby), while the filter iteration is optimized in ASM. Additional CPU gains can be had from Mono4 packing.