4 ms·

Here's someone else's fast-sine code snippet I found and used about a decade ago when I was working on an additive synthesizer I never finished. It basically us

Here's someone else's fast-sine code snippet I found and used about a decade ago when I was working on an additive synthesizer I never finished. It basically uses the same method, Taylor series approximation. Driving it as an oscillator relies on just a single integer addition operation per sample to increment the phase, and relies on integer overflow to reset the phase. I like the comment from someone asking if it could really be so fast since it has 7 whole multiplications in it. Ah I love the realtime dsp world, so different from my python/JS day job. https://www.musicdsp.org/en/latest/Synthesis/261-fast-sin-approximation-for-usage-in-e-g-additive-synthesizers.html https://www.musicdsp.org/en/latest/Synthesis/261-fast-sin-ap...

IIAOPSW 4y ago
You might be interested to know about mod 2^m -1. Look at these numbers: 32640, 15420, 39321, 43690 In binary they look like: 1111111100000000, 0111100001111000, 0011001100110011, 0101010101010101 How cute. They are like little one bit precision sine waves. I assure you that these numbers multiplied by each other mod (2^16 -1) are all equal to zero. When multiplied by themselves they shift position. test it for yourself. Long story short, in the field of mod 2^16 -1, 2^x works just like e^pi i, these numbers function just like sine and cosine, and you can even make linear superposition out of them and use them as a basis to decompose other ints into "frequency space." Even better, the number i is represented here. There's a number int this modulo system which squares to 1111111111111110 (aka -1). Once you see it, you can implement fast sine and cosine directly through 2^{ix} = cos(x) + i sin(x). Cool right. Have fun.
- squeaky-clean 4y ago
  Whoa that is awesome. You and Sesse__ are really inspiring me to get a VST dev environment set up again this weekend.
Sesse__ 4y ago
If you're doing a synth and want a sine wave, you usually won't need to compute cos(x + at) from scratch every time. There's a cheaper and more accurate trick. First, figure out how much the phase (in radians) is going to increase every sample. E.g., for 440 Hz at 44100 Hz sample rate, you want 2pi * 440 / 44100 = 0.06268937721 (or so). Precompute sin(a) and cos(a), once. Now your oscillator is represented as a point in 2D space. Start it at [1,0]. Every sample, you rotate this point a radians around the origin by a radians (this is a 2x2 matrix multiplication, using the aforemented sin(a) and cos(a) constants, so four muladds instead of seven). You can now directly read off your desired cos value; it's the x coordinate. You get sin for free if you need it (in the y coordinate). If you run this for millions or billions of samples (most likely not; few notes are longer than a few seconds), the accumulated numerical errors will cause your [x,y] vector to be of length different from 1; if so, just renormalize every now and then. Of course, if you need to change a during the course of a note, e.g. for frequency or phase modulation (FM/PM), this won't work. But if so, you probably already have aliasing problems and usually can't rely on a simple cos() call anyway.
- squeaky-clean 4y ago
  That is a really cool trick, I think I'll give this a try over the weekend. I'm a bit confused about pre-computing the table though, wouldn't it have to be massive since the sample rate and note frequency don't evenly line up? On my phone at the minute but with a 440hz note under 44.1k sr, doesn't one cycle equal 100.227~ samples? I'm guessing the idea is to make a smaller table and interpolate? And I'm guessing you compute these tables for each musical note (or a limited but realistic range) at compile or load time, and not with each note press. Not being able to change the frequency while the note is running seems like a big deal for a synth though, no? Even without high frequency FM/PM which is often okay to not support, you would still be giving up on things like legato, lfo pitch mod (vibrato) and pitch wheel bending. Or is this only a problem in the higher frequencies of modulation? You know, it sounds like a fun thing to find out myself lol. The accumulating errors could be an issue with free-running oscillators, especially in a studio system left on for a few days. Free-running oscillators aren't that common in additive synths I believe, but it could be neat to see if it makes a difference like it does in a subtractive synth. Either way, periodic renormalization should be pretty straightforward. Also I do somewhat remember when I was doing my original research for an additive synth, the word on KVR was a lookup table based sin was fastest as long as you could fit it into L1 cache. That was over a decade ago and caches are a lot bigger now.
  - IIAOPSW 4y ago
    I think I can explain the trick even easier and hopefully that clears up your confusion. Simply store a complex number s which represents the current output of the synth (cos on the real part, sin on the imag). At each time update (the sample rate), multiply s by a pre-computed constant c (which is also complex valued). You just need for c^(i/sample_rate) to equal e^(i*freq/2pi). It does not matter that freq/2pi doesn't line up with the sample rate. The trick is wholly dependent on just keeping track of the state s and reliably multiplying it at a known update frequency.
  - Sesse__ 4y ago
    You don't need a table for a simple sine oscillator. You just compute sin(a) and cos(a) on note-on, and that's it. You don't store the output in a table, you just apply that 2x2 matrix every time. Tables can definitely have their place, depending on your synthesis method. If you want something like pitch bend, going for a sine table may very well be the fastest (it depends a bit on how much cache you want to burn).

IIAOPSW 4y ago

You might be interested to know about mod 2^m -1. Look at these numbers: 32640, 15420, 39321, 43690 In binary they look like: 1111111100000000, 0111100001111000, 0011001100110011, 0101010101010101 How cute. They are like little one bit precision sine waves. I assure you that these numbers multiplied by each other mod (2^16 -1) are all equal to zero. When multiplied by themselves they shift position. test it for yourself. Long story short, in the field of mod 2^16 -1, 2^x works just like e^pi i, these numbers function just like sine and cosine, and you can even make linear superposition out of them and use them as a basis to decompose other ints into "frequency space." Even better, the number i is represented here. There's a number int this modulo system which squares to 1111111111111110 (aka -1). Once you see it, you can implement fast sine and cosine directly through 2^{ix} = cos(x) + i sin(x). Cool right. Have fun.

squeaky-clean 4y ago
Whoa that is awesome. You and Sesse__ are really inspiring me to get a VST dev environment set up again this weekend.

Sesse__ 4y ago

If you're doing a synth and want a sine wave, you usually won't need to compute cos(x + at) from scratch every time. There's a cheaper and more accurate trick. First, figure out how much the phase (in radians) is going to increase every sample. E.g., for 440 Hz at 44100 Hz sample rate, you want 2pi * 440 / 44100 = 0.06268937721 (or so). Precompute sin(a) and cos(a), once. Now your oscillator is represented as a point in 2D space. Start it at [1,0]. Every sample, you rotate this point a radians around the origin by a radians (this is a 2x2 matrix multiplication, using the aforemented sin(a) and cos(a) constants, so four muladds instead of seven). You can now directly read off your desired cos value; it's the x coordinate. You get sin for free if you need it (in the y coordinate). If you run this for millions or billions of samples (most likely not; few notes are longer than a few seconds), the accumulated numerical errors will cause your [x,y] vector to be of length different from 1; if so, just renormalize every now and then. Of course, if you need to change a during the course of a note, e.g. for frequency or phase modulation (FM/PM), this won't work. But if so, you probably already have aliasing problems and usually can't rely on a simple cos() call anyway.

squeaky-clean 4y ago
That is a really cool trick, I think I'll give this a try over the weekend. I'm a bit confused about pre-computing the table though, wouldn't it have to be massive since the sample rate and note frequency don't evenly line up? On my phone at the minute but with a 440hz note under 44.1k sr, doesn't one cycle equal 100.227~ samples? I'm guessing the idea is to make a smaller table and interpolate? And I'm guessing you compute these tables for each musical note (or a limited but realistic range) at compile or load time, and not with each note press. Not being able to change the frequency while the note is running seems like a big deal for a synth though, no? Even without high frequency FM/PM which is often okay to not support, you would still be giving up on things like legato, lfo pitch mod (vibrato) and pitch wheel bending. Or is this only a problem in the higher frequencies of modulation? You know, it sounds like a fun thing to find out myself lol. The accumulating errors could be an issue with free-running oscillators, especially in a studio system left on for a few days. Free-running oscillators aren't that common in additive synths I believe, but it could be neat to see if it makes a difference like it does in a subtractive synth. Either way, periodic renormalization should be pretty straightforward. Also I do somewhat remember when I was doing my original research for an additive synth, the word on KVR was a lookup table based sin was fastest as long as you could fit it into L1 cache. That was over a decade ago and caches are a lot bigger now.
- IIAOPSW 4y ago
  I think I can explain the trick even easier and hopefully that clears up your confusion. Simply store a complex number s which represents the current output of the synth (cos on the real part, sin on the imag). At each time update (the sample rate), multiply s by a pre-computed constant c (which is also complex valued). You just need for c^(i/sample_rate) to equal e^(i*freq/2pi). It does not matter that freq/2pi doesn't line up with the sample rate. The trick is wholly dependent on just keeping track of the state s and reliably multiplying it at a known update frequency.
- Sesse__ 4y ago
  You don't need a table for a simple sine oscillator. You just compute sin(a) and cos(a) on note-on, and that's it. You don't store the output in a table, you just apply that 2x2 matrix every time. Tables can definitely have their place, depending on your synthesis method. If you want something like pitch bend, going for a sine table may very well be the fastest (it depends a bit on how much cache you want to burn).