Re: [Audacity-devel] Speed up Square, no alias generator

[Roger D. <rb...@cs...>]

This has me curious. For what it's worth, Nyquist *is* C code (at some 
level) and often faster than you could write by hand. There's some 
overhead to interpret the plug-in code, but often there's no 
interpretation once the DSP begins. Some methods to generate a low-alias 
square wave in Nyquist include:

sum of sines (Nyquist has a function called partial that is very fast 
for additive synthesis)

make a highly-oversampled single period using sum of sines, then use 
that as a (linearly interpolated) wave table, the oversampled table 
reduces aliasing you'd get otherwise

make a non-oversampled single period, synthesize it at the closest 
integer period to avoid aliasing, and then resample it to the exact 
period you need (resampling is expensive so I would not place bets on 
this method winning on speed)

-Roger


On 3/30/12 5:02 PM, James Crook wrote:
> On 30/03/2012 20:39, Steve the Fiddle wrote:
>> Thanks for the reply James, but please do try comparing my plug-in
>> generator  with the Audacity no alias square wave generator. I
>> understand the theoretical arguments but in practice my tests indicate
>> that the waveform produced by the Nyquist version is at least as good,
>> and in the case of low frequency square waved it is superior to the
>> built-in effect. Also it is very much faster.
> Would you like to have a go at translating your code to C/C++?
>
> I don't completely understand the nyquist code.  Are you creating the
> wavetable at a higher resolution than you need it, then using FFT (low
> pass filter) to make it alias free?  If so, I don't understand why the
> lowpass parameters differ and why they are absent on tri-wav.  If we try
> to use that technique in tonegen.cpp when the frequency is varying, as
> it can in tonegen, the filter will need to vary too, so will need to be
> applied repeatedly.  Should still be cheaper than calculating zillions
> of sines - up to 400 per sample in Audacity code - which is craziness.
>
> How important is alias free tone generation? Standard square wave
> generation is a lot faster than with no-alias.
>
> --James.
>
>
>> If there is little interest in improving the Audacity version then I
>> suggest that it is replaced by a better quality, faster version
>> written as a Nyquist plug-in (such as the one attached).
>>
>> Steve
>>
>> On 30 March 2012 16:06, James Crook<cr...@in...>   wrote:
>>> Wavetables (in general) work when the period of the wave is an exact
>>> whole number of samples.  Our built-in generators do not assume that, so
>>> we would need to interpolate from wavetables.
>>>
>>> There is an argument for detecting the repeats-after-n-samples case and
>>> doing that quickly.  We can even reference the same audio block
>>> repeatedly if the amplitude and frequency are constant - which would
>>> give very fast generation of very long regular waves.
>>>
>>> Steve, please have a look at our current bugzilla.  I would put this at
>>> P5 if it were a bugzilla enhancement request.  Worth noting, but not a
>>> pressing issue.
>>>
>>> --James.
>>>
>>> On 30/03/2012 14:56, Steve the Fiddle wrote:
>>>> The "Square, no alias" generator is extremely slow, which appears to
>>>> be because the entire waveform is generated by additive synthesis.
>>>> Attached is a Nyquist generator plug-in that is much faster generating
>>>> alias free square waves, particularly when generating waveforms longer
>>>> than 10 seconds.
>>>>
>>>> One would expect a built-in generator to be significantly faster than
>>>> a Nyquist implementation, but the Nyquist version achieves a huge
>>>> speed boost by generating a wavetable for the waveform, then using the
>>>> wavetable to generate the waveform. I presume that this method could
>>>> be implemented in C+ to make the "Square, no alias" generator very
>>>> substantially faster.
>>>>
>
> ------------------------------------------------------------------------------
> This SF email is sponsosred by:
> Try Windows Azure free for 90 days Click Here
> http://p.sf.net/sfu/sfd2d-msazure
> _______________________________________________
> audacity-devel mailing list
> aud...@li...
> https://lists.sourceforge.net/lists/listinfo/audacity-devel