On Mon, 1 Nov 2021 02:05:17 +0100 Ichthyostega <p...@ichthyostega.de> wrote:
>Am 31.10.21 um 00:54 schrieb Ichthyostega: >> When comparing with a build /before/ the "power"-branch, I >> see differences around -125dB, ... > >> But there is again that strange ramp-up (on the sound difference!) on one >> single note, while for all other notes the sound difference is basically >> noise. > > >Hi Will, >Hi Kristian, > >after some more hours of poking around with the debugger >(no, I do not especially like that task...), I consider this case as solved. > >Bottom line: in this case, these differences are the unavoidable consequence >of the change from powf(2.0,...) to power<2>(...). > >By switching to the latter, we (hopefully) get stable computations from now on. > > > >My findings in detail: > >- The AnalogFilter is *not* the root cause. > But the filter somehow increases the problem, and it adds > the seemingly random fluctuations on top of the difference > >- But the root cause of the differences happens already when > (pre)computing the note scale frequencies (Part::PnoteMap). > >The differences arise in Microtonal::getFixedNoteFreq(). > >So the seemingly dramatic "ramp-up" is more like an artefact of computing >the difference between a baseline waveform and an actual waveform. >Simply, because the *frequency* on *some notes* is slightly off, >and thus the visible phase drifts away as the note proceeds. > >We can completely confirm the differing frequencies by calculating >powf(2.0, ...) vs power<2>(...), for example for the first 10 notes: > > >note(-64) frequency = 0.202785403 | 0.202785417 Hz Δ 0.000206379169 cent >note(-63) frequency = 0.21484375 | 0.21484375 Hz Δ 0 cent >note(-62) frequency = 0.227618977 | 0.227618977 Hz Δ 0 cent >note(-61) frequency = 0.241153851 | 0.241153896 Hz Δ 0.000412758313 cent >note(-60) frequency = 0.255493701 | 0.25549373 Hz Δ 0.000206379169 cent >note(-59) frequency = 0.27068609 | 0.270686179 Hz Δ 0.000619137432 cent >note(-58) frequency = 0.286781877 | 0.286781818 Hz Δ -0.000309568799 cent >note(-57) frequency = 0.303834945 | 0.303834856 Hz Δ -0.000515948029 cent >note(-56) frequency = 0.321901828 | 0.321901798 Hz Δ -0.000206379193 cent >note(-55) frequency = 0.341043025 | 0.341043025 Hz Δ 0 cent >note(-54) frequency = 0.361322671 | 0.361322641 Hz Δ -0.000103189594 cent > > >Thus some notes turn out exactly the same, while others are off by a very >minute amount, which, by itself, is way below anything audible. However, >it is conceivable that *beating patterns* of some complex textures might >be changed. > >So this turns out to be a similar situation as recently, when changing >the denormal handling from hardware based to a stable, software based >implementation: a small change, that can be audible under certain conditions, >as the price for getting a numerically stable and reproducible computation. > >-- Hermann Ah! This makes a lot of sense now! It makes me wonder how many other subtle changes we've made over the years that have gone unnoticed but could produce an audible difference or specific extreme circumstances - not that we could do anything about it now. This is clearly one of those "nature of the beast" situations, so thanks for such a deep investigation. I think it's something we just have to accept. -- Will J Godfrey https://willgodfrey.bandcamp.com/ http://yoshimi.github.io Say you have a poem and I have a tune. Exchange them and we can both have a poem, a tune, and a song. _______________________________________________ Yoshimi-devel mailing list Yoshimi-devel@lists.sourceforge.net https://lists.sourceforge.net/lists/listinfo/yoshimi-devel
