Paula documentation/reference wanted!

[8bitbubsy]

Aciddose (a dude on IRC) is helping me getting a very accurate A1200 sound emulation in my ProTracker clone, but we need more info on Paula! Like how many bits it takes in the sample input... Some sites say 8 bits, some say 9, and some say 12... And does paula do rounding on the mixing? Etc.
Any PDF or reference of Paula? I want as much information as possible, not just "Paula is the sound chip in the Amiga series"

[Minuous]

The official Amiga Hardware Reference Manual has this info...

[pandy71]

Minimig Verilog sources should be ok from Your point of view - however notice that for perfect Paula emulation You must work with sampling freq approx 1.79MHz.

Also this can be useful http://www.bel.fi/~alankila/modguide/interpolate.txt

btw i strongly disagree with Sinc resampling - Sinc introduce ALWAYS pre and post ringing - even if this is optimal from mathematical point of view it is also not optimal from subjective quality point of view.

"subjective quality" is a great term for that, since i suppose you can't describe it in scientific terms?

actually, in terms of quality sinc is ideal. the ringing can be minimized as in my implementation in bubsy's protracker clone. try this: close your eyes and listen to the audio, can you hear ringing, or can you hear exactly what comes out of a paula? have you attempted to record the output of your paula using your pc's soundcard, and did you notice any ringing then?

now we've completed our postulate and can move forward to the experimental stage:

record the output of a module playing in protracker on your amiga, select a specific section of audio, you only need about 50ms at most and ideally it would be made up of low frequency playback - samples played at the lowest notes.

look at it in your audio editor - note however that you need to view the samples with interpolation to see an approximation of the information actually represented by the samples and linear interpolation (straight lines) will not help you at all. most audio editors do not support interpolation when drawing between samples - one that does is cooledit96, i recommend you grab a copy of that for this experiment.

now, record the output of bubsy's protracker clone playing the same section of audio in the same module. compare it by listening at first, then look at the waveform to locate any ringing. (i assume you have something like a sb-live/audigy/etc which supports directly sampling pcm data without resampling)

try listening to the output of the pc connected into your commodore monitor while playing back the module in bubsy's tracker, and also with your amiga connected in the same way.

assuming the scientific method is a valid concept at all, and my descriptions for the experiment haven't left anything out your conclusion should be as follows:

- the output from bubsy's tracker looks better and sounds better than the paula sampled/resampled by your audio card. (#1)
- bubsy's tracker sounds closer to the amiga's output than your pc's sample of the amiga itself. (#2)

if not, we definitely need to discuss what went wrong.

(#1 under the assumption you dislike the ringing imposed upon the sampled version of the paula's output and our goal is to eliminate that)
(#2 except in the condition your soundcard has a better minimum phase resampling algorithm or we disregard the need to meet the condition assumed in #1.. which is possible, yet unlikely. most audio cards use filters with lots of ringing and linear phase)

[pandy71]

Originally Posted by aciddose

"subjective quality" is a great term for that, since i suppose you can't describe it in scientific terms?

It can be seen on any common video upscaler - and it can be described by math to - pre and post ringing are error that can be described.
Similar is in audio - Sinc is optimal only in frequency domain as a approximation of the brick filter with some limited (windowed) number of taps.

Originally Posted by aciddose

actually, in terms of quality sinc is ideal. the ringing can be minimized as in my implementation in bubsy's protracker clone.

optimal but only in frequency domain - FFT is not everything...

Originally Posted by aciddose

try this: close your eyes and listen to the audio, can you hear ringing, or can you hear exactly what comes out of a paula? have you attempted to record the output of your paula using your pc's soundcard, and did you notice any ringing then?


now we've completed our postulate and can move forward to the experimental stage:

--cut here--

Why not use decent filter with flat phase and magnitude response? this is only coefficient for software...
All proposed method are OK but they are also OK to prove that Sinc can be replaced by different type of interpolation and number of taps (or rather complexity of interpolator) can be similar or lower without pre and postriniging which not exist in source data.

Interpolation is like guessing - maybe this sample is 0.1 or maybe 0.0975 (but it can be 0.97 to)... there is no one optimal way to solve such problem - and this is personal - subjective choice - until You filter out aliasing and fulfil Nyquist criteria You are ok but way how it can be made or done it is different story.

there is no guessing involved, and this isn't interpolation.
flat phase and magnitude? that would be no filter at all. why not use that? it sounds like shit.

to perfectly reconstruct a sampled signal each sample must be replaced by a sinc impulse ringing at the target frequency - such is a perfect reconstruction. our eyes (when talking about video) prefer to see some aliasing since the aliasing noise approximates the appearance of blur and noise present in a normal visual signal such as we'd see looking at any ordinary object. if the display had a high enough resolution (dpi) to be beyond the ability of the eye to resolve, sinc would also be the ideal to use there. (also, even with the "annoying" ringing present in a low-res filtered visual signal, sinc provides the maximal amount of information and lowest possible noise. if you had ever designed a visual system which needs to gather information - which you clearly have not - you'd understand that what may be annoying for you is actually extremely beneficial in other cases)

in audio things are quite different. as long as the nyquist frequency is above the range of our hearing, the pre/post ringing has no effect. (it is inaudible) what is important is that the portions of the signal present within our hearing range remain unaffected - this is what a sinc reconstruction guarantees.

why not use a shaped sinc as you say with phase adjusted to minimize ringing? well, i do use such a filter as i already explained - also we're talking about inaudible ringing here so the primary concern is for headroom - to prevent clipping. your concern for things "present in the original signal" doesn't make any sense since the original signal is limited to below the nyquist frequency and we are modifying only above that frequency with such a filter.

another concern is cpu power - the filter i use requires five multiplies and five adds per input impulse, one add per output. if the input impulses occur far less frequently than the target sample rate (which they do, if we're talking about playing a 9241.86543hz sample at 48000hz) there are a fractional number of multiplies per destination sample.

if you want to talk about "subjective quality" have a listen to the output of bubsy's protracker clone.

if you want to talk about scientific properties of filters and such, you'd better study them and use them first - also please take into account that there are far more properties we might be interested in than in your generalized assumptions.

[pandy71]

I can only say - sorry for wasting Your precious time - feel free to code Sinc with "the filter i use requires five multiplies and five adds per input impuls".