Queries on AGA Copper and 14-bit Audio

[pandy71]

Originally Posted by meynaf

I real life you get a little more than 14 bits, in fact. You should really have a look at how the calibration works.

I will be glad to read description (technical) but i think you are not electronics engineer when you said that 8 bit DAC may have over 14 bit accuracy (it can be done but then it must be 8 bit DAC made internally with over 14 bit accuracy which is usually not the case for technology in mid of 80's previous century).


Check popular DAC in those time (70's design i believe used in Paula prototyping stage) DAC08 - depends on version (grade) linearity was between +-0.1 to +-0.39%, +-0.1 means approx 0.2% linearity i.e. around 9 bit linearity - this is for best case scenario, now imagine 4 DAC like this in cheap non analogue process (as we know CBM/CSG was specialized in digital IC not analogue one) - well i would say it is highly doubtful that CBM/CSG will do expensive laser trimming for every 4 Paula DAC and in real life scenario Paula DAC linearity is more close to+-.39% or even more non-linearity and it is not possible to stack two 8 bit nonlinear DAC and expect higher linearity (unless nonlinearities in both DAC will compensate but this imply differential topology and lot of hassle at the output stage)
Gain from so called 14 bit is that every added bit reduce quantization noise by 6.0206dB and 10 - 12 dB is easily perceivable (for most of people 11 - 12 bit will sound unperceivable same as 16 bit, add to this power supply noise where in Amiga overall power plane analogue design is far from CD standard).

[pandy71]

please delete

[paul1981]

You think 11-12 bit audio sounds the same as 16-bit audio to the average user? I'd say this very much depends on the users' hifi system (or lo-fi system if one can't tell the difference between 12-bit audio and 16-bit). I tested 14-bit audio 20 years ago on the Amiga and it was 14-bit then, and it still is now! Listen to song fade outs or quiet sections or classical music (being as classical music is the most dynamic music available) and you'll plainly hear the extra resolution which is leaps and bounds beyond 8-bit....sounds a lot like 14-bit to me. Of course, I'd expect a proper calibration to have taken place first which takes about 30 or more minutes to do properly. It really is quite exciting to hear the better quality audio from good ol' 1985 Paula ;-)

[paul1981]

Originally Posted by meynaf

I real life you get a little more than 14 bits, in fact. You should really have a look at how the calibration works.

In real life there's less than 16-bits in 16-bit audio unless the signal is a constant full peak sinewave, so for 14-bit it's going to be less still, but it doesn't change the fact that the resolution capability still 16-bit respectively and 14-bit respectively.

[pandy71]

Originally Posted by paul1981

You think 11-12 bit audio sounds the same as 16-bit audio to the average user?

More or less yes - nowadays 1 bit audio converters deliver over 100dB linearity - (remember that i'm referring to DAC linearity/accuracy not resolution) - bits if properly used are not so important.

Originally Posted by paul1981

I'd say this very much depends on the users' hifi system (or lo-fi system if one can't tell the difference between 12-bit audio and 16-bit). I tested 14-bit audio 20 years ago on the Amiga and it was 14-bit then, and it still is now! Listen to song fade outs or quiet sections or classical music (being as classical music is the most dynamic music available) and you'll plainly hear the extra resolution which is leaps and bounds beyond 8-bit....sounds a lot like 14-bit to me. Of course, I'd expect a proper calibration to have taken place first which takes about 30 or more minutes to do properly. It really is quite exciting to hear the better quality audio from good ol' 1985 Paula ;-)

I can tell you only one thing - adding real 2 - 3 bits is more than necessary to dramatically improve perceived audio quality as quantization noise level drop bellow normal auditory threshold (normal listening environment conditions). 8 bit system has relatively high level of quantization noise and if you add to this sub-Nyquist sampling rate then you realize that Paula sound is not even close to HiFi type.

Slightly modifying Paula (introducing for example sample interleaving will first reduce amount of DAC's - from 4 to 2, secondly it will allow for increased sample rates by using two folded/time interleaved channels - this with proper dithering+noiseshaping even on 8 bit Paula will push quality close to 14 - 16 bit system without any or almost any CPU load - in theory 4 times oversampled, 8 bit PCM DAC is capable to beat DSD so high end, high resolution modern audio format).

Originally Posted by paul1981

In real life there's less than 16-bits in 16-bit audio unless the signal is a constant full peak sinewave, so for 14-bit it's going to be less still, but it doesn't change the fact that the resolution capability still 16-bit respectively and 14-bit respectively.

I strongly disagree - you should check Gerzon-Craven theorem - modern audio system with 16 bit DAC may easily have more than 120 - 130dB dynamics with noiseshaping matching human ATH (personally tested signals with level -140dBFS in 16 bit PCM but in theory you can easily put lower levels - limit is only Gerzon-Craven theorem).

[meynaf]

I maintain that you can get slightly more than 14 bit with Paula (with real Paula ; not in emulators which are really limited to 14 bit).

Remember that we use two 8-bit DAC for playing single 16-bit sample. If we could shift the volume 8 positions we would get full 16 bit. But we can only shift 6 bits (volume 64 vs volume 1). So in theory this is 14 bit, but the output is nonlinear so we have more combinations than just 16384 ; by reordering them cleverly we get over 14bit accuracy.

A good 14-bit routine plays 16-bit samples, not 14-bit samples.
Sure i'm not electronics engineer but i've seen the code and you have not.

[paul1981]

16-bit 44.1 kHz - 24-bit 192 KHz is crap, semi-modern D/A converters do a good job of covering this up...a very good job. It's trickery though.

1-bit DAC, are you talking about sdm? Paula doesn't perform any tricks to give a perceived enhanced output, so in this respect it will not sound as good as a CD player even, but the extra bit depth of 14-bit over 8 is very welcome indeed. Cd's were going to be 8-bit originally, heaven forbid...

[paul1981]

Sub-nyquist? I think we're all talking about 44.1 KHz here, which Paula will happily run at in 30KHz screen modes.

[pandy71]

Originally Posted by meynaf

I maintain that you can get slightly more than 14 bit with Paula (with real Paula ; not in emulators which are really limited to 14 bit).

Explain me how 6+8 can give more than 14 in terms of DAC accuracy if you
have only 8 bit accuracy at best.

Originally Posted by meynaf

Remember that we use two 8-bit DAC for playing single 16-bit sample. If we could shift the volume 8 positions we would get full 16 bit. But we can only shift 6 bits (volume 64 vs volume 1). So in theory this is 14 bit, but the output is nonlinear so we have more combinations than just 16384 ; by reordering them cleverly we get over 14bit accuracy.

It must very clever trick behind it - explain me this please - i believe lot of people can be interested to understand how it works - also it is a bit contradictory to my knowledge about real life circuits (not simple adding and subtracting in CPU register) as real DAC may have very complex non linearity distortions characteristic that require also nonlinear pre-correction.

Originally Posted by meynaf

A good 14-bit routine plays 16-bit samples, not 14-bit samples.
Sure i'm not electronics engineer but i've seen the code and you have not.

So once again asking for knowledge - please share knowledge - my engineer knowledge says:
16 bit sample must fit in 14 bits, this produce unavoidable quantization error (2 bits), such error must be processed somehow - worst case scenario is truncation but this lead in unavoidable way to harmonics distortions which are highly correlated with signal so they can be easily perceived - to avoid this type distortion you need to de-correlate quantization distortions and usually it is done by adding TPDF dither (at least 2LSB) to signal before re-quantization, noiseshaping can be used instead or together with dither to improve perceived SNR and perceived system dynamics.

Originally Posted by paul1981

16-bit 44.1 kHz - 24-bit 192 KHz is crap, semi-modern D/A converters do a good job of covering this up...a very good job. It's trickery though.

Are you able to provide explanation for this? I'm really surprised to read that 24-bit 192 KHz is crap.

Originally Posted by paul1981

1-bit DAC, are you talking about sdm? Paula doesn't perform any tricks to give a perceived enhanced output, so in this respect it will not sound as good as a CD player even, but the extra bit depth of 14-bit over 8 is very welcome indeed. Cd's were going to be 8-bit originally, heaven forbid...

1 bit is extreme example to illustrate that number of bits may not be so important.
Everything is more complex and trust me - 8 bit may be better than 16 if proper signal processing is applied to it.

Originally Posted by paul1981

Sub-nyquist? I think we're all talking about 44.1 KHz here, which Paula will happily run at in 30KHz screen modes.

Well... Paula may run with increased sample rates but nominal system specification can be considered as sub-Nyquist DAC - to run Paula with increased sample rate you need to increase horizontal frequency rate, not sure if someone tried to do this in software on OCS/ICS (in theory this is possible in similar fashion as fake 60Hz modes only more difficult), so practically increased horizontal frequency require ECS/AGA or you need to start using audio DAC in non-DMA mode (so involve or CPU or/and Copper).

[meynaf]

Originally Posted by pandy71

Explain me how 6+8 can give more than 14 in terms of DAC accuracy if you have only 8 bit accuracy at best.

I already explained.
Two 8-bit samples give 65536 combinations. Not all of them are useful but we can get 14bit and more.

Originally Posted by pandy71

It must very clever trick behind it - explain me this please - i believe lot of people can be interested to understand how it works - also it is a bit contradictory to my knowledge about real life circuits (not simple adding and subtracting in CPU register) as real DAC may have very complex non linearity distortions characteristic that require also nonlinear pre-correction.

The 8+6 that is made isn't simple 8+6 (unless no calibration file is found).
The input is a 16-bit sample. The output is two 8-bit samples, both fetched from a table. You can call that nonlinear pre-correction if you want.
I can't really tell the theory behind this, only that it works fine.

Originally Posted by pandy71

So once again asking for knowledge - please share knowledge - my engineer knowledge says:
16 bit sample must fit in 14 bits, this produce unavoidable quantization error (2 bits), such error must be processed somehow - worst case scenario is truncation but this lead in unavoidable way to harmonics distortions which are highly correlated with signal so they can be easily perceived - to avoid this type distortion you need to de-correlate quantization distortions and usually it is done by adding TPDF dither (at least 2LSB) to signal before re-quantization, noiseshaping can be used instead or together with dither to improve perceived SNR and perceived system dynamics.

With normal volume levels you don't hear any quantization at all, as the quantization error is weak - 2bit in a 16bit sample is very quiet - and i suspect most of it will be filtered out by the hardware (after all this is just high frequencies).

Have you heard an Amiga playing 14bit ?

[paul1981]

@pandy71

Here's the 14-bit 'Cybersound' software enclosed in the Play16 archive. With Cybersound you can calibrate your particular Amiga, and with Play16 you can utilise that calibration via a tooltype (paula14c IIRC) and play some cd quality wavs or aiff's - read docs! The Cybersound docs explain exactly how it works, I think.

http://aminet.net/package/mus/play/Play16

[pandy71]

Originally Posted by meynaf

I already explained.
Two 8-bit samples give 65536 combinations. Not all of them are useful but we can get 14bit and more.

Seem you are confusing adding in register with DAC accuracy - going this way by stacking 4 8 bit DAC together we can have 32 bit audio converter...

Originally Posted by meynaf

The 8+6 that is made isn't simple 8+6 (unless no calibration file is found).
The input is a 16-bit sample. The output is two 8-bit samples, both fetched from a table. You can call that nonlinear pre-correction if you want.
I can't really tell the theory behind this, only that it works fine.

Well i have impression that 8+6=14 however if you need to do calibration i.e. not adding 8+6 but less then you will never have 14.
By calibration you accepting fact that you have less than 14 bits useful.

Originally Posted by meynaf

With normal volume levels you don't hear any quantization at all, as the quantization error is weak - 2bit in a 16bit sample is very quiet - and i suspect most of it will be filtered out by the hardware (after all this is just high frequencies).

Please elaborate on this - calibration make less than 14 bit usable...

Originally Posted by meynaf

Have you heard an Amiga playing 14bit ?

Have you heard 6 bit DAC or 3 bit playing? Once again there is no 14 bit audio in Amiga as after calibration you will have less than 14 bits usable, yes, i've heard Amiga in so called 14 bit mode and i have impression that this mode is somewhere between 10 and 12 bits maximum (for 12 assumption is properly made calibration with spectrum analyzer and improved power supply section for Paula).

Originally Posted by paul1981

@pandy71

Here's the 14-bit 'Cybersound' software enclosed in the Play16 archive. With Cybersound you can calibrate your particular Amiga, and with Play16 you can utilise that calibration via a tooltype (paula14c IIRC) and play some cd quality wavs or aiff's - read docs! The Cybersound docs explain exactly how it works, I think.

http://aminet.net/package/mus/play/Play16

Thx Paul, will try to read, already found tool for 14 bit calibration from Christian Buchner and based on brief reading i can say that after calibration there can't be 14 bits but OK - will try read source.

[meynaf]

Originally Posted by pandy71

Seem you are confusing adding in register with DAC accuracy - going this way by stacking 4 8 bit DAC together we can have 32 bit audio converter...

No confusion, no. At least not in my side.
What is a 8 bit DAC for you ? Isn't it the same as 8 1-bit DAC stacked together, each one giving 1/2 of the previous one's output level ?
Listen to Atari ST sampled sound - it is 3 4-bit logarithmic DAC, all together giving close to 8 bit accuracy. Here it's more or less the same trick.

Originally Posted by pandy71

Well i have impression that 8+6=14 however if you need to do calibration i.e. not adding 8+6 but less then you will never have 14.
By calibration you accepting fact that you have less than 14 bits useful.

As i explained the calibrated output uses 8+8, not 8+6.

Originally Posted by pandy71

Please elaborate on this - calibration make less than 14 bit usable...

For 14 bit see above, and my previous posts. I have explained it.
For the quantization noise, don't forget that the final output isn't exactly the same as the input signal.

Originally Posted by pandy71

Have you heard 6 bit DAC or 3 bit playing?

Irrelevant. What are you trying to do ?

Originally Posted by pandy71

Once again there is no 14 bit audio in Amiga as after calibration you will have less than 14 bits usable, yes, i've heard Amiga in so called 14 bit mode and i have impression that this mode is somewhere between 10 and 12 bits maximum (for 12 assumption is properly made calibration with spectrum analyzer and improved power supply section for Paula).

There is more than 14 bit usable, not less. I have explained it several times.
Well, obviously you're standing at your theoretical level and you haven't tried.

My guess is that if you were listening at some true 16 bit sound, then the same with a properly configured 14-bit Amiga config, you wouldn't be able to tell the difference.
(Pretty much like some folk - you know who you are my friend - who pretended downsample by average was very poor, and yet has been proven unable to tell who's who with a blind testing.)

[daxb]

Unfortunately, the Play16/Cybersound docs don`t explain in detail how it works. "Only" the usual: "1/2 channel with full volume mixed with 1/2 channel with min volume". All 4 audio channels used.

[Thorham]

Originally Posted by meynaf

Pretty much like some folk - you know who you are my friend - who pretended downsample by average was very poor, and yet has been proven unable to tell who's who with a blind testing.

I didn't pretend anything. I made a claim based on a test I had run before, and I probably screwed it up. The result of the test was quite perceptible. I just redid the test, but this time with a simple 2:1 average, and sure enough, it doesn't sound like crap. People make mistakes you know

[meynaf]

Originally Posted by Thorham

I didn't pretend anything. I made a claim based on a test I had run before, and I probably screwed it up. The result of the test was quite perceptible. I just redid the test, but this time with a simple 2:1 average, and sure enough, it doesn't sound like crap. People make mistakes you know

Sorry if i've been rude, my friend. It's been quite a long time i play things with 14-bit, downsampled or not, so i know how it sounds. I have the code to do that myself. And so, what can i say when i read claims saying it doesn't work this way ?

[pandy71]

Originally Posted by meynaf

No confusion, no. At least not in my side.
What is a 8 bit DAC for you ? Isn't it the same as 8 1-bit DAC stacked together, each one giving 1/2 of the previous one's output level ?
Listen to Atari ST sampled sound - it is 3 4-bit logarithmic DAC, all together giving close to 8 bit accuracy. Here it's more or less the same trick.

Well it looks like you not confused but OK let assume that 8 bit DAC is 8 1-bit DAC stacked together (it is not like this from circuit perspective and they are various circuit topologies) - what is source for multibit DAC nonlinearities accrodingly to you ? - as a coder you see only numbers and registers assuming that DAC is plain and linear extension where in real life analogy between arithmetic in ALU and real electric current is not direct.
I know how SID and AY tricks works but how this is linked with Paula?

Originally Posted by meynaf

As i explained the calibrated output uses 8+8, not 8+6.

Problem is that you didn't explain anything - side to this you are ignoring fact that calibration assume less bits than in theory available (did you perform re-quantization on 8/6 sample part? - asked you about algorithm used in calibration)

Originally Posted by meynaf

For 14 bit see above, and my previous posts. I have explained it.
For the quantization noise, don't forget that the final output isn't exactly the same as the input signal.

You didn't explained anything and this is problem from my perspective as you expect me to believe when i'm trying to understand.
How output is different on final output when compared to input signal (assumption is 16 bit samples at input) - please briefly describe steps from input to output.

Originally Posted by meynaf

Irrelevant. What are you trying to do ?

Understand how to achieve real 14 bit DAC accuracy on Paula.

Originally Posted by meynaf

There is more than 14 bit usable, not less. I have explained it several times.
Well, obviously you're standing at your theoretical level and you haven't tried.

How there can be more than 14 usable in Paula with calibration - please explain as you didn't explain anything.
My theoretical level is that on Paula with calibration you may have between 10 and 12 bit DAC accuracy - you can fit even 64 bit data but at the end it will not more than 12 bits.

Originally Posted by meynaf

My guess is that if you were listening at some true 16 bit sound, then the same with a properly configured 14-bit Amiga config, you wouldn't be able to tell the difference.

I can tell you when you connecting Amiga to amplifier input without any signal as power plane is so noisy in Amiga... And there is substantial difference between real audio system capable to deliver real 16 bit performance and Amiga.

[NorthWay]

Erm... if the volume control scaling is not linear (not using same size steps as the DAC), does that not make it possible to get better than 14 bit result? (Or worse for that matter.)

I.e. smallest sample value (well, isn't it the difference between two values really?) at full volume is louder than the biggest sample value at minimum volume?

ISTR that I did some tuning for a 14-bit playback thing that involved turning up the stereo something scary.

[robinsonb5]

Originally Posted by pandy71

How there can be more than 14 usable in Paula with calibration - please explain as you didn't explain anything.

I think the point Meynaf's trying to make is that when two channels are combined to do 14-bit, the channel that's at 1/64 volume is still an eight-bit DAC, and some of its range overlaps the range of the full-volume channel.

This means we can input 65536 unique codes into the combined DAC, rather than the 16384 that a true 14-bit DAC could accept.

It's highly likely that non-linearities in the DACs mean that the vast majority of those 65536 codes give a unique output level, and with careful calibration (that the current calibration software may well not be doing sufficiently well) it may be possible to achieve higher than 14-bit resolution, but only in very specific (and not very useful!) parts of the range.

[pandy71]

Originally Posted by NorthWay

Erm... if the volume control scaling is not linear (not using same size steps as the DAC), does that not make it possible to get better than 14 bit result? (Or worse for that matter.)

It is not important as we have 8 bit DAC with level 1 and we not touching 8 bit samples (with exception of calibration as they partially overlapped will address this later as pointed byrobinsonb5).

Originally Posted by NorthWay

I.e. smallest sample value (well, isn't it the difference between two values really?) at full volume is louder than the biggest sample value at minimum volume?

Nope - accordingly to HRM for AUDxVOL=1 attenuation is approx -36.1dB so this is maximum dynamics that can be added to (theoretical) 8 bit DAC dynamics.

Originally Posted by robinsonb5

I think the point Meynaf's trying to make is that when two channels are combined to do 14-bit, the channel that's at 1/64 volume is still an eight-bit DAC, and some of its range overlaps the range of the full-volume channel.

Yes, fully agree but it is not a problem - problem is how to select from 65356 sample values corresponding 16383 and with calibration less than 16383 (as this is lack of linearity - some values are not linear and they are lost i.e. for 1024 voltage or current should have particular value x when for 1025 it should be x+(1/16383) but it can be x i.e. missing code i.e. non-linearity)

Originally Posted by robinsonb5

This means we can input 65536 unique codes into the combined DAC, rather than the 16384 that a true 14-bit DAC could accept.

Nope - this means that 2 LSB from upper 8 bit overlapping partially with lower 8.

Originally Posted by robinsonb5

It's highly likely that non-linearities in the DACs mean that the vast majority of those 65536 codes give a unique output level, and with careful calibration (that the current calibration software may well not be doing sufficiently well) it may be possible to achieve higher than 14-bit resolution, but only in very specific (and not very useful!) parts of the range.

Nope as you reducing volume by only -36dB so ranges from upper 8 bit and lower 8 bit are partially overlapping, also both 8 bit DAC suffer from non-linearity and based on fact that they share same die is more likely like they will have very similar characteristic of non-linearity i.e. can't be nullified by simple combining DAC (it could be compensated if you switch DAC's to pseudo-differential configuration - feeding one of DAC's with negated sample values - this will extend they dynamic range by 1 bit so two 8 bit DAC in pseudo-differential configuration will behave like 9 bit DAC) - stacking two very similar non linear DAC will only increase issue so they will be never close to theoretical 16 bits - more realistically they will be rather 12 bit.
Calibration is very important but in fact it will further reduce number of available sample values... this is quite.