Re: [hatari-devel] New cycle exact YM2149 emulation

[Nicolas Pomarède <npomarede@xxxxxxxxxxxx>]

Le 19/11/2017 à 20:31, Nicolas Pomarède a écrit :
> Le 19/11/2017 à 11:55, Eero Tamminen a écrit :
> > Hi,
> >
> > On 06/18/2017 07:01 AM, David Savinkoff wrote:
> > > ----- Nicolas Pomarède wrote:
> > > >   From a few comparisons with my STF, I would say that Hatari result 
> > > > is a
> > > > little bit too attenuated and should be a little more "sharp", but not
> > > > as much as when turning filter off completely.
> > >
> > > This patch changes filter coefficients for a sharper sound.
> > > I can make more patches until the timbre is right.
> >
> > Nicolas, David, was there any follow up for this?
>
> I have a new ym2149 emulation ready, need to push it to main repo soon.

Hi

I just pushed the new YM2149 emulation code.

So far, no change in the coefficients were necessary (but maybe it will 
need adjustements later if someone reports diffrerences with real HW).

As for this new emulation, it now generates all the samples the YM2149 
would output when running at 250 KHz ; the previous improved version 
only updated the YM2149 counters with a simulated 250 KHz freq, but the 
output samples were converted on the fly to the output freq (44.1 KHz by 
default).

This resulted in too much filtering and loss of high frequencies.

With the new method, all frequency in the input signal are kept (because 
we generate a signal at 250 KHz like the YM2149 would do).

The critic part is then do downsample this 250 KHz signal to the chosen 
output freq of 44.1 KHz. This can be much harder than it sounds, 
there're tons of very technical paper/thesis on how to do this 
accurately when the Nyquist criteria is not respected.

Also, there's no miracle, if input signal is 250 KHz and output freq is 
44.1 KHz, then some harmonics will be lost.

Still, I added 2 downsampling methods (more to come later) :
 - a very simple one taking the nearest sample ; it's fast, but not 
very accurate for high frequency.
 - a more efficient one using a weighted average of the 2 input samples

The latest is the default method in Hatari for now (this is not user 
configurable at the moment, will see later, not sure it's useful)

CPU load should be a little higher with this new 250 KHz quality, but 
compared to the time to emulate CPU and video, I don't think the 
increase should be noticeable.

If you want to compare the improvements, you can run 2 versions of 
Hatari side by side playing the same music / demo (run Hatari 2.0 and 
compare with the latest devel version) ; pause the 1st, listen to the 
2nd for a few seconds, then pause the 2nd and listen to the 1st one.

The difference should be hearable, especially with musics using volume 
modulation with MFP (so called SID sound effect). The result is a little 
less louder, high freq can be heard more distinctively.

Let me know what you think and if you find some musics that sound worse 
(hopefully not :) )

Nicolas