Re: [hatari-devel] Improved YM2149 cycle accuracy

[Nicolas Pomarède <npomarede@xxxxxxxxxxxx>]

Le 26/07/2021 à 21:46, David Savinkoff a écrit :
> On Sun, Jul 25, 2021 at 12:57 PM David Savinkoff wrote:
>
>     Hi Nicolas,
>
>     I looked at the code and found that some filter coefficients need to
>     be converted to 250 KHz. The enclosed patch does the job.
>     I made the patch without testing or compiling because it is
>     simple and I'm not set up for Hatari at the moment.
>
>     Once the patch is applied, only YM2149_Next_Resample_Nearest()
>     is needed (the best choice). The IIR filters are as efficient as
>     the YM2149_Next_Resample_Weighted_Average_N() method.
>
>     Hatari should have the same good timbre as it did when the filters
>     were implemented a few years ago
>
>     David Savinkoff
>
>     ps. If the sound is wrong, I'll have to double check the coefficients.
>
> More comments and coefficients improved to 1% of target fc @ 250KHz

Hi

I tried your patch and the sound is slighly less "sharp", as if some 
high freqs were more filtered, but the difference is not huge (you can 
check in STF mode with syncsqu.snd I attached earlier)

Hard to tell which one is the closest to a real STF :)


As for YM2149_Next_Resample_Nearest(), I don't think it will give the 
best result, because it will downsample a 250 kHz input to 44.1 kHz for 
example, using only 1 input sample every 250/44.1=5.66 samples, thus not 
taking into account a big part of the input samples.

This might be OK when emulating YM2149 for sound having period>=6, but 
with very high freq sound, this gives bad results.

IIRC this was hearable in the game Lethal Xcess in STF mode where they 
play digi sound by setting period=0 on one voice and modulating volume 
for this voice. In that case YM2149_Next_Resample_Weighted_Average_N() 
gives much better result as it uses all samples from the 250kHz input 
signal.

Nicolas