Friday, March 26, 2010

SID waveform captures

Well well well, I knew that the 6581 and the 8580/6582 generated different combined waveforms, but I didn't know that not only each single chip generates slightly different bits from each other, but you also wont get twice the exact same waveform on two separate playbacks on a single chip.

This is a binary diff of two OSC3 sampling runs of a combined waveform on a 6581 CBM (r3) chip:

Thanks for kevtris again for the tip (but no thanks in a way, since I had to spend lots of evenings to make the code to generate these data files on a real c64 lol.)

I took care in recording the normal triangular and saw waveforms in each session for comparisons, and they match across all chips.

You can basically think of the SID chip as a 4096 (4kb) sized wavetable synthesizer with each entry being 12bits in precision, only instead of actually indexing a table (which would have been too long to do with the tight schedule given to Yannes when working at MOS), each index in the table is given to a function that generates output "samples" ; A simple counter in the case of the SAW, a comparator for the Pulse+PW, etc.  Only later did Yannes/Ensoniq actually implement this as a real wavetable in the DOC chip used in the ESQ-1.

The combined waveforms are still a matter of study as to how they are generated, (see the work of Antti Lankila) . After recording a huge bunch of very different ones however I cant help but feel that
there is no "perfect" way to go at this. As each SID will generate something different, why not add some non-deterministic aspects into the generation?

In the mean time we can reasonably emulate the combined waveforms of the SID (which are really a odd mixture of bits in the analog world) by indexing a pre recorded table such as the one I've captured using the SID's 3rd oscillator "read" functionality. As you know the C64 is an 8 bit machine so we only can read a approximation of the real result (8 most significant bits out of the 'real' 12) but it doesn't really matter, since even at 8bits, we can prove that no two reads are the same, so who cares really if we lose 4bits of precision. Those data files for those combined waveforms will be included in my new emu code and you will be able to choose the version of the chip you want. That way you could simulate a wide range of different "runs".

Note1: the waveforms are $11,$21,(...)$81,
           frequency=1, CIA timer=$FFF

Note2: I don't know what is "wrong" with that r2's P_T waveform. seems like its phase starts halfway compared to R3,R4 and 6582... I'm waiting for other R2's from Ebay so I'll retry when I get them.

Note3: The 6582's noise captures are all in phase, but not with my 6581 recordings... weird

more notes to come...

Monday, March 15, 2010

SID 6581R3 ADSR tables, up close.

In the center of any SID emulation there is the bare waveform generation, and a very accurate explanation of how the this works internally has been published through an interview with its creator, which can be read here. Fascinating read for any geek head!

However some very important details are missing, including the exact maths behind the main ADSR clocking but also its pseudo exponential decay/release stages.

It was time to have a look at the SID's DIE itself!
This is exaclty what kevtris and Lord Nightmare have done here

The SID chip has two obvious ROM based lookup tables on the DIE,  for each voice. Here is the bigger one, as described in kevtris's blog entry:

I basically just cropped parts of the picture of the chip, which is available here, and placed the bit values on top of it.

However the blog post doesnt mention, nor explain the purpose of the second table, which I assumed was one of the "exp" tables, as mentionned in the Yannes Interview:

Yannes: "In order to more closely model the exponential decay of sounds, another look-up table on the output of the Envelope Generator would sequentially divide the clock to the Envelope Generator by two at specific counts in the Decay and Release cycles. This created a piece-wise linear approximation of an exponential. I was particularly happy how well this worked considering the simplicity of the circuitry.".

The SID patent's figure 10 mentions the following dividers are used: 30,16,4,2,1 . Hum not quite divide by two heh? 32,16,8,4,2,1 would have been more logical!

So lets try to decode the exact table on the DIE...

Well I didnt have a big clue myself, as reading bits of a DIE is all new to me, but after discussing with Lord Nightmare and kevtris, it appears this is also a LFSR counting trick, this time with a 5bit long LFSR and taps on bits 2 and 4:

Here is that table in plain ascii:
SSSSS  A B  A B  A B  A B  A B
00100  0 1  0 1  1 0  0 1  0 1    
00001  0 1  0 1  0 1  0 1  1 0    
01000  0 1  1 0  1 0  1 0  0 1    
00010  1 0  1 0  1 0  0 1  1 0    
10000  1 0  1 0  0 1  0 1  0 1    

Kevtris explained that the left hand part is a "Selector" of sorts, since there is only one bit active on each line/column. The second part seems to contain inverted pairs of bits..... hmm puzzling...

Using similar code to what they provided, this time for the second table, and only taking the 'B' bits as LSB:

const unsigned short exp_lfsr[5] = {
    0x1B, 0x0F, 0x11 , 0x08 ,0x1C

for (size_t i=0;i<5;i++){   
  unsigned int LFSR=0x1F;   
  size_t c=0;   
    if (LFSR == exp_lfsr[i]){            
       exptable[i]= c;       
        LFSR = ((LFSR << 1) | (((LFSR >> 2) 
          ^ (LFSR >> 4)) & 1)) & 0x1F;

the results in exptable is 8,30,4,16,2, (which mostly matches the numbers in SID patent, except for the missing 8... and the weird order.

Ok so we know that at some points in the decay of the envelope, the clock divider changes... what does that mean exaclty and what are those "points"????

Heres what we can try:

We know the SID chip has two readable registers which are tied to the 3rd voice: One for the 8bit ENV value ($D41C) and the other for the 8 most significant bits of the waveform generator (which is 12 bits internally) at $D41B

So by setting the 3rd SID Oscillator's release at longest rate ($F), and by hooking up a CIA Timer to callback at each $7a13 cycles (which comes from the phase2 counter lenght for release 'F' according to kevtris/LN), we can, in theory, get a synchronized sampling of the envelope and store the results for analysis. As im lasy and that seeing is believing, I just programmed to display the values on the screen while the note decays.

Here is a live capture of that code running on a BreadBox c64 stuffed with a 6581R4AR:

If you know your Screen Codes, you can see that the envelope goes from 255 to 0, and that some values start to repeat at certain points...

The chars to look for are:
"|": (93)  switches to 2  waits before a drop
"6": (54)  switches to 4  waits before a drop 
"Z": (26)  switches to 8  waits before a drop
"N": (14)  switches to 16 waits before a drop (Thanks Frank!)
"F": (06)  switches to 30 waits before a drop

So while I can now go on with my emu code, knowing what happens, I'm still clueless on WHERE/HOW this comparison happens on the DIE. So if you have a clue, or spot anything wrong in my logic, please add a comment!

Sunday, March 14, 2010

MESS 0.137 is out!


I'm proud to announce my little personal contribution to the best emulator project in the world.

Since a bit after chipsounds 1.0 was released, I started contributing some of my recent research to the open source 'MESS' project on the sound front. My contributions are "without strings attached" as I feel that the research in MAME/MESS is crucial to the good preservation of the history of computing. Besides, the accumulated knowledge in there will surely outlive me :)


New System Drivers Supported:
- Casio PV-1000 [Wilbert Pol, plgDavid]

System Driver Changes:
- [SCV] Implemented upd177c audio. [plgDavid]"

The SCV audio still needs work, so that's not the last effort I will put into it. I've also tweaked the Arcadia 2001 audio code and made it much closer to the real thing. I also plan on revisiting a few other "drivers", when I get the chance, namely the VIC-20.

The MAME/MESS Teams members are very passionate and knowledgeable. I want to take the moment to greet Wilbert Pol, kevtris and Lord Nightmare especially, and to thank them for their time and near infinite knowledge.

Get it NOW