|
|
DevelopmentSega Master System / Mark III / Game Gear |
Home - Forums - Games - Scans - Maps - Cheats - Credits |
YM 2413 Reverse Engineering Notes 2016 - 02 - 10
by andete. Original documents available at: https://github.com/andete/ym2413/tree/master/results
<< YM2413 Reverse Engineering Notes 2015-12-31 | YM2413 Reverse Engineering Notes | YM2413 Reverse Engineering Notes 2016-07-16 >>
It's been a while since my last post. Partly because I took a little break and partly because it took me quite a while to figure this stuff out. Even though, in retrospect, the underlying mechanism isn't too difficult at all.
Introduction
In this post I'll investigate the transitions between the attack-decay-sustain- release states in the ADSR-envelope in more detail. Take a look at this image:
This shows the attack- and decay-part of the envelope for a waveform with AR=6 and DR=3. Notice that in both the attack and the decay phases there are segments of two different widths, following the pattern 'wide narrow narrow, wide narrow narrow, ...'. Also notice that the segment between attack and decay, encircled in green (more specifically this is the first segment of the decay phase), still has a different width than either of the repeating widths in the attack or decay phase.
The same is true for the first segment of the release phase, or for the first segment after any transition: almost always this segment has a different width than the surrounding segments.
Before going into more detail, let's first repeat what we already know about the envelope generator from previous investigations:
From the (4-bit) AR,DR,RR settings, the channel frequency and the KSR setting, we calculate a 6-bit effective rate. This rate determines how quickly the envelope changes.
The upper 4 bits of this rate select a bit in the global counter (this counter increases every sample, so at ~49,7kHz). Only if the selected bit toggles (in other words, after a selectable power-of-2 number of samples), we execute the stuff in the next paragraph.
The lower 2 bits of the effective rate select 1 of 4 possible sequences:
0: 0,1,0,1,0,1,0,1 (4 out of 8)
1: 0,1,0,1,1,1,0,1 (5 out of 8)
2: 0,1,1,1,0,1,1,1 (6 out of 8)
3: 0,1,1,1,1,1,1,1 (7 out of 8)
From the global counter we take 3 bits, starting from the selected bit (previous paragraph) and going to the left. Or in other words: take the global counter, shift it down so that the selected bit is in position 0 and take the lower 3 bits of that number. The resulting value (between 0-7) selects a bit from the above sequence. Now only if that selected bit is '1' we advance to the next envelope level.
(For very high rates, the rules are more complicated, but I'll refer to the previous posts for those details. They aren't important for this post).
Measurements
If we repeat the experiment from the introduction multiple times, we see that the width of the first segments varies. This section is all about measuring that phenomena.
If we zoom in on the segment encircled in green in the above image, we get this:
As I annotated in this image, I measure the width of a segment from the first peak of the segment to the first peak of the next segment. These peaks are the oscillations of the basis waveform (a sine wave in this case). So choosing a higher frequency allows to more accurately measure the segment width.
In the hope of making the analysis simpler, I started measuring with the
simplest pattern: "0,1,0,1,0,1,0,1" (4-out-of-8). This gives each segment in
steady-state the same width. So all wide segments instead of narrow-narrow-wide
like in the above images.
Though this 4-out-of-8 pattern is only used for notes with low frequency (first two octaves, the BLOCK channel setting). So that limits the maximum frequency, and thus also the measuring accuracy. If I combine the maximum frequency (0x3ff) with the maximum multiplication factor (car.ML=15), the distance between two peaks is about 37 samples.
Only after I already made a large number of measurements, I realized I could improve the accuracy by also measuring the corresponding segment mirrored around the x-axis (roughly speaking: this segment is shifted half a period in time and has 'opposite' amplitude). This allows to derive an upper and lower bound for the actual segment width.
In the first set of measurements I looked at the transition between attack and decay for various settings of AR and DR. See the table below for the results:
- In the rows AR varies between 1 and 8. In the columns DR varies between 1 and 7. Though I did not exhaustively measure every combination.
- Per cell I sorted the measured segment widths from small to big. Some values were obtained multiple times. Sometimes I annotated such duplicates (e.g. I put '2x' in front), sometimes I forgot to annotate the duplicates (IOW I made more measurements than shown in the table, but all uniquely observed values should be present).
- For most segments I measured the upper and lower bound. Though for the earlier measurements I didn't do this yet. Those results are marked with '#'.
- The results marked with '?' are likely errors (possibly I made some mistakes while doing this repetitive work). They don't fit the pattern we'll discover in the next section. I also couldn't reproduce those results when I went back to them. (I could have cheated and left out those results ;-)
- When AR is equal to DR, the first DR segment was often not visible. I mean normally, at the end of the attack phase, there is a segment with Y=0 and Y=511 (maximum amplitude), but not for this combination. So it's as-if when the condition to transition from the attack to the decay phase triggers, this also immediately triggers the condition to go from one level in the decay phase to the next level. OTOH sometimes there is a single peak visible at amplitude Y=0 or Y=511. These measurements are only accurate to 37 samples. This means very narrow peaks can easily be missed, so for now I wrote '1' for the segment with (and annotated it with '*'). But this needs more investigation.
AR/DR|| 1 | 2 | 3 | 4 | 5 | 6 | 7
=====++================+==============+==============+==============+==============+============+============
1 || 1(*) | 0x800(#) | | | | |
-----++----------------+--------------+--------------+--------------+--------------+------------+------------
2 ||3x0x07f2-0x0814 | 1(*) | 0x3f1-0x413 | | | |
||3x0x17fb-0x181d | | | | | |
-----++----------------+--------------+--------------+--------------+--------------+------------+------------
|| 0x1bec-0x1c0e | 0x3f1-0x413 | 1(*) |3x0x200(#) | 0x100(#) | 0x80(#) | 0x40(#)
|| 0x0404(#) | 0xbf5-0xc17 | | | | |
|| 0x047c(#) ?| 0x3fc(#) | | | | |
3 || 0x0bfe(#) | 0x402(#) | | | | |
||2x 0x0c04(#) | 0x404(#) | | | | |
|| 0x13fe(#) |2x 0xbfc(#) | | | | |
|| 0x1480(#) ?| 0xc01(#) | | | | |
|| 0x1bfe(#) |2x 0xc03(#) | | | | |
-----++----------------+--------------+--------------+--------------+--------------+------------+------------
|| 0x11f8-0x121a | 0x1f0-0x212 |3x0x1f0-0x212 | 1(*) |2x0x100(#) |2x0x80(#) |2x0x40(#)
4 || 0x15fb-0x161d | 0x5f2-0x614 | 0x5f1-0x613 | | | |
|| 0x19fd-0x1a1f | 0xdf7-0xe19 | 0x200(#) | | | |
|| 0x0a05(#) | 0x200(#) | | | | |
-----++----------------+--------------+--------------+--------------+--------------+------------+------------
|| 0x04f2-0x0514 | 0x2f0-0x313 | 0x0f0-0x112 |2x0x0ef-0x111 | 1(*) |2x0x80(#) |2x0x40(#)
5 || 0x08f3-0x0915 | 0x4f2-0x514 | 0x2f1-0x313 |3x0x2f1-0x312 | | |
|| 0x1aeb-0x1b0d | 0x8f4-0x916 | 0x6f3-0x715 |5x 0x300(#) | | |
|| 0x0915(#) | 0x100(#) | 0x2fd(#) | | | |
|| 0x0cf5(#) | 0x2ff(#) |4x 0x4fc(#) | | | |
|| 0x0efe(#) | 0x4fc(#) |4x 0x700(#) | | | |
|| 0x14fc(#) | 0x902(#) | | | | |
|| 0x16fb(#) | 0x97a(#) ?| | | | |
|| 0x1902(#) | 0xeff(#) | | | | |
|| 0x191d(#) | | | | | |
|| 0x1cfe(#) | | | | | |
-----++----------------+--------------+--------------+--------------+--------------+------------+------------
||2x0x0167-0x0189 | 0x267-0x289 | 0x368-0x38a | 0x066-0x088 | 0x077-0x099 | 1(*) |2x0x40(#)
|| 0x116f-0x1191 | 0x569-0x58b | 0x378-0x39c | 0x078-0x09a | 0x178-0x19a | |
6 || 0x1380-0x13a1 | 0xb51-0xb73?| 0x479-0x49b |2x0x167-0x189 |3x 0x080(#) | |
|| 0x1671-0x1693 |2x0xd7e-0xda0 |2x0x668-0x68a | 0x178-0x19a |2x 0x17f(#) | |
|| 0x1e74-0x1e97 | 0x386(#) | 0x76a-0x78c | 0x379-0x39b | 0x199(#) | |
|| 0x0285(#) | 0x580(#) |2x 0x080(#) |2x 0x080(#) | | |
|| 0x0496(#) | 0xa83(#) | 0x182(#) | 0x281(#) | | |
|| 0x0c84(#) | 0xb84(#) | 0x284(#) |2x 0x284(#) | | |
|| 0x1485(#) | 0xe80(#) | 0x383(#) |2x 0x380(#) | | |
|| 0x1882(#) | 0xf80(#) | 0x484(#) | 0x385 | | |
|| 0x1982(#) | | 0x580(#) | | | |
|| | |2x 0x680(#) | | | |
|| | | 0x77e(#) | | | |
-----++----------------+--------------+--------------+--------------+--------------+------------+------------
|| 0x00bc-0x00de | 0x0ac-0x0ce | 0x0ab-0x0cd | 0x124-0x146 | 0x033-0x055 | 0x33-0x55 | 1(*)
|| 0x0124-0x0146 | 0x1ad-0x1cf | 0x537-0x559 | 0x135-0x157 | 0x0ac-0x0cd | 0xbc-0xde |
7 ||2x0x04bf-0x04e1 | 0x436-0x458 | 0x6af-0x6d0 | 0x224-0x246 | 0x135-0x157 | 0x44(#) |
|| 0x0536-0x0559 | 0xe2a-0xe4c | 0x7af-0x7d1 | 0x3bd-0x3df | 0x044(#) | 0x45(#) |
|| 0x0fb3-0x0fd5 | 0xfb2-0xfd4 | 0x156(#) | 0x044(#) | 0x045(#) | 0xac(#) |
|| 0x143e-0x1460 | | 0x1be(#) | 0x0ce(#) | 0x0cd(#) |2x 0xce(#) |
|| | | 0x334(#) | 0x156(#) | 0x135(#) | |
|| | |2x 0x357(#) | 0x1be(#) | 0x156(#) | |
|| | | 0x4ae(#) | 0x246(#) | 0x1be(#) | |
|| | | 0x648(#) | 0x2cf(#) | 0x1dd(#) | |
|| | | |2x 0x3be(#) | | |
-----++----------------+--------------+--------------+--------------+--------------+------------+------------
|| | | | 0x09a-0x0bc | 0x056-0x078 | 0x45-0x67 |4x0x11-0x33
|| | | | 0x0cd-0x0ef | 0x0cd-0x0ef |3x0x56-0x78 | 0x56-0x78
8 || | | | 0x0de-0x100 |2x0x155-0x177 |2x0x9a-0xbc |
|| | | |2x0x1de-0x200 | 0x189-0x1ab |2x0xde-0xff |
|| | | | 0x2df-0x301 | | |
|| | | | 0x3cf-0x3f1 | | |
Next I made similar measurements for the decay-to-release transition. But because taking these measurements takes a long time (30-60 seconds per data point) I limited myself to DR=4,5. That should be sufficient because the pattern for this 2nd set of measurements (DR/RR) seems to be the same as for the 1st set (AR/DR).
Actually the only difference between the two sets occurs when the row-value is equal to the column-value (AR==DR or DR==RR). Before we had a segment of 0 or 1 sample, but now there's no exception when both rates are equal. It's as-if there's no transition at all.
DR/RR|| 1 | 2 | 3 | 4 | 5 | 6 | 7
=====++================+==============+==============+==============+==============+============+============
|| 0x05e2-0x0604 |7x0x1ef-0x211 | 0x1df-0x201 |4x0x3f1-0x413 |6x0x0f1-0x112 |2x0x67-0x89 |2x0x22-0x44
4 || 0x09f5-0x0a17 |3x0x5f3-0x615 |2x0x1ef-0x211 | | |6x0x78-0x9a |6x0x33-0x55
|| 0x15e8-0x160a |3x0xde6-0xe08 |2x0x5f1-0x613 | | | |
||2x0x19eb-0x1a0d | | | | | |
||2x0x1dee-0x1e10 | | | | | |
-----++----------------+--------------+--------------+--------------+--------------+------------+------------
||2x0x0cf5-0x0d17 | 0x2f0-0x312 | 0x2f1-0x313 |3x0x0ee-0x111 |3x0x1ef-0x211 |5x0x67-0x89 |7x0x34-0x56
5 ||2x0x12e8-0x130a | 0x6f2-0x714 |2x0x4f1-0x514 |3x0x2ef-0x312 | |4x0x78-0x9a |
|| 0x18fb-0x191d | 0x8f3-0x915 |2x0x6f2-0x714 | | | |
|| | 0xef7-0xf19 | | | | |
Finding a pattern in the measurements
Next I tried to pick nice 'round' numbers that fall within the error margins of
all the measurements. E.g. for the combination AR=4/DR=2 such numbers are
'0x200, 0x600 and 0xe00'. That seems to match the formula '0x200 + n * 0x400'.
And when I made more measurements for this combination, I eventually also found
a segment with width 0xa00. This confirms that formula. So in the table below
for AR=4/DR=2 I wrote '0x200 0x600 0xa00 0xe00'.
Actually for all the combinations with DR < AR (or DR < RR) I found a similar pattern: all the observed segment widths seem to match this formula:
first-segment-length = (2n + 1) * L
with: L = 1 << (13 - AR)
Note that L is also the required number of steps of the global counter described in the introduction.
As already mentioned before, when AR == DR, the first segment was often not visible or only 1 sample wide. We still need to investigate this further. On the other hand DR == RR is nothing special.
When 'DR > AR' (or RR > DR), the pattern is simple:
first-segment-length = 1 << (13 - DR)
Note that this is half of all the later segments in the decay (or release) phase. Or in other words, the first segment is always a narrow segment while all later segments are wide (for these low frequencies (4-out-of-8 pattern) we only have wide segments in steady-state).
AR/DR|| 1 | 2 | 3 | 4 | 5 | 6 | 7
=====++===============+=============+=============+=============+=============+======+======
1 || 1(*) | 0x800 | | | | |
-----++---------------+-------------+-------------+-------------+-------------+------+------
2 || 0x0800 0x1800 | 1(*) | 0x400 | | | |
-----++---------------+-------------+-------------+-------------+-------------+------+------
3 || 0x0400 0x1400 | 0x400 | 1(*) | 0x200 | 0x100 | 0x80 | 0x40
|| 0x0c00 0x1c00 | 0xc00 | | | | |
-----++---------------+-------------+-------------+-------------+-------------+------+------
|| 0x0200 0x1200 | 0x200 | 0x200 | 1(*) | 0x100 | 0x80 | 0x40
4 || 0x0600 0x1600 | 0x600 | 0x600 | | | |
|| 0x0a00 0x1a00 | 0xa00 | | | | |
|| 0x0e00 0x1e00 | 0xe00 | | | | |
-----++---------------+-------------+-------------+-------------+-------------+------+------
|| 0x0100 ... | 0x100 0x900 | 0x100 | 0x100 | 1(*) | 0x80 | 0x40
5 || 0x0300 0x1b00 | 0x300 0xb00 | 0x300 | 0x300 | | |
|| 0x0500 0x1d00 | 0x500 0xd00 | 0x500 | | | |
|| ... 0x1f00 | 0x700 0xf00 | 0x700 | | | |
-----++---------------+-------------+-------------+-------------+-------------+------+------
|| 0x0080 ... | 0x080 ... | 0x080 0x480 | 0x080 | 0x080 | 1(*) | 0x40
6 || 0x0180 0x1d80 | 0x180 0xd80 | 0x180 0x580 | 0x180 | 0x180 | |
|| 0x0280 0x1e80 | 0x280 0xe80 | 0x280 0x680 | 0x280 | | |
|| ... 0x1f80 | ... 0xf80 | 0x380 0x780 | 0x380 | | |
-----++---------------+-------------+-------------+-------------+-------------+------+------
|| 0x0040 ... | 0x040 ... | 0x040 ... | 0x040 0x240 | 0x040 | 0x40 | 1(*)
7 || 0x00c0 0x1ec0 | 0x0c0 0xec0 | 0x0c0 0x6c0 | 0x0c0 0x280 | 0x0c0 | 0xc0 |
|| 0x0140 0x1f40 | 0x140 0xf40 | 0x140 0x740 | 0x140 0x340 | 0x140 | |
|| ... 0x1fc0 | ... 0xfc0 | ... 0x7c0 | 0x1c0 0x3c0 | 0x1c0 | |
-----++---------------+-------------+-------------+-------------+-------------+------+------
|| | | | 0x020 ... | 0x020 0x120 | 0x20 | 0x20
8 || | | | 0x060 0x360 | 0x060 0x160 | 0x60 | 0x60
|| | | | 0x0a0 0x3a0 | 0x0a0 0x1a0 | 0xa0 |
|| | | | ... 0x3e0 | 0x0e0 0x1e0 | 0xe0 |
DR/RR|| 1 | 2 | 3 | 4 | 5 | 6 | 7
=====++===============+=============+=============+=============+=============+======+======
|| 0x0200 0x1200 | 0x200 | 0x200 | 0x400 | 0x100 | 0x80 | 0x40
4 || 0x0600 0x1600 | 0x600 | 0x600 | | | |
|| 0x0a00 0x1a00 | 0xa00 | | | | |
|| 0x0e00 0x1e00 | 0xe00 | | | | |
-----++---------------+-------------+-------------+-------------+-------------+------+------
|| 0x0100 ... | 0x100 0x900 | 0x100 | 0x100 | 0x200 | 0x80 | 0x40
5 || 0x0300 0x1b00 | 0x300 0xb00 | 0x300 | 0x300 | | |
|| 0x0500 0x1d00 | 0x500 0xd00 | 0x500 | | | |
|| ... 0x1f00 | 0x700 0xf00 | 0x700 | | | |
A model to explain the measurements
Let's again take AR=4/DR=2 as an example.
- AR=4 means that every 0x200 steps there's an 'event' in the attack curve and for this specific channel frequency 1 out of 2 times (4 out 8 in the 0,1,0,1,0,1,0,1 pattern) we actually advance to the next AR level.
- DR=2 means an event every 0x800 samples and with the 4-out-of-8 stuff, it means we advance a DR level every 0x1000 samples.
Note that, in general, when the attack phase starts, the value of the global
counter is random. In this example, when the attack phase ends we've just
processed an AR event with a '1' in the 4-out-8 table and that occurs only
every 0x400 samples. On the other hand, in this example, DR only effectively
advances every 0x1000 samples. Not 0x10000 starting from the end of attack,
but when the global counter reaches the next multiple of 0x10000 (this is not
100% correct, see next paragraphs). The lower bits of the global counter could
already have reached 0xc00, so the next multiple of 0x1000 is only 0x400 samples
away, But it could also be 0x800, 0xc00 or 0x1000 samples till the next multiple
of 0x10000. This already explains why, in this particular example, there are 4
possible initial segment lengths. It does not yet explain why these lengths are
actually 0x200, 0x600, 0xa00, 0xe00 rather than 0x400, 0x800, 0xc00, 0x1000.
That is '(2n+1) * L' rather than '2n * L'.
To explain the '2n+1' stuff we need to take a more detailed look at the 0,1,0,1,0,1,0,1 stuff. I was writing a c++ model for this phenomena and then I realized that there are two possible (evenly-spaced) 4-out-of-8 sequences:
- 0,1,0,1,0,1,0,1 -> gives 0x200 0x600 0xa00 0xe00 - 1,0,1,0,1,0,1,0 -> gives 0x400 0x800 0xc00 0x1000
One of these gives 0x200,0x600,0xa00,0xe00, the other gives 0x400,0x800,0xc00, 0x10000. We observed the first segment width, so the first sequence must be the correct one. Looking back this behavior is very logical, but it took me some time to realize this.
For reference I've included the c++ code for the model, but I won't discuss it further:
#include <random>
#include <set>
using namespace std;
unsigned counter;
unsigned level; // 0-127
unsigned rate; // 0-63
char incr[4][8] = {
/* 0 */ { 0,1,0,1,0,1,0,1 }, // this matches the measurements
// { 1,0,1,0,1,0,1,0 }, // this does not
// { 1,1,0,1,0,1,0,1 }, //
// { 0,1,1,1,0,1,0,1 }, //
{ 0,1,0,1,1,1,0,1 }, // matches measurements
// { 0,1,0,1,0,1,1,1 }, //
// { 1,1,1,0,1,0,1,0 }, //
// { 1,0,1,1,1,0,1,0 }, //
// { 1,0,1,0,1,1,1,0 }, //
// { 1,0,1,0,1,0,1,1 }, //
// { 1,1,0,1,1,1,0,1 }, //
{ 0,1,1,1,0,1,1,1 }, // matches measurements
// { 1,1,1,0,1,1,1,0 }, //
// { 1,0,1,1,1,0,1,1 }, //
{ 0,1,1,1,1,1,1,1 }, // matches measurements
// { 1,1,0,1,1,1,1,1 }, //
// { 1,1,1,1,0,1,1,1 }, //
// { 1,1,1,1,1,1,0,1 }, //
// { 1,0,1,1,1,1,1,1 }, //
// { 1,1,1,0,1,1,1,1 }, //
// { 1,1,1,1,1,0,1,1 }, //
// { 1,1,1,1,1,1,1,0 }, //
};
inline std::minstd_rand0& global_urng()
{
static std::minstd_rand0 u;
return u;
}
inline uint32_t random_32bit()
{
static std::uniform_int_distribution<uint32_t> d;
using parm_t = decltype(d)::param_type;
return d(global_urng(), parm_t{0, 0xffffffff});
}
void step()
{
unsigned shift = 13 - (rate / 4);
unsigned mask = (1 << shift) - 1;
++counter;
if ((counter & mask) != 0) return;
unsigned sub = (counter >> shift) & 7;
if (incr[rate & 3][sub] == 0) return;
++level;
}
unsigned countSteps()
{
unsigned l = level;
unsigned s = 0;
while (l == level) {
++s;
step();
}
return s;
}
int main()
{
multiset<unsigned> p;
for (int i = 0; i < 5000; ++i) {
counter = random_32bit();
level = 0;
rate = 22;
countSteps();
for (int j = 0; j < 8; ++j) {
cout << hex << " 0x" << countSteps();
}
cout << " | ";
rate = 10;
unsigned s = countSteps();
p.insert(s);
cout << hex << "0x" << s;
cout << " | ";
for (int j = 0; j < 3; ++j) {
cout << hex << " 0x" << countSteps();
}
cout << endl;
}
set<unsigned> p2(begin(p), end(p));
for (unsigned s : p2) {
cout << hex << " 0x" << s << '(' << p.count(s) << ')';
}
cout << endl;
}
Extend to higher frequencies
So far all measurements/models were done for frequency 0x3ff, those use the 4-out-of-8 '0,1,0,1,0,1,0,1' pattern. Higher frequencies use 5,6,7-out-of-8 patterns.
Above I did a lot of (time consuming) measurements for various combinations of AR/DR/RR. But after I understood the underlying model I realized it's probably enough to limit myself to a single AR/DR combination. I choose AR=5/DR=2.
5-out-of-8
I measured lengths:
0x100 0x300 0x400 0x500 0x700 0x900 0xb00 0xc00 0xd00 0xf00
Side note: I relatively quickly obtained 1,3,4,5,7,9,c,d,f. But it literally
took over 50 attempts before I finally also observed length 0xb00, which I
suspected should be there. (The c++ model confirms that some lengths are also
more likely to be observed than others, and 0xb00 is a less likely length),
Possible 5-out-of-8 patterns are:
1,1,0,1,0,1,0,1
0,1,1,1,0,1,0,1
0,1,0,1,1,1,0,1 <-- only this one matches the measurements
0,1,0,1,0,1,1,1
1,1,1,0,1,0,1,0
1,0,1,1,1,0,1,0
1,0,1,0,1,1,1,0
1,0,1,0,1,0,1,1
6-out-of-8
I measured these lengths:
0x100 0x200 0x300 0x500 0x600 0x700 0xa00 0xd00 0xe00
The model predicts 0x900 0xb00 0xf00 should also be present (with lower
probability), but I didn't keep on measuring until I actually observed those
values. The current measurements were already enough to exclude the other
possible patterns (see below).
Possible 6-out-of-8 patterns are:
1,1,0,1,1,1,0,1
0,1,1,1,0,1,1,1 <-- this one matches the measurements
1,1,1,0,1,1,1,0
1,0,1,1,1,0,1,1
7-out-of-8
I measured these lengths:
0x100 0x200 0x300 0x400 0x500 0x600 0x700 0xb00 0xc00
The values 0x{9,a,d,e,f}00 should also be present (only 0x800 and 0x1000 are
missing). But these measurements can already exclude the other patterns below.
Possible 7-out-of-8 patterns are:
0,1,1,1,1,1,1,1 <-- this one matches the measurements
1,1,0,1,1,1,1,1
1,1,1,1,0,1,1,1
1,1,1,1,1,1,0,1
1,0,1,1,1,1,1,1
1,1,1,0,1,1,1,1
1,1,1,1,1,0,1,1
1,1,1,1,1,1,1,0
Conclusion
We found that the model we had for the envelope steps already could explain the initial segment lengths after a AR->DR or DR->RR transition. And although I didn't measure this, I'm confident it will also correctly model the initial segment after a RR->DP or DP->AR transition.
We also found the exact 4,5,6,7-out-of-8 sequences, namely
4-out-of-8: 0,1,0,1,0,1,0,1 5-out-of-8: 0,1,0,1,1,1,0,1 6-out-of-8: 0,1,1,1,0,1,1,1 7-out-of-8: 0,1,1,1,1,1,1,1
Side note: seq-5 OR'ed with seq-6 gives seq-7 OR'ing seq-4 with any other sequence simply give that other sequence This is not relevant for a software implementation, but it might give a clue to how the actual hardware circuit is structured (the 4 patterns are somewhat like a 2-bit binary counter).
And if we look at the various emulator cores written by Jarek Burczynski (e.g. YM2413 but also YM2151 and YMF262), they contain exactly these sequences. So kudos to Jarek!
So actually, if I hadn't done any of the investigations in this post. The emulation would (probably) have been exactly the same. But still, it's nice to better understand why this weird initial segment length stuff is happening. And even better it's nice to know that the sequences used by Jarek Burczynski are really the same sequences as used by the YM2413.
<< YM2413 Reverse Engineering Notes 2015-12-31 | YM2413 Reverse Engineering Notes | YM2413 Reverse Engineering Notes 2016-07-16 >>
