YM 2413 Reverse Engineering Notes 2017 - 01 - 26

by andete. Original documents available at: https://github.com/andete/ym2413/tree/master/results

<< YM2413 Reverse Engineering Notes 2016-07-16 | YM2413 Reverse Engineering Notes | YM2413 Reverse Engineering Notes 2018-05-13 >>

It's (again) been a long time since I last wrote about the YM2413 reverse engineering. The last several months I was trying to figure out the details of the envelope attack curve. Though mostly without success, so there wasn't much to report in the first place.

Last month I've been helping Eugeny Brychkov to add MSX-MUSIC support in GR8NET. Possibly because of the discussions with Eugeny I looked at things from a different perspective and got an idea for a new experiment that led to the breakthrough I've been hoping for. More on this in the second half of this post.

In the first part I'll continue where the last post left off. I'll present a C++ model of the YM2413 that includes all the stuff we've found out so far.

1) C++ model

At the end of the previous post (20160716) we developed a C++ model that can generate output samples for a single YM2413 channel with the following features:

• modulator connected to carrier
• feedback from modulator to itself
• channel volume handling
• different waveforms
• correct (bit-exact) log-sine and exponential tables
• mute for small signals

Though that's far from all the YM2413 features we've investigated so far. In particular the following features are still missing:

• channel frequency setting
• key on/off handling
• vibrato and amplitude modulation
• changing envelope over time

So let's fix that.

I created the new model in a few intermediate steps. I'll quickly highlight the changes in each step. But I won't go over the code in detail because there's no new information in it and I hope the code itself is self-explanatory (at least with the information from these posts in mind).

• step 1

This starts from the previous model (previous post). It doesn't add anything new, it only restructures the code somewhat: data is grouped into structures and free standing functions are turned into member functions. This change is also a preparation to support multiple sound channels in the future.

• step 2

This step adds a global counter to the YM2413 model and uses it to implement (per operator) amplitude modulation.

• step 3

This step adds phase calculations, these are:

• phase step size (controlled by channel frequency)
• (per operator) frequency modulation
• step 4

The main change in this step is the addition of the envelope calculations. This envelope stuff is the least-well understood part of the YM2413 (so far), in particular the attack part of the envelope. But hopefully that also means that the implementation can still be simplified in the future. (See the 2nd part of this post for more info on the attack curve).

These 4 source files (model{1,2,3,4}.cc) share 1 header file 'tables.hh'. As the name suggests, that file contains some tables that are used by the model(s). Though for most of these tables I've discovered a simple hardware implementation. So the real hardware does not store that information in a ROM but instead calculates it with logic functions. OTOH an efficient software implementation might still want to use some of these tables.

Disclaimer: To the best of my knowledge model4.cc can reproduce all the YM2413 features we've investigated, with exception of the envelope attack curves. BUT I must also admit that, so far, I've done far too little verification on this model (it's on my TODO list). So there may still be bugs.

2) Envelope attack curve

Now the more interesting part of this post ...

2.1) Past experiments

Let's first quickly recap how we investigated the attack curve in the past. Check this picture from an older post:

In that experiment we combine a not too fast attack-rate with a high-enough frequency. We see that that there are multiple plateaus in the attack curve. From the maximum amplitude of the signal in each plateau we can derive the corresponding envelope level. Remember envelope level is a value between 0 and 127 where 0 maps to 0dB and 127 maps to -47.625dB (0.375dB per step). So a higher envelope level corresponds to a lower amplitude.

From this experiment we derived that during attack the envelope level seems to follow one of these two paths:

   {112-127}, {91- 95}, {71-72}, 53, 39, 28, 20, 13,  9, 5, 1, 0
   {112-127}, {96-102}, {73-74}, 55, 41, 29, 21, 14, 10, 6, 2, 0

Note that for high envelope levels there are multiple levels that have the same maximum amplitude. So by measuring the amplitude we cannot always uniquely identify the envelope-level. That's why the first 3 items in these paths have a range of possible levels.

First problem:

I've tried very hard, but I've been unable to come up with a simple formula that can explain this sequence ..., 53, 39, 28, 20, 13, ... All other things we've discovered about the YM2413 could be explained by simple hardware circuits, but this sequence is a mystery.

Second problem:

This measurement setup only allows to measure relatively slow attack rates. From measurements on the decay curve we know that the very fast decay rates behave in a special way, so likely there's something similar going on for the attack rates. For example the YM2413 data sheet lists a duration of 140us for the fastest attack rate, that's only 7 samples long (@49.7kHz). But the envelope path shown above has 12 distinct values, so it cannot be used for the fastest attack rates.

The underlying problem of this experiment is that the phase of the sine wave is reset to zero at the start of the attack phase (and sin(0) = 0). And fast attack rates don't last long enough for the sine to reach max amplitude (phase = 90 degrees). If only there was a way to 'postpone' the attack curve until the phase has reached 90 degrees ...

2.2) New experiments

... Yes, if only we could 'postpone' the attack curve ...

Let's try setting the attack rate to zero. Then wait some time for the phase to reach 90 degrees. And only then switch to the desired attack rate. More in detail I ran experiments with these settings:

This results in a measurement like this:

On the left the graph shows a small part of a sine, this belongs to the previous experiment. Around x=1600 we set key=off and the signal drops to zero (signal=0 corresponds to y=255 in this graph). After some time we set key=on, but the signal remains zero because attack=0. Around x=6700 we set attack to a non-zero value and the amplitude rapidly rises to maximum. It's important to choose the delay between key=on and attack=non-zero precisely so that the phase reaches exactly 90 degrees (sine has maximum amplitude).

In the graph above the attack curve was a vertical line. But if we zoom-in the attack curve looks as follows:

As in all my measurements graphs y=255 maps to +0 and y=0 maps to +255, so this graph actually shows a signal that's rising in amplitude.

Notice how the curve stays flat once the maximum amplitude is reached. That's because in this example I had chosen a very low frequency, so each entry from the YM2413's sine table is repeated 512 times. And also the top of the sine itself is 'flat'. So we really have several thousands of samples where the output remains at maximum amplitude (and this is needed to measure the slower attack rates).

Also notice that the curve has 12 plateaus, these match with the envelope levels we already obtained from the old experiment:

{112-127}, {96-102}, {73-74}, 55, 41, 29, 21, 14, 10, 6, 2, 0

BUT if we zoom in closer we see something unexpected:

For clarity I marked all samples with 'x' on this graph. We again see the plateaus. But instead of jumping from one plateau to the next, the signal takes 3 intermediate steps!

This last zoomed graph showed the curve for attack rate 10:0 (effective rate = 4*10+0 = 40). From earlier investigations we expect to see plateaus that are 16 samples long for this rate. Instead we see lengths of 13 plus 3 extra samples. This is true for all attack rates less than 12:0: we always see plateaus of length 'pow(2, n) - 3' plus 3 intermediate steps.

Let's analyze this new attack curve in more detail. We'll first look at the 'slow' attack rates (less than 12:0). We'll also separately look at the timing of the steps and the envelope levels that each step visits.

2.3) Envelope levels

So let's initially ignore the timing of the attack curve. In this subsection we will only look at how the amplitude of the signal evolves. As before we can translate the measured amplitude into envelope levels, and that gives the following sequence (remember for the first few values we can't map the amplitude to a unique envelope level, so there's an uncertainty interval):

   {112-127} {103-111} {103-111} {96-102}
    {87-90}   {84-86}   {78-79}  {73-74}
      68        63        59       55
      51        47        44       41
      38        35        32       29
      27        25        23       21
      19        17        15       14
      13        12        11       10
       9         8         7        6
       5         4         3        2
       1         0

Now guess what happens if you only look at ever 4th value in this sequence (I conveniently formatted the sequence in 4 columns): you get the sequence we observed in our old experiment!

For the old sequence I was unable to come with a simple formula, but for this new 4x longer sequence it's fairly easy. The latter values in the series (for which there is no uncertainty) match exactly with:

f4(x) := x - (x >> 4) - 1

For reasons that will become clear later, we'll call this "formula-4" or just 'f4'. So take the current envelope value, divide that by 16 (and round down). Subtract this from the current value and then subtract one more.

We can also apply this formula in reverse to try to determine a more precise value for the initial steps that have an uncertainty interval. It turns out that the following sequence is the only_one_ that both matches 'f4' and all uncertainty intervals.

   127 119 111 104 97 90 84 78 73 68 63 59 55 51 47 44
    41  38  35  32 29 27 25 23 21 19 17 15 14 13 12 11
    10   9   8   7  6  5  4  3  2  1  0

In other words, if we start at envelope level 127 and iteratively apply 'f4' we get our measured sequence.

But what if we start from a different level? Let's say from level 124. That results in:

   124 116 108 101 94 88 82 76 71 66 61 57 53 49 45 42
    39  36  33  30 28 26 24 22 20 18 16 14 13 12 11 10
     9   8   7   6  5  4  3  2  1

And if we pick every 4th element from this sequence we get the other common sequence we obtained from the old experiment.

2.4) Envelope timing (slow rates)

In the previous subsection we figured out which envelope levels the attack curve visits in sequence. The only thing left is figuring out the timing. In other words: when to jump from one level to the next.

To do this we first need more measurements of different attack rates. Visually all these attack curves look similar to the graphs I've shown above, so I won't show any more graphs. Instead I'll present (some of) the measurements in table form:

   attack: 7:0                           attack: 10:0
     sample   |amplitude|env-level         sample  |amplitude|env-level
   -----------+---------+---------       ----------+---------+---------
       0      |     1   | 112-127           0-12   |     1   | 112-127
       1      |     2   | 103-111          13      |     2   | 103-111
       2      |     2   | 103-111          14      |     2   | 103-111
       3-127  |     3   |  96-102          15      |     3   |  96-102
     128      |     5   |  87-90           16-28   |     5   |  87-90
     129      |     6   |  84-86           29      |     6   |  84-86
     130      |     8   |  78-79           30      |     8   |  78-79
     131-255  |    10   |  73-74           31      |    10   |  73-74
     255      |    13   |  68              32-44   |    13   |  68
     257      |    16   |  63              45      |    16   |  63
     258      |    19   |  59              46      |    19   |  59
     259-383  |    23   |  55              47      |    23   |  55
     384      |    28   |  51              48-60   |    28   |  51
     385      |    33   |  47              61      |    33   |  47
     386      |    37   |  44              62      |    37   |  44
     387-511  |    43   |  41              63      |    43   |  41
     512      |    49   |  38              64-76   |    49   |  38
     513      |    56   |  35              77      |    56   |  35
     514      |    63   |  32              78      |    63   |  32
     515-639  |    72   |  29              79      |    72   |  29
     640      |    79   |  27              80-92   |    79   |  27
     641      |    86   |  25              93      |    86   |  25
     642      |    94   |  23              94      |    94   |  23
     643-767  |   102   |  21              95      |   102   |  21
     768      |   112   |  19              96-108  |   112   |  19
     769      |   122   |  17             109      |   122   |  17
     770      |   133   |  15             110      |   133   |  15
     771-895  |   139   |  14             111      |   139   |  14
     896      |   145   |  13             112-124  |   145   |  13
     897      |   151   |  12             125      |   151   |  12
     898      |   158   |  11             126      |   158   |  11
     899-1023 |   165   |  10             127      |   165   |  10
    1024      |   172   |   9             128-140  |   172   |   9
    1025      |   180   |   8             141      |   180   |   8
    1026      |   188   |   7             142      |   188   |   7
    1027-1151 |   196   |   6             143      |   196   |   6
    1152      |   205   |   5             144-156  |   205   |   5
    1153      |   214   |   4             157      |   214   |   4
    1154      |   224   |   3             158      |   224   |   3
    1155-1279 |   234   |   2             159      |   234   |   2
    1280      |   244   |   1             160-172  |   244   |   1
    1281-...  |   255   |   0             173-...  |   255   |   0

These two tables show the attack curves for respectively attack rate 7:0 and 10:0 (effective rates 7*4+0 = 28 and 10*4+0 = 40). Each table has 3 columns.

• The 1st column is the sample number (the time axis in the graphs). I aligned each measurement so that the first non-zero amplitude corresponds to sample=0. If there's a range in this column it means that multiple samples have the same amplitude.
• The 2nd column is the measured amplitude of the signal for this (range of) sample(s). Remember that in this experiment we eliminated the influence of the sine wave on the amplitude, so the measured amplitude directly corresponds to the envelope level.
• The 3rd column is calculated from the 2nd column. It's the envelope level (or the interval of possible levels) that correspond to the measured amplitude.

The first observation is that the envelope levels for both attack rates are the same (and the same as the levels for any attack rate below 12:0). This sequence was already explained in the previous subsection.

A second observation is that we indeed see repeated groups of 3 single values followed by a plateau. We already saw that (visually) in section 2.2.

Note however that the plateaus are not always located on the same envelope level. To double check this I made a series of measurements of the same attack rate. So the following 3 tables are from running the exact same experiment multiple times.

   attack: 11:0  (multiple measurements of the same rate)
    sample |amplitude|env-level      sample |amplitude|env-level      sample |amplitude|env-level
   --------+---------+---------     --------+---------+---------     --------+---------+---------
     0     |     1   | 112-127        0-4   |     1   | 112-127        0     |   1     | 112-127
     1     |     2   | 103-111        5     |     2   | 103-111        1     |   2     | 103-111
     2     |     2   | 103-111        6     |     2   | 103-111        2-7   |   2     | 103-111
     3-7   |     3   |  96-102        7     |     3   |  96-102        7     |   3     |  96-102
     8     |     5   |  87-90         8-12  |     5   |  87-90         8     |   5     |  87-90
     9     |     6   |  84-86        13     |     6   |  84-86         9     |   6     |  84-86
    10     |     8   |  78-79        14     |     8   |  78-79        10-14  |   8     |  78-79
    11-15  |    10   |  73-74        15     |    10   |  73-74        15     |   10    |  73-74
    16     |    13   |  68           16-20  |    13   |  68           16     |   13    |  68
    17     |    16   |  63           21     |    16   |  63           17     |   16    |  63
    18     |    19   |  59           22     |    19   |  59           18-22  |   19    |  59
    19-23  |    23   |  55           23     |    23   |  55           23     |   23    |  55
    24     |    28   |  51           24-28  |    28   |  51           24     |   28    |  51
    25     |    33   |  47           29     |    33   |  47           25     |   33    |  47
    26     |    37   |  44           30     |    37   |  44           26-30  |   37    |  44
    27-31  |    43   |  41           31     |    43   |  41           31     |   43    |  41
    32     |    49   |  38           32-36  |    49   |  38           32     |   49    |  38
    33     |    56   |  35           37     |    56   |  35           33     |   56    |  35
    34     |    63   |  32           38     |    63   |  32           34-38  |   63    |  32
    35-39  |    72   |  29           39     |    72   |  29           39     |   72    |  29
    40     |    79   |  27           40-44  |    79   |  27           40     |   79    |  27
    41     |    86   |  25           45     |    86   |  25           41     |   86    |  25
    42     |    94   |  23           46     |    94   |  23           42-46  |   94    |  23
    43-47  |   102   |  21           47     |   102   |  21           47     |   102   |  21
    48     |   112   |  19           48-52  |   112   |  19           48     |   112   |  19
    49     |   122   |  17           53     |   122   |  17           49     |   122   |  17
    50     |   133   |  15           54     |   133   |  15           50-54  |   133   |  15
    51-55  |   139   |  14           55     |   139   |  14           55     |   139   |  14
    56     |   145   |  13           56-60  |   145   |  13           56     |   145   |  13
    57     |   151   |  12           61     |   151   |  12           57     |   151   |  12
    58     |   158   |  11           62     |   158   |  11           58-62  |   158   |  11
    59-63  |   165   |  10           63     |   165   |  10           63     |   165   |  10
    64     |   172   |   9           64-68  |   172   |   9           64     |   172   |   9
    65     |   180   |   8           69     |   180   |   8           65     |   180   |   8
    66     |   188   |   7           70     |   188   |   7           66-70  |   188   |   7
    67-71  |   196   |   6           71     |   196   |   6           71     |   196   |   6
    72     |   205   |   5           72-76  |   205   |   5           72     |   205   |   5
    73     |   214   |   4           77     |   214   |   4           73     |   214   |   4
    74     |   224   |   3           78     |   224   |   3           74-78  |   224   |   3
    75-79  |   234   |   2           79     |   234   |   2           79     |   234   |   2
    80     |   244   |   1           80-84  |   244   |   1           80     |   244   |   1
    81-... |   255   |   0           85-... |   255   |   0           81-... |   255   |   0

Again we see the same sequence of envelope levels. We also see 3 single steps followed by a plateau (only a short plateau of 5 samples for this relatively high attack rate).

And indeed the position of the plateaus is different in the 3 tables. In the 1st table the plateau starts on sample 3 while in the 2nd and 3rd table it's respectively at sample 0 and 2. I actually had to make many measurements before I got these 3 different results. Most of the time I got the 1st table, a few times I got table 2 and only on my 10th attempt I got table 3. I'm confident that if I continued making more measurements I'd also see a version with the plateau on sample 1, but I didn't have the patience to do that ;-)

Notice how the total length of the attack curve is slightly different in these 3 versions. Even though it's for the exact same attack rate. But we had something similar for the decay curve. There it could be explained by a global counter (see earlier posts for more details). Here we'll use a similar explanation.

2.5) Envelope timing algorithm

Let's first very briefly recap the algorithm for decay. For more details check the post about EG-decay.

Decay:

a)

From the effective rate calculate two values:

eg_shift = 13 - (rate / 4) eg_select = rate & 3

b)

There's a global (shared) counter that's incremented by one every sample.

c)

Take the value of that counter and shift it down over 'eg_shift' number of bits. Only if all the bits that were shifted out are zero we go to the next step. IOW if any bit was a 1 the envelope level does not change.

d)

Next take this table

       0,1,0,1,0,1,0,1
       0,1,0,1,1,1,0,1
       0,1,1,1,0,1,1,1
       0,1,1,1,1,1,1,1

Select the row indicated by 'eg_select' and the column indicated by the 3 LSB bits of the downshifted global counter value. If that position in the table is a '1' then increment the envelope level by one, if it was '0' then do nothing.

With only small tweaks we can make this algorithm work for attack

Attack:

a)

Calculate 'eg_shift' and 'eg_select' in the same way.

b)

Use the same global counter.

c)

Shift the global counter down over 'eg_shift' number of bits. Check whether all shifted-out bits are zero, but IGNORE the 2 LSB bits. That is the first two shifted out bits can have any value, but all following bits must be zero.

d)

Use the same table and column/row selection mechanism as above. But now when there is a '1' in the table, change the current envelope level via formula 'f4' (if there's a '0' do nothing).

All the measurements I've shown so far had 'eg_select=0'. So they all use the first row in the 8x4-bit-pattern-table and thus they all have a very regular timing pattern. Let's also check the other patterns. Below are measurements for rates 11:0 to 11:3 (effective rates 44 to 47).

   attack : 11:0         attack : 11:1         attack : 11:2         attack : 11:3
  sample |env-level     sample |env-level     sample |env-level     sample |env-level
 --------+---------    --------+---------    --------+---------    --------+---------
   0     | 112-127       0     | 112-127       0     | 112-127       0     | 112-127
   1     | 103-111       1     | 103-111       1     | 103-111       1     | 103-111
   2     | 103-111       2     | 103-111       2     | 103-111       2     | 103-111
   3-7   |  96-102       3-7   |  96-102       3     |  96-102       3     |  96-102
   8     |  87-90        8     |  87-90        4     |  87-90        4     |  87-90
   9     |  84-86        9     |  84-86        5     |  84-86        5     |  84-86
  10     |  78-79       10     |  78-79        6     |  78-79        6     |  78-79
  11-15  |  73-74       11     |  73-74        7     |  73-74        7     |  73-74
  16     |  68          12     |  68           8     |  68           8     |  68
  17     |  63          13     |  63           9-13  |  63           9     |  63
  18     |  59          14     |  59          14     |  59          10     |  59
  19-23  |  55          15     |  55          15     |  55          11     |  55
  24     |  51          16     |  51          16     |  51          12     |  51
  25     |  47          17     |  47          17     |  47          13     |  47
  26     |  44          18     |  44          18     |  44          14     |  44
  27-31  |  41          19-23  |  41          19     |  41          15     |  41
  32     |  38          24     |  38          20     |  38          17     |  38
  33     |  35          25     |  35          21     |  35          18     |  35
  34     |  32          26     |  32          22     |  32          19     |  32
  35-39  |  29          27-31  |  29          23     |  29          20     |  29
  40     |  27          32     |  27          24     |  27          21-24  |  27
  41     |  25          33     |  25          25-29  |  25          25     |  25
  42     |  23          34     |  23          30     |  23          26     |  23
  43-47  |  21          35-39  |  21          31     |  21          27     |  21
  48     |  19          40     |  19          32     |  19          28     |  19
  49     |  17          41     |  17          33     |  17          29     |  17
  50     |  15          42     |  15          34     |  15          30     |  15
  51-55  |  14          43     |  14          35     |  14          31     |  14
  56     |  13          44     |  13          36     |  13          32     |  13
  57     |  12          45     |  12          37     |  12          33     |  12
  58     |  11          46     |  11          38     |  11          34     |  11
  59-63  |  10          47     |  10          39     |  10          35     |  10
  64     |   9          48     |   9          40     |   9          36     |   9
  65     |   8          49     |   8          41-45  |   8          37     |   8
  66     |   7          50     |   7          46     |   7          38     |   7
  67-71  |   6          51-55  |   6          47     |   6          39     |   6
  72     |   5          56     |   5          48     |   5          40     |   5
  73     |   4          57     |   4          49     |   4          41     |   4
  74     |   3          58     |   3          50     |   3          42     |   3
  75-79  |   2          59-63  |   2          51     |   2          43     |   2
  80     |   1          64     |   1          52     |   1          44     |   1
  80-... |   0          65-... |   0          53-... |   0          45-... |   0

If you decipher this table you'll find that it has a similar pattern of wide vs narrow steps that we saw for the decay curve (in earlier posts). And that the attack-timing algorithm shown above can indeed fully explain it.

The different positions of the plateaus we talked about in the previous subsection can be explained by this algorithm: it's because the global counter has a different value when attack starts. It can also explain why the first variant of the 11:0 measurement (in section 2.4) occurred more frequently than the other ones.

2.6) Attack envelope for faster rates

With the information in the sections above we can perfectly explain attack rates 1:0 to 11:3. However that explanation breaks down for faster rates.

But let's first measure how the hardware actually behaves for rates 12:0 to 14:3:

 sample|| 12:0  | 12:1  | 12:2  | 12:3  || 13:0  |13:1 |13:2 |13:3 ||14:0 |14:1 |14:2 |14:3
 ------++-------+-------+-------+-------++-------+-----+-----+-----++-----+-----+-----+-----
    0  ||112-127|103-111|112-127|103-111||103-111|91-95|91-95|91-95||91-95|91-95|91-95|63
    1  ||103-111| 96-102|103-111| 96-102|| 96-102|80-83|80-83|71-72||71-72|71-72|71-72|31
    2  ||103-111| 87-90 |103-111| 84-86 || 84-86 |71-72|71-72|53   ||53   |53   |35   |15
    3  || 96-102| 84-86 | 87-90 | 73-74 || 73-74 |62   |62   |39   ||39   |39   |17   |7
    4  || 87-90 | 78-79 | 78-79 | 63    || 63    |54   |54   |29   ||29   |29   | 8   |3
    5  || 84-86 | 73-74 | 68    | 55    || 55    |47   |40   |21   ||21   |21   | 3   |1
    6  || 78-79 | 68    | 59    | 48    || 48    |41   |29   |15   ||15   |15   | 2   |0
    7  || 73-74 | 63    | 55    | 44    || 41    |35   |21   |13   ||11   |11   | 1
    8  || 68    | 59    | 51    | 41    || 35    |30   |15   |11   || 8   | 8   | 0
    9  || 63    | 55    | 47    | 38    || 30    |26   |13   | 9   || 5   | 5
   10  || 59    | 51    | 44    | 35    || 26    |22   |11   | 7   || 3   | 3
   11  || 55    | 47    | 38    | 30    || 22    |19   | 9   | 5   || 2   | 1
   12  || 51    | 44    | 33    | 26    || 19    |16   | 7   | 3   || 1   | 0
   13  || 47    | 41    | 28    | 22    || 16    |11   | 5   | 2   || 0
   14  || 44    | 35    | 24    | 19    || 13    | 8   | 3   | 1
   15  || 41    | 30    | 22    | 16    || 11    | 5   | 2   | 0
   16  || 38    | 26    | 20    | 13    ||  9    | 3   | 1
   17  || 35    | 22    | 18    | 11    ||  7    | 2   | 0
   18  || 32    | 20    | 16    |  9    ||  6    | 1
   19  || 29    | 18    | 13    |  7    ||  5    | 0
   20  || 27    | 16    | 11    |  6    ||  4
   21  || 25    | 14    |  9    |  5    ||  3
   22  || 23    | 13    |  7    |  4    ||  2
   23  || 21    | 12    |  6    |  3    ||  1
   24  || 19    | 11    |  5    |  2    ||  0
   25  || 17    | 10    |  4    |  1
   26  || 15    |  9    |  3    |  0
   27  || 14    |  8    |  2
   28  || 13    |  7    |  1
   29  || 12    |  6    |  0
   30  || 11    |  5
   31  || 10    |  4
   32  ||  9    |  3
   33  ||  8    |  2
   34  ||  7    |  1
   35  ||  6    |  0
   36  ||  5
   37  ||  4
   38  ||  3
   39  ||  2
   40  ||  1
   41  ||  0

Note: Measuring the same rate multiple times, results in slight variations on these results. (This is the same as for the slower rates).

Observations:

• The envelope level changes every sample. IOW there are no plateaus anymore.
• In column 12:0 we see the familiar sequence generated by formula 'f4'.
• But the other columns cannot be explained by the same formula.

But luckily it isn't too difficult to figure out a new explanation. First we need 3 extra formulas:

   f4(x) := x - (x >> 4) - 1
   f3(x) := x - (x >> 3) - 1
   f2(x) := x - (x >> 2) - 1
   f1(x) := x - (x >> 1) - 1

(And this explains why we gave our original formula the name 'f4').

I'll first give the solution and explain it below. The following table has the same structure as the one above. But now the uncertainty intervals are resolved and each value is annotated with a formula 'f1-f4'.

 sample|| 12:0 | 12:1 | 12:2 | 12:3 || 13:0 | 13:1 | 13:2 | 13:3 || 14:0 | 14:1 | 14:2 | 14:3
 ------++------+------+------+------++------+------+------+------++------+------+------+------
   -1  ||127,f4|127,f3|127,f4|127,f3||127,f3|127,f2|127,f2|127,f2||127,f2|127,f2|127,f2|127,f1
    0  ||119,f4|111,f3|119,f4|111,f3||111,f3| 95,f3| 95,f3| 95,f2|| 95,f2| 95,f2| 95,f2| 63,f1
    1  ||111,f4| 97,f4|111,f4| 97,f3|| 97,f3| 83,f3| 83,f3| 71,f2|| 71,f2| 71,f2| 71,f1| 31,f1
    2  ||104,f4| 90,f4|104,f3| 84,f3|| 84,f3| 73,f3| 72,f3| 53,f2|| 53,f2| 53,f2| 35,f1| 15,f1
    3  || 97,f4| 84,f4| 90,f3| 73,f3|| 73,f3| 62,f3| 62,f3| 39,f2|| 39,f2| 39,f2| 17,f1|  7,f1
    4  || 90,f4| 78,f4| 78,f3| 63,f3|| 63,f3| 54,f3| 54,f2| 29,f2|| 29,f2| 29,f2|  8,f1|  3,f1
    5  || 84,f4| 73,f4| 68,f3| 55,f3|| 55,f3| 47,f3| 40,f2| 21,f2|| 21,f2| 21,f2|  3,f2|  1,f1
    6  || 78,f4| 68,f4| 59,f4| 48,f4|| 48,f3| 41,f3| 29,f2| 15,f3|| 15,f2| 15,f2|  2,f2|  0
    7  || 73,f4| 63,f4| 55,f4| 44,f4|| 41,f3| 35,f3| 21,f2| 13,f3|| 11,f2| 11,f2|  1,f2
    8  || 68,f4| 59,f4| 51,f4| 41,f4|| 35,f3| 30,f3| 15,f3| 11,f3||  8,f2|  8,f2|  0
    9  || 63,f4| 55,f4| 47,f4| 38,f4|| 30,f3| 26,f3| 13,f3|  9,f3||  5,f2|  5,f2
   10  || 59,f4| 51,f4| 44,f3| 35,f3|| 26,f3| 22,f3| 11,f3|  7,f2||  3,f2|  3,f1
   11  || 55,f4| 47,f4| 38,f3| 30,f3|| 22,f3| 19,f3|  9,f3|  5,f2||  2,f2|  1,f
   12  || 51,f4| 44,f4| 33,f3| 26,f3|| 19,f3| 16,f2|  7,f2|  3,f2||  1,f2|  0
   13  || 47,f4| 41,f3| 28,f3| 22,f3|| 16,f3| 11,f2|  5,f2|  2,f2||  0
   14  || 44,f4| 35,f3| 24,f4| 19,f3|| 13,f3|  8,f2|  3,f2|  1,f2
   15  || 41,f4| 30,f3| 22,f4| 16,f3|| 11,f3|  5,f2|  2,f2|  0
   16  || 38,f4| 26,f3| 20,f4| 13,f3||  9,f3|  3,f3|  1,f3
   17  || 35,f4| 22,f4| 18,f4| 11,f3||  7,f3|  2,f3|  0
   18  || 32,f4| 20,f4| 16,f3|  9,f3||  6,f3|  1,f3
   19  || 29,f4| 18,f4| 13,f3|  7,f3||  5,f3|  0
   20  || 27,f4| 16,f4| 11,f3|  6,f3||  4,f3
   21  || 25,f4| 14,f4|  9,f3|  5,f3||  3,f3
   22  || 23,f4| 13,f4|  7,f4|  4,f4||  2,f3
   23  || 21,f4| 12,f4|  6,f4|  3,f4||  1,f3
   24  || 19,f4| 11,f4|  5,f4|  2,f4||  0
   25  || 17,f4| 10,f4|  4,f4|  1,f4
   26  || 15,f4|  9,f4|  3,f3|  0
   27  || 14,f4|  8,f4|  2,f3
   28  || 13,f4|  7,f4|  1,f3
   29  || 12,f4|  6,f3|  0
   30  || 11,f4|  5,f3
   31  || 10,f4|  4,f3
   32  ||  9,f4|  3,f3
   33  ||  8,f4|  2,f4
   34  ||  7,f4|  1,f4
   35  ||  6,f4|  0
   36  ||  5,f4
   37  ||  4,f4
   38  ||  3,f4
   39  ||  2,f4
   40  ||  1,f4
   41  ||  0

I added one extra row to the table (sample=-1) because I assumed that the initial envelope level starts at 127. We also got this start value for the slower rates, so I think it's a reasonable assumption. Next, for each cell, I annotated the formula which transforms the envelope level of this cell to the level in the cell below it. So for example take sample=5 and rate=13:3, the envelope level in that cell is 21. If we apply formula 'f2' on it we obtain 15, and that's indeed the value in the cell below. To go one cell further down we need formula 'f3' to get to 13. Using this technique also allowed to resolve all uncertainty intervals.

Notice that:

• Columns 12:0, 13:0 and 14:0 only apply a single formula.
• All other columns use 2 alternating formulas:
  • columns 12:1 - 12:3: f3 and f4
  • columns 13:1 - 13:3: f2 and f3
  • columns 14:1 - 14:3: f1 and f2
• The pattern for the alternating formulas is:
  • columns <n>:1 4x formula f<m> followed by 12x formula f<m+1>
  • columns <n>:2 4x formula f<m> followed by 4x formula f<m+1>
  • columns <n>:3 12x formula f<m> followed by 4x formula f<m+1>

This pattern of 4-12, 4-4, 12-4 is the same as how the fast decay rates behave. Let's try to put this in a model.

2.7) Model the fast attack rates

Attack rates 12:0 and faster change envelope level every sample. So the sub-problem we had to solve for the slow rates, deciding when to change the level, is trivial here.

The remaining problem is to decide which formula to apply. I believe the following model can fully reproduce the observed behavior. It's possible there are simpler models that can also explain this behavior, but I tried to stay close to the model we already built for the slow attack rates.

• Set m = 16 - (rate / 4)

c = (global_counter >> 1) & 6 eg_select = rate & 3

• From the table

       0,1,0,1,0,1,0,1      (odd columns are not strictly needed, but now it's
       0,1,0,1,1,1,0,1       the same table as above)
       0,1,1,1,0,1,1,1
       0,1,1,1,1,1,1,1

Select the element at row 'eg_select' and column 'c'. If that element is '1', subtract 1 from 'm'.

• Apply formula f<m> to the current envelope level.

2.8) Attack rate 15

So far we've discussed the 'slow' (1:0 - 11:3) and the 'fast' (12:0 - 14:3) attack rates. But what about the 'fastest' attack rates 15:0 - 15:3. According to the YM2413 data sheet these rates are instantaneous. That is the envelope goes from min to max in a single step.

However when I tried to measure this I in my experiment I saw something strange. Quick recap, in the experiment:

• I select a relatively low frequency, with initial attack-rate=0
• After some time, when phase has reached 90 degrees, I switch to the to-be-measured attack rate.

This works fine for rates 1-14, but for rate=15 to signal remains at amplitude zero. Obviously when using attack=15 in the 'normal' way it works fine. Just not when switching to it during an in-progress attack.

My hypothesis is that attack=15 is only checked at the end of the damp phase (so right before the attack phase starts). If attack rate is indeed equal to 15, we set the envelope level to zero and immediately go to the decay phase (so skip attack phase). But during the attack phase itself rate=15 does not work, (and normally this also cannot occur).

To verify this hypothesis I ran two small experiments:

• toggle attack 15:

During the attack phase I rapidly switch between different attack rates. I verified that indeed whenever attack rate 0 or 15 is active the attack curve is paused. Switching to another rate continues the attack. Switching back to 0 or 15 pauses it again.

• attack=15 at start is indeed instantaneous

To verify this I selected a frequency of (3579545/72)/4 = 12329Hz. This makes each sine period exactly 4 samples long. So after 1 sample we have a phase of 90 degrees, thus maximal amplitude. I verified that the very first sample after attack indeed already has maximum amplitude. And this means that the envelope has already reached level 0 on the first sample.

2.9) Start value 124 vs 127

One thing that's still bothering me is that in these set of experiments we always observed the sequence:

  [127 119 111 104 97 90 84 78 73 68 63 59 55 51 47 44
    41  38  35  32 29 27 25 23 21 19 17 15 14 13 12 11
    10   9   8   7  6  5  4  3  2  1  0]

While in earlier measurements the sequence was:

  [124 116 108 101 94 88 82 76 71 66 61 57 53 49 45 42
    39  36  33  30 28 26 24 22 20 18 16 14 13 12 11 10
     9   8   7   6  5  4  3  2  1]

Actually in the old experiments we only observed the plateaus which respectively correspond to these levels:

  [127, 97, 73, 55, 41, 29, 21, 14, 10, 6, 2, 0]
  [124, 94, 71, 53, 39, 28, 20, 13,  9, 5, 1, 0]

So why is the start value sometimes 124 and sometimes 127? After some more investigation I figured it out:

In the experiments in this post I used the settings:

combined with:

• set key=off -> this triggers the release phase
• wait a while -> release has time to 'finish'
• set key=on -> actual start of the experiment

This results in start value 127.

But if I change the experiment to:

(I (only) changed car.RR from 15 to 00)

• set key=off -> triggers release phase, but does nothing (because car.RR=00)
• wait a while -> envelope remains unchanged
• set key=on -> this triggers the damp phase

This results in start value 124.

So apparently the release phase changes the envelope level all the way to 127, while the damp phase only goes up to 124.

To test this hypothesis I ran some variations of this last experiment where I varied car.RR and the time between key-off and key-on:

• Each time when release can 'finish' (thus fast release rate and/or enough time between key-off/on) attack starts from level 127.
• Each time there is a damp phase (release didn't finish: slow release rate and/or not enough time between key-off/on) attack starts from 124.

The fact that release does go down to 127 does falsify an hypothesis I made in an earlier post. There I assumed that envelope only goes down to 124 because that allows for a cheaper hardware implementation: no need for clipping logic, just a 5 bit AND-port. It's still the case that envelope levels 124 and higher all result in a constant-zero output signal. Although to be really sure I should still test whether levels 125 and 126 can ever be observed as start values (e.g. by precisely timing the start of the attack). But that's something for a later post.

2.10) Afterthought

I always found the YM2413 attack curve very rough. It goes from min to max amplitude in only 11 steps. In this post we discovered there are actually 4x as many steps, but these extra steps are "well hidden".

I can't think of any good reason for this behavior. With only small changes in the model the steps could be spread out more evenly, resulting in a much smoother attack curve. So I wonder if this is the result of an implementation bug in the YM2413?

Anyway, it is what it is, and to accurately emulate the YM2413 we have to replicate this rough attack curve.

3) Re-measure fast decay rates

In an earlier post we already looked at the decay rates in detail. I remember the fast decay rates were difficult to measure back then. In this post I found a better way to measure fast rates. So let's use it to re-measure the decay rates.

Another motivation to re-measure is that the results for fast decay rates were a bit strange. And I want to double check that the YM2413 is really behaving like that. OTOH the results we got above for fast attack are also strange and strange in much the same way. So in a way that's already some confirmation of the old results for decay.

Here are the new results:

  11-0:  8 8 8 8
  11-1:  8 8 8 4 4
  11-2:  8 4 4 8 4 4
  11-3:  8 4 4 4 4 4 4

  12-0:  4 4 4 4
  12-1:  4 4 4 2 2
  12-2:  4 2 2 4 2 2
  12-3:  4 2 2 2 2 2 2

  13-0:  2 2 2 2 2 2 2 2
  13-1:  2 2 2 2 2 2 1 1 1 1
  13-2:  2 2 1 1 1 1 2 2 1 1 1 1
  13-3:  2 2 1 1 1 1 1 1 1 1 1 1 1 1

  14-0:  1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1
  14-1:  0.5 0.5 0.5 0.5 1 1 1 1 1 1 1 1 1 1 1 1
  14-2:  0.5 0.5 0.5 0.5 1 1 1 1 0.5 0.5 0.5 0.5 1 1 1 1
  14-3:  0.5 0.5 0.5 0.5 0.5 0.5 0.5 0.5 0.5 0.5 0.5 0.5 1 1 1 1

  15-0:  0.5
  15-1:  0.5
  15-2:  0.5
  15-3:  0.5

The meaning of the values after the rate are the number of samples the envelope level remains constant before it is increased by one. Or if the value is 0.5 it means the level is increased by two.

This is exactly the same result as before.

4) Updated C++ model

Finally I updated the C++ model with these recent findings. Now also the attack curve should be correct.

Now the model doesn't need a table anymore for the attack curve. And because this was the last big table that was used from the table.hh header, I moved the few (small) remaining tables from the header to the c file and then eliminated that header.

Same disclaimer as above: I believe this model is correct, but it needs more verification to be sure.

In the next weeks/months I hope to work on reverse-engineering the content of the YM2413 instrument ROM. Maybe that's also a good opportunity to verify the C++ model.

After that the only(?) remaining step is investigating the rhythm sounds.

<< YM2413 Reverse Engineering Notes 2016-07-16 | YM2413 Reverse Engineering Notes | YM2413 Reverse Engineering Notes 2018-05-13 >>

• back to YM2413 main page

• Back to Development index