ayVGM - A lite version of VGM for AY-3-8910

Ascended (10772)

21-01-2022, 12:55

One more small optimisation:

case 0x00:
	u8 reg = *g_ayVGM_Pointer & 0x0F;

The AND is not needed here because the upper nibble is already 0.

aoineko wrote:

I get an average reduction of 63% (especially thanks to the waveforms data packing), but the resulting data is still too big to be really usable in a game (in ROM anyway).

Can you quantify this? Also, when you say too big, are you considering a 32K ROM or a MegaROM?

Por aoineko

Paladin (1005)

21-01-2022, 15:11

Here are some numbers.

My personal guideline for a music to be usable in a game is to stay under the 8KB limit.
Here, we are often even over 16KB, which makes them difficult to use, even with a MegaROM.

Grauw wrote:

The AND is not needed here because the upper nibble is already 0.

Fixed! Thanks.

Por Grauw

Ascended (10772)

21-01-2022, 16:34

I’m not sure if you want to pursue the SCC any further, but here are some ideas in case you do;

What if you designed the SCC commands exactly like you did the PSG? Omit the PSG. For example:

---------------------------------------
 Frequency
---------------------------------------
	0n nn		0x80 = nn 0x81 = n	Frequency channel 1
	1n nn		0x82 = nn 0x83 = n	Frequency channel 2
	2n nn		0x84 = nn 0x85 = n	Frequency channel 3
	3n nn		0x86 = nn 0x87 = n	Frequency channel 4
	4n nn		0x88 = nn 0x89 = n	Frequency channel 5
---------------------------------------
 Volume
---------------------------------------
	5n		0x8A = n		Volume channel 1
	6n		0x8B = n		Volume channel 2
	7n		0x8C = n		Volume channel 3
	8n		0x8D = n		Volume channel 4
	9n		0x8E = n		Volume channel 5
---------------------------------------
 Mixer
---------------------------------------
	An		0x8F = n		On/off switch channel 1 to 5
	Bn	 	0x8F = n | 0x10		On/off switch channel 1 to 5
---------------------------------------
 Waveform
---------------------------------------
	C0 nn[32]	nn[32] -> 0x00[32]	Waveform channel 1
	C1 nn[32]	nn[32] -> 0x20[32]	Waveform channel 2
	C2 nn[32]	nn[32] -> 0x40[32]	Waveform channel 3
	C3 nn[32]	nn[32] -> 0x60[32]	Waveform channel 4
	C4 nn[32]	nn[32] -> 0xA0[32]	Waveform channel 5 (SCC+ only)
---------------------------------------
 Wait
---------------------------------------
	Dn		Wait 882 * (n & 0x0F + 1) samples (50 Hz)
	En		Wait 735 * (n & 0x0F + 1) samples (60 Hz)
---------------------------------------
 Special
---------------------------------------
	FE nn nn    	End of song with loop
	FF          	End of song (without loop)

You can have a separate PSG and SCC player, and play two separate streams simultaneously. The benefit is that the encoding for SCC is more efficient. The cost is that the timing gets encoded (more or less) twice.

Por Grauw

Ascended (10772)

21-01-2022, 16:36

I think volume & frequency are the most important ones to optimise. For PSG, maybe mixer is also relevant. For SCC, mixer is probably only set once so not worth dedicating short commands to it? But SCC waveform is important.

As for frequency, you can also consider to set frequency MSB and LSB always together. E.g. 0nnn, 2nnn, 4nnn. Then you only need two bytes to set frequency instead of the usual 3. (Changed it in the example above.)

As for the waveform, probably it gets changed a lot to previously seen waveforms. So it would be worth to move the unique waveforms into a separate table, and change the command to specify the index in the waveform table.

Por aoineko

Paladin (1005)

21-01-2022, 16:35

I thought about it.
Another idea I had in mind was to keep a single stream but to mark sections for the AY and for the SCC because in the data I've analyzed, the two are often grouped together in package.
Anyway, it seems like a lot of effort for data that will remain big.

For now, I prefer to look at the TriloTracker format...

Por Grauw

Ascended (10772)

21-01-2022, 16:38

Fair enough, it will never go as low as 8k. I’m interested to hear about your future findings .

Por niek

Expert (112)

21-01-2022, 18:56

No matter the result, cool project.

Por GhostwriterP

Paladin (683)

21-01-2022, 20:27

aoineko wrote:

For now, I prefer to look at the TriloTracker format...

You know you will be probably trading speed for song size here. Also I imagine the TT replayer is probably bigger than a VGM based one. Of course neither needs to be a problem.

Por GhostwriterP

Paladin (683)

22-01-2022, 17:51

Going sub 8k for SCC & PSG - 8 channel encoded VGM would I think be challenging. Though, looking at some numbers from Realfun 3 as reference below, it might be possible when keeping the songs "simple" enough (i.e. without many waveform updates). By encoding repetitions and indexing the waveforms, those big SCC songs from metalgear2 and manbow size will come down a lot, maybe not all sub 8k but quite positive it can go below 16 k.

Name	 Size	  data   waves	 #
-----------------------------------
hydroc	 15842    11298	  4544	142
miners	 17780	  12532	  5248	164
corridor 15105	   7297	  7808	244
staff	 15525	  11717	  3808	119
metalsqd 27298	  21442	  5856	183
battle	 18005	  10421	  7584	237
vector	  8915	   7059	  1856   58

Por Grauw

Ascended (10772)

22-01-2022, 19:16

GhostwriterP wrote:

You know you will be probably trading speed for song size here.

For me the speed is the primary benefit of a register-log format. Another pro is the simplicity and freedom to choose any music production tool. In a MegaROM environment I’m less concerned about file sizes, I don’t mind if it goes above 8K. However smaller is still better, so smart encodings are pretty interesting, and seeing what others are doing and what results they’re getting, too.