Challenges found writing an Apple II chiptune player
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
by DEATER (Vince Weaver, vince@deater.net)
http://www.deater.net/weave/vmwprod/chiptune/
====================================================
11 March 2018
GOAL:
~~~~~
The goal is to design a chiptune player that can play large
(150k+ uncompressed) chiptune files on an Apple II with 48k of RAM
and a Mockingboard sound card.
You in theory could have had an Apple II with 48k in 1977 (if you were rich)
and Mockingboards came around 1981, so this all predates the Commodore 64.
USING:
~~~~~~
Boot disk on a real system, or emulator with Mockingboard support.
Applewin works fine (even under Wine on Linux).
MESS does too, it's harder to setup (ROMs) but the audio sounds clearer.
Key bindings:
Spacebar -- pauses
Left/Right arrow -- switches songs
R -- enables/disables rasterbars
You can load up your own YM5 files. Get the "ym5_to_krw" utility found in
the repository https://github.com/deater/vmw-meter/
Copy the files to the disk image, and edit the filenames in chiptune.s
(sorry, don't have code that CATALOGs automatically. TODO?)
HARDWARE:
~~~~~~~~~
Sound
=====
The Mockingboard card has two AY-3-8910 chips, each interfaced with a
VIA 6522 I/O chip. The 6522 more or less acts as a GPIO expander, plus
provides programmable timer interrupts (something the Apple II lacks).
The AY-3-8910 chip provides three channels of square waves, plus noise.
There is also a (global) envelope generator (though it's typically
not used that much). The Mockingboard has two AY-3-8910s,
so you can have up to six channels of sound (3 on right, 3 on left).
Processor
=========
The Apple II has a 6502 processor running at 1.023 MHz.
RAM
===
You could get Apple IIs with as little as 4k of RAM. Eventually models
with 48k, 64k and 128k were popular, but due to I/O and ROM constraints to
access more than 48k you had to do bank switching.
DISK
====
The typical 5 1/4" floppy was single sided and by the time of DOS3.3 held
140k of data. Roughly 16k was used by DOS though if you wanted a bootable
disk. There are all kinds of ways you can cheat and extend this, as well
as using a "real" O/S like ProDOS. However growing up all I ever really
used was DOS3.3 so I'm using it for the sake of tradition.
Also if you want to run DOS3.3 then RAM from $9600 up through $C000 is
used by the O/S. For this project I use stock DOS3.3 so we lose that
amount of RAM (almost 11k).
SOUND DATA:
~~~~~~~~~~~
The AY-3-8910 chips are very flexible and can be programmed in a wide
variety of ways.
I'm attempting to play YM files, which are chiptune files popular in
the Atari and Spectrum communities. These are RAW register dumps;
every 50Hz (they tend to be European) the contents of the 14 AY-3-8910
registers are written. A raw data stream is 700 bytes (50*14) a second,
so 42k per minute. This means holding a raw, uncompressed, data stream
in RAM becomes a challenge.
COMPRESSION:
~~~~~~~~~~~~
The register values tend to be repetitive so they compress well. Especially
if you interleave the files (have all of the register 0 data in a row,
followed by all the data for register 1, etc. This is a lot harder to play
but you can get compression ratios of over 10 times, see the chart
at the end of this document).
In addition, the file data can be compressed even more if you notice unused
bits in the data. For example, the register data has many unused bits (the
period data is only 12 bits for each channel). Also many songs do not use
the envelope feature at all, freeing up 3 bytes. So custom compression that
can make assumptions about the sound format can free up many bytes even in
a raw register dump format.
A typical ym5 file is compressed with LHA compression which isn't practical
for compression.
The LZ4 algorithm is nearly as good and has existing 6502 implementations
which can be adapted. It isn't really a streaming algorithm though, so
it is hard to decompress only a chunk of the file at a time, usually you
need to decompress the whole file at once (the format works by referencing
bit sequences from earlier decompressed data).
This is especially troublesome with interleaved files, as although they
compress really well, you end up decompressing all of the register-0
data before you get to register-1 so with limited RAM you have to
change how you deal with things.
KRW File Format
~~~~~~~~~~~~~~~
I ended up creating yet another sound file format, and wrote a converter
that can convert YM5 files to this KRW format.
The format assumes you take the raw interleaved data, and then break it
up into 768 byte * 14 register (10.5k) chunks. These chunks are compressed
independently and concatenated together. The player then decompresses
these chunks one by one as it pays through the song. The compression
ratio is not as good as compressing the entire file, but it allows most
reasonable-length ym5 files to be played.
The format is as follows:
3 bytes Magic Number KRW
1 byte Skip Value Bytes to skip to get to first LZ4 data
1 byte Title Center Spaces to print to center on 40col
X bytes Title String 0-terminated ASCII Title of song
1 byte Author Center
X bytes Author String
1 byte Time Center
14 bytes Time String " 0:00 / M:SS\0" with length filled in
Repeated block data
2 bytes Chunk Length Little Endian size of LZ4 block
X bytes LZ4 data
After last block, a value of 0/0 indicates end
For proper end-of-song detection, the file data should be interleaved
and the data at the end should be padded with all $FF characters.
End of song is detected by an FF in register[1] which in theory
is not possible in a valid register dump.
PLAYING THE SONG
~~~~~~~~~~~~~~~~~
An interrupt routine wakes at 50Hz to write the registers and a few other
housekeeping things.
We load the KRW file totally into RAM before playing.
The Disk II controller designed by Woz is amazing, but it is timing
sensitive so interrupts are disabled when loading from disk.
We have to have room in RAM for the player (4k) the KRW file (16k)
and the current uncompressed data (14k). See the memory map diagram
at the end.
We also have some visualization going on that plots the amplitude of
the three channels, plus has a rasterbar type thing going on in the
background. Originally the graphics was done full speed in a loop outside
the interrupt handler, but as we'll see due to glitchy audio we had to
do some hackish things.
The actual player is fairly simple, just reads the interleaved data by
striding through memory and writing out to the registers. A frame only
takes maybe 2400 or so cycles.
I ended up creating a 3-phase state machine to handle co-ordinating the
three modes
A: playing chunk 1 while copying chunk 3 data to extra buffer
B: just playing chunk 2
C: playing from extra buffer while decoding next LZ4 block to 1-2-3
I track these in one variable, with the states in the high bits,
$80, $40, $20. The BIT instruction lets us easily check for these
and a ROL instruction easily switches between the states.
CHALLENGES:
~~~~~~~~~~~
The primary problem is decompression also takes a while, longer than
the 50Hz available (20ms). It turns out the default LZ4 algorithm from
qkumba can often take upwards of 700ms, leading to a long pause in
the playback.
First Attempt
=============
My first attempt to work around this was to load the 3 chunks of data
as in the naive approach, but in the background copy chunk 3 in RAM,
and then play from the copied RAM while decompressing the next LZ4 in
the background.
This first attempt almost worked, but it tried to split up the LZ4
decompression into 1/256th chunks to spread across the last chunk being
played but the LZ4 is too irregular for that. Some file-chunks decompress
in irregular ways that don't split up well.
Second Attempt
==============
One 256-interrupt chunk of data being played takes about 5s and no data
chunk seems to take more than 1s to decode. So we can just cheat and
move the graphics code into the interrupt, and have the decoding happen
in non-interrupt space.
This will work for the chiptune player, but it's not going to work well for
something like a video game where you are truly trying to have the music
playing unattended in the background (unless your music consists only of
15s loops).
FITTING ONTO DISK
~~~~~~~~~~~~~~~~~
Apple II DOS33 filesystem uses 256 byte blocks. Each file has at least
one 256 byte Track/Sector list file (and takes an additional one for each
28k or so of filesize).
DOS itself reserves the first 3 tracks (12k) and in theory the catalog
reserves an entire track (4k) to hold file info (although you only need
on 256 byte sector per 7 files).
In addition usually you have a "HELLO" BASIC file that runs at boot
which is going to take at least 512 bytes.
So even though the Disk II / DOS3.3 can in theory hold 140k, after
DOS (12k), the Catalog track (4k), HELLO(512 bytes), and our chiptune player
(4k) we have 24.5k of overhead, with 115.5k free (462 blocks).
The layout of our disk packed to the max with KRW files can be seen
in the Figure at the end. We do manage to fit over 30 minutes of music
on one disk. It would fit a lot more if we had simple songs that compressed
better rather than the complex chiptune examples I picked.
MEMORY LAYOUT
~~~~~~~~~~~~~
As can be seen from the memory map below, if we assume our player can fit in 4k
we have roughly from $2000 to $9600 for memory. That's $7600 (29.5k).
If we could have single buffered, we could have had 256*3*14 (10.5k) for
decompress and 19k for file size which would let us play most of the
reasonable sized songs on our play list (KRW(3) in table at end).
For double buffer, then we need 256*2*14*2 (14k) for decompress
and 16k for file size which still works.
VISUALIZATION
~~~~~~~~~~~~~
Originally I had the volume bars and rasterbars in userspace running,
so it didn't matter how long they took to draw (they'd just get a worse
frame rate if the interrupts were taking a while).
But then I had to move the decompression to userspace, and the visualization
into the interrupt handler.
Then things got interesting. The visualization was taking so much time that
userspace was starved and decompression was not finishing in time, so the
sound was corrupted and finished early.
Thus it was time for some cycle analysis. Here's what I found.
Approx max 20,000 cycles in an interrupt
1,500 used by music decode
7,500 used by volume bars
16,200 (!) used by raster bars
2,000 for misc rest
So the problem can be seen here! That 16,200 for raster bars was
worst-case, it usually would have been a little less.
If takes roughly 700,000 cycles to LZ4 decode a block, so even with
no interrupt can take 35 frames (0.7s) to finish.
I added a variable TIME_TAKEN ($88) that you can use to find out how
long the last LZ4 decode took.
With rasterbars turned off:
INTRO2: 60@19, 60@36, 62@50 61@1:03 61@1:32 60@2:05 61@2:32
So roughly $60 (96) frames, or about 2 seconds.
I went in and optimized the rasterbars code a lot and got it down to
about 10k cycles worst case (6k probably average case).
So now it takes $A0 (160) frames, or about 3 seconds. This seems to
be workable.
Interesting bugs that were hard to debug:
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+ Bug in qkumba's LZ4 decoder, only happened when a copy-block size was
exactly a multiple of 256, in which case it would copy
an extra time.
+ Bug where the box-drawing was starting at 0 rather than at Y.
Turns out I was padding the filename buffer with A0 but going
one too far and it was writing A0 to the first byte of the
hlin routine, and A0 is a LDY # instruction.
+ Our old friend: forgetting the '#' so we're comparing against some random
zero page value rather than a constant
+ Related, the accidentally put in a $ when I meant for it to be decimal.
I was copying to $14 pages instead of 14, overwriting the DOS buffers
which I didn't notice until I tried to load the next file.
FIGURES/TABLES
~~~~~~~~~~~~~~
Memory Map
==========
(not to scale)
------- $ffff
| ROM/IO|
------- $c000
|DOS3.3 |
-------| $9600
| |
| |
| FREE |
| |
| |
|------- $0c00
|GR pg 1|
|------- $0800
|GR pg 0|
------- $0400
| |
------- $0200
|stack |
------- $0100
|zero pg|
------- $0000
File Sizes
==========
Disk(3)
time ym5 KRW(3) KRW(2) Blocks On
~~~~ ~~~ ~~~~~~ ~~~~~~ ~~~~~~~~~~
KORO.KRW 0:54 ? 2707 3039 12
FIGHTING.KRW 1:40 ? 3061 3316 13
CAMOUFLAGE.KRW 1:32 1162 4013 4972 17 17
DEMO4.KRW 2:05 1393 3824 6336 16 16
SDEMO.KRW 2:12 1635 5215 7598 22 22
CHRISTMAS.KRW 1:32 1751 4973 5811 21 21
HARKONEN.KRW 2:46 1803 7256 ???? 30 30
HOLIDAYS.KRW 2:10 2119 5863 ???? 24
SPUTNIK.KRW 2:05 2164 8384 10779 34 34
DEATH2.KRW 2:27 2560 8056 10295 33 33
CRMOROS.KRW 1:29 2566 8007 9565 33 33
TECHNO.KRW 2:23 2630 8896 11126 36 36
WAVE.KRW 2:52 2655 8365 11318 34 34
LYRA2.KRW 3:04 2870 9816 14418 40 40
INTRO2.KRW 2:59 3217 9214 9294 37 37
MMCM.KRW 2:49 3250 11844 ???? 48
ROBOT.KRW 1:26 3448 7717 8337 32 32
UNIVERSE.KRW 1:49 4320 9957 11225 40 40
RANDOM.KRW 2:33 4814 12415 ???? 50 50
NEURO.KRW 3:47 8681 22328 25168 89
AXELF.KRW 10:55 9692 47971 54420 189
----- -----
475 33:14
Notes: my home-made songs don't have ym5 sizes as I don't have a
working LHA encoder to make a real size.
Disk Usage
~~~~~~~~~~
Detailed sector bitmap:
1111111111111111222
0123456789ABCDEF0123456789ABCDEF012
$0: $$$MLLKKKJJJIIHHG#NNOOOPPQQbCDDEEFF
$1: $$$MLLKKKkJJIIHHG#NNOOpPPQQBCDDEEFF
$2: $$$MLLKKKKJJIIHHG#NNOOPPPQQBCDDEEFF
$3: $$$MmLlKKKJJIIHHG#NNOOPPPQQBCDDEEFF
$4: $$$MMLLKKKJJIIiHG#NNOOPPPQQBCDeEEFF
$5: $$$MMLLKKKJJIIIHG#NNOOPPPQQBCDEEfFF
$6: $$$MMLLKKKJJIIIHh#NNOOPPPQQBCDEEFFg
$7: $$$MMLLKKKJJIIIHH#NNOOPPPQQBCDEEFFG
$8: $$$MMLLKKKJJIIIHH#NNOOPPPQQBCDEEFFG
$9: $$$nMLLKKKJJjIIHH#NNOOPPqQQBCDEEFFG
$A: $$$NMLLKKKJJJIIHH#NNOOPPQQQBCDEEFFG
$B: $$$NMLLKKKJJJIIHH#NNOOPPQQ.BCDEEFFG
$C: $$$NMLLKKKJJJIIHH#NNOOPPQQ.BCDEEFFG
$D: $$$NMLLKKKJJJIIHH@NoOOPPQQ.BCDEEFFG
$E: $$$NMLLKKKJJJIIHH@AOOOPPQQ.cCDEEFFG
$F: $$$NMLLKKKJJJIIHH@aOOOPPQQ.CdDEEFFG
Key: $=DOS, @=catalog used, #=catalog reserved, .=free
As you can see, only 5 sectors (1.25k) free.
a HELLO g DEMO4.KRW m SDEMO.KRW
b CHIPTUNE_PLAYER h HARKONEN.KRW n SPUTNIK.KRW
c CAMOUFLAGE.KRW i INTRO2.KRW o TECHNO.KRW
d CHRISTMAS.KRW j LYRA2.KRW p UNIVERSE.KRW
e CRMOROS.KRW k RANDOM.KRW q WAVE.KRW
f DEATH2.KRW l ROBOT.KRW
YM5 Compression Study
=====================
For example, intro2.ym5
raw: 125440 bytes
Compressed, frame at a time (r0..r13, repeat)
lzss: 44154 bytes
gzip: 17119 bytes
lz4: 14666 bytes (-16) (14685 -9, 21377 default)
bzip2: 12685 bytes
lzma (xz) 5312 bytes
Interleaved then Compressed (all of r0 in a row, then all of r1, etc).
lzss/interleaved: 7981 bytes
lha/interleaved: 3217 bytes <=== default used by ym5 format
lz4/interleaved: 3190 bytes (-16) (8914 default, 3209 -9)
bzip2/interleaved 3017 bytes
gzip/interleaved: 2759 bytes
lzma/interleaved: 2129 bytes
Split up, Interleaved, LZ4
lz4,1024*14 chunks 7971 bytes (-16) (14k output buffer)
lz4,768*14 chunks 9214 bytes (-16) (10.5k output buffer)
lz4,512*14 chunks 9294 bytes (-16) (7k output buffer)
Diff (each frame only update registers that change via bitmask)
This is method I used in the KSP demo
simple diff: 69224 bytes
lzss/diff: 31919 bytes
lz4/diff: 13669 bytes (11431 -9)
gzip/diff: 10821 bytes
bzip2/diff: 10477 bytes
lzma/diff: 7257 bytes
Interrupt Timing / AY write latency:
====================================
Trying to find out why some songs, especially DEMO4 do not sound
as good as they should compared to my pi-chiptune player and
also emulation.
On DEMO4 one issue shows up with the envelope is enabled
on Channel A in the first 10s. In some cases the output just
doesn't happen.
I've analyzed the code in memory and as far as I can tell everything
is being uncompressed and played properly.
I thought maybe it was a timing issue (for example, on my Pi player
it only takes roughly 100us to program all the registers, wheras
on the Apple II it takes at least 3-6x longer).
So it could just be an issue of programming the Envelope register
values mid-output and so you get a few cycles where it's non-atomic?
Really though any inconsistency should be < 1ms.
For the numbers below, we assume 13 regs being written (as it's
unusual in many YM songs to update R13 very often, and we skip
it often (0xff) as writing the register (even with the same value)
apparently resets the counter.
Originally roughly 1500 cycles from start of interrupt
to all AY registers being written.
1600
Moved clock to after, near the visualization stuff, more like
1534 = 13+ (117*13)
Load frame data for next time at end of IRQ, instead of begin
1185 = 13+2+(90*13)
Inlined the mockingboard write routine
1029 = 13+2+(78*13)
Only write registers that change. Added 6 cycles per loop
1107 worst case = 13+2+(10+5+62+7)*13
316 if only one reg changed = 13+2+(18*13)+67
Was unnecessarily saving/restoring value to mem, save 2 cycles/loop
1081 worst case = 13+2+(10+5+60+7)*13
314 if only one reg changed = 13+2+(18*13)+65
Tried unrolling. Just unrolling one channel increased the size
of the executable by 454 bytes. By unrolling you can shave a few
cycles off by hard-coding the current reg rather than using X.
1235 to write both channels
627 worst case one channel = 13 + (10+37)*13
360 one reg both channels
180 one reg = 13 + (10)*13+37
3537900858
