Raspberry Casket

A fast and small open source Pretracker replayer

Raspberry Casket Player V2.x (10-Aug-2023)

Provided by Chris 'platon42' Hodges chrisly@platon42.de

Rewritten by platon42/Desire based on a resourced, binary identical version of the original Pretracker V1.0 replayer binary provided by hitchhikr (thanks!), originally written in C by Pink/Abyss.

This version is the hard work of reverse engineering all the offsets, removing all the C compiler crud, removing dead and surplus code (maybe artefacts from earlier ideas that did nothing), optimizing the code where possible. This resulted in both reduced size of the replayer, faster sample calculation and speeding the tick routine up significantly.

I also added a few optional features that come in handy, such as song-end detection and precalc progress support.

Note that this player can be also used for the playback of Pretracker 1.5 tunes, given you don't use sfx or sub-songs.

Also: Open source. It's 2023, keeping the code closed is just not part of the demoscene spirit (anymore?), at least for a replayer.

Also note that this is not the final state of the source code. I could go over many places still and try to rework them. But I wanted the code to be out in public.

Productions that I know have been using Raspberry Casket so far:

• Smooth Flowing/Dekadence
• Cracking Posadas/Software Failure

Verification

The first versions of the replayer had been verified against about 60 Pretracker tunes to create an identical internal state for each tick and identical samples (if certain optimizations switches are disabled).

During the process this identical state and identical samples promise had to be dropped due to bugs in the original player and optimizations. This is especially the case for the track delay feature of Pretracker that could in some cases cause odd behaviour and unwanted muting that has been fixed in Raspberry Casket.

If you find some problems, please let me know under chrisly@platon42.de. Thank you.

Usage

The new replayer comes as a drop-in binary replacement if you wish. In this case you will get faster sample generation (about 12% faster on 68000) and about 45% less CPU time spent during playback. However, you won't get stuff as song-end detection and precalc progress this way. This mode uses the old CPU DMA wait that takes away 8 raster lines.

If you want to get rid of the unnecessary waiting, you can switch to a copper driven audio control. If you want to use the top portion of the copperlist for this, you probably need to double buffer it. Otherwise, you could also position the copperlist at the end of the display and use single buffering if you call the tick routine during the vertical blank.

Please use the documented sizes for the MySong and MyPlayer data structures, which are the symbols sv_SIZEOF and pv_SIZEOF respectively (about 2KB and 12KB with volume table).

The source needs two common include files to compile (custom.i and dmabits.i). You should leave assembler optimizations enabled.

1. (If you're using copper list mode, call pre_PrepareCopperlist.)

2. Call pre_SongInit with

   • a pointer to MySong (mv_SIZEOF) in a1 and
   • the music data in a2. It will return the amount of sample memory needed in d0.

3. Then call pre_PlayerInit with

   • a pointer to MyPlayer (pv_SIZEOF) in a0
   • a pointer to chip memory sample buffer in a1
   • the pointer to MySong in a2
   • a pointer to a longword for progress information or null in a3

   This will create the samples, too.

4. After that, regularly call pre_PlayerTick with MyPlayer in a0 and optionally the copperlist in a1 if you're using that mode).

Size

The original C compiled code was... just bad. The new binary is less than 1/3 of the original one.

The code has been also optimized in a way that it compresses better. The original code compressed with Blueberry's Shrinkler goes from 18052 bytes down to 9023 bytes.

Raspberry Casket, depending on the features compiled in, is about 5840 bytes and shrinkles down to ~4144 bytes (in isolation).

So this means that the optimization is not just "on the outside".

About 2.4 KB of the code (and data) are spent for the sample generation, the remaining code for playback.

Timing

Sample precalculation

Sample generation is a faster than the original 1.0 player and also faster than the 1.5 player, which got a slightly better performance than the 1.0 one (compiler change?).

According to my measurements on my set of Pretracker tunes, Raspberry Casket needs between 10% to 20% less instructions. Of these instructions, about 5% are muls operations and the new player is only able to shave off between 3% and 8% percent of those, so this is probably the limiting factor.

Playback

Raspberry Casket is about twice as fast as the old replayer for playback.

Unfortunately, the replayer is still pretty slow and has high jitter compared to other standard music replayers.

This means it may take up to 32 raster lines (13-18 on average) which is significant more than a standard Protracker replayer (the original one could take about 60 raster lines worst case and about 34 on average!).

Watch out for Presto, the LightSpeedPlayer variant that should solve this problem.

Known issues

• Behaviour for undefined volume slides with both up- and down nibble specified is different (e.g. A9A, hi Rapture!). Don't do that.
• Don't use loops with odd lengths and offsets (even if Pretracker allows this when dragging the loop points).
• Don't stop the music with F00 and use a note delay (EDx) in the same line.
• Don't try to play music with no waves, instruments or patterns.
• Pattern breaks with target row >= $7f will be ignored.
• Shinobi seemed to have used an early beta version of Pretracker where it was possible to specify a Subloop Wait of 0. That's illegal and unsupported.
• Pattern break (Dxx) + Song pos (Bxx) on the same line does not work in original Pretracker & Player: New Dxx position is ignored. There is code to enable it in the player, so you could in theory make backwards running tracks like in Protracker. But this doesn't make sense as long as the tracker itself does not support it.
• Setting the same track delay multiple times will no longer mute the delayed channel.
• Clearing the track delay (multiple times) will no longer mute the delayed channel nor cause a delay of one tick to the note played in the no-longer delayed channel.

Changelog

V2.x (unreleased)

• Split wave generation out of main file, reorganised content into header files.
• Optimized some more code paths for Raspberry Casket replayer.
• In the wave generator optimized away a table (32 words), replacement code is even smaller!
• Replaced the period table by byte-deltas, saved 36 bytes and compression is even better!
• Optimized some code paths for octave selection.
• Removed two 25 bytes tables each, saving another 42 bytes.
• Completely reworked track delay handling, fixed oddities and improved output quality.
• This removes a big source of cpu jitter when track delay is enabled (no longer clearing the track delay buffer).
• This also fixes usages of illegal period 0 in the lead-in that could cause the replay to miss the first trigger.
• Moved pattern table init from PlayerInit to SongInit, optimized SongInit a bit.
• Wave order table filling moved and optimized in SongInit.
• Added Presto player draft.
• Drop-in replacement code size: 5840 bytes.

V1.x (unreleased)

• Fixed a bug regarding the copper output mode with looping waves having a loop-offset.
• Fixed wrong register use on triggering waves regarding the loop offset.
• Minor code size optimizations.

V1.1 (28-Dec-22)

• Optimized base displacement by reordering variables.
• Further optimized ADSR code.
• Optimized wave loop code.
• Baked in this strange vibrato speed multiplication to precalculated vibrato value (where possible).
• Various small optimizations.
• Store instrument number * 4 on loading to avoid using two adds every frame.
• Optimized speed/shuffle code. Idea of using xor turned out to make things too complicated for pattern breaks/jumps.
• Rearranged code for more short branches.
• Optimized track delay code further.
• Optimized pattern / song advance code.
• Maximum jitter now about one rasterline less, average about 0.5 rasterlines less (measurements, your mileage may vary).
• Drop-in replacement code size: 6228 bytes.

V1.0 (26-Dec-22)

• Initial release.
• Drop-in replacement code size: 6446 bytes.