A cross-architecture and cross-platform modular software synthesizer for small intros, forked from 4klang. Targetable architectures include 386, amd64, and WebAssembly; targetable platforms include Windows, Mac, Linux (and related) + browser.
Pull requests / suggestions / issues welcome, through Github! You can also contact me through email (firstname.lastname@gmail.com).
Sointu is work-in-progress. It is a fork and an evolution of 4klang, a modular software synthesizer intended to easily produce music for 4k intros — small executables with a maximum filesize of 4096 bytes containing realtime audio and visuals. Like 4klang, the sound is produced by a virtual machine that executes small bytecode to produce the audio; however, by now the internal virtual machine has been heavily rewritten and extended. It is actually extended so much that you will never fit all the features at the same time in a 4k intro, but a fairly capable synthesis engine can already be fitted in 600 bytes (386, compressed), with another few hundred bytes for the patch and pattern data.
Sointu consists of two core elements:
- A cross-platform synth-tracker app for composing music, written in go. The app is still heavily work in progress. The app exports the projects as .yml files. There are two versions of the app: cmd/sointu-track/, using a plain Go VM bytecode interpreter, and cmd/sointu-nativetrack/, using cgo to bridge calls to the Sointu compiled VM. The former should be highly portable, the latter currently works only on x86/amd64 platforms.
- A compiler, likewise written in go, which can be invoked from the command line to compile these .yml files into .asm or .wat code. For x86/amd64, the resulting .asm can be then compiled by nasm or yasm. For browsers, the resulting .wat can be compiled by wat2wasm.
This is how the current prototype tracker looks like:
Various aspects of the project have different tool dependencies, which are listed below.
This version of the tracker is the version that uses the bytecode interpreter written in plain Go. Running the tracker:
go run cmd/sointu-track/main.go
Building the tracker:
go build -o sointu-track cmd/sointu-track/main.go
On windows, replace -o sointu-track with -o sointu-track.exe.
Sointu-track uses the gioui for the GUI and oto for the audio, so the portability is currently limited by these.
⚠️ Unlike the x86/amd64 VM compiled by Sointu, the Go written VM bytecode interpreter uses a software stack. Thus, unlike x87 FPU stack, it is not limited to 8 items. If you intent to compile the patch to x86/amd64 targets, make sure not to use too much stack. Keeping at most 5 signals in the stack is presumably fine (reserving 3 for the temporary variables of the opcodes). In future, the app should give warnings if the user is about to exceed the capabilities of a target platform.
The command line interface to it is sointu-compile and the actual code resides in the compiler package, which is an ordinary go package with no other tool dependencies.
Running the compiler:
go run cmd/sointu-compile/main.go
Building the compiler:
go build -o sointu-compile cmd/sointu-compile/main.go
On windows, replace -o sointu-compile with -o sointu-compile.exe.
The compiler can then be used to compile a .yml song into .asm and .h files. For example:
sointu-compile -o . -arch=386 tests/test_chords.yml
nasm -f win32 test_chords.asm
WebAssembly example:
sointu-compile -o . -arch=wasm tests/test_chords.yml
wat2wasm --enable-bulk-memory test_chords.wat
Building the regression tests as executables (testing that they work the same way when you would link them in an intro) requires:
- go
- CMake with CTest
- nasm or yasm
- Your favorite CMake compatible c-compiler & build tool. Results have been obtained using Visual Studio 2019, gcc&make on linux, MinGW&mingw32-make, and ninja&AppleClang.
For example, using ninja:
mkdir build
cd build
cmake .. -GNinja
ninja
ninja test
Note that this builds 64-bit binaries on 64-bit Windows. To build 32-bit binaries on 64-bit Windows, replace in above:
cmake .. -DCMAKE_C_FLAGS="-m32" -DCMAKE_ASM_NASM_OBJECT_FORMAT="win32" -GNinja
Another example: on Visual Studio 2019 Community, just open the folder, choose either Debug or Release and either x86 or x64 build, and hit build all.
The native bridge allows Go to call the sointu compiled virtual machine, through cgo, instead of using the Go written bytecode interpreter. It's likely slightly faster than the interpreter. The version of the tracker that uses the native bridge is sointu-nativetrack. Before you can actually run it, you need to build the bridge using CMake (thus, the nativetrack does not work with go get)
Building the native bridge requires:
- go
- CMake
- nasm or yasm
- cgo compatible compiler e.g. gcc. On windows, you best bet is MinGW. We use the tdm-gcc
The last point is because the command line player and the tracker use cgo to interface with the synth core, which is compiled into a library. The cgo bridge resides in the package bridge.
⚠️ you must build the library inside a directory called 'build' at the root of the project. This is because the path where cgo looks for the library is hard coded to point to build/ in the go files.
So, to build the library, run (this example is using ninja for the build; adapt for other build tools accordingly):
mkdir build
cd build
cmake .. -GNinja
ninja sointu
Running ninja sointu only builds the static library that Go needs. This is a
lot faster than building all the CTests.
You and now run all the Go tests, even the ones that test the native bridge. From the project root folder, run:
go test ./...
Play a song from the command line:
go run cmd/sointu-play/main.go tests/test_chords.yml
Run the tracker using the native bridge
go run cmd/sointu-nativetrack/main.go
⚠️ If you are using MinGW and Yasm: Yasm 1.3.0 (currently still the latest stable release) and GNU linker do not play nicely along, trashing the BSS layout. See here and the fix here. Use a newer nightly build of yasm that includes the fix. The linker had placed our synth object overlapping with DLL call addresses; very funny stuff to debug.
⚠️ The sointu-nativetrack cannot be used with the syncs at the moment. For syncs, use the Go VM (sointu-track).
These are automatically invoked by CTest if node and wat2wasm are found in the path.
- New units. For example: bit-crusher, gain, inverse gain, clip, modulate bpm (proper triplets!), compressor (can be used for side-chaining).
- Compiler. Written in go. The input is a .yml file and the output is an
.asm. It works by inputting the song data to the excellent go
text/templatepackage, effectively working as a preprocessor. This allows quite powerful combination: we can handcraft the assembly code to keep the entropy as low as possible, yet we can call arbitrary go functions as "macros". The templates are here and the compiler lives here. - Tracker. Written in go. A crude version exists.
- Supports 32 and 64 bit builds. The 64-bit version is done with minimal changes to get it work, using template macros to change the lines between 32-bit and 64-bit modes. Mostly, it's as easy as writing {{.AX}} instead of eax; the macro {{.AX}} compiles to eax in 32-bit and rax in 64-bit.
- Supports compiling into WebAssembly. This is a complete reimplementation of the core, written in WebAssembly text format (.wat).
- Supports Windows, Linux and MacOS. On all three 64-bit platforms, all tests are passing. Additionally, all tests are passing on windows 32.
- Per instrument polyphonism. An instrument has the possibility to have any number of voices, meaning that multiple voices can reuse the same opcodes. So, you can have a single instrument with three voices, and three tracks that use this instrument, to make chords. See here for an example and here for the implementation. The maximum total number of voices is 32: you can have 32 monophonic instruments or any combination of polyphonic instruments adding up to 32.
- Any number of voices per track. A single track can trigger more than one voice. At every note, a new voice from the assigned voices is triggered and the previous released. Combined with the previous, you can have a single track trigger 3 voices and all these three voices use the same instrument, useful to do polyphonic arpeggios (see here). Not only that, a track can even trigger voices of different instruments, alternating between these two; maybe useful for example as an easy way to alternate between an open and a closed hihat.
- Easily extensible. Instead of %ifdef hell, the primary extension mechanism is through new opcodes for the virtual machine. Only the opcodes actually used in a song are compiled into the virtual machine. The goal is to try to write the code so that if two similar opcodes are used, the common code in both is reused by moving it to a function. Macro and linker magic ensure that also helper functions are only compiled in if they are actually used.
- Songs are YAML files. These markup files are simple data files, describing the tracks, patterns and patch structure (see here for an example). The sointu-compile then reads these files and compiles them into .asm code. This has the nice implication that, in future, there will be no need for a binary format to save patches, nor should you need to commit .o or .asm to repo: just put the .yml in the repo and automate the .yml -> .asm -> .o steps using sointu-compile & nasm.
- Harmonized support for stereo signals. Every opcode supports a stereo variant: the stereo bit is hidden in the least significant bit of the command stream and passed in carry to the opcode. This has several nice advantages: 1) the opcodes that don't need any parameters do not need an entire byte in the value stream to define whether it is stereo; 2) stereo variants of opcodes can be implemented rather efficiently; in some cases, the extra cost of stereo variant is only 5 bytes (uncompressed). 3) Since stereo opcodes usually follow stereo opcodes (and mono opcodes follow mono opcodes), the stereo bits of the command bytes will be highly correlated and if crinkler or any other modeling compressor is doing its job, that should make them highly predictable i.e. highly compressable.
- Test-driven development. Given that 4klang was already a mature project, the first thing actually implemented was a set of regression tests to avoid breaking everything beyond any hope of repair. Done, using go test (runs the .yml regression tests through the library) and CTest (compiles each .yml into executable and ensures that when run like this, the test case produces identical output). The tests are also ran in the cloud using github actions.
- Arbitrary signal routing. SEND (used to be called FST in 4klang) opcode normally sends the signal as a modulation to another opcode. But with the new RECEIVE opcode, you just receive the plain signal there. So you can connect signals in an arbitrary way. Actually, 4klang could already do this but in a very awkward way: it had FLD (load value) opcode that could be modulated; FLD 0 with modulation basically achieved what RECEIVE does, except that RECEIVE can also handle stereo signals. Additionally, we have OUTAUX, AUX and IN opcodes, which route the signals through global main or aux ports, more closer to how 4klang does. But this time we have 8 mono ports / 4 stereo ports, so even this method of routing is unlikely to run out of ports in small intros.
- Pattern length does not have to be a power of 2.
- Sample-based oscillators, with samples imported from gm.dls. Reading gm.dls is obviously Windows only, but with some effort the sample mechanism can be used also without it, in case you are working on a 64k and have some kilobytes to spare. See this example, and this go generate program parses the gm.dls file and dumps the sample offsets from it.
- Unison oscillators. Multiple copies of the oscillator running slightly detuned and added up to together. Great for trance leads (supersaw). Unison of up to 4, or 8 if you make stereo unison oscillator and add up both left and right channels. See this example.
- Compiling as a library. The API is very rudimentary, a single function render, and between calls, the user is responsible for manipulating the synth state in a similar way as the actual player does (e.g. triggering/ releasing voices etc.)
- Calling Sointu as a library from Go language. The Go API is slighty more sane than the low-level library API, offering more Go-like experience.
- A bytecode interpreter written in pure go. It's slightly slower than the hand-written assembly code by sointu compiler, but with this, the tracker is ultraportable and does not need cgo calls.
- Using Sointu as a sync-tracker. Similar to GNU Rocket, but (ab)using the tracker we already have for music. We use the Go rpc package to send current sync values from the new SYNC opcode + optionally the current fractional row the song is on. The syncs are saved every 256th sample (approximately 172 Hz). For 4k intro development, the idea is to write a debug version of the intro that merely loads the shader and listens to the RPC messages, and then draws the shader with those as the uniforms. Then, during the actual 4k intro, one can get the sync data from Sointu: if the song uses syncs, su_render_song writes the syncs to a float array. During each time step, a slice of this array can be sent to the shader as a uniform float array. A track with two voices, triggering an instrument with a single envelope and a slow filter, can even be used as a cheap smooth interpolation mechanism, provided the syncs are added to each other in the shader.
- Find a more general solution for skipping opcodes / early outs. It might be a new opcode "skip" that skips from the opcode to the next out in case the signal entering skip and the signal leaving out are both close to zero. Need to investigate the best way to implement this.
- Even more opcodes. Some potentially useful additions could be:
- Equalizer / more flexible filters
- Very slow filters (~ DC-offset removal). Can be implemented using a single bit flag in the existing filter
- Arbitrary envelopes; for easier automation.
- MIDI support for the tracker.
- Find a solution for denormalized signals. Denormalized floating point numbers (floating point numbers that are very very small) can result in 100x CPU slow down. We got hit by this already: the damp filters in delay units were denormalizing, resulting in the synth being unusable in real time. Need to investigate a) where denormalization can happen; b) how to prevent it: add & substract value; c) make this optional to the user. For quick explanation about the potential massive CPU hit, see https://stackoverflow.com/questions/36781881/why-denormalized-floats-are-so-much-slower-than-other-floats-from-hardware-arch
- Hack deeper into audio sources from the OS. Speech synthesis, I'm eyeing at you.
- Ability to run Sointu as a DAW plugin (VSTi3). Now that Renoise supports VSTi3, there's no fundamental objection to compiling Sointu as a VSTi3. But don't expect it any soon; I need to digest the idea of having to learn the horrors of the VSTi3 C++ API.
- Make sure the assembly code is readable after compiling: it should have liberally comments in the outputted .asm file. This allows humans to study the outputted code and figure out more easily if there's still way to squeeze out instructions from the code.
- Instead of prematurely adding %ifdef toggles to optimize away unused features, start with the most advanced featureset and see if you can implement it in a generalized way. For example, all the modulations are now added into the values when they are converted from integers, in a standardized way. This got rid of most of the %ifdefs in 4klang. Also, with no %ifdefs cluttering the view, many opportunities to shave away instructions became apparent. Also, by making the most advanced synth cheaply available to the scene, we promote better music in future 4ks :)
- Size first, speed second. Speed will only considered if the situation becomes untolerable.
- Benchmark optimizations. Compression results are sometimes slightly nonintuitive so alternative implementations should always be benchmarked e.g. by compiling and linking a real-world song with Leviathan and observing how the optimizations affect the byte size.
4klang development was started in 2007 by Dominik Ries (gopher) and Paul Kraus (pOWL) of Alcatraz. The write-up will still be helpful for anyone looking to understand how 4klang and Sointu use the FPU stack to manipulate the signals. Since then, 4klang has been used in countless of scene productions and people use it even today.
However, 4klang seems not to be actively developed anymore and polyphonism was implemented only in a rather limited way (you could have exactly 2 voices per instrument if you enable it). Also, reading through the code, I spotted several avenues to squeeze away more bytes. These observations triggered project Sointu. That, and I just wanted to learn x86 assembly, and needed a real-world project to work on.
"Sointu" means a chord, in Finnish; a reference to the polyphonic capabilities of the synth. I assume we have all learned by now what "klang" means in German, so I thought it would fun to learn some Finnish for a change. And there's enough klangs already.
Adam by brainlez Coders! - My first test-driving of Sointu. Some ideas how to integrate Sointu to the build chain.
The original 4klang was developed by Dominik Ries (gopher) and Paul Kraus (pOWL) of Alcatraz. ❤️
Sointu was initiated by Veikko Sariola (pestis/bC!).
Apollo/bC! put the project on the path to Go, and wrote the prototype of the tracker GUI.
PoroCYon's 4klang fork inspired the macros for better cross-platform support.