Revisiting old assumptions in Specter lately. It's disheartening to find things that need reworking after so much time spent building them.
But correcting those assumptions now isn't falling behind. It's making sure the work ahead is built on something I can trust.
Two oscillators loading the same wavetable were preparing separate copies. ~18 MiB of duplicate data.
Sharing now removes the extra copy and passes synchronous parity.
Async testing exposed an older renderer-equivalence assumption. Investigating that next.
Found Specter preparing the same FM band data seven times per sample: one carrier, six unison lanes.
Shared the invariant work. Output and FM state stay bit-identical.
5.6% faster in that path, 2.2% better throughput on the stress patch.
Still arguing with microseconds.
Some music makes you nostalgic for a place you've never been, or something you can't quite remember.
I think that's the feeling I'm chasing when I sit down to make music.
And somehow, that's led me to engineering my own synthesizer.
This synth doesn’t have a name yet. It’s been running under a codename for months and I’m not attached to it.
Additive/FM, up to 1024 partials, a filter that likes to self-oscillate.
What would you call it?
I just wanted to make cooler bass sounds...
Now I'm threatening MSVC with "dumpbin /DISASM" all because a memory-layout change cost my oscillator 9% performance at 3 AM.
Big Synth forced me to learn the machine's native language.
Building a synth.
Additive/FM, 3 oscillators per voice, up to 1024 partials, wavetables, and a filter I picked mostly for what it does when you push it into self-resonance.
Pre-release.
One person.
No date yet.
I'll post what changes as it changes.
Less glamorous: voices that have gone quiet now do less work, and some modulation math only runs when something's listening to it. Lighter on CPU in some patches. Also cleared dead parameters out of your DAW's automation list. Old presets still load.
Second one: in certain patch setups, unison was collapsing toward mono instead of spreading across the stereo field. Now it spreads the way it should no matter how the patch is set up. Wider pads, same knob.
Two unison bugs down this week. First one: on some patches, turning unison past 1 made part of the sound quietly drop out. You'd turn it up and get less. Fixed.
Fixed a sneaky one: load a preset, play the first note, and the pulse width would glide in from the wrong place instead of starting where the preset saved it.
Now it starts where it should.
Every note after that already did.