Lightnews — Scholar-powered news

LightNews

mmalex

@mmalex.bsky.social

1.9K followers 330 following 130 posts

synth tinkerer (plinky); midjourney; NVResearch (InstantNGP/NeRF); cofounder MediaMolecule (Dreams, LittleBigPlanet); demoscene/vj (statix/bluespoon)

Posts Media Videos Starter Packs

mmalex @mmalex.bsky.social · 4d

kids are into (2^31)-1 these days wcgw

mmalex @mmalex.bsky.social · 4d

dude, 23. come on

mmalex @mmalex.bsky.social · 8d

yeah that's what i ended up using. it came at the right time :)

mmalex @mmalex.bsky.social · 8d

still messin' with this thing. so many yaks to shave when you start from scratch. autocomplete, filter dsp, sample loading, even json parsing. an eternal pile of todos, but each one small enough I'm making slow but continual forward progress...

mmalex @mmalex.bsky.social · 11d

im sure there are vsts that are as low latency as possible. i guess the vst api / host adds one 'host/plugin' block that can't be avoided. i wonder if anyone would dare make a daw with blocksize 1 lol

mmalex @mmalex.bsky.social · 11d

that's what my thread is about! as described it has 0 latency :)

mmalex @mmalex.bsky.social · 11d

nice! yes that's a good one. on dreams i also downsampled later parts of long irs. i never bothered exploiting reuse between overlapping ffts of different sizes. i wonder if that's a big win.

mmalex @mmalex.bsky.social · 11d

what would you say is the biggest one missing from my outline?

mmalex @mmalex.bsky.social · 11d

(eg ffts must be zero padded & thus double the size of the stated block size, outputs overlap add etc) but i wanted to sketch the shape. when i remembered it i got muddied up on scheduling details, and it was seeing it as independent parallel chunks that unlocked clarity for me. hope its interesting

mmalex @mmalex.bsky.social · 11d

is computed in blocks of 512 by fft. next 1024, blocks of 1024... etc! success! tldr: we broke our ir into chunks of size 128, 128, 256, 512, 1024... with all but the first computed blockwise by fft. voila, low latency and (relatively) low compute. lots of refinements possible and details elided

mmalex @mmalex.bsky.social · 11d

delayed 92858. hm, useful! now we have all the ingredients! chunk your IR into an initial chunk (say) 256 samples. we'll do these as a brute force FIR. the next 256 sample chunk, needs to be output with 256 samples latency. how convenient, that's what fft gives us! the next chunk of 512, likewise,

mmalex @mmalex.bsky.social · 11d

convolution is linear, which means we can cut up our original IR into any chunks we want, and treat each chunk as a totally separate independent reverb, run them all in parallel and add them back up. key point: if a chunk comes from a point 92858 samples into your IR, you *want* its output to be

mmalex @mmalex.bsky.social · 11d

you can also do this by precopmuting the fourier transform of your IR; then you do the FT of a block of N input samples, hadamard product the two, and inverse fft. since fft is O(NlogN) this is faster for large enough N! sadly, since its block based, it introduces a delay of N samples. but...

mmalex @mmalex.bsky.social · 11d

an ir reverb is just a convolution, like a big FIR filter. if your IR is N samples long, for every output sample you just 'dot' together (multiply and add in a for loop) the N ir samples with the last N input samples (remembered in a buffer). so to output N samples, it costs O(N^2). but...

mmalex @mmalex.bsky.social · 11d

i wrote an efficient convolution reverb for dreams while at mm years ago, and found myself in need of one again. my memory was hazy on the details, and i couldnt find it written out in a simple way i could immediately understand, so i chewed a pencil and reminded myself. heres how i think about it

mmalex @mmalex.bsky.social · 11d

lovely groove

mmalex @mmalex.bsky.social · 12d

dual wielding

mmalex @mmalex.bsky.social · 13d

all the best things only make sense in your head at first:) or something.... i tell myself....

mmalex @mmalex.bsky.social · 14d

also added 'keypoints' (by saying that space means 'unset', and . or A means 0) and actually they're more useful than i expected. (wait for it.... a to z!)

mmalex @mmalex.bsky.social · 14d

(side comment - having an old-school text-mode screen of 8 bit characters & attributes as my only rendering primitive is... freeing, weirdly. i forgot how much I love text mode. just poke a byte in memory to draw a letter. yay.)

mmalex @mmalex.bsky.social · 14d

leaning into this text editor/visual editor dual, i've made the code colourer detect comments like /*====*/ and turn them into sliders. I really like how plain old source code remains the 'source of truth' for all state and the editor just doodles over it.

mmalex @mmalex.bsky.social · 14d

not yet :) maybe one day! one can dream! :)

mmalex @mmalex.bsky.social · 14d

i was kinda meh until i saw the rotating return key and now im ALL IN

mmalex @mmalex.bsky.social · 14d

i really appreciate C99's variable length arrays. so you can write float foo[n]; where n is runtime variable!
in before the safety crowd: obviously just as unsafe as alloca (ie be careful of n's range) but less error prone as you dont need to compute a byte size by hand. and more readable.

mmalex @mmalex.bsky.social · 14d

its a super fun rabbit hole!