Quantizing weights & history samples for the frame header causes audible clicks about qoa HOT 6 CLOSED

Sounds like a good approach to me, especially because it trivially un-nests the history and weight data.

As for better quantization: sqrt only seems to be a better solution because of integer math (for -1 <-> 1 floats sqrt(x) ~= x) but is not worth the added complexity in my opinion. I could not find a simple algorithm for square roots, at least not a non-iterative one. As said, the only approximation known to me only works for normalized rational sample representations.

from qoa.

lolremez is very good at approximating with only a few multiplications.

from qoa.

Btw. in Nuklear we use inverse square root as its usage results in faster generated assembly.

from qoa.

lolremez is very good at approximating with only a few multiplications.

But it requires floating-point math, which QOA explicitly avoids

from qoa.

But it requires floating-point math, which QOA explicitly avoids

It does not. lolremez generates C code with floating point data types but the principle works for integers too (with a small additional effort, of course).

from qoa.

I went ahead and did the obvious thing now.

lolremez looks amazing. I wish I had know about this earlier to different things on the residuals. For this case, the added complexity of yet another encoding scheme is certainly not worth saving 1kbit/s

from qoa.