References : Posted by didid[AT]skynet[DOT]be
Because I still see people adding noise or offset to their signal to avoid slow denormalization, here's a piece of code to zero out (near) tiny numbers instead.
Why zeroing out is better? Because a fully silent signal is better than a little offset, or noise. A host or effect can detect silent signals and choose not to process them in a safe way.
Plus, adding an offset or noise reduces huge packets of denormalization, but still leaves some behind.
Also, truncating is what the DAZ (Denormals Are Zero) SSE flag does.
This code uses integer comparison, and a CMOV, so you need a Pentium Pro or higher.
There's no need for an SSE version, as if you have SSE code you're probably already using the DAZ flag instead (but I advise plugins not to mess with the SSE flags, as the host is likely to have DAZ switched on already). This is for FPU code. Should work much faster than crap FPU comparison.
Den_Thres is your threshold, it cannot be denormalized (would be pointless). The function is Delphi, if you want to adapt, just make sure EAX is the buffer and EDX is length (Delphi register calling convention - it's not the same in C++).
from : scoofy[AT]inf[DOT]elte[DOT]hu
comment : You can zero out denormals by adding and subtracting a small number.
void kill_denormal_by_quantization(float &val)
static const float anti_denormal = 1e-18;
val += anti_denormal;
val -= anti_denormal;
Reference: Laurent de Soras' great article on denormal numbers:
I tend to add DC because it is faster than quantization. A slight DC offset (0.000000000000000001) won't hurt. That's -360 decibels...
from : gol
comment : >>You can zero out denormals by adding and subtracting a small number
But with drawbacks as explained in his paper.
As for the speed, not sure which is the faster. Especially since the FPU speed is too manufacturer-dependant (read: it's crap in pentiums), and mine is using integer.
>>A slight DC offset (0.000000000000000001) won't hurt
As I wrote, it really does.. hurt the sequencer, that can't detect pure silence and optimize things accordingly. A host can detect near-silence, but it can't know which offset value YOU chose, so may use a lower threshold.
from : gol
comment : Btw, I happen to see I had already posted this code, probably years ago, doh!
Anyway this version gives more explanation.
And here's more:
The LEA EBX,[EAX*2] is to get rid of the sign bit.
And the integer comparison of float values can be done providing both are the same sign (I'm not quite sure it works on denormals, but we don't care, since they're the ones we want to zero out, so our threshold won't be denormalized).
from : antiprosynthesis[AT]gmail[DOT]com
comment : You could also add input noise and assure output samples are reset to 0 if they're below a certain treshold, slightly higher than your noise volume. That ensures hosts can do proper tail detection and it's cheap.