As far as the denormal issue goes -- I assume ext_no_denormal is being set to avoid noise being added to the analyzed input?
Anyway -- the addition of 0.0000001 or whatever helps, but ideally the JSFX code generator should be taking care of this. I'm investigating why it is not. Edit: made some slight tweaks the the denormal-prevention code, which fixes (coming in 5.97 or whatever).
Something I noticed in svf_filter:
Code:
nlp > 0 ? (
v3 = v0 - ic2eq0;
v1 = a1 * ic1eq0 + a2 * v3;
v2 = ic2eq0 + a2 * ic1eq0 + a3 * v3;
ic1eq0 = 2 * v1 - ic1eq0;
ic2eq0 = 2 * v2 - ic2eq0;
v0 = m0 * v0 + m1 * v1 + m2 * v2;
);
nlp > 2 ? (
v3 = v0 - lp1.ic2eq0;
v1 = lp1.a1 * lp1.ic1eq0 + lp1.a2 * v3;
v2 = lp1.ic2eq0 + lp1.a2 * lp1.ic1eq0 + lp1.a3 * v3;
lp1.ic1eq0 = 2 * v1 - lp1.ic1eq0;
lp1.ic2eq0 = 2 * v2 - lp1.ic2eq0;
v0 = lp1.m0 * v0 + lp1.m1 * v1 + lp1.m2 * v2;
);
These would be better off nested (the optimizer is not smart enough to detect that nlp wouldn't change, and it will still run each comparison in sequence even if the first one fails...)