Improve errors in summations within FFTs

[xfong]

The results of FFTs can be dependent on GPU employed due to rounding errors. It will be nice if the library can be enhanced such that summations handle the rounding errors. One way could be to check the numbers to determine the rounding error:
// a + b
if fabs(a) > fabs(b) {
s = a;
t = s + b;
y = t - a;
s = y - b;
return t - s;
}
else {
s = b;
t = s + a;
y = t - b;
s = y - a;
return t - s;
}