[mad-dev] MAD 0.15 Accuracy

Hi!

I tested my implementation of some math function for MAD 0.15 library. And I have found this:

function mad_fixed_t III_requantize(unsigned int value, signed int exp) in layer3.c file

input value=8 exp=-23;

that means 8^(4/3)*2(-5.75) correct?

result obtained by mad_f_todouble(requantized)is

0.297241

but calculator says

...

8^(4/3)*2^(-5.75)

ans = 0.29730177875068

for Q15 format its means ~2 bits difference.

...

(0.29730177875068-0.29724100000000)*32768

ans = 1.99159810228230

my Q15 arithmetics gives: 9743/32768=0.29733276367188

...

(0.29730177875068-0.29733276367188)*32768

ans = -1.01531389788215

-1 bit difference

And total difference between MAD math and my = 3 bits!

Maybe I have to compare my math not with MAD - but with some float point implementation?