Benchmark: Savage

[Albert Chan]

Doing Savage simulation, using Python gmpy2 / pdeclib (n=2499, radian mode):

12-decimals (halfway to even) => a = 2499.99942 402, relative error ~ 2.3E-7
48-bits CAS float (simulated) ==> a = 2499.99998 195, relative error ~ 7.2E-9

RoundToNearest:
48-bits => a = 2500.00000 041, relative error ~ 1.7E-10
47-bits => a = 2499.99999 957, relative error ~ 1.7E-10
46-bits => a = 2499.99999 984, relative error ~ 6.6E-11
45-bits => a = 2500.00000 108, relative error ~ 4.3E-10
44-bits => a = 2500.00000 179, relative error ~ 7.2E-10
43-bits => a = 2500.00000 463, relative error ~ 1.9E-9
42-bits => a = 2500.00001 457, relative error ~ 5.8E-9 <== CAS float ~ 42 bits
...
37-bits => a = 2500.00032 413, relative error ~ 1.3E-7 <== 12-decimals ~ 37 bits

CAS float is worse than expected due to its biased toward zero rounding.
12-decimals do not get back 40 bits. Decimals roundings had bigger "gap", not 1 ULP ...

Does above numbers match your Android emulation ?