Benchmark: Savage

[Albert Chan]

(09-30-2018 07:20 AM)StephenG1CMZ Wrote:  Pythonesque version seems about twice as fast as the PPL.

That might explained why Python version is more accurate.

CAS float = 53 bits double, truncated to 48 bits.
Because of truncation, last bit is not as good as round-to-nearest, with slight bias.
Let's say it got effectively 47 bits precision, or about 47/3.322 ~ 14 decimals

PPL is much slower, probably doing BCD math (15 internally, rounded to 12 decimals)

BTW, Python user is more comfortable counting from zero.
For 2499 loops, range(2499) look better than range(1,2500)