GH-100425: Timing experiment: For builtin_sum, try replacing Fast2Sum with 2Sum

[rhettinger]

On the Apple M1 Max, this change makes no difference. I get 303/304 nsec per loop before and after the edit.

Would anyone care to run this on their builds and report back the results?

% ./python.exe -m timeit -r21 -s 'n=100' -s 'from random import expovariate as r' -s 'v1=[r(1000) + r(0.125) for i in range(n)]' 'sum(v1)' 1000000 loops, best of 21: 304 nsec per loop 

• Issue: Improve accuracy of builtin sum() for float inputs #100425

[hauntsaninja]

I see no difference either, on Linux with an AMD Zen 2 chip

[eendebakpt]

Both with and without optimizations I see no difference. System: Linux, gcc version 9.4.0 (Ubuntu 9.4.0-1ubuntu1~20.04.1), Intel(R) Core(TM) i7-4710MQ CPU @ 2.50GHz

[rhettinger]

Thank you both. It would be nice to hear from a Windows person as well.

[eendebakpt]

Thank you both. It would be nice to hear from a Windows person as well.

On Windows (default PCbuild/build.bat, no PGO) the timings vary a lot on my system (Intel(R) Core(TM) i7-4710MQ CPU @ 2.50GHz, Windows 10, VS 2019). For this PR, measurements within 5 minutes:

I can confirm that the minimum time for the test is roughly the same for main and this PR.

[rhettinger]

Thank you. I appreciate it.

@mdickinson Given that 2Sum and Fast2Sum have the same performance in the context of builtin.sum(), do we have a non-performance reason to choose one over the other? Or should I leave the sum() code as-is?

[mdickinson]

@rhettinger Leaving as-is sounds good to me. The two should be functionally identical, so performance is just about the only thing that would justify choosing one over the other.