enh: more realistic benchmark

[zerothi]

I think it is vital that benchmarks are realistic. Using fixed size arrays is not realistic in my opinion.

This PR does many things:

• benchmarks should not rely on convergence, rather a fixed number of iterations is more appropriate, otherwise timings are irrelevant
• The problem of inconsistent iterations between the C and Fortran versions is fixed. It was an indexing error when calculating the cell for the rhoarr array (C is 0-based, F is 1-based). So they now run the same number of iterations, regardless of M.
• Moved timings about to make them more comparable.
• Ensured that all examples now use allocate/malloc
• Calls in C and F code are much more similar

For running benchmarks it is vital that M is so large that the 2D array goes beyond the L3 cache to match realistic cases.

[certik]

@zerothi thanks for this!

I have an idea: why don't you leave poisson2d/optimized.f90 and poisson2d/optimized.c as they are, but submit new files both for Fortran and C with your version. I actually want to have both. I want to see how long it takes when only the computation is performed, but I also want to see when you benchmark everything, including memory allocation, as you did. The same with the changes you made --- I actually would like to have both.

Regarding naming, we can even just do version1.f90, version2.f90, etc.

What do you think?

[zerothi]

Good idea, give me some time :)

[zerothi]

Please have a look at the updated look. It is now more complete and easier to extend with new versions!

[zerothi]

Note that benchmarking the allocations is not really fair I think. In this case I have left all allocation timings out since that should be a benchmark by it-self.
The reason for not doing it is that some things may be allocated on the stack which one can't easily time without complicating matters quite a bit.