Tests: Timings include initialization

[jllanfranchi]

The initialization of the gpu histogrammer with a given ftype is included in the timings; while this is interesting on some level, it is also interesting (and moreso for PISA) seeing the timing of solely the histogramming operation itself. (We will instantiate the gpu histogrammer just once at the start of invoking PISA, probably the first time a PISA module is imported; from there on out, the same histogrammer will be used. So the init time penalty hits us just once, compared to countless uses of the histogrammer there on out.) Note that I made this change in PR #27, to only timing the histogramming and not the init.

One could also time both the init and the histogramming separately and report both. That way the number of trials won't impact the results (where the impact of init diminishes with # of trials if you include both). And then the user can make a fully informed judgement of whether it's worthwhile to use this for his/her application (and how to go about using it in a sufficiently efficient way).

[mhieronymus]

For every configuration (number of bins, dimensions and bins, single or double precision) the histogram object gets initialized once and several histograms are computed (tests = 10 at the moment). It would be better to create tests, where the init and the histogramming are timed separately.

Edit:
I merged your PR #27 into another branch merge_conflict. Once I get pandas working with the mess I call plot_timings we can see more detailed information. Unfortunately it will be this weekend until I have time to work on it again.

[mhieronymus]

The master branch now has your optimizations. I did not include the init in the test but I might do this later. Making a test is now very easy:
python main.py --test --outdir dir
You may take a look at the updated wikipage: creating benchmarks (the image used there is still old).

Edit:
The wiki just got updated benchmarks.