measure_function_glm

[gottfriedleibniz]

In PyGLM vs NumPy.py have you considered compensating for how much time is spent in the NumPy/PyGLM functions versus the ancillary measurement bits: per-iteration time.time, formatting, printing, etc. For example, here is my dot product measurement:

| dot product                            |    6.623M  |    1.512M  |     4.38x |

Commenting out func(*args, **kw) and executing again:

| dot product                            |    8.303M  |    8.282M  |     1.00x |

This seems to suggest the measurement function is unintentionally a bottleneck for the libraries actual (or potential) throughput.

Thanks.

[Zuzu-Typ]

Hey there (:

Thank you for your input!

I had this in the back of my head for a while now.
The way I used to measure the performance was very much flawed, but I never got around to fixing it.
Until now that is.

I'm now using the timeit library for the actual measurements, which should be far more reliable.
This also makes it easier to add new tests.

In case you're curious, here is the new script and here is the output.

Cheers
--Zuzu_Typ--

[gottfriedleibniz]

Nice. These ratios resemble what I was seeing locally. In my case was doing over a billion iterations of each function with some additional logic to, hopefully, ensure I was not measuring just caching effects.

For reference, here were some of my results:

+----------------------------------------+-----------+
| dot product                            |    15.50x |
+----------------------------------------+-----------+
| cross product                          |   372.65x |
+----------------------------------------+-----------+
| L2-Norm of 3 component vector          |    62.89x |
+----------------------------------------+-----------+