On 09.03.2015 12:26, François wrote:
Thanks for these clarifications.
I guess that most of the time is spent in the *copy_property* [link to git https://git.skewed.de/count0/graph-tool/blob/master/src/graph/graph_properties_copy.cc#L33]. Indeed, when I provide a pre-initialized distance vector through the dist_map parameter, the execution of shortest_distance is only 10ms faster.
If I understood well, the *copy_property *is used to build a vector initialized with one single value. If so, one would obtain the same result with python and numpy with
x = np.empty(array_size, dtype=float) x.fill(value)
I did try to time this "numpy way initialization" (altough I'm not sure it corresponds to *copy_property*)
python -m timeit -s "import numpy as np" -n 100 "np.empty(33e6, dtype=int).fill(1)" 100 loops, best of 3: 34.9 msec per loop
These**34.9ms have to be compared to the 300ms (bake of envelope calculus) that takes the *copy_property *function.
Am I right about the way that the copy_property function works ? Could it be improved ?
PropertyMap.copy() is slower than a simple numpy initialization because it needs to deal with possible conversions between the values (such as converting from string to double). However, it is possible to include a specialization that avoids this conversion when the types are the same. I have now included this modification in the git version, which significantly improves the time it takes to copy a property without conversion.