Craig Dillabaugh:
If I had been coding this I would almost certainly have passed
the splitting dimensions as a function parameter, since the
compiler has to generate separate makeTree/findMedian functions
for each dimension that i/idx takes in the program.
So my question is, does anyone have any idea why the author may
have used templates here. It was on Rosettacode so maybe they
are
just there to show off D's nice template abilities, but I am
curious if in idiomatic D using templates this way is somehow
superior to the function parameter method I would have tried.
That code was adapted by me from C code written by another person
(I think Ledrug, a very good C programmer). I am maintaining all
the many D entries of Rosettacode (plus few in other languages).
John Colvin explains the point of passing that argument as
template argument instead of run-time argument. The number of
dimensions is limited (often 2 or 3), to for this program the
amount of generated template bloat is acceptable. Knowing the
axis at compile-time should allows the compiler to perform better
optimizations.
Beside binary size, the other disadvantage of template bloat is a
pressure increase on the 16 kb of the half of L1 code cache of
the CPU. So there are always trade-offs.
If I can find the time perhaps I will run some benchmarks on it.
If you want to do benchmarks don't forget to do them with a
better compiler, like LDC2, that is able to use that compile-time
information to optimize better.
When I have written that code I was using only DMD. I have just
started to use LDC2 and I have not yet timed that particular
program with LDC2.
Bye,
bearophile
