I had some python code which I tried to port to Julia to learn this lovely language. I used generators in python. After porting it seems to me (at this moment) that Julia is really slow in this area!
I made part of my code simplified to this exercise:
Think 4x4 chess board. Find every N-moves long path, chess king could do. In this exercise, the king is not allowed to leap twice at the same position in one path. Don't waste memory -> make a generator of every path.
Algorithm is pretty simple:
if we sign every position with numbers:
0 1 2 3
4 5 6 7
8 9 10 11
12 13 14 16
point 0 has 3 neighbors (1, 4, 5). We could find a table for every neighbor for every point:
NEIG = [[1, 4, 5], [0, 2, 4, 5, 6], [1, 3, 5, 6, 7], [2, 6, 7], [0, 1, 5, 8, 9], [0, 1, 2, 4, 6, 8, 9, 10], [1, 2, 3, 5, 7, 9, 10, 11], [2, 3, 6, 10, 11], [4, 5, 9, 12, 13], [4, 5, 6, 8, 10, 12, 13, 14], [5, 6, 7, 9, 11, 13, 14, 15], [6, 7, 10, 14, 15], [8, 9, 13], [8, 9, 10, 12, 14], [9, 10, 11, 13, 15], [10, 11, 14]]
PYTHON
A recursive function (generator) which enlarge given path from the list of points or from a generator of (generator of ...) points:
def enlarge(path):
if isinstance(path, list):
for i in NEIG[path[-1]]:
if i not in path:
yield path[:] + [i]
else:
for i in path:
yield from enlarge(i)
Function (generator) which give every path with given length
def paths(length):
steps = ([i] for i in range(16)) # first steps on every point on board
for _ in range(length-1):
nsteps = enlarge(steps)
steps = nsteps
yield from steps
We could see that there are 905776 paths with length 10:
sum(1 for i in paths(10))
Out[89]: 905776
JULIA (this code was created by @gggg during our discussion here )
const NEIG_py = [[1, 4, 5], [0, 2, 4, 5, 6], [1, 3, 5, 6, 7], [2, 6, 7], [0, 1, 5, 8, 9], [0, 1, 2, 4, 6, 8, 9, 10], [1, 2, 3, 5, 7, 9, 10, 11], [2, 3, 6, 10, 11], [4, 5, 9, 12, 13], [4, 5, 6, 8, 10, 12, 13, 14], [5, 6, 7, 9, 11, 13, 14, 15], [6, 7, 10, 14, 15], [8, 9, 13], [8, 9, 10, 12, 14], [9, 10, 11, 13, 15], [10, 11, 14]];
const NEIG = [n.+1 for n in NEIG_py]
function enlarge(path::Vector{Int})
(push!(copy(path),loc) for loc in NEIG[path[end]] if !(loc in path))
end
collect(enlarge([1]))
function enlargepaths(paths)
Iterators.Flatten(enlarge(path) for path in paths)
end
collect(enlargepaths([[1],[2]]))
function paths(targetlen)
paths = ([i] for i=1:16)
for newlen in 2:targetlen
paths = enlargepaths(paths)
end
paths
end
p = sum(1 for path in paths(10))
benchmark
In ipython we could time it:
python 3.6.3:
%timeit sum(1 for i in paths(10))
1.25 s ± 15.1 ms per loop (mean ± std. dev. of 7 runs, 1 loop each)
julia 0.6.0
julia> @time sum(1 for path in paths(10))
2.690630 seconds (41.91 M allocations: 1.635 GiB, 11.39% gc time)
905776
Julia 0.7.0-DEV.0
julia> @time sum(1 for path in paths(10))
4.951745 seconds (35.69 M allocations: 1.504 GiB, 4.31% gc time)
905776
Question(s):
We Julians are saying this: It is important to note that the benchmark codes are not written for absolute maximal performance (the fastest code to compute recursion_fibonacci(20) is the constant literal 6765). Instead, the benchmarks are written to test the performance of identical algorithms and code patterns implemented in each language.
In this benchmark, we are using the same idea. Just simple for cycles over arrays enclosed to generators. (Nothing from numpy, numba, pandas or others c-written and compiled python packages)
Is assumption that Julia's generators are terribly slow right?
What could we do to make it really fast?
