sparse() is slow when Range inputs are used as opposed to vectors

[ViralBShah]

When ranges are passed to sparse(), the code runs 10 times slower.

julia> load("sparse.jl")
julia> x = 1:1000000;
julia> y = [x];
julia> @time sparse(x,x,x);
elapsed time: 0.4299280643463135 seconds
julia> @time sparse(y,y,y);
elapsed time: 0.043261051177978516 seconds

[ViralBShah]

I am guessing this is because the codegen is optimizing A[i] for vectors, but there is no similar rule for Range.

[ViralBShah]

This is even more surprising, since range inputs are expanded anyways in _jl_make_sparse

[ViralBShah]

This is still the case even with the new implementation of sparse().

julia> x = 1000000:-1:1
1000000:-1:1

julia> @time sparse(x,x,x);
elapsed time: 0.3297080993652344 seconds

julia> x = [1000000:-1:1];

julia> @time sparse(x,x,x);
elapsed time: 0.09220290184020996 seconds

[JeffBezanson]

Yes, ref is slower on ranges than on arrays, but iterating over them is just as fast.

[ViralBShah]

It is a performance pothole, as people will routinely pass stuff like 1:n around. Can we improve the performance of ref on ranges, or will this be something people have to watch out for?

-viral

On 19-Jun-2012, at 10:15 AM, Jeff Bezanson wrote:

Yes, ref is slower on ranges than on arrays, but iterating over them is just as fast.

---

Reply to this email directly or view it on GitHub:
#938 (comment)