On Thu, Feb 25, 2016 at 9:33 AM, abc <[email protected]> wrote: > Ah, yes, this seems to have fixed the problem, thank you. > > Now, when not in global scope, using eachindex is definitely the fastest > approach. > I made some measurements over 1000 runs for each of the approaches in the > original post, here are the averages: > Using eachindex: 0.0026683 > Using ranges: 0.0041256 > Using in: 0.0031200 > > I guess I'll stick with eachindex. :) > > On Thursday, February 25, 2016 at 1:50:04 PM UTC, Stefan Karpinski wrote: >> >> Can you try it not in global scope? >> >> On Thu, Feb 25, 2016 at 5:17 AM, abc <[email protected]> wrote: >>> >>> For the following matrix >>> my_matrix = randn(10000, 1000) >>> using eachindex to access all elements is much slower than using ranges >>> or even for el in my_matrix, even though it says in the documentation >>> (http://docs.julialang.org/en/release-0.4/stdlib/arrays/#Base.eachindex) >>> that eachindex uses ranges for Arrays. >>> >>> Some code and numbers: >>> julia> sum = 0.0 >>> julia> @time for iter in eachindex(my_matrix) >>> sum += my_matrix[iter] >>> end >>> 1.288944 seconds (50.00 M allocations: 915.519 MB, 3.36% gc time) >>> julia> sum = 0.0 >>> julia> @time for i in 1:10000, j in 1:1000 >>> sum += my_matrix[i,j] >>> end >>> 0.681678 seconds (34.38 M allocations: 524.582 MB, 2.45% gc time)
^^ I believe this one is also iterating in the wrong direction. Try `for j in ..., for i in ...` >>> julia> sum = 0.0 >>> julia> @time for el in my_matrix >>> sum += el >>> end >>> 1.063564 seconds (40.00 M allocations: 762.993 MB, 3.41% gc time) >>> >>> Am I reading the documentation wrong, or is there something strange with >>> the matrix indexing? Because as it is, I don't see any benefit of using >>> anything different than simple ranges for manipulating matrices. >> >> >
