On Fri, Sep 2, 2016 at 7:41 AM, Mauro <[email protected]> wrote: > On Fri, 2016-09-02 at 13:34, Jong Wook Kim <[email protected]> wrote: > > Hi Yichao, what a nice idea :) > > > > But even if I write in the C++ way, @time sqrt(1) yields 5 allocations > of 176 > > bytes, and in inner loops this could be a bottleneck. >
It's the allocation due to doing things in the global scope. try `f() = @time sqrt(1); f()` > > Those are just allocations for the return value of sqrt. Consider: > > julia> function f(n) > out = 0.0 > for i=1:n > out += sin(n) > end > out > end > f (generic function with 1 method) > > julia> @time f(10) # warmup > 0.000008 seconds (149 allocations: 10.167 KB) > -5.440211108893696 > > julia> @time f(10) > 0.000005 seconds (5 allocations: 176 bytes) > -5.440211108893696 > > julia> @time f(10000) > 0.000849 seconds (5 allocations: 176 bytes) > -3056.143888882987 > > > > Is this an inevitable overhead of using ccall, or is it just a bogus > that I can > > ignore? > > > > Jong Wook > > > > > > On Sep 2, 2016, at 7:14 AM, Yichao Yu <[email protected]> wrote: > > > > > > > > On Fri, Sep 2, 2016 at 7:03 AM, Jong Wook Kim <[email protected]> > wrote: > > > > Hi, > > > > I'm using Julia 0.4.6 on OSX El Capitan, and was trying to > normalize > > each column of matrix, so that the norm of each column becomes > 1. Below > > is a isolated and simplified version of what I'm doing: > > > > function foo1() > > local a = rand(1000, 10000) > > @time for i in 1:size(a, 2) > > a[:, i] /= norm(a[:, i]) > > end > > end > > > > foo1() > > 0.165662 seconds (117.44 k allocations: 232.505 MB, 37.08% gc > time) > > > > I thought maybe the array copying is the problem, but this > didn't help > > much: > > > > function foo2() > > local a = rand(1000, 10000) > > @time for i in 1:size(a, 2) > > a[:, i] /= norm(slice(a, :, i)) > > end > > end > > > > foo2() > > 0.131377 seconds (98.47 k allocations: 155.921 MB, 36.66% gc > time) > > > > and then I figured that this ugly one runs the fastest: > > > > function foo3() > > local a = rand(1000, 10000) > > @time for i in 1:size(a, 2) > > setindex!(a, norm(slice(a, :, i)), :, i) > > end > > end > > > > foo3() > > 0.013814 seconds (49.49 k allocations: 1.365 MB, 4.86% gc time) > > > > So I overheard a few times that plain for-loops are faster than > > vectorized code in Julia, and it seems it's allocating slightly > less > > memory, but it's slower than the above. > > > > function foo4() > > local a = rand(1000, 10000) > > @time @inbounds for i in 1:size(a, 2) > > n = norm(slice(a, :, i)) > > @inbounds for j in 1:size(a, 1) > > a[j, i] /= n > > end > > end > > end > > > > foo4() > > 0.055522 seconds (30.00 k allocations: 1.068 MB, 15.14% gc time) > > > > Is there a solution that is faster and less uglier than foo3() > and foo4 > > ()? > > > > Thinking of an equivalent implementation in C/C++, I should be > able to > > write this logic without any heap allocation. Is it possible in > Julia? > > > > > > You can write it in the way you'd write it in c++ and just don't use > `norm > > `. > > > > > > > > Thanks, > > Jong Wook >
