Thanks, that was very helpful. The problem was eliminated by parameterizing wrt the type S<:FESet. The lesson I learned was that knowing the abstract type is not enough: the compiler needs to know the concrete type. That is something I was not used to and it tripped me up again this time. I hope I will remember from now on.
Thanks again, Petr On Friday, January 9, 2015 at 12:08:58 PM UTC-8, Tim Holy wrote: > > On Friday, January 09, 2015 11:51:15 AM Petr Krysl wrote: > > The size() function is from the base. fes.conn is a 2D array of ints. > > No it's not, it's a JFIntMat :-). But seriously, _you_ might know that > corresponds to a 2D array of ints, but the key question is if you've given > enough information in your declaration of JFIntMat for the _compiler_ to > know > that. > > Bottom line is that the behavior you're observing isn't normal. It's also > extremely unlikely to be a bug in julia, because we do this kind of stuff > all > over julia without the kind of allocations you're observing. So it's > almost > certainly something in your code. @code_warntype is your best way of > finding > out. > > --Tim > > > > > P > > > > On Friday, January 9, 2015 at 11:25:34 AM UTC-8, Tim Holy wrote: > > > I wonder if your `size` function isn't returning an Int, or that the > > > compiler > > > can't infer that. Did you try this on julia 0.4 with @code_warntype? > > > > > > --Tim > > > > > > On Friday, January 09, 2015 08:06:43 AM Petr Krysl wrote: > > > > Actually, I realized that the example shows the same problem, but > for a > > > > different function. This time the allocation occurs in the function > > > > CALLING getconn!. > > > > > > > > P > > > > > > > > On Friday, January 9, 2015 at 8:03:00 AM UTC-8, Petr Krysl wrote: > > > > > Tim, > > > > > > > > > > Your explanation is very helpful. There is a complication though: > > > > > The function is actually called from another function (from a > chain > > > > > > of > > > > > > > > functions). Here is the measurement from a simplified situation: > > > > > - using JFFoundationModule > > > > > - using FESetModule > > > > > - > > > > > - n=100; > > > > > - function test(n) > > > > > 0 fes=FESetH8(conn=rand(JFInt,n, > > > > > > > > > > 8)) > > > > > > > > > > 112 conn1::JFIntMat=zeros(JFInt,nfense(fes),1); > > > > > > > > > > 6383704 for j=1:size(fes.conn,1) > > > > > > > > > > 0 getconn!(fes,conn1,j); > > > > > - end > > > > > 0 return true > > > > > - end > > > > > - > > > > > - test(n) > > > > > - clear_malloc_data() > > > > > - n=100000; > > > > > - test(n) > > > > > > > > > > As you can see the loop over the rows of the array fes.conn > allocates > > > > > substantial amount of memory. > > > > > > > > > > Here is the type: > > > > > https://gist.github.com/PetrKryslUCSD/794f521a8e5b057e5e4e > > > > > > > > > > Petr > >
