Hi people, Just to let you know that I have installed the new version of Julia (the one in Github), that has the new implementation of cumsum and the improvement in performance is absolutely amazing!
By performing @profile for my function using the former version of cumsum the resulting backtrace was 1921. For the new version of cumsum the backtrace was 139. One order of magnitude lower!! Thank you again for your huge effort in making Julia such a dynamic project! Long life to Julia! Best, Charles On Sat, Jun 21, 2014 at 1:24 AM, Charles Novaes de Santana < charles.sant...@gmail.com> wrote: > Thank you again, Dahua! I hope it can be changed easy. > > Best, > > Charles > > > On Sat, Jun 21, 2014 at 1:06 AM, Dahua Lin <linda...@gmail.com> wrote: > >> I have looked the codes of cumsum and friends. They are still using >> old-style slice-based implementation, instead of the new cache friendly >> ways (the way we are implementing reduction). >> >> Not sure how quickly these will be resolved. If this is not addressed in >> 2 - 3 weeks, I may take a shot to reimplement them. >> >> Dahua >> >> >> On Friday, June 20, 2014 5:53:44 PM UTC-5, Charles Santana wrote: >> >>> Hi again, >>> >>> Just to let you know about the issue I just opened in Github: >>> >>> https://github.com/JuliaLang/julia/issues/7342 >>> >>> Thank you for everything! >>> >>> Best, >>> >>> Charles >>> >>> >>> On Fri, Jun 20, 2014 at 10:07 PM, Charles Novaes de Santana < >>> charles...@gmail.com> wrote: >>> >>>> Thank you, Dahua! >>>> >>>> I will open an issue in Github as suggested by you. In meanwhile I will >>>> see if by using sum I can get a better performance. >>>> >>>> Best, >>>> >>>> Charles >>>> >>>> >>>> On Fri, Jun 20, 2014 at 5:54 PM, Dahua Lin <lind...@gmail.com> wrote: >>>> >>>>> The cumsum / cummax / cummin / cumprod, etc have suboptimal >>>>> performance currently, which are about 20x slower than the sum/prod etc >>>>> (which we spent a lot of efforts to optimize and tune). >>>>> >>>>> Please open an issue in Github, and we will try to address this >>>>> problem later. >>>>> >>>>> Dahua >>>>> >>>>> >>>>> On Friday, June 20, 2014 10:15:55 AM UTC-5, Charles Santana wrote: >>>>>> >>>>>> Dear Julia users, >>>>>> >>>>>> First of all, Congratulations for this amazing community and for this >>>>>> impressive language! I used to program in C++ and in R, I started to >>>>>> program with Julia 3 months ago and it has changed my life for better!! >>>>>> Thank you!! >>>>>> >>>>>> By checking the profile of a program we are developing we noted that >>>>>> the "bottleneck" seems to be in a cumulative sum along a dimension in a >>>>>> matrix, for what we use the function cumsum. >>>>>> >>>>>> We are doing something like this: >>>>>> >>>>>> DI = rand(5,5); >>>>>> Dc = cumsum(DI,2); >>>>>> >>>>>> Just to try to clarify what we are doing: Imagine that Matrix DI(i,j) >>>>>> represents the probability of an individual to move from a site i to a >>>>>> site >>>>>> j. We use Dc to determine to which site an individual in site i will >>>>>> move, >>>>>> by generating a random number between 0 and maximum(Dc[i,:]). That means, >>>>>> we are trying to perform a Multinomial Distribution. >>>>>> >>>>>> Do you know an alternative to cumsum or do you indicate a good way to >>>>>> use this function. >>>>>> >>>>>> Thanks in advance for any help! >>>>>> >>>>>> Best regards, >>>>>> >>>>>> Charles Novaes de Santana >>>>>> -- >>>>>> Um axé! :) >>>>>> >>>>>> -- >>>>>> Charles Novaes de Santana, PhD >>>>>> http://www.imedea.uib-csic.es/~charles >>>>>> >>>>> >>>> >>>> >>>> -- >>>> Um axé! :) >>>> >>>> -- >>>> Charles Novaes de Santana, PhD >>>> http://www.imedea.uib-csic.es/~charles >>>> >>> >>> >>> >>> -- >>> Um axé! :) >>> >>> -- >>> Charles Novaes de Santana, PhD >>> http://www.imedea.uib-csic.es/~charles >>> >> > > > -- > Um axé! :) > > -- > Charles Novaes de Santana, PhD > http://www.imedea.uib-csic.es/~charles > -- Um axé! :) -- Charles Novaes de Santana, PhD http://www.imedea.uib-csic.es/~charles
