[Rcpp-devel] Overriding 'ARMA_64BIT_WORD 1' in RcppArmadillo
George Vega Yon
g.vegayon at gmail.com
Tue May 3 19:17:47 CEST 2016
I think I've found a solution for this. I've submitted a pull request here
https://github.com/RcppCore/RcppArmadillo/pull/88. It basically now allows
the users to pass -DARMA_64BIT_WORD and compile with it (so ARMA_32BIT_WORD
is not defined but it is still the default behavior of RcppArmadillo). Will
work on tests on this repo https://github.com/gvegayon/arma64bit.
Best,
George G. Vega Yon
+1 (626) 381 8171
http://www.its.caltech.edu/~gvegayon/
On Thu, Apr 28, 2016 at 6:07 PM, George Vega Yon <g.vegayon at gmail.com>
wrote:
> Hey!
>
> Since it seems to be the case that we don't know much about what happens,
> I'm making an experiment which consists on an R package to stress
> RcppArmadillo here: https://github.com/gvegayon/arma64bit. Following
> Dirk's advice, I've included the #define ARMA_64BIT_WORD before #include
> <RcppArmadillo.h> (as seen here
> https://github.com/gvegayon/arma64bit/blob/c10b5d4a80b93088da730e45ee8c583feeff2ab6/src/misc.cpp#L2-L3)
> and included CXX_STD = CXX11 in Makevars. After compiling the package and
> asking for a matrix of size 1e5 x 1e5 I get an error, here (so it doesn't
> seem to be working):
>
> > x <- sp_runif(1e5,1e5,.01)
>
> error: SpMat::init(): requested size is too large
> Error: SpMat::init(): requested size is too large
>
> My guess is that #define ARMA_32BIT_WORD is overriding #define
> ARMA_64BIT_WORD. I'm an R user that jumped to Rcpp a while ago so I still
> don't fully understand how things work under the hood with R + Rcpp. Any
> clues?
>
> Best,
>
> sessionInfo()
>
> R version 3.2.5 (2016-04-14)
> Platform: x86_64-pc-linux-gnu (64-bit)
> Running under: Ubuntu 14.04.4 LTS
>
> locale:
> [1] LC_CTYPE=en_US.UTF-8 LC_NUMERIC=C
> LC_TIME=en_US.UTF-8
> [4] LC_COLLATE=en_US.UTF-8 LC_MONETARY=en_US.UTF-8
> LC_MESSAGES=en_US.UTF-8
> [7] LC_PAPER=en_US.UTF-8 LC_NAME=C LC_ADDRESS=C
>
> [10] LC_TELEPHONE=C LC_MEASUREMENT=en_US.UTF-8
> LC_IDENTIFICATION=C
>
> attached base packages:
> [1] stats graphics grDevices utils datasets methods base
>
> other attached packages:
> [1] arma64bit_0.1.0.9000 Matrix_1.2-5
>
> loaded via a namespace (and not attached):
> [1] Rcpp_0.12.4 lattice_0.20-33 digest_0.6.9 withr_1.0.1
> grid_3.2.5
> [6] covr_2.0.1 git2r_0.14.0 magrittr_1.5 lazyeval_0.1.10
> rstudioapi_0.5
> [11] whisker_0.3-2 rex_1.1.1 rmarkdown_0.9.5 devtools_1.11.1
> tools_3.2.5
> [16] yaml_2.1.13 htmltools_0.3.5 memoise_1.0.0
>
> George G. Vega Yon
> +1 (626) 381 8171
> http://www.its.caltech.edu/~gvegayon/
>
> On Thu, Mar 31, 2016 at 12:37 PM, Dirk Eddelbuettel <edd at debian.org>
> wrote:
>
>>
>> On 29 March 2016 at 15:55, George Vega Yon wrote:
>> | Hi there,
>> |
>> | A couple of days ago I included this comment on a Stackoverflow
>> question about
>> | "too large" matrices in RcppArmadillo
>> | I wonder if we can use this in a package. After trying to compile the
>> package
>> | with and without PKG_CXXFLAGS=-DARMA_64BIT_WORD I'm still not able to
>> create
>> | sparse matrices bigger than 65,535 x 65,535. I've realized that the
>> | RcppArmadilloConfig.h (version 0.6.600.4.0) file has ARMA_32BIT_WORD 1
>> so it
>> | actually prevents using ARMA_64BIT_WORD (I think?). Any way I can solve
>> this,
>> | or I just have to deal with R (3.2.4 Revised (2016-03-16 r70336)) not
>> been able
>> | to handle 64INT?
>> |
>> | http://stackoverflow.com/questions/16159174/
>> | large-spmat-object-with-rcpparmadillo
>> |
>> | For which Dirk answered:
>> |
>> | Try locally setting ARMA_64BIT_WORD (check how) -- I can't at the
>> package level
>> | as it breaks existing code. It will likely help you. If you set it
>> before the #
>> | include <RcppArmadillo.h> things may just work. That at least is the
>> pln. (And
>> | you should have asked on rcpp-devel to get more than two eyeballs on
>> this...)
>> | And yes the other answer basically says the same thing. – Dirk
>> Eddelbuettel 2
>> | days ago
>> |
>> | Now I'm asking again but on rcpp-devel =). The issue is that I'm
>> concerned on
>> | the part that Dirk says that "things may just work". I was reading
>> about R not
>> | been able to handle int64 and its implications for big matrices, and so
>> I'm a
>> | little concerned that if passing ARMA_64BIT_WORD works I may loose some
>> | information on the way. For example, if I allow a users creating and
>> | manipulating 1e6 x 1e6 sparse matrices, is there any chance for them to
>> loose
>> | information?
>>
>> It is entirely possibly that nobody stresses RcppArmadillo this way and
>> that
>> setting ARMA_64BIT_WORD gets unset. I only know that I cannot set it
>> globally now as some packages (still) depend on ARMA_32BIT_WORD.
>>
>> So I encourage you to experiment and to try locally. Maybe something can
>> (and even should) be improved.
>>
>> Dirk
>>
>>
>> |
>> | Thanks,
>> |
>> | George G. Vega Yon
>> | +1 (626) 381 8171
>> | http://www.its.caltech.edu/~gvegayon/
>> | _______________________________________________
>> | Rcpp-devel mailing list
>> | Rcpp-devel at lists.r-forge.r-project.org
>> | https://lists.r-forge.r-project.org/cgi-bin/mailman/listinfo/rcpp-devel
>>
>> --
>> http://dirk.eddelbuettel.com | @eddelbuettel | edd at debian.org
>>
>
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.r-forge.r-project.org/pipermail/rcpp-devel/attachments/20160503/c083538e/attachment.html>
More information about the Rcpp-devel
mailing list