[R-pkg-devel] dtrti2 error on CRAN checks

[peter dalgaard]

Kurt Hornik submitted a bug report to the gfortran maintainers yesterday, and yes, it seems to have a lot to do with this.

He and Brian Ripley have been able to pinpoint it to certain gfortran revisions (specifically gfortran-9 at least r268992, or gfortran-8 at least r269349), but we're lacking a simple test case.

-pd

> On 18 Apr 2019, at 07:53 , Benjamin Christoffersen <boennecd using gmail.com> wrote:
> 
> Thanks for the quick reply. First a correction. I meant Ubuntu 18.04.2
> of course, sorry.
> 
>> In short, you need to look more closely.
> Is there a way to tell which BLAS and LAPACK is used on the Debian
> machines on CRAN? I checked the "CRAN Package Check Flavors" page but
> there is no information there regarding BLAS or LAPACK.
> 
> I have tried to test the package with Atlas, OpenBlas, and the
> reference implementation with gcc 7.3.0 and 8.2.0. It all worked
> without any errors.
> 
> It seems that gcc has changed since January from version 8.2.0 and
> 7.3.0 to 8.3.0 on the tests on CRAN. Further, it is now Debian version
> 8.3.0-2 instead of Debian 8.2.0-7 and Debian 7.3.0-29. I do not know
> whether this can matter.
> 
> Sincerely yours,
> Benjamin Christoffersen
> 
> 
> 
> 
> 
> 
> 
> Den ons. 17. apr. 2019 kl. 11.38 skrev Dirk Eddelbuettel <edd using debian.org>:
>> 
>> 
>> On 17 April 2019 at 11:06, Benjamin Christoffersen wrote:
>> | Since the start of April, I have gotten some errors on the Debian
>> | machines on the check on CRAN for my package dynamichazard. The places
>> | where I get the errors seems to come down to LAPACK's dtrtri routine.
>> | I have tried to reproduce the error on Debian 18.04.2 with clang 8.0.0
>> | but I cannot reproduce them.
>> |
>> | The errors came suddenly and the tests passed before. I see similar
>> | issues (I think) in the following package:
>> |  - RcppHMM
>> |  - BNPmix
>> |  - themetagenomics
>> |  - tidytext
>> |  - stm
>> |
>> | Any ideas what the error may be? Here is the relevant code parts for
>> | one of the failed tests:
>> | https://github.com/boennecd/dynamichazard/blob/c42105feea5e2d16a8d461b7c606245a743b3e0a/src/for_tests.cpp#L216
>> | https://github.com/boennecd/dynamichazard/blob/c42105feea5e2d16a8d461b7c606245a743b3e0a/src/BLAS_LAPACK/arma_BLAS_LAPACK.cpp#L26
>> | https://github.com/boennecd/dynamichazard/blob/c42105feea5e2d16a8d461b7c606245a743b3e0a/src/BLAS_LAPACK/R_BLAS_LAPACK.cpp#L67
>> 
>> There are four or more different sources of LAPACK and BLAS on Debian as we
>> (since the late 1990s !!)  utilize the nature of the _interface_ allowing
>> different packages to fill in.
>> 
>> So there could be
>>  i)   the R internal source with a partial library -- Fedora/CentOS use this
>>       but the Debian distro builds do not
>>  ii)  "reference BLAS", external, unoptimized
>>  iii) OpenBLAS, the successor to Goto, parallel via multithreading
>>  iv)  Atlas, "tuned" but not mulitthreading
>> and more as eg Intel MKL via a quick script (see my blog).
>> 
>> As a historical aside, and way-back-when, ii) earned us year's long growling
>> from the direction of Oxfordshire because some early/old versions of LAPACK
>> had bugs.  But it cuts both ways: the external versions tend to be complete
>> whereas what R ships with internally contains a subset -- and at least one
>> (early) user of RcppArmadillo was bitten when R built from sources using
>> defaults would not have the complex operations he needed. To R Core's credit
>> these missing functions got filled in over the years.
>> 
>> In short, you need to look more closely. On the Ubuntu 18.10 machine that I
>> type this on, sessionInfo()'s second paragraph has
>> 
>>  Matrix products: default
>>  BLAS: /usr/lib/x86_64-linux-gnu/openblas/libblas.so.3
>>  LAPACK: /usr/lib/x86_64-linux-gnu/libopenblasp-r0.3.3.so
>> 
>> which IIRC is also the default (via some Debian/Ubuntu internal 'ordering' of
>> the available alternatives).
>> 
>> Hth, Dirk
>> 
>> --
>> http://dirk.eddelbuettel.com | @eddelbuettel | edd using debian.org
> 
> ______________________________________________
> R-package-devel using r-project.org mailing list
> https://stat.ethz.ch/mailman/listinfo/r-package-devel

-- 
Peter Dalgaard, Professor,
Center for Statistics, Copenhagen Business School
Solbjerg Plads 3, 2000 Frederiksberg, Denmark
Phone: (+45)38153501
Office: A 4.23
Email: pd.mes using cbs.dk  Priv: PDalgd using gmail.com