x86_64, acml-3.5.0-gfortran64 and lme4
On Mon, 16 Oct 2006, Douglas Bates wrote:
I am encountering segfaults when checking the lme4 package on an Athlon64 system if I use the acml blas. R was built as a 64-bit application using the GCC 4.0.3 compiler suite including gfortran. The version of acml is 3.5.0 gfortran64.
gcc 4.1.1 is current, and I would suggest an update would be a good idea: my tests with gcc4 went a lot better with 4.1.0 than 4.0.3. I have an FC5 dual x86_64 system (hence gcc 4.1.1) with ACML 3.5.0 and Goto 1.06 (I cannot get ATLAS to build on that system). I too get a segfault with ACML (I used the _mp version, but also tried with it confined to a single thread), but not with Goto nor R's internal BLAS.
I do not encounter the segfaults when I compile R with R's built-in BLAS. The errors occur in the first example in lme4 in a call to lmer. It looks like they would occur in any call to lmer. Running under the debugger shows that the segfault occurs in a call to dtrsm (a level-3 BLAS routine to solve a triangular system of equations) that is called from within a cholmod (sparse matrix library) routine. Has anyone succeeded in running R CMD check on the lme4 package with accelerated BLAS? I'm trying to pin down is this occurs only with ACML or also with Atlas and/or Goto's BLAS.
______________________________________________ R-devel at r-project.org mailing list https://stat.ethz.ch/mailman/listinfo/r-devel
Brian D. Ripley, ripley at stats.ox.ac.uk Professor of Applied Statistics, http://www.stats.ox.ac.uk/~ripley/ University of Oxford, Tel: +44 1865 272861 (self) 1 South Parks Road, +44 1865 272866 (PA) Oxford OX1 3TG, UK Fax: +44 1865 272595