Arithmetic Error while compiling R with the Intel compilers
On 02/12/2013 1:39 PM, Swapnil Gaikwad wrote:
Hi Team,
I downloaded the R 3.0.2 and, built and installed it
using Intel compilers (icc and ifort) from Intel parallel studio 2013,
sp1. After performing make tried to check it using 'make check'. Most of
the tests passed successfully but while checking a 'stats' package I saw
many lines were printed which did not match the desired output.
e.g.
There were two types of mismatch
Case 1)
6300c6300
< Grand Mean: 291.5937
---
> Grand Mean: 291.5938
Case 2) 12699c12699 < UrbanPop *0.278* -0.873 -0.378 0.134 ---
> UrbanPop *-0.278* -0.873 -0.378 0.134
Case 1 was having slight difference in the values and
assumed due to the floating point error, but Case 2 is having values
with difference signs. Is such behaviour is normal or there is any issue
with the Intel libraries.
You need to look at the context of what is being tested. In this particular case, you are testing principal component loadings: they can all be multiplied by -1 and give exactly the same results, so this one is okay. (The choice of sign is probably being made in the linear algebra routines.)
I used some of the libraries from Intel MKL
like fftw, mkl_lapack etc. I am pasting below the command that I used to
configure this R package.
./configure --prefix=$INSTALL_DIR --with-blas="-lmkl_intel_lp64
-lmkl_intel_thread -lmkl_lapack -lmkl_core -lpthread -lfftw3xf_intel
-lfftw3xc_intel" --with-lapack
I also tried installing using Intel compilers without using
Intel MKL libraries. It was done using below command.
./configure --prefix=$INSTALL_DIR
Exactly same errors were there.
There were no mismatch observed while compiling with gcc.Kindly
provide your suggestions whether it is good idea to use Intel compilers.
Thanks for your kind attention and time.
Can't help you with your final question, I have no experience with them. Duncan Murdoch