[R] precision problems in testing with Intel compilers

[Samuelson, Frank*]

I compiled the 1.9.1 src.rpm with the standard gnu tools and it works.
I tried compiling the 1.9.1 src.rpm with the Intel 8 C and FORTRAN
compilers and it bombs out during the testing phase:

    comparing 'd-p-q-r-tests.Rout' to './d-p-q-r-tests.Rout.save' ...267c267
    < df = 0.5[1] "Mean relative  difference: 5.001647e-10"
    ---
    > df = 0.5[1] TRUE
    make[3]: *** [d-p-q-r-tests.Rout] Error 1
    make[3]: Leaving directory `/usr/src/redhat/BUILD/R-1.9.1/tests'
    make[2]: *** [test-Specific] Error 2
    make[2]: Leaving directory `/usr/src/redhat/BUILD/R-1.9.1/tests'
    make[1]: *** [test-all-basics] Error 1
    make[1]: Leaving directory `/usr/src/redhat/BUILD/R-1.9.1/tests'
    make: *** [check-all] Error 2
    error: Bad exit status from /var/tmp/rpm-tmp.63044 (%build)

looking at the differences between the failed file and the standard, I get:

    fws wolf tests] diff  d-p-q-r-tests.Rout.save d-p-q-r-tests.Rout.fail 
    3c3
    < Version 1.9.0 Patched (2004-04-19), ISBN 3-900051-00-3
    ---
    > Version 1.9.1  (2004-06-21), ISBN 3-900051-00-3
    281c281
    < df = 0.5[1] TRUE
    ---
    > df = 0.5[1] "Mean relative  difference: 5.001647e-10"
    935c935
    < Time elapsed:  7.83 0.04 16.1 0 0 
    ---
    > Time elapsed:  2.49 0.01 2.55 0 0 

Besides being 3 times faster, it's stopping on the following code:

   for(df in c(0.1, 0.5, 1.5, 4.7, 10, 20,50,100)) {
     cat("df =", formatC(df, wid=3))
     xx <- c(10^-(5:1), .9, 1.2, df + c(3,7,20,30,35,38))
     pp <- pchisq(xx, df=df, ncp = 1) #print(pp)
     dtol <- 1e-12 *(if(2 < df && df <= 50) 64 else if(df > 50) 20000 else
500)
     print(all.equal(xx, qchisq(pp, df=df, ncp=1), tol = dtol))# TRUE
     ##or print(mapply(rErr, xx, qchisq(pp, df=df,ncp=1)), digits = 3)
   }

Where dtol used by all.equal is set to be 5e-10,
which the intel compiler misses by 1.6e-13.
This tolerance value seems a bit arbitrary.
The gcc compiled version's passes the test with a 9.3e-11 error.


I am using the -mp option for the intel compilers, which was recommended
on this mailing list previously and would make sense given the docs:

   Floating Point Optimization Options
       -mp    Maintain floating-point  precision  (disables  some
              optimizations).  The -mp option restricts optimiza-
              tion to maintain declared precision and  to  ensure
              that   floating-point   arithmetic   conforms  more
              closely to the ANSI and IEEE  standards.  For  most
              programs,  specifying this option adversely affects
              performance. If  you  are  not  sure  whether  your
              application  needs  this  option, try compiling and
              running your program both with and  without  it  to
              evaluate the effects on both performance and preci-
              sion.


Has anyone else encountered this?  

-Frank