ifc gives wrong result in matrix analysis

[gijs]

Hi,
my problem is somewhat hard to pin down, since I notice it only on a fairly high level in the code. In the code I am developing, a number of matrix operations are performed using the Numerical Recipes routines. ifc gives erroneous and internally inconsistent results for singular value decomposition and for Cholesky decomposition of my positive-definite symmetric matrix. The operations occasionally do involve the evaluation of differences between nearly equal numbers, but nowhere do these get close to machine precision.
I compared the test results with those obtained with the Portland Group Fortran compiler (which is twice as slow; hats off for to you for speed!). This compiler does produce the correct result from the same code.
I would be happy to supply more details of the problem if anyone has a clue where to look.
Thanks,
Gijs

[gijs]

Since last week, I did a little additional research. I found that the actual values returned by my program compiled by ifc *change* according to whether I get external routines from a linked library, are compiled along with the other source files, or are included as internal subroutines. Again, I tested every version against our Portland compiler, which gives the same result every time.
My Linux is v2.4.9-13, and the machine is an SGI Pentium III 800 MHz. Compiler flags are -O0 -mp (though changing these has no effect on the results; haven't tried the -x? flags). Same results for compiling from scratch and running on a similar 900 MHz system. Still stymied.
Thanks for any help.

[Martyn_C_Intel]

The only circumstances I have come across which might give the sort of behavior you describe involved changing the precision by modification of the floating point control word at runtime, leading to inconsistencies with the math library.

Unless this helps, please submit an issue, with a test case if at all possible, to Intel Premier Support at htpps:premier.intel.com .

Martyn

[gijs]

Thanks.
I found a possible cause of the problem: there is a difference in the way the compilers handle array references: an array A(:) which is indexed A(2:1) is handled as a zero length array by the Portland compiler (the desired behaviour in my case), but gives rise to a memory leak with the Intel compiler. Note that this sloppy programming is NOT flagged an error to by compiling with -CB! I will submit an issue to Premier Support with a sample program.
Gijs

[tburns671]

I am having similiar problems and am not sure why. Any updates?