[Csdp] Puzzling result using icc

[Ronald Bruck]

s anybody used icc to compile csdp 6.0.1?  I'm using version 11, basically with the lines

CC=icc
CFLAGs=-O3 -m64 -align -Wall -ansi -openmp -DNOSHORTS -DUSEOPENMP
LIBS=-lm -lmkl -lgomp -lguide -lpthread 

in the makefiles, and everything compiles OK (except for some useless warnings).  But it just doesn't work the same.  (The behavior is the same if I remove the -openmp, BTW).

I have a lengthy file on which this version fails on the first iterate, claiming the system matrix is singular.  (Which it is.)  When I run the binary Brian supplies for the P4, it "solves" the SDP and shows very good numbers overall, although when you examine the solution file there are some BIG numbers.  I'm sure what's happening is there is a sup (or inf) to the problem, and they're equal, but they're not a max/min (i.e. they're not attained).

Curiously, the generic P4 code is about 3.5 times FASTER than the "optimized" icc code.  (It averages about 0.09 seconds vs. about 0.026 sec). That's BAD for a commercial compiler.  Am I doing something wrong?

When I use my GMP version of csdp, it converges and shows the same big numbers.  (csdpgmp is designed to overcome singularities in the problem.)  It takes 3.5 sec at 128 bits, but against that it reports primal and dual values of 
0.3292044793644460771e0, much better than the P4's 3.2920448e-01 and 3.2920449e-01.

This is with the same param.csdp file.  Is there something different about the P4 version of the code?  Has anybody else had such poor performance with icc?

-- Ron Bruck
sdp file attached--sorry, it's big.  I *know* you're not supposed to send attachments to mailing lists, but it takes a big file to trigger this behavior.

-------------- next part --------------
A non-text attachment was scrubbed...
Name: E.sdp
Type: application/vnd.stardivision.impress
Size: 21303 bytes
Desc: not available
Url : http://list.coin-or.org/pipermail/csdp/attachments/20090818/02e40d10/attachment-0001.bin