On May 25, 2010, at 6:58 , Barry Wardell wrote:
Hi,
I have been trying to compile the Einstein Toollkit on Kraken and have encountered a problem with the AHFinder thorn. I looks to me like a compiler problem. The error message I get is:
Preprocessing /nics/a/proj/cactus/barrywar/xt5/Cactus/arrangements/ EinsteinAnalysis/AHFinder/src/AHFinder_pow.F Compiling /nics/a/proj/cactus/barrywar/xt5/Cactus/arrangements/ EinsteinAnalysis/AHFinder/src/AHFinder_pow.F *** glibc detected *** /opt/pgi/9.0.4/linux86-64/9.0-4/bin/pgf902: realloc(): invalid next size: 0x00000000009f2e80 ***
Yes, this is a compiler problem. The usual remedies are (in this order): 0. Delete configuration, build again 1. use a different compiler version 2. use a different compiler 3. change optimisation flags 4. blindly rearrange source code a bit to see whether this changes anything 5. remove thorn from thorn list 6. use a different machine
Round 0: Did you try this? Sometimes there is a transmission error in a file, or some other transient error that goes away if you try again.
Round 1: You could try using PGI 10.3, but when I tried this, I found more compiler problems than with 9.0.4. I'm attaching my option list for 10.3 anyway. I don't expect PGI to be interested in bug reports against 9.0.4 that can't be reproduced with 10.3 any more. PGI 10.4 should be available this month -- apparently PGI/Cray have a monthly release cycle for compiler versions.
Round 2: GCC on Kraken creates much slower code than PGI. I'm not too fond of the PathScale compilers.
Round 3: I was fighting with PGI optimisation options yesterday, and my new preferred optimisation options are now
-O3 -Munroll=c:1 -Mscalarsse -Mcache_align -Mflushz
which leave out some of the options implied by -fastsse as well as the -Mvect=assoc. Do you want to try these?
-erik