On 14 Nov 2013, at 16:43, Roland Haas roland.haas@physics.gatech.edu wrote:
Signed PGP part Hello all,
I suggest that you run anyway. It's likely there will be failures (all intel-compiler-based systems are currently showing failures in GRHydro and other tests), so you may as well set up your tree and
Not all intel systems show the same behaviour. Intel 12/13 based systems seem to have few tests that fail and those that fail fail due to tolerances being possibly too large.
Intel 11 system on the other hand crash catastrophically, from what I can tell so far from the compiler not handling Cray pointers correctly (eg the Hydro_InitExcision gives a segfault). So far I have no good workaround for the issue. I will have to check if all usage of Cray pointers is affected or only the Reconstruction drivers (which is where the segfault happens) since they use array slices.
These are not the usual GRHydro failures unfortunately but instead a compiler bug. A workaround that does not fail on other compilers is required (one workaround is eg to set the size of the leftmost dimension of the pointee in question to zero, it almost seems as if the compiler actually allocates space for the pointees).
If there is a workaround, that is of course good, but another alternative would be to say that Intel 11 is unsupported with the new features in grhydro, and to update the optionlists to a newer version where cray pointers work. On some machines, this won't be possible, e.g. tezpur, where intel11 is the latest version available.
In any case, an autoconf test, or a runtime test, that cray pointers are working correctly, would be very nice to have.