25 Oct
2008
25 Oct
'08
3:25 a.m.
Hm... sorry, this doesn't ring any bells. I've used fftw on 64-bit architectures (fftw3tbx module in cctbx) without a problem. We compile everything with -fno-strict-aliasing. In the past I tried to cut corners by using -fno-strict-aliasing only for some translation units. When I switched to compiling everything with this option I didn't notice an impact on the runtime. If it hangs in the pow() function, maybe there is a floating-point problem? Unfortunately, gcc's default "non-stop" handling of floating-point exceptions is useless IMO. We install proper traps if possible. When we started doing this, we found a whole bunch of numerical problems that went unnoticed before. (The trap stuff is in boost_adaptbx/meta_ext.cpp.) Ralf