QuLogic
commented
10 years ago This likely comes from intrinsic functions such as size() that return and integer8 instead of an integer4, thus leading to incorrect results when used in function calls is the -i8 flag is not added to the compiler options; however, when adding -i8 the code does not compile because the MPI calls then refuse to compile (they need integer4 as arguments).
This explanation doesn't really make sense to me. The size intrinsic should return an integer of the default kind, and changing the memory model shouldn't affect that?
Is there a test case for this?
komatits
mpbl
In SPECFEM3D_GLOBE, fix this known 64-bit issue when -mcmodel=medium -shared-intel is used in order to handle very large static memory size per core (the 3D_Cartesian and 2D packages are fine because that compiler option is never used in them):
If you run very large meshes on a relatively small number of processors, the static memory size needed on each processor might become greater than 2 gigabytes, which is the upper limit for 32-bit addressing (dynamic memory allocation is always OK, even beyond the 2 GB limit; only static memory has a problem). In this case, on some compilers you may need to add -mcmodel=medium (if you do not use the Intel ifort / icc compiler) or -mcmodel=medium -shared-intel (if you use the Intel ifort / icc compiler) to the configure options of CFLAGS, FCFLAGS and LDFLAGS otherwise the compiler will display an error message (for instance 'relocation truncated to fit: R_X86_64_PC32 against .bss' or something similar);
BEWARE that using -mcmodel=medium -shared-intel is known for currently leading to incorrect seismograms, at least when flag ATTENUATION is on (and maybe even without), at least in the case of the Intel ifort / icc compiler. This likely comes from intrinsic functions such as size() that return and integer8 instead of an integer4, thus leading to incorrect results when used in function calls is the -i8 flag is not added to the compiler options; however, when adding -i8 the code does not compile because the MPI calls then refuse to compile (they need integer4 as arguments). It seems that this problem is triggered when 3D models are used and mostly (only?) when attenuation is on, but not triggered for 1D Earth model benchmarks even with attenuation (at least not triggered for EXAMPLES/benchmarks/attenuation_benchmark_GJI_2002_versus_normal_modes nor for Anne Sieminski's three 1D benchmarks), which makes it uneasy to locate. Since most current users run the code with less than 2 GB of static memory per core, we have not investigated that problem carefully for now, and thus recommend that you do NOT use these compiler flags until this problem is located and fixed.
(todo_list_please_dont_remove.txt: suggestion 45)