Compiling VASP With Gfortran

For some time, VASP has been centered on the x86 processors and Intel’s Fortran compiler. Inside the VASP source distribution, you can find some makefiles for other compilers, but they seldom work nowadays, and in many cases you need to make modifications to the source code to make it work with other compilers.

In particular, recent versions of Gfortran cannot compile VASP. If you try, the compiler would stop at errors concerning a circular dependency of a module (i.e. the module includes itself), and some output formatting errors.

From what I can understand, these problems are actually related to violations of the Fortran language standard, which are allowed by the Intel compiler. There are no compiler flags for gfortran that let you “relax” the standard like this to let it compile VASP, so you need to modify the source to make it compliant.

When I tested with gcc 4.7.2 and gcc 4.8.0, four files needed to be modified: us.F, vdwforcefield.F, finite_diff.F, and spinsym.F. I have prepared the patches as a “patch file” which you can download. To apply the patches to the source code, locate your VASP 5.3.3 source code directory and do

cd vasp.5.3
patch -p0 < vasp533gcc.patch

In the makefile, you need to set the following compiler flags for gfortran.

FC = mpif90 (or similar depending on the MPI)
FFLAGS = -ffree-form -ffree-line-length-0  -fno-second-underscore
OFLAG=-O3 -march=corei7-avx -mtune=corei7-avx

Global -O3 optimization seems to work for me on Triolith (Xeon E5 processors), but I haven’t tested all functionality of the gfortran version yet. As with the Intel compiler, you may have to decrease the optimization or disable aggressive inlining in certain files.

In the preprocessor section, put something like this. Note that you should not use the -DPGF90 flag when compiling with gfortran.

    -DMPI_BLOCK=262144 \
    -Duse_collective -DCACHE_SIZE=12000 -Davoidalloc -DNGZhalf\

These tricks made it for me, and I now have a reference version of VASP compiled with Gfortran on Triolith. The speed seems to be about same as when compiled with Intel Fortran, since VASP relies heavily on FFTWs and BLAS calls and I still link with MKL and Intel’s MPI.

Later, I will try to make a longer guide how to compile VASP with a fully free software stack, and compare performance and stability.