[petsc-users] problem with installation using quad precision

[Jed Brown]

Santiago Andres Triana <repepo at gmail.com> writes:

> Dear petsc-users,
>
> I am trying to install petsc with the option --with-precision=__float128.
> The ./configure goes fine, as well as the make all stage. However, the make
> check step to test the libraries fails with the following error:
>
> /usr/bin/ld: home/spin/petsc-3.9.3/arch-linux2-c-opt/lib/libpetsc.so:
> undefined reference to `dgemv_'

Please send complete configure.log and make.log.  Possibly fighting
between f2cblaslapack and Fortran kernels.

>
> this is the configure command I use:
>
> ./configure --with-scalar-type=complex --with-precision=__float128
> --with-debugging=0 --with-fortran-kernels=1 COPTFLAGS='-O3 -march=native
> -mtune=native' CXXOPTFLAGS='-O3 -march=native -mtune=native' FOPTFLAGS='-O3
> -march=native -mtune=native' --download-f2cblaslapack

Note that -mtune=native above is redundant -- you only need to specify
it explicitly if you want to tune to a more restrictive class of
processors.  For example

  -march=sandybridge -mtune=skylake

would not use AVX2/FMA instructions, but would optimize code on the
assumption of an extra load/store unit (present in Haswell and Skylake;
may mean more cycles waiting on load/store when run on Sandy Bridge).