[petsc-users] error on Spock
bsmith at petsc.dev
Tue Jul 13 21:09:15 CDT 2021
The expected behavior on Cray systems is you load "appropriate" modules and then run ./configure without needing to provide compiler and MPI information and it "just works".
So one should not need to provide --with-cc=cc --with-cxx=CC --with-fc=ftn on Cray systems. Loading "appropriate" modules is suppose to define the compilers (and MPI) you want to use so on should not need to be passed manually to PETSc's configure this kind of information.
The --HIPPPFLAGS=-I/opt/cray/pe/mpich/8.1.4/ofi/crayclang/9.1include is horrific. needing to have pass MPI information to the HIP compiler likely means PETSc does not have a proper separation of HIP code from "plain old C code" that is mistakenly put in hip files. This is definitely currently true with CUDA and likely carried over to the HIP interfaces. (.i.e. most of the functions in the .cu files in PETSc should just be in .c files).
The various rocm information should be handled automatically by ./configure and not be required to be provided by users. As it is currently handled for the CUDA libraries such as cubBLAS and cuSparse.
> On Jul 13, 2021, at 4:47 PM, Mark Adams <mfadams at lbl.gov> wrote:
More information about the petsc-users