CMake Error: string sub-command REPLACE requires at least four arguments

03/02/2021
CMake was trying to gather a list of Fortran source files for the libtinker.a static library but failed to locate and/or parse the Makefile inside the tinker9/tinker git submodule. The most likely reason is, the second installation step "Build the Canonical Tinker" was incomplete.

Reference: #138

CMake Error: undefined reference to __boxes_MOD_box34 or other _MOD_ variables

03/18/2021
This usually happened when the Fortran Tinker library libtinker.a was not correctly compiled. Possible reasons are:

• The tinker9/tinker submodule is outdated. Try to synchronize with the required Fortran Tinker version.
• Tinker9 was untarred from a zip file. You need to clone the required version from GitHub:TinkerTools/tinker to tinker9/tinker.

Reference: https://github.com/TinkerTools/tinker9/blob/master/doc/manual/m/install/tinker.rst

CMake Error: The C++ compiler "nvc++" is not able to compile a simple test program

04/04/2021

Check the text file bin/localrc in the OpenACC installation. This file is likely to show that the ACC compiler is configured to another/older GCC version (possibly 4.8.5) than what you are currently using. The file localrc was generated when the ACC compilers were first installed. If a newer GCC version had not been made default, the ACC compilers would be configured to the default/older GCC version. When you load/switch to a newer GCC version, compatibility issues may show up.

There are four possible fixes:

1. After deleting the current OpenACC installation, you switch to the newer GCC version (gcc, g++, gfortran) and reinstall OpenACC. This will require root privilege.
2. Switch to the newer GCC version (gcc, g++, gfortran) and run makelocalrc (which is under $(dirname $(which nvc++))) with root privilege.
3. Contact your system admin to do 1 or 2 for you, and make a module that bundles the consistent GCC version therein, so you can load them altogether in the future.
4. Create another localrc file in your home directory and source it in your shell resource file. The details are included in the link below.

To "re-register" the default GCC version for OpenACC compilers, please refer to:

More About Using PGI Compiler on the Clusters (Installation Step 1)

/usr/bin/ld: Cannot find -lcufft -lcublas -lcublasLt

04/04/2021

The short answer is: check your $PATH, your canonical cuda/bin directory is likely after NVHPC. If so, move it in front of NVHPC.

The full picture is: NVHPC has a different directory structure than the canonical CUDA installation and the old PGI installation, thus some of the required CUDA runtime libraries expected in cuda/targets and cuda/lib64 cannot be found in the same relative path within the NVHPC installation.

This problem should be solved once we optimize our CMake files. (We have to wait until new CMake fully supports NVHPC.)