Build fails due to missing Eigen3

[susilehtola]

The addition of libint2 causes Psi4 to not build. Libint2 is trying to compile the hartree-fock++ test, which depends on Eigen3, which it is not finding for some reason even though I have it installed as a system package.

Libint2 tests should be disabled in the Psi4 build, or the Eigen3 dependency added so that the build works even in the lack of a system package.

(Note: libint2 is also suffering from the -march issue; the libint2 CMake files need to be updated.)

[loriab]

The addition of libint2 causes Psi4 to not build. Libint2 is trying to compile the hartree-fock++ test, which depends on Eigen3, which it is not finding for some reason even though I have it installed as a system package.

Are you using the latest psi4-dev package (.dev22)? That adds the mpfr and eigen deps to the build environment. I definitely acknowledge that the Libint2Config.cmake is not handling find_dependency right for eigen and mpfr (really, gmp) yet. I do have cmake edits for those locally, but I have not yet bent cmake to my will. So having the conda packages in place patches the problem. Does your system package have a Eigen3Config.cmake? Maybe expand CMAKE_MODULE_PATH?

Libint2 tests should be disabled in the Psi4 build, or the Eigen3 dependency added so that the build works even in the lack of a system package.

All of Psi4 master is now running Libint2 -- any references to Libint1 or Libint1_t (a hack to allow both to coexist) are only waiting for removal.

(Note: libint2 is also suffering from the -march issue; the libint2 CMake files need to be updated.)

Thanks for the reminder. I've just edited it locally.

[susilehtola]

I am using the latest git snapshot of the master branch.

[susilehtola]

And yes, the system package comes with CMake files:

/usr/share/cmake/Modules/FindEigen3.cmake
/usr/share/eigen3/Eigen3Config.cmake
/usr/share/eigen3/Eigen3ConfigVersion.cmake
/usr/share/eigen3/Eigen3Targets.cmake
/usr/share/eigen3/UseEigen3.cmake

[loriab]

I am using the latest git snapshot of the master branch.

gtk. this is what I was talking about. It hides dependency tracking from many psi builders and seeds cmake configuration. I was hoping that could carry most builds until FindMPFR called by Libint2Config and I come to terms. Are you building libint2 through psi or using the libint2 conda-package or detecting other pre-built (will need to silence find_apckage(... components) in that case)? For now, I'd add find_package(eigen) and hack cmake_module_path as needed. And this should stay open until I get it patched correctly upstream.

ambit >=0.5.1,<0.6.0a0, chemps2 >=1.8.9,<1.8.10.0a0, cmake >=3.8, dkh >=1.2,<1.3.0a0, eigen, gau2grid >=2.0.3,<3.0a0, gcc_linux-64 7.*, gdma >=2.2.6,<2.2.7.0a0, gfortran_linux-64 7.*, gxx_linux-64 7.*, hdf5 >=1.10.6,<1.10.7.0a0, libint2 2.6.0 hc9558a2_9, libxc >=4.3.4,<5.0a0, mkl-devel 2019.*, mpfr, msgpack-python, networkx >=2.4, numpy >=0.12.0, pcmsolver >=1.2.1,<1.2.2.0a0, pybind11-headers >=2.5.0,<2.6.0a0, pytest, pytest-xdist, python >=3.8,<3.9.0a0, qcelemental >=0.17.0,<1.0, qcengine >=0.17.0,<1.0, simint >=0.7,<0.8.0a0

[susilehtola]

I'm building Libint2 via Psi4; I wouldn't have this error otherwise ;)

I'm waiting on Ed to release a stable updated version of Libint2, so that I can package it in Fedora.

[loriab]

I'm building Libint2 via Psi4; I wouldn't have this error otherwise ;)

The hartree-fock++ test case won't happen, but the eigen3 detection can still be a problem even with pre-built because the l2 c++ interface needs it to compile p4 against.

I'm waiting on Ed to release a stable updated version of Libint2, so that I can package it in Fedora.

I've seen some clues that he's waiting on this. I need to get the mpfr and eigen3 detection fixed before I can start ready-to-merge negotiations.

Also, I don't know if EFV is planning the sph-ordering-as-runtime-option for the 2.7 release so that psi can use the Fedora distribution.

[JonathonMisiewicz]

Moved to 1.5 per discussion on conference call. It's not worth the further delay to the 1.4 release, given that most users will get Psi from conda.

[loriab]

The switch to 2022 CMake fork of Libint2 in #2413 should have healed this issue, as it has better dependency handling and demands fewer dependencies. Try it out when you get a chance, please. You may find this dependency table helpful: https://github.com/evaleev/libint/blob/9f12ee61e1ce52420fe3020712c3584cb3e9a1b4/INSTALL.md#prerequisites

[loriab]

#2413 should have fixed this. please reopen if needed.