Summit ppc64 build fails on develop branch with gcc 10.2

[james-simone]

The ppc64 gcc 10.2 compiler is not happy with Eigen v3.4.0. I expect this has been an issue since commit b640c1f44b8 "bump Eigen to 3.4.0".

[ 31%] Building CXX object lib/CMakeFiles/quda_cpp.dir/eig_iram.cpp.o In file included from /ccs/home/simonej/ks_rhmc/build_quda/_deps/eigen-src/Eigen/Core:19, from /ccs/home/simonej/ks_rhmc/build_quda/_deps/eigen-src/Eigen/Eigenvalues:11, from /ccs/home/simonej/ks_rhmc/quda_git/lib/../include/eigen_helper.h:12, from /ccs/home/simonej/ks_rhmc/quda_git/lib/eig_iram.cpp:15: /ccs/home/simonej/ks_rhmc/build_quda/_deps/eigen-src/Eigen/src/Core/arch/AltiVec/MatrixProductMMA.h: In instantiation of 'void Eigen::internal::gemm_complex_unrolled_MMA_iteration(const DataMapper&, const Scalar*, const Scalar*, Index, Index, Index, Index, Index&, Index, const Packet&, const Packet&) [with int unroll_factor = 4; Scalar = double; Packet = __vector(2) double; Packetc = Eigen::internal::Packet1cd; RhsPacket = Eigen::internal::PacketBlock<__vector(2) double, 2>; DataMapper = Eigen::internal::blas_data_mapper<std::complex, long int, 0, 0, 1>; Index = long int; Index accRows = 4; Index accCols = 2; bool ConjugateLhs = false; bool ConjugateRhs = false; bool LhsIsReal = false; bool RhsIsReal = false]' ...... /ccs/home/simonej/ks_rhmc/build_quda/_deps/eigen-src/Eigen/src/Core/arch/AltiVec/MatrixProductMMA.h:533:5: error: invalid conversion from type '* __vector_pair' 533 | MICRO_COMPLEX_MMA_ONE

[weinbe2]

Not strictly an answer, but do you have a specific need for GCC 10.*? I believe I've seen that flavor of gcc have issues for "no good reason" before, and have empirically found 7.* and 9.* to be reliably safe.

[weinbe2]

Not saying that the specific issue above is an error for "no good reason", I'm more implying that gcc 10 has a (personal) track record of being particularly difficult.

[james-simone]

OK, this report may just be informational. I happened upon this issue since gcc 10.2 is one of the later compilers on our ppc64 worker. I was able to reproduce the issue with the gcc 10.2 compiler on Summit.

[weinbe2]

Digging into the Eigen repo, it looks like there were some known issues related to PowerPC (the top of that header queries the CPU arch).

Fixes were merged only ~3 weeks ago (even though Eigen 3.4.0 itself is ~6 months old): https://gitlab.com/libeigen/eigen/-/merge_requests/847

This diff from the commit looks particularly relevant:

- __vector_pair* a0 = (__vector_pair *)(&a.packet[0]);
+ __vector_pair* a0 = reinterpret_cast<__vector_pair *>(const_cast<Packet2d *>(&a.packet[0]));

Since the issue isn't happening with other versions of gcc on PowerPC, I recommend just using gcc 7 or 9 for now, and as soon as a new version of Eigen is released we'll update it. (If a specific need for gcc 10 does arise, though, let us know.)

[weinbe2]

Re-opening it for visibility, will close once a new Eigen release has been merged in.

[mathiaswagner]

I would prefer to stick to released Eigen versions but we could add an option for using any git commit if there is a strong need. For now it seems we can wait on the next release.

[bjoo]

I fell afoul of this today. I got around it by downloading eigen-3.3.9 and unzipping it manually. Then setting

  -DQUDA_DOWNLOAD_EIGEN=OFF \
   -DEIGEN_INCLUDE_DIR=${SRCROOT}/eigen-3.3.9 \

on the Cmake command line. I am not sure why 3.4.0 is giving the problem. Working with GCC-10 because of the desire/need to use Concepts in QDP-JIT.

[mathiaswagner]

Is this related to #1256 ?

You should also be able to use something like -DQUDA_EIGEN_VERSION=3.3.9 instead of the more tedious download / unzip.