Inline-friendly dot product.

[RudolfWeeber]

No description provided.

[RudolfWeeber]

@jngrad,

I observed the not-inlined dot-product (probably originating from v.norm() with GCC13 on the shared memory pr (5097):

From perf Output:

     4.24%  python3         espresso_core.so                                   [.] auto Utils::operator*<3ul, double, double, void>(Utils::Vector<double, 3ul> const&, Utils::Vector<double, 3ul> const&)

[jngrad]

In a debug build with the python branch:

$ nm -D src/core/espresso_core.so | grep inner | c++filt
double boost::inner_product<std::vector<double>, std::vector<double>, double>(std::vector<double> const&, std::vector<double> const&, double)
double std::inner_product<double const*, double const*, double>(double const*, double const*, double const*, double)
double std::inner_product<int const*, double const*, double>(int const*, int const*, double const*, double)
int std::inner_product<int const*, int const*, int>(int const*, int const*, int const*, int)
$ nm -D src/core/espresso_core.so | grep operator | c++filt
P3MStateCommon<double>::calc_differential_operator()
P3MStateCommon<float>::calc_differential_operator()
boost::iterators::detail::operator_arrow_dispatch<BondView, BondView*>::apply(BondView const&)
boost::iterators::detail::operator_arrow_dispatch<BondView, BondView*>::proxy::proxy(BondView const&)
boost::iterators::detail::operator_arrow_dispatch<BondView, BondView*>::proxy::proxy(BondView const&)
boost::iterators::detail::operator_arrow_dispatch<BondView, BondView*>::proxy::operator->()
boost::iterators::detail::operator_arrow_dispatch<Cell&, Cell*>::apply(Cell&)

In a release build, these symbols disappear. The boost::inner_product comes from the Utils::Matrix class. PR 5097 has extra operator symbols from Kokkos.

@RudolfWeeber Which build options are you currently using?

[RudolfWeeber]

@jngrad,
CMake says: release.
Here is one of the compilation commands form the json:

{
  "directory": "/tikhome/weeber/es/build_minimal/src/core",
  "command": "/usr/bin/g++-13 -DBOOST_MPI_DYN_LINK -DBOOST_MPI_NO_LIB -DBOOST_SERIALIZATION_DYN_LINK -DBOOST_SERIALIZATION_NO_LIB -DKOKKOS_DEPENDENCE -DR123_USE_MULHILO64_C99 -Despresso_core_EXPORTS -I/tikhome/weeber/es/src/core -I/tikhome/weeber/es/src/config/include -I/tikhome/weeber/es/build_minimal/src/config/include -I/tikhome/weeber/es/src/utils/include -I/tikhome/weeber/es/src/shapes/include -I/tikhome/weeber/es/build_minimal/_deps/cabana-src/core/src -I/tikhome/weeber/es/build_minimal/_deps/cabana-build/core/src -I/tikhome/weeber/es/src/instrumentation/include -I/tikhome/weeber/es/build_minimal/_deps/caliper-src/include -I/tikhome/weeber/es/build_minimal/_deps/caliper-build/include -I/tikhome/weeber/es/src/walberla_bridge/include -isystem /tikhome/weeber/.local/include -isystem /usr/include/hdf5/openmpi -isystem /usr/lib/x86_64-linux-gnu/openmpi/include -isystem /usr/lib/x86_64-linux-gnu/openmpi/include/openmpi -isystem /tikhome/weeber/es/libs/Random123-1.09/include -isystem /tikhome/weeber/es/src/particle_observables/include -O3 -DNDEBUG -std=c++20 -fPIC -Wall -Wextra -pedantic -Wfloat-conversion -Wdelete-non-virtual-dtor -Wcast-qual -Wcast-align -Wunused-macros -Wpointer-arith -Winit-self -Wformat-signedness -Wformat=2 -Wduplicated-branches -Wshadow=compatible-local -Wno-sign-compare -Wno-unused-function -Wno-unused-parameter -Wno-array-bounds -Wno-restrict -Wno-clobbered -Wno-cast-function-type -Wno-implicit-fallthrough -Wno-format-nonliteral -Wno-float-conversion -Wno-cast-qual -DBOOST_MPL_CFG_NO_PREPROCESSED_HEADERS -DBOOST_MPL_LIMIT_LIST_SIZE=30 -fopenmp -o CMakeFiles/espresso_core.dir/communication.cpp.o -c /tikhome/weeber/es/src/core/communication.cpp",
  "file": "/tikhome/weeber/es/src/core/communication.cpp",
  "output": "src/core/CMakeFiles/espresso_core.dir/communication.cpp.o"
},

[jngrad]

To reproduce your build, I must use -D ESPRESSO_BUILD_WITH_SHARED_MEMORY_PARALLELISM=OFF.

Still, command for f in $(find src -name "*.so"); do nm -D $f | grep inner | c++filt; done returns nothing.

[jngrad]

To reproduce the issue, an Intel CPU is needed. I could reproduce it on i7-12700, i7-12700K, i7-14700K.