Intel compiler icpx 2024 takes a long time to compile deal.II with -O3

[gassmoeller]

Installing deal.II with NATIVE_OPTIMIZATIONS=ON on the new TACC Stampede3 system takes an extremely long time (~6.5 hours on 48 cores). After some experimentation and discussion with @tjhei I found that it is caused by the -O3 flag which is set by NATIVE_OPTIMIZATIONS, switching to -O2 brings compile time back to a reasonable ~1h. The problematic files that hold up the compilation are the various source/matrix_free/evaluation_template_factory_inst files.

System information:
TACC Stampede
icpx: Intel(R) oneAPI DPC++/C++ Compiler 2024.0.0 (2024.0.0.20231017)

module list output:

  1) intel/24.0      4) cmake/3.28.1   7) netcdf/4.9.2     10) phdf5/1.14.3
  2) impi/21.11      5) xalt/3.0.1     8) p4est/2.8.5
  3) autotools/1.3   6) TACC           9) trilinos/14.4.0

local.cfg content:

NATIVE_OPTIMIZATIONS=ON # or OFF
BUILD_EXAMPLES=OFF
USE_DEAL_II_CMAKE_MPI_COMPILER=OFF

DEAL_II_CONFOPTS="-D DEAL_II_WITH_COMPLEX_VALUES=OFF -D DEAL_II_COMPONENT_EXAMPLES=OFF -D DEAL_II_WITH_64BIT_INDICES=ON -D DEAL_II_CXX_FLAGS='-Wno-tautological-constant-compare -Wno-deprecated-declarations -march=native'"

PACKAGES="load:dealii-prepare once:astyle once:sundials dealii"

MKL=ON
MKL_DIR=$TACC_MKL_LIB
P4EST_DIR=$TACC_P4EST_DIR
TRILINOS_DIR=$TACC_TRILINOS_DIR
HDF5_DIR=$TACC_PHDF5_DIR
NETCDF_DIR=$TACC_NETCDF_DIR

DEAL_II_VERSION=v9.5.2

For me everything is working now, but I thought it is useful to document this behavior in case others encounter the same problem.

[tjhei]

The candi versions of p4est and trilinos fail?

[gassmoeller]

I havent tried the candi versions of p4est and trilinos, because the preinstalled modules worked.