Unidata
Test failure with new Intel compilers
I'm building the 4.9.2 version with new intel compiler icx. Specifically in nczarr_test in the run_ncgen4.sh test But if I use the icc compiler, it works ok. Is it any known bug or still not completely compatible with the new compilers?
I did a very simple build: export CC=icc ./configure --prefix=$path_netcdf/netcdf --disable-shared --with-pic --enable-largefile --disable-filter-testing --disable-dap --disable-netcdf4
We do not typically test against the Intel compilers; I'll take a look at icx and see if I can determine what is going on.
`############################## Install NETCDF-C Library ############################
cd $WRF_FOLDER/Downloads tar -xzvf v4.9.2.tar.gz cd netcdf-c-4.9.2/
these variables need to be set for the NetCDF-C install to work
export CPPFLAGS=-I$DIR/grib2/include export LDFLAGS=-L$DIR/grib2/lib export LIBS="-lhdf5_hl -lhdf5 -lz -lcurl -lgfortran -lgcc -lm -ldl -lpnetcdf"
autoreconf -i
CC=$MPICC FC=$MPIFC CXX=$MPICXX F90=$MPIF90 F77=$MPIF77 CFLAGS=$CFLAGS ./configure --prefix=$DIR/NETCDF --disable-dap --enable-netcdf-4 --enable-netcdf4 --enable-shared --enable-pnetcdf --enable-cdf5 --enable-parallel-tests | tee netcdf.configure.log
make -j $CPU_HALF_EVEN 2>&1 | tee netcdf.make.log make -j $CPU_HALF_EVEN install 2>&1 | tee make.install.log | tee netcdf.makeinstall.log
other libraries below need these variables to be set
export PATH=$DIR/NETCDF/bin:$PATH export NETCDF=$DIR/NETCDF
echo " "`
This works for GNU but not intel
@WardF
`############################## Install NETCDF-C Library ############################
cd $WRF_FOLDER/Downloads tar -xzvf v4.9.2.tar.gz cd netcdf-c-4.9.2/
these variables need to be set for the NetCDF-C install to work
export CPPFLAGS=-I$DIR/grib2/include export LDFLAGS=-L$DIR/grib2/lib export LIBS="-lhdf5_hl -lhdf5 -lz -lcurl -lgfortran -lgcc -lm -ldl -lpnetcdf"
autoreconf -i
CC=$MPICC FC=$MPIFC CXX=$MPICXX F90=$MPIF90 F77=$MPIF77 CFLAGS=$CFLAGS ./configure --prefix=$DIR/NETCDF --disable-dap --enable-netcdf-4 --enable-netcdf4 --enable-shared --enable-pnetcdf --enable-cdf5 --enable-parallel-tests | tee netcdf.configure.log
make -j $CPU_HALF_EVEN 2>&1 | tee netcdf.make.log make -j $CPU_HALF_EVEN install 2>&1 | tee make.install.log | tee netcdf.makeinstall.log
other libraries below need these variables to be set
export PATH=$DIR/NETCDF/bin:$PATH export NETCDF=$DIR/NETCDF
echo " "`
This works for GNU but not intel
@WardF
config.log netcdf.configure.log
@WardF @cross85 @edwardhartnett
So I found a solution for NETCDF-C(4.9.2) annd NETCDF-Fortran (4.6.0)
cd $WRF_FOLDER/Downloads
tar -xzvf v4.9.2.tar.gz
cd netcdf-c-4.9.2/
autoreconf -i -f 2>&1 | tee autoreconf.log
# these variables need to be set for the NetCDF-C install to work
export CPPFLAGS=-I$DIR/grib2/include
export LDFLAGS=-L$DIR/grib2/lib
export LIBS="-lhdf5_hl -lhdf5 -lz -lcurl -lgcc -lm -ldl -lpnetcdf"
CC=$MPICC FC=$MPIFC CXX=$MPICXX F90=$MPIF90 F77=$MPIF77 CFLAGS=$CFLAGS ./configure --prefix=$DIR/NETCDF --disable-dap --enable-netcdf-4 --enable-netcdf4 --enable-shared --enable-pnetcdf --enable-cdf5 --enable-parallel-tests 2>&1 | tee configure.log
automake -a -f 2>&1 | tee automake.log
make -j $CPU_HALF_EVEN 2>&1 | tee make.log
make -j $CPU_HALF_EVEN install 2>&1 | tee make.install.log
# other libraries below need these variables to be set
export PATH=$DIR/NETCDF/bin:$PATH
export NETCDF=$DIR/NETCDF
echo " "
############################## NetCDF-Fortran library ############################
cd $WRF_FOLDER/Downloads
tar -xvzf v4.6.0.tar.gz
cd netcdf-fortran-4.6.0/
autoreconf -i -f 2>&1 | tee autoreconf.log
# these variables need to be set for the NetCDF-Fortran install to work
export LD_LIBRARY_PATH=$DIR/NETCDF/lib:$LD_LIBRARY_PATH
export CPPFLAGS="-I$DIR/NETCDF/include -I$DIR/grib2/include"
export LDFLAGS="-L$DIR/NETCDF/lib -L$DIR/grib2/lib"
export LIBS="-lnetcdf -lpnetcdf -lcurl -lhdf5_hl -lhdf5 -lz -lm -ldl"
CC=$MPICC FC=$MPIFC CXX=$MPICXX F90=$MPIF90 F77=$MPIF77 CFLAGS=$CFLAGS ./configure --prefix=$DIR/NETCDF --enable-netcdf-4 --enable-netcdf4 --enable-parallel-tests --enable-hdf5 2>&1 | tee configure.log
automake -a -f 2>&1 | tee automake.log
make -j $CPU_HALF_EVEN 2>&1 | tee make.log
make -j $CPU_HALF_EVEN install 2>&1 | tee make.install.log
echo " "
export CC=icx
export CXX=icpx
export FC=ifx
export F77=ifx
export F90=ifx
export MPIFC='mpiifort -fc=ifx'
export MPIF77='mpiifort -fc=ifx'
export MPIF90='mpiifort -fc=ifx'
export MPICC='mpiicc -cc=icx'
export MPICXX='mpiicpc -cxx=icpx'
export CFLAGS="-fPIC -fPIE -O3 -Wno-implicit-function-declaration -Wno-absolute-value -Wno-incompatible-function-pointer-types -Wno-unknown-warning-option -Wno-unused-command-line-argument "
export FFLAGS="-m64"
export FCFLAGS="-m64"
Relevant Log Files: NETCDF-C make.log config.log automake.log configure.log autoreconf.log make.install.log
NETCDF-Fortran: make.log config.log automake.log configure.log autoreconf.log make.install.log
Hope This helps.
Building with icx seems to work, but 'make check' runs into failures with infinities and NaNs. From tst_netcdf4.log:
*** Testing NaN
*** creating NaN test file tst_nans.nc...ok.
*** Tests successful!
6,7c6,7
< fvar:_FillValue = nanf ;
< fvar:att = -inff, nanf, inff ;
---
> fvar:_FillValue = NaNf ;
> fvar:att = -Infinityf, NaNf, Infinityf ;
9,10c9,10
< dvar:_FillValue = nan ;
< dvar:att = -inf, nan, inf ;
---
> dvar:_FillValue = NaN ;
> dvar:att = -Infinity, NaN, Infinity ;
13c13
< fvar = -inf, _, _ ;
---
> fvar = -Infinityf, _, Infinityf ;
15c15
< dvar = -inf, _, _ ;
---
> dvar = -Infinity, _, Infinity ;
FAIL tst_netcdf4.sh (exit status: 1)
Intel(R) oneAPI DPC++/C++ Compiler 2023.2.3 (2023.2.3.20231204) netcdf-c at tag v4.9.2
So this error will take some thought of how to approach, as it appears to reflect a compiler-specific representation rather than an error in the code. Thank you for the updated report!
I was wrong. The decision to print "Infinity" or "NaN" is made by the code in ncdump in the function pr_att_algs (and probably elsewhere in ncdump). The relevant code looks roughly like this:
if(isnan(ff))
printf("NaN");
else if(isinf(ff))
printf("Infinity");
So my current hypothesis is that isnan() and isinf() are not working right so the print in ncdump defaults to (roughly)
printf("%lf",ff);
Again, a simple program could test this.
It looks like printf() for icx is handling NaN and Infinity in a
different way than other compilers.
That is:
for icx: printf("%f",(float)nan) produces "nanf" for e.g. gcc: printf("%f",(float)nan) produces "NaNf"
You easily create a program to test this. There is probably something somewhere in the C language standard about this. I suspect that icx is wrong.
Building with icx seems to work, but 'make check' runs into failures with infinities and NaNs. From tst_netcdf4.log:
*** Testing NaN *** creating NaN test file tst_nans.nc...ok. *** Tests successful! 6,7c6,7 < fvar:_FillValue = nanf ; < fvar:att = -inff, nanf, inff ; --- > fvar:_FillValue = NaNf ; > fvar:att = -Infinityf, NaNf, Infinityf ; 9,10c9,10 < dvar:_FillValue = nan ; < dvar:att = -inf, nan, inf ; --- > dvar:_FillValue = NaN ; > dvar:att = -Infinity, NaN, Infinity ; 13c13 < fvar = -inf, _, _ ; --- > fvar = -Infinityf, _, Infinityf ; 15c15 < dvar = -inf, _, _ ; --- > dvar = -Infinity, _, Infinity ; FAIL tst_netcdf4.sh (exit status: 1)Intel(R) oneAPI DPC++/C++ Compiler 2023.2.3 (2023.2.3.20231204) netcdf-c at tag v4.9.2
I've just tried -fhonor-infinities compiler flag and I've passed all the tests. I'm not 100% sure if it's a great solution. I'm using version 2024.1.0. This flag has following description: "Specify that floating-point optimizations are not allowed that assume arguments and results are not +-inf."
Did you try the test program I suggested to see what isinf() and isnan() are doing?
Thank you for the efforts. I've encountered this issue in make check just now. We're porting into a new machine. Currently, it seems -fhonor-infinities as suggested by @baneta passes the tests. I wonder about the implications for this. I'm no expert, but looking forward to its resolution.
Building with icx seems to work, but 'make check' runs into failures with infinities and NaNs. From tst_netcdf4.log:
*** Testing NaN *** creating NaN test file tst_nans.nc...ok. *** Tests successful! 6,7c6,7 < fvar:_FillValue = nanf ; < fvar:att = -inff, nanf, inff ; --- > fvar:_FillValue = NaNf ; > fvar:att = -Infinityf, NaNf, Infinityf ; 9,10c9,10 < dvar:_FillValue = nan ; < dvar:att = -inf, nan, inf ; --- > dvar:_FillValue = NaN ; > dvar:att = -Infinity, NaN, Infinity ; 13c13 < fvar = -inf, _, _ ; --- > fvar = -Infinityf, _, Infinityf ; 15c15 < dvar = -inf, _, _ ; --- > dvar = -Infinity, _, Infinity ; FAIL tst_netcdf4.sh (exit status: 1)Intel(R) oneAPI DPC++/C++ Compiler 2023.2.3 (2023.2.3.20231204) netcdf-c at tag v4.9.2
I've just tried -fhonor-infinities compiler flag and I've passed all the tests. I'm not 100% sure if it's a great solution. I'm using version 2024.1.0. This flag has following description: "Specify that floating-point optimizations are not allowed that assume arguments and results are not +-inf."
seems more like a bandaid then a fix for the issue. Because it is supressing a warning right?
The failure being reported in the test appears to be a difference in infinity representation; thanks for reporting the flag -fhonor-inifinities; I'll need to dig more into this to see what's going on and the extent to which this is a potential issue vs. (my original assumption) a difference in compilers.
After reading this: https://github.com/abseil/abseil-cpp/issues/1451 I was able to pass all "make check" tests by using the flag -fp-model=precise I hope this also helps.
Thank you all; @Maduvi I appreciate the reference, I will see about adding a compiler check and setting the flag appropriately in our CMakeLists.txt file!