Quda work merge master

[kostrzewa]

Here's my attempt to synchronise master and quda_work. I've not tested compilation and hence also not done any regression tests. To be honest, I never wanted to touch this code again...

[kostrzewa]

It would be nice if someone could set up github actions to at least automatically test the build system...

[sunpho84]

what do you mean? isn't travis cehcking the build?

[kostrzewa]

what do you mean? isn't travis cehcking the build?

that has been broken for ages.

[kostrzewa]

Anyway, this builds now on my machine but one has to be careful to create a new build directory (or appropriately empty the current one) because there is a change to the files generated in ${top_build_dir}/include.

[kostrzewa]

Doing the following:

CC=mpicc CXX=mpicxx F77=f77 \                                                                                                                                
CFLAGS="-mtune=znver2 -march=znver2 -O3 -mavx2 -mfma -fopenmp" \
LDFLAGS="-fopenmp" \
~/code/tmLQCD.merge_quda_work/configure \
  --enable-mpi \
  --with-mpidimension=4 \
  --disable-sse2 --disable-sse3 \
  --with-cudadir=/usr/local/cuda/lib64 \
  --with-qudadir=$(pwd)/../quda_develop/install_dir \
  --with-limedir=$(pwd)/../lime/install_dir \
  --with-lapack="-lblas -llapack"

in an empty build directory works for me, but maybe there is still some compiler dependence...

[sunpho84]

Doing the following:

CC=mpicc CXX=mpicxx F77=f77 \                                                                                                                                
CFLAGS="-mtune=znver2 -march=znver2 -O3 -mavx2 -mfma -fopenmp" \
LDFLAGS="-fopenmp" \
~/code/tmLQCD.merge_quda_work/configure \
  --enable-mpi \
  --with-mpidimension=4 \
  --disable-sse2 --disable-sse3 \
  --with-cudadir=/usr/local/cuda/lib64 \
  --with-qudadir=$(pwd)/../quda_develop/install_dir \
  --with-limedir=$(pwd)/../lime/install_dir \
  --with-lapack="-lblas -llapack"

in an empty build directory works for me, but maybe there is still some compiler dependence...

yes actually it almost does, now I get

../../io/utils_write_first_message.c: In function ‘write_first_messages’:
../../io/utils_write_first_message.c:27:115: error: ‘TMLQCD_PACKAGE_VERSION’ undeclared (first use in this function); did you mean ‘LIME_PACKAGE_VERSION’?
   27 |   snprintf(message, 1024, "This is the %s code for twisted mass Wilson QCD\n\nVersion %s, commit %s\n",executable,TMLQCD_PACKAGE_VERSION,git_hash);
      |                                                                                                                   ^~~~~~~~~~~~~~~~~~~~~~
      |                                                                                                                   LIME_PACKAGE_VERSION
../../io/utils_write_first_message.c:27:115: note: each undeclared identifier is reported only once for each function it appears in

[kostrzewa]

yes actually it almost does, now I get

delete the file include/tmlqcd_config.h in your build dir, it's a left-over from a previous build which is not deleted by the build system (since now a different file is auto-generated)

[sunpho84]

ok I see, a make distclean did the trick

[kostrzewa]

Had to make one more modif, moving the non-generated file into the include/ directory in the source dir. This is in order to be able to grab the settings when using tmLQCD as a library in an external program as in that case, one doesn't want to have to specify the source directory itself as an include dir (as there might be include files with the same names between the client app and tmLQCD). In the end, both tmlqcd_config.h and tmlqcd_config_internal.h should of course be installed.

I've checked this now linked against my cvc_depgraph extension of CVC and the changes merged in from master don't seem to break the QUDA behaviour.

[sunpho84]

what do you mean? isn't travis cehcking the build?

that has been broken for ages.

? travis is checking my fork and he seems to be rather happy...

https://travis-ci.com/github/sunpho84/tmLQCD/builds/222046903