Move to gcc10 as baseline compiler (OMP build error on ggttgg with gcc9.2)

[valassi]

While working on epochX4 PR #267 I got a build error for ggttgg using gcc9.2

ccache /cvmfs/sft.cern.ch/lcg/releases/gcc/9.2.0-afc57/x86_64-centos7/bin/g++  -O3  -std=c++17 -I. -I../../src -I../../../../../tools -DUSE_NVTX -Wall -Wshadow -Wextra -fopenmp -ffast-math   -DMGONGPU_FPTYPE_DOUBLE -DMGONGPU_CURAND_ONDEVICE -I/usr/local/cuda-11.4/include/ -c CPPProcess.cc -o CPPProcess.o
CPPProcess.cc: In function ‘void Proc::calculate_wavefunctions(int, const fptype_sv*, fptype_sv*, int)’:
CPPProcess.cc:1866:24: error: ‘*.LC0’ not specified in enclosing ‘parallel’
 1866 |       constexpr fptype cf[ncolor][ncolor] = {
      |                        ^~
CPPProcess.cc:108:9: error: enclosing ‘parallel’
  108 | #pragma omp parallel for default(none) shared(allmomenta,allMEs,cHel,cIPC,cIPD,ihel,npagV) private (amp_sv,w_sv,jamp_sv)
      |         ^~~
make: *** [CPPProcess.o] Error 1

It appears that this is a gcc9 error
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=91920
The bug report mentions that this is fixed in gcc10 and backported to gcc9.3.

Indeed I tested gcc10.2 and the issue seems gone, I no longer get the build error above.

It is probably a good idea to upgrade my baseline compiler anyway (even for AVX512 issue #173). I am using cuda 11.4 and I think that gcc11 is no longer supported by it. I will try to upgrade to the latest gcc10 I have available (gcc10.3) and check that I get no performance hit.

[valassi]

Note also that only gcc10 includes the lto-dump tool, useful for LTO studies issue #229

[valassi]

I had noted somewhere (in commit log 6ff529c and in my private bashMG.sh) that gcc10.3 was giving issues (a segmentation fault?) with cuda 10.4, but presently I see no such issue. I will move to gcc10.3.

[valassi]

Using gcc10.3 I see no real performance benefit on the default "no inline" build, but I see a 10% gain in the "aggressive inlining" mode (#229)...

[valassi]

Ok this is now done - closing it.

I am using

. /cvmfs/sft.cern.ch/lcg/releases/gcc/10.3.0/x86_64-centos7/setup.sh

I have updated all logs to use gcc10.3 as new baseline

The code now builds AND REQUIRES gcc10 (or at least, it will not build on gcc9.2 due to the OMP bug).

Closing

[valassi]

Note that because of the issues discussed in #277, #281, #282, I have downgraded back from cuda 11.4 to cuda 11.1 as baseline. Unlike cuda 11.0, this allows using gcc10, but I am using now gcc10.2 (as my cvmfs 10.3 has some issues). I am using c++17 in gcc and c++14 in cuda.