Segmentation fault with Benchmark_wilson about grid

[paboyle]

Thanks -- I've reproduced this.

[paboyle]

For info -- I'm seeing it working under Clang and ICPC, but failing under G++ on Cori.

Amusingly, there is a bug in gdb on Cori that makes gdb itself segv when it reads the symbols in
the debug binary which is a pain... Apparently an infinite recursion loop on demangling C++11 symbols.

gdb gdb

[paboyle]

Hmm... it's associated with the use of std::thread in the comms overlapped with compute
portion.

I can disable this -- it went slower rather than faster on the Cray hardware I was targeting.

However, for now it does look to me like GCC's fault and not mine. I've committed a patch,
simply switching off the use of std::thread()

Peter

[i-kanamori]

Thank you very much, I have confirmed that the new version runs without any problems.
I actually had the same problem with Clang 3.5, but has disappeared now.

The latest version also runs with more realistic configurations with icpc:

• without the debug option: CXX=icpc --enable-simd=AVX --enable-precision=single CXXFLAGS="-std=c++11 -O3 -mavx" LIBS="-lgmp -lmpfr" --enable-comms=none
• with openmp: CXX=icpc --enable-simd=AVX --enable-precision=single CXXFLAGS="-fopenmp -std=c++11 -O3 -mavx " LIBS="-lgmp -lmpfr" --enable-comms=none
• with openmp and mpi: CXX=mpic++ --enable-simd=AVX --enable-precision=single CXXFLAGS="-fopenmp -std=c++11 -O3 -mavx " LIBS="-lgmp -lmpfr" --enable-comms=mpi
  ( my environment for the mpi is open MPI 1.10.2)

Issaku

[paboyle]

Thanks Issaku,

I will now return to the IMCI compilation.

Peter

[paboyle]

From Babbage at NERSC.

[pboyle@bint01 icpc-imci-none]$ icpc --version
icpc (ICC) 16.0.0 20150815
Copyright (C) 1985-2015 Intel Corporation. All rights reserved.

[pboyle@bint01 icpc-imci-none]$ gcc (GCC) 4.8.0

$ ../../configure --host=none --enable-precision=single --enable-simd=IMCI CXXFLAGS=-mmic -openmp -O3 -std=c++11 LDFLAGS= LIBS=-mmic -openmp --enabl
e-comms=none CXX=icpc

Clean compile:
make[1]: Entering directory `/chos/global/u1/p/pboyle/nightly/Grid/builds/icpc-imci-none/benchmarks'
CXX Benchmark_comms.o
CXX Benchmark_dwf.o
CXX Benchmark_memory_asynch.o
CXX Benchmark_memory_bandwidth.o
CXX Benchmark_su3.o
CXX Benchmark_wilson.o
CXX Benchmark_zmm.o
CXXLD Benchmark_comms
CXXLD Benchmark_su3
CXXLD Benchmark_memory_bandwidth
CXXLD Benchmark_memory_asynch
CXXLD Benchmark_zmm
CXXLD Benchmark_dwf
CXXLD Benchmark_wilson

[paboyle]

Also an aside on clang-3.5 -- Clang only integrated openMP into the main trunk on v3.8 .

I think, but don't know clang-omp++, the non-standard OpenMP development branch would have
been ok on 3.5, but not stock 3.5.

I've been using clang 3.8 since it was released since it was the first clang with OMP support
integrated in the main branch.

[i-kanamori]

I will try the new version with KNC, although I have no time till Friday. Also I will update my Clang to 3.8.
Thank you very much,
Issaku

[i-kanamori]

Now the latest version (Grid-f8f9fd6f22aede4c78284b912bee0ae79d68fb3c) runs on KNC, thank you very much.

../configure CXX=icpc --enable-simd=IMCI --enable-precision=single CXXFLAGS="-fast -mmic -fopenmp -std=c++11 -I/opt/mic/include/" LDFLAGS="-mmic" LIBS="-mmic -L/opt/mic/lib -lmpfr -lgmp" --enable-comms=none --host=none
(I also tried -O0 and -O3 instead of -fast, all of them works)