Build reference build

[ilectra]

In /home/dp208/dp208/shared/RAC16. Build also the dependencies there, so that everyone else can point to.

DONE:

• build in shared folder - both CPU and GPU
• documentation in wiki
• Unit tests pass (make check) #6

[ilectra]

Write documentation of how it works in wiki

[qiUip]

Currently fails at SU3 tests, check log in shared folder/ed_test.

shared folder/mg_test has no-sp3 branch that is the combination of the commit that works with A100 + patch to guard against SP3 - Created issue #6 to try and fix the build. We should not block this issue if we can get the first two tasks done.

[qiUip]

These are the flags I'm using for CPU. @asifsamiarain are these the same?

 ../configure \
    --prefix=${prefix} \
    --enable-comms=mpi \
    --enable-simd=AVX2 \
    --enable-gen-simd-width=256 \
    --disable-gparity \
    --disable-fermion-reps \
    --enable-Sp \
    --enable-Nc=4 \
    --with-lime=${prefix} \
    --with-gmp=${prefix} \
    --with-mpfr=${prefix} \
    CXX=mpicxx \
    CXXFLAGS="-std=c++17"

[asifsamiarain]

Here we have a script for testing:

/home/dp208/dp208/shared/RAC16/ma_test/deploy.sh

[qiUip]

This flag is GPU specific for unmanaged memory --disable-unified.

[asifsamiarain]

These are the flags I'm using for CPU. @asifsamiarain are these the same?

 ../configure \
    --prefix=${prefix} \
    --enable-comms=mpi \
    --enable-simd=AVX2 \
    --enable-gen-simd-width=256 \
    --disable-gparity \
    --disable-fermion-reps \
    --enable-Sp \
    --enable-Nc=4 \
    --with-lime=${prefix} \
    --with-gmp=${prefix} \
    --with-mpfr=${prefix} \
    CXX=mpicxx \
    CXXFLAGS="-std=c++17"

Just to mention; did a fresh compilation without make check in one go and it took nearly 2 hours at gpu-a100-80 partition and directory size in the end is 14G.

[qiUip]

Just to mention; did a fresh compilation without make check in one go and it took nearly 2 hours at gpu-a100-80 partition and directory size in the end is 14G.

I built with the above configure options on the login node with make -j32 and it took 29 min (including the configure step). The size at least is the same as in your case (14G). Based on this, I think building on the GPU nodes is a significant waste of resources (GPU compute allocation).