Add memory pool for Random123 streams.

[olupton]

This speeds up initialisation when running on GPU if Boost is available.

Previously many small Random123 stream objects were allocated separately using (ultimately) cudaMallocManaged in GPU builds. This is very slow, and makes setup on GPU much slower than on CPU.

This change places a pool allocator "in front of" cudaMallocManaged, which both makes allocation faster and (hopefully) reduces the number of unified memory page faults during simulation.

In a small channel-benchmark-based test this makes model setup 3x faster.

Use certain branches for the SimulationStack CI

CI_BRANCHES:NEURON_BRANCH=master,

[pramodk]

👎 boost

[bbpbuildbot]

Logfiles from GitLab pipeline #27488 (:no_entry:) have been uploaded here!

Status and direct links:

• ✅ spack_setup
• ✅ build:nmodl:gpu
• ⛔ build:coreneuron+nmodl:gpu
• ⏩ test:coreneuron+nmodl:gpu
• ⏩ build:neuron+nmodl:gpu
• ⏩ test:neuron+nmodl:gpu

[bbpbuildbot]

Logfiles from GitLab pipeline #27530 (:white_check_mark:) have been uploaded here!

Status and direct links:

• ✅ spack_setup
• ✅ build:nmodl:gpu
• ✅ build:coreneuron+nmodl:gpu
• ✅ test:coreneuron+nmodl:gpu
• ✅ build:neuron+nmodl:gpu
• ✅ test:neuron+nmodl:gpu

[pramodk]

LGTM

[pramodk]

As reported on Hackathon slack, this fails with:

[ 87%] Building CXX object coreneuron/CMakeFiles/coreneuron.dir/mpi/core/nrnmpidec.cpp.o
/ccsopen/home/PCARRIER/NEURON/CoreNeuron/coreneuron/utils/randoms/nrnran123.cu(72): error: identifier "nrnran123_State" is undefined

/autofs/nccsopen-svm1_sw/ascent/gcc/6.4.0/include/c++/6.4.0/bits/unique_ptr.h(171): error: no instance of constructor "std::tuple<_T1, _T2>::tuple [with _T1=<error-type>, _T2=coreneuron::alloc_deleter<<unnamed>::random123_allocator>]" matches the argument list
          detected during instantiation of "std::unique_ptr<_Tp, _Dp>::unique_ptr(std::unique_ptr<_Tp, _Dp>::pointer) [with _Tp=coreneuron::nrnran123_State, _Dp=coreneuron::alloc_deleter<<unnamed>::random123_allocator>]"
/ccsopen/home/PCARRIER/NEURON/CoreNeuron/coreneuron/utils/randoms/nrnran123.cu(292): here

/ccsopen/home/PCARRIER/NEURON/CoreNeuron/coreneuron/utils/memory.h(90): error: no instance of function template "std::allocator_traits<_Alloc>::destroy [with _Alloc=<unnamed>::random123_allocator]" matches the argument list
            argument types are: (<unnamed>::random123_allocator, <error-type>)
          detected during:
            instantiation of "void coreneuron::alloc_deleter<Alloc>::operator()(coreneuron::alloc_deleter<Alloc>::pointer) const [with Alloc=<unnamed>::random123_allocator]"
/autofs/nccsopen-svm1_sw/ascent/gcc/6.4.0/include/c++/6.4.0/bits/unique_ptr.h(239): here
            instantiation of "std::unique_ptr<_Tp, _Dp>::~unique_ptr() [with _Tp=coreneuron::nrnran123_State, _Dp=coreneuron::alloc_deleter<<unnamed>::random123_allocator>]"
/ccsopen/home/PCARRIER/NEURON/CoreNeuron/coreneuron/utils/randoms/nrnran123.cu(292): here

3 errors detected in the compilation of "/tmp/tmpxft_0000e191_00000000-6_nrnran123.cpp1.ii".
gmake[2]: *** [coreneuron/CMakeFiles/coreneuron.dir/utils/randoms/nrnran123.cu.o] Error 1
gmake[2]: *** Waiting for unfinished jobs....
[ 88%] Linking CXX static library ../lib/libscopmath.a
[ 88%] Built target scopmath
gmake[1]: *** [coreneuron/CMakeFiles/coreneuron.dir/all] Error 2
gmake: *** [all] Error 2