* Disable cmake-format and clang-format checks.
* Disable GitLab CI except for NMODL + GPU.

* Add a hackathon-specific argument for benchmarks.
* Add a reference comparison for channel-benchmark.

* create build/benchmark folder before trying to use it
* run nrnivmodl-core in parallel than serially (too slow)

…umber and update the related documentation (#700)

* Add memory pool for Random123 streams.
   This speeds up initialisation when running on GPU.
* Make Boost optional.

This was a silly bug in #702.

* Simplify unified memory logic.
* Pass -mp=gpu when we pass -acc
* Pass -gpu=lineinfo for better debug information.
* Pass -Minfo=accel,mp for better compile time diagnostics.
* Add nrn_pragma_{acc,omp} macros for single-source Open{ACC,MP} support.
* Call omp_set_default_device.
* Drop cc60 because of OpenMP offload incompatibility.
* Add --gpu to test.
* Default (BB5-valid) CORENRN_EXTERNAL_BENCHMARK_DATA.
* Remove cuda_add_library.
* Don't print number of GPUs when quiet.
* Set OMP_NUM_THREADS=1 for lfp_test.
* Update NMODL to emit nrn_pragma{acc,omp} macros.

Co-authored-by: Pramod Kumbhar <[email protected]>

* Add wrapper functions for using OpenMP or OpenACC API
* Add -mp=gpu in order to link gpu runtime with tests as well
* Avoid copying VecPlay members twice otherwise association fails with OpenMP
     * IvocVect members t_ and y_ were copied twice
     * only discon_indices_ is pointer and hence that
        needs to be copied

…erGrid & threadsPerBlock (#710)

* Use #pragma omp instead of runtime API in `cnrn_target_{copyin,delete}`
* Fix `VecPlayContinuous::discon_indices_` device transfer.
* Name `cnrn_target_` wrappers more consistently.

Co-authored-by: Olli Lupton <[email protected]>

We prefer selective host-to-device updates.

Code fixes for XLC and Clang execution without build system changes.
This mainly adds missing OpenMP pragmas and makes cnrn_target_
wrappers visible to NMODL.

omp_get_mapped_ptr was added in OpenMP 5.1 and is not widely supported.

With this change then calling cnrn_target_deviceptr on a pointer that is not
present on the device is a hard error instead of returning nullptr, so avoid
calling it for artificial cells.

* Set nwarp to very big number for optimal parallelization and improve a bit grid config of CUDA solve_interleaved2

* Re-enable GitLab CI.
* Add NMODL + OpenACC test.
* Restore {clang,cmake}-format checks.
* Prefer OpenACC with MOD2C.
* Do not enable OpenACC in NMODL + OpenMP mode.
* Convert more #pragma acc to nrn_pragma_acc(...).
* Call cudaSetDevice in OpenMP mode.

Co-authored-by: Ioannis Magkanaris <[email protected]>

Presumably this was working before because our nvhpc localrc files
accidentally included CUDA include directories before
BlueBrain/spack#1392.

* Compile NVHPC+Open{ACC,MP} with -cuda.
* Pull in NMODL+Eigen fixes to make this work.