Add unit tests for the Hines solver.

[olupton]

Description
Add some tests of the various CPU/OpenACC/OpenMP/CUDA implementations of the Hines matrix solver.

These use toy input data that is generated on the fly. Tests cover scenarios such as:

• many small cells
• few large cells

both with hardcoded constant values for the matrix elements and pseudorandom number generation.

The results (d and rhs) from all available solver implementations are compared to one another in each scenario.

See also #826, as some of these tests do not seem to pass at -O0 using NVHPC 22.5. See: #841 (comment).

Also update NMODL to the latest master, which now requires Python 3.7+.

How to test this?
Build CoreNEURON with tests enabled (and Boost), run the test-solver test. Or just execute bin/test-solver.

Test System

• OS: BB5
• Compiler: NVHPC 22.3/22.5
• Version: this feature branch
• Backend: GPU

Use certain branches in CI pipelines.

CI_BRANCHES:NEURON_BRANCH=master,NMODL_BRANCH=master,SPACK_BRANCH=develop

[codecov-commenter]

Codecov Report

Merging #841 (a5e2497) into master (d6ba964) will increase coverage by 0.79%.
The diff coverage is 90.49%.

@@            Coverage Diff             @@
##           master     #841      +/-   ##
==========================================
+ Coverage   57.56%   58.36%   +0.79%     
==========================================
  Files         101      102       +1     
  Lines        9189     9410     +221     
==========================================
+ Hits         5290     5492     +202     
- Misses       3899     3918      +19     

Impacted Files	Coverage Δ
coreneuron/mechanism/eion.cpp	55.90% <ø> (ø)
coreneuron/sim/solve_core.cpp	100.00% <ø> (ø)
tests/unit/solver/test_solver.cpp	90.49% <90.49%> (ø)
coreneuron/gpu/nrn_acc_manager.cpp	42.50% <0.00%> (+2.50%)	⬆️

---

Continue to review full report at Codecov.

Legend - Click here to learn more
Δ = absolute <relative> (impact), ø = not affected, ? = missing data
Powered by Codecov. Last update d6ba964...a5e2497. Read the comment docs.

[bbpbuildbot]

Logfiles from GitLab pipeline #65143 (:no_entry:) have been uploaded here!

Status and direct links:

• ✅ spack_setup
• ✅ build:nmodl
• ✅ build:coreneuron:mod2c:nvhpc:acc
• ✅ build:coreneuron:mod2c:nvhpc:acc:unified
• ✅ build:coreneuron:nmodl:nvhpc:omp
• ✅ build:coreneuron:nmodl:nvhpc:acc
• ✅ build:coreneuron:mod2c:intel
• ✅ build:coreneuron:nmodl:intel
• ✅ build:neuron:mod2c:nvhpc:acc
• ✅ build:neuron:nmodl:nvhpc:omp
• ✅ build:neuron:nmodl:nvhpc:acc
• ✅ build:neuron:mod2c:intel
• ✅ build:neuron:nmodl:intel
• ✅ test:coreneuron:mod2c:nvhpc:acc
• ✅ test:coreneuron:mod2c:nvhpc:acc:unified
• ✅ test:coreneuron:nmodl:nvhpc:omp
• ✅ test:coreneuron:nmodl:nvhpc:acc
• ✅ test:coreneuron:mod2c:intel
• ✅ test:coreneuron:nmodl:intel
• ✅ test:neuron:mod2c:nvhpc:acc
• ✅ test:neuron:nmodl:nvhpc:omp
• ✅ test:neuron:nmodl:nvhpc:acc
• ✅ test:neuron:mod2c:intel
• ⛔ test:neuron:nmodl:intel

[olupton]

cc: @nrnhines in case you have any particular suggestions about toy data characteristics/topologies that would provide useful extra coverage.

So far the number/size of cells varies, as do the a/b/d/rhs values, but each cell is just a binary tree (possibly a bit unbalanced, depending on the total number of segments per cell that is chosen).

[olupton]

With the additional fixes in #842, I can build and run the new tests with NVHPC 22.5 and -O0 -g plus -tp haswell for nvc/nvc++ and -lineinfo for nvcc.

The CellPermute2_GPU configuration does not match the reference (CellPermute0_CPU) in the SingleCellAndThread, UnbalancedCellSingleThread, LargeCellSingleThread and LargeCellSingleThreadRandom tests, but the ones with smaller cells do pass. It appears that there is some data race issue with parallelisation across a single cell.

At -O1, the issue disappears and the tests added in this PR all pass with NVHPC 22.5.

[alexsavulescu]

LGTM 🚀