Add unit tests for the Hines solver (BlueBrain/CoreNeuron#841)

* As well as improving test coverage, these aim to
  demonstrate problems with nvc++ 22.5 at -O0 (BlueBrain/CoreNeuron#826).
* OpenMP target offload support for the
  `cell_permute=0` solver implementation.
* Unrelated: update NMODL submodule, which now
  requires Python 3.7+

CoreNEURON Repo SHA: BlueBrain/CoreNeuron@9523270

Author: olupton

external/nmodl

@@ -1 +1 @@
-Subproject commit 58456d1018a177691ce4d63ac2ef436d1e535000
+Subproject commit 7000ff612208ed8b27837438731903c58d1786e3

src/coreneuron/mechanism/eion.cpp

@@ -200,7 +200,6 @@ void nrn_wrote_conc(int type,
         pe[0] = nrn_nernst(pe[1 * _STRIDE], pe[2 * _STRIDE], gimap[type][2], celsius);
     }
 }
-nrn_pragma_omp(end declare target)
 
 static double efun(double x) {
     if (fabs(x) < 1e-4) {
@@ -210,6 +209,8 @@ static double efun(double x) {
     }
 }
 
+nrn_pragma_omp(end declare target)
+
 double nrn_ghk(double v, double ci, double co, double z) {
     double temp = z * v / ktf;
     double eco = co * efun(temp);

src/coreneuron/sim/solve_core.cpp

@@ -42,6 +42,7 @@ static void triang(NrnThread* _nt) {
     nrn_pragma_acc(parallel loop seq present(
         vec_a [0:i3], vec_b [0:i3], vec_d [0:i3], vec_rhs [0:i3], parent_index [0:i3])
                        async(_nt->stream_id) if (_nt->compute_gpu))
+    nrn_pragma_omp(target if (_nt->compute_gpu))
     for (int i = i3 - 1; i >= i2; --i) {
         double p = vec_a[i] / vec_d[i];
         vec_d[parent_index[i]] -= p * vec_b[i];
@@ -62,13 +63,15 @@ static void bksub(NrnThread* _nt) {
 
     nrn_pragma_acc(parallel loop seq present(vec_d [0:i2], vec_rhs [0:i2])
                        async(_nt->stream_id) if (_nt->compute_gpu))
+    nrn_pragma_omp(target if (_nt->compute_gpu))
     for (int i = i1; i < i2; ++i) {
         vec_rhs[i] /= vec_d[i];
     }
 
     nrn_pragma_acc(
         parallel loop seq present(vec_b [0:i3], vec_d [0:i3], vec_rhs [0:i3], parent_index [0:i3])
             async(_nt->stream_id) if (_nt->compute_gpu))
+    nrn_pragma_omp(target if (_nt->compute_gpu))
     for (int i = i2; i < i3; ++i) {
         vec_rhs[i] -= vec_b[i] * vec_rhs[parent_index[i]];
         vec_rhs[i] /= vec_d[i];

test/coreneuron/unit/CMakeLists.txt

@@ -31,6 +31,7 @@ if(Boost_FOUND)
     add_subdirectory(unit/interleave_info)
     add_subdirectory(unit/alignment)
     add_subdirectory(unit/queueing)
+    add_subdirectory(unit/solver)
     # lfp test uses nrnmpi_* wrappers but does not load the dynamic MPI library TODO: re-enable
     # after NEURON and CoreNEURON dynamic MPI are merged
     if(NOT CORENRN_ENABLE_MPI_DYNAMIC)

test/coreneuron/unit/solver/CMakeLists.txt

@@ -0,0 +1,20 @@
+# =============================================================================
+# Copyright (C) 2022 Blue Brain Project
+#
+# See top-level LICENSE file for details.
+# =============================================================================
+
+include_directories(${CMAKE_SOURCE_DIR}/coreneuron ${Boost_INCLUDE_DIRS})
+add_executable(test-solver test_solver.cpp)
+target_link_libraries(test-solver ${Boost_UNIT_TEST_FRAMEWORK_LIBRARY} coreneuron
+                      ${corenrn_mech_library})
+target_include_directories(test-solver SYSTEM
+                           PRIVATE ${CORENEURON_PROJECT_SOURCE_DIR}/external/CLI11/include)
+
+# Tell CMake *not* to run an explicit device code linker step (which will produce errors); let the
+# NVHPC C++ compiler handle this implicitly.
+set_target_properties(test-solver PROPERTIES CUDA_RESOLVE_DEVICE_SYMBOLS OFF)
+target_compile_options(test-solver PRIVATE ${CORENEURON_BOOST_UNIT_TEST_COMPILE_FLAGS})
+add_dependencies(test-solver nrniv-core)
+add_test(NAME test-solver COMMAND $<TARGET_FILE:test-solver>)
+cpp_cc_configure_sanitizers(TARGET test-solver TEST test-solver)