Give up debugging and remove parallelism for hpe_vjp_vjp

mops/examples/CMakeLists.txt

@@ -4,10 +4,6 @@ add_executable(example_cpp example.cpp)
 target_link_libraries(example_cpp mops)
 add_test(NAME example_cpp COMMAND ./example_cpp)
 
-add_executable(hpe_vjp_vjp_cpp hpe_vjp_vjp.cpp)
-target_link_libraries(hpe_vjp_vjp_cpp mops)
-add_test(NAME hpe_vjp_vjp_cpp COMMAND ./hpe_vjp_vjp_cpp)
-
 if (CMAKE_CUDA_COMPILER AND MOPS_CUDA)
     add_executable(example_cuda example.cu)
     set_target_properties(example_cuda PROPERTIES CUDA_ARCHITECTURES native)

mops/src/hpe/cpu.tpp

@@ -339,7 +339,10 @@ void _homogeneous_polynomial_evaluation_vjp_vjp_templated_polynomial_order(
     scalar_t* grad_grad_a_ptr_i = nullptr;
     scalar_t* grad_o_ptr_i = nullptr;
     scalar_t* a_ptr_i = nullptr;
-    #pragma omp parallel for
+    // #pragma omp parallel for
+    // Enabling parallelism here leads to non-reentrant gradients of gradients on
+    // the GitHub CI. Not clear what the problem is. To be investigated, possibly with
+    // thread sanitizer.
     for (size_t i = 0; i < size_first_dimension_interleft; i++) {
         if (compute_grad_grad_output) {
             grad_grad_o_ptr_i = grad_grad_o_ptr + i * simd_element_count;

tox.ini

@@ -81,10 +81,6 @@ deps =
 cmake-options =
     -DCMAKE_BUILD_TYPE=Debug \
     -DMOPS_TESTS=ON \
-    -DCMAKE_CXX_FLAGS="-fsanitize=thread -fno-omit-frame-pointer" \
-    -DCMAKE_SHARED_LINKER_FLAGS="-fsanitize=thread" \
-    -DCMAKE_EXE_LINKER_FLAGS="-fsanitize=thread"
-
 
 commands =
     cmake {toxinidir}/mops -B {envdir}/build {[testenv:cxx-tests]cmake-options}