NMODL generates OpenACC code that crashes on machines without GPUs.

[olupton]

If we use NMODL and NVHPC to translate and compile mechanisms for GPU execution, the resulting special-core crashes when run on a machine without a GPU when --gpu is not passed.

The expected behaviour is:

• --gpu + no attached GPU: meaningful error message
• --gpu + attached GPU: correct results, calculated on GPU
• no --gpu: correct results, calculated on CPU, regardless of whether a GPU is attached

The problem is, presumably, that NMODL emits code like

    static inline void* mem_alloc(size_t num, size_t size, size_t alignment = 16) {
        void* ptr;
        cudaMallocManaged(&ptr, num*size);
        cudaMemset(ptr, 0, num*size);
        return ptr;
    }


    static inline void mem_free(void* ptr) {
        cudaFree(ptr);
    }

that does not check --gpu. This could be changed to use these utilities in CoreNEURON:
https://github.com/BlueBrain/CoreNeuron/blob/9c821d51e90237827e5e70fe76748bc54612913a/coreneuron/utils/memory.h#L26-L114

cc: @iomaganaris

[olupton]

Note that after various changes (BlueBrain/CoreNeuron#595, BlueBrain/CoreNeuron#606 and BlueBrain/CoreNeuron#609) then CPU execution of GPU-enabled builds should work, although there are still some non-fatal cases to be fixed as noted in BlueBrain/CoreNeuron#599.

The model I was testing runs successfully on CPU in a GPU build that uses mod2c.