`v5_sonata` CoreNEURON tests fail since Neuron commit 0e7570e50

[WeinaJi]

Since Neuron 0e7570e50, two v5_sonata coreneuron unit tests fail

1. test_v5_sonata_multisteps
2. test_v5_sonata_lfp

I first saw in the workflow with Neuron 942d5ef, and later reproduced locally down to Neuron commit 0e7570e50. It seems there have been some changes in nmodl for code generation for CoreNEURON.

The change of results happens only with CoreNEURON, the NEURON results for v5_sonata remain unchanged, which means that we have discrepancy between CoreNEURON and NEURON results.

@jamesgking @JCGoran

[jamesgking]

Doing, some investigation using --dump-cell-state=4

I see differences in the SK_E2 initial values for cai

In file 5_cpu_t0.000000.corenrn
earlier NEURON: cai = 5e-05
latest NEURON: cai = 0

whereas for both versions, the value in the NEURON dump 5_py_Neuron_t0.0.nrndat
cai = 5e-05

So, further examination is need for why latest coreneuron has different value

[WeinaJi]

Comparing the compiled corenrn/mod2c/SK_E2.cpp with neuron 0e7570e50 vs a34dfd3, I notice the different lines on the mem_alloc function.

1. 0e7570e50, NMODL Compiler : 0.6 [800d098cc 2024-10-31 09:22:42 +0100]

    static inline void* mem_alloc(size_t num, size_t size, size_t alignment = 64) {
        size_t aligned_size = ((num*size + alignment - 1) / alignment) * alignment;
        void* ptr = aligned_alloc(alignment, aligned_size);
        memset(ptr, 0, aligned_size);
        return ptr;
    }

2. a34dfd3, NMODL Compiler : 0.6 [791014679 2024-07-05 16:14:29 +0200]

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

I don't find the definition of posix_memalign. @JCGoran , would this cause any difference ?

[jamesgking]

Thanks for highlighting where the core neuron files are generated, Weina. I took a look at other mod files that impact ca ion

In CaDynamics_E2.cpp and CaDynamics_DC0.cpp

The newer version has an additional line:
inst->cai[id] = inst->global->cai0;
In function
void nrn_init_CaDynamics_E2(NrnThread* nt, Memb_list* ml, int type)

This line doesn't exist in the older version

[WeinaJi]

With the attempts in #143 and my local tests that could make the compiled corenrn/mod2c/SK_E2.cpp identical between neuron 0e7570e50 and a34dfd3, the test_lfp.py still fails with the same discrepancy to the reference file.
It seems that the issue is somewhere else.

[jamesgking]

Trying to pick this task up again.

I edit the generated CaDynamics_E2.cpp and CaDynamics_DC0.cpp files to remove the line

inst->cai[id] = inst->global->cai0;

And then regenerate the special. When I execute, in file 5_cpu_t0.000000.corenrn, the value matches the earlier commit

cai = 5e-05

[jamesgking]

When I run with '--coreneuron-direct-mode', then my dump state files are back to having the same values for cai