www.neuron.yale.edu

Hi.

I was looking to this old thread (viewtopic.php?p=14858&hilit=avx#p14858) about compiling NEURON with AVX.
Now that AVX-256 and AVX-512 are common on desktop CPU, I was wondering if it was possible to add a flag, during the configure, to compile NEURON for one instruction set or the other.

I wrote to pkumbhar and hope Pramod can chime in on this.

bremen wrote: ↑Wed Oct 16, 2024 8:52 am Now that AVX-256 and AVX-512 are common on desktop CPU, I was wondering if it was possible to add a flag, during the configure, to compile NEURON for one instruction set or the other.

In NEURON, we don't have auto-vectorization capability (yet). So, even with AVX-256/AVX-512 specific flags, the code generated from MOD files won't use AVX-256/AVX-256 today. This is something we want to fix in near future.

We worked on CoreNEURON library inside NEURON and using compiler like Intel and standard -DCMAKE_BUILD_TYPE=RelWithDebInfo CMake flag is sufficient to generate vectorized code. But again, this is CoreNEURON specific and it doesn't support all features (see Supported features on this page and here is how to).

HI.

Thank you for the explanation.