[GSoC 2026 Interest] Profiling Synapse Compilation & Backend Optimization

Development
1

Hi Brian Community,

My name is Ayush, and I am a researcher working on Physics-Constrained Neural Networks and Spiking Dynamics. I have been following Brian2’s development and am interested in contributing to the C++/Cython backend optimizations for GSoC 2026.

I have set up the dev environment on Windows and started profiling Synapses.connect() behavior. I noticed some interesting caching artifacts where scaling the neuron count triggers unexpected re-compilation spikes.

As shown in the attached benchmark:

  • 1k neurons: ~0.3s (Expected)

  • 5k neurons: ~0.02s (Cached/Fast)

  • 10k neurons: Spikes to ~11.8s (Unexpected Re-compilation)

  • 20k neurons: Drops back to ~0.2s

Given my background in graph modeling and numerical methods, I am looking to work on issues related to backend performance or codegen optimization.

Could you point me toward any specific “pain points” in the current C++ standalone mode that you’d prioritize for the next cycle?

Best, Ayush

benchmark_result
benchmark_result640×480 25.2 KB

3

Update: PR Submitted

I dug into the compilation flags and found that MSVC on Windows defaults to /Ox (Full Optimization), which was causing the 23s compile times.

I have submitted a Pull Request that adds a codegen.cpp.msvc_debug preference. Toggling this switches to /Od, which brings the compile time down to 0.3s (a ~76x speedup).

PR Link: https://github.com/brian-team/brian2/pull/1730