The debugger is stopping at “weird” places because the compiler inlined/optimized template/device functions (and/or you don’t have device debug symbols), so source binary mapping is not 1:1.

Quick steps to fix (do these in your Debug build)

1. Build with device debug info and disable device optimizations:

   • nvcc: add -G (Generate debug info for device) and disable optimizations for CUDA code.

   • Only use -G for local debug builds (it drastically changes code/performance).

2. Build host with debug info and no optimizations:

   • MSVC: C/C++ → Optimization = Disabled (/Od) and Debug Info = /Zi.

   • Clean + full rebuild.

3. Force a non-inlined function where you want a reliable breakpoint:

   • Use noinline on the functions you debug:

     • GCC/Clang/nvcc: attribute((noinline))

     • MSVC: __declspec(noinline)

   • Example:
     device host attribute((noinline)) TVector3<T> operator+(...) { ... }

4. Start the correct debug session in Nsight:

   • Use “Start CUDA Debugging” / Nsight CUDA debug to hit device code. The normal VS debugger only handles host.

5. Verify symbols and sources:

   • Check Modules window / Nsight logs to ensure PDB / debug symbols for the binary are loaded and match source files.

6. Short-run kernel for easier stepping:

   • Run a tiny grid (1 block, 1 thread) while debugging so it’s reproducible and easier to step.