I ran the command “ncu -o profile --set full ./executable” and analyzed my kernel launch using the Nvidia Visual Profiler, version 2023.2.0.0, and I am trying to reduce my bank conflicts. However, when I looked under Memory Workload Analysis > Shared Memory, I noticed that almost all of my bank conflicts come from the “Other” category, as shown in the image attached. I’ve tried looking here for what the Other category may include, but I’m having a hard time finding the answer to that. Does anyone know that category includes?
There’s some information in this GTC presentation How to Understand and Optimize Shared Memory Accesses using Nsight Compute | NVIDIA On-Demand about shared memory accesses. If you want to attach the full report, there may be some additional clues to help uncover what is causing this.