An NVIDIA Nsight Systems plugin specifically designed for H3P PCIe switch boxes. It monitors and records switch utilization, throughput, error counters, and device temperature onto the Nsight Systems timeline.

• CUDA Toolkit (for NVTX headers, usually located at /usr/local/cuda/include)
• libh3ppci / libph3ppci: This plugin interfaces with the H3P PCIe switch using the libh3ppci.so shared library.
• g++ and Make

1. Be sure to check the Makefile to ensure it points to the correct location of your libh3ppci.so / libh3ppci.a library.
2. Before compiling, you must export the NVTX header paths:

   export NVTX_PATH="/opt/nvidia/nsight-systems/2026.1.1/target-linux-x64/nvtx/include"
   export NVTXEXT_PATH="/opt/nvidia/nsight-systems/2026.1.1/target-linux-x64/nvtx/include"

3. Run make to compile:

   make

   This generates the h3_sw_counters executable.

Nsight Systems loads the plugin via the configuration provided in the h3_sw_counters.yaml manifest.

1. Ensure the h3_sw_counters executable was built successfully.
2. Set the NSYS_PLUGIN_SEARCH_DIRS environment variable to point to the $(pwd) subdirectory inside this repository. You can use the following command in the project root:

   export NSYS_PLUGIN_SEARCH_DIRS=$(pwd)

   (Note: The variable must point to the folder containing the h3_sw_counters.yaml file)
3. Run nsys profile and pass the --enable flag with the name of the plugin (h3_sw_counters):

   nsys profile --enable h3_sw_counters <your_target_application>

When using the --enable flag, you can pass custom arguments to the h3_sw_counters (such as the module type or sampling interval).

The syntax requires the plugin name followed by its arguments, all separated by commas (no spaces) inside double quotes:

nsys profile --enable "h3_sw_counters,<arg1>,<val1>,<arg2>,<val2>" <your_target_application>

• Monitor Error Counters instead of Throughput:

  nsys profile --enable "h3_sw_counters,-m,error" ./my_app

• Set Sampling Interval to 500ms:

  nsys profile --enable "h3_sw_counters,-t,500" ./my_app

• Combine both (Error module + 500ms interval):

  nsys profile --enable "h3_sw_counters,-m,error,-t,500,-p,0,32" ./my_app

• Monitor Multiple Modules Simultaneously:

  # Monitor throughput and temperature
  nsys profile --enable "h3_sw_counters,-m,throughput,temperature,-i,0,-p,0" ./my_app
  
  # Monitor everything (throughput, error, temperature)
  nsys profile --enable "h3_sw_counters,-m,all,-i,0,-p,0" ./my_app

• Filter by Device Index (e.g., Device 0 only):

  nsys profile --enable "h3_sw_counters,-i,0" ./my_app

These arguments are passed directly to the h3_sw_counters executable by Nsight Systems during the profiling session.