Sonar build and performance

Multibeam sonar uses CUDA. Prepare the NVIDIA driver, CUDA compiler/runtime, cuFFT, cuBLAS and Gazebo rendering.

Check the toolchain

On the machine or inside the container that will run sonar:

nvidia-smi
nvcc --version

GPU visibility and compiler availability are separate checks. For containers, configure NVIDIA Container Toolkit on the host as well as the required libraries inside the image.

Select a CUDA architecture

The sonar CMake configuration exposes CUDA_ARCHITECTURE, defaulting to 60. Set it to the SM version supported by your GPU and CUDA toolkit. From a sourced workspace, rebuild after choosing that value:

colcon build --merge-install --executor sequential --symlink-install \
  --cmake-args -DCUDA_ARCHITECTURE=YOUR_GPU_SM
source install/setup.bash

Replace YOUR_GPU_SM with the numeric architecture. Check CUDA's supported architectures before selecting it; a toolkit may have dropped an older target. Do not confuse this project-specific singular variable with CMake's CMAKE_CUDA_ARCHITECTURES.

Verify installed targets

ros2 pkg prefix multibeam_sonar
ros2 pkg prefix multibeam_sonar_system
ros2 pkg prefix posim_multibeam_sonar_demo

Check library files to confirm CUDA targets were built. Packages can register even when CUDA targets are skipped.

test -f "$(ros2 pkg prefix multibeam_sonar)/lib/multibeam_sonar/libmultibeam_sonar.so"
test -f "$(ros2 pkg prefix multibeam_sonar_system)/lib/multibeam_sonar_system/libmultibeam_sonar_system.so"

Each command must exit with status 0. Check build/launch logs for skipped targets or load errors, then receive image/raw data in Multibeam sonar.

Tune one variable at a time

Setting	Effect to measure
Horizontal beams / vertical rays	Angular sampling and computation cost
Range / `maxDistance`	Scene coverage and range processing
`raySkips`	Sampling versus computation
Sensor `update_rate`	Requested sensor throughput
`writeLog`, `debugFlag`	Disk and console overhead
Sensor gain	Image presentation; distinguish it from physical return strength

Record frame processing time, received topic rate, Gazebo real-time factor, CPU/GPU utilization and memory with the same scene and pose. Separate physics, rendering, CUDA computation, publication and logging costs. Warm up the renderer before timing.

Retain a baseline dataset when changing numerical kernels. Compare intensity/range outputs, not only rendered screenshots. Upstream research and attribution are listed under Citation.

GPU processing flow

Sonar processing includes ray summation, window/correction and FFT stages. Separate stage timings from GPU memory transfers when profiling. Figure: POSIM Notion Wiki.