Basic Info

Description of contribution in a few bullet points

• Added sonar_compute_wgpu.hh defining a SonarComputeBackend abstract interface and WgpuComputeBackend implementation — the sensor can swap between CUDA, wgpu, and CPU backends at runtime via an SDF <compute_backend> parameter without recompiling
• Added sonar_compute_wgpu.cc implementing the wgpu backend through a Rust static library over a stable C FFI boundary — GPU init failures fall back to CPU automatically without crashing the sensor
• Added wgpu_backend/ — a Rust crate with three WGSL compute shaders implementing Choi et al. (2021) Algorithm 1, plus a cbindgen-generated C header that stays in sync with the Rust source automatically
• Updated CMakeLists.txt to optionally invoke cargo build --release when cargo is present — if cargo is absent the existing CUDA build is completely unaffected
• Existing sonar_calculation_cuda.cu and the CUDA path are untouched

Description of the physics pipeline

Three WGSL compute shaders implement Choi et al. (2021) Algorithm 1:

• backscatter.wgsl — Eq. 14 (scatter amplitude per ray) + Eq. 8 (coherent phase summation across frequency bins). Uses fixed-point atomicAdd for parallel ray accumulation since WGSL has no native float atomics. Philox4x32-10 RNG (Salmon et al. 2011, same constants as CUDA curand) + Box-Muller transform generates the complex Gaussian speckle (ξ_xi + i·ξ_yi)/√2 per ray
• matmul.wgsl — tiled 16×16 shared-memory GEMM for beam pattern correction W·P/sum(W), replacing cuBLAS SGEMM
• fft.wgsl — Cooley-Tukey radix-2 FFT converting beam spectrum P_j(f) to time-domain echo p_j(t). Range maps as r = t·c/(2B)

wgpu selects Vulkan on Linux, Metal on macOS, and DirectX 12 on Windows — the same WGSL shaders run unchanged on any backend.

Description of how this change was tested

Validated on a standalone physics pipeline using the Blueview P900-90 sonar configuration from Choi et al. (2021) Table 1 (512 beams, 114 rays, 512 freq bins, B=2.95 kHz):

GPU              : NVIDIA GeForce RTX 3050 6GB Laptop GPU (Vulkan)
range resolution : 0.254 m/bin  (c=1500 m/s, B=2.95 kHz)
expected bin     : 39  (10 m floor, r = t·c/2B)
observed peak    : 39
result           : ✓ PASS (diff=0)
avg dispatch     : 11.1 ms/frame over 10 runs  (11.1% of 10 Hz budget)

The peak bin check validates the full pipeline end-to-end: scatter amplitude (Eq. 14) → phase sweep (Eq. 8) → FFT → range. A flat floor at 10 m must produce a peak at bin t = 10 × 2B/c = 39. Getting diff=0 confirms the frequency grid convention, phase accumulation, and FFT output indexing are all correct.

Rust crate builds clean with cargo build --release on Ubuntu 22.04.

Description of documentation updates required

No documentation updates at this stage. Usage notes for the <compute_backend> SDF parameter will be added once the hookup into MultibeamSonarSensor.cc is reviewed and merged.

Known limitations and follow-up work

• MultibeamSonarSensor.cc is not yet wired to call the new backend — intentionally left for a follow-up PR once this interface is reviewed by maintainers
• GPU FFT shader requires power-of-2 n_freq; rustfft crate integration for arbitrary sizes is planned as a follow-up
• D(θ,φ) beam pattern and S(f) source spectrum are set to 1 in this PR (ideal beam, flat spectrum) — full values are applied through the matmul pass in a follow-up
• MultibeamSonarSensor.cc integration not tested on Gazebo Harmonic locally (Ubuntu 22.04 constraint) — CI will validate on the target platform