uw-gpu-klt

CUDA port of the uw_ct_vio KLT stereo front end: a drop-in GpuKltStereoTracker that reproduces the algorithm and OpenCV's numerics on the GPU.

track() latency per frame, CPU vs GPU	End-to-end ATE with CPU vs GPU tracks
NTNU traj_06, 6058 frames, shared RTX 5080 (other users' jobs inflate GPU tails)	all 6 NTNU 2023 trajectories, one run each

Results

Front end, NTNU traj_06, 6058 frames, closed loop, workstation RTX 5080 (shared machine, 2026-11-04):

ms/frame	mean	P50	P95	P99	max
CPU track()	19.28	19.01	26.31	31.58	49.97
GPU track()	9.53	7.91	20.69	24.24	37.57

Speed-up 2.0x mean, 2.4x P50 (laptop, quiet period: 4.9x mean). Not measured on Jetson Orin (sm_87 compiled, never run).

End to end, uw_ct_vio best config (h05) on all 6 NTNU 2023 trajectories, same binary and flags:

tracks	mean ATE (m)	mean RPE (%)
CPU KLT (h05)	0.401	6.75
GPU KLT	0.341	6.71

No accuracy loss on any trajectory. One run per trajectory, no variance estimate: the lower GPU mean is NOT claimed as an improvement.

Quick start

git clone https://github.com/Blackcipher101/uw-vio-dvl ../uw-vio-dvl   # CPU reference tracker and types
bash -lc 'source /opt/ros/jazzy/setup.bash && \
  cmake -S . -B build -DCMAKE_BUILD_TYPE=Release -DGTSAM_PREFIX=$HOME/deps/gtsam-4.2.0 \
        -DCMAKE_CUDA_COMPILER=/usr/local/cuda-12.8/bin/nvcc && cmake --build build -j2'
LD_LIBRARY_PATH=$HOME/deps/gtsam-4.2.0/lib ctest --test-dir build --output-on-failure
bench/run_bench.sh --gpu-keepalive    # NTNU root from $NTNU

More

• docs/DESIGN.md: which stages run on the GPU and their measured fidelity to the CPU tracker, unsupported options, synchronisation and allocation behaviour, CUDA architectures.
• docs/BENCHMARK.md: build and test, benchmark options, the UWTRK1 track replay, the full per-trajectory ATE table and figure regeneration.

Known limits: map_predict, reacquire, snr_sigma, ncc_noise_aware not implemented (constructor throws); max 2048 features; single-threaded use; only sm_120 has been run.