CUDA SfmLevenbergMarquardtOptimizer

Created by Codex.

Run BAL-style bundle adjustment with device-resident linearization, Schur elimination, solving, retraction, and error evaluation.

Elimination and backend selection¶

CUDA exposes the same SfmEliminationMode axis as the CPU optimizer. Schur is the default and uses the existing bespoke CUDA landmark-elimination kernels. Full is reserved for a later implementation and currently throws immediately at the optimizer boundary; it never falls back to a CPU-linearized solve.

The reduced camera system can be solved with dense Cholesky, cuDSS, or PCG. Dense Cholesky remains the default.

import gtsam

if hasattr(gtsam, "cuda"):
    params = gtsam.cuda.SfmLevenbergMarquardtParams.ceresDefaults()
    params.setEliminationMode(gtsam.SfmEliminationMode.Schur)
    params.setLinearSolver(gtsam.cuda.LinearSolverType.Cudss)
    print(params.getEliminationMode(), params.getLinearSolver())

Ordering and supported graphs¶

A CUDA Schur ordering contains camera keys only. Backends that do not consume an ordering reject one. The graph must contain supported BAL projection or point-batched projection factors over SfmCamera and Point3 values. Other values are carried through unchanged.

# optimizer = gtsam.cuda.SfmLevenbergMarquardtOptimizer(
#     graph, initial, params
# )
# result = optimizer.optimize()
# diagnostics = optimizer.result()

Reserved Full hook¶

Setting SfmEliminationMode.Full is useful for feature detection and configuration round trips, but optimization raises: CUDA SFM Full elimination mode is not implemented; select Schur. The public API will not need to change when full-system CUDA solving is added.