CUDA graphs

step and reset can be captured in a CUDA graph, which replays a whole frame with one call instead of one kernel launch per kernel. A few rules follow from capture:

Figure 12Eager stepping against a captured graph.

Schematic timeline of one frame. Tutorial: CUDA graph capture · Mass-matrix update interval · Changing the model at runtime

• Most options are fixed at construction. Nine options are read every step and may be assigned between steps: baumgarte, relaxation, warn_constraint_overflow, velocity_drive_mode, contact_warmstart_decay, restitution_velocity_threshold, contact_speculative_scale, contact_gap_gate and contact_friction_gap_max. A captured graph keeps the value it was captured with; recapture after a change. contact_friction_gap_max can change between finite values, but switching to or from None (or inf) raises; build a new solver for that.
• Arrays are read in place. Edits to model arrays the solver reads every step (drive targets, loop-closure anchors, joint friction, adhesion) take effect on the next replay. Replacing an array object needs a recapture.
• Host reads stay outside capture. check_constraint_capacity(), constraint_row_watermarks() and the compliance counts read the device and raise during capture. snapshot_constraint_status() copies the status without waiting, for a later check_constraint_capacity(snapshot=True).
• Host cadence is baked in. update_mass_matrix_interval should divide the number of steps per captured graph.
• Contact torsion needs contact_torsion_device=True to be captured.
• Observables are allocated with solver.observables(...) before capture.