Glossary#

The terms of art this repo uses — several are overloaded, and two different things are both called a “tier”. When a doc and this page disagree, fix the doc.

resource tier#: The shared-memory budget a kernel is compiled for: TIER_SHARED (everything in smem), then the spill rungs (TIER_LITE / TIER_MINIMAL) that move progressively more scratch to the global d_workspace. Selected per algorithm per robot at codegen time (Resource-Tier System (v2.0)), dispatched via the RESOURCE_TIER template parameter. This is what “tier” means unqualified.
precompile variants#: The OTHER historical use of “tier”: grid_rbd.precompile(..., tiers=...) builds several pre-compiled .so option-variants of one robot. Unrelated to resource tiers — prefer calling these variants.
arena#: A kernel’s flat shared-memory scratch pool (s_temp), laid out at codegen time with compile-time offsets. The per-tier *_DYNAMIC_SHARED_MEM_BYTES macros size it.
spill rung#: One step of a tier’s shrink ladder: a specific buffer (or group) rerouted from the smem arena to d_workspace. Rungs are applied in a fixed order until the arena fits the target.
_inner / _device / _kernel / host#: The four emission layers per algorithm — placement-free computational core; the __device__ orchestrator owning scratch placement; the __global__ batch entry point; the H↔D-copying host wrapper. See Codegen Architecture.
mjx / twin / pin#: pin = Pinocchio-convention (GRiD-native) kernels and I/O. mjx = MuJoCo-convention (wxyz quaternion, global-linear free-joint velocity). A twin is the mjx variant of a kernel (MUJOCO_OUTPUT=true instantiation), emitted only for floating-base non-mimic/skew robots.
receipt / shard / carry#: The signed gpu-proof.json at the repo root attests GPU test outcomes so CPU-only CI can gate merges. A shard is one crash-isolated slice of the split suite with its own fingerprint; a carried shard is one whose files are unchanged since an ancestor commit’s green run and is reused rather than re-executed (everyday policy only — a release receipt refuses carries).
fingerprint#: A content hash binding GPU evidence to its inputs. The global receipt covers the generator, bindings, peer-library gitlinks, configuration, and GPU test harness. Narrow shard fingerprints support the separately guarded everyday refresh workflow. Website wording and figures are outside the global scope, so committed presentation-only changes can retain a completed receipt; code changes cannot silently reuse it.
GLASS#: The first-party GPU linear-algebra header library (external/GLASS). Extend it rather than working around it.
ffi_bases#: The per-algo {tier, threads} block of a robot’s tuned launch config (config/launch_configs/<robot>/<gpu>.json) used by the jax/torch FFI launch path; pybind_bases / torch_bases are runtime overlays over the same bake.
batch-to-land#: Bench metric: wall time for a batch INCLUDING the device→host copy of the result (what a consumer actually waits for), vs. kernel-only time.
mimic joint#: A URDF joint slaved to another via <mimic> (q_child = m·q_parent + o). Collapses nv below the joint count and makes many per-joint kernels take the reduced path; mimic robots never get mjx twins.
subset build / algorithm_list#: Building only named algorithms (plus transitive deps) into a robot’s .so. Un-built methods raise the clean “not built into this robot .so — add to algorithm_list” error (C-ABI rc=3).