Implicit flex elasticity in the CG constraint solver via an effective metric
This CL replaces the post-hoc implicit flex correction (`flexInterp_cgsolve`) with a **linearly-implicit effective metric** `M̃ = M + (h² + h·damping)·K` carried by the CG constraint solver itself. Contact/friction forces and implicit flex elasticity are now computed against one consistent metric, instead of the solver seeing `M` and a post-solve correction changing `qacc` behind its back. Gate (unchanged semantics): `solver="CG"` + implicit/implicitfast integrator + pyramidal cones + flex stiffness present. Newton and PGS are untouched. `solver="CG"` remains the user-facing contract — the factorization is an implementation detail of the preconditioner. ### What's in the metric - **mjData `efm_*`** (arena, efc-like lifetime/skip semantics; built in `mj_fwdPosition`, value-refreshed in `mj_fwdVelocity`): the per-step stiffness CSR `efm_B_*`, its reverse-Cholesky factor `efm_dofid` + `efm_L_*` (nested-dissection ordered, separators-first for the reverse factorization), and the smooth-force shift `efm_c = h·K·qvel`. - **`mjd_flexStiff_assemble`** now assembles stretch (Gauss–Newton), standard dim-2 bending, and — via the cached corotated stiffness `d->flexelem_krot` — interp stiffness (all node bodies on simple sliders: point Jacobian is I₃, `flex_centered` not required; fixed nodes drop like pins) into one dof-level CSR. `mjd_effMulAdd`/`mjd_effSolve` apply the metric, with matrix-free operator fallbacks where assembly does not apply. - **mjModel `efm0_*`** (`nefm0dof`/`nefm0L`): the constant part of the metric factor — currently the dim-2 bending factor, computed once in `mj_setConst` — so bending-only models pay zero per-step factorization cost. Naming mirrors mjData's `efm_*` with the standard `0`-suffix (reference/constant) idiom, and is deliberately not bending-specific: future constant contributors extend it without renames. - The solver consumes the metric through pre-shifted `qfrc_smooth` and the metric products `Ma`/`Mv`/`Mgrad`; `qacc_smooth` becomes the unconstrained minimizer of the implicit dynamics, which makes the no-constraint shortcut and the warmstart choice consistent by construction. - **`mj_inverse` adds `B·qacc − c`**, making inverse dynamics discrete-consistent with the gated forward dynamics — exact, since the gated path has no qDeriv term (new test `ForwardTest.GatedFlexInverseConsistency`). ### Performance All numbers: ms/step over the same 2000-step window, models as shipped on each side (old code with the old model settings vs this CL with the new ones). The new solver path activates on exactly two shipped models — the ponchos, the only flex models that need an implicit integrator (poncho on Euler degenerates to >200 ms/step). For them, this CL trades speed for consistency: the implicit bending solve now runs inside every solver iteration, where the contact solve can see the stiffness, instead of once after the solve. Solver iterations drop because the curvature is visible, but each iteration pays for the implicit solve: | model | before | after | solver iters/step | |---|---|---|---| | poncho | 2.47 | 3.30 (1.33×) | 16.8 → 11.8 | | poncho_edgeequality | 1.96 | 2.72 (1.39×) | 13.2 → 10.0 | What that price buys: contact forces consistent with the implicit elasticity (previously the post-hoc correction changed `qacc` after the constraint solve), discrete-consistent inverse dynamics, and the removal of the post-hoc special case from the integration path. Raising poncho's timestep from 2 to 5 ms leaves its per-step cost nearly flat, so the consistency price can be recovered by taking fewer steps where accuracy allows. Every other flex model was measured stable on Euler at its shipped timestep and switches to it (these models predate the post-hoc integrator; implicit was never load-bearing for them). They end up equal or faster than before: bunny_multicell 0.47 → 0.40, trampoline 0.28 → 0.25, plate 1.02 → 0.99, pancake 0.34 → 0.33. Finally, the per-step factorization makes configurations practical that the old code could only integrate explicitly: implicit stretch elasticity (`elastic2d="stretch"`/`"both"`, dim-3 solids) and factorized interp stiffness. No before/after exists for these — stock has no implicit treatment of stretch at all. ### Behavior changes - With the post-hoc correction deleted, interp/bending models running `solver="Newton"` (or elliptic cones, or islands) now integrate flex elasticity **explicitly** (previously: post-hoc implicit). Affects e.g. `gripper_trilinear` (stable, and faster, but different semantics). Follow-up options: Newton-side metric support, or a documented fallback. - With the gate on, `mj_forward` outputs are timestep-dependent for gated models (they answer the linearly-implicit discrete problem); `qacc_smooth` and `mj_inverse` change accordingly. Non-gated models are bit-identical (full suite green throughout). ### Validation - 1737/1737 tests, including new: `FlexStretchDerivatives` (FD-validated GN operator), `FlexStiffAssemble`/`FlexStiffAssembleInterp` (CSR ≡ operators), `GatedFlexInverseConsistency` (fails pre-change), equivalence tests vs the old post-hoc treatment (bending matches to 2e-11). - Fingerprint discipline throughout: bending-only models bit-exact across every refactor; permutation/kernel changes verified iteration-identical. ### Known follow-ups (not in this CL) 3×3-block sparse Cholesky kernel (the numeric factorization is index-bound; projected ~3× on the factor); mjModel persistence of the factor's symbolic pattern (rest-pose ND makes sizes compile-time); the general effective-metric mode (all solvers, all PSD-safe force classes, behind an enable flag). PiperOrigin-RevId: 948561856 Change-Id: I8b8e32ebd0428042af71647d0470d10773bf6daf
This commit is contained in:
Alessio Quaglino
committed by
Copybara-Service
Copybara-Service
parent
f0fa3d8260
commit
ea230a950c
@@ -23,11 +23,13 @@
|
||||
#include <mujoco/mjsan.h> // IWYU pragma: keep
|
||||
#include "engine/engine_core_smooth.h"
|
||||
#include "engine/engine_core_util.h"
|
||||
#include "engine/engine_derivative.h"
|
||||
#include "engine/engine_forward.h"
|
||||
#include "engine/engine_io.h"
|
||||
#include "engine/engine_memory.h"
|
||||
#include "engine/engine_util_blas.h"
|
||||
#include "engine/engine_util_errmem.h"
|
||||
#include "engine/engine_util_solve.h"
|
||||
#include "engine/engine_util_misc.h"
|
||||
#include "engine/engine_util_sparse.h"
|
||||
#include "engine/engine_util_spatial.h"
|
||||
@@ -1320,6 +1322,165 @@ static void setSpring(mjModel* m, mjData* d) {
|
||||
|
||||
|
||||
// entry point: set all remaining constant fields of mjModel, except for lengthrange
|
||||
|
||||
// constant part of the implicit effective metric factor (currently dim-2 bending): sparse
|
||||
// reverse-Cholesky of M + (h^2 + h*damping)*K_bend over
|
||||
// the dofs of unpinned vertices of standard dim-2 flexes with bending. The matrix is constant
|
||||
// (flat-rest bending stiffness, point masses), so the factor is computed here once and reused
|
||||
// by the implicit-flex constraint solve every step. Bending couples only same-coordinate dofs,
|
||||
// so the pattern is three interleaved copies of the vertex flap adjacency. Row order (flex
|
||||
// order, vertex order, coordinate fastest) and the resulting fill count must match the
|
||||
// compiler's symbolic sizing (checked below).
|
||||
static void setEfm0Factor(mjModel* m, mjData* d) {
|
||||
int nbd = m->nefm0dof;
|
||||
if (!nbd) {
|
||||
return;
|
||||
}
|
||||
mj_markStack(d);
|
||||
mjtNum h = m->opt.timestep;
|
||||
|
||||
// enumerate covered vertices: compact slot per unpinned vertex of qualifying flexes
|
||||
// (filter matches the compiler's sizing and, for bending, flexStiff_active in
|
||||
// engine_derivative.c: bending data exists only for dim-2 flexes)
|
||||
int* vslot = mjSTACKALLOC(d, m->nflexvert > 0 ? m->nflexvert : 1, int);
|
||||
for (int i=0; i < m->nflexvert; i++) {
|
||||
vslot[i] = -1;
|
||||
}
|
||||
int nfree = 0;
|
||||
for (int f=0; f < m->nflex; f++) {
|
||||
if (m->flex_interp[f] || m->flex_rigid[f] || m->flex_dim[f] != 2 ||
|
||||
m->flex_bendingadr[f] < 0) {
|
||||
continue;
|
||||
}
|
||||
for (int lv=0; lv < m->flex_vertnum[f]; lv++) {
|
||||
int gv = m->flex_vertadr[f] + lv;
|
||||
if (m->body_dofnum[m->flex_vertbodyid[gv]] == 3) {
|
||||
vslot[gv] = nfree;
|
||||
nfree++;
|
||||
}
|
||||
}
|
||||
}
|
||||
if (3*nfree != nbd) {
|
||||
mj_freeStack(d);
|
||||
mjERROR("constant metric factor dof count mismatch: compiler sized %d, engine found %d",
|
||||
nbd, 3*nfree);
|
||||
}
|
||||
|
||||
// fill row -> dof address (row 3*slot + k, coordinate fastest)
|
||||
for (int gv=0; gv < m->nflexvert; gv++) {
|
||||
if (vslot[gv] >= 0) {
|
||||
int da = m->body_dofadr[m->flex_vertbodyid[gv]];
|
||||
for (int k=0; k < 3; k++) {
|
||||
m->efm0_dofid[3*vslot[gv] + k] = da + k;
|
||||
}
|
||||
}
|
||||
}
|
||||
|
||||
// assemble the bending-only stiffness K = (h^2 + h*damping)*K_bend over all dofs with the
|
||||
// shared stencil walker from engine_derivative: bending values are configuration-independent
|
||||
// and stretch/interp are gated off, so the call is valid at set-constants time (d is used
|
||||
// for stack scratch only)
|
||||
int nv = m->nv;
|
||||
int* K_rownnz = mjSTACKALLOC(d, nv, int);
|
||||
int* K_rowadr = mjSTACKALLOC(d, nv, int);
|
||||
int nK = mjd_flexStiff_assemble(m, d, K_rownnz, K_rowadr, NULL, NULL, h*h, h,
|
||||
/*flg_bend=*/1, /*flg_stretch=*/0, NULL);
|
||||
int* K_colind = mjSTACKALLOC(d, nK > 0 ? nK : 1, int);
|
||||
mjtNum* K_val = mjSTACKALLOC(d, nK > 0 ? nK : 1, mjtNum);
|
||||
mjd_flexStiff_assemble(m, d, K_rownnz, K_rowadr, K_colind, K_val, h*h, h, 1, 0, NULL);
|
||||
|
||||
// inverse map: dof address -> compact factor row (monotone: slots follow dof order)
|
||||
int* dofrow = mjSTACKALLOC(d, nv, int);
|
||||
for (int i=0; i < nv; i++) {
|
||||
dofrow[i] = -1;
|
||||
}
|
||||
for (int r=0; r < nbd; r++) {
|
||||
dofrow[m->efm0_dofid[r]] = r;
|
||||
}
|
||||
|
||||
// compact B to covered rows, keeping same-coordinate entries only: bending blocks are
|
||||
// isotropic (q * I3), so the off-coordinate entries of assemble's 3x3 block pattern are
|
||||
// structurally zero and dropping them preserves the pattern the compiler sized.
|
||||
// H = M + (h^2+h*d)*K_bend in compact dof indices: lower CSR (values) + upper CSR (pattern)
|
||||
int nHl = 0, nHu = 0;
|
||||
for (int r=0; r < nbd; r++) {
|
||||
int dof = m->efm0_dofid[r];
|
||||
for (int c=0; c < K_rownnz[dof]; c++) {
|
||||
int rc = dofrow[K_colind[K_rowadr[dof] + c]];
|
||||
if (rc < 0 || (rc - r) % 3 != 0) continue; // uncovered or off-coordinate
|
||||
if (rc < r) nHl++;
|
||||
else if (rc > r) nHu++;
|
||||
}
|
||||
}
|
||||
nHl += nbd; // diagonals: always present, also for rows without bending entries
|
||||
int* Hl_rownnz = mjSTACKALLOC(d, nbd, int);
|
||||
int* Hl_rowadr = mjSTACKALLOC(d, nbd, int);
|
||||
int* Hl_colind = mjSTACKALLOC(d, nHl, int);
|
||||
mjtNum* Hl_val = mjSTACKALLOC(d, nHl, mjtNum);
|
||||
int* Hu_rownnz = mjSTACKALLOC(d, nbd, int);
|
||||
int* Hu_rowadr = mjSTACKALLOC(d, nbd, int);
|
||||
int* Hu_colind = mjSTACKALLOC(d, nHu > 0 ? nHu : 1, int);
|
||||
|
||||
int ladr = 0, uadr = 0;
|
||||
for (int r=0; r < nbd; r++) {
|
||||
int dof = m->efm0_dofid[r];
|
||||
Hl_rowadr[r] = ladr;
|
||||
Hu_rowadr[r] = uadr;
|
||||
mjtNum diag = 0;
|
||||
// B columns are dof-ascending, so filtered lower/upper columns stay ascending
|
||||
for (int c=0; c < K_rownnz[dof]; c++) {
|
||||
int adr = K_rowadr[dof] + c;
|
||||
int rc = dofrow[K_colind[adr]];
|
||||
if (rc < 0 || (rc - r) % 3 != 0) continue;
|
||||
if (rc < r) {
|
||||
Hl_colind[ladr] = rc;
|
||||
Hl_val[ladr++] = K_val[adr];
|
||||
} else if (rc > r) {
|
||||
Hu_colind[uadr++] = rc;
|
||||
} else {
|
||||
diag = K_val[adr];
|
||||
}
|
||||
}
|
||||
// diagonal last: point mass + armature + bending diagonal
|
||||
Hl_colind[ladr] = r;
|
||||
Hl_val[ladr++] = m->body_mass[m->dof_bodyid[dof]] + m->dof_armature[dof] + diag;
|
||||
Hl_rownnz[r] = ladr - Hl_rowadr[r];
|
||||
Hu_rownnz[r] = uadr - Hu_rowadr[r];
|
||||
}
|
||||
|
||||
// symbolic factorization: counting phase (from the upper-triangle pattern)
|
||||
int* LT_rownnz = mjSTACKALLOC(d, nbd, int);
|
||||
int* LT_rowadr = mjSTACKALLOC(d, nbd, int);
|
||||
int nnz = mju_cholFactorSymbolic(NULL, m->efm0_L_rownnz, m->efm0_L_rowadr,
|
||||
NULL, LT_rownnz, LT_rowadr, NULL,
|
||||
Hu_rownnz, Hu_rowadr, Hu_colind, nbd, d);
|
||||
if (nnz != m->nefm0L) {
|
||||
mj_freeStack(d);
|
||||
mjERROR("constant metric factor size mismatch: compiler sized %d, symbolic found %d",
|
||||
(int)m->nefm0L, nnz);
|
||||
}
|
||||
|
||||
// symbolic factorization: filling phase
|
||||
int* LT_colind = mjSTACKALLOC(d, nnz, int);
|
||||
int* LT_map = mjSTACKALLOC(d, nnz, int);
|
||||
mju_cholFactorSymbolic(m->efm0_L_colind, m->efm0_L_rownnz, m->efm0_L_rowadr,
|
||||
LT_colind, LT_rownnz, LT_rowadr, LT_map,
|
||||
Hu_rownnz, Hu_rowadr, Hu_colind, nbd, d);
|
||||
|
||||
// numeric factorization
|
||||
int rank = mju_cholFactorNumeric(m->efm0_L, nbd, mjMINVAL,
|
||||
m->efm0_L_rownnz, m->efm0_L_rowadr, m->efm0_L_colind,
|
||||
LT_rownnz, LT_rowadr, LT_colind, LT_map,
|
||||
Hl_val, Hl_rownnz, Hl_rowadr, Hl_colind, d);
|
||||
if (rank != nbd) {
|
||||
mj_freeStack(d);
|
||||
mjERROR("constant metric factor is rank-deficient (%d of %d)", rank, nbd);
|
||||
}
|
||||
|
||||
mj_freeStack(d);
|
||||
}
|
||||
|
||||
|
||||
void mj_setConst(mjModel* m, mjData* d) {
|
||||
// compute npolygonmax and nmeshdegmax
|
||||
m->npolygonmax = 0;
|
||||
@@ -1354,6 +1515,9 @@ void mj_setConst(mjModel* m, mjData* d) {
|
||||
|
||||
// set quantities that depend qpos_spring
|
||||
setSpring(m, d);
|
||||
|
||||
// precompute the constant part of the implicit effective metric factor
|
||||
setEfm0Factor(m, d);
|
||||
}
|
||||
|
||||
|
||||
|
||||
Reference in New Issue
Block a user