Skip to content

solver warmstart #5

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Draft
wants to merge 1 commit into
base: main
Choose a base branch
from
Draft

solver warmstart #5

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
1 commit
Select commit
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
3 changes: 3 additions & 0 deletions mujoco_warp/_src/io.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -346,14 +346,17 @@ def _constraint(mjm: mujoco.MjModel, nworld: int, njmax: int) -> types.Constrain
efc.worldid = wp.zeros((njmax,), dtype=wp.int32)

efc.Jaref = wp.empty(shape=(njmax,), dtype=wp.float32)
efc.Jaref_warmstart = wp.empty(shape=(njmax,), dtype=wp.float32)
efc.Ma = wp.empty(shape=(nworld, mjm.nv), dtype=wp.float32)
efc.Ma_warmstart = wp.empty(shape=(nworld, mjm.nv), dtype=wp.float32)
efc.grad = wp.empty(shape=(nworld, mjm.nv), dtype=wp.float32)
efc.grad_dot = wp.empty(shape=(nworld,), dtype=wp.float32)
efc.Mgrad = wp.empty(shape=(nworld, mjm.nv), dtype=wp.float32)
efc.search = wp.empty(shape=(nworld, mjm.nv), dtype=wp.float32)
efc.search_dot = wp.empty(shape=(nworld,), dtype=wp.float32)
efc.gauss = wp.empty(shape=(nworld,), dtype=wp.float32)
efc.cost = wp.empty(shape=(nworld,), dtype=wp.float32)
efc.cost_warmstart = wp.empty(shape=(nworld,), dtype=wp.float32)
efc.prev_cost = wp.empty(shape=(nworld,), dtype=wp.float32)
efc.solver_niter = wp.empty(shape=(nworld,), dtype=wp.int32)
efc.active = wp.empty(shape=(njmax,), dtype=wp.int32)
Expand Down
153 changes: 137 additions & 16 deletions mujoco_warp/_src/solver.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -23,15 +23,14 @@
from .warp_util import kernel_copy


def _create_context(m: types.Model, d: types.Data, grad: bool = True):
def _create_context(m: types.Model, d: types.Data):
@kernel
def _init_context(d: types.Data):
worldid = wp.tid()
d.efc.cost[worldid] = wp.inf
d.efc.solver_niter[worldid] = 0
d.efc.done[worldid] = False
if grad:
d.efc.search_dot[worldid] = 0.0
d.efc.search_dot[worldid] = 0.0

@kernel
def _jaref(m: types.Model, d: types.Data):
Expand All @@ -47,13 +46,6 @@ def _jaref(m: types.Model, d: types.Data):
d.efc.J[efcid, dofid] * d.qacc[worldid, dofid] - d.efc.aref[efcid] / float(m.nv),
)

@kernel
def _search(d: types.Data):
worldid, dofid = wp.tid()
search = -1.0 * d.efc.Mgrad[worldid, dofid]
d.efc.search[worldid, dofid] = search
wp.atomic_add(d.efc.search_dot, worldid, search * search)

wp.launch(_init_context, dim=(d.nworld), inputs=[d])

# jaref = d.efc_J @ d.qacc - d.efc_aref
Expand All @@ -65,11 +57,128 @@ def _search(d: types.Data):
support.mul_m(m, d, d.efc.Ma, d.qacc, d.efc.done)

_update_constraint(m, d)
if grad:
_update_gradient(m, d)

# search = -Mgrad
wp.launch(_search, dim=(d.nworld, m.nv), inputs=[d])

def _create_context_warmstart(m: types.Model, d: types.Data):
@kernel
def _init(d: types.Data):
worldid = wp.tid()
d.efc.cost[worldid] = 0.0
d.efc.cost_warmstart[worldid] = 0.0
d.efc.solver_niter[worldid] = 0
d.efc.done[worldid] = False
d.efc.search_dot[worldid] = 0.0

@kernel
def _jaref(m: types.Model, d: types.Data):
efcid, dofid = wp.tid()

if efcid >= d.nefc[0]:
return

worldid = d.efc.worldid[efcid]
J = d.efc.J[efcid, dofid]
aref = d.efc.aref[efcid]
scl = 1.0 / float(m.nv)

wp.atomic_add(
d.efc.Jaref,
efcid,
J * d.qacc_smooth[worldid, dofid] - aref * scl,
)

wp.atomic_add(
d.efc.Jaref_warmstart,
efcid,
J * d.qacc_warmstart[worldid, dofid] - aref * scl,
)

@kernel
def _cost(d: types.Data):
efcid = wp.tid()

if efcid >= d.nefc[0]:
return

worldid = d.efc.worldid[efcid]

Jaref = d.efc.Jaref[efcid]
Jaref_warmstart = d.efc.Jaref_warmstart[efcid]
efc_D = d.efc.D[efcid]

# cost = 0.5 * sum(efc_D * Jaref * Jaref * active)
# TODO(team): active and conditionally active constraints
if Jaref < 0.0:
wp.atomic_add(d.efc.cost, worldid, 0.5 * efc_D * Jaref * Jaref)

if Jaref_warmstart < 0.0:
wp.atomic_add(
d.efc.cost_warmstart, worldid, 0.5 * efc_D * Jaref_warmstart * Jaref_warmstart
)

@kernel
def _gauss(d: types.Data):
worldid, dofid = wp.tid()
qfrc_smooth = d.qfrc_smooth[worldid, dofid]
qacc_smooth = d.qacc_smooth[worldid, dofid]

gauss_cost = (
0.5
* (d.efc.Ma[worldid, dofid] - qfrc_smooth)
* (d.qacc_smooth[worldid, dofid] - qacc_smooth)
)
wp.atomic_add(d.efc.cost, worldid, gauss_cost)

gauss_cost_warmstart = (
0.5
* (d.efc.Ma_warmstart[worldid, dofid] - qfrc_smooth)
* (d.qacc_warmstart[worldid, dofid] - qacc_smooth)
)
wp.atomic_add(d.efc.cost_warmstart, worldid, gauss_cost_warmstart)

@kernel
def _qacc_ma_warmstart(d: types.Data):
worldid, dofid = wp.tid()

if d.efc.cost_warmstart[worldid] < d.efc.cost[worldid]:
d.qacc[worldid, dofid] = d.qacc_warmstart[worldid, dofid]
d.efc.Ma[worldid, dofid] = d.efc.Ma_warmstart[worldid, dofid]
else:
d.qacc[worldid, dofid] = d.qacc_smooth[worldid, dofid]

@kernel
def _jaref_warmstart(d: types.Data):
efcid = wp.tid()

if efcid >= d.nefc[0]:
return

worldid = d.efc.worldid[efcid]

if d.efc.cost_warmstart[worldid] < d.efc.cost[worldid]:
d.efc.Jaref[efcid] = d.efc.Jaref_warmstart[efcid]

wp.launch(_init, dim=(d.nworld), inputs=[d])

# jaref = d.efc_J @ qacc - d.efc_aref
d.efc.Jaref.zero_()
d.efc.Jaref_warmstart.zero_()
wp.launch(_jaref, dim=(d.njmax, m.nv), inputs=[m, d])

# Ma = qM @ qacc
support.mul_m(m, d, d.efc.Ma, d.qacc_smooth, d.efc.done)
support.mul_m(m, d, d.efc.Ma_warmstart, d.qacc_warmstart, d.efc.done)

# cost = 0.5 * sum(efc_D * Jaref * Jaref * active)
wp.launch(_cost, dim=(d.njmax,), inputs=[d])

# gauss = 0.5 * (Ma - qfrc_smooth).T @ (qacc - qacc_smooth)
wp.launch(_gauss, dim=(d.nworld, m.nv), inputs=[d])

wp.launch(_qacc_ma_warmstart, dim=(d.nworld, m.nv), inputs=[d])
wp.launch(_jaref_warmstart, dim=(d.njmax), inputs=[d])

_update_constraint(m, d)


def _update_constraint(m: types.Model, d: types.Data):
Expand Down Expand Up @@ -874,6 +983,13 @@ def solve(m: types.Model, d: types.Data):
"""Finds forces that satisfy constraints."""
ITERATIONS = m.opt.iterations

@kernel
def _init_search(d: types.Data):
worldid, dofid = wp.tid()
search = -1.0 * d.efc.Mgrad[worldid, dofid]
d.efc.search[worldid, dofid] = search
wp.atomic_add(d.efc.search_dot, worldid, search * search)

@kernel
def _zero_search_dot(d: types.Data):
worldid = wp.tid()
Expand Down Expand Up @@ -970,9 +1086,14 @@ def _beta(d: types.Data):
)

# warmstart
kernel_copy(d.qacc, d.qacc_warmstart)
if not m.opt.disableflags & types.DisableBit.WARMSTART:
_create_context_warmstart(m, d)
else:
wp.copy(d.qacc, d.qacc_smooth)
_create_context(m, d)

_create_context(m, d, grad=True)
_update_gradient(m, d)
wp.launch(_init_search, dim=(d.nworld, m.nv), inputs=[d])

for i in range(m.opt.iterations):
_linesearch(m, d)
Expand Down
2 changes: 2 additions & 0 deletions mujoco_warp/_src/solver_test.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -67,6 +67,8 @@ def _load(
d = mjwarp.put_data(mjm, mjd, nworld=nworld, njmax=njmax)
return mjm, mjd, m, d

# TODO(team): test disableflags warmstart

@parameterized.parameters(
(mujoco.mjtCone.mjCONE_PYRAMIDAL, mujoco.mjtSolver.mjSOL_CG, 25, 5, False, False),
(
Expand Down
9 changes: 8 additions & 1 deletion mujoco_warp/_src/types.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -46,11 +46,12 @@ class DisableBit(enum.IntFlag):
PASSIVE = mujoco.mjtDisableBit.mjDSBL_PASSIVE
GRAVITY = mujoco.mjtDisableBit.mjDSBL_GRAVITY
CLAMPCTRL = mujoco.mjtDisableBit.mjDSBL_CLAMPCTRL
WARMSTART = mujoco.mjtDisableBit.mjDSBL_WARMSTART
ACTUATION = mujoco.mjtDisableBit.mjDSBL_ACTUATION
REFSAFE = mujoco.mjtDisableBit.mjDSBL_REFSAFE
EULERDAMP = mujoco.mjtDisableBit.mjDSBL_EULERDAMP
FILTERPARENT = mujoco.mjtDisableBit.mjDSBL_FILTERPARENT
# unsupported: EQUALITY, FRICTIONLOSS, MIDPHASE, WARMSTART, SENSOR
# unsupported: EQUALITY, FRICTIONLOSS, MIDPHASE, SENSOR


class TrnType(enum.IntEnum):
Expand Down Expand Up @@ -274,14 +275,17 @@ class Constraint:
aref: reference pseudo-acceleration (njmax,)
force: constraint force in constraint space (njmax,)
Jaref: Jac*qacc - aref (njmax,)
Jaref_warmstart: Jac*qacc_warmstart - aref (njmax,)
Ma: M*qacc (nworld, nv)
Ma_warmstart: M*qacc_warmstart (nworld, nv)
grad: gradient of master cost (nworld, nv)
grad_dot: dot(grad, grad) (nworld,)
Mgrad: M / grad (nworld, nv)
search: linesearch vector (nworld, nv)
search_dot: dot(search, search) (nworld,)
gauss: gauss Cost (nworld,)
cost: constraint + Gauss cost (nworld,)
cost_warmstart: constraint + Gauss cost (nworld,)
prev_cost: cost from previous iter (nworld,)
solver_niter: number of solver iterations (nworld,)
active: active (quadratic) constraints (njmax,)
Expand Down Expand Up @@ -322,14 +326,17 @@ class Constraint:
aref: wp.array(dtype=wp.float32, ndim=1)
force: wp.array(dtype=wp.float32, ndim=1)
Jaref: wp.array(dtype=wp.float32, ndim=1)
Jaref_warmstart: wp.array(dtype=wp.float32, ndim=1)
Ma: wp.array(dtype=wp.float32, ndim=2)
Ma_warmstart: wp.array(dtype=wp.float32, ndim=2)
grad: wp.array(dtype=wp.float32, ndim=2)
grad_dot: wp.array(dtype=wp.float32, ndim=1)
Mgrad: wp.array(dtype=wp.float32, ndim=2)
search: wp.array(dtype=wp.float32, ndim=2)
search_dot: wp.array(dtype=wp.float32, ndim=1)
gauss: wp.array(dtype=wp.float32, ndim=1)
cost: wp.array(dtype=wp.float32, ndim=1)
cost_warmstart: wp.array(dtype=wp.float32, ndim=1)
prev_cost: wp.array(dtype=wp.float32, ndim=1)
solver_niter: wp.array(dtype=wp.int32, ndim=1)
active: wp.array(dtype=wp.int32, ndim=1)
Expand Down
Loading