1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
|
use super::{DeltaVel, ParallelInteractionGroups, ParallelVelocitySolver};
use crate::dynamics::solver::ParallelPositionSolver;
use crate::dynamics::{IntegrationParameters, JointGraphEdge, JointIndex, RigidBodySet};
use crate::geometry::{ContactManifold, ContactManifoldIndex};
use crate::math::Isometry;
use crate::utils::WAngularInertia;
use rayon::Scope;
use std::sync::atomic::{AtomicUsize, Ordering};
#[macro_export]
#[doc(hidden)]
macro_rules! concurrent_loop {
(let batch_size = $batch_size: expr;
for $elt: ident in $array: ident[$index_stream:expr,$index_count:expr] $f: expr) => {
let max_index = $array.len();
if max_index > 0 {
loop {
let start_index = $index_stream.fetch_add($batch_size, Ordering::SeqCst);
if start_index > max_index {
break;
}
let end_index = (start_index + $batch_size).min(max_index);
for $elt in &$array[start_index..end_index] {
$f
}
$index_count.fetch_add(end_index - start_index, Ordering::SeqCst);
}
}
};
(let batch_size = $batch_size: expr;
for $elt: ident in $array: ident[$index_stream:expr] $f: expr) => {
let max_index = $array.len();
if max_index > 0 {
loop {
let start_index = $index_stream.fetch_add($batch_size, Ordering::SeqCst);
if start_index > max_index {
break;
}
let end_index = (start_index + $batch_size).min(max_index);
for $elt in &$array[start_index..end_index] {
$f
}
}
}
};
}
pub(crate) struct ThreadContext {
pub batch_size: usize,
// Velocity solver.
pub constraint_initialization_index: AtomicUsize,
pub num_initialized_constraints: AtomicUsize,
pub joint_constraint_initialization_index: AtomicUsize,
pub num_initialized_joint_constraints: AtomicUsize,
pub warmstart_contact_index: AtomicUsize,
pub num_warmstarted_contacts: AtomicUsize,
pub warmstart_joint_index: AtomicUsize,
pub num_warmstarted_joints: AtomicUsize,
pub solve_interaction_index: AtomicUsize,
pub num_solved_interactions: AtomicUsize,
pub impulse_writeback_index: AtomicUsize,
pub joint_writeback_index: AtomicUsize,
pub body_integration_index: AtomicUsize,
pub num_integrated_bodies: AtomicUsize,
// Position solver.
pub position_constraint_initialization_index: AtomicUsize,
pub num_initialized_position_constraints: AtomicUsize,
pub position_joint_constraint_initialization_index: AtomicUsize,
pub num_initialized_position_joint_constraints: AtomicUsize,
pub solve_position_interaction_index: AtomicUsize,
pub num_solved_position_interactions: AtomicUsize,
pub position_writeback_index: AtomicUsize,
}
impl ThreadContext {
pub fn new(batch_size: usize) -> Self {
ThreadContext {
batch_size, // TODO perhaps there is some optimal value we can compute depending on the island size?
constraint_initialization_index: AtomicUsize::new(0),
num_initialized_constraints: AtomicUsize::new(0),
joint_constraint_initialization_index: AtomicUsize::new(0),
num_initialized_joint_constraints: AtomicUsize::new(0),
num_warmstarted_contacts: AtomicUsize::new(0),
warmstart_contact_index: AtomicUsize::new(0),
num_warmstarted_joints: AtomicUsize::new(0),
warmstart_joint_index: AtomicUsize::new(0),
solve_interaction_index: AtomicUsize::new(0),
num_solved_interactions: AtomicUsize::new(0),
impulse_writeback_index: AtomicUsize::new(0),
joint_writeback_index: AtomicUsize::new(0),
body_integration_index: AtomicUsize::new(0),
num_integrated_bodies: AtomicUsize::new(0),
position_constraint_initialization_index: AtomicUsize::new(0),
num_initialized_position_constraints: AtomicUsize::new(0),
position_joint_constraint_initialization_index: AtomicUsize::new(0),
num_initialized_position_joint_constraints: AtomicUsize::new(0),
solve_position_interaction_index: AtomicUsize::new(0),
num_solved_position_interactions: AtomicUsize::new(0),
position_writeback_index: AtomicUsize::new(0),
}
}
pub fn lock_until_ge(val: &AtomicUsize, target: usize) {
if target > 0 {
// let backoff = crossbeam::utils::Backoff::new();
std::sync::atomic::fence(Ordering::SeqCst);
while val.load(Ordering::Relaxed) < target {
// backoff.spin();
// std::thread::yield_now();
}
}
}
}
pub struct ParallelIslandSolver {
mj_lambdas: Vec<DeltaVel<f32>>,
positions: Vec<Isometry<f32>>,
parallel_groups: ParallelInteractionGroups,
parallel_joint_groups: ParallelInteractionGroups,
parallel_velocity_solver: ParallelVelocitySolver,
parallel_position_solver: ParallelPositionSolver,
thread: ThreadContext,
}
impl ParallelIslandSolver {
pub fn new() -> Self {
Self {
mj_lambdas: Vec::new(),
positions: Vec::new(),
parallel_groups: ParallelInteractionGroups::new(),
parallel_joint_groups: ParallelInteractionGroups::new(),
parallel_velocity_solver: ParallelVelocitySolver::new(),
parallel_position_solver: ParallelPositionSolver::new(),
thread: ThreadContext::new(8),
}
}
pub fn solve_island<'s>(
&'s mut self,
scope: &Scope<'s>,
island_id: usize,
params: &'s IntegrationParameters,
bodies: &'s mut RigidBodySet,
manifolds: &'s mut Vec<&'s mut ContactManifold>,
manifold_indices: &'s [ContactManifoldIndex],
joints: &'s mut Vec<JointGraphEdge>,
joint_indices: &[JointIndex],
) {
let num_threads = rayon::current_num_threads();
let num_task_per_island = num_threads; // (num_threads / num_islands).max(1); // TODO: not sure this is the best value. Also, perhaps it is better to interleave tasks of each island?
self.thread = ThreadContext::new(8); // TODO: could we compute some kind of optimal value here?
self.parallel_groups
.group_interactions(island_id, bodies, manifolds, manifold_indices);
self.parallel_joint_groups
.group_interactions(island_id, bodies, joints, joint_indices);
self.parallel_velocity_solver.init_constraint_groups(
island_id,
bodies,
manifolds,
&self.parallel_groups,
joints,
&self.parallel_joint_groups,
);
self.parallel_position_solver.init_constraint_groups(
island_id,
bodies,
manifolds,
&self.parallel_groups,
joints,
&self.parallel_joint_groups,
);
self.mj_lambdas.clear();
self.mj_lambdas
.resize(bodies.active_island(island_id).len(), DeltaVel::zero());
self.positions.clear();
self.positions
.resize(bodies.active_island(island_id).len(), Isometry::identity());
for _ in 0..num_task_per_island {
// We use AtomicPtr because it is Send+Sync while *mut is not.
// See https://internals.rust-lang.org/t/shouldnt-pointers-be-send-sync-or/8818
let thread = &self.thread;
let mj_lambdas = std::sync::atomic::AtomicPtr::new(&mut self.mj_lambdas as *mut _);
let positions = std::sync::atomic::AtomicPtr::new(&mut self.positions as *mut _);
let bodies = std::sync::atomic::AtomicPtr::new(bodies as *mut _);
let manifolds = std::sync::atomic::AtomicPtr::new(manifolds as *mut _);
let joints = std::sync::atomic::AtomicPtr::new(joints as *mut _);
let parallel_velocity_solver =
std::sync::atomic::AtomicPtr::new(&mut self.parallel_velocity_solver as *mut _);
let parallel_position_solver =
std::sync::atomic::AtomicPtr::new(&mut self.parallel_position_solver as *mut _);
scope.spawn(move |_| {
// Transmute *mut -> &mut
let mj_lambdas: &mut Vec<DeltaVel<f32>> =
unsafe { std::mem::transmute(mj_lambdas.load(Ordering::Relaxed)) };
let positions: &mut Vec<Isometry<f32>> =
unsafe { std::mem::transmute(positions.load(Ordering::Relaxed)) };
let bodies:
|
