sparse ten_J

[thowell]

(always) sparse tendon Jacobian

tl;dr
performance is about the same for scenes with few tendons (eg, 2) and performance (steps per second and memory) is significantly better for a scene with many tendons (eg, 67)

---

myoarm (#937)

sps: 789,390 -> 952,616
memory (MiB): 1852.00 -> 1820.00

ntendon: 67 nwrap: 464

mjwarp-testspeed benchmark/myo_sim/arm/myoarm.xml --nworld=8192 --nconmax=16 --njmax=48 -o "opt.ccd_iterations=50" --event_trace --memory

this pr

Loading model from: benchmark/myo_sim/arm/myoarm.xml...

Model
  nq: 38 nv: 38 nu: 63 nbody: 40 ngeom: 161
Option
  integrator: EULER
  cone: PYRAMIDAL
  solver: NEWTON iterations: 100 ls_iterations: 50
  is_sparse: True
  ls_parallel: False
  broadphase: NXN broadphase_filter: PLANE|SPHERE|OBB
Data
  nworld: 8192 naconmax: 131072 njmax: 48
Rolling out 1000 steps at dt = 0.002...

Summary for 8192 parallel rollouts

Total JIT time: 0.51 s
Total simulation time: 8.60 s
Total steps per second: 952,616
Total realtime factor: 1,905.23 x
Total time per step: 1049.74 ns
Total converged worlds: 8192 / 8192

Event trace:

step: 1047.62
  forward: 922.01
    fwd_position: 489.25
      kinematics: 57.90
      com_pos: 15.00
      camlight: 1.93
      flex: 0.17
      tendon: 290.97
      crb: 25.08
      tendon_armature: 11.07
      collision: 12.40
        nxn_broadphase: 4.32
        convex_narrowphase: 5.90
        primitive_narrowphase: 1.28
      make_constraint: 31.15
      transmission: 41.66
    sensor_pos: 0.17
    fwd_velocity: 106.62
      com_vel: 7.16
      passive: 15.46
      rne: 13.90
      tendon_bias: 30.54
    sensor_vel: 0.17
    fwd_actuation: 20.11
    fwd_acceleration: 128.90
      xfrc_accumulate: 4.76
    solve: 174.80
      mul_m: 5.75
    sensor_acc: 0.17
  euler: 125.09

Model memory 17.28 MiB (0.95% of used memory):
 (no field >= 1% of used memory)
Data memory 796.09 MiB (43.74% of used memory):
 geom_xmat: 45.28 MiB (2.49%)
 site_xpos: 47.53 MiB (2.61%)
 site_xmat: 142.59 MiB (7.83%)
 ten_J: 37.56 MiB (2.06%)
 wrap_obj: 29.00 MiB (1.59%)
 wrap_xpos: 87.00 MiB (4.78%)
 actuator_moment: 74.81 MiB (4.11%)
 efc.J: 72.00 MiB (3.96%)
Other memory: 1006.64 MiB (55.31% of used memory)
Total memory: 1820.00 MiB (3.74% of total device memory)

0df83a0 + a put_data fix + tendon event scope

Loading model from: benchmark/myo_sim/arm/myoarm.xml...

Model
  nq: 38 nv: 38 nu: 63 nbody: 40 ngeom: 161
Option
  integrator: EULER
  cone: PYRAMIDAL
  solver: NEWTON iterations: 100 ls_iterations: 50
  is_sparse: True
  ls_parallel: False
  broadphase: NXN broadphase_filter: PLANE|SPHERE|OBB
Data
  nworld: 8192 naconmax: 131072 njmax: 48
Rolling out 1000 steps at dt = 0.002...

Summary for 8192 parallel rollouts

Total JIT time: 0.51 s
Total simulation time: 10.38 s
Total steps per second: 789,390
Total realtime factor: 1,578.78 x
Total time per step: 1266.80 ns
Total converged worlds: 8192 / 8192

Event trace:

step: 1264.61
  forward: 1138.18
    fwd_position: 681.06
      kinematics: 58.03
      com_pos: 14.95
      camlight: 1.91
      flex: 0.17
      tendon: 448.51
      crb: 25.24
      tendon_armature: 11.46
      collision: 12.39
        nxn_broadphase: 4.28
        convex_narrowphase: 5.92
        primitive_narrowphase: 1.29
      make_constraint: 31.60
      transmission: 74.88
    sensor_pos: 0.17
    fwd_velocity: 132.26
      com_vel: 7.35
      passive: 15.30
      rne: 13.28
      tendon_bias: 29.89
    sensor_vel: 0.18
    fwd_actuation: 20.58
    fwd_acceleration: 126.29
      xfrc_accumulate: 4.75
    solve: 175.65
      mul_m: 5.76
    sensor_acc: 0.17
  euler: 125.91

Model memory 17.27 MiB (0.93% of used memory):
 (no field >= 1% of used memory)
Data memory 838.09 MiB (45.25% of used memory):
 geom_xmat: 45.28 MiB (2.44%)
 site_xpos: 47.53 MiB (2.57%)
 site_xmat: 142.59 MiB (7.70%)
 ten_J: 79.56 MiB (4.30%)
 wrap_obj: 29.00 MiB (1.57%)
 wrap_xpos: 87.00 MiB (4.70%)
 actuator_moment: 74.81 MiB (4.04%)
 efc.J: 72.00 MiB (3.89%)
Other memory: 996.64 MiB (53.81% of used memory)
Total memory: 1852.00 MiB (3.81% of total device memory)

---

mujoco/model/humanoid/humanoid.xml
https://github.com/google-deepmind/mujoco/blob/main/model/humanoid/humanoid.xml

sps: 3,786,138 -> 3,786,812
memory (MiB): 402.00 -> 402.00

ntendon: 2 nwrap: 4

mjwarp-testspeed mujoco/model/humanoid/humanoid.xml --nworld=8192 --nconmax=24 --njmax=64 --event_trace --memory

this pr

Loading model from: mujoco/model/humanoid/humanoid.xml...

Model
  nq: 28 nv: 27 nu: 21 nbody: 17 ngeom: 20
Option
  integrator: EULER
  cone: PYRAMIDAL
  solver: NEWTON iterations: 100 ls_iterations: 50
  is_sparse: False
  ls_parallel: False
  broadphase: NXN broadphase_filter: PLANE|SPHERE|OBB
Data
  nworld: 8192 naconmax: 196608 njmax: 64
Rolling out 1000 steps at dt = 0.005...

Summary for 8192 parallel rollouts

Total JIT time: 0.32 s
Total simulation time: 2.16 s
Total steps per second: 3,786,812
Total realtime factor: 18,934.06 x
Total time per step: 264.07 ns
Total converged worlds: 8192 / 8192

Event trace:

step: 262.33
  forward: 252.52
    fwd_position: 84.97
      kinematics: 9.94
      com_pos: 5.67
      camlight: 1.77
      flex: 0.17
      tendon: 0.97
      crb: 13.30
      tendon_armature: 0.73
      collision: 9.84
        nxn_broadphase: 3.60
        convex_narrowphase: 0.17
        primitive_narrowphase: 5.11
      make_constraint: 38.31
      transmission: 2.34
    sensor_pos: 0.17
    fwd_velocity: 28.13
      com_vel: 5.93
      passive: 1.67
      rne: 7.08
      tendon_bias: 1.57
    sensor_vel: 0.18
    fwd_actuation: 2.28
    fwd_acceleration: 11.21
      xfrc_accumulate: 1.59
    solve: 123.55
      mul_m: 6.08
    sensor_acc: 0.17
  euler: 9.29

Model memory 0.03 MiB (0.01% of used memory):
 (no field >= 1% of used memory)
Data memory 258.50 MiB (64.30% of used memory):
 xmat: 4.78 MiB (1.19%)
 ximat: 4.78 MiB (1.19%)
 geom_xmat: 5.62 MiB (1.40%)
 cdof: 5.06 MiB (1.26%)
 cinert: 5.31 MiB (1.32%)
 actuator_moment: 17.72 MiB (4.41%)
 crb: 5.31 MiB (1.32%)
 qM: 24.50 MiB (6.09%)
 qLD: 22.78 MiB (5.67%)
 cdof_dot: 5.06 MiB (1.26%)
 contact.frame: 6.75 MiB (1.68%)
 efc.J: 56.00 MiB (13.93%)
Other memory: 143.47 MiB (35.69% of used memory)
Total memory: 402.00 MiB (0.83% of total device memory)

0df83a0 + a put_data fix + tendon event scope

Loading model from: mujoco/model/humanoid/humanoid.xml...

Model
  nq: 28 nv: 27 nu: 21 nbody: 17 ngeom: 20
Option
  integrator: EULER
  cone: PYRAMIDAL
  solver: NEWTON iterations: 100 ls_iterations: 50
  is_sparse: False
  ls_parallel: False
  broadphase: NXN broadphase_filter: PLANE|SPHERE|OBB
Data
  nworld: 8192 naconmax: 196608 njmax: 64
Rolling out 1000 steps at dt = 0.005...

Summary for 8192 parallel rollouts

Total JIT time: 0.77 s
Total simulation time: 2.16 s
Total steps per second: 3,786,138
Total realtime factor: 18,930.69 x
Total time per step: 264.12 ns
Total converged worlds: 8192 / 8192

Event trace:

step: 262.44
  forward: 252.63
    fwd_position: 84.31
      kinematics: 9.86
      com_pos: 5.67
      camlight: 1.76
      flex: 0.17
      tendon: 0.89
      crb: 13.33
      tendon_armature: 0.70
      collision: 9.71
        nxn_broadphase: 3.56
        convex_narrowphase: 0.17
        primitive_narrowphase: 5.02
      make_constraint: 38.01
      transmission: 2.30
    sensor_pos: 0.17
    fwd_velocity: 28.96
      com_vel: 5.93
      passive: 1.67
      rne: 7.07
      tendon_bias: 1.56
    sensor_vel: 0.17
    fwd_actuation: 2.28
    fwd_acceleration: 11.21
      xfrc_accumulate: 1.58
    solve: 123.49
      mul_m: 6.09
    sensor_acc: 0.18
  euler: 9.29

Model memory 0.03 MiB (0.01% of used memory):
 (no field >= 1% of used memory)
Data memory 260.06 MiB (64.69% of used memory):
 xmat: 4.78 MiB (1.19%)
 ximat: 4.78 MiB (1.19%)
 geom_xmat: 5.62 MiB (1.40%)
 cdof: 5.06 MiB (1.26%)
 cinert: 5.31 MiB (1.32%)
 actuator_moment: 17.72 MiB (4.41%)
 crb: 5.31 MiB (1.32%)
 qM: 24.50 MiB (6.09%)
 qLD: 22.78 MiB (5.67%)
 cdof_dot: 5.06 MiB (1.26%)
 contact.frame: 6.75 MiB (1.68%)
 efc.J: 56.00 MiB (13.93%)
Other memory: 141.91 MiB (35.30% of used memory)
Total memory: 402.00 MiB (0.83% of total device memory)

[quagla]

Nice improvement!