MuJoCo Warp documentation: Memory.
PiperOrigin-RevId: 857148640 Change-Id: I7f21bebb2186a5dd4287a8d2ee9b430f823b9051
This commit is contained in:
committed by
Copybara-Service
parent
77d1ac10fe
commit
9ba545b8b0
+35
-1
@@ -220,7 +220,7 @@ Minimal example
|
|||||||
Command line scripts
|
Command line scripts
|
||||||
--------------------
|
--------------------
|
||||||
|
|
||||||
Benchmark an environment with testspeed
|
Benchmark an environment with _`testspeed`
|
||||||
|
|
||||||
.. code-block:: shell
|
.. code-block:: shell
|
||||||
|
|
||||||
@@ -379,6 +379,40 @@ To enable this routine set ``Model.opt.ls_parallel=True`` or add a custom numeri
|
|||||||
|
|
||||||
.. _mjwBatch:
|
.. _mjwBatch:
|
||||||
|
|
||||||
|
|
||||||
|
Memory
|
||||||
|
------
|
||||||
|
|
||||||
|
Simulation throughput is often limited by memory requirements for large numbers of worlds. Considerations for optimizing
|
||||||
|
memory utilization include:
|
||||||
|
|
||||||
|
- CCD colliders require more memory than primitive colliders, see MuJoCo's :ref:`pair-wise colliders table <coPairwise>`
|
||||||
|
for information about colliders.
|
||||||
|
- :ref:`multiccd <option-flag-multiccd>` requires more memory than CCD.
|
||||||
|
- CCD memory requirements scale linearly with :ref:`Option.ccd_iterations <option-ccd-iterations>`.
|
||||||
|
- A scene with at least one mesh geom and using :ref:`multiccd <option-flag-multiccd>` will have memory requirements
|
||||||
|
that scale linearly with the maximum number of vertices per face and with the maximum number of edges per vertex,
|
||||||
|
computed over all meshes.
|
||||||
|
|
||||||
|
`testspeed`_ provides the flag ``--memory`` for reporting a simulation's total memory utilization and information about
|
||||||
|
:class:`mjw.Model <mujoco_warp.Model>` and :class:`mjw.Data <mujoco_warp.Data>` fields that require significant memory.
|
||||||
|
Memory allocated inline, including for CCD and the constraint solver, can also be significant and is reported as
|
||||||
|
``Other memory``.
|
||||||
|
|
||||||
|
.. admonition:: Maximum number of contacts per collider
|
||||||
|
:class: note
|
||||||
|
|
||||||
|
Some MJWarp colliders have a different maximum number of contacts compared to MuJoCo:
|
||||||
|
|
||||||
|
- ``PLANE<>MESH``: 4 versus 3
|
||||||
|
- ``HFieldCCD``: 4 versus ``mjMAXCONPAIR``
|
||||||
|
|
||||||
|
.. admonition:: Sparsity
|
||||||
|
:class: note
|
||||||
|
|
||||||
|
Sparse Jacobians can enable significant memory savings. Updates for this feature are tracked in GitHub issue
|
||||||
|
`#88 <https://github.com/google-deepmind/mujoco_warp/issues/88>`__.
|
||||||
|
|
||||||
Batched :class:`Model <mujoco_warp.Model>` Fields
|
Batched :class:`Model <mujoco_warp.Model>` Fields
|
||||||
=================================================
|
=================================================
|
||||||
|
|
||||||
|
|||||||
Reference in New Issue
Block a user