- mjTIMER_ACCELERATION is not very useful, it's usually very tiny (0.5% for humanoid, 1% for 22humanoids). It basically times a single call to `mj_solveM`.
- Added `mjTIMER_ADVANCE` to time `mj_Euler` and `mj_implicit` (which can be significant).
- Moved `mj_transmission` to `mjTIMER_KINEMATICS` where it belongs (and used to be).
PiperOrigin-RevId: 565936880
Change-Id: I26b8a067133594e5500ca1c809beb3518b4a8d9c
- See public description below.
- Stopped incrementing the iteration count in saveStats().
- `sizeof(mjSolverStat) == 40`, so this ends up costing 160KB, up from 40KB.
BEGIN_PUBLIC
Changed the size of `mjData.solver`, the structure used to collect solver diagnostic information. The array is now of length `mjNISLAND * mjNSOLVER`, where each row of length `mjNSOLVER` contains separate solver statistics for each constraint island. Until solver islanding is implemented, only row 0 is used.
- The new constant `mjNISLAND` is set to 20.
- `mjNSOLVER` is reduced from 1000 to 200.
- Added `mjData.solver_nisland`, the number of islands for which the solver ran.
- `mjData.solver_niter` (renamed from mjData.solver_iter) and `mjData.solver_nnz` are now integer vectors of length `mjNISLAND`.
END_PUBLIC
PiperOrigin-RevId: 565030093
Change-Id: I773e918805c6ced79f0dab5f19ea23956760c8c5
- Make mjTask non-opaque and remove C++ Task class.
- Make C++ thread pool a subclass of a skeletal mjThreadPool C struct.
- Make the mjTask status enum more consistent with the rest of MuJoCo.
- Change mju_threadPoolEnqueue to take just the mjTask. Users must now prepare the mjTask by assigning the function pointer and argument into the struct.
- Rename files in thread/ to be more consistent with the rest of MuJoCo.
- Run threading tests in CMake.
- Allow use of C++20 designated initializers.
Otherwise the functionality remains identical.
PiperOrigin-RevId: 564374675
Change-Id: I37c9894566bc39faf217e5aa97a4e2713a70e467
std::function can have an allocation cost when it's created, and also adds overhead to function calls. Since this API must be used from C, using a raw function pointer is good enough.
PiperOrigin-RevId: 562760324
Change-Id: I2b714a13e07a2b71566d013f21f76c506e817a52
Also add asan instrumentation to detect stack frame leakages (i.e. `mj_markStack` without a corresponding `mj_freeStack` in the same caller function).
PiperOrigin-RevId: 562625645
Change-Id: I4e3ff66ca0b9d08ed0a95cef45393db8e3053e22
This implementation explicitly avoids calling division instructions (which are generally expensive) even when link-time optimization is not active or if the alloc functions aren't inlined for some reason.
Also make mj_stackAlloc cheaper for external callers who may be linking against MuJoCo dynamically (and therefore cannot benefit from inlining).
PiperOrigin-RevId: 561299565
Change-Id: Iea4be5bf6d351ddf0c9d1fd8bd89a0c207b39309
Also fix bug in allocation for constraint islands, added allocation for limit constraints.
PiperOrigin-RevId: 561083128
Change-Id: I9e01d5a5b7454ec3cf60bccadaa0f96f2c0d87d1
At the leaves of the midphase trees in `mj_collideTree`, a bounding sphere test is performed. The margin was incorrectly set to zero instead of reading it from `geom_margin`.
PiperOrigin-RevId: 559782262
Change-Id: Icfceb8445bbf6977582e0061566d7bd4def2f0ae
Required for constraint solver islanding.
Replace linked lists with `island_{dof,efc}_{num,adr,ind}`, corresponding to the standard `{rownnz,rowadr,colind}` sparse matrix representation. By effectively defining two sparse matrix structures of dimensions `nisland x nv` and `nisland x nefc`, respectively, this representation is more conducive to reuse of existing sparse matrix utility functions, while being cache-friendlier by making sequential indices adjacent in memory.
PiperOrigin-RevId: 559704957
Change-Id: I919362cbef0d5fe5acc4aa2bb4ffb17eee8223a2
Use only the first `mjMAXCONPAIR` contacts in SDF-mesh collisions after sorting them using the penetration depth.
PiperOrigin-RevId: 558176217
Change-Id: Ia1ee12bd5ceb81e4c52f143831ad00426e8cf174
Yuval Tassa