crba: avoid temporaries with .noalias()

[alaurenzi]

Dear maintainers,
for a few projects we'd like to avoid allocating memory inside function running inside a real time loop.

I have found out that a few temporaries are generated inside the crba function. Adding a .noalias() here and there was enough to keep Eigen from generating temporaries during matrix multiplications.

Let me know what you think!

Arturo

[ManifoldFR]

Hi,

Thanks for reaching out!
You found that a malloc happened at the lines you modified ? I thought the involved matrices had fixed Eigen sizes.

[alaurenzi]

For some of them you may be right.. I belive the problematic line is

data.M.block(jmodel.idx_v(),jmodel.idx_v(),jmodel.nv(),data.nvSubtree[i]).noalias() = ... * ...

where the LHS block has dynamic size.

[ManifoldFR]

You're right, the LHS is a dynamic-sized Eigen expression.

Have you run the tests/benchmarks with this change and the CHECK_RUNTIME_MALLOC flag (I think you need a Debug build type for that)?

[jcarpent]

@alaurenzi Thanks for reaching us. Could you provide a robot model that triggers allocation?
It would be indeed required to check on our side.

[alaurenzi]

Hi @jcarpent , please have a look to this minimal self contained example: pinocchio_malloc_test.zip

• wget https://github.com/stack-of-tasks/pinocchio/files/13740137/pinocchio_malloc_test.zip
• unzip pinocchio_malloc_test.zip
• cd pinocchio_malloc_test
• bash run.bash
• The program will compile and execute under gdb. It is supposed to crash on malloc, then bt to inspect backtrace

On my machine the interesting line is

#16 0x00005555555acb3b in pinocchio::CrbaBackwardStep<double, 0, pinocchio::JointCollectionDefaultTpl>::algo<pinocchio::JointModelRevoluteUnalignedTpl<double, 0> > (jmodel=..., jdata=..., mo
del=..., data=...)
    at /home/alaurenzi/code/core_ws/install/include/pinocchio/algorithm/crba.hxx:81

which corresponds to

/* M[i,SUBTREE] = S'*F[1:6,SUBTREE] */
      data.M.block(jmodel.idx_v(),jmodel.idx_v(),jmodel.nv(),data.nvSubtree[i]) 
      = jdata.S().transpose()*data.Fcrb[i].middleCols(jmodel.idx_v(),data.nvSubtree[i]);

Adding .noalias() fixes it.

[jcarpent]

See #2126 for the final solving + test

[jcarpent]

@alaurenzi I've provided a clean revision in #2126 with a unit test. Could you check on your side if it does what you expect?