Closed cmastalli closed 4 years ago
-march=native is changing the memory allocator in Eigen. So, if all the libraries that are using Eigen and are also Pinocchio's dependencies, then they should be compiled with the same compilations flags when allowing vectorization and so on.
Please redo your test by not considering hpp-fcl
.
I think I already tried to compile both hpp-fcl and pinocchio with -march=native
and I encountered alignment issues. This was a few months ago so I don't remember the details and I may have made a mistake when testing it.
I suspect there is a missing EIGEN_MAKE_ALIGNED_OPERATOR_NEW
somewhere.
I suspect there is a missing
EIGEN_MAKE_ALIGNED_OPERATOR_NEW
somewhere.
This should be in hpp-fcl, not in Pinocchio then.
On my side, everything works fine with alignment and so, and without setting hpp-fcl. By the way, @cmastalli it seems that you're linking to an installed pinocchio lib. Be sure that this is the correct one.
I suspect there is a missing
EIGEN_MAKE_ALIGNED_OPERATOR_NEW
somewhere.This should be in hpp-fcl, not in Pinocchio then.
Or in the part of Pinocchio which is enabled when compiled with hpp-fcl. At least, all the unit test in hpp-fcl were ok with native flags.
The Python bindings need alignment to work properly. So, in this way, it has been checked.
And I've checked twice, and all the geometry containers of Pinocchio have EIGEN_MAKE_ALIGNED_OPERATOR_NEW
.
Let see if solving humanoid-path-planner/hpp-fcl#166 fixes the issue in Pinocchio.
The Python bindings need alignment to work properly. So, in this way, it has been checked.
Isn't this the same for hpp-fcl ?
I don't think so because you are using shared_ptr to wrap them. This is not the case for Pinocchio, where Python is managing the memory.
I must have been fooled by the linker too when I tried it. I can't reproduce the issue I had and the unit test passes.
Please redo your test by not considering
hpp-fcl
.
If I compile Pinocchio without hpp-fcl
, then the unittest are passing.
I must have been fooled by the linker too when I tried it. I can't reproduce the issue I had and the unit test passes.
This happens to me with hpp-fcl
. Then I didn't decide to install it with -march=native
which is strictly necessary for compiling Pinocchio with collision support. @jcarpent there is no issue in hpp-fcl
, neither in Pinocchio. Again this comes due to linking to installed libraries.
Note that the unittests with collision support are passing as well.
@jcarpent @jmirabel thank you so much for your time. I didn't know how important is to uninstall the libraries.
I didn't know how important is to uninstall the libraries.
This is new (from ubuntu 18.04). This issue is very annoying and even if I am aware of it, I am regularly trapped. It should be fixed in the JRL cmake module.
I have compiled the latest version of Pinocchio with all the options enables (with the exception of CasADi). When I run the tests, I encountered the follows:
I decided to inspect a little the
test-cpp-urdf
usinggdb
. It seems there is an issue with the geometry stuff. This is the messagein particular, the problem seems to be in
details::parseTreeForGeom
.Note that I didn't try to compile
hpp-fcl
using-march=native
since the unittest are not passing (for more details see https://github.com/humanoid-path-planner/hpp-fcl/issues/166)