A few tests fail with missing dependencies even when the precompiled snopt is present

[RussTedrake]

http://drake006.csail.mit.edu/drake/pull-RussTedrake-snopt_precompiled_for_all-into-RobotLocomotion-master_20151008-1111-Continuous

There are 5 that list missing dependency 'SNOPT' and one that lists 'snopt'

To reproduce, simply run the build using WITH_SNOPT off and WITH_SNOPT_PRECOMPILED on.

[RussTedrake]

@hongkai-dai - could you take a look?

[hongkai-dai]

Yes, will look into it

On Thu, Oct 8, 2015 at 10:29 AM, Russ Tedrake [email protected]
wrote:

@hongkai-dai https://github.com/hongkai-dai - could you take a look?

—
Reply to this email directly or view it on GitHub
#1402 (comment)
.

[hongkai-dai]

This is a problem with the old inverseKinTraj.m function. The other IK functions (InverseKinematics.m, inverseKin.cpp, etc) either has C++ implementation directly (inverseKin.cpp, inverseKinTraj.cpp), or they are derived from NonlinearProgram.m, so they can call NonlinearProgramsnoptmex. But inverseKinTraj.m still calls inverseKinBackend.m, which calls MATLAB snopt directly. This approach does not work with precompiled snopt.

As our goal is to get rid of the RigidBodyConstraint class, and uses only Constraint class and NonlinearProgram as the interface for IK, I suggest not to test the inverseKinTraj.m function when we are using the precompiled version of snopt.

[hongkai-dai]

The reason why inverseKinTraj.m, inverseKinPointwise.m and inverseKin.m are looking for the MATLAB version of inverseKinBackend.m, is that we do not compile inverseKinmex, inverseKinTrajmex and inverseKinPointwisemex with SNOPT. Is it possible that we can precompile these three files also, as we did for NonlinearProgramSnoptmex?

[RussTedrake]

as discussed, let's just fix the easy ones first. we eventually want to distribute the drake libs which depend on libsnoptc , but that will take longer I know.

[RussTedrake]

this will be resolved by #245 and the eventual move to everything running through our c++ solver wrappers.