Runtime error at ipopt example problem

[omersahintas]

Dear Bradley,

The execution of ipopt_solve_get_started.cpp during make check fails with the following error.

/usr/local/include/cppad/utility/nan.hpp:188:17: runtime error: division by zero
cppad-20210417 error from a known source:
vector: index greater than or equal vector size
Error detected by false result for
    i < length_
at line 360 in the file 
    /usr/local/include/cppad/utility/vector.hpp

Can you reproduce this error?

[bradbell]

What system are you using ?

I am using a Fedora 33 system with the master version
559c841

Using the following comamnds:

bin/get_optional.sh
bin/run_cmake.sh --debug_all
cd build
make check_example_ipopt_solve

The last command results in the following output:

... snip ...
[100%] Linking CXX executable example_ipopt_solve
[100%] Built target example_ipopt_solve
Begin test group example/ipopt_solve
get_started         OK
ode_inverse         OK
retape              OK
memory_leak         OK
All 4 tests passed.
End test group example/ipopt_solve
[100%] Built target check_example_ipopt_solve

[omersahintas]

I'm using Ubuntu 18.04.4 LTS. I have tried to compile both with gcc/g++ 8.4.0 and gcc/g++ 7.5.0, but got the same result.

I've further tried linking with both Ipopt 3.13.5 which is the current master head and latest stable release 3.13.3, but that didn't change anything.

[bradbell]

I fixed a problem when ipopt was the only optional package installed; see the heading 04-26 on
https://coin-or.github.io/CppAD/doc/whats_new_21.htm#04-26

No the following bash runs the tests in a docker image using ubuntu 18.04:

#! /bin/bash -e
cat << EOF > Dockerfile
From dokken/ubuntu-18.04
RUN apt-get update
RUN  DEBIAN_FRONTEND=noninteractive apt-get install -y  \
    cmake \
    git \
    gfortran \
    libblas-dev \
    liblapack-dev \
    pkg-config \
    vim \
    wget
RUN git clone https://github.com/coin-or/cppad.git cppad.git
RUN cd cppad.git && \
    bin/get_ipopt.sh
ENV PKG_CONFIG_PATH='/cppad.git/build/prefix/lib/pkgconfig'
RUN cd cppad.git/build && \
    cmake -D CMAKE_VERBOSE_MAKEFILE=1 -D include_ipopt=true .. && \
    make check_example_ipopt_solve
EOF
docker build --tag ubuntu_cppad .
echo 'test_ipopt.sh: OK'
exit 0

I get the following output when I run this script:

... snip ...
cd /cppad.git/build/example/ipopt_solve && ./example_ipopt_solve
Begin test group example/ipopt_solve
get_started         OK
ode_inverse         OK
retape              OK
memory_leak         OK
All 4 tests passed.
End test group example/ipopt_solve
... snip ...

Please try the corresponding commands on your system; e.g., use bin/get_ipopt.sh to install ipopt etc.

[omersahintas]

Thanks for your detailed instructions. It's my bad to install ipopt separately and not with the shell script you provided . By using get_ipopt.sh I get the correct results.

[bradbell]

You can change the install settings in
https://coin-or.github.io/CppAD/doc/get_ipopt.sh.htm
So that it installs ipopt with the prefix and version that you want.

You also should be able to use any valid ipopt options using your own script with bin/get_ipopt.sh as a staring point.

[omersahintas]

Thanks for the hints!

By the way, I see another runtime error when I run the example ipopt problem with gtest. I get

/usr/local/include/cppad/utility/nan.hpp:188:17: runtime error: division by zero

But, the optimization works fine and I get the correct results. I didn't encounter this error while executing make check.

To inspect this error more in detail, I've set print_level 10 and have run the test case with gdb, here is the output:

Running main() from gtest_main.cc
[==========] Running 1 test from 1 test case.
[----------] Global test environment set-up.
[----------] 1 test from CppAD
[ RUN      ] CppAD.IpoptExample
/usr/local/include/cppad/utility/nan.hpp:188:17: runtime error: division by zero

List of options:

                                    Name   Value                # times used
                         derivative_test = second-order              1
                                max_iter = 10                        1
               point_perturbation_radius = 0                         1
                             print_level = 10                        2
                                      sb = yes                       1
                                     tol = 1e-06                     2
This is Ipopt version 3.13.2, running with linear solver mumps.
NOTE: Other linear solvers might be more efficient (see Ipopt documentation).

Starting derivative checker for first derivatives.

Starting derivative checker for second derivatives.


No errors detected by derivative checker.
(...)

So it happens even before the derivative checker. Do you have an idea why this error occurs? I don't think this is something serious.

[bradbell]

The file
https://github.com/coin-or/CppAD/blob/master/include/cppad/utility/nan.hpp
uses zero / zero to generate a nan for any data type. Using this function was deprecated one 2015-10-04; see
https://coin-or.github.io/CppAD/doc/nan.htm#nan(zero).Deprecated%202015-10-04
Do you know which test this error is occurring in ?

[bradbell]

I am going to change this from a bug report to a discussion.

[omersahintas]

To inspect the source of this error, I have tried to catch it with SIGFPE, but that did not work well. I have added an assert(false) before return zero / zero; and backtraced with gdb:

#4  0x0000555555922cf9 in CppAD::nan<double> (zero=@0x7fffffffba60: 0) at /usr/local/include/cppad/utility/nan.hpp:190
#5  0x00005555558b0d3d in CppAD::ipopt::solve_callback<Eigen::Matrix<double, -1, 1, 0, -1, 1>, Eigen::Matrix<CppAD::AD<double>, -1, 1, 0, -1, 1>, (anonymous namespace)::FG_eval>::solve_callback (this=0x61b000000080, nf=1, nx=4, ng=2, xi=..., xl=..., xu=..., gl=..., gu=..., 
    fg_eval=..., retape=false, sparse_forward=false, sparse_reverse=false, solution=...)
    at /usr/local/include/cppad/ipopt/solve_callback.hpp:257

I obtain this log when I run ipopt_solve_get_started.cpp in gtest and call CppAD::ipopt::solve<Eigen::VectorXd, OptimizationProblem>(options_, xi_, xl_, xu_, gl_, gu_, optProblem_, solution);

Maybe UBSan finds an undefined behavior?

[bradbell]

I think I have removed the uses of the deprecated version of nan with the following commit:
fd95f9a

[omersahintas]

Yes, I can confirm that. That error doesn't occur anymore.

[omersahintas]

If the division in div.hpp returns a nan, how is it treated internally? Are there rules to catch certain types of exceptions?

[bradbell]

Perhaps the following examples make this clear ?
#89 (comment)
https://coin-or.github.io/CppAD/doc/cond_exp.cpp.htm