test-cpp-contact-cholesky failure with GCC 13.3.0 #2304
Assignees
Comments
|
Same version of Eigen? |
|
yes |
13 tasks
|
And, only on aarch64: Not sure if related to #2277 or not, but we'll deactivate also this one, for aarch64, for now. |
|
I would check if we're initializing all matrices/vectors with zeros. |
|
I think I can confirm this is an issue in GCC 13.3.0 with this reproducer: FROM debian
RUN --mount=type=cache,sharing=locked,target=/var/cache/apt \
--mount=type=cache,sharing=locked,target=/var/lib/apt \
apt-get update -y && DEBIAN_FRONTEND=noninteractive apt-get install -qqy --no-install-recommends \
autoconf \
build-essential \
bzip2 \
cmake \
g++-multilib \
gcc-multilib \
git \
libboost-all-dev \
libeigen3-dev \
liburdfdom-dev \
libtinyxml-dev \
make \
wget \
xz-utils
ARG GCC_VERSION=13.2.0
ENV GCC_VERSION=$GCC_VERSION
WORKDIR /src
ADD https://gmplib.org/download/gmp/gmp-6.3.0.tar.xz .
ADD https://www.mpfr.org/mpfr-current/mpfr-4.2.1.tar.xz .
ADD https://ftp.gnu.org/gnu/mpc/mpc-1.3.1.tar.gz .
ADD https://gcc.gnu.org/pub/gcc/infrastructure/isl-0.24.tar.bz2 .
RUN tar xf gmp* \
&& tar xf mpfr* \
&& tar xf mpc* \
&& tar xf isl*
ADD https://gcc.gnu.org/pub/gcc/releases/gcc-$GCC_VERSION/gcc-$GCC_VERSION.tar.xz .
RUN tar xf gcc-$GCC_VERSION.tar.xz
WORKDIR gcc-$GCC_VERSION
RUN mv ../gmp-6.3.0 gmp \
&& mv ../mpfr-4.2.1 mpfr \
&& mv ../mpc-1.3.1 mpc \
&& mv ../isl-0.24 isl
RUN ./configure --enable-languages=c,c++ \
&& make -j 8 \
&& make -j 8 install
WORKDIR /src
ADD https://github.com/stack-of-tasks/pinocchio/releases/download/v3.0.0/pinocchio-3.0.0.tar.gz .
RUN tar xf pinocchio-3.0.0.tar.gz
WORKDIR pinocchio-3.0.0
RUN cmake -B build -S . -DCMAKE_BUILD_TYPE=Release -DBUILD_PYTHON_INTERFACE=OFF -DCMAKE_CXX_STANDARD=14
RUN cmake --build build -j 4
ENV LD_LIBRARY_PATH=/usr/local/lib64
#CMD ./build/unittest/test-cpp-contact-choleskyrunning |
|
the top right corners in l. 359 start with: vs |
|
Taking another look at this, it seems we have a RUN cmake -B build -S . -DCMAKE_BUILD_TYPE=Release -DBUILD_PYTHON_INTERFACE=OFF -DCMAKE_CXX_STANDARD=14
RUN sed -i 's/-O3/-O2/' build/CMakeCache.txt
RUN cmake --build build -j 4 -t test-cpp-contact-cholesky
ENV LD_LIBRARY_PATH=/usr/local/lib64
RUN ./build/unittest/test-cpp-contact-cholesky
Digging further, looking at https://gcc.gnu.org/onlinedocs/gcc/Optimize-Options.html#index-O3, I tried to replace the failing The good solution would obviously be to trash all our C/C++ code base and start again in a sane language. But this might not be on the roadmap for now, so I guess I'll just go for |
|
Conda-forge is now using gcc 13.3 as default and we can now reproduce this issue. Here all the failing tests:
|
|
As stated here -On is not equivalent to using all the -f optimization flag. There is some unamed optimization activated by the -On options. |
|
Ok, thanks. It looks once again like a GCC issue then. But to report that we should probably work on a MRE. For that we can try to dump the matrices used in one of those failing tests. |
|
I'm working on it :) |
|
Note for tomorrow:
|
|
I found some piece of code that run differently between O2 and O3 with g++ 13.3 : https://github.com/stack-of-tasks/pinocchio/blob/master/include/pinocchio/algorithm/contact-info.hpp#L747-L750 for (int k = 0; k < joint1.nv(); ++k, ++current1_col_id)
{
colwise_joint1_sparsity[current1_col_id] = true;
}This loop doesn't run in O3 for the first joint. Then, colwise_joint1_sparsity is not well initialized. Also, if I remove the following else clause the code is also working in O2 and O3 (in the case the else clause is not needed). else
{
const JointModel & joint2 = model.joints[current2_id];
joint2_span_indexes.push_back((Eigen::DenseIndex)current2_id);
Eigen::DenseIndex current2_col_id = joint2.idx_v();
for (int k = 0; k < joint2.nv(); ++k, ++current2_col_id)
{
colwise_joint2_sparsity[current2_col_id] = true;
}
current2_id = model.parents[current2_id];
}@nim65s, @jcarpent do you see a particular undefined behavior that can explain why G++ don't run the last iteration of the for loop when the else clause is here ? I think I have three options:
I think I will try the last option first, but I'm afraid it will not be so easy, maybe the optimization is activated because of some previous code (contact-info.hpp code is all inline). |
|
As I feared it will be difficult to have a MRE. The following doesn't reproduce the bug. MRE.cpp #include <cstddef>
#include <iostream>
#include <Eigen/Core>
#include <vector>
extern std::size_t JOINT1_ID;
extern std::size_t JOINT2_ID;
extern int NV;
extern int NJOINT;
extern int JOINT_IDX_V[];
extern int JOINT_NV[];
extern int JOINT_PARENTS[];
void printVector(const Eigen::VectorX<bool>& v);
int main()
{
Eigen::VectorX<bool> colwise_joint1_sparsity(NV);
Eigen::VectorX<bool> colwise_joint2_sparsity(NV);
std::vector<int> joint1_span_indexes;
joint1_span_indexes.reserve(NJOINT);
std::vector<int> joint2_span_indexes;
joint2_span_indexes.reserve(NJOINT);
static const bool default_sparsity_value = false;
colwise_joint1_sparsity.fill(default_sparsity_value);
colwise_joint2_sparsity.fill(default_sparsity_value);
std::size_t current1_id = 0;
if (JOINT1_ID > 0)
current1_id = JOINT1_ID;
std::size_t current2_id = 0;
if (JOINT2_ID > 0)
current2_id = JOINT2_ID;
while (current1_id != current2_id)
{
if (current1_id > current2_id)
{
int current1_col_id = JOINT_IDX_V[current1_id];
joint1_span_indexes.push_back(current1_id);
for (int k = 0; k < JOINT_NV[current1_id]; ++k, ++current1_col_id)
{
colwise_joint1_sparsity[current1_col_id] = true;
}
current1_id = JOINT_PARENTS[current1_id];
}
else
{
int current2_col_id = JOINT_IDX_V[current2_id];
joint2_span_indexes.push_back(current2_id);
for (int k = 0; k < JOINT_NV[current2_id]; ++k, ++current2_col_id)
{
colwise_joint2_sparsity[current2_col_id] = true;
}
current2_id = JOINT_PARENTS[current2_id];
}
}
printVector(colwise_joint1_sparsity);
}MRE_module.cpp #include <cstddef>
#include <iostream>
#include <Eigen/Core>
std::size_t JOINT1_ID = 5;
std::size_t JOINT2_ID = 0;
int NV = 10;
int NJOINT = 5;
int JOINT_IDX_V[] = {0, 0, 6, 7, 8, 9};
int JOINT_NV[] = {0, 6, 1, 1, 1, 1};
int JOINT_PARENTS[] = {0, 0, 1, 2, 3, 4};
void printVector(const Eigen::VectorX<bool>& v)
{
std::cout << v.transpose() << std::endl;
}compile.sh #! /bin/sh
#
FLAGS="-O3 -DNDEBUG -std=gnu++17"
g++ $FLAGS -c MRE_module.cpp -o MRE_module.o -isystem $CONDA_PREFIX/include/eigen3/
g++ $FLAGS -c MRE.cpp -o MRE.o -isystem $CONDA_PREFIX/include/eigen3/
g++ $FLAGS MRE.o MRE_module.o -o mre |
|
Current devel: caching nv with: --- a/include/pinocchio/algorithm/contact-info.hpp
+++ b/include/pinocchio/algorithm/contact-info.hpp
@@ -742,9 +742,10 @@ namespace pinocchio
if (current1_id > current2_id)
{
const JointModel & joint1 = model.joints[current1_id];
+ const int j1nv = joint1.nv();
joint1_span_indexes.push_back((Eigen::DenseIndex)current1_id);
Eigen::DenseIndex current1_col_id = joint1.idx_v();
- for (int k = 0; k < joint1.nv(); ++k, ++current1_col_id)
+ for (int k = 0; k < j1nv; ++k, ++current1_col_id)
{
colwise_joint1_sparsity[current1_col_id] = true;
}
@@ -753,9 +754,10 @@ namespace pinocchio
else
{
const JointModel & joint2 = model.joints[current2_id];
+ const int j2nv = joint2.nv();
joint2_span_indexes.push_back((Eigen::DenseIndex)current2_id);
Eigen::DenseIndex current2_col_id = joint2.idx_v();
- for (int k = 0; k < joint2.nv(); ++k, ++current2_col_id)
+ for (int k = 0; k < j2nv; ++k, ++current2_col_id)
{
colwise_joint2_sparsity[current2_col_id] = true;
}
@@ -770,10 +772,11 @@ namespace pinocchio
while (current_id > 0)
{
const JointModel & joint = model.joints[current_id];
+ const int jnv = joint.nv();
joint1_span_indexes.push_back((Eigen::DenseIndex)current_id);
joint2_span_indexes.push_back((Eigen::DenseIndex)current_id);
Eigen::DenseIndex current_row_id = joint.idx_v();
- for (int k = 0; k < joint.nv(); ++k, ++current_row_id)
+ for (int k = 0; k < jnv; ++k, ++current_row_id)
{
colwise_joint1_sparsity[current_row_id] = true;
colwise_joint2_sparsity[current_row_id] = true;I get: There is something fishy in that in |
|
this seems to fix all those: |
nim65s
added a commit
to nim65s/pinocchio
that referenced
this issue
Sep 14, 2024
|
The issue doesn't appear anymore with GCC14.1. So I will close this issue. I have created a minimal example (but with pinocchio, so it's hard to share with gcc team). #include <iostream>
#include "pinocchio/multibody/sample-models.hpp"
#include <boost/test/unit_test.hpp>
BOOST_AUTO_TEST_SUITE(BOOST_TEST_MODULE)
using namespace Eigen;
using namespace pinocchio;
template<typename Scalar>
struct ContactModel
{
JointIndex joint1_id;
JointIndex joint2_id;
Eigen::Matrix<bool, Eigen::Dynamic, 1> colwise_joint1_sparsity;
Eigen::Matrix<bool, Eigen::Dynamic, 1> colwise_joint2_sparsity;
template<int OtherOptions, template<typename, int> class JointCollectionTpl>
ContactModel(
const ModelTpl<Scalar, OtherOptions, JointCollectionTpl> & model,
const JointIndex joint1_id,
const JointIndex joint2_id)
: joint1_id(joint1_id)
, joint2_id(joint2_id)
, colwise_joint1_sparsity(model.nv)
, colwise_joint2_sparsity(model.nv)
{
init(model);
}
template<int OtherOptions, template<typename, int> class JointCollectionTpl>
void init(const ModelTpl<Scalar, OtherOptions, JointCollectionTpl> & model)
{
typedef ModelTpl<Scalar, OtherOptions, JointCollectionTpl> Model;
typedef typename Model::JointModel JointModel;
static const bool default_sparsity_value = false;
colwise_joint1_sparsity.fill(default_sparsity_value);
colwise_joint2_sparsity.fill(default_sparsity_value);
JointIndex current1_id = 0;
if (joint1_id > 0)
current1_id = joint1_id;
JointIndex current2_id = 0;
if (joint2_id > 0)
current2_id = joint2_id;
while (current1_id != current2_id)
{
if (current1_id > current2_id)
{
const JointModel & joint1 = model.joints[current1_id];
const int j1nv = joint1.nv();
Eigen::DenseIndex current1_col_id = joint1.idx_v();
for (int k = 0; k < j1nv; ++k, ++current1_col_id)
{
colwise_joint1_sparsity[current1_col_id] = true;
}
current1_id = model.parents[current1_id];
}
else
{
const JointModel & joint2 = model.joints[current2_id];
const int j2nv = joint2.nv();
Eigen::DenseIndex current2_col_id = joint2.idx_v();
for (int k = 0; k < j2nv; ++k, ++current2_col_id)
{
colwise_joint2_sparsity[current2_col_id] = true;
}
current2_id = model.parents[current2_id];
}
}
}
};
template<typename Scalar>
struct ContactModelNotWorking
{
JointIndex joint1_id;
JointIndex joint2_id;
Eigen::Matrix<bool, Eigen::Dynamic, 1> colwise_joint1_sparsity;
Eigen::Matrix<bool, Eigen::Dynamic, 1> colwise_joint2_sparsity;
template<int OtherOptions, template<typename, int> class JointCollectionTpl>
ContactModelNotWorking(
const ModelTpl<Scalar, OtherOptions, JointCollectionTpl> & model,
const JointIndex joint1_id,
const JointIndex joint2_id)
: joint1_id(joint1_id)
, joint2_id(joint2_id)
, colwise_joint1_sparsity(model.nv)
, colwise_joint2_sparsity(model.nv)
{
init(model);
}
template<int OtherOptions, template<typename, int> class JointCollectionTpl>
void init(const ModelTpl<Scalar, OtherOptions, JointCollectionTpl> & model)
{
typedef ModelTpl<Scalar, OtherOptions, JointCollectionTpl> Model;
typedef typename Model::JointModel JointModel;
static const bool default_sparsity_value = false;
colwise_joint1_sparsity.fill(default_sparsity_value);
colwise_joint2_sparsity.fill(default_sparsity_value);
JointIndex current1_id = 0;
if (joint1_id > 0)
current1_id = joint1_id;
JointIndex current2_id = 0;
if (joint2_id > 0)
current2_id = joint2_id;
while (current1_id != current2_id)
{
if (current1_id > current2_id)
{
const JointModel & joint1 = model.joints[current1_id];
Eigen::DenseIndex current1_col_id = joint1.idx_v();
for (int k = 0; k < joint1.nv(); ++k, ++current1_col_id)
{
colwise_joint1_sparsity[current1_col_id] = true;
}
current1_id = model.parents[current1_id];
}
else
{
const JointModel & joint2 = model.joints[current2_id];
Eigen::DenseIndex current2_col_id = joint2.idx_v();
for (int k = 0; k < joint2.nv(); ++k, ++current2_col_id)
{
colwise_joint2_sparsity[current2_col_id] = true;
}
current2_id = model.parents[current2_id];
}
}
}
};
BOOST_AUTO_TEST_CASE(test_gcc13_3)
{
using namespace Eigen;
using namespace pinocchio;
typedef JointCollectionDefaultTpl<double, pinocchio::context::Options> JC;
using buildModels::details::addJointAndBody;
Inertia Ijoint(.1, Inertia::Vector3::Zero(), Inertia::Matrix3::Identity() * .01);
Model model;
auto ffidx = model.addJoint(0, typename JC::JointModelFreeFlyer(), SE3::Identity(), "root_joint");
model.lowerPositionLimit.template segment<4>(3).fill(-1.);
model.upperPositionLimit.template segment<4>(3).fill(1.);
model.appendBodyToJoint(ffidx, Ijoint);
model.addJointFrame(ffidx);
buildModels::details::addManipulator(model, ffidx);
const std::string LF = "wrist2_joint";
const Model::JointIndex LF_id = model.getJointId(LF);
ContactModel ci_LF(model, LF_id, 0);
std::cout << ci_LF.colwise_joint1_sparsity.transpose() << std::endl;
ContactModelNotWorking ci_LF_not_work(model, LF_id, 0);
std::cout << ci_LF_not_work.colwise_joint1_sparsity.transpose() << std::endl;
}
BOOST_AUTO_TEST_SUITE_END() |
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Hi,
This is a preliminary bug report, I'm trying a few things to check that.
But so far, with nix, since gcc was upgraded from 13.2.0 to 13.3.0, test-cpp-contact-cholesky fail with big differences in some matrices (ie. 0.207 != 5e-17).
I suspect this is an issue in GCC, so I will deactivate this test on nix, and try to make a proper bug report to GCC.
The text was updated successfully, but these errors were encountered: