Eigen allocation segmentation fault (Open3D) - linux

I'm trying to make some operations on Pointcloud but when calling VoxelDownSample function i'm getting Segmentation fault signal. I see that the problem is that program want to allocate some astronomic amount of data but i don't know where it come from as actuall PointCloud object size is 120.
Code:
open3d::geometry::PointCloud open3d_source_original;
open3d::geometry::PointCloud open3d_source_down;
createPointCloud(open3d_source_original);
open3d_source_down = *open3d_source_original.VoxelDownSample(0.01);
createPointCloud is my function which creates PointCloud from csv file and it works fine.
gdb output:
#0 __GI_raise (sig=sig#entry=6) at ../sysdeps/unix/sysv/linux/raise.c:51
#1 0x00007ffff6eb4921 in __GI_abort () at abort.c:79
#2 0x00007ffff7ad6957 in ?? () from /usr/lib/x86_64-linux-gnu/libstdc++.so.6
#3 0x00007ffff7adcae6 in ?? () from /usr/lib/x86_64-linux-gnu/libstdc++.so.6
#4 0x00007ffff7adcb21 in std::terminate() () from /usr/lib/x86_64-linux-gnu/libstdc++.so.6
#5 0x00007ffff7adcd54 in __cxa_throw () from /usr/lib/x86_64-linux-gnu/libstdc++.so.6
#6 0x00007ffff7b05012 in std::__throw_bad_alloc() () from /usr/lib/x86_64-linux-gnu/libstdc++.so.6
#7 0x0000555555563476 in __gnu_cxx::new_allocator<Eigen::Matrix<double, 3, 1, 0, 3, 1> >::allocate(unsigned long, void const*) ()
#8 0x0000555555562bec in std::allocator_traits<std::allocator<Eigen::Matrix<double, 3, 1, 0, 3, 1> > >::allocate(std::allocator<Eigen::Matrix<double, 3, 1, 0, 3, 1> >&, unsigned long) ()
#9 0x0000555555562088 in std::_Vector_base<Eigen::Matrix<double, 3, 1, 0, 3, 1>, std::allocator<Eigen::Matrix<double, 3, 1, 0, 3, 1> > >::_M_allocate(unsigned long) ()
#10 0x00005555555613ac in Eigen::Matrix<double, 3, 1, 0, 3, 1>* std::vector<Eigen::Matrix<double, 3, 1, 0, 3, 1>, std::allocator<Eigen::Matrix<double, 3, 1, 0, 3, 1> > >::_M_allocate_and_copy<__gnu_cxx::__normal_iterator<Eigen::Matrix<double, 3, 1, 0, 3, 1> const*, std::vector<Eigen::Matrix<double, 3, 1, 0, 3, 1>, std::allocator<Eigen::Matrix<double, 3, 1, 0, 3, 1> > > > >(unsigned long, __gnu_cxx::__normal_iterator<Eigen::Matrix<double, 3, 1, 0, 3, 1> const*, std::vector<Eigen::Matrix<double, 3, 1, 0, 3, 1>, std::allocator<Eigen::Matrix<double, 3, 1, 0, 3, 1> > > >, __gnu_cxx::__normal_iterator<Eigen::Matrix<double, 3, 1, 0, 3, 1> const*, std::vector<Eigen::Matrix<double, 3, 1, 0, 3, 1>, std::allocator<Eigen::Matrix<double, 3, 1, 0, 3, 1> > > >) ()
#11 0x00005555555606e2 in std::vector<Eigen::Matrix<double, 3, 1, 0, 3, 1>, std::allocator<Eigen::Matrix<double, 3, 1, 0, 3, 1> > >::operator=(std::vector<Eigen::Matrix<double, 3, 1, 0, 3, 1>, std::allocator<Eigen::Matrix<double, 3, 1, 0, 3, 1> > > const&) ()
#12 0x000055555555fcfa in open3d::geometry::PointCloud::operator=(open3d::geometry::PointCloud const&) ()
#13 0x000055555555f400 in main ()
Valgrind output:
==14534== Invalid read of size 8
==14534== at 0x114D59: std::vector<Eigen::Matrix<double, 3, 1, 0, 3, 1>, std::allocator<Eigen::Matrix<double, 3, 1, 0, 3, 1> > >::begin() const (in /home/rufus/CLionProjects/isolatedSegFault/build/isolatedSegFault)
==14534== by 0x1147EA: std::vector<Eigen::Matrix<double, 3, 1, 0, 3, 1>, std::allocator<Eigen::Matrix<double, 3, 1, 0, 3, 1> > >::operator=(std::vector<Eigen::Matrix<double, 3, 1, 0, 3, 1>, std::allocator<Eigen::Matrix<double, 3, 1, 0, 3, 1> > > const&) (in /home/rufus/CLionProjects/isolatedSegFault/build/isolatedSegFault)
==14534== by 0x113CF9: open3d::geometry::PointCloud::operator=(open3d::geometry::PointCloud const&) (in /home/rufus/CLionProjects/isolatedSegFault/build/isolatedSegFault)
==14534== by 0x1133FF: main (in /home/rufus/CLionProjects/isolatedSegFault/build/isolatedSegFault)
==14534== Address 0x61da940 is 0 bytes after a block of size 112 alloc'd
==14534== at 0x4C3217F: operator new(unsigned long) (in /usr/lib/valgrind/vgpreload_memcheck-amd64-linux.so)
==14534== by 0x12C358: allocate (new_allocator.h:111)
==14534== by 0x12C358: allocate (alloc_traits.h:436)
==14534== by 0x12C358: __allocate_guarded<std::allocator<std::_Sp_counted_ptr_inplace<open3d::geometry::PointCloud, std::allocatoropen3d::geometry::PointCloud, (__gnu_cxx::_Lock_policy)2> > > (allocated_ptr.h:104)
==14534== by 0x12C358: __shared_count<open3d::geometry::PointCloud, std::allocatoropen3d::geometry::PointCloud > (shared_ptr_base.h:635)
==14534== by 0x12C358: __shared_ptr<std::allocatoropen3d::geometry::PointCloud > (shared_ptr_base.h:1295)
==14534== by 0x12C358: shared_ptr<std::allocatoropen3d::geometry::PointCloud > (shared_ptr.h:344)
==14534== by 0x12C358: allocate_shared<open3d::geometry::PointCloud, std::allocatoropen3d::geometry::PointCloud > (shared_ptr.h:691)
==14534== by 0x12C358: make_sharedopen3d::geometry::PointCloud (shared_ptr.h:707)
==14534== by 0x12C358: open3d::geometry::PointCloud::VoxelDownSample(double) const (PointCloud.cpp:306)
==14534== by 0x1133DB: main (in /home/rufus/CLionProjects/isolatedSegFault/build/isolatedSegFault)
==14534==
==14534==
==14534== HEAP SUMMARY:
==14534== in use at exit: 8 bytes in 1 blocks
==14534== total heap usage: 19 allocs, 18 frees, 107,576 bytes allocated
==14534==
==14534== LEAK SUMMARY:
==14534== definitely lost: 0 bytes in 0 blocks
==14534== indirectly lost: 0 bytes in 0 blocks
==14534== possibly lost: 0 bytes in 0 blocks
==14534== still reachable: 8 bytes in 1 blocks
==14534== suppressed: 0 bytes in 0 blocks
==14534== Rerun with --leak-check=full to see details of leaked memory
==14534==
==14534== For counts of detected and suppressed errors, rerun with: -v
==14534== Use --track-origins=yes to see where uninitialised values come from
==14534== ERROR SUMMARY: 6 errors from 6 contexts (suppressed: 0 from 0)
Operating system: Ubuntu 18.04
Open3D build from source
Compiler version: gcc 7.5
I tried to change macros from Eigen and change compilers but nothing worked.

This part shows a SIGABRT due to too much memory, not a SIGSEGV:
#0 __GI_raise (sig=sig#entry=6) at ../sysdeps/unix/sysv/linux/raise.c:51
#1 0x00007ffff6eb4921 in __GI_abort () at abort.c:79
Valgrind is indicating that uninitialized data was used, not unmapped memory:
==14534== Invalid read of size 8
It appears to be pointing at the internal pointer to the start of the array which underlies the vector:
==14534== at 0x114D59: std::vector<Eigen::Matrix<double, 3, 1, 0, 3, 1>, std::allocator<Eigen::Matrix<double, 3, 1, 0, 3, 1> > >::begin() const (in /home/rufus/CLionProjects/isolatedSegFault/build/isolatedSegFault)
Then it says that the vector was created by make_shared inside voxelDownSample:
==14534== by 0x12C358: make_sharedopen3d::geometry::PointCloud (shared_ptr.h:707)
==14534== by 0x12C358: open3d::geometry::PointCloud::VoxelDownSample(double) const (PointCloud.cpp:306)
My intuition says,
If a weak_ptr is involved, then perhaps the vector was destroyed by losing all its shared_ptr references. Valgrind can't see this because it wouldn't be deallocated until the weak_ptr references are gone, too.
Otherwise, there's some other kind of memory error, such as
trying to use the vector before it has actually been constructed
using a raw pointer after the shared_ptr is destroyed
a buffer overflow, or other randomly spraying memory (rather less likely to hit a shared_ptr object, but you never know)
In any case, it looks like a library bug, if this object was created after the last time any of your code ran, and your parameters were numerically valid.

Related

Pack tensor containing padding on both sides

I have an LSTM that receives as input a padded tensor with pad values on both sides (assume that pad values are 0):
[0, 0, 0, 3, 2, 1, 4, 0, 0, 0]
[1, 9, 5, 3, 2, 1, 4, 7, 2, 3]
[0, 0, 0, 3, 2, 1, 4, 2, 3, 3]
Typically, one would use TORCH.NN.UTILS.RNN.PACK_PADDED_SEQUENCE to let the LSTM know which values to pad, but I've only seen it used for cases where tensors are right-padded. There's also no indication from the documentation that it will be able to recognize padded values on the left-hand side. Is there an alternative?

How to initialize columns in hybrid sparse tensor

How initialize in pytorch hybrid tensor torch.sparse_coo_tensor (one dimension is sparse and other is not), which have the following dense representation?
array([[1, 0, 5, 0],
[2, 0, 6, 0],
[3, 0, 7, 0],
[4, 0, 8, 0]])
What should I put into the indices argument?
How to initialize
Something like this:
import torch
indices = torch.tensor([[0, 0, 1, 1, 2, 2, 3, 3], [0, 2, 0, 2, 0, 2, 0, 2]])
tensor = torch.sparse_coo_tensor(
indices, torch.tensor([1, 2, 3, 4, 5, 6, 7, 8]), size=(4, 4)
)
Given above:
indices - first dimension specifies row, second column, where non-zero value(s) will be located. Those become pairs, in this case: (0, 0), (0, 2), (1, 0), (1, 2)... and so on
values - values located at those pairs, so 1 will be under (0, 0) coordinate, 2 under (0, 2) and so it goes.
size - total size of the matrix, optional, might be inferred in this case from your input
8 pairs, 8 values, there are also other ways to specify it, but the idea holds.
And a quick check:
print(tensor)
print(tensor.to_dense())
Gives us:
tensor(indices=tensor([[0, 0, 1, 1, 2, 2, 3, 3],
[0, 2, 0, 2, 0, 2, 0, 2]]),
values=tensor([1, 2, 3, 4, 5, 6, 7, 8]),
size=(4, 4), nnz=8, layout=torch.sparse_coo)
tensor([[1, 0, 2, 0],
[3, 0, 4, 0],
[5, 0, 6, 0],
[7, 0, 8, 0]])
Why to initialize
If your actual data is 50% sparse, you shouldn't use COO tensor.
It will save some memory, but operations will be way slower, so keep that in mind.

Fio results are steadily increasing IOPS, not what I expected

I'm trying to somehow test my rbd storage with random read, random write, mixed randrw, but the output is not correct, it is a sequential growing number.
What is wrong with my steps?
This is the fio file that I ran:
; fio-rand-write.job for fiotest
[global]
name=fio-rand-write
filename=fio-rand-write
rw=randwrite
bs=4K
direct=1
write_iops_log=rand-read
[file1]
size=1G
ioengine=libaio
iodepth=16
And the result is this:
head rand-read_iops.1.log
2, 1, 1, 4096, 0
2, 1, 1, 4096, 0
2, 1, 1, 4096, 0
2, 1, 1, 4096, 0
2, 1, 1, 4096, 0
3, 1, 1, 4096, 0
4, 1, 1, 4096, 0
5, 1, 1, 4096, 0
5, 1, 1, 4096, 0
5, 1, 1, 4096, 0
tail rand-read_iops.1.log
30700, 1, 1, 4096, 0
30700, 1, 1, 4096, 0
30700, 1, 1, 4096, 0
30700, 1, 1, 4096, 0
30700, 1, 1, 4096, 0
30700, 1, 1, 4096, 0
30700, 1, 1, 4096, 0
30700, 1, 1, 4096, 0
30700, 1, 1, 4096, 0
30700, 1, 1, 4096, 0
I'm using fio 3.18.
Why I don't get the iops that is the real one?
(Note this isn't really a programming question so Stackoverflow is the wrong place to ask this... Maybe Super User or Serverfault would be a better choice and get faster answers?)
but the output is not correct, it is a sequential growing number
Which column are you referring to? If you mean the left most column then isn't that time per the fio Log File Formats documentation?
Fio supports a variety of log file formats, for logging latencies, bandwidth, and IOPS. The logs share a common format, which looks like this:
time (msec), value, data direction, block size (bytes), offset (bytes)
Doesn't time generally monotonically increase relative to prior readings (accounting for precision)?
Also see the documentation for write_iops_log that says:
Because fio defaults to individual I/O logging, the value entry in the IOPS log will be 1 unless windowed logging (see log_avg_msec) has been enabled

How to use Numpy .tobytes() to serialize objects

How do you serialized/deserialize a numpy array?
A = np.random.randint(0, 10, 40).reshape(8, 5)
print(A)
print (A.dtype)
snapshot = A
serialized = snapshot.tobytes()
[[9 5 5 7 4]
[3 8 8 1 0]
[5 7 1 0 2]
[2 2 7 1 2]
[2 6 3 5 4]
[7 5 4 8 3]
[2 4 2 4 7]
[3 4 2 6 2]]
int64
Returns
deserialized = np.frombuffer(serialized).astype(np.int64)
print (deserialized)
[0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0
0 0 0]
There is a mismatch between the default dtype used to generate A and in np.frombuffer. Works as expected when using the correct dtype (may depend on the machine / Python / numpy version):
# Python 3.6 64-bits with numpy 1.12.1 64-bits
A = np.random.randint(0, 10, 40).reshape(8, 5)
print(A)
>>> array([[3, 3, 5, 3, 9],
[1, 4, 7, 1, 8],
[1, 7, 4, 3, 0],
[9, 2, 9, 1, 2],
[2, 8, 9, 1, 1],
[3, 3, 5, 2, 6],
[5, 0, 2, 7, 6],
[2, 8, 8, 0, 7]])
A.dtype
>>> dtype('int32')
deserialized = np.frombuffer(A.tobytes(), dtype=np.int32).reshape(A.shape)
print(deserialized)
>>> array([[3, 3, 5, 3, 9],
[1, 4, 7, 1, 8],
[1, 7, 4, 3, 0],
[9, 2, 9, 1, 2],
[2, 8, 9, 1, 1],
[3, 3, 5, 2, 6],
[5, 0, 2, 7, 6],
[2, 8, 8, 0, 7]])

Memory leak in Eigen Conjugate Gradient?

I have spent the last day looking for a memory leak in my code. For each frame in my loop I can see in the task manager that about 200 Mb of memory is lost. Eventually the program of course crashes.
The pseudo-code looks like this:
for frame = 0:NBROFFRAMES
cv::Mat = getImage(frame);
cout<<"Before optimization"<<endl;
getchar();
optFunction.optimizeColor(img);
cout<<"After optimization"<<endl;
getchar();
Do other things
end
Just looking in the task manager shows that in optimizeColor approximately 200 Mb are lost for every frame.
In optimizeColorI run Eigens ConjugateGradient several times. When I changed to Cholesky Factorization, SimplicialLDLT, the memory consumption is between 4.5-4.7 Gb over the entire sequence. Using ConjugateGradient it starts with 4.5 Gb and after 10 images the amount of memory used according to my Task Manager is 11.2 Gb.
I have also run Valgrind and one of the messages I get is the following:
==22576== 110,430,720 bytes in 18 blocks are possibly lost in loss record 30,186 of 30,189
==22576== at 0x4C2E80F: operator new[](unsigned long) (in /usr/lib/valgrind/vgpreload_memcheck-amd64-linux.so)
==22576== by 0x913CCD: Eigen::internal::scoped_array<int>::scoped_array(long) (Memory.h:666)
==22576== by 0x910329: Eigen::internal::CompressedStorage<float, int>::reallocate(long) (CompressedStorage.h:230)
==22576== by 0x910695: Eigen::internal::CompressedStorage<float, int>::resize(long, double) (CompressedStorage.h:96)
==22576== by 0x91CC41: Eigen::SparseMatrix<float, 0, int>& Eigen::SparseMatrix<float, 0, int>::operator=<Eigen::SparseMatrix<float, 1, int> >(Eigen::SparseMatrixBase<Eigen::SparseMatrix<float, 1, int> > const&) (SparseMatrix.h:1085)
==22576== by 0x91A911: Eigen::SparseMatrix<float, 0, int>::SparseMatrix<Eigen::SparseMatrix<float, 1, int> >(Eigen::SparseMatrixBase<Eigen::SparseMatrix<float, 1, int> > const&) (SparseMatrix.h:674)
==22576== by 0x91837C: void Eigen::Ref<Eigen::SparseMatrix<float, 0, int> const, 0, Eigen::OuterStride<-1> >::construct<Eigen::SparseMatrix<float, 1, int> >(Eigen::SparseMatrix<float, 1, int> const&, Eigen::internal::false_type) (SparseRef.h:217)
==22576== by 0x91631C: Eigen::Ref<Eigen::SparseMatrix<float, 0, int> const, 0, Eigen::OuterStride<-1> >::Ref<Eigen::SparseMatrix<float, 1, int> >(Eigen::SparseMatrixBase<Eigen::SparseMatrix<float, 1, int> > const&) (SparseRef.h:184)
==22576== by 0x913801: void Eigen::internal::generic_matrix_wrapper<Eigen::SparseMatrix<float, 0, int>, false>::grab<Eigen::SparseMatrix<float, 1, int> >(Eigen::EigenBase<Eigen::SparseMatrix<float, 1, int> > const&) (IterativeSolverBase.h:81)
==22576== by 0x90FEFE: void Eigen::IterativeSolverBase<Eigen::ConjugateGradient<Eigen::SparseMatrix<float, 0, int>, 3, Eigen::DiagonalPreconditioner<float> > >::grab<Eigen::SparseMatrix<float, 1, int> >(Eigen::SparseMatrix<float, 1, int> const&) (IterativeSolverBase.h:375)
==22576== by 0x90C03D: Eigen::ConjugateGradient<Eigen::SparseMatrix<float, 0, int>, 3, Eigen::DiagonalPreconditioner<float> >& Eigen::IterativeSolverBase<Eigen::ConjugateGradient<Eigen::SparseMatrix<float, 0, int>, 3, Eigen::DiagonalPreconditioner<float> > >::analyzePattern<Eigen::SparseMatrix<float, 1, int> >(Eigen::EigenBase<Eigen::SparseMatrix<float, 1, int> > const&) (IterativeSolverBase.h:199)
==22576== by 0x9077F3: OptimizeAlbedo::optimize_rho(cv::Mat const&, std::vector<Eigen::Matrix<float, -1, 1, 0, -1, 1>, std::allocator<Eigen::Matrix<float, -1, 1, 0, -1, 1> > > const&, std::vector<Eigen::Matrix<int, 2, 1, 0, 2, 1>, std::allocator<Eigen::Matrix<int, 2, 1, 0, 2, 1> > > const&, Eigen::Matrix<float, -1, 1, 0, -1, 1> const&, float, float, Eigen::Matrix<float, -1, 1, 0, -1, 1> const&) (OptimizeAlbedo.cpp:586)
The only thing I do that allocates memory is that I define the sizes of some matrices when I create the instance optFunction and reserve space for the entries. That is of course only done once and that I have checked several times. The only explanation I see is that ConjugateGradient is consuming memory.
Here is my code to call ConjugateGradient:
function optmize()
Eigen::ConjugateGradient<Eigen::SparseMatrix<float>, Eigen::Lower | Eigen::Upper> cg;
cg.setTolerance(0.001);
cg.setMaxIterations(200);
cg.analyzePattern(A_tot);
cg.factorize(A_tot);
Eigen::VectorXf opt = cg.solveWithGuess(b_tot, rho_current);
end
I would have expected that cg is cleared when the function goes out of scope, but it seems that is not the case.
I think I use Eigen version: 3.2.92, this is in the Macros.h:
#define EIGEN_WORLD_VERSION 3
#define EIGEN_MAJOR_VERSION 2
#define EIGEN_MINOR_VERSION 92
I am using Ubuntu 16.04 and g++ 5.4.0.
EDIT: As pointed out by Avi Ginsburg, I was using a beta-version of Eigen. By upgrading to 3.3.2 it works without memory leaks.

Resources