Hello,

At work I try to update our simulation lib to use eigen. So I switched the Solver for sparse linear systems to the SparseLU solver from eigen. First I used a test programm, it worked perfectly and very fast. then i puut it in our dll (the same code no changes, just that the matrix didnt come from a test file, instead it comes from the programm, which uses the simulation dll). Suddenly I get Access Violation Errors in GeneralBlockPanelKernel.h:

Code: Select all

EIGEN_STRONG_INLINE void unpackRhs(DenseIndex n, const RhsScalar* rhs, RhsScalar* b)
  {
    for(DenseIndex k=0; k<n; k++)
      pstore1<RhsPacket>(&b[k*RhsPacketSize], rhs[k]);
  }


So I looked at the position with the Debugger and everything seems to be in order (at least for the debugging tool from visual studio 2013). So i controlled the input, the matrix and everything else. And if i protocol step by step in short:

Code: Select all

SparseMatrix<double> matrix;
typedef Eigen::Triplet < double > T;
std::vector<T, Eigen::aligned_allocator<T>> triplets;
SparseLU<SparseMatrix<double>> solver;
VectorXd rhs, solution;

this->matrix.resize(this->rows, this->cols);
matrix.reserve(VectorXi::Constant(rows, 12));

this->rhs.resize(rows);
this->solution.resize(cols);

this->matrix.makeCompressed();

this->solver.analyzePattern(matrix);

this->solver.factorize(this->matrix);

this->solver._solve(this->rhs, this->solution);


I get the following info:

- Matrix get initialised successfully
- Matrix get compressed successfully
- analyzePattern and factorize are successfull

and by solve I get the Access Violation. Has somebody an Idea what I do wrong? Or is it a problem, because eigen is inside a dll?

Thanks in Advance,

Kai Gödde

First, make sure that both the DLL and your application are compiled with the same flags.

Second, could you show the values of b (the address, not the values pointed by b), k and n? It might be that b is not aligned. It would also be interesting to test with EIGEN_DONT_VECTORIZE predefined for both the DLL and your program.

Thank you for your fast answer.

- If I compile with EIGEN_DONT_VECTORIZE it just works. But how much speed did I loose than?

- If i dont use the compiler flag, than

b = 0x068bb62c
k = 0
n = 2

Alright, so the problem is that b is not aligned. Maybe the stack is not aligned and so alloca does not return aligned memory within DLLs. You can verify this guess by adding:

#undef EIGEN_ALLOCA

right after it has been defined in Eigen/src/Core/util/Memory.h

Thanks,

If I use the flag

#undef EIGEN_ALLOCA

it works fine, but it is really slow. How can I fix that?

In the Memory.h file, you can modify the lines:

Code: Select all

  #ifdef __arm__
    #define EIGEN_ALIGNED_ALLOCA(SIZE) reinterpret_cast<void*>((reinterpret_cast<size_t>(EIGEN_ALLOCA(SIZE+16)) & ~(size_t(15))) + 16)
  #else
    #define EIGEN_ALIGNED_ALLOCA EIGEN_ALLOCA
  #endif


to

Code: Select all

  #if defined(__arm__) || defined(_MSC_VER)
    #define EIGEN_ALIGNED_ALLOCA(SIZE) reinterpret_cast<void*>((reinterpret_cast<size_t>(EIGEN_ALLOCA(SIZE+16)) & ~(size_t(15))) + 16)
  #else
    #define EIGEN_ALIGNED_ALLOCA EIGEN_ALLOCA
  #endif


If that work, we'll apply the fix upstream.

Thank you

Now it works fine (but slow, because i have created some redundancies, which i now have to clean up). SO now i just have to show my superior how much faster we can get

Kind regards,

Kai Gödde

https://bitbucket.org/eigen/eigen/commits/95bf819b5849/
Changeset: 95bf819b5849
User: ggael
Date: 2014-02-14 00:04:38
Summary: alloca is not necessarily alligned on windows