Error when compiling tensor with NDEBUG

[dubek]

I'm compiling a simple program which includes the tensor.hpp header:

#include <boost/numeric/ublas/tensor.hpp>
#include <ostream>

int main()
{
    auto A = boost::numeric::ublas::tensor<float>{3,4,2};
    std::cout << "A=" << A << std::endl;
    return 0;
}

This works fine if I don't define NDEBUG. However, in release mode with (-DNDEBUG argument to the compiler), this fails:

In file included from /celibs/boost_1_72_0/boost/numeric/ublas/tensor.hpp:22,
                 from <source>:1:
/celibs/boost_1_72_0/boost/numeric/ublas/tensor/strides.hpp:215:9: error: expected unqualified-id before '<' token
  215 | template<class size_type, class layout_type>
      |         ^
/celibs/boost_1_72_0/boost/numeric/ublas/tensor/strides.hpp:223:2: error: expected ';' before 'inline'
  223 | }
      |  ^
      |  ;
In file included from /celibs/boost_1_72_0/boost/numeric/ublas/tensor.hpp:24,
                 from <source>:1:
/celibs/boost_1_72_0/boost/numeric/ublas/tensor/tensor.hpp:333:2: error: expected unqualified-id before 'template'
  333 |  template<class other_layout>
      |  ^~~~~~~~
/celibs/boost_1_72_0/boost/numeric/ublas/tensor/tensor.hpp:355:2: error: expected unqualified-id before 'template'
  355 |  template<class derived_type>
      |  ^~~~~~~~

...

As you can see in compiler explorer, this happens both on gcc 9.3 and clang 10.0.0 with these command-line options: -DNDEBUG -std=c++17

I also found that adding #define BOOST_UBLAS_INLINE before including tensor.hpp solves the compilation problem (but, of course, cancels inlining which probably hurts performance).

Am I using the wrong compilation options? ublas' documentation says I should define NDEBUG preprocessor symbol to get proper performance (release mode). I assume tensor was tested with -DNDEBUG and -O3 etc. BTW I think that the error reported in #79 might also be related to -DNDEBUG.

[bassoy]

As you can see in compiler explorer, this happens both on gcc 9.3 and clang 10.0.0 with these command-line options: -DNDEBUG -std=c++17

Thanks @dubek for reporting this issue. Your observation is correct and this is a bug. Sorry for the inconvenience. I doubt that you will have severe performance penalties for tensor operations when not using the NDEBUG option. We have not yet optimized tensor operations such as done in this repo.

In general, Boost.uBlas is in a transition and we are about to generalize the tensor concept. The BOOST_UBLAS_INLINE pragma will be depecrated for tensor after Boost.uBlas version 1.73.

[dubek]

Thanks, I'll look at TTV. The problem with compiling without NDEBUG is that I lose optimizations for other parts of Boost that I use in my program.

[bassoy]

Thanks, I'll look at TTV. The problem with compiling without NDEBUG is that I lose optimizations for other parts of Boost that I use in my program.

Sorry to hear that. We do our best to yield a high-code coverage with unit-tests.
Could you tell me just roughly what applications you use for tensor operations. Are their any missing tensor operations or functions?

[yjwen]

Met the same compilation error with NDEBUG defined today. After some digging into tensor's code, I guess it is due to the wrong position of BOOST_UBLAS_INLINE placed before template keyword in some tensor/tensor.hpp and tensor/strides.hpp.

Take the code at line 214-216 of strides.hpp for example.

BOOST_UBLAS_INLINE
template<class size_type, class layout_type>
auto access(std::vector<size_type> const& i, basic_strides<size_type,layout_type> const& w)

When NDEBUG is defined, BOOST_UBLAS_INLINE is defined as inline in ublas/detail/config.hpp. The above code is expanded to

inline
template<class size_type, class layout_type>
auto access(std::vector<size_type> const& i, basic_strides<size_type,layout_type> const& w)

which cause syntax error.

The proper code should be inline following template, that requires BOOST_UBLAS_INLINE being placed after template

There are several such BOOST_UBLAS_INLINE before template should be moved to after template.

The issue is still seen with boost 1.77.0 release.

Regards,
Yujie

[bassoy]

This is changed with commit 231ba5f in branch develop.

Yet not merged into the master branch which is why you still see the issue in boost version 1.77.0.

[huang-lihao]

This problem still exits in v1.80.0