[FEA]: Improve binary function objects and replace thrust implementation

[miscco]

Is this a duplicate?

• I confirmed there appear to be no duplicate issues for this request and that I agree to the Code of Conduct

Area

General CCCL

Is your feature request related to a problem? Please describe.

Currently the binary function objects like cuda::std::plus do not universally default their template argument in all standard modes:

#if _CCCL_STD_VER > 2011
template <class _Tp = void>
#else
template <class _Tp>
#endif
struct _LIBCUDACXX_TEMPLATE_VIS plus;

Furthermore those objects are duplicated in thrust.

Describe the solution you'd like

We should replace thrusts versions with an alias declaration and default the cuda::std ones universally

Describe alternatives you've considered

No response

Additional context

No response

[srinivasyadav18]

Hi,

I am interested to work on this issue.

However, I need further context to get started with it.
I have understood the first part, which is defaulting the template argument universally.

Regarding replacing the thrust version with an alias declaration using cuda::std, should I be replacing this code
https://github.com/NVIDIA/cccl/blob/main/thrust/thrust/functional.h#L213-L237

template <typename T = void>
struct plus
{
  typedef T first_argument_type;
  typedef T second_argument_type;
  typedef T result_type;

  _CCCL_EXEC_CHECK_DISABLE
  _CCCL_HOST_DEVICE constexpr T operator()(const T& lhs, const T& rhs) const
  {
    return lhs + rhs;
  }
}; 

to use cuda::std::plus ?

Thanks!

[jrhemstad]

Hey @srinivasyadav18, thanks for your interest in contributing to CCCL!

You are correct. The basic idea would be to do the following:

Before:

namespace thrust{
template <typename T = void>
struct plus
{
  typedef T first_argument_type;
  typedef T second_argument_type;
  typedef T result_type;

  _CCCL_EXEC_CHECK_DISABLE
  _CCCL_HOST_DEVICE constexpr T operator()(const T& lhs, const T& rhs) const
  {
    return lhs + rhs;
  }
}; 
}

After:

namespace thrust {
   template <typename  T = void>
   using plus = ::cuda::std::plus<T>;
}

[srinivasyadav18]

Hi @jrhemstad,

Thank you providing more context on the issue!

I have replaced most of the function objects in thrust to use cuda::std one's.
But, currently I am stuck with replacing binary_negate function (here), also unary_negate.
As in C++ 20, negators are disabled by default in cuda::std. Hence, its not possible to use it.

Also, the existing thrust::binary_negate does not compile because, the Predicate type does not have any first_argument_type or second_argument_type nested type, becasue these functions objects now use cuda::std:: one's, which are dervided from __binary_function in binary_funcion.h that disable first_argument_type and second_argument_type by default.

Can you please provide more insights on how to proceed with this issue ?
Should we disable the negators in thrust now, like how cuda::std disabled it based on CXX standard ?

Thanks!

[jrhemstad]

@miscco can you answer @srinivasyadav18's questions?

[miscco]

@srinivasyadav18 That is indeed unfortunate.

I would for now suggest to just ignore negate and keep them at is. We can move incrementally if we like

[bernhardmgruber]

@srinivasyadav18 how are you getting along? I could use your work, even if it's just partial :)

[bernhardmgruber]

I opened a PR to enable the transparent functors (the = void) also in C++11: #1851

[srinivasyadav18]

@bernhardmgruber Apologies for late response. I have replace most of function objects to use the ::cuda::std one's. But there are still lot more places where I need to remove the use of first_argument_type, second_arguement_type and result_type nested types from the binary functions.

[bernhardmgruber]

Awesome! So your changes should not interfere with any of mine. Make sure to rebase on the latest tip of main when #1851 lands!