Use camp tuple and methods

[adayton1]

• Replaces the following in serac with camp implementations that are fully variadic: ** tuple ** tuple_size ** tuple_element ** several methods relating to tuples

[adayton1]

It appears that the issue I was running into comes from using camp::make_tuple instead of serac::make_tuple. Rather than holding this merge request up, I decided to continue using serac::make_tuple and determine how to replace it in a follow on pull request.

[adayton1]

I think the failures are from using public packages without the patch I have for camp. @white238, do I need to run the build tpls script without any arguments to install new public packages?

[adayton1]

I ran the build_tpls script. @white238, do I need to run something else to update the host configs?

[adayton1]

nvcc is crashing when trying to compile the following code. A compiler bug has been submitted through LC.

#include "camp/tuple.hpp"
#include <utility>

template <typename... S, typename... T, int... i>
CAMP_HOST_DEVICE constexpr auto plus_helper(const camp::tuple<S...>& x,
                                            const camp::tuple<T...>& y,
                                            std::integer_sequence<int, i...>)
{
     return camp::tuple{camp::get<i>(x) + camp::get<i>(y)...};
}

template <typename... S, typename... T>
CAMP_HOST_DEVICE constexpr auto operator+(const camp::tuple<S...>& x,
                                          const camp::tuple<T...>& y)
{
   static_assert(sizeof...(S) == sizeof...(T));
   return plus_helper(x, y, std::make_integer_sequence<int, static_cast<int>(sizeof...(S))>());
}

int main(int argc, char* argv[]) {
   camp::tuple a{0};
   camp::tuple b{1};
   camp::tuple c = a + b;
   (void) c;
   return 0;
}

[adayton1]

I have another data point. I tried using camp functionality instead of std::integer_sequence and std::make_integer_sequence, but nvcc still crashes.

#include "camp/tuple.hpp"
#include <utility>

template <typename... S, typename... T, camp::idx_t... i>
CAMP_HOST_DEVICE constexpr auto plus_helper(const camp::tuple<S...>& x,
                                            const camp::tuple<T...>& y,
                                            camp::idx_seq<i...>)
{
     return camp::tuple{camp::get<i>(x) + camp::get<i>(y)...};
}

template <typename... S, typename... T>
CAMP_HOST_DEVICE constexpr auto operator+(const camp::tuple<S...>& x,
                                          const camp::tuple<T...>& y)
{
   static_assert(sizeof...(S) == sizeof...(T));
   return plus_helper(x, y, camp::make_idx_seq_t<sizeof...(S)>{});
}

int main(int argc, char* argv[]) {
   camp::tuple a{0};
   camp::tuple b{1};
   camp::tuple c = a + b;
   (void) c;
   return 0;
}

[white238]

~~Data point from @trws, removing all CTAD from camp::tuple should allow this to compile.~~ Correction below

[trws]

Small clarification, the CTAD is mostly fine, it's just in copy deduction that we have problems. In this example it's specifically the camp::tuple c = a + b; line. Changing that out for auto should let it work.

[white238]

Small clarification, the CTAD is mostly fine, it's just in copy deduction that we have problems. In this example it's specifically the camp::tuple c = a + b; line. Changing that out for auto should let it work.

Apparently my reading comprehension is low today. Thanks for the correction @trws !

[white238]

Hi Chris,

Just heard back from the Nvidia guy at the lab. The fix for the CAMP tuple bug is slated for the CUDA 12.1 release.

Thanks, Alan Dayton

[jamiebramwell]

Closing as significant refactoring is happening. This PR will be used as reference once the refactoring is complete.