Use implicit id counter for shared memory allocation

[BenjaminW3]

This branch implements a compile time counter for the shared memory allocation that is compiling on all platforms.
It is based on the code from Filip Roséen.
At least in C++11 and C++14 this code is valid. The C++ Standards Committee is currently debating on wheter this should be legal or not in C++17. See here.

• Suppress the high number of warnings. (-Wundefined-internal / -Wnon-template-friend)
• Increase the maximum counter value from 64 to something reasonable. However, this could exceed the maximum template instantiation depth.
• Add the solution to the Stackoverflow thread.

[psychocoderHPC]

Thanks for this workaround I will test it next week.

[psychocoderHPC]

There is no need to add the id to the CPU alloc structs. The use of an id for shared memory can fully hidden in the GPU specialization.

[psychocoderHPC]

I tested you current branch with cuda7.0 and two shared memory variables. Both are equal in the kernel which means the current implementation is not working.

[BenjaminW3]

I might have an Idea how to fix this and will test it in the next days.

[psychocoderHPC]

btw: I tested the original solution of Filip Roséen with g++4.9 on the host and always get the some ID.

[psychocoderHPC]

OK the original id generation on CPU is not running because of a problem that always the path for BOOST_COMP_MSVC is used also in linux with gcc.
update: Found the problem #ifdef instead od #if is used

[psychocoderHPC]

This must be #if else always the first code path is used

[BenjaminW3]

My mistake...

[psychocoderHPC]

I pushed a small update to your pull request that fix some small mistakes. BenjaminW3#1

Shared memory id generation on is working but not with cuda 7.0

[BenjaminW3]

Thanks, moving the uniqueId() call to the default template parameter was the same idea I had.

[psychocoderHPC]

Yes if it is not a default parameter at this place no new ids are generated.

[psychocoderHPC]

The current problem with NVCC is that it not supports noexcept on this way we need it here. NVCC implements this.
The second pass in line 92 is also not working with nvcc :-(

[BenjaminW3]

I tried a lot of different things but did not get it working with nvcc (neither 7.0 nor 7.5). Therefore I will close this for now. However, clang, MSVC and gcc did work as expected.

[ax3l]

The C++ Standards Committee is currently debating on wheter this should be legal or not in C++17.

@BenjaminW3 btw, do you know how the discussion ended? ;)

[BenjaminW3]

@ax3l You can have a look at the C++ core issue here