[GPU] Replace CV_XADD by C++11 method

[SunBlack]

Currently we are using proprietary code here

pcl/gpu/containers/src/device_memory.cpp

Lines 79 to 94 in 7118b62

	#ifdef __GNUC__
	#if !defined WIN32 && (defined __i486__ || defined __i586__ || defined __i686__ || defined __MMX__ || defined __SSE__ || defined __ppc__)
	#define CV_XADD __sync_fetch_and_add
	#else
	#include <ext/atomicity.h>
	#define CV_XADD __gnu_cxx::__exchange_and_add
	#endif
	#elif defined WIN32 || defined _WIN32
	#include <intrin.h>
	#define CV_XADD(addr,delta) _InterlockedExchangeAdd((long volatile*)(addr), (delta))
	#else
	template<typename _Tp> static inline _Tp CV_XADD(_Tp* addr, _Tp delta)
	{ int tmp = *addr; *addr += delta; return tmp; }
	#endif

I don't know why there is a thread-safety reference counter, but we can replace CV_XADD by atomic operations of C++11 (I believe atomic_fetch_add should fit)

[kunaltyagi]

If these are used in host, 👍

For device however, proceed carefully since some versions of CUDA still don't support atomic operations.

[SunBlack]

This looks more like host code for me ;-).

[stale]

Marking this as stale due to 30 days of inactivity. It will be closed in 7 days if no further activity occurs.

[stale]

Marking this as stale due to 30 days of inactivity. It will be closed in 7 days if no further activity occurs.

[rabaho]

CV_XADD is host code for sure. First, the code was copied from OpenCV without acknowledgement, there is even comments left behind tying it to OpenCV. Second, I am not sure why we're reinventing the wheel with ref counting, shared_ptr would do the work.

[mvieth]

Fixed by the pull request linked above