add special kernel methods

[BenjaminW3]

• clock_t clock();
  long long int clock64();
• T __ldg(const T* address);
  The read-only data cache load function is only supported by devices of compute capability 3.5 and higher.
• __shfl_sync, __shfl_up_sync, __shfl_down_sync, __shfl_xor_sync
  Exchange a variable between threads within a warp.

[felicepantaleo]

just checking if there is any plan on this

[psychocoderHPC]

just checking if there is any plan on this

You increased the priority for this features by +1

[BenjaminW3]

Are you more interested in ldg or shfl?

[psychocoderHPC]

The request in our internal Mattermost was for __ballot_sync() and __shfl_up_sync(). I also see shuffle as more important as __ldg()

Note: CUDA has __ballot (asynchronous) and __ballot_sync()

[j-stephan]

We will implement these intrinsics for alpaka 0.7.0.

• Implementing __ldg assigned to @j-stephan (as part of abstract the memory access inside kernels #38).
• Implementing __shfl assigned to @sbastrakov.

[sbastrakov]

The shuffle part was implemented in #1273 (not by me, but it's now in).

[j-stephan]

2/3 of the methods mentioned here have already been implemented. The remaining functionality will be tracked in #1312 (for easier project management).