add a float atomics sample

[bashbaug]

Adds a sample demonstrating how to use floating-point atomics in an OpenCL kernel.

[bashbaug]

This is currently a "draft" because the emulated floating-point atomic add that returns a value is not producing correct intermediate results. The final results of all floating-point atomic adds is correct, though, and the intermediate results using cl_ext_float_atomics and other device-specific solutions is working.

To see the issue with intermediate results, pass the -c option to add "intermediate results checking", and perhaps the -e option to force the emulated atomic add codepath. Example:

$ ./floatatomics -p3 -c -e
Running on platform: Intel(R) OpenCL Graphics
Running on device: Intel(R) Arc(TM) A750 Graphics
Forcing emulation.
Finished in 0.007057 seconds
Basic Validation: Success.
Error at index 0: expected 0.000000, got 1.000000!
Error at index 1: expected 1.000000 > 1.000000!
<snip>
Intermediate Results Validation: Found 64797 mismatches / 65536 values!!!

[bashbaug]

After some discussion, I confirmed that the emulated float atomic implementation that uses atomic_xchg does not reliably return the previous value in memory ("intermediate results"). So, I've added a slower (but safer, and correct) emulated version that does reliably return the previous value in memory. The faster version is still the default version chosen by passing -e, but now the slower version can be selected by passing -s.

The non-emulated versions are still preferred and are chosen by default, for supporting devices.