add a float16_t data type for half-float support

[jaredhoberock]

The half-float [1] type would interoperate with standard 32-bit and 64-bit floats. Also, to avoid redundant conversions the arithmetic operators should yield a float32_t result. This behavior is consistent with the use of half-floats as a compressed storage format.

This enhancement would likely depend on issue NVIDIA/libcudacxx#216 .

class float16_t
{
unsigned short s;

public:

// constructors
host device
float16_t()
: s(float32_t()) {}

host device
float16_t(const float32_t& f)
: s(__float2half_rn(f)) {}

// arithmetic
host device
float32_t operator+(const float16_t& f) const
{
// note: return a float
return static_cast<float32_t>(*this) + static_cast<float32_t>(f);
}

// cast to float
host device
operator float32_t(void) const
{
return __half2float(s);
}
};

[1] http://en.wikipedia.org/wiki/Half_precision_floating-point_format

Forwarded from http://code.google.com/p/thrust/issues/detail?id=259

[marcinz]

Just a note, I was looking for complex<__half> support today. I guess this would be a logical continuation of adding float16_t.

[jrhemstad]

We will be adding extended floating point types to libcu++ in the near future. NVIDIA/cccl#1011