float vs double

[Matthias Kretz]

On Monday 20 August 2012 22:42:57 Kulakov, Igor wrote:
> From time to time we want to check our single precision algorithms in double
> precision. This is interesting from both precision point of view and time
> (scalability) point of view.
> 
> I think It would be nice (at least for us) if Vc provide easy way to do
> this.
> 
> What do you think?

Ah, you mean because of the Vector::Size? There are many possible ways to do 
this, and it very much depends on the problem which way is best for a fair 
comparison...

I'd rather not have a flag to redefine the meaning of float_v. I'm more 
thinking of a new type that will always have a size of max(float_v::Size, 
double_v::Size) and be implemented with different precisions depending on some 
flag.

Or even more flexible: provide a template class like this:
template<typename MemoryEntryType, typename OperationEntryType, int Size> 
class CustomVector

Then you could do this:

#ifdef USE_REAL_16_64
typedef Vc::CustomVector<float16, double, float_v::Size> real_v;
#elif defined(USE_REAL_32_64)
typedef Vc::CustomVector<float, double, float_v::Size> real_v;
#elif defined(USE_REAL_64)
typedef Vc::CustomVector<double, double, float_v::Size> real_v;
#elif defined(USE_REAL_16_32)
typedef Vc::CustomVector<float16, float, float_v::Size> real_v;
#else
typedef Vc::float_v real_v;
#endif

typedef real_v::EntryType real_t;
typedef real_v::Mask real_m;

That would require quite a bit of work in Vc, but would certainly be 
interesting for benchmarking.

I can't promise to work on this anytime soon, though. This could be a nice 
task for somebody to get into Vc hacking, though. :)

Cheers,
	Matthias
-- 
Dipl.-Phys. Matthias Kretz
http://compeng.uni-frankfurt.de/?mkretz

SIMD easy and portable: http://compeng.uni-frankfurt.de/?vc